content
stringlengths 1
15.9M
|
|---|
\section{Tree-like product of graphs}\label{sec-constr}
Let $G$ and $H$ be graphs, and let $|V(G)|=n$ and $V(H)=\{u_1,\ldots, u_k\}$.
Let $T_{n,k}$ denote the rooted tree of depth $k+1$ such that each vertex at depth at most $k$ has
precisely $n$ children (the \emph{depth} of the tree is the number of vertices of a longest path
starting with its root, and the depth of a vertex $x$ is the number of vertices of the path from the
root to $x$; i.e., the root has depth $1$). For each non-leaf vertex $x\in V(T_{n,k})$, let $G_x$ be a distinct copy of the graph $G$
and let $\theta_x$ be a bijection from $V(G_x)$ to the children of $x$ in $T_{n,k}$. If $v\in V(G_x)$, $y$ is a non-leaf
vertex of the subtree of $T_{n,k}$ rooted in $\theta_x(v)$, and $z\in V(G_y)$, then we say that $v$ is a \emph{progenitor} of $z$.
The \emph{level} of $v$ is defined to be the depth of $x$ in $T_{n,k}$. Note that a vertex at level $j$ has exactly one
progenitor at level $i$ for all positive $i<j$.
The graph $T(G,H)$ is obtained from the disjoint union of the graphs $G_x$ for non-leaf vertices $x\in V(T_{n,k})$
by, for each edge $u_iu_j\in E(H)$ with $i<j$, adding all edges from vertices of $T(G,H)$ at level $j$ to their
progenitors at level $i$.
Note that the graph $T(G,H)$ depends on the ordering of the vertices of $H$, which we consider to be fixed arbitrarily.
\begin{lemma}\label{lemma-propgen}
Let $G$ and $H$ be graphs with $|V(G)|=n\ge 2$ and $V(H)=\{u_1,\ldots, u_k\}$. Let $q\le k$ be the maximum integer
such that $\{u_{k-q+1},\ldots,u_k\}$ is an independent set in $H$. The graph $T(G,H)$ has $O(n^k)$ vertices
and $\omega(T(G,H))=\omega(G)+\omega(H)-1$. Furthermore, if $F$ is a minor of $T(G,H)$ and $F$ is $(k-q+1)$-connected,
then there exists a set $X\subseteq V(F)$ of size at most $k-q$ such that $F-X$ is a minor of $G$.
\end{lemma}
\begin{proof}
The tree $T_{n,k}$ has $1+n+n^2+\ldots+n^{k-1}\le 2n^{k-1}$ non-leaf vertices, and thus $|V(T(G,H))|\le 2n^k$.
Consider a clique $K$ in $T(G,H)$, and let $v$ be a vertex of $K$ of largest level. Let $x$ be the vertex of $T_{n,k}$
such that $v\in V(G_x)$. Note that all vertices of $K\setminus V(G_x)$ are progenitors of $v$, and the vertices of $H$
corresponding to their levels are pairwise adjacent. Consequently, $|K\setminus V(G_x)|\le \omega(H)-1$
and $|K\cap V(G_x)|\le \omega(G)$. Therefore, each clique in $T(G,H)$ has size at most $\omega(G)+\omega(H)-1$.
A converse argument shows that cliques $K_G$ in $G$ and $K_H$ in $H$
give rise to a clique in $T(G,H)$ of size $|K_G|+|K_H|-1$, implying that $\omega(T(G,H))=\omega(G)+\omega(H)-1$.
For $i=0,\ldots, k-q$, let $G_i$ denote the graph obtained from $G$ by adding $i$ universal vertices.
Observe that $T(G,H)$ is obtained from copies of $G_0$, \ldots, $G_{k-q}$ by clique-sums on cliques of size
at most $k-q$ (consisting of the progenitors whose level is most $k-q$). Hence, each $(k-q+1)$-connected minor $F$
of $T(G,H)$ is a minor of one of $G_0$, \ldots, $G_{k-q}$,
and thus a minor of $G$ can be obtained from $F$ by removing at most $k-q$ vertices.
\end{proof}
For an integer $p\ge 1$, the \emph{$p$-blowup} of a graph $H_0$ is the graph $H$ obtained from $H_0$ by replacing every
vertex $u$ by an independent set $S_u$ of $p$ vertices, and by adding all edges $zz'$ such that $z\in S_u$ and $z'\in S_{u'}$
for some $uu'\in E(H_0)$. For the purposes of constructing the graph $T(G,H)$, we order the vertices of $H$ so that
for each $u\in V(H_0)$, the vertices of $S_u$ are consecutive in the ordering.
The \emph{strong $p$-blowup} is obtained from the $p$-blowup by making the sets $S_u$ into cliques
for each $u\in V(H_0)$. For integers $a\ge b\ge 1$, an \emph{$(a:b)$-coloring} of $H_0$ is a function $\varphi$ that to each
vertex of $H_0$ assigns a subset of $\{1,\ldots, a\}$ of size $b$ such that $\varphi(u)\cap\varphi(v)=\emptyset$ for each edge
$uv$ of $H_0$. The \emph{fractional chromatic number} $\chi_f(H_0)$ is the infimum of $\{a/b:\text{$H_0$ has an $(a:b)$-coloring}\}$.
Note that if $H$ is the strong $p$-blowup of a graph $H_0$, then a $c$-coloring of $H$ gives a $(c:p)$-coloring of $H_0$.
Consequently, we have the following.
\begin{observation}\label{obs-sbl}
If $H$ is the strong $p$-blowup of a graph $H_0$, then $\chi(H)\ge p\chi_f(H_0)$.
\end{observation}
We now state a key result concerning the chromatic number of the graph $T(G,H)$.
\begin{lemma}\label{lemma-propcolor}
Let $p,c\ge 1$ be integers and let $G$ be a graph.
Let $H_0$ be a graph such that $\chi(H_0)=\chi_f(H_0)=c$, and let $H$ be the $p$-blowup of $H_0$.
Then
$$\chi(T(G,H))\le c\chi(G)$$
and if $\chi(G)\ge p$, then
$$\chi(T(G,H))\ge cp.$$
\end{lemma}
\begin{proof}
Let $V(H)=\{u_1,\ldots, u_k\}$, where $k=p|V(H_0)|$.
Note that $\chi(H)\le \chi(H_0)=c$. Let $\varphi_H$ be a proper coloring of $H$ using $c$ colors.
Let $C_1$, \ldots, $C_c$ be pairwise disjoint sets of $\chi(G)$ colors.
For each non-leaf vertex $x$ of $T_{n,k}$ of depth $i$, color $G_x$ properly using the colors in $C_{\varphi_H(u_i)}$.
Observe that this gives a proper coloring of $T(G,H)$ using at most $c\chi(G)$ colors,
and thus $\chi(T(G,H))\le c\chi(G)$.
Suppose now that $\chi(G)\ge p$ and consider a proper coloring $\varphi$ of $T(G,H)$.
Let $P=x_1x_2\ldots x_{k+1}$ be a path in $T_{n,k}$ from its root $x_1$ to one of the leaves and let $\psi$ be a coloring
of $H$ constructed as follows.
Suppose that we already selected $x_1$, \ldots, $x_i$ for some $i\le k$. Let $Z_i$ denote the set of progenitors
of level at least $i-p+1$ of the vertices of $G_{x_i}$. Since $|Z_i|\le p-1$ and $\varphi$ uses at least $\chi(G)\ge p$ distinct
colors on $G_{x_i}$, there exists $v\in V(G_{x_i})$ such that $\varphi(v)$ is different from the colors of all vertices of $Z_i$.
We define $x_{i+1}=\theta_{x_i}(v)$ be the child of $x_i$ in $T_{n,k}$ corresponding to $v$, and set $\psi(u_i)=\varphi(v)$.
Note that $\psi$ is a proper coloring of $H$ such that for each $u\in V(H_0)$, $\psi$ assigns vertices in $S_u$
pairwise distinct colors. Consequently, $\psi$ is a proper coloring of the strong $p$-blowup of $H_0$, and
thus $\psi$ (and $\varphi$) uses at least $p\chi_f(H_0)=cp$ distinct colors by Observation~\ref{obs-sbl}. We conclude that $\chi(T(G,H))\ge cp$.
\end{proof}
For positive integers $p$ and $k$, let $K_{k\times p}$ denote the $p$-blowup of $K_k$,
i.e., the complete $k$-partite graph with parts of size $p$. Let us summarize the results of this section
in the special case of the graph $T(G_0,K_{k\times 4})$ with $G_0$ planar.
\begin{corollary}\label{cor-prop}
Let $G_0$ be a planar graph with $n$ vertices and let $k_0$ be a positive integer. Let $G=T(G_0,K_{k_0\times 4})$.
The graph $G$ has $O(n^{4k_0})$ vertices. If $G_0$ is $3$-colorable, then $\chi(G)\le 3k_0$, and otherwise $\chi(G)\ge 4k_0$.
Furthermore, every $(4k_0-3)$-connected graph appearing as a minor in $G$ is $(4k_0-4)$-apex.
\end{corollary}
\begin{proof}
Note that $\chi(K_{k_0})=\chi_f(K_{k_0})=k_0$, $|V(K_{k_0\times 4})|=4k_0$ and the last $4$ vertices of $K_{k_0\times 4}$ form an independent set.
The claims follow from Lemma~\ref{lemma-propgen} (with $H=K_{k_0\times 4}$, $k=4k_0$ and $q=4$, using the fact that every minor of a planar graph is planar)
and Lemma~\ref{lemma-propcolor} (with $H_0=K_{k_0}$, $p=4$ and $c=k_0$).
\end{proof}
\section{Non-approximability}\label{sec-nonapprox}
The main non-approximability result is a simple consequence of Corollary~\ref{cor-prop} and NP-hardness
of testing $3$-colorability of planar graphs.
\begin{proof}[Proof of Theorem~\ref{thm-nonapprox}]
Suppose for a contradiction that there exists such a polynomial-time algorithm $\mathcal{A}$,
taking as an input an $F$-minor-free graph $G$ and returning an integer $c$ such that $\chi(G)\le c<\beta\chi(G)+(4-3\beta)k_0$.
Let $G_0$ be a planar graph, and let $G=T(G_0,K_{k_0\times 4})$. By Corollary~\ref{cor-prop},
the size of $G$ is polynomial in the size of $G_0$ and $G$ is $F$-minor-free.
Furthermore, if $G_0$ is $3$-colorable, then $\chi(G)\le 3k_0$, and otherwise $\chi(G)\ge 4k_0$.
Hence, if $G_0$ is $3$-colorable, then the value returned by the algorithm $\mathcal{A}$
applied to $G$ is less than $\beta\chi(G)+(4-3\beta)k_0\le 4k_0$, and if $G_0$ is not $3$-colorable, then the value returned is
at least $\chi(G)\ge 4k_0$. This gives a polynomial-time algorithm to decide whether $G_0$ is $3$-colorable.
However, it is NP-hard to decide whether a planar graph is $3$-colorable~\cite{garey1979computers},
which gives a contradiction unless $\text{P}=\text{NP}$.
\end{proof}
Note that the graphs $T(G_0,K_{k_0\times 4})$ used in the proof of Theorem~\ref{thm-nonapprox} have large cliques
(of size greater than $k_0$). This turns out not to be essential---we can prove somewhat weaker non-approximability result
even for graphs with clique number $3$.
To do so, we need to apply the construction with both $G$ and $H_0$ being triangle-free. A minor issue is that testing $3$-colorability
of triangle-free planar graphs is trivial by Gr\"otzsch' theorem~\cite{grotzsch1959}. However, this can be easily worked around.
\begin{lemma}\label{lemma-npcnotria}
Let $\mathcal{G}$ denote the class of graphs such that all their $3$-connected minors with at least $12$ vertices are planar.
The problem of deciding whether a triangle-free graph $G\in\mathcal{G}$ is $3$-colorable is NP-hard.
\end{lemma}
\begin{proof}
Let $R_0$ be the Gr\"otzsch graph ($R_0$ is a triangle-free graph with $11$ vertices and chromatic number $4$,
and all its proper subgraphs are $3$-colorable).
Let $R$ be a graph obtained from $R_0$ by removing any edge $uv$. Note that $R$ is $3$-colorable and the vertices
$u$ and $v$ have the same color in every $3$-coloring.
Let $G_1$ be a planar graph. Let $G_2$ be obtained from $G_1$ by replacing each edge $xy$ of $G_1$
by a copy of $R$ whose vertex $u$ is identified with $x$ and an edge is added between $v$ and $y$
(i.e., $G_2$ is obtained from $G_1$ by a sequence of Haj\'os sums with copies of $R_0$).
Clearly, $G_2$ is triangle-free, it is $3$-colorable if and only if $G_1$ is $3$-colorable, and
$|V(G_2)|=|V(G_1)|+10|E(G_1)|$.
Furthermore, $G_2$ is obtained from a planar graph by clique-sums with $R_0$ on cliques of size two,
and thus every $3$-connected minor of $G_2$ is either planar or a minor of $R_0$
(and thus has at most $11$ vertices). Hence, $G_2$ belongs to $\mathcal{G}$.
Since testing $3$-colorability of planar graphs is NP-hard, it follows that testing $3$-colorability of triangle-free graphs from $\mathcal{G}$
is NP-hard.
\end{proof}
We also need a graph $H_0$ which is triangle-free and its fractional chromatic number is large and equal to its ordinary
chromatic number. We are not aware of such a construction being previously studied; in Appendix, we prove the following.
\begin{lemma}\label{lemma-constr}
For every positive integer $m$, there exists a triangle-free graph $H_m$ with $O(m^4\log^2 m)$ vertices
such that $\chi(H_m)=\chi_f(H_m)=m$.
\end{lemma}
Theorem~\ref{thm-k4free} now follows in the same way as Theorem~\ref{thm-nonapprox}, using the graphs from Lemma~\ref{lemma-constr}
instead of cliques.
\begin{proof}[Proof of Theorem~\ref{thm-k4free}]
Suppose for a contradiction that there exists such a polynomial-time algorithm $\mathcal{A}$,
taking as an input an $F$-minor-free graph $G$ with $\omega(G)\le 3$
and returning an integer $c$ such that $\chi(G)\le c<\beta\chi(G)+d$.
Recall that $m=\lceil d/(4-3\beta)\rceil$.
Let $H_m$ be the graph from Lemma~\ref{lemma-constr}. Let $k_0=|V(H_m)|$ and
let $H$ be the $4$-blowup of $H_m$. Let $\mathcal{G}$ be the class of graphs such that all their $3$-connected minors with at least $12$
vertices are planar.
Consider a triangle-free graph $G_0\in \mathcal{G}$, and let $G=T(G_0,H)$. By Lemma~\ref{lemma-propgen},
the size of $G$ is polynomial in the size of $G_0$. Consider any $(4k_0-1)$-connected minor $F'$ of $G$.
By Lemma~\ref{lemma-propgen}, there exists a set $X$ of size at most $4k_0-4$ such that $F'-X$ is a minor of $G_0$.
Since $F'-X$ is $3$-connected and $G_0\in \mathcal{G}$, we conclude that either $|V(F')|\le |X|+11\le 4k_0+7$ or $F'-X$ is planar.
Consequently, $F'\neq F$, and thus $G$ does not contain $F$ as a minor.
Furthermore, $\omega(G)\le\omega(G_0)+\omega(H)-1\le 3$.
Recall that $\chi(H_m)=\chi_f(H_m)=m$.
By Lemma~\ref{lemma-propcolor}, if $G_0$ is $3$-colorable, then $\chi(G)\le 3m$, and otherwise $\chi(G)\ge 4m$.
Hence, if $G_0$ is $3$-colorable, then the value returned by the algorithm $\mathcal{A}$
applied to $G$ is less than $\beta\chi(G)+d\le 3\beta m+d\le 4m$, and if $G_0$ is not $3$-colorable, then the value returned is
at least $\chi(G)\ge 4m$. This gives a polynomial-time algorithm to decide whether $G_0$ is $3$-colorable,
in contradiction to Lemma~\ref{lemma-npcnotria} unless $\text{P}=\text{NP}$.
\end{proof}
\section{Approximation algorithms}\label{sec-approx}
Let us now turn our attention to the additive approximation algorithms.
The algorithms we present use ideas similar to the ones of the $2$-approximation algorithm of Demaine et al.~\cite{demaine2005algorithmic}
and of the additive approximation algorithms of Kawarabayashi et al.~\cite{kawarabayashi2009additive} and Demaine et al.~\cite{demaine2009approximation}.
We find a partition of the vertex set of the input graph $G$ into parts $L$ and $C$ such that $G[L]$ has bounded tree-width
(and thus its chromatic number can be determined exactly) and $G[C]$ has bounded chromatic number, and color the parts
using disjoint sets of colors.
The existence of such a decomposition is proved using the minor structure theorem~\cite{robertson2003graph}, in the variant
limiting the way apex vertices attach to the surface part of the decomposition. The proof of this stronger version can be
found in~\cite{apex}. Let us now introduce definitions necessary to state this variant of the structure theorem.
A \emph{tree decomposition} of a graph $G$ is a pair $(T,\beta)$, where $T$ is a tree and $\beta$ is a function
assigning to each vertex of $T$ a subset of $V(G)$, such that for each $uv\in E(G)$ there exists $z\in V(T)$ with
$\{u,v\}\subseteq \beta(z)$, and such that for each $v\in V(G)$, the set $\{z\in V(T):v\in\beta(z)\}$ induces a non-empty
connected subtree of $T$. The \emph{width} of the tree decomposition is $\max\{|\beta(z)|:z\in V(T)\}-1$,
and the \emph{tree-width} of a graph is the minimum of the widths of its tree decompositions.
The decomposition is \emph{rooted} if $T$ is rooted.
For a rooted tree decomposition $(T,\beta)$ and a vertex $z$ of $T$ distinct from the root,
if $w$ is the parent of $z$ in $T$, we write $\beta\uparrow z\colonequals \beta(z)\cap\beta(w)$ and
$\beta\downarrow z\colonequals \beta(z)\setminus \beta(w)$. If $z$ is the root of $T$, then
$\beta\uparrow z\colonequals\emptyset$ and $\beta\downarrow z\colonequals\beta(z)$.
The \emph{torso expansion} of a graph $G$ with respect to its rooted tree decomposition $(T,\beta)$
is the graph obtained from $G$ by adding edges of cliques on $\beta\uparrow z$ for all $z\in V(T)$.
A \emph{path decomposition} is a tree decomposition $(T,\beta)$ where $T$ is a path.
A \emph{vortex with boundary sequence $v_1$, \ldots, $v_s$ and depth $d$} is a graph with a path decomposition
$(p_1p_2\ldots p_s,\beta)$ such that $|\beta(p_i)|\le d+1$ and $v_i\in \beta(p_i)$ for $i=1,\ldots,s$.
An embedding of a graph in a surface is \emph{$2$-cell} if each face of the embedding is homeomorphic to an open disk.
\begin{theorem}\label{thm-struct}
For every non-negative integer $t$ and a $t$-apex graph $H$, there exists a constant $a_H$ such that the following holds.
For every $H$-minor-free graph $G$, there exists a rooted tree decomposition $(T,\beta)$ of $G$ with the following properties.
Let $G'$ denote the torso expansion of $G$ with respect to $(T,\beta)$.
For every $v\in V(T)$, there exists a set $A_v\subseteq \beta(v)$ of size at most $a_H$ with $\beta\uparrow v\subseteq A_v$,
a set $A'_v\subseteq A_v$ of size at most $t-1$, and
subgraphs $G_v$, $G_{v,1}$, \ldots, $G_{v,m}$ of $G'[\beta(v)\setminus A_v]$ for some $m\le a_H$ such that
\begin{itemize}
\item[(a)] $G'[\beta(v)\setminus A_v]=G_v\cup G_{v,1}\cup \ldots\cup G_{v,m}$, and for $1\le i<j\le m$,
the graphs $G_{v,i}$ and $G_{v,j}$ are vertex-disjoint and $G'$ contains no edges between $V(G_{v,i})$ and $V(G_{v,j})$,
\item[(b)] the graph $G_v$ is $2$-cell embedded in a surface $\Sigma_v$ in that $H$ cannot be drawn,
\item[(c)] for $1\le i\le m$, $G_{v,i}$ is a vortex of depth $a_H$ intersecting $G_v$ only in its boundary sequence,
and this sequence appears in order in the boundary of a face $f_{v,i}$ of $G_v$, and $f_{v,i}\neq f_{v,j}$ for $1\le i<j\le m$,
\item[(d)] vertices of $G_v$ have no neighbors in $A_v\setminus A'_v$, and
\item[(e)] if $w$ is a child of $v$ in $T$ and $\beta(w)\cap V(G_v)\neq \emptyset$, then
$\beta\uparrow w\subseteq V(G_v)\cup A'_v$.
\end{itemize}
Furthermore, the tree decomposition and the sets and subgraphs as described can be found in polynomial time.
\end{theorem}
Informally, the graph $G$ is a clique-sum of the graphs $G'[\beta(v)]$ for $v\in V(T)$, and $\beta(v)$ contains a bounded-size set $A_v$
of apex vertices such that $G'[\beta(v)]-A_v$ can be embedded in $\Sigma_v$ up to a bounded number of vortices of bounded depth.
Furthermore, at most $t-1$ of the apex vertices (forming the set $A'_v$) can have neighbors in the part $G_v$ of $G'[\beta(v)]-A_v$
drawn in $\Sigma_v$, or in the other bags of the decomposition that intersect $G_v$. Note that it is also possible that
$\Sigma_v$ is the null surface, and consequently $A_v=\beta(v)$.
We need the following observation on graphs embedded up to vortices.
\begin{lemma}\label{lemma-remove}
Let $\Sigma$ be a surface of Euler genus $g$ and let $a$ be a non-negative integer. Let $G$ be a graph and let $G_0$, $G_1$, \ldots, $G_m$
be its subgraphs such that $G=G_0\cup G_1\cup \ldots \cup G_m$, the subgraphs $G_1$, \ldots, $G_m$
are pairwise vertex-disjoint and $G$ contains no edges between them, $G_0$ is $2$-cell embedded in $\Sigma$,
and there exist pairwise distinct faces $f_1$, \ldots, $f_m$ of this embedding such that for $1\le i\le m$,
$G_i$ intersects $G_0$ only in a set of vertices contained in the boundary of $f_i$.
If the graphs $G_1$, \ldots, $G_m$ have tree-width at most $a$, then there exists a subset $L_0$ of vertices of $G_0$
such that $G_0-L_0$ is planar and the graph $G_0[L_0]\cup G_1\cup\ldots\cup G_m$ has tree-width at most $26g+9m+a$.
\end{lemma}
\begin{proof}
If $\Sigma$ is the sphere, then we can set $L_0=\emptyset$; hence, we can assume that $g\ge 1$.
Let $G'_0$ be the graph obtained from $G_0$ by, for $1\le i\le m$, adding a new vertex $r_i$ drawn inside $f_i$
and joined by edges to all vertices of $V(G_0)\cap V(G_i)$. Note that $G'_0$ has a $2$-cell embedding in $\Sigma$ extending the embedding of $G$.
Let $F$ be a subgraph of $G'_0$ such that the embedding of $F$ in $\Sigma$ inherited from the embedding of $G'_0$
is $2$-cell and $|V(F)|+|E(F)|$ is minimum. Then $F$ has only one face, since otherwise it is possible to remove an edge separating distinct faces of $F$,
and $F$ has minimum degree at least two, since otherwise we can remove a vertex of degree at most $1$ from $F$.
By generalized Euler's formula, we have $|E(F)|=|V(F)|+g-1$, and thus $F$ contains at most $2(g-1)$ vertices of degree
greater than two. By considering the graph obtained from $F$ by suppressing vertices of degree two, we see that
$F$ is either a cycle (if $g=1$) or a subdivision of a graph with at most $3(g-1)$ edges.
Let $M_0$ be the set of vertices of $F$ of degree at least three and their neighbors.
We claim that each vertex of $V(F)\setminus M_0$ is adjacent in $G'_0$ to only two vertices of $V(F)$.
Indeed, suppose that a vertex $x\in V(F)\setminus M_0$ has at least three neighbors in $G'_0$ belonging to $V(F)$.
Let $w$ and $y$ be the neighbors of $x$ in $F$, and let $z$ be a vertex distinct from $w$ and $y$ adjacent to $x$ in $G'_0$.
The graph $F+xz$ has two faces, and by symmetry, we can assume that the edge $xy$ separates them. Since $x\not\in M_0$,
the vertex $y$ has degree two in $F$, and thus the embedding of $F+xz-y$ is $2$-cell, contradicting the minimality of $F$.
Let $M_1$ be the set of vertices of $F$ at distance at most $4$ from a vertex of degree greater than two. Note that $|M_1|\le 26g$.
For $1\le i\le m$, let $N_i$ denote the set of vertices of $F-M_1$ that are in $G'_0$ adjacent to a vertex of $V(G_i)\cap V(G_0)\setminus V(F)$.
We claim that $|N_i|\le 9$. Indeed, suppose for a contradiction that $|N_i|\ge 10$ and consider a path $w_4w_3w_2w_1xy_1y_2y_3y_4$ in $F$ such that
$x$ belongs to $N_i$ (vertices $x$, $w_1$, \ldots, $w_4$, $y_1$, \ldots, $y_4$ have degree two in $F$, since $x\not\in M_1$).
If $r_i\in V(F)$, then let $Q$ be a path in $G'_0$ of length two between $x$ and $r_i$ through a vertex of $V(G_i)\cap V(G_0)\setminus V(F)$; note that
$r_i\not\in\{w_1,w_2,y_1,y_2\}$, since $r_i$ has at least $10$ neighbors in $V(F)$ and belongs to $M_0$ by the previous paragraph
and $x\not\in M_1$. If $r_i\not\in V(F)$, then there exists a vertex $z\in N_i\setminus \{x,w_1,\ldots, w_4,y_1,\ldots,y_4\}$;
we let $Q$ be a path of length at most four between $x$ and $z$ passing only through their neighbors in $V(G_i)\cap V(G_0)\setminus V(F)$ and possibly through $r_i$.
By symmetry, we can assume that the edge $xy_1$ separates the two faces of $F+Q$, and the graph
$F+Q-\{y_1,y_2\}$ if $r_i\in V(F)$ or $F+Q-\{y_1,y_2,y_3,y_4\}$ if $r_i\not\in V(F)$ contradicts the minimality of $F$.
Let $M=(M_1\cup \bigcup_{i=1}^m N_i)\cap V(G_0)$; we have $|M|\le 26g+9m$. Let $L_0=V(F)\cap V(G_0)$.
Clearly, $G_0-L_0\subseteq G'_0-V(F)$ is planar. Let $G'$ be the graph obtained from $G_1\cup\ldots\cup G_m$ by adding
vertices of $M$ as universal ones, adjacent to all other vertices of $G'$. The tree-width of $G'$ is at most $a+|M|\le 26g+9m+a$.
Note that each vertex of $L_0\setminus V(G')$ has degree two in $G_0[L_0]\cup G_1\cup\ldots\cup G_m$,
and thus $G_0[L_0]\cup G_1\cup\ldots\cup G_m$ is a subgraph of a subdivision of $G'$. We conclude that
the tree-width of $G_0[L_0]\cup G_1\cup\ldots\cup G_m$ is also at most $26g+9m+a$.
\end{proof}
Note that the set $L_0$ can be found in polynomial time. For the clarity of presentation of the proof we selected $F$ with $|V(F)|+|E(F)|$
minimum; however, it is sufficient to start with an arbitrary inclusionwise-minimal subgraph with exactly one $2$-cell face
(obtained by repeatedly removing edges that separate distinct faces and vertices of degree at most $1$) and
repeatedly perform the reductions described in the proof until each vertex of $V(F)\setminus M_0$ is adjacent in $G'_0$ to only
two vertices of $V(F)$ and until $|N_i|\le 9$ for $1\le i\le m$.
For positive integers $t$ and $a$, we say that a rooted tree decomposition $(T,\beta)$ of a graph $G$ is \emph{$(t,a)$-restricted}
if for each vertex $v$ of $T$, the subgraph of the torso expansion of $G$ induced by $\beta\downarrow v$
is planar, $|\beta\uparrow v|\le t-1$, and each vertex of $\beta\downarrow v$ has at most $a-1$ neighbors in $G$ that belong to
$\beta\uparrow v$. Using the decomposition from Theorem~\ref{thm-struct}, we now partition the considered graph
into a part of bounded tree-width and a $(t,t)$-restricted part.
\begin{theorem}\label{thm-decomp}
For every positive integer $t$ and a $t$-apex graph $H$, there exists a constant $c_H$ with the following property.
The vertex set of any $H$-minor-free graph $G$ can be partitioned in polynomial time into two parts $L$ and $C$ such that
$G[L]$ has tree-width at most $c_H$ and $G[C]$ has a $(t,t)$-restricted rooted tree decomposition.
Additionally, for any such graph $H$ and positive integers $a\le b$, there exists a constant $c_{H,a,b}$ such that
if $G$ is $H$-minor-free and does not contain $K_{a,b}$ as a subgraph, then $L$ and $C$ can be chosen so that $G[L]$ has tree-width at
most $c_{H,a,b}$ and $G[C]$ has a $(t,a)$-restricted rooted tree decomposition.
\end{theorem}
\begin{proof}
Let $a_H$ be the constant from Theorem~\ref{thm-struct} for $H$, and let $g$ be the maximum Euler genus of a surface in that
$H$ cannot be embedded. Let $c_H=26g+11a_H$ and $c_{H,a,b}=c_H+\binom{t-1}{a}(b-1)$.
Since $G$ is $H$-minor-free, we can in polynomial time find its rooted tree decomposition $(T,\beta)$, its torso expansion $G'$,
and for each $v\in V(T)$, find $A_v$, $A'_v$, $G_v$, $G_{v,1}$, \ldots, $G_{v,m}$, and $\Sigma_v$ as described
in Theorem~\ref{thm-struct}. Let $L'_v$ be the set of vertices obtained by applying Lemma~\ref{lemma-remove}
to $G_v$, $G_{v,1}$, \ldots, $G_{v,m}$; i.e., $G_v-L'_v$ is planar and the graph
$G_v[L'_v]\cup G_{v,1}\cup\ldots\cup G_{v,m}$ has tree-width at most $26g+10a_H$.
When considering the case that $G$ does not contain $K_{a,b}$ as a subgraph, let $S_v$ be the set of vertices of $G_v$ that have at least $a$ neighbors
in $G$ belonging to $A_v$ (and thus to $A'_v$); otherwise, let $S_v=\emptyset$.
Since there are at most $\binom{t-1}{a}$ ways how to choose a set of $a$ neighbors in $A'_v$ and no $b$ vertices can have the
same set of $a$ neighbors in $A'_v$, we have $|S_v|\le \binom{t-1}{a}(b-1)$.
Let $L_v=(A_v\setminus \beta\uparrow v)\cup V(G_{v,1}\cup \ldots\cup G_{v,m})\cup L'_v\cup S_v$.
We define $L=\bigcup_{v\in V(T)} L_v$. Note that $L\cap \beta(v)\subseteq L_v\cup (\beta\uparrow v)\subseteq L_v\cup A_v$.
Consequently, $G'[L\cap \beta(v)]$ is obtained from $G_v[L'_v]\cup G_{v,1}\cup\ldots\cup G_{v,m}$ by adding $S_v$ and some of the
vertices of $A_v$, and consequently $G'[L\cap \beta(v)]$ has tree-width at most $c_{H,a,b}$ when considering the case that $G$ does not contain $K_{a,b}$ as a subgraph
and at most $c_H$ otherwise. The graph $G[L]$ is a clique-sum of the graphs
$G'[L\cap\beta(v)]$ for $v\in V(T)$, and thus the tree-width of $G[L]$ is also at most $c_{H,a,b}$ or $c_H$.
Let $C=V(G)\setminus L$, and consider the graph $G[C]$. For $v\in V(T)$, let $\beta'(v)=\beta(v)\cap C$.
Then $(T,\beta')$ is a rooted tree decomposition of $G[C]$ such that for every $v\in V(T)$,
the graph $G'[\beta'\downarrow v]\subseteq G_v-L'_v$ is planar, all vertices of $\beta'\uparrow v$ adjacent in $G$
to a vertex of $\beta'\downarrow v$ belong to $A'_v$ (and thus there are at most $t-1$ such vertices), and
when considering the case that $G$ does not contain $K_{a,b}$ as a subgraph, each vertex of $\beta'\downarrow v$ has at most $a-1$
neighbors in $G$ belonging to $\beta'\uparrow v$.
Note that $\beta'\uparrow v$ can contain vertices not belonging
to $A'_v$, and thus $\beta'\uparrow v$ can have size larger than $t-1$, and
the tree decomposition $(T,\beta')$ is not necessarily $(t,t)$-restricted.
However, by the condition (e) from the statement of Theorem~\ref{thm-struct},
the vertices of $(\beta'\uparrow v)\setminus A'_v$ can only be contained in the
bags of descendants of $v$ which are disjoint from $V(G_v)$, and thus we can
fix up this issue as follows.
If $w$ is a child of $v$ and $\beta'(w)\cap V(G_v)=\emptyset$, we say that the edge $vw$ is \emph{skippable}; note that
in that case $\beta'\uparrow w\subseteq \beta'\uparrow v$. For each vertex $w$ of $T$, let $f(w)$ be the nearest
ancestor of $w$ such that the first edge on the path from $f(w)$ to $w$ in $T$ is not skippable.
Let $T'$ be the rooted tree with vertex set $V(T)$ where the parent of each vertex $w$ is $f(w)$. Observe that $(T',\beta')$
is a tree decomposition of $G[C]$. Furthermore, denoting by $z$ the child of $f(w)$ on
the path from $f(w)$ to $w$ in $T$, note that if a vertex $x\in \beta'\uparrow f(w)$ is contained in $\beta'(w)$, then
$x\in \beta(z)$, and since the edge $f(w)z$ is not skippable, the condition (e) from the statement of Theorem~\ref{thm-struct}
implies that $x\in A'_{f(w)}$.
Hence, letting $\beta''(v)=(\beta'\downarrow v)\cup (A'_v\cap C)$
for each vertex $v$ of $T'$, we conclude that $(T',\beta'')$ is a rooted tree decomposition of $G[C]$ which
is $(t,t)$-restricted, and when considering the case that $G$ does not contain $K_{a,b}$ as a subgraph, the decomposition is $(t,a)$-restricted.
\end{proof}
Let us now consider the chromatic number of graphs with a $(t,a)$-restricted tree decomposition.
\begin{lemma}\label{lemma-color}
Let $a$ and $t$ be positive integers.
Let $G$ be a graph with a $(t,a)$-restricted rooted tree decomposition $(T,\beta)$.
The chromatic number of $G$ is at most $\min(t+3,a+4)$.
Additionally, if $G$ is triangle-free, then the chromatic number of $G$ is at most $\lceil (13t+172)/14\rceil$.
\end{lemma}
\begin{proof}
We can color $G$ using $t+3$ colors, starting from the root of the tree decomposition, as follows.
Suppose that we are considering a vertex $v\in V(T)$ such that $\beta\uparrow v$ is already colored.
Since $|\beta\uparrow v|\le t-1$, this leaves at least $4$ other colors to be used on
$G[\beta\downarrow v]$. Hence, we can extend the coloring to $G[\beta\downarrow v]$ by the Four Color Theorem.
We can also color $G$ using $a+4$ colors, starting from the root of the tree decomposition, as follows.
For each vertex $x$ of $\beta\downarrow v$, at most $a-1$ colors are used on its neighbors in
$\beta\uparrow v$, leaving $x$ with at least $5$ available colors not appearing on its neighbors. Since
$G[\beta\downarrow v]$ is planar, we can color it from these lists of size at least $5$ using the result
of Thomassen~\cite{thomassen1994}, again extending the coloring to $G[\beta\downarrow v]$.
Finally, suppose that $G$ is triangle-free. Let $G'$ be the torso expansion of $G$ with respect to $(T,\beta)$, and let $c=\lceil (13t+172)/14\rceil$.
We again color $G$ starting from the root of the tree decomposition using at most
$c$ colors. Additionally, we choose the coloring so that the following invariant is satisfied: ($\star$) for each vertex $w$ of $T$
and for each independent set $I$ in $G[\beta(w)]$ such that $I\cap \beta\downarrow w$ induces a clique in $G'$,
at most $c-6$ distinct colors are used on $I$.
Let $v$ be a vertex of $T$. Suppose we have already colored
$\beta\uparrow v$, and we want to extend the coloring to $\beta\downarrow v$.
Note that the choice of this coloring may only affect the validity of the invariant ($\star$) at $v$ and
at descendants of $v$ in $T$. Consider any descendant $w$ of $T$. Coloring $\beta\downarrow v$
can only assign color to vertices of $\beta\uparrow w$, and since
$G'$ is the torso expansion of $G$, the set $\beta\uparrow w\cap \beta\downarrow v$ induces a clique in $G'$.
Consequently, the validity of ($\star$) at $v$ implies the validity at $w$ (until more vertices of $G$ are assigned colors),
and thus when choosing the coloring of $\beta\downarrow v$, we only need to ensure that ($\star$) holds at $v$.
The graph $G'[\beta\downarrow v]$ is planar, and thus it is $5$-degenerate; i.e., there exists
an ordering of its vertices such that each vertex is preceded by at most $5$ of its neighbors.
Let us color the vertices of $\beta\downarrow v$ according to this ordering, always preserving the validity
of ($\star$) at $v$. Suppose that we are choosing a color for a vertex $x\in V(\beta\downarrow v)$. Let $P_x$
consist of the neighbors of $x$ in $G'$ belonging to $\beta\downarrow v$ that precede it in the ordering;
we have $|P_x|\le 5$. Note that all cliques in $G'[\beta\downarrow v]$ containing $x$ and with all other vertices already colored
are subsets of $P_x\cup \{x\}$. Let $Q_x=P_x\cup \beta\uparrow v$; we have $|Q_x|\le t+4$.
Let $N_x$ consist of vertices of $Q_x$ that are adjacent to $x$ in $G$.
We say that a color $a$ is forbidden at $x$ if there exists an independent set
$A_a\subseteq Q_x\setminus N_x$ of $G$ such that $A_a\cap P_x$ is a clique in $G'$ and $c-6$ colors distinct from $a$ appear on $A_a$.
Observe that assigning $x$ a color which neither appears on $N_x$ nor is forbidden results
in a proper coloring that preserves the invariant ($\star$) at $v$.
Suppose first that no color is forbidden at $x$. Since $G$ is triangle-free, $N_x\setminus P_x$ is an independent set in
$G[\beta\uparrow v]$, and by ($\star$), at most $c-6$ colors appear on $N_x\setminus P_x$. Since $|P_x|\le 5$, it follows that
some color does not appear on $N_x$, as required.
Hence, we can assume that some color is forbidden at $x$, and thus there exists an independent
set $Z_1\subseteq Q_x\setminus N_x$ of size at least $c-6$ such that vertices of $Z_1$ are assigned pairwise distinct colors.
Since $|Q_x|\le t+4$, at most $t+4-(c-6)=t+10-c$ of these colors appear at least twice
on $Q_x$, and thus there exists a set $Z_2\subseteq Z_1$ of size at least $c-6-(t+10-c)=2c-t-16$
such that the color of each vertex of $Z_2$ appears exactly once on $Q_x$ (and thus does not appear on $N_x$). Let $Z=Z_2\setminus P_x$;
we have $|Z|\ge 2c-t-21$. We claim that not all colors appearing on $Z$ are forbidden; since
such colors do not appear on $N_x$, we can use them to color $x$.
For contradiction, assume that colors of all vertices of $Z$ are forbidden at $x$. Let $Z=\{z_1,\ldots, z_m\}$
for some $m\ge 2c-t-21$, and for $a=1,\ldots,m$, let $a$ be the color of $z_a$. Since $a$ is forbidden at $x$,
there exists an independent set $A_a\subseteq Q_x\setminus N_x$ such that $A_a\cap P_x$ is a clique in $G'$ and $c-6$ colors
distinct from $a$ appear on $A_a$. Note that $A_a\cup \{z_a\}$ is not an independent set, as otherwise
this set contradicts the invariant ($\star$) at $v$. Hence, we can choose a neighbor $f(a)$ of $z_a$ in $A_a$.
Since $Z$ is an independent set, we have $f(a)\not\in Z$. Furthermore, we claim that the preimage in $f$ of each vertex
has size at most $6$: if say $f(z_1)=\ldots=f(z_7)=y$, then for $i=1,\ldots,7$, the vertex $z_i$ would have a neighbor $y$ in the independent set
$A_1$, and thus $z_1,\ldots,z_7\not\in A_1$; however, the only appearance of colors $1$, \ldots, $7$ in $Q_x$ is on the vertices
$z_1$, \ldots, $z_7$, and thus at most $c-7$ colors would appear on $A_1$. We conclude that $|f(Z)|\ge |Z|/6$,
and thus $t+4\ge |Q_x|\ge |Z|+|f(Z)|\ge \tfrac{7}{6}|Z|\ge \tfrac{7}{6}(2c-t-21)\ge (6t+25)/6$. This is a contradiction.
\end{proof}
Combining Lemma~\ref{lemma-color} with Theorem~\ref{thm-decomp} (coloring $G[L]$ using at most $\chi(G)$ colors in linear time~\cite{coltweasy},
using the fact that $G[L]$ has bounded tree-width, and coloring $G[C]$ using a disjoint set of colors),
we obtain Theorems~\ref{thm-alg}, \ref{thm-algnotri}, and \ref{thm-algnobip}.
\bibliographystyle{siam}
|
\section{Introduction}
The theoretical prediction and the subsequent experimental discovery of topological insulators \citep{wsm1,wsm45,wsm46,wsm47,wsm48,wsm49}revolutionized the band theory of solids. Topological insulators exhibit a full spectral gap in the bulk but robust gapless conducting states with non-degenerate spin texture protected by topology on the surface. Such topological systems with full spectral gap in the bulk have been classified into ten symmetry classes based on the presence or absence of discrete non-spatial symmetries such as time reversal, particle-hole, and the chiral symmetry \citep{wsm39, wsm40, wsm41, wsm42}. It has also been shown recently \citep{wsm3,wsm4,wsm5,wsm6,wsm7,wsm8,xu2015,wsm9,wsm10,wsm11,wsm12,wsm13,wsm14} that topological states of matter can exist even when the bulk spectrum is gapless, and these systems fall outside the ten-fold classification of topological insulators and superconductors. Three dimensional topological Weyl semimetals (WSM), a new state of quantum matter with gapless spectrum at bulk nodal points and spin non-degenerate open Fermi arc states on the surface, fall into this class and have recently sparked enormous interest in condensed matter physics \citep{wsm3,wsm4,wsm5,wsm6,wsm7,wsm8,xu2015,wsm9,wsm10,wsm11,wsm12,wsm13,wsm14}.
In WSMs a pair of non-degenerate energy bands touch at isolated points in the momentum space, and the low energy excitations near the band touching points, known as Weyl points, disperse linearly along all three momentum directions. The Weyl points act as source and sink of Abelian Berry curvature, which is an analog of magnetic field but defined in the momentum space \citep{wsm21,wsm22,wsm23,wsm24}. Weyl semimetals are different from Dirac semimetals, which are topologically protected in the presence of time reversal, space inversion and a lattice symmetry of the underlying crystal, in that WSMs violate space inversion and/or time reversal symmetry and are topologically protected by a non-zero quantized flux of Berry curvature across the Fermi surface known as the Chern number \citep{wsm34,volovik}. The Chern number is equal to the strength of the Berry magnetic monopole enclosed by the Fermi surface, while the monopole charge (also called chirality) summed over all the Weyl points in the Brillouin zone vanishes \citep{wsm32,wsm33}. In recent studies several materials have been theoretically predicted to be Weyl semimetals, with experimental confirmation in several candidate materials such as, TaAs, NbAs, TaP, NbP \citep{wsm5,wsm6,wsm7,wsm8,wsm28,wsm29,wsm30,wsm31}.
In WSMs violation of Lorentz invariance, existence of non-trivial Berry curvature in the momentum space, and the (approximate) separate conservation of the number of Weyl electrons of different chiralities, lead to many anomalous transport and optical properties, such as anomalous Hall effect, dynamic chiral magnetic effect which is related to optical gyrotropy, and most importantly, nagative longitudional magnetoresistance in the presence of parallel electric and magnetic fields, due to non-conservation of separate electron numbers of opposite chirality for relativistic massless Fermions, an effect known as the chiral Adler-Bell-Jackiw anomaly \citep{volovik,wsm3,xu2011,wsm32,wsm33,bell1969,xu2011,wsm13,goswami2015,aji2012,adler1969}.
In recent work \citep{xu2015,wsm9} it has been proposed that based on the symmetry and fermiology WSMs can be broadly classified into two types, type-I and type-II Weyl semimetals. While the conventional type-I WSMs have point-like Fermi surface and vanishing density of states at the Fermi energy, the WSMs of type-II break Lorentz symmetry explicitly, resulting in a tilted conical spectra with electron and hole pockets producing finite density of states at the Fermi level \citep{xu2015,wsm9,wsm44}. The tilting can be generated in many different ways, e.g., by change in chemical doping or strain in different directions \citep{wsm37}. The tilted conical spectra and the finite density of states at the Fermi level in type-II WSMs have been shown to produce interesting effects such as chiral anomaly induced longitudinal magnetoresistance which is strongly anisotropic in space and a novel anomalous Hall effect\citep{wsm16,wsm17}. In this work we consider the anomalous Nernst effect in type-II Weyl semimetals in the framework of semi-classical Boltzmann theory\citep{wsm16,wsm19,wsm15,lundgren2014,kim2014,son2013}. Based on both a linearized dispersion model of time-reversal breaking WSMs and the corresponding lattice model, we show that the anomalous Nernst response in type-II WSMs is finite, sharply increases with decreasing chemical potential, and is strongly anisotropic in space, which can serve as a reliable signature of type-II Weyl semimetals in a host of magnetic systems with spontaneously broken time reversal symmetry.
The rest of the paper is organized as follows: In section II we introduce a linearized dispersion model for type-II WSMs with a higher energy cut-off. In section III we briefly discuss the derivation of the semiclassical Boltzmann distribution formula which we have used for calculating the anomalous Nernst response. The results of our calculation of Nernst signal and anisotropy have been included in section IV, where we have considered both the low energy linearized Hamiltonian of a type-II WSM as well as a more realistic lattice model. We include discussion of our central results in section IV. We end with a brief conclusion in section V.
\section{Low energy linearized Hamiltonian for topological WSM}
The low-energy linearized model for a time reversal broken type-II WSM with a single pair of Weyl points separated and tilted along the $k_z$ direction is described by the Hamiltonian,
\begin{equation}
H_{1,2}(\bm{k})=\hslash C_{1,2}(k_{z}\mp Q)\mp \hslash v \bm{\sigma}.(\bm{k}\mp Q\bm{e_{z}})
\label{Eq:H1}
\end{equation}
\begin{figure}[h]
\begin{frame}{}
\hbox{\hspace{-3ex}\includegraphics[width=.52\textwidth]{weyl1.png}}
\end{frame}
\caption{Schematic illustrations for type-I WSM (a) Two untilted Weyl cones of type-I with the Weyl points located at $+Q$ and $-Q$ along the $k_z$ axis and tilting parameters $C_1=C_2=0$. (b) Weyl cones of type-I located at $\pm Q$ along the $k_z$ axis with tilting parameters $C_1/v=0.5, C_2/v=-0.5$. (c) Weyl cones of type-I located as before with tilting parameters $C_1/v=-0.5, C_2/v=0.5$}
\label{Fig:TypeI}
\end{figure}
\begin{figure}[h]
\begin{frame}{}
\hbox{\hspace{-1ex}\includegraphics[width=.52\textwidth]{weyl2.png}}
\end{frame}
\caption{Schematic illustrations for type-II Weyl cones. (a) and (b) are at the critical point between transition from type-I to type-II WSM with Weyl cones just touching the Fermi surface. Tilting parameters are taken as $C_1/v=1$ and $C_2/v=-1$ (a) and $C_1/v=-1, C_2/v=1$ (b). (c) and (d) are WSMs of type II. Weyl cones have gone through the Fermi surface and there is finite density of states at the Fermi level arising from electron and hole pockets coexisting with Weyl points. Tilting parameters are taken $C_1/v=1.3, C_2/v=-1.3$ (c) and $C_1/v=-1.3, C_2/v=1.3$ (d)}
\label{Fig:TypeII}
\end{figure}
where $\bm\sigma$'s are the Pauli matrices and $\sigma_0$ is the $2\times 2$ unit matrix. As unbounded linear dispersion model is not realistic we have used higher energy and momentum cut-offs in the following calculations. Fermi velocity of electrons in the absence of tilting is written as $v$. For numerical calculations the Fermi velocity $v$ is chosen as $10^6 $ m/s and the energy cutoff taken as $0.3$ eV. In Eq.~(\ref{Eq:H1}) the two Weyl points of opposite chirality are located at $(0,0,\pm Q)$ in the momentum space. We assume the momentum space splitting between the Weyl points as $Q=5 $ nm. The type of Weyl points is defined by the tilting parameters $C_{1,2}$. For $|C_{1,2}| < v$ the Weyl cones do not touch the Fermi surface. So for these values of the tilting parameters the system is a WSM of type-I. When $|C_{1,2}| = v$, the Weyl cones touch Fermi surface and is the critical point between transition from WSM-I to WSM-II. Finally, for $|C_{1,2}| > v$ the Weyl cones go inside the Fermi surface and the system is a WSM of type-II with electron and hole pockets and a finite density of states at the Fermi level. In Fig.~(\ref{Fig:TypeI} b,c) we have shown WSM of type-I where the Weyl cones, though tilted, do not touch the Fermi surface. In Fig.~(\ref{Fig:TypeII} a,b) Weyl cones touch the Fermi surface so are at the critical point between transition from type-I to type-II. Fig.~(\ref{Fig:TypeII} c,d) show type-II WSM with electron and hole pockets coexisting with topologically protected Weyl points. In the following, for calculations using the linearized dispersion model we will take $C_2 = -C_1$, as is usually the case (but not always, see Sec.~IV C) when the linearized model arises from a microscopic lattice model of type-II WSMs.
\section{Nernst effect in the presence of Berry curvature}
Generally Nernst response can be seen as a generation of electric field in the presence of a transverse temperature gradient. Conventionally Nernst effect can be observed in the presence of an external magnetic field transverse to the applied temperature gradient, but a non-trivial Berry curvature \citep{wsm21,wsm22,wsm23,wsm24,wsm26,wsm36} can also give rise to anomalous contribution to Hall and the Nernst signal. In the presence of electric field ($\textbf{E}$) and temperature gradient ($\bm\nabla$T) the linear response equations can be written as,
\begin{equation}
\begin{bmatrix}
\textbf{J} \\ \textbf{Q}
\end{bmatrix}
=
\begin{bmatrix}
\bm\sigma & \bm\alpha \\
\bm{\bar{\alpha}} & \bm\kappa
\end{bmatrix}
\begin{bmatrix}
\bm{E}\\
-\bm{\nabla} T
\end{bmatrix}
\label{Eq:Nernst}
\end{equation}
From Eq.~(\ref{Eq:Nernst}) the Nernst coefficient can be defined as,
\begin{equation}
\nu= \frac{E_y}{(-\frac{dT}{dx})}=\frac{\alpha_{xy} \sigma_{xx}-\alpha_{xx} \sigma_{xy}}{\sigma_{xx}^2+\sigma_{xy}^2}\\
\end{equation}
where $(\frac{dT}{dx})$ is the applied temperature gradient in $\hat{x}$ direction and $E_y$ is the electric field generated in the $\hat{y}$ direction. Usually, the longitudinal conductivity $\sigma_{xx}$ is much greater than the Hall conductivity $\sigma_{xy}$ ($\sigma_{xx}>>\sigma_{xy}$), as is the case in this paper, so we neglect $\sigma_{xy}^2$ in the denominator.
With this approximation, the expression for the Nernst coefficient becomes \\
\begin{equation}
\nu=\frac{\alpha_{xx}}{\sigma_{xx}}\left(\theta_P-\theta_H\right)\\
\label{Eq:Theta}
\end{equation}
where $\theta_H=\sigma_{xy}/\sigma_{xx}$ is called the Hall angle and $\theta_P=\alpha_{xy}/\alpha_{xx}$ is called the Peltier angle.
It is now well known that the presence of non-trivial Berry curvature in systems with broken time reversal symmetry can give rise to an anomalous contribution to the Hall signal even in the absence of an external magnetic field. In systems with broken inversion symmetry but in which the time reversal symmetry remains unbroken the anomalous contribution to the Hall effect vanishes.
In the presence of Berry curvature, the semi-classical equations of motion for the position coordinate of an electron wave packet is given by,\\
\begin{equation}
\bm{\dot{r}}=\frac{1}{\hslash}\frac{\partial{\varepsilon(\bm k)}}{\partial \bm k}+
\frac{\dot{\bm p}}{\hslash}\times \bm\Omega_k\\
\label{Eq:req}
\end{equation}
Similarly the equation of motion for the momentum coordinate in the presence of both electric and magnetic fields is given by,
\begin{equation}
\bm{\dot{p}}=e\bm{E}+e\bm{\dot{r}}\times \bm{B}
\label{Eq:peq}
\end{equation}
It is clear from Eqs.~(\ref{Eq:req},\ref{Eq:peq}) that a non-zero Berry curvature $\Omega_k$ acts as a pseudo magnetic field in the equations of motion of an electron, albeit one that is defined in the momentum space.
Solving the coupled equations, Eqs.~(\ref{Eq:req},\ref{Eq:peq}), for $\bm{\dot{r}}$ and $\bm{\dot{p}}$ and simplifying them we get,\\
\begin{equation}
\bm{\dot{r}}=D(\bm B,\bm \Omega_k)\left(\bm v_k+\frac{e}{\hslash}(\bm E\times \bm\Omega_k)+\frac{e}{\hslash} (\bm v_k.\bm\Omega_k)\bm B\right)\\
\label{Eq:req1}
\end{equation}
\begin{equation}
\bm{\dot{p}}=D(\bm B,\bm\Omega_k)\left(e\bm E+\frac{e}{\hslash}(\bm v_k\times \bm B)+\frac{e^2}{\hslash}(\bm E.\bm B)\bm\Omega_k\right)\\
\label{Eq:peq1}
\end{equation}
where,
\begin{equation}
D(\bm B,\bm\Omega_k) =(1+e(\bm B.\bm\Omega_k)/\hslash)^{-1}
\label{Eq:Deq}
\end{equation}
is a nontrivial phase-space factor arising from non-zero Poisson brackets of the coordinates \citep{duval2006}. In Eqs.~(\ref{Eq:req1},\ref{Eq:peq1}) we have used $\bm v_k=\hslash^{-1}\dfrac{\partial \varepsilon_{\bm k}}{\partial \bm k}$ as band velocity.
We now use Eqs.~(\ref{Eq:req1},\ref{Eq:peq1}) for $\bm{\dot{r}}$ and $\bm{\dot{p}}$ to solve the semi-classical Boltzmann equations in the relaxation time approximation \citep{wsm25,wsm35}, resulting in,
\begin{equation}
(\bm{\dot{r}}.\bm{\nabla_r}+\bm{\dot{k}}.\bm{\nabla_k})\\f_{\bm k}=-\frac{f_{\bm k}-f_{eq}}{\tau}
\label{Eq:Boltzmann}
\end{equation}
where $\tau$ is the scattering time of electrons, which for simplicity we take as independent of momentum. In Eq.~(\ref{Eq:Boltzmann}) $f_{\bm k}$ is the Fermi-Dirac distribution function with perturbation and $f_{eq}$ is the equilibrium Fermi-Dirac distribution. The linear response relations between the charge and the applied fields can be stated as,
\begin{equation}
J_a=\sigma_{ab} E_b+\alpha_{ab} (-\nabla_b T)
\label{Eq:Jeq}
\end{equation}
We will consider the case when $\bm E=0$ and derive the longitudinal and transverse conductivities. We consider particular arrangement relevant for experimental measurements i.e. $\bm \nabla T=\nabla_x T \hat{x}, \bm B=B \hat{z}$ and $\bm E=0$. After making substitutions for $\dot r $ and $\dot p$ from Eqs.~(\ref{Eq:req1},\ref{Eq:peq1}) in Boltzmann Eq.~(\ref{Eq:Boltzmann}) we get,
\begin{multline}
v_x\tau \nabla_x T\frac{\varepsilon-\mu}{T}\left(-\frac{\partial f_{eq} }{\partial \varepsilon}\right)+\frac{eB}{\hslash}\left(-v_x\frac{\partial }{\partial k_y}+v_y \frac{\partial}{\partial k_x}\right) f_{\bm{k}}\\=-\frac{f_{\bm {k}}-f_{eq}}{D(B,\Omega_k)\tau}
\label{Eq:Boltzmann1}
\end{multline}
We use the following expression of $f_k$ with correction factor $\Lambda$ for finite magnetic field,
\begin{multline}
f_{\bm {k}}=f_{eq}-\left(D\tau v_x \nabla_x T \frac{\varepsilon-\mu }{T}-\bm v.\bm \Lambda\right)\left(-\frac{\partial f_{eq}}{\partial \varepsilon}\right)
\label{Eq:Boltzmann2}
\end{multline}
Using this expression [Eq.~(\ref{Eq:Boltzmann2})] of $f_k$ in Eq.~(\ref{Eq:Boltzmann1}) we find,
\begin{multline}
\frac{eB}{\hslash}\left(v_y \frac{\partial}{\partial k_x}-v_x\frac{\partial}{\partial k_y}\right)\left(-D\nabla_x T\frac{\varepsilon-\mu}{T}v_x \tau+\bm v.\bm \Lambda\right)\\=-\frac{\bm v. \bm \Lambda}{\tau}
\end{multline}
We find that $\Lambda_z=0$ as this equation must be valid for all values of $\bm v$. The equation can be simplified as,
\begin{multline}
eB\nabla_x T \frac{\varepsilon -\mu}{T}D\tau \left(\frac{v_x}{m_{xy}}-\frac{v_y}{m_{xx}}\right)+eB\left(\frac{v_y \Lambda_x}{m_{xx}}-\frac{v_x \Lambda_y}{m_{yy}}\right)\\
=-v_x \Lambda_x \left(-\frac{eB}{m_{xy}}+\frac{1}{D\tau}\right)-v_y \Lambda_y \left(\frac{eB}{m_{xy}}+\frac{1}{D\tau}\right)
\end{multline}
We now introduce two complex variables $V=v_x+i v_y$ and $\Lambda=\Lambda_x-i \Lambda_y$ to solve this equation and can be rewritten in this way,
\begin{multline}
Re\left[eB\tau D \nabla_x T\frac{\varepsilon-\mu }{T}V\left(\frac{1}{m_{xy}}+\frac{i}{m_{xx}}\right)\right]\\
=Re\left[V \Lambda\left(\frac{ieB}{m_{xx}}-\frac{1}{D\tau}\right)+\frac{eBV\Lambda}{m_{xy}}\right]
\end{multline}
For convenience of notation we define, $\Lambda_i=\tau \nabla_x T\dfrac{\varepsilon -\mu }{T}c_i$. We now write lengthy expressions for $c_x, c_y$ as,
\begin{widetext}
\begin{equation}
\bm {c_x}=eB D(B,\Omega_k)\dfrac{\left(\dfrac{v_x}{m_{xy}}-\dfrac{v_y}{m_{xx}}\right)\left(\dfrac{eB v_y}{m_{xx}}+\dfrac{eBv_x}{m_{xy}}+\dfrac{v_x}{D\tau}\right)-\left(\dfrac{v_y}{m_{xy}}+\dfrac{v_x}{m_{xx}}\right)\left(\dfrac{eBv_x}{m_{xx}}-\dfrac{v_y}{D\tau}-\dfrac{eBv_y}{m_{xy}}\right)}{\left(\dfrac{eBv_x}{m_{xy}}\right)^2-\left(\dfrac{v_x}{D\tau}+\dfrac{eBv_y}{m_{xx}}\right)^2-\left(\dfrac{eBv_x}{m_{xx}}-\dfrac{v_y}{D\tau}\right)^2+\left(\dfrac{eBv_y}{m_{xy}}\right)^2}
\label{Eq:cx}
\end{equation}
\begin{equation}
\bm {c_y}=eBD(B,\Omega_k) \dfrac{\left(\dfrac{v_x}{m_{xy}}-\dfrac{v_y}{m_{xx}}\right)\left(\dfrac{eBv_y}{m_{xy}}+\dfrac{eBv_x}{m_{xx}}-\dfrac{v_y}{D\tau}\right)-\left(\dfrac{v_y}{m_{xy}}+\dfrac{v_x}{m_{xx}}\right)\left(\dfrac{eBv_x}{m_{xy}}-\dfrac{v_x}{D\tau}-\dfrac{eBv_y}{m_{xx}}\right)}{\left(\dfrac{eBv_x}{m_{xx}}-\dfrac{v_y}{D\tau}\right)^2-\left(\dfrac{eBv_y}{m_{xy}}\right)^2-\left(\dfrac{eBv_x}{m_{xy}}\right)^2+\left(\dfrac{v_x}{D\tau}+\dfrac{eBv_y}{m_{xx}}\right)^2}
\label{Eq:cy}
\end{equation}
\end{widetext}
In the presence of $\bm B$ and $\Omega_k$, the expression for the charge current $\bm J$ also modified\citep{xiao2006berry} by the factor $D(B,\Omega_k)$ [Eq.~\ref{Eq:Deq}]. The modified $\bm J$ can be rewritten as,
\begin{equation}
\bm J=-e\int[d \bm{k}]D^{-1} \bm{\dot r}f+\frac{k_B e \nabla T}{\hslash}\times\int [d \bm k]\Omega_{\bm k} s_{\bm k}
\label{Eq:feq1}
\end{equation}
Using expressions for $c_x, c_y$ [Eqs .~(\ref{Eq:cx},\ref{Eq:cy})] in the previously mentioned expressions of $f_k$ [Eq .~(\ref{Eq:Boltzmann2})], $f_k$ can be written in terms of $c_x, c_y$ and D as,
\begin{equation}
f_{\bm k}=f_{eq}-\left[\tau \nabla_x T \dfrac{\varepsilon- \mu }{T}\left(\frac{\partial f_{eq}}{\partial \varepsilon}\right)\right][(c_x-D)v_x+c_y v_y]
\label{Eq:Jeq1}
\end{equation}
Now substituting Eq.~(\ref{Eq:Jeq1}) in Eq.~(\ref{Eq:feq1}) and comparing with the previous linear response relation of J [Eq.~(\ref{Eq:Jeq})], we finally find thermal conductivities as,
\begin{equation}
\alpha_{xx}=e\int \bm{[dk]} v_x^2 \tau \frac{(\varepsilon-\mu)}{T}\left(-\frac{\partial f_{eq}}{\partial \varepsilon}\right)(c_x-D)
\end{equation}
\begin{multline}
\alpha_{xy}=e\int \bm{[dk]}\frac{(\varepsilon-\mu)}{T}\tau\left(-\frac{\partial f_{eq}}{\partial \varepsilon}\right)\left(v_y^2 c_y+v_x v_y(c_x-D)\right)\\
+ \frac{k_B e}{\hslash}\int [dk]\Omega_z s_k
\end{multline}
In the above expression, $v_x=\hslash^{-1}\dfrac{\partial \varepsilon}{\partial k_x}, v_y=\hslash^{-1}\dfrac{\partial \varepsilon}{\partial k_y}$ are the band velocities of electrons and $\bm {[dk]} = \dfrac{d^3 \bm k}{(2 \pi)^3}$ denotes integration over the 3D momentum space.
Similarly the electrical conductivity tensors in the presence of electric field can be obtained as,
\begin{equation}
\sigma_{xx}=-e^2\int \bm{[dk]} v_{x}^2\tau \left(-\frac{\partial f_{eq}}{\partial \varepsilon}\right)(c_x-D)
\end{equation}
\begin{multline}
\sigma_{xy}=-e^2\int \bm{[dk]}\left(-\frac{\partial f_{eq}}{\partial\varepsilon}\right) \tau\left(v_y^2 c_y+v_x v_y (c_x-D)\right)\\
+\frac{e^2}{\hslash}\int [dk]\Omega_z f_k
\end{multline}
For realistic values of the chemical potentials $c_x, c_y, D$ reduce to, $c_x-D\rightarrow -1$ and $c_y \rightarrow \omega \tau$ where $\omega$ is the cyclotron frequency, $\omega=\frac{eB}{m}$ (m is the effective mass of electrons). By realistic values of chemical potential we mean $\mu$ in range of 15 meV-100 meV \citep{zhang2015tantalum,mar1992metal,wsm9}. According to this approximation longitudinal conductivity tensors can be simplified to,
\begin{equation}
\sigma_{xx}=e^2\int \bm{[dk]} v_{x}^2\tau \left(-\frac{\partial f_{eq}}{\partial \varepsilon}\right)
\label{Eq:Sigmax}
\end{equation}
\begin{equation}
\alpha_{xx}=-\frac{e}{T}\int \bm{[dk]} v_x^2 \tau (\varepsilon-\mu)\left(-\frac{\partial f_{eq}}{\partial \varepsilon}\right)
\label{Eq:Alphax}
\end{equation}
Using full expressions of $c_x, c_y$ [Eqs.~(\ref{Eq:cx},\ref{Eq:cy})], D [Eq.~(\ref{Eq:Deq})] and keeping terms only upto linear order in B the transverse conductivity tensors can be obtained as,
\begin{multline}
\sigma_{xy}=-\frac{e^3 \tau^2 B}{\hslash^2}\int \bm{[dk]}\left(-\frac{\partial f_{eq}}{\partial\varepsilon}\right) \left(\frac{v_x^2 \partial^2\varepsilon}{\partial k_y^2}-\frac{v_x v_y \partial^2 \varepsilon}{\partial k_x \partial k_y }\right)\\
+\frac{e^2}{\hslash}\int \bm{[dk]}\Omega_z f_k
\label{Eq:Sigmay}
\end{multline}
\begin{multline}
\alpha_{xy}=\frac{e^2 \tau^2 B}{T \hslash^2}\int \bm{[dk]}(\varepsilon-\mu)\left(-\frac{\partial f_{eq}}{\partial \varepsilon}\right)\left(\frac{v_x^2\partial^2\varepsilon}{ \partial k^2_y}-\frac{v_x v_y \partial^2 \varepsilon}{\partial k_x \partial k_y}\right)\\
+\frac{k_B e}{\hslash}\int \bm{[dk]}\Omega_z s_k
\label{Eq:Alphay}
\end{multline}
As in this paper we are focusing only on anomalous contributions, in the absence of an external magnetic field, the transverse conductivity tensors simplifies to,
\begin{equation}
\sigma_{xy}=\frac{e^2}{\hslash}\int \bm{[dk]}\Omega_z f_k
\label{Eq:Sigmay1}
\end{equation}
\begin{equation}
\alpha_{xy}= \frac{k_B e}{\hslash}\int \bm{[dk]}\Omega_z s_k
\label{Eq:Alphay1}
\end{equation}
For the calculations of the Berry curvatures we use the expression \citep{wsm26},\\\\
\begin{equation}
\Omega^{n}_{a,b}=i\sum_{m\neq n} \frac{\langle n\vert \frac{\partial H}{\partial k_{a}}\vert m\rangle \langle m\vert \frac{\partial H}{\partial k_b}\vert n\rangle -a\leftrightarrow b}{(\varepsilon_n-\varepsilon_m )^2}
\label{Eq:BC}
\end{equation}
where $\varepsilon_n$ is the energy of the $n$-th band and the sum is over all eigenstates $\vert m\rangle$ of the Hamiltonian excluding the eigenstate $\vert n\rangle$.
From the denominator of the expression it is clear that the Berry curvature peaks up near the band touching points which in the present work are precisely the Weyl points. So in our calculations with the linearized model the Berry curvature has peaks at $(0,0,\pm Q)$ in the momentum space, which give rise to the anomalous Hall and Nernst effects even in the absence of an external magnetic field.
\section{Results}
In this section we first discuss the anomalous Nernst signal from the low energy linearized model with a high energy cutoff followed by the results from a lattice model of WSMs where the tilt is in the $k_z$ direction. We also point out the role of the entropy density $s_k$ in the results for the Nernst signal in the linearized model for opposite tilts of the Weyl cones. We then consider a lattice model for type-II WSMs with the tilt in the $k_x$ direction and illustrate the main result of this work, a significant anisotropy in the anomalous Nernst signal depending on the direction of the temperature gradient, which can be used as a signature of type-II Weyl semimetals with broken time reversal symmetry.
\subsection{Anomalous Nernst Signal in the low energy linearized model}
The anomalous Nernst response in the presence of Berry curvature can be calculated from the Eqs.~(\ref{Eq:Theta},\ref{Eq:Sigmax},\ref{Eq:Alphax},\ref{Eq:Sigmay1},\ref{Eq:Alphay1},\ref{Eq:BC}).
\begin{figure}[h]
\begin{frame}{}
\hbox{\hspace{-3ex}\includegraphics[width=.53\textwidth]{nernsttype2.pdf}}
\end{frame}
\caption{Normalized anomalous Nernst coefficient ($\nu$) for a linearized model of WSM, Eq.~(\ref{Eq:H1}), as a function of the tilting parameter of Weyl cones. The Nernst signal has been calculated for different values of the chemical potential. When the tilting is zero the anomalous Nernst signal vanishes, which is an artifact of the linearized model (see Fig.~(\ref{Fig:NLattice})). The slopes of the curves change when the system makes the transition from type-I to type-II WSM, i.e., upto $|C_1|<v$ the slope is negative whereas for $|C_1|>v $ the slope is positive, which indicates a qualitative change in the behavior of the anomalous Nernst signal at the critical point $|C_1|=v$. It is important to note that for a fixed value of the tilting parameter the anomalous Nernst signal increases with decreasing values of the chemical potential, i.e., as the system approaches the undoped limit}
\label{Fig:Nlinear}
\end{figure}
The anomalous Nernst signal ($\nu$) on the tilting parameter $C_1$ (we take $C_2 = -C_1$ in the linearized model) for a set of finite values of the chemical potential is shown in Fig.~(\ref{Fig:Nlinear}). As mentioned before, here we have introduced an energy and momentum cut-off as unbounded linear dispersion is not realistic for physical Weyl semimetals. In the realistic systems there will be intrinsic energy-momentum cutoff arising from the lattice parameter in a realistic lattice model description of the WSMs. Here in the approximate linearized model we observe that the anomalous Nernst signal has a negative slope with the tilting parameter for WSM type-I region from $C_1/v=-1$ to $C_1/v=1$. For WSM of type-II $C_1/v>1$ and $C_1/v<-1$ the Nernst signal has a positive slope with the tilting parameter. It is important to note that for a fixed value of the tilting parameter the anomalous Nernst signal increases with decreasing values of the chemical potential, i.e., with the system approaching the undoped limit. In the limit of $C_1/v\longrightarrow \infty $ or $C_1/v\longrightarrow -\infty$ the Nernst signal converges to a finite value which approximately vanishes.
\begin{figure}[h]
\begin{frame}{}
\hbox{\hspace{-3ex}\includegraphics[width=.5\textwidth]{entropy.png}}
\end{frame}
\caption{Entropy density in the linearized model of a WSM. (a) shows the entropy density for the tilt parameter $C_1/v=-1.5$, and (b) shows the entropy density for the tilt parameter $C_1/v=1.5$. The entropy density moves significantly in the momentum space with change in the tilting of the Weyl cones from $C_1/v =-1.5$ to $C_1/v = 1.5$ and, coupled with the fact that the Berry curvature distribution is independent of the tilting parameter and depends only on the location of the Weyl nodes, results in opposite anomalous Nernst signal for opposite tilting of the Weyl cones. As it is a linearized model with a higher energy cut-off the plot of entropy density does not look like a periodic function.}
\label{Fig:Entropy}
\end{figure}
The entropy density
\begin{equation}
s_{\bm k}=-f_{\bm k} \log{f_{\bm k}}-(1-f_{\bm k})\log(1-f_{\bm k})
\end{equation}
for the linearized model is shown in Fig.~(\ref{Fig:Entropy}). We observe a significant change in the entropy density along the $k_z$ axis. As in the low energy linearized model the Berry curvature does not change with the tilt of the Weyl cones, change in entropy density is what causes the opposite Nernst signal for $C_1/v=1.5$ and $C_1/v=-1.5$.
It can be seen from Fig.~(\ref{Fig:Nlinear}) that the anomalous Nernst signal vanishes \citep{wsm18} when there is no tilting. As has been discussed before \citep{wsm18} this is an artifact of the linearized model and a more realistic lattice model for the time reversal breaking WSM produces a non-zero anomalous Nernst signal even in the absence of tilting of the Weyl cones.
\subsection{Anomalous Nernst signal for a lattice model of time reversal breaking WSM}
For a lattice model of the time reversal breaking Weyl semimetal with an intrinsic energy-momentum cut-off provided by the lattice spacing we take the Hamiltonian,
\begin{multline}
H(\bm k)=\gamma (\cos(k_z)-\cos(k_0))\sigma_0\\
-\left( m[2-\cos(k_y)-\cos(k_x)]+2t_x[\cos(k_z)-\cos(k_0)]\right)\sigma_1\\
-2t\sin(k_y)\sigma_2+2t\sin(k_x)\sigma_3
\label{Eq:Hlattice}
\end{multline}
where $\gamma$ is the tilt parameter along the $k_z$ direction, $\sigma$'s are the Pauli matrices and $\sigma_0$ is the $2\times 2$ unit matrix. We take the lattice parameter $k_0=\pi/2$, $t_x=t$ and $m=2t$ for our numerical calculations. The Hamiltonian in Eq.~(\ref{Eq:Hlattice}) produces two Weyl points located at $(0,0,\pm\pi/2)$ in the momentum space and makes a transition from type-I to type-II at $\gamma/t=\pm2$ when the chemical potential $\mu$ is zero.
\begin{figure}[h]
\begin{frame}{}
\hbox{\hspace{-2ex}\includegraphics[width=.53\textwidth]{lattice.png}}
\hbox{\hspace{-4ex}\includegraphics[width=.53\textwidth]{nernstlattice.pdf}}
\end{frame}
\caption{Normalized anomalous Nernst signal for a lattice model of time reversal breaking WSM. Upper panel: (a) indicates the energy spectrum for a lattice model, Eq.~(\ref{Eq:Hlattice}), when there is no tilting in the Weyl cones ($\gamma=0$). (b) shows the Weyl cones tilted from each other in opposite direction with the tilting parameter $\gamma/t=1.5$. In this case the tilts are in $k_z$ direction similar to the linearized model, Eq.~(\ref{Eq:H1}). The Weyl points are located at $(0,0,\pm\pi/2)$ in the momentum space. Lower panel shows the anomalous Nernst coefficient($\nu$) for the lattice model as a function of the tilt parameter $\gamma$ of the Weyl cones. The Nernst signal has been calculated for different values of the chemical potential and increases monotonically with decrease in the chemical potential. The anomalous Nernst signal is finite even in the absence of tilting of the Weyl cones \citep{wsm18}, in contrast to the case of the linearized model (see Fig.~(\ref{Fig:Nlinear})). But near the end points i.e. with very large tilting, the signals from both the linearized model and the lattice model matche with each other and the anomalous Nernst signal approximately vanishes.}
\label{Fig:NLattice}
\end{figure}
We calculate Nernst signal for this lattice model according to Eqs~.(\ref{Eq:Theta},\ref{Eq:Sigmax},\ref{Eq:Alphax},\ref{Eq:Sigmay1},\ref{Eq:Alphay1},\ref{Eq:BC}). The dependence of the anomalous Nernst signal on the tilt parameter $\gamma$ has been plotted in Fig.~(\ref{Fig:NLattice}) for different values of the chemical potential. We find that the lattice model results in the anomalous Nernst effect which is considerably different from that in the linearized model. The most important difference is at zero tilting ($\gamma=0$), for which we find a finite Nernst signal (in fact a peak plotted as a function of $\gamma$), in contrast to zero Nernst signal for the untilted Weyl nodes in the linearized model. Fig.~\ref{Fig:NLattice}, therefore, establishes the fact that the vanishing of the anomalous Nernst effect for $\gamma=0$ is an artifact of the linearized model. This result agrees with Ref .~[\onlinecite{wsm18}], where only the case of $\gamma=0$ was considered, and disagrees with the results for anomalous Nernst signal in Ref .~[\onlinecite{lundgren2014}]. We also observe that for $\gamma/t\longrightarrow \infty$ or $\gamma/t\longrightarrow-\infty$ the Nernst signal converges to a finite value which approximately vanishes as in the linearized model. It is important to note that, as in the linearized model, for a fixed value of the tilting parameter the anomalous Nernst signal monotonically increases with decreasing values of the chemical potential.
\subsection{Anisotropy in the anomalous Nernst signal for type-II WSM}
As the previous models were symmetric with respect to $k_x, k_y$ there was no anisotropy in the anomalous Nernst signals between the cases when the temperature gradient is applied along $\hat{x}$ or $\hat{y}$ axes. To illustrate the anisotropy when the temperature gradient is applied along and perpendicular to the tilt axis,
we now consider a lattice model for a time reversal broken Weyl semimetal with the Weyl cones tilted along the $k_x$ axis,
\begin{multline}
H(\bm k)=\gamma (\sin(k_x)-\sin(k_0))\sigma_0\\
-\left( m[2-\cos(k_y)-\cos(k_x)]+2t_x[\cos(k_z)-\cos(k_0)]\right)\sigma_1\\
-2t\sin(k_y)\sigma_2+2t\sin(k_x)\sigma_3
\label{Eq:HLattice1}
\end{multline}
Here, as before, $\sigma$'s are conventional Pauli matrices, $\sigma_0$ is $2\times 2$ unit matrix, and $\gamma$ is the tilt parameter along the $k_x$ direction. We take the lattice parameters $k_0=0$, $t_x=t$, and $m=2t$. The Hamiltonian in Eq.~(\ref{Eq:HLattice1}) produces two Weyl cones at $(0,0,\pm \pi/2)$ still separated along the $k_z$ axis but now tilted along $k_x$.\\
\begin{figure}[h]
\begin{frame}{}
\hbox{\hspace{-3ex}\includegraphics[width=.53\textwidth]{anislattice.png}}
\hbox{\hspace{-5ex}\includegraphics[width=.53\textwidth]{nernstanis.pdf}}
\end{frame}
\caption{Normalised anomalous Nernst signal for a lattice model of type-II WSM. Upper panel: (a) indicates the energy spectrum of lattice model Eq.~(\ref{Eq:HLattice1}) when there is no tilting in the Weyl cones. (b) shows both Weyl cones tilted in the same direction along the $k_x$ axis when the tilt parameter $\gamma/t=1.5$. Lower panel: Anomalous Nernst coefficient ($\nu$) is plotted with the tilt parameter. The Nernst signal has been calculated for different values of the chemical potential. When the tilt is zero we get a finite Nernst signal similar to the previous lattice model with the tilt in the $k_z$ direction. As both the Weyl cones are tilted in the same direction, in contrast to Fig.~(\ref{Fig:NLattice}), the curves are symmetric about the tilting axis and the anomalous Nernst signal has the same value whether the tilting parameter is positive or negative.}
\label{Fig:Nlattice1}
\end{figure}
\begin{figure}[h]
\begin{frame}{}
\hbox{\hspace{-4ex}\includegraphics[width=.53\textwidth]{anis.pdf}}
\end{frame}
\caption{Normalised difference in Nernst signal ($\Delta \nu$) between the cases when the temperature gradient is applied parallel and perpendicular to the tilt direction plotted as a function of the tilt parameter in time reversal breaking Weyl semimetal described in Eq.~(\ref{Eq:HLattice1}). The anisotropy has been calculated for different values of the chemical potential. As expected, when the tilt is zero there is no anisotropy and the anisotropy increases with increase in the tilt of the Weyl cones.}
\label{Fig:Anisotropy}
\end{figure}
The anomalous Nernst response in the presence of Berry curvature can be calculated from Eqs.~(\ref{Eq:Theta},\ref{Eq:Sigmax},\ref{Eq:Alphax},\ref{Eq:Sigmay1},\ref{Eq:Alphay1},\ref{Eq:BC}). The dependence of Nernst signal for Hamiltonian in Eq.~(\ref{Eq:HLattice1}) has been plotted in Fig.~(\ref{Fig:Nlattice1}). For $\gamma/t\longrightarrow \infty$ or $\gamma/t\longrightarrow-\infty$ the anomalous Nernst signal converges to zero. It is important to note that in the model of Eq.~(\ref{Eq:HLattice1}) both the Weyl cones are tilted in the same direction along the $k_x$ axis, which is different from the model in Eq.~(\ref{Eq:Hlattice}) where the tilts are in opposite directions. As the Weyl cones are tilted in the same direction, we get the same Nernst signal whether the tilt parameter is positive or negative, and in contrast to Fig.~(\ref{Fig:NLattice}) the curves in Fig.~(\ref{Fig:Nlattice1}) are symmetrical.
The Hamiltonian in Eq.~(\ref{Eq:HLattice1}) is not symmetric in $k_x$ and $k_y$, as the $\sigma_0$ term responsible for the tilting of the Weyl cones contains only $k_x$. Consequently, we expect the system to show strong anisotropy between the cases when the applied temperature gradient is parallel or perpendicular to the tilt axis.
For calculating the Nernst coefficient when the temperature gradient is along $x$ direction and the generated electric field is in the $y$ direction, we use the formula,
\begin{equation}
\nu_x= \frac{E_y}{-\frac{dT}{dx}}=\frac{\alpha_{xy} \sigma_{xx}-\alpha_{xx} \sigma_{xy}}{\sigma_{xx}^2+\sigma_{xy}^2}\\
\label{Eq:Theta1}
\end{equation}
Similarly, for calculating the Nernst signal with temperature gradient along $y$ direction and the generated electric field is in the $x$ direction, we use the expression,
\begin{equation}
\nu_y= \frac{E_x}{-\frac{dT}{dy}}=\frac{\alpha_{xy} \sigma_{yy}-\alpha_{yy} \sigma_{xy}}{\sigma_{yy}^2+\sigma_{xy}^2}\\
\label{Eq:Theta2}
\end{equation}
Here $\sigma_{xy}=\sigma_{yx}\quad $and$ \quad \alpha_{xy}=\alpha_{yx}$ due to symmetry in the anomalous terms.
Finally, we calculate $\Delta \nu= \nu_x-\nu_y$ as the signature of anisotropy in the anomalous Nernst signal in type-II Weyl semimetals that can be measured in experiments. In our model $\Delta \nu$ was calculated according to Eqs.~(\ref{Eq:Sigmax},\ref{Eq:Alphax},\ref{Eq:Sigmay1},\ref{Eq:Alphay1},\ref{Eq:BC},\ref{Eq:Theta1},\ref{Eq:Theta2}).
The anisotropy in the anomalous Nernst coefficient for this lattice model is illustrated in Fig.~(\ref{Fig:Anisotropy}). In the absence of tilting, i.e $\gamma/t=0$, we don't observe any anisotropy, as expected for type-I WSM without tilting of Weyl cones. We find that the anisotropy increases with the tilting of the Weyl cones as the direction of tilt introduces a preferred direction on the $x-y$ plane. Thus, the system responds differently when the applied temperature gradient is along the direction of tilt or perpendicular to it.
\section{Conclusion}
We study the anomalous Nernst signal of time reversal broken type-II Weyl semimetals for both low-energy linearized dispersion model with a high energy cut-off as well as for more realistic lattice models. We observe that the anomalous Nernst signal vanishes in the linearized model for zero tilting but for more realistic lattice models the anomalous Nernst effect is finite (in fact has a peak) when the tilt parameter vanishes. The vanishing of anomalous Nernst effect in the linearized model is thus an artifact of the low energy description and this agrees with a similar result in Ref.~[\onlinecite{wsm18}] (which treats only the case of zero tilting) and disagrees with Ref.~[\onlinecite{lundgren2014}]. We also study the anomalous Nernst signal in more realistic lattice models where the tilts are along the $k_z$ and $k_x$ axes. Generically, we find that the Nernst effect has a peak for zero tilting and decreases with increasing tilts of the Weyl cones. Importantly, we also find that the anomalous Nernst signal increases monotonically with decreasing values of the chemical potential, i.e., as the system approaches the undoped limit. A central result of this work is a pronounced anisotropy in the anomalous Nernst effect in type-II Weyl semimetals between the cases when the external temperature gradient is applied parallel and perpendicular to the tilt axis. The tilt in the Weyl cones introduces a preferred direction in momentum space and the corresponding symmetry breaking introduces strong anisotropy in topological response functions which can be used as an important marker for type-II WSMs in magnetic systems with spontaneously broken time reversal symmetry.
\acknowledgements
The authors appreciate access to the computing facilities in the Department of Physics, IIT Kharagpur, India. We would like to thank G. Sharma, M. Chakraborty for some useful discussions.
\subsection*{Author contribution statement}
S Tewari conceived the problem, S Saha performed the calculations, S Saha and S Tewari wrote the paper.
\\\\
\textit{Note added:} After our paper was completed, we became aware of a recent preprint Ref.~[\onlinecite {ferreiros2017arxiv}], which also discusses anomalous Nernst effect in tilted Weyl semimetals, but only from a low energy linearised dispersion model.
|
\section{Desired properties} \label{desired}
For this new generation of ECPs, we envision including the following set of properties.
a) {\bf Many-body construction} is the key and perhaps the most salient point in this list. Some ideas in this direction have been tried before,
for example, Acioli and Ceperley \cite{Acioli:1994jcp} explored correlated wave functions to generate ECPs and required that the pseudoatom's density matrix matched that of the full atom beyond a given core radius.
Extending on this methodology, Trail and Needs have generated tables of ECPs from MCHF wave functions \cite{Trail:2013jcp,Trail:2015jcp} and CCSD(T) excitation energies \cite{Trail:2017jcp} to improve their previously generated table \cite{Trail:2005jcp}.
Additionally, attempts were made to build core effects into ECPs in the works of Maron, et al.\cite{Maron:cp1998} and Fromager et al.\cite{Fromager:jcp2004}.
We believe that a more systematic and accurate approach here can still be achieved, especially when considering properties outside of equilibrium (see below).
In our construction, we use Coupled Cluster (CC) and Configuration Interaction (CI) methods that are very effective for atomic and small molecular systems, and moreover, we analyze how our generated ECPs perform in non-equilibrium geometries.
b) {\bf The simplicity of the ECP operator} has significant benefits since it allows use in many methods (QMC, quantum chemistry based on expansions in basis sets, and DFT, using established codes).
We, therefore, build upon a simple semi-local form that has been in use for some time.
Per need, a number of extensions could prove to be useful and could be incorporated in future.
This might involve different choices of core-valence partitions, several types of representations such as numerical on radial meshes besides usual Gaussian expansions, fixed radial cutoffs, etc.
More desired properties could be added as the project develops, e.g., how to optimally deal with ``fat" cores of heavy elements, possible inclusion of core polarization and core relaxation terms, minimizing the fixed-node errors generated by the core part of the wave function without compromising
the accuracy, minimizing the locality approximation in QMC applications following \cite{krogel_kent_2017}, etc. For plane wave applications, the suitability of recasting the ECP into the Kleinman-Bylander form \cite{Kleinman1982} could be included as well.
c) {\bf Testing} on a set of systems in order to delineate the accuracy limits for energy differences, equilibrium bond lengths and other properties. Potentially, some of these systems could be included into the retuning set, if important or necessary. For use and further improvements, it will be very useful to have documented systematic errors of generated ECPs documented upfront.
d) {\bf Systematic labeling and updates}, i.e., keeping the data, developments, and history on a website \cite{website} that can be eventually updated by interested contributors at large.
\section{Effective ECP Hamiltonian: Isospectrality on a subspace of valence states} \label{form}
One way to approach the ECP construction
is to formulate it as
an inverse problem, i.e.,
finding an effective operator that produces
a subset of valence atomic states such that outside
the core they reproduce the
all-electron properties as closely as possible.
We assume that for this subset both the all-electron and pseudized atoms can be solved in the same systematic and consistent
framework, ideally exactly, or, in practice, nearly exactly, using the state-of-the-art many-body methods.
What makes the two operators, the all-electron Hamiltonian and ECP Hamiltonian, close? Assuming we are interested in the valence subset of states and properties there are the following two key aspects:
\smallskip
\noindent
i) the two spectra should be the same/very close; and,
\smallskip
\noindent
ii) the spatial characteristics (one- and multi-particle many-body density matrices) of the corresponding two sets of eigenstates outside the core region should be the same/very close.
Note that, in general, the two Hamiltonians differ substantially, in the number of particles/degrees of freedom, presence/absence of ionic Coulomb singularities and relativity.
Since in the reference calculations all correlations (including cores) are assumed to be present, i) and ii) implicitly demand that the ECPs capture, as much as possible, also the {\it impact of core-core and core-valence correlations
on the valence properties}. Our results below show that to a certain extent, this is indeed
the case.
Finding the desired solution in this setting essentially defines an inverse problem. We can expect that this problem will be ill-conditioned with many nearly optimal or non-unique solutions.
Why is this the case? The inclusion of many states into the optimization set often leads to a frustrated optimization problem, e.g., improvement for one state increases penalties for other states. With sufficient number of such frustrating couplings one ends up with a problem that is qualitatively similar, say, to finding a ground state of a spin glass (a well-known problem in statistical mechanics).
One way how to deal with such
ill-conditioned problems is to impose appropriate constraints
that limit the space of possible solutions. That makes the problem solvable but could result in compromises on the
resulting accuracy or incurred biases.
Often, it is quite difficult to find the right set of constraints. If the objective function is over-constrained the search for a minimum is fast but
the solution might be too biased by the constraints.
On the other hand, in an under-constrained formulation, the optimization could be very inefficient.
Therefore, the goal is to find the best trade-off(s) between these two limits.
\subsection{ECP Form}
The valence-only electronic Born-Oppenheimer Hamiltonian we consider has
the following form
\begin{equation}
H_{\rm val} = \sum_i[T^{\rm kin}_i + V^{\rm pp}_i] +\sum_{i<j} 1/r_{ij}
\end{equation}
For this work, we use a semi-local ECP form with a minimal number of parameters \cite{Burkatzki:2007jcp}
\begin{equation}
V^{\rm pp}_i = V_{loc}(r_i) + \sum_{l=0}^{l_{max}} V_l(r_i) \sum_{m}|lm\rangle\langle lm|,
\end{equation}
where $r_i$ is the radial distance of
electron $i$ from the core's origin. The non-local terms contain the projectors on $lm$-th angular momentum state.
The local term, $V_{loc}$, is chosen as
\begin{equation}
V_{loc}(r) = -\frac{Z_{\rm eff}}{r}(1 -e^{-\alpha r^2}) + \alpha Z_{\rm eff} re^{-\beta r^2} + \gamma e^{-\delta r^2},
\end{equation}
where $Z_{\rm eff}$ is the effective core charge, $Z_{\rm eff}=Z-Z_{core}$.
The $V_l$ potential was chosen to consist of a single gaussian term
\begin{equation}
V_l(r) = \beta_le^{-\alpha_lr^2}.
\end{equation}
All variables labeled by Greek letters are treated as optimization parameters in the minimization of a chosen objective function.
Additionally, a constraint that forces a concave shape at the origin is imposed \cite{Burkatzki:2007jcp}
\begin{equation}
\gamma\delta + \alpha_l\beta_l > 0 , \quad \forall l .
\end{equation}
In the case(s) of combined constructions below,
the number of gaussians in each channel is extended.
Note that the chosen form imposes a very significant restriction on the variational freedom of the effective operator. We have used this simple form not only to simplify
the optimization problem but also to investigate the true many-body accuracy limit for this ``minimal model" version.
The results are actually very encouraging.
As presented in the following sections, even with such restricted variational freedom we were able to construct more accurate effective potentials than the existing ones as well as to illustrate the presented concepts above.
\section{Optimization methods and constructions} \label{optimization}
In this section, we give a summary of a few atomic and molecular properties predicted from our constructed ECPs for several atoms from the first and second rows.
For comparison, along with our results, we juxtapose the predictions of these properties from correlated all-electron, uncorrelated core/correlated valence for a handful of other ECPs that have been used in many-body calculations in recent years.
Furthermore, we present results that utilize different strategies to construct the ECPs,
namely, ECPs built from all-electron spectral data only,
from spatial data only as well as combined and iterated constructions.
For correlated calculations, a number of codes can be used in a non-relativistic setting. For relativistic calculations, however, the choices are more limited.
Nevertheless, it is interesting to see how far one could push
the accuracy limits with currently existing
methods and codes.
\subsection{Objective function with atomic spectral discrepancies only}
For the all-electron spectral-only references, we used the \textsc{Molpro} quantum chemistry package \cite{MOLPRO-WIREs} to calculate a subset of states from each atom's spectrum using the CCSD(T) method.
To account for scalar relativistic effects, in the all-electron reference we used the Douglas-Kroll-Hess Hamiltonian throughout.
For each atom, the uncontracted aug-cc-pCV5Z basis set (which includes core state correlation functions) was chosen
in order to minimize finite basis set errors as much as possible.
We used this same basis for both pseudo-atom and all-electron cases so that the basis set errors would largely cancel when gap discrepancies between the two cases were calculated. We carried out limited tests with even larger basis (6Z) and we found out that the gained accuracy had a marginal effect on the spectral differences when compared with other discrepancies
such as binding curves for molecular systems as explained below. In the supplementary material, we show again marginal level of errors from the extrapolation to complete basis set limit for both BFD and all-electron spectral gaps for the first and second row atoms N and S, respectively.
To generate the ECPs, we followed an energy consistent scheme, as described in reference \cite{Burkatzki:2007jcp} and references therein, and minimized the differences in the all-electron and ECP excitation energies in a least-squares way, with our objective function defined as
\begin{equation}
f = \sum_s \left( \Delta E_s^{AE} - \Delta E_s^{pp} \right)^2,
\label{eqn:spectral}
\end{equation}
where $s$ labels a given excited state and $\Delta E_s$ is the energy gap between the excited state and the neutral atomic ground state.
We chose to include as many ionizations as possible in the atomic references; from the most deeply ionized single-valence state up to (at least) the $M^\mathrm{th}$ anionic state where $M$ is the predicted number of bound anions from the all-electron CCSD(T) spectrum.
The spectrum here therefore includes not only neutral atomic states, but also numerous states for the anions and cations.
In our investigation, we observed that when the most deeply ionized states were not included in the reference, their corresponding excitations tended to have significant discrepancies from the all-electron excitations and consequently the pseudoatom's transferability could be negatively affected.
Additionally, for each ionization, we expanded the reference to include the bound ground states of all possible total spin channels of that ion, e.g., for an ion with four electrons in the valence space, we included the ground states of the quintet, triplet and singlet spin channels into the reference, provided that each state was predicted to be bound from the all-electron calculation.
The latter choice was motivated by a desire to minimize any possible contamination from energetically lower states within the same symmetry channel.
\subsection{ Objective function with spectral and spatial density matrix discrepancies}
In order to incorporate spatial information into the pseudopotential, we explored the utilization of the single-body density matrix following the previous work of Acioli and Ceperley.
For the all-electron reference and pseudoatom, we generated the single-body density matrices from a CISD wave function.
We expect that if the pseudoatom density matrix and the all-electron density matrix agree beyond a chosen core region defined by
cut-off radius $r_c$, then the pseudoatom valence will mimic that of the all-electron atom.
Given an AE atom and a pseudoatom, we can measure the
difference between the density matrices outside of $r_c$ as
\begin{equation}
\Delta\rho_{r_c} = \iint\limits_{r,r'>r_c} d\mathbf{r} d\mathbf{r}' |\rho_{AE}(\mathbf{r},\mathbf{r}') - \rho_{ECP}(\mathbf{r},\mathbf{r}')|^2 .
\label{eqn:dm}
\end{equation}
If we express each density matrix through the natural orbitals $\{\phi_i\}$ and occupation numbers $n_i$ of a correlated calculation such as CISD, we can rewrite the discrepancy between the density matrices as
\begin{eqnarray}
\Delta\rho_{r_c} = \sum_{ij}^N n^{AE}_i n^{AE}_j \left|\langle \phi_i^{AE} | \phi_j^{AE}\rangle_{r_c}\right|^2
\nonumber\\
+ \sum_{i,j}^M n_i^{ECP}n_j^{ECP} \left| \langle \phi_i^{ECP}|\phi_j^{ECP}\rangle_{r_c}\right|^2 \nonumber\\
-2\sum_i^N \sum_j^M n_i^{AE} n_j^{ECP} \left| \langle\phi_i^{AE} | \phi_j^{ECP} \rangle_{r_c}\right|^2
\end{eqnarray}
where the expression $\langle \phi_i | \phi_j \rangle_{r_c}$ is the overlap of two natural orbitals evaluated from $r_c$ to infinity.
An important point worth noting here is that
one should not expect to rigorously match the all-electron and pseudoatom density matrices.
The reason is that the density matrix of the original
atom generally contains a contribution from the tails
of the core states beyond the cutoff radius. Since it is difficult to disentangle the tails in the many-body setting this ``contamination" will be present.
Note that in traditional one-particle formulations this is a non-issue since
HF or Kohn-Sham self-consistent orbitals
decompose
the core and valence states into distinct sets.
The closest analogy to the
one-particle decomposition comes from the natural orbitals, however, even there the occupation numbers
and the natural orbitals will slightly differ
(reflecting thus the presence or the absence of the core states).
This becomes particularly relevant whenever smaller core radii are imposed.
The fits can be further elaborated to approximately take this into
account, however, here we opted for a simple agreement
assuming that this effect will not dominate the achieved
discrepancies.
We further constructed ECPs where we include both the spectral and spatial information into the objective function.
Using the integration of the density matrix beyond a cutoff radius as defined above, we define our spatial objective function as
\begin{equation}
g = \sum_s \Delta \rho_{r_c}^s
\label{eqn:spatial}
\end{equation}
where we sum over all states considered.
We construct a new objective function where we include both spatial and spectral information as
\begin{equation}
h = \frac{f}{f_0} + \frac{g}{g_0}
\label{eqn:spec/space}
\end{equation}
where $f_0$ and $g_0$ are the optimal values of the respective objective functions. In the case of sulfur below, the two objective functions were left unweighted given the similarity between $f_0$ and $g_0$ in that case.
\subsection{Optimization methods}
Throughout, we have utilized the nonlinear DONLP2 optimization code of Spellucci \cite{DONLP2} for generating the final parameter values of our ECPs with respect to the various objective functions which we have considered.
The specifics of the method implemented in DONLP2 are outlined in two papers \cite{SpellucciA,SpellucciB}, where generally, the solver extends the sequential quadratic programming method, an iterative method that relies on the second derivatives of both the objective function and the constraints, such that it can be applied to nonlinear problems.
We considered this choice for the chosen solver to be appropriate given the nonlinear ``smoothness" constraint that has been imposed on the parameter sets of our ECPs.
\subsection{ Constructed, Combined and Iterated Schemes}
The final procedure that we pursued took molecular data into account.
Having generated molecular binding curve predictions from an assortment of ECP constructions, we noticed that there was a potential to improve the ECPs further. In particular, the optimization could be guided to fulfill additional criteria such as reproducing the dimer binding curve in a few iterations. This type of optimization loop produced very small discrepancies from the all-electron CCSD(T) curves along the desired range of geometries even at the steep repulsive side of the binding curves.
In particular, for the corresponding dimers, we found that this strategy is able to produce discrepancies as small as $0.05$~eV or lower along the binding curve all the way up to the dissociation limit at the repulsive side. Due to time demands of correlated calculations and rather slow coupling between different codes we guided this level of optimization by interventions to speed up the search. That was also very useful to understand the qualitative relationships between shapes/amplitudes of the ECP versus its properties and therefore we leave the automation of this part for future. We comment on these constructions at each presented atomic case.
In all of the systems presented below, we compare our ECPs to all-electron CCSD(T), all-electron CCSD(T) with an uncorrelated core (UC), i.e., no excitations from the core, and various ECPs often used in QMC calculations, namely Burkatzki-Filippi-Dolg (BFD) \cite{Burkatzki:2007jcp}, Trail-Needs (TN-DF) \cite{Trail:2005jcp} DF ECPs, and where applicable, the shape-consistent correlated Trail-Needs (TN-CEPP) ECPs \cite{Trail:2013jcp} as well as the recent shape and energy consistent correlated Trail-Needs (TN-eCEPP) ECPs \cite{Trail:2017jcp}. Note that when utilizing the TN-CEPP and TN-eCEPP, we only test the effective core potential and do not include the semi-empirical core polarization potentials utilized by Trail and Needs,
given that we were interested in comparing all effective cores within a consistent level of approximation.
A key question that we chose to pursue was how well an ECP alone would capture core-valence effects without the need for additional adjustments or approximations. By not including CPPs along with the ECPs allowed us to isolate this effect.
Note that, Trail and Needs have studied the effect of the CPP in Figure 6 of \cite{Trail:2017jcp}, illustrating that CPPs do have only minor impact on the dissociation energies over a wide variety of molecules. As explained in results, we found that much larger discrepancies almost invariably appear at shorter molecular bond lengths and therefore we have focused
on addressing this aspect in
our constructions.
\section{Results} \label{results}
Here we present our initial results for a number of explicitly correlated ECPs using the various objective functions described above. In the first row, we present optimized ECPs for Boron, Carbon, Nitrogen, and Oxygen. For the second row elements, we show our results for Sulfur.
\subsection{Boron}
\begin{table*}[ht!]
\centering
\caption{ Atomic and ionic excitations and corresponding discrepancies for Boron. IP denotes the first ionization potential while EA is the electron affinity. Q is the ionization charge, 2S+1 the usual total spin multiplicity. AE denotes the calculated all-electron values while the rest of columns shows the discrepancies.
UC means all-electron valence-only correlation with self-consistent but uncorrelated core, as explained in the text.
All energies in eV. The MAD is the mean absolute difference over all of the discrepancies.
Note that all gaps are calculated with reference to the ground state, namely Q=0 and 2S+1 = 2.
The same notation applies to all the atomic/ionic data tables throughout the paper. }
\begin{tabular}{rcrrrrrrrr}
\hline
\hline
Q & 2S+1 & AE Gap & UC & BFD & TN-DF & TN-CEPP & TN-eCEPP & Spectral & Constructed \\
\hline
+2 & 2 & $ 33.4290$ & $ -0.0726 $ & $ 0.0048 $ & $ -0.0316 $ & $ 0.1186 $ & $ 0.0224 $ & $ 0.0027 $ & $ 0.0109 $ \\
(IP)+1 & 1 & $ 8.2771$ & $ -0.0379 $ & $ -0.0670 $ & $ -0.0711 $ & $ -0.0030 $ & $ -0.0280 $ & $ -0.0077 $ & $ -0.0301 $ \\
+1 & 3 & $ 12.9083$ & $ -0.0085 $ & $ 0.0249 $ & $ 0.0121 $ & $ 0.0326 $ & $ 0.0032 $ & $ -0.0042 $ & $ -0.0129 $ \\
0 & 4 & $ 3.5752$ & $ 0.0187 $ & $ 0.0202 $ & $ 0.0523 $ & $ 0.0184 $ & $ 0.0107 $ & $ 0.0018 $ & $ -0.0079 $ \\
$-$1 & 1 & $ 0.2854$ & $ -0.0026 $ & $ 0.0067 $ & $ 0.0092 $ & $ 0.0079 $ & $ 0.0046 $ & $ -0.0034 $ & $ 0.0031 $ \\
(-EA)$-$1 & 3 & $ -0.2482$ & $ 0.0043 $ & $ 0.0142 $ & $ 0.0220 $ & $ 0.0144 $ & $ 0.0117 $ & $ 0.0006 $ & $ 0.0098 $ \\
$-$1 & 5 & $ 2.5144$ & $ 0.0207 $ & $ 0.0183 $ & $ 0.0683 $ & $ 0.0301 $ & $ 0.0193 $ & $ -0.0004 $ & $ -0.0027 $ \\
\hline
MAD & & & 0.0236 & 0.0223 & 0.0381 & 0.0321 & 0.0143 & 0.0030 & 0.0110\\
\hline
\hline
\end{tabular}
\label{tab:boron}
\end{table*}
\begin{figure}[ht!]
\caption{Boron dimer potential energy surface. UC represents an all-electron CCSD(T) calculation with a self-consistent but {\it uncorrelated} core, i.e., with
no excitations from the core states. Spectral represents the optimization for the atomic spectrum alone, and Constructed indicates the ECP driven iteratively to minimize the dimer discrepancy while accepting a small increase in the spectrum discrepancy.}
\includegraphics[width=8cm]{figures/1_b2}
\label{fig:b2}
\centering
\end{figure}
\begin{figure}[ht!]
\caption{Boron dimer binding energy discrepancies compared to the all-electron CCSD(T) binding curve. The gray envelope represents a $0.05$~eV window for the discrepancy. The vertical line indicates the equilibrium bond length as predicted by the all-electron CCSD(T) calculation. }
\includegraphics[width=8cm]{figures/2_b2diffs}
\label{fig:b2_diffs}
\centering
\end{figure}
Boron is perhaps the lightest element where time savings from an appropriate ECP become significant while the valence space is sufficiently large to represent the atom in chemical settings.
Table \ref{tab:boron} shows
several constructions of the ECP obtained by minimizing the atomic spectral error only (Spectral) and another that has been adjusted to reproduce the binding
curve of the ground state of B$_2$ in the range of bond lengths (Constructed).
Note the very high accuracy that was obtained with minimizing only the spectral discrepancy. A significantly improved dimer solution has been
found by minimizing a compromise between reproducing the spectrum
and the binding curve. Note that the shortest
bond length corresponds approximately to the dissociation point due to nucleus-nucleus repulsion. In solids, such distance between
the atoms would correspond to very high pressures roughly beyond
500~GPa.
Figure \ref{fig:b2} shows the B$_2$ potential energy surface for the $^3\Sigma_g$ state and \ref{fig:b2_diffs} provides a set of discrepancies from the all-electron CCSD(T) for our constructed ECPs compared against
previous ECPs: BFD, TN-DF, the recently correlated constructions by Trail and Needs (TN-CEPP and TN-eCEPP) as well as the all-electron uncorrelated core result.
\begin{table}[!ht]
\setlength{\tabcolsep}{12pt}
\centering
\caption{ECP parameters for Constructed Boron. The parametrization for each channel is given by $V_l(r) = \sum_k \beta_{lk} r^{n_{lk}-2}e^{-\alpha_{lk}r^2}$. The corresponding correlation consistent basis sets are included in the Supplementary Material. }
\begin{tabular}{c r r r}
\hline\hline
Channel & $n_{lk}$ & $\alpha_{lk}$ & $\beta_{lk}$ \\
\hline
$p$ & 1 & 31.49298 & 3.00000\\
& 3 & 22.56509 & 94.47895\\
& 2 & 8.64669 & -9.74800\\
\hline
$s-p$ & 2 & 4.06246 & 20.74800\\
\hline\hline
\end{tabular}
\label{tab:pp_b}
\end{table}
Right away we can observe the following:
a) The constructed ECPs provide a significantly more accurate picture of the molecular system, both for the valence spectrum optimization (Spectral) as well
as more accurate binding curve optimization (Constructed), than the existing examples. Remarkably, our ECPs show
{\it smaller errors} for both the spectrum and for binding curve than the all-electron correlated valence-valence calculation with an {\it uncorrelated
core} (UC). In UC all one-particle states are solved self-consistently including the cores (i.e., it is not the atomic frozen core in that sense), however, only valence-valence correlations are invoked in the CCSD(T) method while core-valence and core-core correlations are neglected. The fact that our effective Hamiltonians provide a better description than UC for the dimer and atomic spectra is quite surprising. We, therefore, observe that the inverse problem formulation enables to mimic, within a certain level of accuracy, effects of core excitations on the valence space.
b) The simple form of the ECP described above is able to accommodate higher accuracy demands.
Note that overall accuracy in reproducing correlated
properties is roughly at the level of $0.02$~eV, significantly higher than
in previous constructions.
Both of these observations are quite unexpected and will be further
elaborated in the cases of other elements.
Since these solutions were satisfactory we have not pursued further
improvements by including spatial correlations explicitly into
the objective function. Note that the spectrum and refinement
alone were sufficient to determine the high accuracy solution
within the minimal and significantly restricted representation.
Our best ECP parameters (Constructed) are given in Table \ref{tab:pp_b}.
{\bf Labeling.}
For the sake of clear identification and recording of progress we introduce here the label for this and subsequent ECPs as ``correlation consistent ECP",
version 0.1, or {\bf ccECP.0.1}, in short.
The same label is assigned to all tabulated ECPs in this paper.
Note that the notation can be further expanded, for example, as ttECP.xx.yy.ENSCO.
Here the theory would be labeled as
tt=cc,hf,df,..., while major.minor release is represented as above by xx.yy, respectively. In addition, fitted quantities could include ENSCO=excitations-norms/shapes-spatialDM-combined/iterated-other approaches. As far as systems used for fit is concerned one could possibly affix a label ADHOCG=atom-dimer-hydride-oxide-cluster-general/other system. As an example, the full label for our Boron case
would be ccECP.0.1.EC.
\subsection{Carbon}
\begin{figure}[ht!]
\caption{Carbon dimer binding energy discrepancies compared to the all-electron CCSD(T) binding curve.}
\includegraphics[width=8cm]{figures/3_c2diffs}
\label{fig:c2_all_diffs}
\centering
\end{figure}
Since the spectral only optimization worked quite well for B, only needing a slight readjustment with our constructed ECP, we test the same for the C atom.
The atomic spectrum for various ECPs is given in Table \ref{tab:carbon} and dimer discrepancies in the $^1\Sigma_g$ state are given in Figure \ref{fig:c2_all_diffs}.
The spectral only construction does remarkably well on minimizing the energy differences for the atomic spectra and produces an ECP that is improved over other ECPs. While there is slight overbinding, the Spectral curve is relatively flat indicating accurate vibrational frequencies compared to the all-electron. The electron affinity (EA) for the spectral differs with the all electron CCSD(T) value by only $-1$~meV. The ionization potential (IP) is an order of magnitude better and only differs by $+0.1$~meV.
In addition, a minor compromise in reproducing the atomic spectra produces a much better dimer binding that is within $0.05$~eV across the entire curve and has a flat discrepancy, i.e., the same vibrational frequencies as the all-electron curve.
\begin{table*}[ht!]
\centering
\caption{Atomic data for Carbon, similar to Table \ref{tab:boron}. Energies in eV. Note that all gaps are calculated with reference to the ground state, namely Q=0 and 2S+1 = 3.}
\begin{tabular}{rcrrrrrrrr}
\hline
\hline
Q & 2S+1 & AE Gap & UC & BFD & TN-DF & TN-CEPP & TN-eCEPP & Spectral & Constructed \\
\hline
$+3$ & 2 & $ 83.4895 $ & $ -0.1469 $ & $ -0.1090 $ & $ -0.1561 $ & $ -0.0824 $ & $ 0.2544 $ & $ 0.0005 $ & $ -0.0024 $\\
$+2$ & 1 & $ 35.6041 $ & $ -0.1020 $ & $ -0.2208 $ & $ -0.1723 $ & $ -0.0953 $ & $ 0.0326 $ & $ -0.0007 $ & $ 0.0110 $\\
$+2$ & 3 & $ 42.1035 $ & $ -0.0561 $ & $ -0.1083 $ & $ -0.1080 $ & $ -0.0368 $ & $ 0.0751 $ & $ -0.0009 $ & $ -0.0061 $\\
(IP)$+1$ & 2 & $ 11.2452 $ & $ -0.0334 $ & $ -0.0725 $ & $ -0.0631 $ & $ -0.0277 $ & $ -0.0073 $ & $ 0.0016 $ & $ 0.0027 $\\
$+1$ & 4 & $ 16.5590 $ & $ -0.0022 $ & $ -0.0955 $ & $ -0.0552 $ & $ -0.0173 $ & $ 0.0075 $ & $ 0.0001 $ & $ 0.0019 $\\
$0$ & 1 & $ 1.3950 $ & $ -0.0143 $ & $ 0.0013 $ & $ 0.0000 $ & $ 0.0045 $ & $ 0.0095 $ & $ 0.0006 $ & $ -0.0009 $\\
$0$ & 5 & $ 4.1491 $ & $ 0.0231 $ & $ -0.0743 $ & $ -0.0085 $ & $ -0.0030 $ & $ 0.0126 $ & $ 0.0005 $ & $ 0.0084 $\\
(-EA)$-1$ & 4 & $ -1.2421 $ & $ 0.0072 $ & $ 0.0259 $ & $ 0.0249 $ & $ 0.0098 $ & $ 0.0270 $ & $ -0.0010 $ & $ -0.0006 $\\
\hline
MAD & & & 0.0481 & 0.0884 & 0.0735 & 0.0532 & 0.0346 & 0.0008 & 0.0046 \\
\hline
\hline
\end{tabular}
\label{tab:carbon}
\end{table*}
The parameters for our best Carbon ECP (Constructed) is given in Table \ref{tab:pp_c}.
\begin{table}[h]
\setlength{\tabcolsep}{12pt}
\centering
\caption{ECP parameters for Constructed Carbon. The parametrization for each channel is given by $V_l(r) = \sum_k \beta_{lk} r^{n_{lk}-2}e^{-\alpha_{lk}r^2}$. The corresponding correlation consistent basis sets are included in the Supplementary Material. }
\begin{tabular}{c r r r}
\hline\hline
Channel & $n_{lk}$ & $\alpha_{lk}$ & $\beta_{lk}$ \\
\hline
$p$ & 1 & 14.43502 & 4.00000 \\
& 3 & 8.39889 & 57.74008 \\
& 2 & 7.38188 & -25.81955 \\
\hline
$s-p$ & 2 & 7.76079 & 52.13345 \\
\hline\hline
\end{tabular}
\label{tab:pp_c}
\end{table}
\subsection{Nitrogen}
\begin{table*}[ht!]
\centering
\caption{Atomic data for Nitrogen, similar to Table \ref{tab:boron}. Energies in eV. Note that all gaps are calculated with reference to the ground state, namely Q=0 and 2S+1 = 4.}
\scriptsize
\begin{tabular}{rcrrrrrrrrrr}
\hline
\hline
Q & 2S+1 & AE Gap & UC & BFD & TN & TN-CEPP & TN-eCEPP & Spectral & Spatial & Spec/Space & Constructed \\
\hline
$+4$ & $2$ & $169.0094$ & $-0.2308$ & $ -0.2139$ & $ -0.3301$ & $ 0.5136 $ & $-0.3450 $ & $ 0.0024 $ & $ -0.2259 $& $ 0.0022 $ & $ -0.2758 $ \\
$+3$ & $1$ & $ 91.5371$ & $-0.1784$ & $ -0.3903$ & $ -0.2882$ & $ 0.2012 $ & $-0.1921 $ & $ -0.0043 $ & $ -0.2368 $& $ -0.0046 $ & $ -0.0494 $ \\
$+3$ & $3$ & $ 99.8869$ & $-0.1205$ & $ -0.2582$ & $ -0.2868$ & $ 0.2291 $ & $-0.2538 $ & $ -0.0011 $ & $ -0.0157 $& $ -0.0009 $ & $ -0.2531 $ \\
$+2$ & $2$ & $ 44.1212$ & $-0.0866$ & $ -0.1827$ & $ -0.1429$ & $ 0.0792 $ & $-0.0862 $ & $ 0.0018 $ & $ -0.0635 $& $ 0.0017 $ & $ 0.0014 $ \\
$+2$ & $4$ & $ 51.1908$ & $-0.0455$ & $ -0.2695$ & $ -0.2339$ & $ 0.0645 $ & $-0.1732 $ & $ -0.0023 $ & $ 0.0548 $& $ -0.0019 $ & $ -0.2317 $ \\
$+1$ & $1$ & $ 16.5790$ & $-0.0448$ & $ -0.0472$ & $ -0.0370$ & $ 0.0380 $ & $-0.0168 $ & $ 0.0105 $ & $ 0.0028 $& $ 0.0105 $ & $ 0.0232 $ \\
(IP)$+1$ & $3$ & $ 14.5319$ & $-0.0297$ & $ -0.0646$ & $ -0.0547$ & $ 0.0125 $ & $-0.0291 $ & $ -0.0013 $ & $ -0.0070 $& $ -0.0013 $ & $ 0.0100 $ \\
$+1$ & $5$ & $ 20.5319$ & $ 0.0009$ & $ -0.2524$ & $ -0.1729$ & $-0.0041 $ & $-0.1077 $ & $ 0.0011 $ & $ 0.0559 $& $ 0.0015 $ & $ -0.2154 $ \\
$ 0$ & $2$ & $ 2.6789 $ & $-0.0295$ & $ -0.0037$ & $ -0.0056$ & $ 0.0046 $ & $-0.0104 $ & $ -0.0084 $ & $ -0.0080 $& $ -0.0085 $ & $ -0.0068 $ \\
\hline
MAD & & & 0.0852 & $0.1870$ & 0.1725 & 0.1274 & 0.1349 & 0.0037 & 0.0745 & 0.0037 & 0.1185\\
\hline
\hline
\end{tabular}
\label{tab:nitrogen}
\end{table*}
\begin{figure}[ht!]
\caption{Nitrogen dimer binding energy discrepancies compared to the all-electron CCSD(T) binding curve.}
\includegraphics[width=8cm]{figures/4_n2diffs}
\label{fig:n2_all_diffs}
\centering
\end{figure}
In constructions of nitrogen, we explored several objective functions and refinement strategies, as shown in Figure \ref{fig:n2_all_diffs} for the molecule in the $^1\Sigma_g$ state and Table \ref{tab:nitrogen} for the atomic properties.
The spectrum only optimization performs rather well and is a significant improvement over many of the previous constructions.
The results show that using the spatial information only
such as one-particle density matrices appears
less favorable overall. We conjecture that this
is caused due to very weak correlation ``signal" in the corresponding density matrices. The difference
between the mean-field and correlated density matrices is very tiny so that the correlation effect is overwhelmed by the one-particle character that is dominant.
On the other hand, the spectrum alone
provides a much stronger signal since the eigenvalues basically determine the wave function tails essentially exactly.
When combining the spatial and spectral information, the spatial signal was not sufficient to significantly change the parameters, and we obtain essentially the same ECP for both constructions.
Although the spectral ECP has excellent atomic properties and a reasonable dimer, we sought to construct a better ECP overall.
The constructed ECP was generated by adding additional constraints that were not included in the original objective function in order to alter systematic trends we observed in our various constructions to minimize the dimer discrepancy.
In particular, we observed that pairs of ECPs with dimer discrepancies of opposite trends could be produced based on different constraints and subsequently the ECPs could be combined as a linear sum leading to increased agreement with the molecular properties from all-electron CCSD(T) while also preserving a high level of accuracy on the atomic spectrum.
This procedure could certainly be used to define a new objective function where the dimer fit is directly included, however, this proved to be computationally inefficient and we therefore chose to improve the molecular properties by hand as described.
The constructed ECP has a discrepancy of below $0.05$~eV across the entire curve, although we compromise the spectrum when compared to our spectral construction, particularly for high-spin atomic states.
However, if the high-spin atomic states are not considered, the atomic spectrum is better than all previous generations of ECPs. Note that in molecular systems the nitrogen almost invariably appears in low/lowest spin configurations.
Another observation is that ECPs with a combination of minimizing the dimer binding and spectrum discrepancies could be constructed without generating an overly large negative impact to the atomic spectra, provided an appropriate constraint is utilized.
Though some accuracy of the atomic properties would diminish as the dimer binding is improved, as can be seen in Table \ref{tab:nitrogen}, we found that a reasonable balance could be obtained between the two when compared to other ECP constructions.
The parameters for our best Nitrogen ECP (Constructed) is given in Table \ref{tab:pp_n}.
\begin{table}[h]
\setlength{\tabcolsep}{12pt}
\centering
\caption{ECP parameters for Constructed Nitrogen. The parametrization for each channel is given by $V_l(r) = \sum_k \beta_{lk} r^{n_{lk}-2}e^{-\alpha_{lk}r^2}$.
The corresponding correlation consistent basis sets are included in the Supplementary Material. }
\begin{tabular}{c r r r}
\hline\hline
Channel & $n_{lk}$ & $\alpha_{lk}$ & $\beta_{lk}$ \\
\hline
$p$ & 1 & 12.91881 & 3.25000 \\
& 1 & 9.22825 & 1.75000 \\
& 3 & 12.96581 & 41.98612 \\
& 3 & 8.05477 & 16.14945 \\
& 2 & 12.54876 & -26.09522 \\
& 2 & 7.53360 & -10.32626 \\
\hline
$s-p$ & 2 & 9.41609 & 34.77692 \\
& 2 & 8.16694 & 15.20330 \\
\hline\hline
\end{tabular}
\label{tab:pp_n}
\end{table}
\subsection{Oxygen}
As in the previous case, we test a variety of objective functions when constructing an ECP for O.
We present the results in Figure \ref{fig:o2_all_diffs} for the molecule in the $^3\Sigma_g$ state and in Table \ref{tab:oxygen} for the atomic properties.
Using a spectral-only construction, we are able to construct an ECP that provides a binding curve discrepancy to the all-electron CCSD(T) curve within $0.05$~eV along the entire curve.
Since the Spectral ECPs are reasonably flat compared to other ECP constructions and the all-electron uncorrelated core (UC), the vibrational frequencies agree quite well to the all-electron CCSD(T) calculations. As in N, the spatial information alone appears insufficient in generating a high-quality ECP.
When combining both spatial and spectral information (Spec/Space), we actually obtain an ECP that is essentially flat across a wide range of bond lengths and shows the best dissociation energy. This also results in a slight compromise on the atomic properties.
However, this ECP begins to overbind significantly in the shorter bond length regime which would correspond to high pressure in solids. Based on this and the quality of our spectral only ECP, we believe that the spectral only optimization for Oxygen produces the best ECP overall and we do not pursue iterated constructions. The parameters for Spectral Oxygen are given in Table \ref{tab:pp_o}
\begin{figure}[ht!]
\caption{Oxygen dimer binding energy discrepancies compared to the all-electron CCSD(T) binding curve.}
\includegraphics[width=8cm]{figures/5_o2diffs}
\label{fig:o2_all_diffs}
\centering
\end{figure}
\begin{table*}[ht!]
\centering
\caption{Atomic data for Oxygen, similar to Table I. Energies in eV. Note that all gaps are calculated with reference to the ground state, namely Q=0 and 2S+1 = 3.}
\scriptsize
\begin{tabular}{rcrrrrrrrrr}
\hline
\hline
Q & 2S+1 & AE Gap & UC & BFD & TN-DF & TN-CEPP & TN-eCEPP & Spectral & Spatial & Spec/Space \\
\hline
$+5$ & $2$ & $294.8903$ & $ -0.3059$ & $ -0.4334$ & $ -0.6186$ & $ 0.7414$ & $-0.5643 $ & $-0.0021$ & $ -0.4531$ & $ -0.0228$ \\
$+4$ & $1$ & $180.9900$ & $ -0.2478$ & $ -0.6864$ & $ -0.4477$ & $ 0.3505$ & $-0.1826 $ & $ 0.0017$ & $ -0.2573$ & $ -0.0328$ \\
$+4$ & $3$ & $191.1861$ & $ -0.1808$ & $ -0.4825$ & $ -0.5444$ & $ 0.3467$ & $-0.4327 $ & $-0.0037$ & $ -0.2651$ & $ 0.0252$ \\
$+3$ & $2$ & $103.6268$ & $ -0.1383$ & $ -0.3375$ & $ -0.2391$ & $ 0.1617$ & $-0.0780 $ & $ 0.0060$ & $ -0.0801$ & $ 0.0282$ \\
$+3$ & $4$ & $112.4585$ & $ -0.0894$ & $ -0.5304$ & $ -0.4781$ & $ 0.0894$ & $-0.3264 $ & $-0.0021$ & $ -0.2007$ & $ 0.0209$ \\
$+2$ & $1$ & $ 51.3705$ & $ -0.0760$ & $ -0.1056$ & $ -0.0811$ & $ 0.0815$ & $-0.0026 $ & $ 0.0200$ & $ 0.0153$ & $ 0.0541$ \\
$+2$ & $3$ & $ 48.7002$ & $ -0.0627$ & $ -0.1445$ & $ -0.1185$ & $ 0.0349$ & $-0.0297 $ & $-0.0035$ & $ -0.0223$ & $ 0.0264$ \\
$+2$ & $5$ & $ 56.1201$ & $ -0.0257$ & $ -0.5550$ & $ -0.4073$ & $-0.0454$ & $-0.2398 $ & $ 0.0061$ & $ -0.1885$ & $ -0.0026$ \\
$+1$ & $2$ & $ 17.1953$ & $ -0.0621$ & $ -0.0367$ & $ -0.0448$ & $-0.0097$ & $-0.0183 $ & $-0.0206$ & $ -0.0102$ & $ -0.0005$ \\
(IP)$+1$ & $4$ & $ 13.5653$ & $ -0.0142$ & $ -0.0438$ & $ -0.0436$ & $-0.0192$ & $-0.0053 $ & $-0.0058$ & $ -0.0118$ & $ 0.0083$ \\
$+1$ & $6$ & $ 43.9706$ & $ -0.0290$ & $ -0.7264$ & $ -0.4557$ & $-0.1001$ & $-0.2281 $ & $-0.0075$ & $ -0.2131$ & $ -0.0299$ \\
$ 0$ & $1$ & $ 2.1816$ & $ -0.0210$ & $ 0.0057$ & $ 0.0007$ & $ 0.0069$ & $-0.0052 $ & $-0.0055$ & $ 0.0039$ & $ -0.0017$ \\
$ 0$ & $5$ & $ 9.4860$ & $ -0.0182$ & $ -0.1259$ & $ -0.0750$ & $-0.0545$ & $-0.0110 $ & $-0.0250$ & $ -0.0350$ & $ -0.0164$ \\
$ 0$ & $7$ & $ 39.4276$ & $ -0.0247$ & $ -0.6879$ & $ -0.4390$ & $-0.0948$ & $-0.2207 $ & $ 0.0025$ & $ -0.1962$ & $ -0.0152$ \\
(-EA)$-1$ & $2$ & $ -1.4209$ & $ 0.0017$ & $ 0.0105$ & $ 0.0092$ & $ 0.0259$ & $-0.0083 $ & $ 0.0044$ & $ -0.0012$ & $ -0.0036$ \\
\hline
MAD & & & $0.0865$ & $ 0.3275$ & $ 0.2669$ & $0.1442$ & 0.1434 & $0.0078$ & $0.1303$ & $ 0.0192$ \\
\hline
\hline
\end{tabular}
\label{tab:oxygen}
\end{table*}
\begin{table}[h]
\setlength{\tabcolsep}{12pt}
\centering
\caption{ECP parameters for Spectral Oxygen. The parametrization for each channel is given by $V_l(r) = \sum_k \beta_{lk} r^{n_{lk}-2}e^{-\alpha_{lk}r^2}$. The corresponding correlation consistent basis sets are included in the Supplementary Material. }
\begin{tabular}{c r r r}
\hline\hline
Channel & $n_{lk}$ & $\alpha_{lk}$ & $\beta_{lk}$ \\
\hline
$p$ & 1 & 12.30997 & 6.00000 \\
& 3 & 14.76962 & 73.85984 \\
& 2 & 13.71419 & -47.87600 \\
\hline
$s-p$ & 2 & 13.65512 & 85.86406 \\
\hline\hline
\end{tabular}
\label{tab:pp_o}
\end{table}
\subsection{Sulfur}
As a last illustration, we include our progress from the second-row atom, Sulfur.
Construction of the ECP for sulfur has followed similar steps as in previous cases. In Fig.\ref{fig:s2_spectral} we show the impact on the S$_2$ dimer its ground state ($^3\Sigma_g^{(-)}$)
of leaving the core uncorrelated on the overall accuracy and we see that near equilibrium
its agreement with the all-electron prediction is at the level of $\approx0.03$~eV.
\begin{table*}[ht!]
\caption{Atomic data for Sulfur, similar to Table I. Energies in eV. Note that all gaps are calculated with reference to the ground state, namely Q=0 and 2S+1 = 3.}
\begin{tabular}{rcrrrrrrrrrrrrrrrrrrrrrrr}
\hline\hline
Q & 2S+1 & AE Gap & UC & BFD & TN-DF & Spectral & Spatial & Spec/Space & Constructed \\
\hline
$+5$ & $2$ & $188.3560$ & $-0.4975$ & $-0.3201$ & $-0.6243$ & $-0.0469$ & $-0.3620$ & $-0.9587$ & $-0.5141$ & & & & \\
$+4$ & $1$ & $115.7750$ & $-0.2143$ & $-0.0922$ & $-0.1412$ & $ 0.0123$ & $ 0.0693$ & $-0.4847$ & $-0.1838$ & & & & \\
$+4$ & $3$ & $126.1610$ & $-0.3461$ & $-0.1881$ & $-0.3452$ & $ 0.0580$ & $ 0.2173$ & $-0.3349$ & $-0.2871$ & & & & \\
$+3$ & $2$ & $ 68.5115$ & $-0.1152$ & $-0.0295$ & $-0.0381$ & $ 0.0290$ & $ 0.2974$ & $-0.1429$ & $-0.0926$ & & & & \\
$+3$ & $4$ & $ 77.3129$ & $-0.2375$ & $-0.1931$ & $-0.2359$ & $ 0.0318$ & $ 0.3573$ & $-0.0642$ & $-0.1816$ & & & & \\
$+2$ & $1$ & $ 35.1649$ & $-0.1012$ & $-0.0399$ & $-0.0338$ & $-0.0130$ & $ 0.2559$ & $-0.0326$ & $-0.0717$ & & & & \\
$+2$ & $3$ & $ 33.6900$ & $-0.0515$ & $-0.0332$ & $-0.0177$ & $-0.0076$ & $ 0.2565$ & $-0.0246$ & $-0.0582$ & & & & \\
$+2$ & $5$ & $ 40.8647$ & $-0.1675$ & $-0.2658$ & $-0.2325$ & $-0.0577$ & $ 0.2273$ & $-0.0197$ & $-0.1568$ & & & & \\
$+1$ & $2$ & $ 12.4609$ & $-0.0920$ & $-0.0732$ & $-0.0636$ & $-0.0673$ & $ 0.0794$ & $-0.0484$ & $-0.0830$ & & & & \\
(IP)$+1$ & $4$ & $ 10.2976$ & $-0.0159$ & $-0.0581$ & $-0.0377$ & $-0.0536$ & $ 0.0826$ & $-0.0320$ & $-0.0559$ & & & & \\
$+1$ & $6$ & $ 30.7228$ & $-0.1557$ & $-0.2981$ & $-0.3143$ & $-0.0025$ & $ 0.2499$ & $ 0.1068$ & $-0.0134$ & & & & \\
$+0$ & $1$ & $ 1.3319$ & $-0.0399$ & $-0.0029$ & $-0.0084$ & $-0.0015$ & $ 0.0090$ & $-0.0011$ & $-0.0094$ & & & & \\
$+0$ & $5$ & $ 8.9838$ & $-0.0223$ & $-0.0563$ & $-0.0648$ & $-0.0244$ & $ 0.0749$ & $-0.0076$ & $-0.0136$ & & & & \\
$+0$ & $7$ & $ 27.8996$ & $-0.1643$ & $-0.3088$ & $-0.3504$ & $ 0.0094$ & $ 0.2134$ & $ 0.1109$ & $ 0.0170$ & & & & \\
(-EA)$-1$ & $2$ & $ -2.0494$ & $-0.0018$ & $ 0.0277$ & $ 0.0064$ & $ 0.0302$ & $-0.0741$ & $-0.0124$ & $ 0.0149$ & & & & \\
\hline
MAD & & & $ 0.1482$ & $ 0.1325$ & $ 0.1676$ & $ 0.0297$ & $ 0.1884$ & $ 0.1588$ & $ 0.1169$ & & & & \\
\hline\hline
\end{tabular}
\label{tab:sulfur}
\end{table*}
\begin{figure}[ht!]
\caption{Potential energy surfaces of the S$_2$ molecule from CCSD(T). We have plotted the predictions from various treatments of the sulfur cores. Shown are the all-electron core (AE), all-electron uncorrelated core (UC), Burkatzki-Filippi-Dolg (BFD), Dirac-Fock Trail-Needs (TN) and the CCSD(T) spectrum matched (Spectral) ECPs described in the text.}
\includegraphics[width=8cm]{figures/6_s2}
\label{fig:s2_spectral}
\centering
\end{figure}
\begin{figure}[ht!]
\caption{Sulfur dimer binding energy discrepancies compared to the all-electron CCSD(T) binding curve.}
\includegraphics[width=8cm]{figures/7_s2diffs}
\label{fig:s2_all_diffs}
\centering
\end{figure}
\begin{figure}[ht!]
\caption{Sulfur dimer binding energy discrepancies compared to the all-electron CCSD(T) binding curve.}
\includegraphics[width=8cm]{figures/8_s2spatialdiffs}
\label{fig:s2_spatial_diffs}
\centering
\end{figure}
For this case, we utilized a Ne core ECP and attempted a set of optimization strategies similar to the investigation of N and O.
In this case, we again consider constructing ECPs to mimic all-electron many-body spectral, spatial and combined properties, in turn.
Figures \ref{fig:s2_all_diffs} and \ref{fig:s2_spatial_diffs} show CCSD(T) binding energy discrepancies from the all-electron S$_2$ molecule ($^3\Sigma_g^{(-)}$) for various approximations to the sulfur cores including our generated ECPs.
For the spectral case, we again have used an all-electron reference formed from bound excitations and included all possible valence ground states of total spin and charge whereby we minimized the spectral discrepancies exclusively with respect to ECP's parameters.
In Table \ref{tab:sulfur}, it is shown that the MAD from the all-electron excitation energies that is no more than $0.03$~eV in this case.
This agreement is nearly an order of magnitude improvement over the other approximations to the core.
Additionally, with the spectral objective function, we observe that the agreement with the all-electron binding energy at equilibrium is very good with an error of no more than $0.01$~eV.
For shorter separations of the dimer, however, the spectral ECP undershoots the all-electron binding energy by tenths of eV and as a result of this large change from equilibrium, we see a non-negligible slope and curvature in the discrepancy at equilibrium which negatively impacts the agreement with the all-electron prediction of the ground state vibrational frequency.
Using the spatial information for the same set of states, we generated the single-body density matrices from all-electron CISD wave functions and subsequently imposed that the single-body density matrix of the pseudoatom's CISD wave function match beyond a core radius.
The resulting ECPs constructed from this procedure were not adequately transferable which we attribute again to marginal electronic correlations signal in the natural orbitals as we argued for the cases of N and O.
It can also be observed from Table \ref{tab:sulfur} that the atomic properties are generally negatively affected by optimizing spatial data alone.
For sulfur, we also attempted to match both the all-electron excitation energies and single-body density matrices, simultaneously.
For this case, the binding energy discrepancies from the all-electron atom are shown in Fig. \ref{fig:s2_all_diffs}.
Here we see that the error is quite uniform over the entire region plotted, and moreover, it is no more than about $0.06$ eV.
Lastly, we again considered additional molecular constraints on the objective function in order to further improve the ECP's dimer properties as described for the nitrogen atom.
The constraints were constructed in such a way as to reach a balance between the spectral and molecular properties.
We show its discrepancies in Fig. \ref{fig:s2_all_diffs}.
For this ECP, we see that the dimer's error is mostly within $0.05$ eV throughout the plotted region and the spectral properties are an improvement over both UC and previously generated ECPs.
The parameters for sulfur's constructed ECP are shared in Table \ref{tab:pp_s}.
\begin{table}[h]
\setlength{\tabcolsep}{12pt}
\centering
\caption{ECP parameters for Constructed ECP for Sulfur. The parametrization for each channel is given by $V_l(r) = \sum_k \beta_{lk} r^{n_{lk}-2}e^{-\alpha_{lk}r^2}$. The corresponding correlation consistent basis sets are included in the Supplementary Material. }
\begin{tabular}{c r r r}
\hline\hline
Channel & $n_{lk}$ & $\alpha_{lk}$ & $\beta_{lk}$ \\
\hline
$d$ & 1 & 4.23812 & 3.06000 \\
& 1 & 2.19773 & 2.94000 \\
& 3 & 1.71348 & 12.96866 \\
& 3 & 10.20072 & 6.46132 \\
& 2 & 3.41487 & -10.45671 \\
& 2 & 1.40439 & -9.79751 \\
\hline
$s-d$ & 2 & 3.91958 & 23.19840 \\
& 2 & 3.91388 & 22.28866 \\
\hline
$p-d$ & 2 & 2.71232 & 8.39601 \\
& 2 & 3.20078 & 11.15610 \\
\hline\hline
\end{tabular}
\label{tab:pp_s}
\end{table}
\section{Transferability}\label{transferability}
One of the desired properties for our ECPs is transferability, i.e., high-quality performance in systems that were not used in our optimization procedure.
In the case of our spectral ECPs, the optimization of the parameters only involved atomic properties; for these ECPs, the dimer discrepancies compared to the all-electron results illustrate the transferability of our ECPs to a degree.
However, for our constructed ECPs, the parameters were tuned in order to produce improved dimer properties while not sacrificing the atomic properties and thus the transferability should be verified also on independent examples.
Therefore, it is desirable to test those cases in other bonding environments in order to illustrate their transferability beyond the dimers.
For this purpose, we have calculated potential energy surface discrepancies for hydrides, oxides and a handful of additional molecules BH$_3$, BN, BS and CN and in order to ascertain their level of transferability.
All molecular calculations were performed with either the \textsc{Molpro} quantum chemistry package \cite{MOLPRO-WIREs} or the \textsc{Gaussian09} code \cite{g09}.
We show binding energy discrepancies with respect to all-electron CCSD(T) for the NH, OH, NO, SH and SO molecules in Figs. \ref{fig:nh_diffs},\ref{fig:oh_diffs},\ref{fig:no_diffs},\ref{fig:sh_diffs} and \ref{fig:so_diffs}, respectively.
Furthermore, we summarize the discrepancies in the binding parameters, $D_e$, $r_e$, and $\omega_e$ of all molecules considered in this work with respect to all-electron CCSD(T) values in Fig. \ref{fig:global_transferability}, where $D_e$ is the dissociation energy, $r_e$ is the equilibrium bond length, and $\omega_e$ is the vibrational frequency.
For each case, the parameters and their errors were obtained from fitting the potential energy surface to the Morse potential near equilibrium, where the potential is given as
\begin{equation}
V(r) = D_e(e^{-2 a (r-r_e)}-2e^{-a(r-r_e)})
\label{eqn:morse}
\end{equation}
where $a$ is related to the vibrational frequency by
\begin{equation}
\omega_e = \sqrt{\frac{2 a^2 D_e}{\mu}}
\label{eqn:freq}
\end{equation}
and $\mu$ is the reduced mass of the molecule.
To make the comparison of the pseudoatom transferabilities easier, we also share the MADs of these parameters, both at equilibrium and at the dissociation limit at short bond lengths corresponding to high pressures, in Table \ref{tab:global_mads}.
From the table, it is shown that our ECPs perform the best overall, where the mean-absolute deviation from the all-electron dissociation energy is smaller than all considered core approximations;
furthermore, this improvement
and better overall balance is achieved with
very limited variational freedom as given by
the choice of the ECP form.
The only exception is a marginally better MAD for the vibrational frequencies of TN-CEPP with 5(3) versus ours 7(2) cm$^{-1}$.
\begin{figure}[ht!]
\caption{NH binding energy discrepancies for various ECPs}
\includegraphics[width=8cm]{figures/9_nhdiffs}
\label{fig:nh_diffs}
\centering
\end{figure}
\begin{figure}[ht!]
\caption{OH binding energy discrepancies for various ECPs. For Oxygen, we use our spectral ECP.}
\includegraphics[width=8cm]{figures/10_ohdiffs}
\label{fig:oh_diffs}
\centering
\end{figure}
\begin{figure}[ht!]
\caption{NO binding energy discrepancies for various ECPs. For Nitrogen, we use our constructed ECP and for Oxygen, we use our spectral ECP.}
\includegraphics[width=8cm]{figures/11_nodiffs}
\label{fig:no_diffs}
\centering
\end{figure}
\begin{figure}[ht!]
\caption{SH binding energy discrepancies compared to the all-electron CCSD(T) binding curve. For Sulfur, we use our constructed ECP and for Oxygen, we use our spectral ECP.}
\includegraphics[width=8cm]{figures/12_shtransferability}
\label{fig:sh_diffs}
\centering
\end{figure}
\begin{figure}[ht!]
\caption{SO binding energy discrepancies compared to the all-electron CCSD(T) binding curve. For Sulfur, we use our constructed ECP.}
\includegraphics[width=8cm]{figures/13_sotransferability}
\label{fig:so_diffs}
\centering
\end{figure}
\begin{figure*}[ht!]
\centering
\caption{Discrepancies of molecular binding parameters of our ECPs, UC and previous constructions with respect to all-electron CCSD(T) calculations. Parameters were obtained from Morse potential fits in all cases.
}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/14a_de.pdf}
\caption{Dissociation energy discrepancies}
\end{subfigure
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/14b_re.pdf}
\caption{Equilibrium bond length discrepancies.}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/14c_we.pdf}
\caption{Vibrational frequency discrepancies.}
\end{subfigure}
\label{fig:global_transferability}
\end{figure*}
\begin{table*}[ht!]
\centering
\caption{Mean absolute deviations of discrepancies of binding parameters at equilibrium ($D_e$, $r_e$ and $\omega_e$) and near the dissociation threshold ($D_{diss}$) at short bond lengths for our ECPs and previous constructions with respect to all-electron CCSD(T) calculations. The system sets correspond to Fig.\ref{fig:global_transferability} except for the BH$_3$ molecule which was omitted from the MADs of the dissociation threshold energy.}
\begin{tabular}{c | cccccc}
\hline
\hline
& UC & BFD & TN-DF & TN-CEPP & TN-eCEPP & This Work \\
\hline
$D_e$ (eV/$10^2$) & 2.9(1) & 2.3(1) & 7.8(1) & 1.9(2) & 2.7(2) & 1.8(1) \\
$r_e$ (\AA/$10^3$) & 2.9(2) & 2.9(2) & 3.6(2) & 3.2(3) & 1.1(2) & 1.0(2) \\
$\omega_e$ (cm$^{-1}$) & 9(2) & 12(2) & 23(2) & 5(3) & 12(3) & 7(2) \\
$D_{diss}$ (eV/$10^2$) & 11.94 & 23.76 & 9.78 & 22.64 & 7.10 & 6.13\\
\hline
\hline
\end{tabular}
\label{tab:global_mads}
\end{table*}
\section{Conclusions}\label{conclusions}
Our paper presents an advancement in the construction of effective core potentials for accurate, correlated valence-only calculations.
A key difference from previous constructions being the consistent use of nearly exact many-body approaches to build the ECPs
and balancing this
with refinement by molecular
data that improves the transferability.
We introduce isospectrality of all-electron and ECP Hamiltonians on the subspace of valence states as a foundation in formulating the objective functions.
Additional criteria that were explored included many-body spatial information from density matrices.
This was followed by analyzing the accuracy on dimers and if significant differences were observed we included the dimer information to boost the accuracy and transferability.
This was done in an iterative manner and combined constructions were used for efficiency reasons.
We were especially careful in obtaining consistent binding curves within a large portion of the bond, i.e., from separations near the dissociation limit to at least equilibrium.
We further ascertained the transferability in those cases on hydride, oxide molecules and selected other molecules that involved the considered element.
The calculations are done at the relativistic level, therefore, energy differences are very close to the actual experimental values for the relevant quantities.
We were able to decrease the atomic spectral errors for B, C, N, O, and S atoms to significantly smaller values when compared to previous constructions.
Interestingly, the spectral discrepancies are also {\it smaller} than in uncorrelated core calculations.
Therefore the constructed operators are able to effectively accommodate at least some of the effects from core-core and core-valence correlations as it is clear from comparisons with uncorrelated core calculations.
Similar argument applies to the improved description of simple molecular systems for a range of bond lengths including very short bonds that are relevant for calculations of materials at high pressures.
Rather surprisingly, this implies that for a range of valence properties these ECPs will provide the same or {\it even more accurate} results then all-electron treatments with uncorrelated cores.
Note the caveat ``for a range of valence properties" in the previous statement since this is true only for states/properties that are not too far from the one used in optimization.
Clearly, inclusion of larger range of energies and states that can be described would require either additional terms in the effective ECP Hamiltonian or, in general, including more subshells into the valence space (as is common for heavier elements).
Our experience from these constructions shows:
a) the accuracy of even the simplest forms of semi-local ECPs can be boosted very significantly by appropriate many-body constructions;
b) optimization is quite involved since minor changes in the inputs and the objective function (e.g., increasing the size of the basis set, boosting the spectral accuracy from CCSD to CCSD(T), etc)
complicate the optimizations with many local minima solutions, i.e., we see a clear hallmark of an ill-conditioned problem;
c) there are differences from element to element influenced by the extent and choice of excitations, occupations and other details showing thus further complications from the ill-conditioned character of the task.
The obtained results and experience offers a route for constructing comparable quality
ECPs for more involved elements such as $3d$ transition series and beyond. Some adjustments in the constructions and expanding the parameter sets can be expected. Overall prospects for heavier elements would be limited basically only by the availability of
codes that can do accurate correlated all-electron relativistic calculations for atomic and small molecular systems.
In this work we have pursued the construction of ECPs from a many-body framework as opposed to one-particle schemes.
In addition, we have included further considerations into the
construction such as molecular systems at and away of the equilibrium for better overall transferability.
One-particle constructions have generally involved norm/shape and charge conservations, one-particle eigenvalues, logarithmic derivatives, differences between approximate total energies, etc.
Unfortunately, all of these were subject to the biases of the (approximate) method used to solve the underlying atomic problem.
The expectation and hope were that in subsequent calculations the systematic cancellations of the underlying biases would be the same as in all-electron setting.
In many cases, this worked reasonably well. However, this was not true systematically and for some elements, states and bonding environments the errors are significant.
For example, in DFT all-electron and ECP calculations for just a single atom could lead to large differences. In addition,
special adjustments might be necessary such as the nonlinear core corrections for transition metal atoms \cite{Louie:1982prb}, etc.
These complications, however, defy the universal use of such ECPs in other methods, what is one of our stated goals.
In addition, there has always been a lingering (and valid) question why an ECP constructed in DFT or HF methods should be any good in a correlated many-body approach.
Clearly, that was one of our motivations to use a many-body framework consistently throughout the construction, in testing and also in further developments.
The validity and limits of the constructed ECPs also become much more transparent if the {\it information about systematic errors built into ECPs is provided upfront} so as to clear the stage for subsequent calculations.
The simplicity of the basic ECP form is also highly desirable since that enables wide use in many settings and approaches.
Note that presented ECPs are of the simplest semilocal type with minimal form and size, with only one or two gaussians per channel.
It is realistic to expect that more careful and more elaborate fits could further tune the properties beyond accuracies we have currently obtained.
The corresponding optimizations are still challenging and are hampered by complicated couplings between the atomic solvers and optimization methods and require further refinement, higher efficiency, and more robustness.
At the same time, the presented results offer encouraging examples of the accuracy that can be obtained with present day capabilities and show that there is significant room for further improvements and expansions.
Input and output files for the calculations performed in this study are available \cite{data_doi} via the Materials Data Facility.
\section*{Supplementary Material}
See supplementary material for basis set extrapolation analysis and a listing of correlation consistent basis sets for each pseudoatom presented in this work.
\smallskip
{\bf Acknowledgments.}
We are grateful for stimulating discussions with Jerry Whitten during this study. We also would like to thank Paul Kent for a careful reading of the manuscript while in preparation which lead to a number of vital improvements.
The majority of this work (development of the methods, calculations, tests, and writing of the paper) has been supported by
the U.S. Department of Energy, Office of Science, Basic Energy Sciences, Materials Sciences and Engineering Division, as part of the Computational Materials Sciences Program and Center for Predictive Simulation of Functional Materials.
The initial theoretical and conceptual considerations were supported by ORNL/UT Batelle, LLC, subcontract N. 4000144475.
The calculations for this work were performed mostly at Sandia National Laboratories, while some of the calculation have been carried out at TACC under XSEDE.
Sandia National Laboratories is a multimission laboratory managed and operated by National Technology and Engineering Solutions of Sandia LLC, a wholly owned subsidiary of Honeywell International Inc. for the U.S. Department of Energy National Nuclear Security Administration under contract DE-NA0003525.
|
\section{Introduction}
Deep Reinforcement Learning is an emerging subfield of Reinforcement Learning (RL) that relies on deep neural networks as function approximators that can scale RL algorithms to complex and rich environments. One key work in this direction was the introduction of DQN \cite{Mnih2015Human} which is able to play many games in the ATARI suite of games \cite{bellemare13arcade} at above human performance. However the agent requires a fairly large amount of time and data to learn effective policies and the learning process itself can be quite unstable, even with innovations introduced to improve wall clock time, data efficiency, and robustness by changing the
learning algorithm \cite{Schaul2015Prioritzed,Hasselt2015Deep} or by improving the optimizer \cite{MniBadMir2016a, Schulman2015Trust}.
A different approach was introduced by \cite{Jaderber2016Reinforcement,Mirowski2016Learning,Lample2016Playing}, whereby data efficiency is improved by training additional auxiliary tasks jointly with the RL task.
With the success of deep RL has come interest in increasingly complex tasks and a shift in focus towards scenarios in which a single agent must solve multiple related problems, either simultaneously or sequentially. Due to the large computational cost, making progress in this direction requires robust algorithms which do not rely on task-specific algorithmic design or extensive hyperparameter tuning. Intuitively, solutions to related tasks should facilitate learning since the tasks share common structure, and thus one would expect that individual tasks should require less data or achieve a higher asymptotic performance. Indeed this intuition has long been pursued in the multitask and transfer-learning literature~\cite{Bengio12deeplearning,AAAIMag11-Taylor,yosinski-nips2014,Caruana:1997}.
Somewhat counter-intuitively, however, the above is often not the result encountered in practice, particularly in the RL domain~\cite{RusColGul2015a,ParBaSal2016a}. Instead, the multitask and transfer learning scenarios are frequently found to pose additional challenges to existing methods. Instead of making learning easier it is often observed that training on multiple tasks can negatively affect performances on the individual tasks, and additional techniques have to be developed to counteract this \cite{RusColGul2015a,ParBaSal2016a}.
It is likely that gradients from other tasks behave as noise, interfering with learning, or, in another extreme, one of the tasks might dominate the others.
In this paper we develop an approach for multitask and transfer RL that allows effective sharing of behavioral structure across tasks, giving rise to several algorithmic instantiations. In addition to some instructive illustrations on a grid world domain, we provide a detailed analysis of the resulting algorithms via comparisons to A3C \cite{MniBadMir2016a} baselines on a variety of tasks in a first-person, visually-rich, 3D environment (DeepMind Lab \cite{beattie2016deepmind}).
We find that the Distral algorithms learn faster and achieve better asymptotic performance, are significantly more robust to hyperparameter settings, and learn more stably than multitask A3C baselines.
\section{Distral: Distill and Transfer Learning}
\label{sec:Method}
\begin{wrapfigure}{r}{.5\textwidth}
\vspace*{-4em}
\includegraphics[width=.5\textwidth]{distill.pdf}
\vspace*{-2em}
\caption{Illustration of the Distral framework.}\label{fig:Distral}
\vspace*{-1em}
\end{wrapfigure}
We propose a framework for simultaneous reinforcement learning of multiple tasks which we call Distral. Figure~\ref{fig:Distral} provides
a high level illustration involving four tasks. The method is founded on the notion of a shared policy (shown in the centre) which distills (in the sense of Bucila and Hinton et al.~\cite{BucCarNic06,HinVinDea2014}) common behaviours or representations from task-specific policies \cite{RusColGul2015a,ParBaSal2016a}. Crucially, the distilled policy is then used to guide task-specific policies via regularization using a Kullback-Leibler (KL) divergence. The effect is akin to a shaping reward which can, for instance, overcome random walk exploration bottlenecks. In this way, knowledge gained in one task is distilled into the shared policy, then transferred to other tasks.
\subsection{Mathematical framework}
In this section we describe the mathematical framework underlying Distral.
A multitask RL setting is considered where there are $n$ tasks, where for simplicity we assume an infinite horizon with discount factor $\gamma$.\footnote{The method can be easily generalized to other scenarios like undiscounted finite horizon.} We will assume that the action $A$ and state $S$ spaces are the same across tasks; we use $a\in A$ to denote actions, $s\in S$ to denote states. The transition dynamics $p_i(s'|s,a)$ and reward functions $R_i(a,s)$ are different for each task $i$. Let $\pi_i$ be task-specific stochastic policies.
The dynamics and policies give rise to joint distributions over state and action trajectories starting from some initial state, which we will also denote by $\pi_i$ by an abuse of notation.
Our mechanism for linking the policy learning across tasks is via optimising an objective which consists of expected returns and policy regularizations. We designate $\pi_0$ to be the \emph{distilled policy} which we believe will capture agent behaviour that is common across the tasks. We regularize each task policy $\pi_i$ towards the distilled policy using $\gamma$-discounted
KL divergences $\mathbb{E}_{\pi_i}[\sum_{t\ge 0} \gamma^t \log\frac{\pi_i(a_t|s_t)}{\pi_0(a_t|s_t)}]$. In addition, we also use a $\gamma$-discounted entropy regularization to further encourage exploration. The resulting objective to be maximized is:
\begin{align}
J(\pi_0,\{\pi_i\}_{i=1}^n) =
&\sum_i \mathbb{E}_{\pi_i}\left[ \sum_{t\ge 0} \gamma^t R_i(a_t,s_t)
- c_\text{KL}\gamma^t\log \frac{\pi_i(a_t|s_{t})}{\pi_0(a_t|s_{t})}
- c_\text{Ent}\gamma^t\log {\pi_i(a_t|s_{t})}
\right] \nonumber \\
=&\sum_i \mathbb{E}_{\pi_i}\left[ \sum_{t\ge 0} \gamma^t R_i(a_t,s_t)
+ \frac{\gamma^t\alpha}{\beta}\log \pi_0(a_t|s_{t})
- \frac{\gamma^t}{\beta}\log {\pi_i(a_t|s_{t})}
\right]
\label{eq:objective}
\end{align}
where $c_{\text{KL}}, c_\text{Ent}\ge 0$ are scalar factors which determine the strengths of the KL and entropy regularizations, and $\alpha=c_\text{KL}/(c_\text{KL}+c_\text{Ent})$ and $\beta=1/(c_\text{KL}+c_\text{Ent})$.
The $\log \pi_0(a_t|s_t)$ term can be thought of as a reward shaping term which encourages actions which have high probability under the distilled policy, while the entropy term $-\log \pi_i(a_t|s_{t})$ encourages exploration. In the above we used the same regularization costs $c_\text{KL}$, $c_\text{Ent}$ for all tasks. It is easy to generalize to using task-specific costs; this can be important if tasks differ substantially in their reward scales and amounts of exploration needed, although it does introduce additional hyperparameters that are expensive to optimize.
\subsection{Soft Q-Learning and Distillation}
A range of optimization techniques in the literature can be applied to maximize the above objective, which we will expand on below.
To build up intuition for how the method operates, we will start in the simple case of a tabular representation and an alternating maximization procedure which optimizes over $\pi_i$ given $\pi_0$ and over $\pi_0$ given $\pi_i$. With $\pi_0$ fixed, \eqref{eq:objective} decomposes into separate maximization problems for each task, and is an entropy regularized expected return with redefined (regularized) reward $R'_i(a,s) := R_i(a,s) + \frac{\alpha}{\beta}\log\pi_0(a|s)$. It can be optimized using soft Q-learning aka G learning, which are based on deriving the following ``softened'' Bellman updates for the state and action values (see \cite{Rawlik2012On,FoxPakTis2017a,SchAbbChe2017a,NacNorXu2017a} for derivations):
\begin{align}
V_i(s_t) &=
\frac{1}{\beta} \log\sum_{a_{t}} \pi_{0}^\alpha(a_{t}|s_t)\exp\left[\beta Q_i(a_{t},s_t)\right] \\
Q_i(a_t,s_{t}) & = R_i(a_t,s_t) + \gamma \sum_{s_t} {p_i(s_{t+1}|s_{t}, a_{t})}
V_i(s_{t+1}) \label{eq:Qdef}
\end{align}
The Bellman updates are softened in the sense that the usual max operator over actions for the state values $V_i$ is replaced by a soft-max at inverse temperature $\beta$, which hardens into a max operator as $\beta\rightarrow \infty$.
The optimal policy $\pi_i$ is then a Boltzmann policy at inverse temperature $\beta$:
\begin{align}
\pi_i(a_t|s_{t}) &= \pi_{0}^\alpha(a_t|s_{t}) e^{\beta Q_i(a_t|s_{t})-\beta V_i(s_{t})}
= \pi_{0}^\alpha(a_t|s_{t}) e^{\beta A_i(a_t|s_{t})}
\label{eq:boltzmann}
\end{align}
where $A_i(a,s) = Q_i(a,s) - V_i(s)$ is a softened advantage function.
Note that the softened state values $V_i(s)$ act as the log normalizers in the above. The distilled policy $\pi_0$ can be interpreted as a policy prior, a perspective well-known in the literature on RL as probabilistic inference \cite{Toussaint2006Probabilistic,kappen2012optimal,Rawlik2012On,FoxPakTis2017a}. However, unlike in past works, it is raised to a power of $\alpha\le 1$. This softens the effect of the prior $\pi_0$ on $\pi_i$, and is the result of the additional entropy regularization beyond the KL divergence.
Also unlike past works, we will learn $\pi_0$ instead of hand-picking it (typically as a uniform distribution over actions). In particular, notice that the only terms in \eqref{eq:objective} depending on $\pi_0$ are:
\begin{align}
\frac{\alpha}{\beta}\sum_i \mathbb{E}_{\pi_i}\left[\sum_{t\ge 0} \gamma^t \log \pi_0(a_t|s_t)\right]
\label{eq:distillation}
\end{align}
which is simply a log likelihood for fitting a model $\pi_0$ to a mixture of $\gamma$-discounted state-action distributions, one for each task $i$ under policy $\pi_i$. A maximum likelihood (ML) estimator can be derived from state-action visitation frequencies under roll-outs in each task, with the optimal ML solution given by the mixture of state-conditional action distributions. Alternatively, in the non-tabular case, stochastic gradient ascent can be employed, which leads precisely to an update which distills the task policies $\pi_i$ into $\pi_0$ \cite{BucCarNic06,HinVinDea2014,RusColGul2015a,ParBaSal2016a}. Note however that in our case the distillation step is derived naturally from a KL regularized objective on the policies. Another difference from \cite{RusColGul2015a,ParBaSal2016a} and from prior works on the use of distillation in deep learning \cite{BucCarNic06,HinVinDea2014} is that the distilled policy is ``fed back in'' to improve the task policies when they are next optimized, and serves as a conduit in which common and transferable knowledge is shared across the task policies.
It is worthwhile here to take pause and ponder the effect of the extra entropy regularization. First suppose that there is no extra entropy regularization, $\alpha=1$, and consider the simple scenario of only $n=1$ task.
Then \eqref{eq:distillation} is maximized when the distilled policy $\pi_0$ and the task policy $\pi_1$ are equal, and the KL regularization term is 0. Thus the objective reduces to an unregularized expected return, and so the task policy $\pi_1$ converges to a greedy policy which locally maximizes expected returns. Another way to view this line of reasoning is that the alternating maximization scheme is equivalent to trust-region methods like natural gradient or TRPO \cite{PascanuNatural14, Schulman2015Trust} which use a KL ball centred at the previous policy, and which are understood to converge to greedy policies.
If $\alpha<1$, there is an additional entropy term in \eqref{eq:objective}. So even with $\pi_0=\pi_1$ and $\mathsf{KL}(\pi_1\|\pi_0)=0$, the objective \eqref{eq:objective} will no longer be maximized by greedy policies. Instead \eqref{eq:objective} reduces to an entropy regularized expected returns with entropy regularization factor $\beta'= \beta/(1-\alpha) = 1/c_\text{Ent}$, so that the optimal policy is of the Boltzmann form with inverse temperature $\beta'$ \cite{Rawlik2012On,FoxPakTis2017a,SchAbbChe2017a,NacNorXu2017a}. In conclusion, by including the extra entropy term, we can guarantee that the task policy will not turn greedy, and we can control the amount of exploration by adjusting $c_\text{Ent}$ appropriately.
This additional control over the amount of exploration is essential when there are more than one task. To see this, imagine a scenario where one of the tasks is easier and is solved first, while other tasks are harder with much sparser rewards.
Without the entropy term, and before rewards in other tasks are encountered, both the distilled policy and all the task policies can converge to the one that solves the easy task. Further, because this policy is greedy, it can insufficiently explore the other tasks to even encounter rewards, leading to sub-optimal behaviour.
For single-task RL, the use of entropy regularization was recently popularized by Mnih et al.~\cite{MniBadMir2016a} to counter premature convergence to greedy policies, which can be particularly severe when doing policy gradient learning. This carries over to our multitask scenario as well, and is the reason for the additional entropy regularization.
\subsection{Policy Gradient and a Better Parameterization}
The above method alternates between maximization of the distilled policy $\pi_0$ and the task policies $\pi_i$, and is reminiscent of the EM algorithm \cite{DemLaiRub1977a} for learning latent variable models, with $\pi_0$ playing the role of parameters, while $\pi_i$ plays the role of the posterior distributions for the latent variables. Going beyond the tabular case, when both $\pi_0$ and $\pi_i$ are parameterized by, say, deep networks, such an alternating maximization procedure can be slower than simply optimizing \eqref{eq:objective} with respect to task and distilled policies jointly by stochastic gradient ascent. In this case the gradient update for $\pi_i$ is simply given by policy gradient with an entropic regularization \cite{MniBadMir2016a,SchAbbChe2017a}, and can be carried out within a framework like advantage actor-critic \cite{MniBadMir2016a}.
A simple parameterization of policies would be to use a separate network for each task policy $\pi_i$, and another one for the distilled policy $\pi_0$. An alternative parameterization, which we argue can result in faster transfer, can be obtained by considering the form of the optimal Boltzmann policy \eqref{eq:boltzmann}. Specifically, consider parameterizing the distilled policy using a network with parameters $\theta_0$,
\begin{align}
\hat{\pi}_{0}(a_t|s_{t}) &= \frac
{\exp(h_{\theta_0}(a_t|s_{t})}
{\sum_{a'} \exp(h_{\theta_0}(a'|s_{t}))}
\end{align}
and estimating the soft advantages\footnote{In practice, we do not actually use these as advantage estimates. Instead we use \eqref{eq:pi_i} to parameterize a policy which is optimized by policy gradients.} using another network with parameters $\theta_i$:
\begin{align}
\hat{A}_{i}(a_t|s_t)
&= f_{\theta_i}(a_t|s_t) - \frac{1}{\beta}\log\sum_a \hat{\pi}_{0}^\alpha(a|s_t)\exp(\beta f_{\theta_i}(a|s_t))
\label{eq:advantages}
\end{align}
We used hat notation to denote parameterized approximators of the corresponding quantities.
The policy for task $i$ then becomes parameterized as,
\begin{align}
\hat{\pi}_{i}(a_t|s_{t}) &= \hat{\pi}_{0}^\alpha(a_t|s_{t}) \exp(\beta \hat{A}_{i}(a_t|s_{t}) ) = \frac
{\exp(\alpha h_{\theta_0}(a_t|s_{t}) + \beta f_{\theta_i}(a_t|s_t))} {\sum_{a'} \exp((\alpha h_{\theta_0}(a'|s_{t}) + \beta f_{\theta_i}(a'|s_t))}
\label{eq:pi_i}
\end{align}
This can be seen as a two-column architecture for the policy, with one column being the distilled policy, and the other being the adjustment required to specialize to task $i$.
Given the parameterization above, we can now derive the policy gradients. The gradient wrt to the task specific parameters $\theta_i$ is given by the standard policy gradient theorem \cite{sutton1999policy},
\begin{align}
\nabla_{\theta_i} J =&
\mathbb{E}_{\hat{\pi}_{i}}\left[ \left(\textstyle \sum_{t\ge 1}
\nabla_{\theta_i} \log \hat{\pi}_{i}(a_t|s_{t}) \right)
\left (\textstyle \sum_{u\ge 1} \gamma^u (R^\text{reg}_i(a_u,s_u)
) \right)
\right] \nonumber \\
=&
\mathbb{E}_{\hat{\pi}_{i}}\left[ \textstyle \sum_{t\ge 1}
\nabla_{\theta_i} \log \hat{\pi}_{i}(a_t|s_{t})
\left (\sum_{u\ge t} \gamma^u (R^\text{reg}_i(a_u,s_u)
) \right)
\right]
\label{eq:policygrad}
\end{align}
where $R^\text{reg}_i(a,s)=R_i(a,s)
+ \frac{\alpha}{\beta}\log \hat{\pi}_{0}(a|s)
- \frac{1}{\beta}\log {\hat{\pi}_{i}(a|s)}$ is the regularized reward.
Note that the partial derivative of the entropy in the integrand has expectation $\mathbb{E}_{\hat{\pi}_{i}}[\nabla_{\theta_i}\log\hat{\pi}_{i}(a_t|s_t)]=0$ because of the log-derivative trick. If a value baseline is estimated, it can be subtracted from the regularized returns in order to reduce gradient variance.
The gradient wrt $\theta_0$ is more interesting:
\begin{align}
\nabla_{\theta_0} J =&
\sum_i \mathbb{E}_{\hat{\pi}_{i}}\left[ \textstyle
\sum_{t\ge 1}
\nabla_{\theta_0} \log \hat{\pi}_{i}(a_t|s_{t})
\left (\sum_{u\ge t} \gamma^u( R^\text{reg}_i(a_u,s_u)
\right)
\right] \label{eq:policygrad_2} \\
& + \frac{\alpha}{\beta} \sum_i \mathbb{E}_{\hat{\pi}_{i}}\left[ \textstyle
\sum_{t\ge 1} \gamma^t
\sum_{a'_t}(\hat{\pi}_{i}(a'_t|s_{t}) - \hat{\pi}_{0}(a'_t|s_{t}))
\nabla_{\theta_0} h_{\theta_0}(a'_t|s_{t}) \right]
\nonumber
\end{align}
Note that the first term is the same as for the policy gradient of $\theta_i$. The second term tries to match the probabilities under the task policy $\hat{\pi}_{i}$ and under the distilled policy $\hat{\pi}_{0}$. The second term would not be present if we simply parameterized $\pi_i$ using the same architecture $\hat{\pi}_{i}$, but do not use a KL regularization for the policy. The presence of the KL regularization gets the distilled policy to learn to be the centroid of all task policies, in the sense that the second term would be zero if $\hat{\pi}_{0}(a'_t|s_t) = \frac{1}{n}\sum_i \hat{\pi}_{i}(a'_t|s_t)$, and helps to transfer information quickly across tasks and to new tasks.
The centroid and star-shaped structure of DisTraL is reminiscent of ADMM \cite{Boyd2011}, elastic-averaging SGD \cite{ZhaChoLec2015} and hierarchical Bayes \cite{gelman2014bayesian}. Though a crucial difference is that while ADMM, EASGD and hierarchical Bayes operate in the space of parameters, in Distral the distilled policy learns to be the centroid in the space of policies.
We argue that this is semantically more meaningful, and may contribute to the observed robustness of Distral by stabilizing learning. In our experiments we find indeed that absence of the KL regularization significantly affects the stability of the algorithm.
Our approach is also reminiscent of recent work on option learning \cite{fox2016principled}, but with a few important differences. We focus on using deep neural networks as flexible function approximators, and applied our method to rich 3D visual environments, while Fox et al.~\cite{fox2016principled} considered only the tabular case. We argue for the importance of an additional entropy regularization besides the KL regularization. This lead to an interesting twist in the mathematical framework allowing us to separately control the amounts of transfer and of exploration.
On the other hand Fox et al.~\cite{fox2016principled} focused on the interesting problem of learning multiple options (distilled policies here). Their approach treats the assignment of tasks to options as a clustering problem, which is not easily extended beyond the tabular case.
\section{Algorithms}\label{sec:Algorithms}
\begin{figure}[t]
\centering
\includegraphics[width=1\textwidth]{DisTraL_1.pdf}
\caption{Depiction of the different algorithms and baselines. On the left are two of the Distral algorithms and on the right are the three A3C baselines. Entropy is drawn in brackets as it is optional and only used for KL+ent 2col and KL+ent 1col.}
\label{fig:trajectories}
\end{figure}
The framework we just described allows for a number of possible algorithmic instantiations, arising as combinations of objectives, algorithms and architectures, which we describe below and summarize in Table~\ref{table:algos} and Figure~\ref{fig:trajectories}.
\textit{KL divergence vs entropy regularization}: With $\alpha=0$, we get a purely entropy-regularized objective which does not couple and transfer across tasks \cite{MniBadMir2016a,SchAbbChe2017a}. With $\alpha=1$, we get a purely KL regularized objective, which does couple and transfer across tasks, but might prematurely stop exploration if the distilled and task policies become similar and greedy. With $0<\alpha<1$ we get both terms.
\textit{Alternating vs joint optimization}: We have the option of jointly optimizing both the distilled policy and the task policies, or optimizing one while keeping the other fixed. Alternating optimization leads to algorithms that resemble policy distillation/actor-mimic \cite{ParBaSal2016a,RusColGul2015a}, but are iterative in nature with the distilled policy feeding back into task policy optimization. Also, soft Q-learning can be applied to each task, instead of policy gradients.
While alternating optimization can be slower, evidence from policy distillation/actor-mimic indicate it might learn more stably, particularly for tasks which differ significantly.
\textit{Separate vs two-column parameterization}: Finally, the task policy can be parameterized to use the distilled policy \eqref{eq:pi_i} or not. If using the distilled policy, behaviour distilled into the distilled policy is ``immediately available'' to the task policies so transfer can be faster. However if the process of transfer occurs too quickly, it might prevent effective exploration of individual tasks.
From this spectrum of possibilities we consider four concrete instances which differ in the underlying network architecture and distillation loss, identified in Table~\ref{table:algos}. In addition, we compare against three A3C baselines.
In initial experiments we explored two variants of A3C: the original method~\cite{MniBadMir2016a} and the variant of Schulman et al. ~\cite{SchAbbChe2017a} which uses entropy regularized returns. We did not find significant differences for the two variants in our setting, and chose to report only the original A3C results for clarity in Section~\ref{sec:Experiments}. Further algorithmic details are provided in the Appendix.
\begin{table}
\begin{center}
\begin{tabular}{@{}l|lll@{}}
\toprule
&$h_{\theta_0}(a| s)$ & $f_{\theta_i}(a|s)$ &
$\alpha h_{\theta_0}(a| s) + \beta f_{\theta_i}(a|s)$\\
\midrule
$\alpha=0$&\textrm{\algoname{A3C multitask}}&\textrm{\algoname{A3C}}&\textrm{\algoname{A3C 2col}} \\
$\alpha=1$&&\textrm{\algoname{KL 1col}}&\textrm{\algoname{KL 2col}}\\
$0<\alpha<1$&&\textrm{\algoname{KL+ent 1col}}&\textrm{\algoname{KL+ent 2col}}\\
\bottomrule
\end{tabular}
\end{center}
\caption{
The seven different algorithms evaluated in our experiments. Each column describes a different architecture, with the column headings indicating the logits for the task policies.
The rows define the relative amount of KL vs entropy regularization loss, with the first row comprising the A3C baselines (no KL loss).
}
\label{table:algos}
\end{table}
\section{Experiments}\label{sec:Experiments}
We demonstrate the various algorithms derived from our framework, firstly using alternating optimization with soft Q-learning and policy distillation on a set of simple grid world tasks. Then all seven algorithms will be evaluated on three sets of challenging RL tasks in partially observable 3D environments \cite{beattie2016deepmind}.
\subsection{Two room grid world}
To give better intuition for the role of the distilled behaviour policy, we considered a set of tasks in a grid world domain with two rooms connected by a corridor (see Figure~\ref{fig:gridworld}) \cite{fox2016principled}. Each task is distinguished by a different randomly chosen goal location and each MDP state consists of the map location, the previous action and the previous reward. A Distral agent is trained using only the KL regularization and an optimization algorithm which alternates between soft Q-learning and policy distillation. Each soft Q-learning iteration learns using a rollout of length 10.
To determine the benefit of the distilled policy, we compared the Distral agent to one which soft Q learns a separate policy for each task.
The learning curves are shown in Figure~\ref{fig:gridworld} (left). We see that the Distral agent is able to learn significantly faster than single-task agents.
Figure~\ref{fig:gridworld} (right) visualizes the distilled policy (probability of next action given position and previous action), demonstrating that the agent has learnt a robust policy which guides the agent to move consistently in the corridor in order to reach the other room. This allows the agent to reach the other room faster and helps exploration, if the agent is shown new test tasks. In Fox et al.~\cite{fox2016principled} two separate options are learnt, while here we learn a single distilled policy which conditions on more past information.
\begin{figure}
\begin{center}
\begin{tabular}{ccccc}
\includegraphics[height=1.25in, trim=0 0 5in 0, clip]{gw-meanlearncurve.pdf} & &
\includegraphics[height=1.25in]{maze_random_goal.pdf} & &
\includegraphics[height=1.25in]{maze_layout2.pdf}
\end{tabular}
\end{center}
\caption{\textbf{Left:} Learning curves on two room grid world.
The DisTraL agent (blue) learns faster, converges towards better policies, and demonstrates more stable learning overall. \textbf{Center:} Example of tasks. Green is goal position which is uniformly sampled for each task. Starting position is uniformly sampled at the beginning of each episode. \textbf{Right:} depiction of learned distilled policy $\pi_0$ only in the corridor, conditioned on previous action being left/right and no previous reward. Sizes of arrows depict probabilities of actions. Note that up/down actions have negligible probabilities. The model learns to preserve direction of travel in the corridor. }
\label{fig:gridworld}
\end{figure}
\subsection{Complex Tasks}
\label{sec:lab_results}
To assess Distral under more challenging conditions, we use a complex first-person partially observed 3D environment with a variety of visually-rich RL tasks \cite{beattie2016deepmind}.
All agents were implemented with a distributed Python/TensorFlow code base, using 32 workers for each task and learnt using asynchronous RMSProp. The network columns contain convolutional layers and an LSTM and are uniform across experiments and algorithms. We tried three values for the entropy costs $\beta$ and three learning rates $\epsilon$. Four runs for each hyperparameter setting were used. All other hyperparameters were fixed to the single-task A3C defaults and, for the \algoname{KL+ent 1col} and \algoname{KL+ent 2col} algorithms, $\alpha$ was fixed at $0.5$.
\textbf{Mazes} In the first experiment, each of $n=8$ tasks is a different maze containing randomly placed rewards and a goal object. Figure~\ref{fig:maze}.A1 shows the learning curves for all seven algorithms. Each curve is produced by averaging over all 4 runs and 8 tasks, and selecting the best settings for $\beta$ and $\epsilon$ (as measured by the area under the learning curves). The Distral algorithms learn faster and achieve better final performance than all three A3C baselines. The two-column algorithms learn faster than the corresponding single column ones. The Distral algorithms without entropy learn faster but achieve lower final scores than those with entropy, which we believe is due to insufficient exploration towards the end of learning.
We found that both multitask A3C and two-column A3C can learn well on some runs, but are generally unstable---some runs did not learn well, while others may learn initially then suffer degradation later. We believe this is due to negative interference across tasks, which does not happen for Distral algorithms.
The stability of Distral algorithms also increases their robustness to hyperparameter selection. Figure~\ref{fig:maze}.A2 shows the final achieved average returns for all 36 runs for each algorithm, sorted in decreasing order. We see that Distral algorithms have a significantly higher proportion of runs achieving good returns, with \algoname{KL+ent\_2col} being the most robust.
DisTraL algorithms, along with multitask A3C, use a distilled or common policy which can be applied on all tasks. Panels B1 and B2 in Figure~\ref{fig:maze} summarize the performances of the distilled policies. Algorithms that use two columns (\algoname{KL\_2col} and \algoname{KL+ent\_2col}) obtain the best performance, because policy gradients are also directly propagated through the distilled policy in those cases. Moreover, panel B2 reveals that Distral algorithms exhibit greater stability as compared to traditional multitask A3C. We also observe that \algoname{KL} algorithms have better-performing distilled policies than \algoname{KL+ent} ones. We believe this is because the additional entropy regularization allows task policies to diverge more substantially from the distilled policy. This suggests that annealing the entropy term or increasing the KL term throughout training could improve the distilled policy performance, if that is of interest.
\textbf{Navigation} We experimented with $n=4$ navigation and memory tasks. In contrast to the previous experiment, these tasks use random maps which are procedurally generated on every episode.
The first task features reward objects which are randomly placed in a maze, and the second task requires to return these objects to the agent's start position. The third task has a single goal object which must be repeatedly found from different start positions, and on the fourth task doors are randomly opened and closed to force novel path-finding. Hence, these tasks are more involved than the previous navigation tasks.
The panels C1 and C2 of Figure~\ref{fig:maze} summarize the results. We observe again that Distral algorithms yield better final results while having greater stability (Figure~\ref{fig:maze}.C2). The top-performing algorithms are, again, the 2 column Distral algorithms (\algoname{KL\_2col} and \algoname{KL+ent\_2col}).
\textbf{Laser-tag} In the final set of experiments, we use $n=8$ laser-tag levels from DeepMind Lab.
These tasks require the agent to learn to tag bots controlled by a built-in AI, and differ substantially: fixed versus procedurally generated maps, fixed versus procedural bots, and complexity of agent behaviour (e.g. learning to jump in some tasks). Corresponding to this greater diversity, we observe (see panels D1 and D2 of Figure~\ref{fig:maze}) that the best baseline is the A3C algorithm that is trained independently on each task.
Among the Distral algorithms, the single column variants perform better, especially initially, as they are able to learn task-specific features separately. We observe again the early plateauing phenomenon for algorithms that do not possess an additional entropy term. While not significantly better than the A3C baseline on these tasks, the Distral algorithms clearly outperform the multitask A3C. Considering the 3 different sets of complex 3D experiments, we argue that the Distral algorithms are the most promising solution to the multitask RL problem.
\begin{figure}[!t]
\begin{center}
\includegraphics[width=\linewidth]{maze_final_summary_4_a.png}
\includegraphics[width=\linewidth]{rat_and_lt_final_summary_4.png}
\caption{Panels A1, C1, D1 show task specific policy performance (averaged across all the tasks) for the maze, navigation and laser-tag tasks, respectively. The $x$-axes are total numbers of training environment steps per task. Panel B1 shows the mean scores obtained with the distilled policies (\algoname{A3C} has no distilled policy, so it is represented by the performance of an untrained network.). For each algorithm, results for the best set of hyperparameters (based on the area under curve) are reported. The bold line is the average over 4 runs, and the colored area the average standard deviation over the tasks.
Panels A2, B2, C2, D2 shows the corresponding final performances for the 36 runs of each algorithm ordered by best to worst (9 hyperparameter settings and 4 runs).}
\label{fig:maze}
\end{center}
\end{figure}
\section{Discussion}\label{sec:Discussion}
We have proposed Distral, a general framework for distilling and transferring common behaviours in multitask reinforcement learning. In experiments we showed that the resulting algorithms learn quicker, produce better final performances, and are more stable and robust to hyperparameter settings. We have found that Distral significantly outperforms the standard way of using shared neural network parameters for multitask or transfer reinforcement learning.
Two ideas might be worth reemphasizing here. We observe that distillation arises naturally as one half of an optimization procedure when using KL divergences to regularize the output of task models towards a distilled model. The other half corresponds to using the distilled model as a regularizer for training the task models. Another observation is that parameters in deep networks do not typically by themselves have any semantic meaning, so instead of regularizing networks in parameter space, it is worthwhile considering regularizing networks in a more semantically meaningful space, e.g.\ of policies.
Possible directions of future research include: combining Distral with techniques which use auxiliary losses \cite{Jaderber2016Reinforcement,Mirowski2016Learning,Lample2016Playing}, exploring use of multiple distilled policies or latent variables in the distilled policy to allow for more diversity of behaviours, exploring settings for continual learning where tasks are encountered sequentially, and exploring ways to adaptively adjust the KL and entropy costs to better control the amounts of transfer and exploration.
\newpage
\footnotesize
\setlength{\bibsep}{5pt}
\bibliographystyle{plain}
|
\section{Introduction}
\input{intro.tex}
\section{Theoretical results}\label{sec:theo}
\subsection{Case where $\Sigma$ is known}
Let us first introduce some notations. Let
\begin{equation} \label{eq:C:J}
C=\frac{1}{nq}\mathcal{X}'\mathcal{X} \textrm{\, and \,}
J=\{1 \leq j \leq pq, \mathcal{B}_j \neq 0\},
\end{equation}
where $\mathcal{X}$ is defined in (\ref{eq:def:Y_X_E}) and where $\mathcal{B}_j$ denotes the $j$th component of the vector $\mathcal{B}$ defined in \eqref{eq:def:Y_X_E}.
Let also define
\begin{equation} \label{eq:CJJ}
C_{J,J} = \frac{1}{nq} (\mathcal{X}_{\bullet,J})'\mathcal{X}_{\bullet,J} \textrm{\, and \,} C_{J^c,J} = \frac{1}{nq} (\mathcal{X}_{\bullet,J^c})'\mathcal{X}_{\bullet,J},
\end{equation}
where $\mathcal{X}_{\bullet,J}$ and $\mathcal{X}_{\bullet,J^c}$ denote the columns of $\mathcal{X}$ belonging to the set $J$ defined in (\ref{eq:C:J}) and to its complement $J^c$, respectively.
More generally, for any matrix $A$, $A_{I,J}$ denotes the partitioned matrix extracted from $A$ by considering the rows of $A$ belonging to the set $I$ and
the columns of $A$ belonging to the set $J$, with $\bullet$ indicating all the rows or all the columns.
The following theorem gives some conditions under which the estimator $\widehat{\mathcal{B}}$ defined in (\ref{eq:lasso}) is sign-consistent as defined by \cite{zhao:yu:2006}, namely,
$$
\mathbb{P}\left(\ensuremath{\rm sign}(\widehat{\mathcal{B}}) = \ensuremath{\rm sign}(\mathcal{B})\right)\to 1, \textrm{\, as \,} n\to\infty,
$$
where the $\ensuremath{\rm sign}$ function maps positive entries to 1, negative entries to -1 and zero to 0.
\begin{theo}\label{th1}
Assume that $\mathcal{Y}=(\mathcal{Y}_1,\mathcal{Y}_2,\ldots,\mathcal{Y}_{nq})'$ satisfies Model \eqref{eq:model_vec}. Assume also that
there exist some positive constants $M_1$, $M_2$, $M_3$ and positive numbers $c_1$, $c_2$ such that $0<c_1 +c_2<1/2$ satisfying:
\begin{enumerate}[label=\textbf{\rm (A\arabic*)}]
\item \label{th1(i)} for all $n\geq 1$, for all $j\in\{1,\ldots,pq\}$, $\frac{1}{n} (\mathcal{X}_{\bullet,j})'\mathcal{X}_{\bullet,j} \leq M_1$, where $\mathcal{X}_{\bullet,j}$ is the $j$th column of $\mathcal{X}$ defined in (\ref{eq:def:Y_X_E}),
\item \label{th1(ii)} for all $n\geq 1$, $\frac{1}{n}\lambda_{\rm{min}}\left((\mathcal{X}'\mathcal{X})_{J,J}\right)\geq M_2$, where $\lambda_{\rm{min}}(A)$ denotes the smallest eigenvalue of $A$,
\item \label{th1(iii)} $|J|=O(q^{c_1})$, where $J$ is defined in (\ref{eq:C:J}) and $|J|$ is the cardinality of the set $J$,
\item \label{th1(iv)} $q^{c_2} \min_{j \in J} |\mathcal{B}_j| \geq M_3$.
\end{enumerate}
Assume also that the following strong Irrepresentable Condition holds:
\begin{enumerate}[label=\textbf{\rm (IC)}]
\item \label{eq:irrep} There exists a positive constant vector $\eta$ such that
$$
\left|(\mathcal{X}'\mathcal{X})_{J^c,J}((\mathcal{X}'\mathcal{X})_{J,J})^{-1} \, \ensuremath{\rm sign}(\mathcal{B}_J)\right|\leq \mathbf{1}-\eta,
$$
where $\mathbf{1}$ is a $(pq-|J|)$ vector of 1 and the inequality holds element-wise.
\end{enumerate}
Then, for all $\lambda$ that satisfies
\begin{enumerate}[label=\textbf{\rm (L)}]
\item \label{assum:lambda_nq}
$\displaystyle \quad q=q_n=o\left(n^{\frac{1}{2(c_1+c_2)}}\right), \quad \frac{\lambda}{\sqrt{n}}\to\infty \textrm{\, and \,} \frac{\lambda}{n} =o\left(q^{-(c_1+c_2)}\right),\textrm{ as } n\to\infty,$
\end{enumerate}
we have
$$
\mathbb{P}\left(\ensuremath{\rm sign}(\widehat{\mathcal{B}}(\lambda)) = \ensuremath{\rm sign}(\mathcal{B})\right)\to 1, \textrm{\, as \,} n\to\infty,
$$
where $\widehat{\mathcal{B}}(\lambda)$ is defined by (\ref{eq:lasso}).
\end{theo}
\begin{remark}
Observe that if $c_1+c_2<(2k)^{-1}$, for some positive $k$, then the first condition of \ref{assum:lambda_nq} becomes $q=o(n^k)$. Hence for large values of $k$, the size $q$ of $\Sigma$
is much larger than $n$.
\end{remark}
The proof of Theorem~\ref{th1} is given in Section \ref{sec:proofs}.
It is based on Proposition~\ref{prop1} which is an adaptation to the multivariate case of Proposition 1 in \cite{zhao:yu:2006}.
\begin{prop}\label{prop1}
Let $\widehat{\mathcal{B}}(\lambda)$ be defined by (\ref{eq:lasso}). Then
$$
\mathbb{P}\left(\ensuremath{\rm sign}(\widehat{\mathcal{B}}(\lambda))=\ensuremath{\rm sign}(\mathcal{B})\right)\geq\mathbb{P}(A_n\cap B_n),
$$
where
\begin{equation}\label{eq:An}
A_n=\left\{\left|(C_{J,J})^{-1}W_J\right|<\sqrt{nq}\left(|\mathcal{B}_J|-\frac{\lambda}{2nq}|(C_{J,J})^{-1}\ensuremath{\rm sign}(\mathcal{B}_J)|\right)\right\}
\end{equation}
and
\begin{equation}\label{eq:Bn}
B_n=\left\{\left|C_{J^c,J}(C_{J,J})^{-1}W_J-W_{J^c}\right|\leq \frac{\lambda}{2\sqrt{nq}}\left(\mathbf{1}-\left|C_{J^c,J}(C_{J,J})^{-1}\ensuremath{\rm sign}(\mathcal{B}_J)\right|\right)
\right\},
\end{equation}
with $
W=\mathcal{X}'\mathcal{E}/\sqrt{nq}.
$
In (\ref{eq:An}) and (\ref{eq:Bn}), $C_{J,J}$ and $C_{J^c,J}$ are defined in (\ref{eq:CJJ}) and $W_J$ and $W_{J^c}$ denote the components of $W$ being in $J$ and $J^c$, respectively.
Note that the previous inequalities hold element-wise.
\end{prop}
The proof of Proposition \ref{prop1} is given in Section \ref{sec:proofs}.
We give in the following proposition which is proved in Section \ref{sec:proofs} some conditions on $X$ and $\Sigma$ under which Assumptions~\ref{th1(i)}~and~\ref{th1(ii)} of Theorem~\ref{th1} hold.
\begin{prop}\label{propcond}
If there exist some positive constants $M'_1$, $M'_2$, $m_1$, $m_2$ such that, for all $n\geq 1$,
\begin{enumerate}[label=\textbf{\rm (C\arabic*)}]
\item \label{cond1} for all $j\in\{1,\ldots,p\}$, $\frac{1}{n} (X_{\bullet,j})'X_{\bullet,j}\leq M'_1$,
\item \label{cond2} $\frac{1}{n}\lambda_{\rm{min}}(X'X) \geq M'_2$,
\item \label{cond3} $\lambda_{\rm{max}}(\Sigma^{-1}) \leq m_1$,
\item \label{cond4} $\lambda_{\rm{min}}(\Sigma^{-1}) \geq m_2$,
\end{enumerate}
then Assumptions~\ref{th1(i)}~and~\ref{th1(ii)} of Theorem~\ref{th1} are satisfied.
\end{prop}
\begin{remark}
Observe that \ref{cond1} and \ref{cond2} hold in the case where the columns of the matrix $X$ are orthogonal.
\end{remark}
We give in Proposition \ref{prop_IC} in Section \ref{subsec:AR1} some conditions under which Condition \ref{eq:irrep} holds in the specific case where $\Sigma$ is the covariance
matrix of an AR(1) process.
\subsection{Case where $\Sigma$ is unknown}
Similarly as in (\ref{eq:C:J}) and (\ref{eq:CJJ}), we introduce the following notations:
\begin{equation}\label{eq:C_tilde}
\widetilde{C}=\frac{1}{nq}\widetilde{\mathcal{X}}'\widetilde{\mathcal{X}}
\end{equation}
and
\begin{equation} \label{eq:CJJ_tilde}
\widetilde{C}_{J,J} = \frac{1}{nq} (\widetilde{\mathcal{X}}_{\bullet,J})'\widetilde{\mathcal{X}}_{\bullet,J} \textrm{\, and \,} \widetilde{C}_{J^c,J} = \frac{1}{nq} (\widetilde{\mathcal{X}}_{\bullet,J^c})'\widetilde{\mathcal{X}}_{\bullet,J},
\end{equation}
where $\widetilde{\mathcal{X}}_{\bullet,J}$ and $\widetilde{\mathcal{X}}_{\bullet,J^c}$ denote the columns of $\widetilde{\mathcal{X}}$ belonging to the set $J$ defined in \eqref{eq:C:J} and to its complement $J^c$, respectively.
A straightforward extension of Proposition \ref{prop1} leads to the following proposition for Model~(\ref{eq:model_vec_tilde}).
\begin{prop} \label{prop2}
Let $\widetilde{\mathcal{B}}(\lambda)$ be defined by (\ref{eq:lasso_tilde}). Then
$$
\mathbb{P}\left(\ensuremath{\rm sign}(\widetilde{\mathcal{B}}(\lambda))=\ensuremath{\rm sign}(\mathcal{B})\right)\geq\mathbb{P}(\widetilde{A_n}\cap \widetilde{B_n}),
$$
where
\begin{equation}\label{eq:An_tilde}
\widetilde{A_n}=\left\{\left|(\widetilde{C}_{J,J})^{-1}\widetilde{W}_J\right|<\sqrt{nq}\left(|\mathcal{B}_J|-\frac{\lambda}{2nq}|(\widetilde{C}_{J,J})^{-1}\ensuremath{\rm sign}(\mathcal{B}_J)|\right)\right\}
\end{equation}
and
\begin{equation}\label{eq:Bn_tilde}
\widetilde{B_n}=\left\{\left|\widetilde{C}_{J^c,J}(\widetilde{C}_{J,J})^{-1}\widetilde{W}_J-\widetilde{W}_{J^c}\right|\leq \frac{\lambda}{2\sqrt{nq}}\left(\mathbf{1}-\left|\widetilde{C}_{J^c,J}(\widetilde{C}_{J,J})^{-1}\ensuremath{\rm sign}(\mathcal{B}_J)\right|\right)
\right\},
\end{equation}
with $
W=\widetilde{\mathcal{X}}'\widetilde{\mathcal{E}}/\sqrt{nq}.
$
In (\ref{eq:An_tilde}) and (\ref{eq:Bn_tilde}), $\widetilde{C}_{J,J}$ and $\widetilde{C}_{J^c,J}$ are defined in (\ref{eq:CJJ_tilde}) and $\widetilde{W}_J$ and $\widetilde{W}_{J^c}$ denote the components of $\widetilde{W}$ being in $J$ and $J^c$, respectively.
Note that the previous inequalities hold element-wise.
\end{prop}
The following theorem extends Theorem \ref{th1} to the case where $\Sigma$ is unknown and gives some conditions under which the estimator $\widetilde{\mathcal{B}}$ defined in
(\ref{eq:lasso_tilde}) is sign-consistent. The proof of Theorem \ref{th2}
is given in Section \ref{sec:proofs} and is based on Proposition \ref{prop2}.
\begin{theo}\label{th2}
Assume that Assumptions \ref{th1(i)}, \ref{th1(ii)}, \ref{th1(iii)}, \ref{th1(iv)}, \ref{eq:irrep} and \ref{assum:lambda_nq} of Theorem~\ref{th1} hold. Assume also that, there exist some positive
constants $M_4$, $M_5$, $M_6$ and $M_7$, such that for all $n\geq 1$,
\begin{enumerate}[label=\textbf{\rm (A\arabic*)}] \setcounter{enumi}{4}
\item \label{th2(v)} $\|(X'X)/n\|_\infty\leq M_4$,
\item \label{th2(vi)} $\lambda_{\rm{min}}((X'X)/n)\geq M_5$,
\item \label{th2(vii)} $\lambda_{\rm{max}}(\Sigma^{-1}) \leq M_6$,
\item \label{th2(viii)} $\lambda_{\rm{min}}(\Sigma^{-1}) \geq M_7$.
\end{enumerate}
Suppose also that
\begin{enumerate}[label=\textbf{\rm (A\arabic*)}] \setcounter{enumi}{8}
\item \label{th2(ix)} $\|\Sigma^{-1}-\widehat{\Sigma}^{-1}\|_{\infty}=O_P((nq)^{-1/2})$, as $n$ tends to infinity,
\item \label{th2(x)} $\rho(\Sigma-\widehat{\Sigma})=O_P((nq)^{-1/2})$, as $n$ tends to infinity.
\end{enumerate}
Let $\widetilde{\mathcal{B}}(\lambda)$ be defined by (\ref{eq:lasso_tilde}), then
$$
\mathbb{P}\left(\textrm{sign}(\widetilde{\mathcal{B}}(\lambda))=\textrm{sign}(\mathcal{B})\right)\to 1,\textrm{ as } n\to\infty.
$$
In the previous assumptions, $\lambda_{\rm{max}}(A)$, $\lambda_{\rm{min}}(A)$, $\rho(A)$ and $\|A\|_\infty$ denote the largest eigenvalue, the smallest
eigenvalue, the spectral radius and the infinite norm (induced by the associated vector norm) of the matrix $A$.
\end{theo}
\begin{remark}
Observe that Assumptions \ref{th2(v)} and \ref{th2(vi)} hold in the case where the columns of the matrix $X$ are orthogonal.
Note also that \ref{th2(vii)} and \ref{th2(viii)} are the same as \ref{cond3} and \ref{cond4} in Proposition \ref{propcond}.
\end{remark}
In order to estimate $\Sigma$, we propose the following strategy:
\begin{itemize}
\item Fitting a classical linear model to each column of the matrix
$Y$ in order to have access to an
estimation $\widehat{E}$ of the random error matrix
$E$. It is possible since $p$ is assumed to be fixed and smaller than $n$.
\item Estimating $\Sigma$ from $\widehat{E}$ by assuming that $\Sigma$ has a particular structure, Toeplitz for instance.
\end{itemize}
More precisely, $\widehat{E}$ defined in the first step is such that:
\begin{equation}\label{eq: Ehat}
\widehat{E}=\left(\textrm{Id}_{\mathbb{R}^n}-X(X'X)^{-1}X'\right) E=:\Pi E,
\end{equation}
which implies that
\begin{equation}\label{eq:Ehat_1}
\widehat{\mathcal{E}}=vec(\widehat{E})=\left[\textrm{Id}_{\mathbb{R}^q}\otimes\Pi\right]\mathcal{E},
\end{equation}
where $\mathcal{E}$ is defined in (\ref{eq:def:Y_X_E}).
We prove in Proposition \ref{prop:cond_th_AR} below that our strategy for estimating $\Sigma$ provides an estimator satisfying the assumptions of Theorem \ref{th2}
in the case where $(E_{1,t})_t$, $(E_{2,t})_t$, ..., $(E_{n,t})_t$ are assumed to be independent AR(1) processes.
\subsection{The AR(1) case}\label{subsec:AR1}
\subsubsection{Sufficient conditions for Assumption \ref{eq:irrep} of Theorem \ref{th1}}
The following proposition gives some conditions under which the strong Irrepresentable Condition \ref{eq:irrep} of Theorem \ref{th1} holds.
\begin{prop}\label{prop_IC}
Assume that $(E_{1,t})_t$, $(E_{2,t})_t$, ..., $(E_{n,t})_t$ in Model (\ref{eq:model:matriciel}) are independent AR(1) processes satisfying:
$$
E_{i,t}-\phi_1 E_{i,t-1}=Z_{i,t},\; \forall i\in\{1,\dots,n\},
$$
where the $Z_{i,t}$'s are zero-mean i.i.d. Gaussian random variables with variance $\sigma^2$ and $|\phi_1|<1$.
Assume also that $X$ defined in (\ref{eq:model:matriciel}) is such that $X'X=\nu\textrm{Id}_{\mathbb{R}^p}$, where $\nu$ is a positive constant.
Moreover, suppose that if $j\in J$, then $j>p$ and $j< pq -p$. Suppose also that for all $j$, $j-p$ or $j+p$ is not in $J$.
Then, the strong Irrepresentable Condition \ref{eq:irrep} of Theorem \ref{th1} holds.
\end{prop}
The proof of Proposition \ref{prop_IC} is given in Section \ref{sec:proofs}.
\subsubsection{Sufficient conditions for Assumptions \ref{th2(vii)}, \ref{th2(viii)}, \ref{th2(ix)} and \ref{th2(x)} of Theorem \ref{th2}}
The following proposition establishes that in the particular case where the $(E_{1,t})_t$, $(E_{2,t})_t$, ..., $(E_{n,t})_t$ are independent AR(1) processes, our strategy for estimating $\Sigma$
provides an estimator satisfying the assumptions of Theorem \ref{th2}.
\begin{prop}\label{prop:cond_th_AR}
Assume that $(E_{1,t})_t$, $(E_{2,t})_t$, ..., $(E_{n,t})_t$ in Model (\ref{eq:model:matriciel}) are independent AR(1) processes satisfying:
$$
E_{i,t}-\phi_1 E_{i,t-1}=Z_{i,t},\; \forall i\in\{1,\dots,n\},
$$
where the $Z_{i,t}$'s are zero-mean i.i.d. Gaussian random variables with variance $\sigma^2$ and $|\phi_1|<1$.
Let
$$
\widehat{\Sigma}=\frac{1}{1-\widehat{\phi}_1^2}\left(
\begin{array}{cccccc}
1 & \widehat{\phi}_1 & \widehat{\phi}_1^2 & \dots & \widehat{\phi}_1^{q-1}\\
\widehat{\phi}_1 & 1 & \widehat{\phi}_1 & \dots & \widehat{\phi}_1^{q-2}\\
\vdots & \ddots & \ddots & \ddots & \vdots \\
\vdots & \ddots & \ddots & \ddots & \vdots \\
\widehat{\phi}_1^{q-1} & \dots & \dots & \dots & 1 \\
\end{array}
\right),
$$
where
\begin{equation}\label{eq:phi_1_hat}
\widehat{\phi}_1=\frac{\sum_{i=1}^n\sum_{\ell=2}^q \widehat{E}_{i,\ell} \widehat{E}_{i,\ell-1}}{\sum_{i=1}^n\sum_{\ell=1}^{q-1} \widehat{E}_{i,\ell}^2},
\end{equation}
where $\widehat{E}=(\widehat{E}_{i,\ell})_{1\leq i\leq n,1\leq \ell\leq q}$ is defined in (\ref{eq: Ehat}).
Then, Assumptions \ref{th2(vii)}, \ref{th2(viii)}, \ref{th2(ix)} and \ref{th2(x)} of Theorem \ref{th2} are valid.
\end{prop}
The proof of Proposition \ref{prop:cond_th_AR} is given in Section \ref{sec:proofs}. It is based on the following lemma.
\begin{lemma}\label{lem:AR:estim}
Assume that $(E_{1,t})_t$, $(E_{2,t})_t$, ..., $(E_{n,t})_t$ in Model (\ref{eq:model:matriciel}) are independent AR(1) processes satisfying:
$$
E_{i,t}-\phi_1 E_{i,t-1}=Z_{i,t},\; \forall i\in\{1,\dots,n\},
$$
where the $Z_{i,t}$'s are zero-mean i.i.d. Gaussian random variables with variance $\sigma^2$ and $|\phi_1|<1$. Let
$$
\widehat{\phi}_1=\frac{\sum_{i=1}^n\sum_{\ell=2}^q \widehat{E}_{i,\ell} \widehat{E}_{i,\ell-1}}{\sum_{i=1}^n\sum_{\ell=1}^{q-1} \widehat{E}_{i,\ell}^2},
$$
where $\widehat{E}=(\widehat{E}_{i,\ell})_{1\leq i\leq n,1\leq \ell\leq q}$ is defined in (\ref{eq: Ehat}).
Then,
$$
\sqrt{nq_n}(\widehat{\phi}_1-\phi_1)=O_p(1),\textrm{ as } n\to\infty.
$$
\end{lemma}
Lemma \ref{lem:AR:estim} is proved in Section \ref{sec:proofs}. Its proof is based on Lemma \ref{lem:denom} in Section \ref{sec:lemmas}.
\input{numexp}
\section{Proofs}\label{sec:proofs}
\begin{proof}[Proof of Proposition \ref{prop1}]
For a fixed nonnegative $\lambda$, by \eqref{eq:lasso},
$$
\widehat{\mathcal{B}}=\widehat{\mathcal{B}}(\lambda)=\textrm{Argmin}_{\mathcal{B}}\left\{\|\mathcal{Y}-\mathcal{X}\mathcal{B}\|_2^2+\lambda\|\mathcal{B}\|_1\right\}.
$$
Denoting $\widehat{u}=\widehat{\mathcal{B}}-\mathcal{B}$, we get
\begin{align*}
\|\mathcal{Y}-\mathcal{X}\widehat{\mathcal{B}}\|_2^2+\lambda\|\widehat{\mathcal{B}}\|_1
&=\|\mathcal{X}\mathcal{B}+\mathcal{E}-\mathcal{X}\widehat{\mathcal{B}}\|_2^2+\lambda \|\widehat{u}+\mathcal{B}\|_1
=\|\mathcal{E}-\mathcal{X}\widehat{u}\|_2^2+\lambda \|\widehat{u}+\mathcal{B}\|_1\\
&=\|\mathcal{E}\|_2^2-2\widehat{u}'\mathcal{X}'\mathcal{E}+\widehat{u}'\mathcal{X}'\mathcal{X}\widehat{u}+\lambda \|\widehat{u}+\mathcal{B}\|_1.
\end{align*}
Thus,
$$
\widehat{u}=\textrm{Argmin}_u\; V(u),
$$
where
$$
V(u)=-2 (\sqrt{nq}u)' W+(\sqrt{nq}u)' C (\sqrt{nq}u)+\lambda \|u+\mathcal{B}\|_1.
$$
Since the first derivative of $V$ with respect to $u$ is equal to
$$
2\sqrt{nq}\big(C(\sqrt{nq}u)-W\big)+\lambda\,\ensuremath{\rm sign}(u+\mathcal{B}),$$
$\widehat{u}$ satisfies
\[
C_{J,J}(\sqrt{nq}\widehat{u}_J)-W_J=-\frac{\lambda}{2\sqrt{nq}}\ensuremath{\rm sign}(\widehat{u}_J+\mathcal{B}_J)
=-\frac{\lambda}{2\sqrt{nq}}\ensuremath{\rm sign}(\widehat{\mathcal{B}}_J), \textrm{\, if \,}\widehat{u}_J+\mathcal{B}_J=\widehat{\mathcal{B}}_J\neq 0
\]
and
\[
\big|C_{J^c,J}(\sqrt{nq}\widehat{u}_J)-W_{J^c}\big|\leq\frac{\lambda}{2\sqrt{nq}}.\]
Note that, if $|\widehat{u}_J| < |\mathcal{B}_J|$, then $\widehat{\mathcal{B}}_J\neq 0$ and $\textrm{sign}(\widehat{\mathcal{B}}_J)=\textrm{sign}(\mathcal{B}_J)$.
Let us now prove that when $A_n$ and $B_n$, defined in (\ref{eq:An}) and (\ref{eq:Bn}), are satisfied then there exists
$\widehat{u}$ satisfying:
\begin{align}
C_{J,J}(\sqrt{nq}\widehat{u}_J)-W_J&=-\frac{\lambda}{2\sqrt{nq}}\textrm{sign}(\mathcal{B}_J),\label{eq:support1}\\
|\widehat{u}_J| & < |\mathcal{B}_J|,\label{eq:support2}\\
\big|C_{J^c,J}(\sqrt{nq}\widehat{u}_J)-W_{J^c}\big|&\leq\frac{\lambda}{2\sqrt{nq}}\label{eq:support3}.
\end{align}
Note that $A_n$ implies:
\begin{equation}\label{eq:u1}
\sqrt{nq}\left(-|\mathcal{B}_J|+\frac{\lambda}{2nq}(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right)<(C_{J,J})^{-1}W_J<\sqrt{nq}\left(|\mathcal{B}_J|+\frac{\lambda}{2nq}(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right).
\end{equation}
By denoting
\begin{equation}\label{eq:u_J_hat}
\widehat{u}_J=\frac{1}{\sqrt{nq}}(C_{J,J})^{-1}W_J-\frac{\lambda}{2nq}(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J),
\end{equation}
we obtain from (\ref{eq:u1}) that (\ref{eq:support1}) and (\ref{eq:support2}) hold.
Note that $B_n$ implies:
\begin{multline*}
-\frac{\lambda}{2\sqrt{nq}}\left(\mathbf{1}-C_{J^c,J}(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right)\\
\leq C_{J^c,J}(C_{J,J})^{-1}W_J-W_{J^c}\leq\frac{\lambda}{2\sqrt{nq}}\left(\mathbf{1}+C_{J^c,J}(C_{J,J})^{-1}
\textrm{sign}(\mathcal{B}_J)\right).
\end{multline*}
Hence,
$$
\left|C_{J^c,J}\left((C_{J,J})^{-1}W_J-\frac{\lambda}{2\sqrt{nq}}(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right)-W_{J^c}\right|\leq\frac{\lambda}{2\sqrt{nq}},
$$
which is (\ref{eq:support3}) by (\ref{eq:u_J_hat}). This concludes the proof.
\end{proof}
\begin{proof}[Proof of Theorem \ref{th1}]
By Proposition \ref{prop1},
$$
\mathbb{P}\left(\textrm{sign}(\widehat{\mathcal{B}}(\lambda))=\textrm{sign}(\mathcal{B})\right)\geq\mathbb{P}(A_n\cap B_n)
=1-\mathbb{P}(A_n^c\cup B_n^c)\geq 1-\mathbb{P}(A_n^c)-\mathbb{P}(B_n^c),
$$
where $A_n$ and $B_n$ are defined in (\ref{eq:An}) and (\ref{eq:Bn}). It is thus enough to prove that $\mathbb{P}(A_n^c)$ and $\mathbb{P}(B_n^c)$ tend to zero as $n$ tends to infinity.
By definition of $A_n$,
\begin{align}\label{eq:P_Anc}
\mathbb{P}(A_n^c)&=\mathbb{P}\left(\left|(C_{J,J})^{-1}W_J\right|\geq\sqrt{nq}\left(|\mathcal{B}_J|-\frac{\lambda}{2nq}|(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)|\right)\right)\nonumber\\
&\leq\sup_{j\in J} \mathbb{P}\left(|\xi_j|\geq\sqrt{nq}\left(|\mathcal{B}_j|-\frac{\lambda}{2nq}|b_j|\right)\right),
\end{align}
where
$$
\xi=(\xi_j)_{j\in J}=(C_{J,J})^{-1}W_J=\frac{1}{\sqrt{nq}}(C_{J,J})^{-1}(\mathcal{X}_{\bullet,J})'\mathcal{E} =: H_A\;\mathcal{E},
$$
and
$$
b=(b_j)_{j\in J}=(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J).
$$
By definition of $B_n$ and \ref{eq:irrep},
\begin{align}\label{eq:P_Bnc}
\mathbb{P}(B_n^c)&=\mathbb{P}\left(\left|C_{J^c,J}(C_{J,J})^{-1}W_J-W_{J^c}\right|>\frac{\lambda}{2\sqrt{nq}}\left(\mathbf{1}-\left|C_{J^c,J}(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right|\right)\right)\nonumber\\
&\leq \mathbb{P}\left(\left|C_{J^c,J}(C_{J,J})^{-1}W_J-W_{J^c}\right|>\frac{\lambda}{2\sqrt{nq}}\eta\right)\nonumber\\
&\leq \sup_{j\in J^c} \mathbb{P}\left(|\zeta_j|>\frac{\lambda}{2\sqrt{nq}}\eta\right),
\end{align}
where
$$
\zeta=(\zeta_j)_{j\in J^c}=C_{J^c,J}(C_{J,J})^{-1}W_J-W_{J^c}=\frac{1}{\sqrt{nq}}\big(C_{J^c,J}(C_{J,J})^{-1}(\mathcal{X}_{\bullet,J})'-(\mathcal{X}_{\bullet,J^c})'\big)\mathcal{E}
=: H_B\;\mathcal{E}.
$$
Note that, for all $j$ in $J$,
$$
|b_j|\leq\sum_{j\in J} |b_j|\leq \sqrt{|J|}\left(\sum_{j\in J} b_j^2\right)^{1/2}=\sqrt{|J|}\|b\|_2.
$$
Moreover,
$$
\|b\|_2=\|(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\|_2\leq \|(C_{J,J})^{-1}\|_2\sqrt{|J|}:=\lambda_{\textrm{max}}((C_{J,J})^{-1})\sqrt{|J|},
$$
where $\lambda_{\textrm{max}}(A)$ denotes the largest eigenvalue of the matrix $A$. Observe that
\begin{equation}\label{eq:lambda_max_CJJ-1}
\lambda_{\textrm{max}}((C_{J,J})^{-1})=\frac{1}{\lambda_{\rm{min}}(C_{J,J})}=\frac{q}{\lambda_{\rm{min}}((\mathcal{X}'\mathcal{X})_{J,J})/n}\leq \frac{q}{M_2},
\end{equation}
by Assumption~\ref{th1(ii)} of Theorem~\ref{th1}. Thus, for all $j$ in $J$,
\begin{equation}\label{eq:bj}
|b_j|\leq \frac{q|J|}{M_2}.
\end{equation}
By Assumption~\ref{th1(iv)} of Theorem~\ref{th1}, we get thus that for all $j$ in $J$,
\begin{equation}\label{eq:minoration_terme_A_th1}
\sqrt{nq}\left(|\mathcal{B}_j|-\frac{\lambda}{2nq}\left|\left((C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right)_j\right|\right)=\sqrt{nq}\left(|\mathcal{B}_j|-\frac{\lambda}{2nq}|b_j|\right)
\geq \sqrt{nq}\left(M_3 q^{-c_2} -\frac{\lambda q |J|}{2nq M_2}\right).
\end{equation}
Thus,
\begin{equation}\label{eq:Anc_bis}
\mathbb{P}(A_n^c)\leq\sup_{j\in J} \mathbb{P}\left(|\xi_j|\geq\sqrt{nq}\left(M_3 q^{-c_2} -\frac{\lambda q |J|}{2nq M_2}\right)\right).
\end{equation}
Since $\mathcal{E}$ is a centered Gaussian random vector having a covariance matrix equal to identity,
$\xi= H_A\;\mathcal{E}$ is a centered Gaussian random vector with a covariance matrix equal to:
$$
H_A H_A'=\frac{1}{nq}(C_{J,J})^{-1}(\mathcal{X}_{\bullet,J})' \mathcal{X}_{\bullet,J}(C_{J,J})^{-1}=(C_{J,J})^{-1}.
$$
Hence, by (\ref{eq:lambda_max_CJJ-1}), we get that for all $j$ in $J$,
$$
\mathop{\rm Var}\nolimits(\xi_j)=\left((C_{J,J})^{-1}\right)_{jj}\leq\lambda_{\textrm{max}}(C_{J,J}^{-1})\leq \frac{q}{M_2}.
$$
Thus,
$$
\mathbb{P}\left(|\xi_j|\geq\sqrt{nq}\left(M_3 q^{-c_2} -\frac{\lambda q |J|}{2nq M_2}\right)\right)
\leq \mathbb{P}\left(|Z|\geq\frac{\sqrt{M_2}}{\sqrt{q}}\left(M_3 q^{-c_2}\sqrt{nq}-\frac{\lambda q |J|}{2\sqrt{nq} M_2}\right)\right),
$$
where $Z$ is a standard Gaussian random variable.
By Chernoff inequality, we thus obtain that for all $j$ in $J$,
$$
\mathbb{P}\left(|\xi_j|\geq\sqrt{nq}\left(M_3 q^{-c_2} -\frac{\lambda q |J|}{2nq M_2}\right)\right)
\leq 2\exp\left(-\frac{M_2}{2q}\left\{M_3 q^{-c_2}\sqrt{nq}-\frac{\lambda q |J|}{2\sqrt{nq} M_2}\right\}^2\right).
$$
By Assumption~\ref{th1(iii)} of Theorem~\ref{th1}, we get that under the last condition of \ref{assum:lambda_nq},
\begin{equation}\label{eq:lambda_J_sqrt_nq}
\frac{\lambda q |J|}{\sqrt{nq}}=o\left(q^{-c_2}\sqrt{nq}\right), \textrm{ as } n\to\infty.
\end{equation}
Thus,
\begin{equation}\label{eq:Anc:limit}
\mathbb{P}(A_n^c)\to 0,\; \textrm{ as } n\to\infty.
\end{equation}
Let us now bound $\mathbb{P}(B_n^c)$. Observe that
$\zeta= H_B\;\mathcal{E}$ is a centered Gaussian random vector with a covariance matrix equal to:
\begin{align*}
H_B H_B' &=\frac{1}{nq}(C_{J^c,J}(C_{J,J})^{-1}(\mathcal{X}_{\bullet,J})'-\mathcal{X}_{\bullet,J^c}')(\mathcal{X}_{\bullet,J}(C_{J,J})^{-1}C_{J,J^c}-\mathcal{X}_{\bullet,J^c})\\
&=C_{J^c,J^c}-C_{J^c,J}(C_{J,J})^{-1}C_{J,J^c}=\frac{1}{nq} (\mathcal{X}_{\bullet,J^c})'\left(\textrm{Id}_{\mathbb{R}^{nq}}-\mathcal{X}_{\bullet,J}((\mathcal{X}_{\bullet,J})'\mathcal{X}_{\bullet,J})^{-1}(\mathcal{X}_{\bullet,J})'\right)\mathcal{X}_{\bullet,J^c}\\
&=\frac{1}{nq} (\mathcal{X}_{\bullet,J^c})' \, \left(\textrm{Id}_{\mathbb{R}^{nq}}-\Pi_{\textrm{Im}(\mathcal{X}_{\bullet,J})}\right) \, \mathcal{X}_{\bullet,J^c},
\end{align*}
where $\Pi_{\textrm{Im}(\mathcal{X}_{\bullet,J})}$ denotes the orthogonal projection onto the column space of $\mathcal{X}_{\bullet,J}$.
Note that, for all $j$ in $J^c$,
\begin{align*}
\mathop{\rm Var}\nolimits(\zeta_j)&=\frac{1}{nq}\left((\mathcal{X}_{\bullet,J^c})'\; \left(\textrm{Id}_{\mathbb{R}^{nq}}-\Pi_{\textrm{Im}(\mathcal{X}_{\bullet,J})}\right)\; \mathcal{X}_{\bullet,J^c}\right)_{jj}\\
&=\frac{1}{nq}\left((\mathcal{X}_{\bullet,J^c})' \mathcal{X}_{\bullet,J^c}\right)_{jj}-\frac{1}{nq}\left((\mathcal{X}_{\bullet,J^c})'\;\Pi_{\textrm{Im}(\mathcal{X}_{\bullet,J})}\mathcal{X}_{\bullet,J^c}\right)_{jj}\\
&\leq\frac{1}{nq}\left((\mathcal{X}_{\bullet,J^c})' \mathcal{X}_{\bullet,J^c}\right)_{jj}\leq \frac{M_1}{q},
\end{align*}
where the inequalities come from Lemma~\ref{lem:proj} and Assumption~\ref{th1(i)} of Theorem~\ref{th1}.
Thus, for all $j$ in $J^c$,
$$
\mathbb{P}\left(|\zeta_j|>\frac{\lambda}{2\sqrt{nq}}\eta\right)\leq \mathbb{P}\left(|Z|>\frac{\lambda\sqrt{q}}{2\sqrt{M_1}\sqrt{nq}}\eta\right),
$$
where $Z$ is a standard Gaussian random variable. By Chernoff inequality, for all $j$ in $J^c$,
$$
\mathbb{P}\left(|\zeta_j|>\frac{\lambda}{2\sqrt{nq}}\eta\right)\leq 2\exp\left\{-\frac12\left(\frac{\lambda}{2\sqrt{M_1}\sqrt{n}}\eta\right)^2\right\}.
$$
Hence, under the following assumption
$$
\frac{\lambda}{\sqrt{n}}\to\infty,
$$
which is the second condition of \ref{assum:lambda_nq},
\begin{equation}\label{eq:Bnc:limit}
\mathbb{P}(B_n^c)\to 0,\; \textrm{ as } n\to\infty.
\end{equation}
\end{proof}
\begin{proof}[Proof of Proposition \ref{propcond}]
Let us first prove that \ref{cond1} and \ref{cond3} imply \ref{th1(i)}.
For $j\in\{1,\ldots,pq\}$, by considering the Euclidian division of $j-1$ by $p$ given by $(j-1)=pk_j+r_j$, we observe that
\begin{align*
(\mathcal{X}_{\bullet,j})'\mathcal{X}_{\bullet,j}&=(((\Sigma^{-1/2})' \otimes X)_{\bullet,j})'((\Sigma^{-1/2})' \otimes X)_{\bullet,j}\\
&=((\Sigma^{-1/2}) \otimes X')_{j,\bullet})((\Sigma^{-1/2})' \otimes X)_{\bullet,j}\\
&=((\Sigma^{-1/2})_{k_j+1,\bullet} \otimes (X_{\bullet,r_j+1})')(((\Sigma^{-1/2})_{\bullet,k_j+1})' \otimes X_{\bullet,r_j+1})\\
&=(\Sigma^{-1/2})_{k_j+1,\bullet}((\Sigma^{-1/2})_{\bullet,k_j+1})' \otimes (X_{\bullet,r_j+1})'X_{\bullet,r_j+1}\\
&=(\Sigma^{-1})_{k_j+1,k_j+1} \otimes(X_{\bullet,r_j+1})'X_{\bullet,r_j+1}\\
&=(\Sigma^{-1})_{k_j+1,k_j+1}(X_{\bullet,r_j+1})'X_{\bullet,r_j+1}.
\end{align*}
Hence, using \ref{cond1}, we get that for all $j$ in $\{1,\ldots,pq\}$,
\begin{align*}
\frac{1}{n}(\mathcal{X}_{\bullet,j})'\mathcal{X}_{\bullet,j}&\leq M'_1(\Sigma^{-1})_{k_j+1,k_j+1}
\leq M'_1\sup_{k \in \{0,\ldots,q-1\}}((\Sigma^{-1})_{k+1,k+1})\\
& \leq M'_1\lambda_{\rm{max}}(\Sigma^{-1})\leq M'_1 m_1,
\end{align*}
where the last inequality comes from~\ref{cond3}, which gives (A1).
Let us now prove that \ref{cond2} and \ref{cond4} imply \ref{th1(ii)}.
Note that
\begin{align*
(\mathcal{X}'\mathcal{X})_{J,J}&=(((\Sigma^{-1/2})' \otimes X)'((\Sigma^{-1/2})' \otimes X))_{J,J}\\
&=(\Sigma^{-1/2}(\Sigma^{-1/2})'\otimes X'X)_{J,J}\\
&=(\Sigma^{-1} \otimes X'X)_{J,J}.
\end{align*}
Then, by Theorem 4.3.15 of \cite{Horn:Johnson:1985},
\begin{align}
\lambda_{\min}((\mathcal{X}'\mathcal{X})_{J,J})&=\lambda_{\min}((\Sigma^{-1} \otimes X'X)_{J,J}) \nonumber\\
& \geq \lambda_{\min}(\Sigma^{-1} \otimes X'X) \nonumber\\
&=\lambda_{\min}(X'X)\lambda_{\min}(\Sigma^{-1}).\label{lmbgreat}
\end{align}
Finally, by using Conditions~\ref{cond2}~and~\ref{cond4}, we obtain
\[
\frac{1}{n}\lambda_{\min}(\mathcal{X}'\mathcal{X})_{J,J}\geq \frac{1}{n}\lambda_{\min}(X'X)\lambda_{\min}(\Sigma^{-1})\geq M'_2m_2,
\]
which gives \ref{th1(ii)}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{th2}]
By Proposition \ref{prop2},
$$
\mathbb{P}\left(\textrm{sign}(\widetilde{\mathcal{B}}(\lambda))=\textrm{sign}(\mathcal{B})\right)\geq\mathbb{P}(\widetilde{A}_n\cap\widetilde{B}_n)
=1-\mathbb{P}(\widetilde{A}_n^c\cup \widetilde{B}_n^c)\geq 1-\mathbb{P}(\widetilde{A}_n^c)-\mathbb{P}(\widetilde{B}_n^c),
$$
where $\widetilde{A}_n$ and $\widetilde{B}_n$ are defined in (\ref{eq:An_tilde}) and (\ref{eq:Bn_tilde}).
By definition of $\widetilde{A}_n$, we get
$$
\mathbb{P}(\widetilde{A}_n^c)=\mathbb{P}\left(\left\{\left|(\widetilde{C}_{J,J})^{-1}\widetilde{W}_J\right|\geq\sqrt{nq}\left(|\mathcal{B}_J|
-\frac{\lambda}{2nq}|(\widetilde{C}_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)|\right)\right\}\right).
$$
Observing that
\begin{align*}
(\widetilde{C}_{J,J})^{-1}\widetilde{W}_J&=(C_{J,J})^{-1}W_J+(C_{J,J})^{-1}\left(\widetilde{W}_J-W_J\right)\\
&+\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)W_J\\
&+\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)\left(\widetilde{W}_J-W_J\right),
\end{align*}
$$
(\widetilde{C}_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)=(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)+\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)
\textrm{sign}(\mathcal{B}_J),
$$
and using the triangle inequality, we obtain that
\begin{align}\label{eq:An_tilde_bound}
&\mathbb{P}(\widetilde{A}_n^c)\leq\mathbb{P}\left(\left\{\left|(C_{J,J})^{-1}W_J\right|\geq\frac{\sqrt{nq}}{5}\left(|\mathcal{B}_J|
-\frac{\lambda}{2nq}\left|(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right|\right)\right\}\right)\nonumber\\
&+\mathbb{P}\left(\left\{\left|(C_{J,J})^{-1}\left(\widetilde{W}_J-W_J\right)\right|
\geq\frac{\sqrt{nq}}{5}\left(|\mathcal{B}_J|-\frac{\lambda}{2nq}\left|(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right|\right)\right\}\right)\nonumber\\
&+\mathbb{P}\left(\left\{\left|\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)W_J\right|
\geq\frac{\sqrt{nq}}{5}\left(|\mathcal{B}_J|-\frac{\lambda}{2nq}\left|(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right|\right)\right\}\right)\nonumber\\
&+\mathbb{P}\left(\left\{\left|\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)\left(\widetilde{W}_J-W_J\right)\right|
\geq\frac{\sqrt{nq}}{5}\left(|\mathcal{B}_J|-\frac{\lambda}{2nq}\left|(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right|\right)\right\}\right)\nonumber\\
&+\mathbb{P}\left(\left\{\frac{\lambda}{2\sqrt{nq}}\left|\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)\textrm{sign}(\mathcal{B}_J)\right|\geq
\frac{\sqrt{nq}}{5}\left(|\mathcal{B}_J|-\frac{\lambda}{2nq}\left|(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right|\right)\right\}\right).\nonumber\\
\end{align}
The first term in the r.h.s of (\ref{eq:An_tilde_bound}) tends to 0 by the definition of ${A}_n^c$ and (\ref{eq:Anc:limit}).
By (\ref{eq:minoration_terme_A_th1}), the last term of (\ref{eq:An_tilde_bound}) satisfies, for all $j\in J$:
\begin{align*}
&\mathbb{P}\left(\left|\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)\textrm{sign}(\mathcal{B}_J)\right|\geq
\frac{2 nq}{5\lambda}\left(|\mathcal{B}_J|-\frac{\lambda}{2nq}\left|(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right|\right)\right)\\
&\leq \mathbb{P}\left(\left|\left(\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)\textrm{sign}(\mathcal{B}_J)\right)_j\right|\geq
\frac{2 nq}{5\lambda}\left(M_3 q^{-c_2} -\frac{\lambda q|J|}{2nq M_2}\right)\right).
\end{align*}
Let $U=(\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}$ and $s=\textrm{sign}(\mathcal{B}_J)$ then for all $j$ in $J$:
\begin{equation}\label{eq:U_s_j}
|(U s)_j|=\left|\sum_{k\in J} U_{jk} s_k\right|\leq \sqrt{|J|}\|U\|_{2}.
\end{equation}
We focus on
\begin{align*}
&\|(\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\|_2=\|(\widetilde{C}_{J,J})^{-1}(C_{J,J}-\widetilde{C}_{J,J})(C_{J,J})^{-1}\|_{2}
\leq \|(\widetilde{C}_{J,J})^{-1}\|_{2}\;\|C_{J,J}-\widetilde{C}_{J,J}\|_{2}\;\|(C_{J,J})^{-1}\|_{2}\\
&\leq \frac{\rho(C_{J,J}-\widetilde{C}_{J,J})}{\lambda_{\rm{min}}(\widetilde{C}_{J,J})\lambda_{\rm{min}}({C}_{JJ})}
\leq \frac{\rho(C_{J,J}-\widetilde{C}_{J,J})}{\lambda_{\rm{min}}(\widetilde{C}_{J,J}) (M_2/q)},
\end{align*}
where the last inequality comes from Assumption \ref{th1(ii)} of Theorem \ref{th1}, which gives that
\begin{equation}\label{eq:CJJ-1:bound}
\|(C_{J,J})^{-1}\|_{2}\leq \frac{q}{M_2}.
\end{equation}
Using Theorem 4.3.15 of \cite{Horn:Johnson:1985}, we get
$$
\|(\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\|_2\leq \frac{q\rho(C-\widetilde{C})}{\lambda_{\textrm{min}}(\widetilde{C}) M_2}.
$$
By definition of $C$ and $\widetilde{C}$ given in (\ref{eq:C:J}) and (\ref{eq:C_tilde}), respectively, we get
\begin{equation}\label{eq:C_C_tilde}
C=\frac{\Sigma^{-1}\otimes (X'X)}{nq} \textrm{ and } \widetilde{C}=\frac{\widehat{\Sigma}^{-1}\otimes (X'X)}{nq}.
\end{equation}
By using that the eigenvalues of the Kronecker product of two matrices is equal to the product of the eigenvalues of the two matrices, we obtain
\begin{align}\label{eq:norme2_diff_CJJ}
\|(\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\|_{2}
&\leq\frac{\rho(\Sigma^{-1}-\widehat{\Sigma}^{-1})\lambda_{\textrm{max}}((X'X)/n)q}{\lambda_{\textrm{min}}(\widehat{\Sigma}^{-1})
\lambda_{\textrm{min}}((X'X)/n) M_2}\leq \frac{\rho(\Sigma^{-1}-\widehat{\Sigma}^{-1})\lambda_{\textrm{max}}(\widehat{\Sigma}) \lambda_{\textrm{max}}((X'X)/n)q}
{\lambda_{\textrm{min}}((X'X)/n) M_2}\nonumber\\
&\leq\frac{\rho(\Sigma^{-1}-\widehat{\Sigma}^{-1})\left(\rho(\widehat{\Sigma}-\Sigma)+\lambda_{\textrm{max}}(\Sigma)\right) \lambda_{\textrm{max}}((X'X)/n)q}
{\lambda_{\textrm{min}}((X'X)/n) M_2},
\end{align}
where the last inequality follows from Theorem 4.3.1 of \cite{Horn:Johnson:1985}. Thus, by Assumptions \ref{th2(v)}, \ref{th2(vi)}, \ref{th2(viii)}, \ref{th2(ix)} and \ref{th2(x)}, we get that
\begin{equation}\label{eq:norm2_Ctilde_JJ-1_-C_JJ-1}
\|(\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\|_{2}=O_P(q(nq)^{-1/2}), \textrm{ as } n\to\infty.
\end{equation}
Hence, by (\ref{eq:U_s_j}), we get for all $j$ in $J$ that
\begin{align*}
&\mathbb{P}\left(\left|\left(\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)\textrm{sign}(\mathcal{B}_J)\right)_j\right|\geq
\frac{2 nq}{5\lambda}\left(M_3 q^{-c_2} -\frac{\lambda q|J|}{2nq M_2}\right)\right)\\
&\leq\mathbb{P}\left(\sqrt{|J|}\;\|(\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\|_{2}\geq
\frac{2 \sqrt{nq}}{5\lambda}\left(M_3 q^{-c_2}\sqrt{nq} -\frac{\lambda q|J|}{2\sqrt{nq} M_2}\right)\right).
\end{align*}
By (\ref{eq:lambda_J_sqrt_nq}), (\ref{eq:norm2_Ctilde_JJ-1_-C_JJ-1}) and \ref{th1(iii)}, it is enough to prove that
$$
\mathbb{P}\left(q^{c_1/2}q(nq)^{-1/2}\geq \frac{nq}{\lambda}q^{-c_2}\right)\to 0, \textrm{ as } n\to\infty.
$$
By the last condition of \ref{assum:lambda_nq},
$$
\frac{\frac{nq}{\lambda}q^{-c_2}}{q^{1+c_1}}\to\infty, \textrm{ as } n\to\infty,
$$
and the result follows since $n$ tends to infinity.
Hence, the last term of (\ref{eq:An_tilde_bound}) tends to zero as $n$ tends to infinity.
Let us now study the second term in the r.h.s of (\ref{eq:An_tilde_bound}).
\begin{align}\label{eq:diff_Wn}
&\widetilde{W}_J-W_J=\frac{1}{\sqrt{nq}}\left(\left(\widetilde{\mathcal{X}}'\widetilde{\mathcal{E}}\right)_J-\left(\mathcal{X}'\mathcal{E}\right)_J\right)
=\frac{1}{\sqrt{nq}}\left(\widetilde{\mathcal{X}}'\widetilde{\mathcal{E}}-\mathcal{X}'\mathcal{E}\right)_J\nonumber\\
&=\frac{1}{\sqrt{nq}}\left[\left(\widehat{\Sigma}^{-1/2}\otimes X'\right)\left((\widehat{\Sigma}^{-1/2})'\otimes \textrm{Id}_{\mathbb{R}^n}\right)\textrm{Vec}(E)
-\left(\Sigma^{-1/2}\otimes X'\right)\left((\Sigma^{-1/2})'\otimes \textrm{Id}_{\mathbb{R}^n}\right)\textrm{Vec}(E)\right]_J\nonumber\\
&=\frac{1}{\sqrt{nq}}\left[\left\{\left(\widehat{\Sigma}^{-1}-\Sigma^{-1}\right)\otimes X'\right\}\textrm{Vec}(E)\right]_J
\stackrel{d}{=}AZ,
\end{align}
where $Z$ is a centered Gaussian random vector having a covariance matrix equal to identity and
\begin{equation}\label{eq:def_A}
A=\frac{1}{\sqrt{nq}}\left[\left\{\left(\widehat{\Sigma}^{-1}-\Sigma^{-1}\right)\otimes X'\right\}\left\{(\Sigma^{1/2})'\otimes\textrm{Id}_{\mathbb{R}^n}\right\}\right]_{J,\bullet}.
\end{equation}
By Cauchy-Schwarz inequality, we get for all $K\times nq$ matrix $B$, and all $nq\times 1$ vector $U$ that for all $k$ in $\{1,\dots,K\}$,
\begin{equation}\label{eq:Cauchy-Schwarz}
\left|(BU)_k\right|=\left|\sum_{\ell=1}^{nq} B_{k,\ell} U_\ell\right|\leq \|B\|_2\;\|U\|_2.
\end{equation}
Thus, for all $j$ in $J$, for all $\gamma$ in $\mathbb{R}$ and all $|J|\times |J|$ matrix $D$,
\begin{equation}\label{eq:B_WJ-WJtilde}
\mathbb{P}\left(\left|\left(D\left(\widetilde{W}_J-W_J\right)\right)_j\right|\geq \gamma\right)
=\mathbb{P}\left(\left|\left(DAZ\right)_j\right|\geq \gamma\right)
\leq\mathbb{P}\left(\|D\|_2\;\|A\|_2 \|Z\|_2\geq \gamma\right),
\end{equation}
where $A$ is defined in (\ref{eq:def_A}) and $Z$ is a centered Gaussian random vector having a covariance matrix equal to identity. Hence, for all $j$ in $J$,
\begin{align*}
&\mathbb{P}\left(\left|\left((C_{J,J})^{-1}\left(\widetilde{W}_J-W_J\right)\right)_j\right|
\geq\frac{\sqrt{nq}}{5}\left(|\mathcal{B}_j|-\frac{\lambda}{2nq}\left|\left((C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right)_j\right|\right)\right)\\
&\leq \mathbb{P}\left(\|(C_{J,J})^{-1}\|_2\|A\|_2\|Z\|_2\geq \frac{\sqrt{nq}}{5}\left(M_3 q^{-c_2} -\frac{\lambda q|J|}{2nq M_2}\right)\right).
\end{align*}
Let us bound $\|A\|_{2}$.
Observe that
\begin{align}\label{eq:calculnorme2A}
& \left\|\left[\left\{\left(\widehat{\Sigma}^{-1}-\Sigma^{-1}\right)\otimes X'\right\}\left\{(\Sigma^{1/2})'\otimes\textrm{Id}\right\}\right]_{J,\bullet}\right\|_2\nonumber\\
&=\rho\left(\left[\left(\widehat{\Sigma}^{-1}-\Sigma^{-1}\right)\Sigma\left(\widehat{\Sigma}^{-1}-\Sigma^{-1}\right)\otimes (X'X)\right]_{J,J}\right)^{1/2}\nonumber\\
&\leq \rho\left(\left(\widehat{\Sigma}^{-1}-\Sigma^{-1}\right)\Sigma\left(\widehat{\Sigma}^{-1}-\Sigma^{-1}\right)\right)^{1/2}\lambda_{\rm{max}}\left(X'X\right)^{1/2}\nonumber\\
&\leq \rho\left(\widehat{\Sigma}^{-1}-\Sigma^{-1}\right)\lambda_{\rm{max}}(\Sigma)^{1/2}\lambda_{\rm{max}}\left(X'X\right)^{1/2},\nonumber\\
\end{align}
where the first inequality comes from Theorem 4.3.15 of \cite{Horn:Johnson:1985}. Hence, by \ref{th2(v)}, \ref{th2(viii)} and \ref{th2(ix)}
\begin{equation}\label{eq:norme2_A}
\|A\|_{2}=\frac{1}{\sqrt{nq}}\left\|\left[\left\{\left(\widehat{\Sigma}^{-1}-\Sigma^{-1}\right)\otimes X'\right\}\left\{(\Sigma^{1/2})'\otimes\textrm{Id}\right\}\right]_{J,\bullet}\right\|_2
=O_P(q^{-1/2}(nq)^{-1/2}).
\end{equation}
By (\ref{eq:lambda_J_sqrt_nq}), (\ref{eq:CJJ-1:bound}) and (\ref{eq:norme2_A}), it is enough to prove that
$$
\mathbb{P}\left(\sum_{k=1}^{nq} Z_k^2\geq nq\; n\; q^{-2c_2}\right)\to 0,\textrm { as } n\to\infty.
$$
The result follows from the Markov inequality and the first condition of \ref{assum:lambda_nq}.
Let us now study the third term in the r.h.s of (\ref{eq:An_tilde_bound}). Observe that
\begin{align}\label{eq:WJ}
&W_J =\frac{1}{\sqrt{nq}}\left[\left(\Sigma^{-1/2}\otimes X'\right)
\left((\Sigma^{-1/2})'\otimes\textrm{Id}_{\mathbb{R}^n}\right)\textrm{Vec}(E)\right]_J\nonumber\\
&\stackrel{d}{=}\frac{1}{\sqrt{nq}}\left[\left(\Sigma^{-1}\otimes X'\right)\left((\Sigma^{1/2})'\otimes\textrm{Id}_{\mathbb{R}^n}\right)\right]_{J,\bullet}Z=:A_1 Z,
\end{align}
where $Z$ is a centered Gaussian random vector having a covariance matrix equal to identity and
\begin{equation}\label{eq:def_A1}
A_1=\frac{1}{\sqrt{nq}}\left[\left(\Sigma^{-1}\otimes X'\right)\left((\Sigma^{1/2})'\otimes\textrm{Id}_{\mathbb{R}^n}\right)\right]_{J,\bullet}.
\end{equation}
Using (\ref{eq:Cauchy-Schwarz}), we get for all $j$ in $J$, for all $\gamma$ in $\mathbb{R}$ and all $|J|\times |J|$ matrix $D$,
\begin{equation}\label{eq:D_WJ}
\mathbb{P}\left(\left|\left(D\;W_J\right)_j\right|\geq \gamma\right)
=\mathbb{P}\left(\left|\left(DA_1Z\right)_j\right|\geq \gamma\right)
\leq\mathbb{P}\left(\|D\|_2\;\|A_1\|_2\; \|Z\|_2\geq \gamma\right),
\end{equation}
where $A_1$ is defined in (\ref{eq:def_A1}) and $Z$ is a centered Gaussian random vector having a covariance matrix equal to identity. Hence, for all $j$ in $J$,
\begin{align*}
&\mathbb{P}\left(\left|\left(\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)W_J\right)_j\right|
\geq\frac{\sqrt{nq}}{5}\left(|\mathcal{B}_j|-\frac{\lambda}{2nq}\left|\left((C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right)_j\right|\right)\right)\\
&\leq \mathbb{P}\left(\left\|(\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right\|_2\;\|A_1\|_2\; \|Z\|_2\geq\frac{\sqrt{nq}}{5}\left(M_3 q^{-c_2} -\frac{\lambda q|J|}{2nq M_2}\right)\right).
\end{align*}
Let us now bound $\|A_1\|_2$. Note that
\begin{multline*}
\left\|\left[\left(\Sigma^{-1}\otimes X'\right)\left((\Sigma^{1/2})'\otimes\textrm{Id}_{\mathbb{R}^n}\right)\right]_{J,\bullet}\right\|_2
=\left\|\left[\left(\Sigma^{-1/2}\otimes X'\right)\right]_{J,\bullet}\right\|_2\\
=\rho\left(\left[\Sigma^{-1}\otimes (X'X)\right]_{J,J}\right)^{1/2}\leq \rho\left(\left[\Sigma^{-1}\otimes (X'X)\right]\right)^{1/2}
\leq\lambda_{\rm{max}}(\Sigma^{-1})^{1/2}\lambda_{\rm{max}}(X'X)^{1/2},
\end{multline*}
where the first inequality comes from Theorem 4.3.15 of \cite{Horn:Johnson:1985}. Hence, by \ref{th2(v)} and \ref{th2(vii)},
\begin{equation}\label{eq:norme2_A1}
\|A_1\|_2\leq \frac{1}{nq}\lambda_{\rm{max}}(\Sigma^{-1})^{1/2}\lambda_{\rm{max}}(X'X)^{1/2}=O_P(q^{-1/2}).
\end{equation}
By (\ref{eq:lambda_J_sqrt_nq}), (\ref{eq:norm2_Ctilde_JJ-1_-C_JJ-1}) and (\ref{eq:norme2_A1}) it is thus enough to prove that
$$
\mathbb{P}\left(\sum_{k=1}^{nq}Z_k^2\geq nq\; n\; q^{-2c_2}\right)\to 0,\textrm{ as } n\to\infty.
$$
The result follows from the Markov inequality and the first condition of \ref{assum:lambda_nq}.
Let us now study the fourth term in the r.h.s of (\ref{eq:An_tilde_bound}).
By (\ref{eq:B_WJ-WJtilde}), for all $j$ in $J$,
\begin{align*}
&\mathbb{P}\left(\left|\left(\left((\widetilde{C}_{J,J})^{-1}-C_{J,J})^{-1}\right)\left(\widetilde{W}_J-W_J\right)\right)_j\right|
\geq\frac{\sqrt{nq}}{5}\left(|\mathcal{B}_j|-\frac{\lambda}{2nq}\left|\left((C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right)_j\right|\right)\right)\\
&\leq \mathbb{P}\left(\left\|(\widetilde{C}_{J,J})^{-1}-C_{J,J})^{-1}\right\|_2\; \|A\|_2\; \|Z\|_2
\geq \frac{\sqrt{nq}}{5}\left(M_3 q^{-c_2} -\frac{\lambda q|J|}{2nq M_2}\right)\right),
\end{align*}
where $A$ is defined in (\ref{eq:def_A}).
By (\ref{eq:lambda_J_sqrt_nq}), (\ref{eq:norm2_Ctilde_JJ-1_-C_JJ-1}) and (\ref{eq:norme2_A}), it is thus enough to prove that
$$
\mathbb{P}\left(\sum_{k=1}^{nq}Z_k^2\geq (nq)\; n^2\; q^{1-2c_2}\right)\to 0,\textrm{ as } n\to\infty.
$$
The result follows from the Markov inequality and the fact that $c_2<1/2$.
Let us now study $\mathbb{P}(\widetilde{B}_n)$. By definition of $\widetilde{B}_n$, we get that
$$
\mathbb{P}(\widetilde{B}_n^c)=\mathbb{P}\left(\left\{\left|\widetilde{C}_{J^c,J}(\widetilde{C}_{J,J})^{-1}\widetilde{W}_J-\widetilde{W}_{J^c}\right|\geq\frac{\lambda}{2\sqrt{nq}}\left(1-|\widetilde{C}_{J^c,J}(\widetilde{C}_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)|\right)\right\}\right).
$$
Observe that
\begin{align*}
\widetilde{C}_{J^c,J}(\widetilde{C}_{J,J})^{-1}\widetilde{W}_J-\widetilde{W}_{J^c}&=C_{J^c,J}(C_{J,J})^{-1}W_J-W_{J^c}\\
&+C_{J^c,J}(C_{J,J})^{-1}\left(\widetilde{W}_J-W_J\right)\\
&+C_{J^c,J}\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)W_J\\
&+C_{J^c,J}\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)\left(\widetilde{W}_J-W_J\right)\\
&+\left(\widetilde{C}_{J^c,J}- C_{J^c,J}\right)(C_{J,J})^{-1}W_J\\
&+\left(\widetilde{C}_{J^c,J}- C_{J^c,J}\right)(C_{J,J})^{-1}\left(\widetilde{W}_J-W_J\right)\\
&+\left(\widetilde{C}_{J^c,J}- C_{J^c,J}\right)\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)W_J\\
&+\left(\widetilde{C}_{J^c,J}- C_{J^c,J}\right)\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)\left(\widetilde{W}_J-W_J\right)\\
&+W_{J^c}-\widetilde{W}_{J^c}.
\end{align*}
Moreover,
\begin{align*}
\widetilde{C}_{J^c,J}(\widetilde{C}_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)&=C_{J^c,J}(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\\
&+C_{J^c,J}\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)\textrm{sign}(\mathcal{B}_J)\\
&+\left(\widetilde{C}_{J^c,J}-C_{J^c,J}\right)(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\\
&+\left(\widetilde{C}_{J^c,J}-C_{J^c,J}\right)\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)\textrm{sign}(\mathcal{B}_J).
\end{align*}
By \ref{eq:irrep} and the triangle inequality, we obtain that
\begin{align}\label{eq:Bn_tilde_bound}
&\mathbb{P}(\widetilde{B}_n^c)\leq\mathbb{P}\left(\left|C_{J^c,J}(C_{J,J})^{-1}W_J-W_{J^c}\right|\geq\frac{\lambda}{24\sqrt{nq}}\eta\right)\nonumber\\
&+\mathbb{P}\left(\left|C_{J^c,J}(C_{J,J})^{-1}\left(\widetilde{W}_J-W_J\right)\right|\geq\frac{\lambda}{24\sqrt{nq}}\eta\right)\nonumber\\
&+\mathbb{P}\left(\left\{\left|C_{J^c,J}\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)W_J\right|\geq\frac{\lambda}{24\sqrt{nq}}\eta\right\}\right)\nonumber\\
&+\mathbb{P}\left(\left\{\left|C_{J^c,J}\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)\left(\widetilde{W}_J-W_J\right)\right|\geq\frac{\lambda}{24\sqrt{nq}}\eta\right\}\right)\nonumber\\
&+\mathbb{P}\left(\left\{\left|\left(\widetilde{C}_{J^c, J}- C_{J^c,J}\right)(C_{J,J})^{-1}W_J\right|\geq\frac{\lambda}{24\sqrt{nq}}\eta\right\}\right)\nonumber\\
&+\mathbb{P}\left(\left\{\left|\left(\widetilde{C}_{J^c ,J}- C_{J^c,J}\right)(C_{J,J})^{-1}\left(\widetilde{W}_J-W_J\right)\right|\geq\frac{\lambda}{24\sqrt{nq}}\eta\right\}\right)\nonumber\\
&+\mathbb{P}\left(\left\{\left|\left(\widetilde{C}_{J^c, J}- C_{J^c,J}\right)\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)W_J\right|\geq\frac{\lambda}{24\sqrt{nq}}\eta\right\}\right)\nonumber\\
&+\mathbb{P}\left(\left\{\left|\left(\widetilde{C}_{J^c, J}- C_{J^c,J}\right)\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)\left(\widetilde{W}_J-W_J\right)\right|\geq\frac{\lambda}{24\sqrt{nq}}\eta\right\}\right)\nonumber\\
&+\mathbb{P}\left(\left\{\left|W_{J^c}-\widetilde{W}_{J^c}\right|\geq\frac{\lambda}{24\sqrt{nq}}\eta\right\}\right)\nonumber\\
&+\mathbb{P}\left(\left\{\left|C_{J^c,J}\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)\textrm{sign}(\mathcal{B}_J)\right|\geq\frac{\eta}{12}\right\}\right)\nonumber\\
&+\mathbb{P}\left(\left\{\left|\left(\widetilde{C}_{J^c,J}-C_{J^c,J}\right)(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right|\geq\frac{\eta}{12}\right\}\right)\nonumber\\
&+\mathbb{P}\left(\left\{\left|\left(\widetilde{C}_{J^c,J}-C_{J^c,J}\right)\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)\textrm{sign}(\mathcal{B}_J)\right|\geq\frac{\eta}{12}\right\}\right).\nonumber\\
\end{align}
The first term in the r.h.s of (\ref{eq:Bn_tilde_bound}) tends to 0 by (\ref{eq:Bnc:limit}).
Let us now study the second term of (\ref{eq:Bn_tilde_bound}).
By (\ref{eq:B_WJ-WJtilde}), we get that for all $j$ in $J^c$,
\begin{align*}
&\mathbb{P}\left(\left|\left(C_{J^c,J}(C_{J,J})^{-1}\left(\widetilde{W}_J-W_J\right)\right)_j\right|
\geq\frac{\lambda}{24\sqrt{nq}}\eta\right)\\
&\leq \mathbb{P}\left(\left\|C_{J^c,J}\right\|_2\|(C_{J,J})^{-1}\|_2\|A\|_2\|Z\|_2\geq \frac{\lambda}{24\sqrt{nq}}\eta\right).
\end{align*}
Observe that
\begin{align}\label{eq:n2_Cj_jc}
\left\|C_{J^c,J}\right\|_2 &= \rho\left(\frac{(\mathcal{X}_{\bullet,J^c})'\mathcal{X}_{\bullet,J}}{nq}\frac{(\mathcal{X}_{\bullet,J})'\mathcal{X}_{\bullet,J^c}}{nq}\right)^{1/2}
=\frac{1}{nq}\left\|(\mathcal{X}_{\bullet,J^c})'\mathcal{X}_{\bullet,J}\right\|_2\leq \frac{\left\|(\mathcal{X}_{\bullet,J^c})'\right\|_2}{\sqrt{nq}}\frac{\left\|\mathcal{X}_{\bullet,J}\right\|_2}{\sqrt{nq}}\nonumber\\
&\leq\rho\left(\frac{(\mathcal{X}_{\bullet,J^c})'\mathcal{X}_{\bullet,J^c}}{nq}\right)^{1/2}
\rho\left(\frac{(\mathcal{X}_{\bullet,J})'\mathcal{X}_{\bullet,J}}{nq}\right)^{1/2} \nonumber \\
&=\rho(C_{J^cJ^c})^{1/2}\rho(C_{J,J})^{1/2}\leq\rho(C)=\frac{\lambda_{\rm{max}}(\Sigma^{-1})}{q}\lambda_{\rm{max}}(X'X/n)=O_P(q^{-1}).
\end{align}
In (\ref{eq:n2_Cj_jc}) the last inequality and the fourth equality come from Theorem 4.3.15 of \cite{Horn:Johnson:1985} and (\ref{eq:C_C_tilde}), respectively. The last equality
comes from \ref{th2(v)} and \ref{th2(vii)}.
By (\ref{eq:CJJ-1:bound}), (\ref{eq:norme2_A}) and (\ref{eq:n2_Cj_jc}), it is thus enough to prove that
$$
\mathbb{P}\left(\sum_{k=1}^{nq} Z_k^2\geq \left((nq)^{1/2} \sqrt{q}\frac{\lambda}{\sqrt{nq}}\right)^2\right)=\mathbb{P}\left(\sum_{k=1}^{nq} Z_k^2\geq (nq)
\left(\frac{\lambda}{\sqrt{n}}\right)^2\right)\to 0,\textrm{ as } n\to\infty,
$$
which holds true by the second condition of \ref{assum:lambda_nq} and Markov inequality. Hence, the second term of (\ref{eq:Bn_tilde_bound}) tends to zero as $n$ tends to infinity.
Let us now study the third term of (\ref{eq:Bn_tilde_bound}).
By (\ref{eq:D_WJ}), we get that for all $j$ in $J^c$,
\begin{align*}
&\mathbb{P}\left(\left|\left(C_{J^c,J}\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)W_J\right)_j\right|\geq\frac{\lambda}{24\sqrt{nq}}\eta\right)\\
&\leq\mathbb{P}\left(\left\|C_{J^c,J}\right\|_2\|(\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\|_2\|A_1\|_2\|Z\|_2\geq \frac{\lambda}{24\sqrt{nq}}\eta\right).
\end{align*}
By (\ref{eq:norm2_Ctilde_JJ-1_-C_JJ-1}), (\ref{eq:norme2_A1}) and (\ref{eq:n2_Cj_jc}), it is thus enough to prove that
$$
\mathbb{P}\left(\sum_{k=1}^{nq} Z_k^2\geq \left((nq)^{1/2} \sqrt{q}\frac{\lambda}{\sqrt{nq}}\right)^2\right)=\mathbb{P}\left(\sum_{k=1}^{nq} Z_k^2\geq (nq)
\left(\frac{\lambda}{\sqrt{n}}\right)^2\right)\to 0,\textrm{ as } n\to\infty,
$$
which holds true by the second condition of \ref{assum:lambda_nq} and Markov inequality. Hence, the third term of (\ref{eq:Bn_tilde_bound}) tends to zero as $n$ tends to infinity.
Let us now study the fourth term of (\ref{eq:Bn_tilde_bound}). By (\ref{eq:B_WJ-WJtilde}), it amounts to prove that
$$
\mathbb{P}\left(\left\|C_{J^c,J}\right\|_2\|(\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\|_2\|A\|_2\|Z\|_2\geq \frac{\lambda}{24\sqrt{nq}}\eta\right)\to 0,\textrm{ as } n\to\infty.
$$
By (\ref{eq:n2_Cj_jc}), (\ref{eq:norm2_Ctilde_JJ-1_-C_JJ-1}) and (\ref{eq:norme2_A}) it is enough tho prove that
$$
\mathbb{P}\left(\sum_{k=1}^{nq} Z_k^2\geq(nq) \;(nq)\left(\frac{\lambda}{\sqrt{n}}\right)^2\right)\to 0,\textrm{ as } n\to\infty,
$$
which holds true by the second condition of \ref{assum:lambda_nq}. Hence, the fourth term of (\ref{eq:Bn_tilde_bound}) tends to zero as $n$ tends to infinity.
Let us now study the fifth term of (\ref{eq:Bn_tilde_bound}).
By (\ref{eq:D_WJ}), proving that the fifth term of (\ref{eq:Bn_tilde_bound}) tends to 0 amounts to proving that
$$
\mathbb{P}\left(\left\|C_{J^c,J}-\widetilde{C}_{J^c, J}\right\|_2\|(C_{J,J})^{-1}\|_2\|A_1\|_2\|Z\|_2\geq \frac{\lambda}{24\sqrt{nq}}\eta\right)\to 0,\textrm{ as } n\to\infty.
$$
Let us now bound $\|C_{J^c,J}-\widetilde{C}_{J^c, J}\|_2$.
\begin{align}\label{eq:norme2_CJcJ-CJcJ_tilde}
&\left\|C_{J^c,J}-\widetilde{C}_{J^c ,J}\right\|_2=\left\|\left(C-\widetilde{C}\right)_{J^c, J}\right\|_2
=\rho\left(\left(C-\widetilde{C}\right)_{J^c, J}\left(C-\widetilde{C}\right)_{J^c, J}\right)^{1/2}\nonumber\\
&\leq\left\|\left(C-\widetilde{C}\right)_{J^c, J}\left(C-\widetilde{C}\right)_{J^c, J}\right\|_{\infty}^{1/2}
\leq \left\|\left(C-\widetilde{C}\right)\left(C-\widetilde{C}\right)\right\|_{\infty}^{1/2}
\leq\left\|C-\widetilde{C}\right\|_{\infty}\nonumber\\
&=\frac{1}{q}\left\|\Sigma^{-1}-\widehat{\Sigma}^{-1}\right\|_{\infty}\;\left\|\frac{X'X}{n}\right\|_{\infty}=O_P(q^{-1}(nq)^{-1/2}),
\end{align}
as $n$ tends to infinity, where the last equality comes from \ref{th2(v)} and \ref{th2(ix)}.
By (\ref{eq:CJJ-1:bound}), (\ref{eq:norme2_A1}) and (\ref{eq:norme2_CJcJ-CJcJ_tilde}), to prove that the fifth term of (\ref{eq:Bn_tilde_bound}) tends to zero as $n$ tends to infinity, it is enough
to prove that
$$
\mathbb{P}\left(\sum_{k=1}^{nq} Z_k^2\geq nq\left(\frac{\lambda}{\sqrt{n}}\right)^2\right)\to 0,\textrm{ as } n\to\infty,
$$
which holds using Markov's inequality and the second condition of \ref{assum:lambda_nq}.
Using similar arguments as those used for proving that the second, third and fourth terms of (\ref{eq:Bn_tilde_bound}) tend to zero,
we get that the sixth, seventh and eighth terms of (\ref{eq:Bn_tilde_bound}) tend to zero, as $n$ tends to infinity, by replacing (\ref{eq:n2_Cj_jc}) by (\ref{eq:norme2_CJcJ-CJcJ_tilde}).
Let us now study the ninth term of (\ref{eq:Bn_tilde_bound}).
Replacing $J$ by $J^c$ in (\ref{eq:diff_Wn}), (\ref{eq:def_A}), (\ref{eq:B_WJ-WJtilde}, (\ref{eq:calculnorme2A}) and (\ref{eq:norme2_A}) in order to prove that the ninth term of
(\ref{eq:Bn_tilde_bound}) tends to 0 it is enough to prove that
$$
\mathbb{P}\left(\sum_{k=1}^{nq} Z_k^2\geq nq\left(\frac{\lambda}{\sqrt{n}}\right)^2\right)\to 0,\textrm{ as } n\to\infty,
$$
which holds using Markov's inequality and the second condition of \ref{assum:lambda_nq}.
Let us now study the tenth term of (\ref{eq:Bn_tilde_bound}). Using the same idea as the one used for proving (\ref{eq:U_s_j}), we get that
\begin{align*}
&\mathbb{P}\left(\left\{\left|C_{J^c,J}\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)\textrm{sign}(\mathcal{B}_J)\right|\geq\frac{\eta}{12}\right\}\right)\\
&\leq\mathbb{P}\left(\sqrt{|J|}\left\|C_{J^c,J}\right\|_2\; \left\|(\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right\|_2\geq\frac{\eta}{12}\right),
\end{align*}
which tends to zero as $n$ tends to infinity by \ref{th1(iii)}, (\ref{eq:norm2_Ctilde_JJ-1_-C_JJ-1}), (\ref{eq:n2_Cj_jc}) and the fact that $c_1<1/2$.
Let us now study the eleventh term of (\ref{eq:Bn_tilde_bound}). Using the same idea as the one used for proving (\ref{eq:U_s_j}), we get that
\begin{align*}
&\mathbb{P}\left(\left\{\left|\left(\widetilde{C}_{J^c,J}-C_{J^c,J}\right)(C_{J,J})^{-1}\textrm{sign}(\mathcal{B}_J)\right|\geq\frac{\eta}{12}\right\}\right)\\
&\leq \mathbb{P}\left(\sqrt{|J|}\left\|\widetilde{C}_{J^c,J}-C_{J^c,J}\right\|_2 \|(C_{J,J})^{-1}\|_2\geq\frac{\eta}{12}\right),
\end{align*}
which tends to zero as $n$ tends to infinity by \ref{th1(iii)}, (\ref{eq:CJJ-1:bound}) and (\ref{eq:norme2_CJcJ-CJcJ_tilde}) and the fact that $c_1<1/2$.
Finally, the twelfth term of (\ref{eq:Bn_tilde_bound}) can be bounded as follows:
\begin{align*}
&\mathbb{P}\left(\left\{\left|\left(\widetilde{C}_{J^c,J}-C_{J^c,J}\right)\left((\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right)\textrm{sign}(\mathcal{B}_J)\right|\geq\frac{\eta}{12}\right\}\right)\\
&\leq \mathbb{P}\left(\sqrt{|J|} \left\|\widetilde{C}_{J^c,J}-C_{J^c,J}\right\|_2\left\|(\widetilde{C}_{J,J})^{-1}-(C_{J,J})^{-1}\right\|_2\geq\frac{\eta}{12}\right),
\end{align*}
which tends to zero as $n$ tends to infinity by \ref{th1(iii)}, (\ref{eq:norm2_Ctilde_JJ-1_-C_JJ-1}) and (\ref{eq:norme2_CJcJ-CJcJ_tilde}) and the fact that $c_1<1/2$.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop_IC}]
Observe that
\begin{equation}\label{eq:Sigma-1}
\Sigma^{-1} =\left(
\begin{matrix}
1 & -\phi_1 & 0 & \cdots & 0\\
-\phi_1 & 1+\phi_1^2 & -\phi_1 & \cdots & 0 \\
0 &-\phi_1 & \ddots & \ddots & \vdots \\
\vdots & \vdots & \ddots & 1+\phi_1^2 & -\phi_1 \\
0 & 0 & \cdots & -\phi_1 & 1 &
\end{matrix}
\right).
\end{equation}
Let
$
S=\mathcal{X'}\mathcal{X}=\Sigma^{-1} \otimes X'X.
$
Then,
\begin{equation*}
S_{i,j} =\left\{
\begin{tabular}{cl}
$n_{r_i+1}$ & if $j=i$ and $k_i\in\{0,q-1\}$\\
$(1+\phi_1^2)n_{r_i+1}$ & if $j=i$ and $k_i\notin\{ 0,q-1\}$ \\
$-\phi_1 n_{r_i+1}$ & if $j=i+p$ or if $j=i-p$ \\
0 & \textrm{otherwise}
\end{tabular}
\right.,
\end{equation*}
where $i-1=(p-1)k_i+r_i$ corresponds to the Euclidean division of $(i-1)$ by $(p-1)$.
In order to prove \ref{eq:irrep}, it is enough to prove that
$$
\|S_{J^c,J} (S_{J,J})^{-1}\|_{\infty}\leq 1-\eta,
$$
where $\eta\in (0,1)$.
Since for all $j$, $(j-p)\in J^c$ or $(j+p)\in J^c$,
$$
\|S_{J^c,J}\|_\infty=\nu |\phi_1|.
$$
Let $A=S_{J,J}$. Since $A=(a_{i,j})$ is a diagonally dominant matrix, then, by Theorem 1 of \cite{varah:1975},
$$
\|A^{-1}\|_{\infty}\leq\frac{1}{\min_k (a_{k,k}-\displaystyle{\sum_{\stackrel{1\leq j\leq |J|}{j \neq k}}a_{k,j}})}.
$$
Using that for all $j$, $(j-p)\in J^c$ or $(j+p)\in J^c$,
$$
\sum_{\stackrel{1\leq j\leq |J|}{j \neq k}} a_{k,j}\leq \nu |\phi_1|.
$$
If $k\in J$ then $k>p$ and $k<pq -p$. Thus,
$$
a_{k,k}\geq \nu(1+\phi_1^2).
$$
Hence,
$$
\|A^{-1}\|_{\infty}\leq\frac{1}{\nu(1 + \phi_1^2- |\phi_1|)}
$$
and
$$
\|S_{J^c,J}(S_{J,J})^{-1}\|_\infty \leq \|S_{J^c,J}\|_\infty \|(S_{J,J})^{-1}\|_\infty \leq \frac{ |\phi_1| }{1 + \phi_1^2-|\phi_1|}.
$$
Since $|\phi_1|<1$, the strong Irrepresentability Condition holds when
$$
|\phi_1| \leq (1-\eta)(1 +|\phi_1|^2 - |\phi_1|),
$$
which is true for a small enough $\eta$.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop:cond_th_AR}]
Since $|\phi_1|<1$,
$$
\|\Sigma^{-1}\|_\infty\leq |\phi_1|+|1+\phi_1^2|\leq 3,
$$
which gives \ref{th2(vii)} by Theorem 5.6.9 of \cite{Horn:Johnson:1985}.
Observe that
$$
\|\Sigma\|_{\infty}\leq \frac{1}{1-\phi_1^2}\left(1+2\sum_{h=1}^{q-1}|\phi_1|^h\right)\leq\frac{1}{1-\phi_1^2}\left(1+\frac{2}{1-|\phi_1|}\right)=\frac{3-|\phi_1|}{1-\phi_1^2}
\leq \frac{3}{1-\phi_1^2},
$$
which gives \ref{th2(viii)} by Theorem 5.6.9 of \cite{Horn:Johnson:1985}.
Since $\widehat{\Sigma}^{-1}$ has the same expression as $\Sigma^{-1}$ defined in (\ref{eq:Sigma-1}) except that $\phi_1$ is replaced by $\widehat{\phi_1}$ defined in
(\ref{eq:phi_1_hat}), we get that
$$
\left\|\Sigma^{-1}-\widehat{\Sigma}^{-1}\right\|_{\infty}\leq 2\left|\phi_1-\widehat{\phi}_1\right|+\left(\phi_1-\widehat{\phi}_1\right)^2,
$$
which implies Assumption \ref{th2(ix)} of Theorem \ref{th2} by Lemma \ref{lem:AR:estim}.
Let us now check Assumption \ref{th2(x)} of Theorem \ref{th2}. Since, by Theorem 5.6.9 of \cite{Horn:Johnson:1985},
$\rho(\Sigma-\widehat{\Sigma})\leq \|\Sigma-\widehat{\Sigma}\|_{\infty}$, it is enough to prove that
$$
\left\|\Sigma-\widehat{\Sigma}\right\|_{\infty}=O_P((nq)^{-1/2}),\textrm{ as } n\to\infty.
$$
Observe that
\begin{align*}
&\left\|\Sigma-\widehat{\Sigma}\right\|_{\infty}\leq \left|\frac{1}{1-\phi_1^2}-\frac{1}{1-\widehat{\phi}_1^2}\right|
+2\sum_{h=1}^{q-1}\left|\frac{\phi_1^h}{1-\phi_1^2}-\frac{\widehat{\phi}_1^h}{1-\widehat{\phi}_1^2}\right|\\
&\leq \left|\frac{\phi_1^2-\widehat{\phi}_1^2}{(1-\phi_1^2)(1-\widehat{\phi}_1^2)}\right|+2\sum_{h=1}^{q-1}\left|\frac{\phi_1^h-\widehat{\phi}_1^h}{1-\phi_1^2}\right|
+2\sum_{h=1}^{q-1}\left|\widehat{\phi}_1^h\left(\frac{1}{1-\phi_1^2}-\frac{1}{1-\widehat{\phi}_1^2}\right)\right|\\
&\leq \left|\frac{(\phi_1-\widehat{\phi}_1)(\phi_1+\widehat{\phi}_1)}{(1-\phi_1^2)(1-\widehat{\phi}_1^2)}\right|+2\sum_{h=1}^{q-1}\left|\frac{\phi_1^h-\widehat{\phi}_1^h}{1-\phi_1^2}\right|
+2\sum_{h=1}^{q-1}\left|\left(\widehat{\phi}_1^h-\phi_1^h\right)\left(\frac{1}{1-\phi_1^2}-\frac{1}{1-\widehat{\phi}_1^2}\right)\right|\\
&\hspace{50mm}+2\sum_{h=1}^{q-1}\left|\phi_1^h \left(\frac{1}{1-\phi_1^2}-\frac{1}{1-\widehat{\phi}_1^2}\right)\right|\\
&\leq \left|\frac{(\phi_1-\widehat{\phi}_1)(\phi_1+\widehat{\phi}_1)}{(1-\phi_1^2)(1-\widehat{\phi}_1^2)}\right|\left(1+\frac{2}{1-|\phi_1|}\right)
+2\left(\frac{1}{|1-\phi_1^2|}+\left|\frac{(\phi_1-\widehat{\phi}_1)(\phi_1+\widehat{\phi}_1)}{(1-\phi_1^2)(1-\widehat{\phi}_1^2)}\right|\right)
\sum_{h=1}^{q-1}\left|\widehat{\phi}_1^h-\phi_1^h\right|.
\end{align*}
Moreover,
\begin{align*}
&\sum_{h=1}^{q-1}\left|\widehat{\phi}_1^h-\phi_1^h\right|\leq \left|\widehat{\phi}_1-\phi_1\right|\sum_{h=1}^{q-1}\sum_{k=0}^{h-1}
|\phi_1|^k |\widehat{\phi}_1|^{h-k-1} \leq \left|\widehat{\phi}_1-\phi_1\right|\left(\frac{1-|\widehat{\phi}_1|^{q-1}}{1-|\widehat{\phi}_1|}\right)
\left(\frac{1-|\phi_1|^{q-1}}{1-|\phi_1|}\right)\\
&\leq \left|\widehat{\phi}_1-\phi_1\right|\left(\frac{1}{1-|\widehat{\phi}_1|}\right)
\left(\frac{1}{1-|\phi_1|}\right).
\end{align*}
Thus, by Lemma \ref{lem:AR:estim},
$$
\left\|\Sigma-\widehat{\Sigma}\right\|_{\infty}=O_P((nq)^{-1/2}),
$$
which implies Assumption \ref{th2(x)} of Theorem \ref{th2}.
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:AR:estim}]
In the following, for notational simplicity, $q=q_n$.
Observe that
$$
\sqrt{nq}\widehat{\phi}_1=\frac{\frac{1}{\sqrt{nq}}\sum_{i=1}^n\sum_{\ell=2}^q \widehat{E}_{i,\ell} \widehat{E}_{i,\ell-1}}{\frac{1}{nq}\sum_{i=1}^n\sum_{\ell=1}^{q-1} \widehat{E}_{i,\ell}^2}.
$$
By (\ref{eq: Ehat}),
\begin{align}
\sum_{i=1}^n\sum_{\ell=2}^q \widehat{E}_{i,\ell} \widehat{E}_{i,\ell-1}&=\sum_{\ell=2}^q (\widehat{E}_{\bullet,\ell})'\widehat{E}_{\bullet,\ell-1}
=\sum_{\ell=2}^q (\Pi E_{\bullet,\ell})' (\Pi E_{\bullet,\ell-1})\nonumber\\
&=\sum_{\ell=2}^q (\phi_1 \Pi E_{\bullet,\ell-1} +\Pi Z_{\bullet,\ell})' (\Pi E_{\bullet,\ell-1})\label{eq:phi1_hat_def_E}\\
&=\phi_1\sum_{\ell=1}^{q-1}(\Pi E_{\bullet,\ell})'(\Pi E_{\bullet,\ell})+\sum_{\ell=2}^q (\Pi Z_{\bullet,\ell})'(\Pi E_{\bullet,\ell-1})\nonumber,
\end{align}
where (\ref{eq:phi1_hat_def_E}) comes from the definition of $(E_{i,t})$.
Hence,
$$
\sqrt{nq}(\widehat{\phi}_1-\phi_1)=\frac{\frac{1}{\sqrt{nq}}\sum_{\ell=2}^q (\Pi Z_{\bullet,\ell})'(\Pi E_{\bullet,\ell-1})}{\frac{1}{nq}\sum_{i=1}^n\sum_{\ell=1}^{q-1} \widehat{E}_{i,\ell}^2}.
$$
In order to prove that $\sqrt{nq}(\widehat{\phi}_1-\phi_1)=O_P(1)$, it is enough to prove that
\begin{equation}\label{eq:den:approx}
\frac{1}{nq}\sum_{i=1}^n\sum_{\ell=1}^{q-1} E_{i,\ell}^2-\frac{1}{nq}\sum_{i=1}^n\sum_{\ell=1}^{q-1} \widehat{E}_{i,\ell}^2=o_P(1),\textrm{ as }
n\to\infty,
\end{equation}
by Lemma \ref{lem:denom} and
\begin{equation}\label{eq:num:approx}
\frac{1}{\sqrt{nq}}\sum_{\ell=2}^q (\Pi Z_{\bullet,\ell})'(\Pi E_{\bullet,\ell-1})=O_P(1),\textrm{ as } n\to\infty.
\end{equation}
Let us first prove (\ref{eq:den:approx}). By (\ref{eq:Ehat_1}),
$$
\widehat{\mathcal{E}}=\left[\textrm{Id}_{\mathbb{R}^q}\otimes\Pi\right]\mathcal{E}:= A\mathcal{E}.
$$
Note that
$$
\mathop{\rm Cov}\nolimits(\widehat{\mathcal{E}})=A(\Sigma\otimes\textrm{Id}_{\mathbb{R}^n})A'=\Sigma\otimes\Pi.
$$
Hence, for all $i$
$$
\mathop{\rm Var}\nolimits(\widehat{\mathcal{E}}_i)\leq\lambda_{\rm{max}}(\Sigma).
$$
Since the covariance matrix of $\mathcal{E}$ is equal to $\Sigma\otimes\textrm{Id}_{\mathbb{R}^n}$, for all $i$
$$
\mathop{\rm Var}\nolimits(\mathcal{E}_i)\leq\lambda_{\rm{max}}(\Sigma).
$$
By Markov inequality,
\begin{multline*}
\frac{1}{nq}\sum_{i=1}^n\sum_{\ell=1}^{q-1} E_{i,\ell}^2-\frac{1}{nq}\sum_{i=1}^n\sum_{\ell=1}^{q-1} \widehat{E}_{i,\ell}^2
=\frac{1}{nq}\sum_{i=1}^n\sum_{\ell=1}^{q} E_{i,\ell}^2-\frac{1}{nq}\sum_{i=1}^n\sum_{\ell=1}^{q} \widehat{E}_{i,\ell}^2+o_P(1)\\
=\frac{1}{nq}\left(\|\mathcal{E}\|_2^2-\|\widehat{\mathcal{E}}\|_2^2\right)+o_P(1).
\end{multline*}
Observe that
\begin{align*}
&\|\mathcal{E}\|_2^2-\|\widehat{\mathcal{E}}\|_2^2=\|\mathcal{E}\|_2^2-\|A\mathcal{E}\|_2^2=\mathcal{E}'\mathcal{E}-\mathcal{E}'A'A\mathcal{E}
=\mathcal{E}'\left(\textrm{Id}_{\mathbb{R}^{nq}}-\textrm{Id}_{\mathbb{R}^q}\otimes\Pi\right)\mathcal{E}\\
&=\mathcal{E}'\left(\textrm{Id}_{\mathbb{R}^q}\otimes\left(\textrm{Id}_{\mathbb{R}^n}-\Pi\right)\right)\mathcal{E}=\sum_{i=1}^{pq}\widetilde{\mathcal{E}}_i^2,
\end{align*}
where $\widetilde{\mathcal{E}}=O\mathcal{E}$, where $O$ is an orthogonal matrix.
Using that
$$
\mathbb{E}(\widetilde{\mathcal{E}}_i^2)=\mathop{\rm Cov}\nolimits(\widetilde{\mathcal{E}})_{i,i}\leq\lambda_{\rm{max}}(\Sigma),
$$
and Markov inequality, we get (\ref{eq:den:approx}).
Let us now prove (\ref{eq:num:approx}).
By definition of $(E_{i,t})$ and since $|\phi_1|<1$, $\mathbb{E}[(\Pi Z_{\bullet,\ell})'(\Pi E_{\bullet,\ell-1})]=0$. Moreover,
\begin{align*}
&\mathbb{E}\left[\left(\sum_{\ell=2}^q (\Pi Z_{\bullet,\ell})'(\Pi E_{\bullet,\ell-1})\right)^2\right]
=\mathbb{E}\left[\left(\sum_{\ell=2}^q\sum_{i=1}^n \left(\sum_{k=1}^n\Pi_{i,k} Z_{k,\ell}\right)\left(\sum_{j=1}^n\Pi_{i,j} E_{j,\ell-1}\right)\right)^2\right]\\
&=\sum_{2\leq\ell,\ell'\leq q}\;\sum_{1\leq i,j,k,i',j',k'\leq n} \Pi_{i,k}\Pi_{i',k'}\Pi_{i,j}\Pi_{i',j'}\mathbb{E}\left(Z_{k,\ell}Z_{k',\ell'}E_{j,\ell-1}E_{j',\ell'-1}\right)\\
&=\sum_{2\leq\ell,\ell'\leq q}\;\sum_{1\leq i,j,k,i',j',k'\leq n} \Pi_{i,k}\Pi_{i',k'}\Pi_{i,j}\Pi_{i',j'}\sum_{r,s\geq 0}\phi_1^r \phi_1^s
\mathbb{E}\left(Z_{k,\ell}Z_{k',\ell'}Z_{j,\ell-1-r}Z_{j',\ell'-1-s}\right),
\end{align*}
since the $(E_{i,t})$ are AR(1) processes with $|\phi_1|<1$. Note that $\mathbb{E}(Z_{k,\ell}Z_{k',\ell'}Z_{j,\ell-1-r}Z_{j',\ell'-1-s})=0$ except when $\ell=\ell'$, $k=k'$, $j=j'$ and $r=s$.
Thus,
\begin{align*}
&\mathbb{E}\left[\left(\sum_{\ell=2}^q (\Pi Z_{\bullet,\ell})'(\Pi E_{\bullet,\ell-1})\right)^2\right]
=\sigma^4\left(\sum_{r\geq 0}\phi_1^{2r}\right)\sum_{\ell=2}^q\sum_{1\leq i,j,k,i'\leq n}\Pi_{i,k}\Pi_{i',k}\Pi_{i,j}\Pi_{i',j}\\
&=\frac{q\sigma^4}{1-\phi_1^2}\textrm{Tr}(\Pi)\leq\frac{nq\sigma^4}{1-\phi_1^2},
\end{align*}
where $\textrm{Tr}(\Pi)$ denotes the trace of $\Pi$, which concludes the proof of (\ref{eq:num:approx}) by Markov inequality.
\end{proof}
\section{Technical lemmas}\label{sec:lemmas}
\begin{lemma}\label{lem:proj}
Let $A\in\mathcal{M}_n(\mathbb{R})$ and $\Pi$ an orthogonal projection matrix.
For any $j$ in $\{1,\ldots,n\}$
$$
(A'\Pi A)_{jj}\geq 0.
$$
\end{lemma}
\begin{proof}[Proof of Lemma \ref{lem:proj}]
Observe that
$$
(A'\Pi A)=A'\Pi'\Pi A=(\Pi A)'(\Pi A),
$$
since $\Pi$ is an orthogonal projection matrix.
Moreover,
$$
(A'\Pi A)_{jj}=e_j'(\Pi A)'(\Pi A)e_j\geq 0,
$$
since $(\Pi A)'(\Pi A)$ is a positive semidefinite symmetric matrix, where $e_j$ is a vector containing null entries except the $j$th entry which is equal to 1.
\end{proof}
\begin{lemma}\label{lem:denom}
Assume that $(E_{1,t})_t$, $(E_{2,t})_t$, ..., $(E_{n,t})_t$ are independent AR(1) processes satisfying:
$$
E_{i,t}-\phi_1 E_{i,t-1}=Z_{i,t},\; \forall i\in\{1,\dots,n\},
$$
where the $Z_{i,t}$'s are zero-mean i.i.d. Gaussian random variables with variance $\sigma^2$ and $|\phi_1|<1$.
Then,
$$
\frac{1}{nq_n}\sum_{i=1}^n\sum_{\ell=1}^{q_n-1} E_{i,\ell}^2\stackrel{P}{\longrightarrow}\frac{\sigma^2}{1-\phi_1^2},\textrm{ as } n\to\infty.
$$
\end{lemma}
\begin{proof}[Proof of Lemma \ref{lem:denom}]
In the following, for notational simplicity, $q=q_n$.
Since $\mathbb{E}(E_{i,\ell}^2)=\sigma^2/(1-\phi_1^2)$, it is enough to prove that
$$
\frac{1}{nq}\sum_{i=1}^n\sum_{\ell=1}^{q-1} \left(E_{i,\ell}^2-\mathbb{E}(E_{i,\ell}^2)\right)\stackrel{P}{\longrightarrow} 0,\textrm{ as } n\to\infty.
$$
Since
$$
E_{i,\ell}^2=\left(\sum_{j\geq 0}\phi_1^j Z_{i,\ell-j}\right)^2=\sum_{j,j'\geq 0}\phi_1^j\phi_1^{j'} Z_{i,\ell-j} Z_{i,\ell-j'},
$$
\begin{align}\label{eq:variance}
&\mathop{\rm Var}\nolimits\left(\frac{1}{nq}\sum_{i=1}^n\sum_{\ell=1}^{q-1} \left(E_{i,\ell}^2-\mathbb{E}(E_{i,\ell}^2)\right)\right)
=\frac{1}{(nq)^2}\sum_{i=1}^n\;\sum_{1\leq\ell,\ell'\leq q-1} \mathop{\rm Cov}\nolimits(E_{i,\ell}^2 ; E_{i,\ell'}^2)\nonumber\\
&=\frac{1}{(nq)^2}\sum_{i=1}^n\;\sum_{1\leq\ell,\ell'\leq q-1}\;\sum_{j,j'\geq 0}\;\sum_{k,k'\geq 0}\;\phi_1^j\phi_1^{j'}\phi_1^k\phi_1^{k'}\mathop{\rm Cov}\nolimits(Z_{i,\ell-j} Z_{i,\ell-j'} ; Z_{i,\ell'-k} Z_{i,\ell'-k'}).
\end{align}
By Cauchy-Schwarz inequality $|\mathop{\rm Cov}\nolimits(Z_{i,\ell-j} Z_{i,\ell-j'} ; Z_{i,\ell'-k} Z_{i,\ell'-k'})|$ is bounded by a positive constant. Moreover $\sum_{j\geq 0} |\phi_1|^j<\infty$, hence
(\ref{eq:variance}) tends to zero as $n$ tend to infinity, which concludes the proof of the lemma.
\end{proof}
\bibliographystyle{chicago}
\section{Numerical experiments}\label{sec:num_exp}
\noindent The goal of this section is twofold: $i)$ to provide sanity
checks for our theoretical results in a well-controlled framework;
and $ii)$ to investigate the robustness of our estimator to some violations of the
assumptions of our theoretical results. The latter may reveal a broader scope of
applicability for our method than the one guaranteed by the theoretical results.
We investigate $i)$ in the AR(1) framework presented in
Section~\ref{subsec:AR1}. Indeed, all assumptions made in
Theorems~\ref{th1} and \ref{th2} can be specified with
well-controllable simulation parameters in the AR(1) case with
balanced design matrix $X$.
Point $ii)$ aims to explore the limitations of our theoretical
framework and assess its robustness. To this end, we propose two
numerical studies relaxing some of the assumptions of our theorems:
first, we study the effect of an unbalanced design -- which violates
the sufficient condition of the irrepresentability condition \textbf{\rm (IC)} given in Proposition \ref{prop_IC} -- on the
sign-consistency; and second, we study the effect of other types of
dependence than an AR(1).
In all experiments, the performance are assessed in terms of sign-consistency.
In other words, we evaluate the probability for the sign of
various estimators to be equal to $\ensuremath{\rm sign}(\mathcal{B})$. We compare the
performance of three different estimators:
\begin{itemize}
\item $\widehat{\mathcal{B}}$ defined in (\ref{eq:lasso}), which corresponds to the LASSO criterion applied
to the data whitened with the true covariance matrix $\Sigma$;
we call this estimator \texttt{oracle}. Its theoretical properties are established in
Theorem~\ref{th1}.
\item $\widetilde{\mathcal{B}}$ defined in (\ref{eq:lasso_tilde}), which corresponds to the LASSO criterion
applied to the data whitened with an estimator of the covariance matrix $\widehat{\Sigma}$; we refer to this estimator as
\texttt{whitened-lasso}. Its theoretical properties are established in Theorem~\ref{th2}.
\item the LASSO criterion applied to the raw data, which we call
\texttt{raw-lasso} hereafter. Its theoretical properties are established only in the univariate case in \cite{alquier:doukhan:2011}.
\end{itemize}
\subsection{AR(1) dependence structure with balanced one-way ANOVA}\label{sec:num:goodcase}
In this section, we consider Model~\eqref{eq:model:matriciel} where
$X$ is the design matrix of a one-way ANOVA with two balanced
groups. Each row of the random error matrix $E$ is distributed as a centered
Gaussian random vector as in
Equation~\eqref{eq:def_E} where the matrix $\Sigma$ is the covariance matrix of an AR(1) process
defined in Section \ref{subsec:AR1}.
In this setting, Assumptions \ref{th1(i)}, \ref{th1(ii)} and Condition~\ref{eq:irrep}
of Theorem \ref{th1} are satisfied, see Propositions \ref{propcond} and \ref{prop_IC}.
The three remaining assumptions~\ref{th1(iii)}, \ref{th1(iv)} and
\ref{assum:lambda_nq} are related to more practical quantities:
\ref{th1(iii)} controls the sparsity level of the problem, involving
$c_1$; \ref{th1(iv)} basically controls the signal-to-noise
ratio, involving $c_2$ and \ref{assum:lambda_nq} links
the sample size $n$, $q$ and the two
constants $c_1$, $c_2$, so that an appropriate range of penalty $\lambda$ exists for
having a large probability of support recovery. This latter
assumption is used in our experiments to tune the difficulty of the
support recovery as follows: we consider different values of $n$, $q$, $c_1$, $c_2$
and we choose a sparsity level $|J|$ and a
minimal magnitude in $\mathcal{B}$ such that Assumptions \ref{th1(iii)}
and \ref{th1(iv)} are fulfilled. Hence,
the problem difficulty is essentially driven by the validity of
Assumption \ref{assum:lambda_nq} where $q=o(n^{k})$ with
$c_1+c_2=1/2k$, and so by the relationship between $n$, $q$ and $k$.
We consider a large range of sample sizes $n$ varying from $10$ to
$1000$ and three different values for
$q$ in $\{10, 50, 1000 \}$. The constants $c_1$, $c_2$ are chosen such
that $c_1 + c_2=1/2k$ with $c_1=c_2$ and
$k$ in $\{1,2,4\}$. Additional values of $c_1$ and $c_2$ have also been considered
and the corresponding results are available upon request. Finally, we consider two values for
the parameter $\phi_1$ appearing in the definition of the AR(1) process: $\phi_1\in\{0.5, 0.95\}$.
Note that in this AR(1) setting with the estimator $\widehat{\phi}_1$ of $\phi_1$ defined in \eqref{eq:phi_1_hat} , all
the assumptions of Theorem~\ref{th2} are fulfilled, see Proposition~\ref{prop:cond_th_AR}.
The frequencies of support recovery for the three estimators
averaged over 1000 replications is displayed in
Figure~\ref{fig:ar1bal}.
\begin{figure}[htbp!]
\centering
\includegraphics[scale=1]{ANOVA1_BALANCE_group_AR1}
\caption{Frequencies of support recovery in a multivariate one-way
ANOVA model with two balanced groups and an AR(1) dependence.}
\label{fig:ar1bal}
\end{figure}
We observe from Figure~\ref{fig:ar1bal} that \texttt{whitened-lasso} and \texttt{oracle}
have similar performance since $\phi_1$ is well estimated. These two approaches always exhibit better performance
than \texttt{raw-lasso}, especially when $\phi_1=0.95$. In this case, the sample size $n$ required to reach the same
performance is indeed ten time larger for \texttt{raw-lasso} than for
\texttt{oracle} and \texttt{whitened-lasso}.
Finally, the performance of all estimators are altered when $n$ is too
small, especially in situations
where the signal to noise ratio (SNR) is small and the signal is not sparse enough, these two
characteristics corresponding to small values of $k$.
\subsection{Robustness to unbalanced designs and correlated features}\label{sec:num:robustness}
The goal of this section is to study some particular design matrices
$X$ in Model~\eqref{eq:model:matriciel} that may lead to violation of
the Irrepresentability Condition~\ref{eq:irrep}.
To this end, we consider the multivariate linear
model~\eqref{eq:model:matriciel} with the same AR(1) dependence
as the one considered in Section \ref{sec:num:goodcase}. Then, two
different matrices $X$ are considered:
First, an one-way ANOVA model with two unbalanced groups with
respective sizes $n_1$ and $n_2$ such that $n_1+n_2=n$; and second, a
multiple regression model with $p$ correlated Gaussian predictors such
that the rows of $X$ are i.i.d. $\mathcal{N}(0,\Sigma^X)$.
For the one-way ANOVA, violation
of \ref{eq:irrep} may occur when
$r=n_1/n$ is too different from 1/2, as stated in
Proposition~\ref{prop_IC}. For the regression model, we choose for
$\Sigma^X$ a $9\times 9$ matrix ($p=9$) such that $\Sigma_{i,i}^X=1$,
$\Sigma_{i,j}^X = \rho$, when $i\neq j$.
The other simulation parameters
are fixed as in Section~\ref{sec:num:goodcase}.
We report in
Figure \ref{fig:ic_violation} the results for the case where $q=1000$ and $k=2$
both for unbalanced one-way ANOVA (top panels) and regression with
correlated predictors (bottom panels). For the one-way ANOVA,
$r$ varies in $\{0.4, 0.2, 0.1\}$. For the regression
case, $\rho$ varies in $\{0.2, 0.6, 0.9\}$. In
both cases, the gray lines correspond to the ideal
situation (that is, either unbalanced or uncorrelated) denoted \texttt{Ideal} in the legend of Figure \ref{fig:ic_violation}.
The probability of support recovery is estimated over 1000 runs.
\begin{figure}[htbp!]
\centering
\includegraphics{violIC.pdf}
\caption{Frequencies of support recovery in general linear models
with unbalanced designs: one-way ANOVA and regression.}
\label{fig:ic_violation}
\end{figure}
From this figure, we note that correlated features or unbalanced
designs deteriorate the support recovery of all estimators. This was
expected for these LASSO-based methods which all suffer from the violation of
the irrepresentability condition~\ref{eq:irrep}. However, we also
note that \texttt{whitened-lasso} and \texttt{oracle} have similar performance, which
means that the estimation of $\Sigma$ is
not altered, and that whitening always improves the support recovery.
\subsection{Robustness to more general autoregressive processes}
In this section, we consider the case where $X$ is the design matrix of a one-way ANOVA with two balanced
groups and where $\Sigma$ is the covariance matrix of an AR($m$) process with $m$ in $\{5,10\}$.
Figure \ref{fig:arlong} displays the performance of the different estimators when $q=500$.
Here, for computing $\widehat{\Sigma}$ in \texttt{whitened-lasso}, the parameters $\phi_1,\dots,\phi_m$ of the AR($m$) process are estimated as follows.
They are obtained by averaging over the $n$ rows of $\widehat{E}$ defined in (\ref{eq: Ehat}) the estimations
$\widehat{\phi}_1^{(i)},\dots,\widehat{\phi}_m^{(i)}$ obtained for the $i$th row of $\widehat{E}$ by using standard estimation approaches for AR processes described in \cite{Brockwell:1990}.
As previously, we observe from this figure that \texttt{whitened-lasso} and \texttt{oracle} have better performance than \texttt{raw-lasso}.
\begin{figure}[htbp!]
\centering
\includegraphics[scale=1]{ARlong.pdf}
\caption{Frequencies of support recovery in one-way ANOVA with AR($m$) covariance matrix.}
\label{fig:arlong}
\end{figure}
|
\section{Introduction}
\subsection{Motivations and state of the art}
%
%
Relativistic quantum particles (electrons or quarks) confined in planar or spatial regions are efficiently described by a Hamiltonian given by the Dirac operator in a domain of the two- or three-dimensional Euclidean space with adequate boundary conditions. The question we address in this paper is related to two models of mathematical physics involving such Hamiltonians: the so-called \emph{graphene quantum dots} and the \emph{MIT bag model}. We discuss both of them below.
\paragraph{Graphene quantum dots}
These two-dimensional models come into play when investigating graphene, that is a two-dimensional allotrope of carbon where the atoms are located on an infinite hexagonal lattice (see, for instance, \cite{RevModPhys.81.109}). It turns out that the massless Dirac operator is the effective Hamiltonian describing low-energy properties of electrons in such a structure but, as in practice only finite size sheets of graphene can be obtained, one has to consider the Dirac operator in a bounded domain. The shape of this domain may varies, according to whether one is interested in nano-tubes, nano-ribbons or flakes and the bounded confining systems are called graphene quantum dots. Mathematically, this translates to the study of the massless Dirac operator imposing particular boundary conditions on the boundary of a domain (the quantum dot). On the one hand, boundary conditions can be obtained by specific cuts of the carbon sheet, the most common ones being the \emph{zigzag} and \emph{armchair} boundary conditions (see \cite{PhysRevB.77.085423}). On the other hand, another confining system can be formally obtained \emph{via} a coupling of the massless Dirac operator with a mass potential that is zero inside the quantum dot and infinite elsewhere (see for instance \cite{PhysRevB.78.195427} where this boundary condition is physically justifed and \cite{2016arXiv160309657S} for a rigorous mathematical derivation). The resulting boundary conditions are called \emph{infinite mass} boundary conditions and the two-dimensional Dirac operator in a domain with these precise boundary conditions is the operator we are interested in in this paper.
\paragraph{MIT bag model}
In the mid-70's, physicists in the MIT proposed a phenomenological model to describe the confinement of quarks inside hadrons (see \cite{MIT061974,MIT101974,MIT101975,hosaka2001quarks,johnson}) and this model, called the MIT bag model, has predicted successfully many properties of hadrons (see, for instance, \cite{PhysRevD.12.2060}). It involves the three-dimensional Dirac operator with a mass term in a bounded domain of the Euclidean space with adequate boundary conditions. These conditions can be seen as the three-dimensional counterpart of the infinite mass boundary conditions for graphene quantum dots and our interest in this three-dimensional model has drawn our attention to its two-dimensional analogue.
From a mathematical point of view, the first challenge studying Dirac operators in bounded domains is to understand on which domain they are self-adjoint. Because the Dirac operator is an elliptic operator of order one, one expect this domain to be contained in the usual Sobolev space $H^1$. Of course, it depends on the boundary conditions and it is true for the MIT bag model for sufficiently smooth domains, as proved in \cite{arrizabalaga:hal-01540149} for $\mathcal{C}^{2,1}$-smooth domains and in \cite{2016arXiv161207058O} for $\mathcal{C}^2$-smooth domains. Moreover, when one deals with $\mathcal{C}^\infty$-smooth domains more general results can be found in \cite[Thm. 4.11]{MR3618047} and in \cite{MR2536846} where the authors use pseudo-differential techniques and Calder\'on projectors. In dimension two, the question of self-adjointness is addressed in \cite{BFSV17} for $\mathcal{C}^2$-smooth domains of the Euclidean plane. For a large class of boundary conditions, the authors prove that indeed, the domain of self-adjointness is contained in $H^1$. However, we would like to point out that it is known to be false for zigzag boundary conditions (see \cite{FS14,Sch95}) and it has important consequences for the spectral features of the problem.
In this paper, we tackle the question of self-adjointness for the two-dimensional Dirac operator with infinite mass boundary conditions in sectors and, as a byproduct of the analysis, we deduce a self-adjointness result on polygonal domains. To our knowledge, this is the first attempt to mathematically handle this question for corner domains although \emph{polygonal graphene quantum dots} have drawn attention among the physicists community in the past few years (see for instance \cite{PhysRevB.77.193410,PhysRevB.81.033403,0957-4484-19-43-435401,PhysRevB.81.085432} for triangular, rectangular, trapezoidal and hexagonal graphene quantum dots). Actually, the question of self-adjointness is the first step toward future investigations. First, in a perspective of numerical applications, it is rather natural to consider polygonal domains because any two-dimensional domain, even smooth, is meshed with polygons. Second, we have in mind the investigation of the MIT bag operator in polyhedral domains in the regime of infinite mass. This is motivated by the work \cite{MR3663620} where it is proved that for smooth domains, the asymptotics of the eigenvalues in the usual Dirac gap are driven by a Laplace-Beltrami operator on the boundary with a curvature induced potential. As corners can be thought of as points of ``infinite'' curvature we aim to understand their influence on the spectrum of the MIT bag operator in this asymptotic regime. Because the geometry is less involved in dimension two, in a first attempt to shed some light on this question, we focus on the two-dimensional counterpart of the MIT bag model that is the Dirac operator with a mass term and infinite mass boundary conditions. This motivates the part of the present paper concerning basic spectral properties of such an operator.
Finally, let us describe the techniques we use in this paper. They are reminiscent of \cite[Section 4.6]{Thaller1992} and \cite{esteban2017domains} where the three-dimensional Dirac-Coulomb operator is studied as well as \cite{MR1025228} which deals with the case of a radial $\delta$-shell interaction. The key point in all these works is to investigate the restrictions of the operator to stable subspaces of functions of fixed angular momentum. Then, the restricted operators only act in the radial variable and their self-adjoint extensions can be studied using classic ODE techniques \cite{MR923320}. We obtain the result for the whole operator using the standard result \cite[Theorem X.11]{reed-2}.
%
%
\subsection{The Dirac operator with infinite mass boundary conditions in sectors}
%
%
For $\omega\in(0,\pi)$, let $\Omega_\omega$ denote the two-dimensional sector of half-aperture $\omega$
%
\begin{equation}\label{def:sector}
%
\Omega_\omega = \{
r(\cos(\theta),\sin(\theta))\in \mathbb{R}^2\,: r>0, \; |\theta|<\omega
\}\,.
%
\end{equation}
%
Let $(D, \mathcal{D}(D))$ denote the Dirac operator with mass $m\in\mathbb{R}$ and infinite mass boundary conditions in $\Omega_\omega$. It is defined by
%
\begin{equation}\label{def:DiracMIT}
%
\begin{split}
%
&
\mathcal{D}(D) =
\{
u\in H^1({\Omega_\omega}\, ;\ \mathbb{C}^2): \mathcal{B}_{\mathbf{n}} u = u \mbox{ on } \partial \Omega_\omega
\},
\\
&
D u = -i\sigma\cdot \nabla u+m\sigma_3 u, \mbox{ for all }u \in \mathcal{D}(D),
\end{split}
%
\end{equation}
%
%
where the Pauli matrices $\sigma = (\sigma_1,\sigma_2,\sigma_3)$ are $2\times2$ hermitian matrices defined by
%
\[
\sigma_1 =
\begin{pmatrix}
0&1\\1&0
\end{pmatrix}
,\,
\sigma_2 =
\begin{pmatrix}
0&-i\\i&0
\end{pmatrix}
%
\mbox{ and }
\sigma_3 =
\begin{pmatrix}
1&0\\0&-1
\end{pmatrix}.
\]
For $a\in\mathbb{R}^d$ (for $d=2,3$), we set
%
\[
\sigma\cdot a= \sum_{k=1}^d\sigma_k a_k.
\]
%
Remark that the Pauli matrices satisfy
%
\begin{equation}\label{eq:DiracMatrixProp}
%
(\sigma\cdot a)(\sigma\cdot b) = 1_2(a\cdot b) + i\sigma\cdot (a\times b),\quad \text{for } a,b \in \mathbb{R}^3.
%
\end{equation}
%
%
For almost all $s\in \partial \Omega_\omega$, $\mathbf{n}(s)$ denotes the outer unit normal at the point $s$. Let $\mathbf{v}\in\mathbb{R}^2$ be a unit vector, the matrix $\mathcal{B}_{\mathbf{v}} $ is defined by
%
\begin{equation}\label{def:BoundaryCond}
%
\mathcal{B}_\mathbf{v} = -i\sigma_3\sigma\cdot \mathbf{v}.
%
\end{equation}
%
%
%
%
%
Let us remark that $\mathcal{B}_\mathbf{v}$ satisfies
%
\begin{equation}\label{eq:propB}
%
{\mathcal{B}_\mathbf{v}}^* = \mathcal{B}_\mathbf{v},
\quad
{\mathcal{B}_\mathbf{v}}^2 = 1_2,
\quad
\Sp (\mathcal{B}_\mathbf{v}) = \{\pm 1\},
%
\end{equation}
%
where $1_2$ denotes the $2\times2$ identity matrix.
%
%
%
%
\begin{remark}
The operator $(D, \mathcal{D}(D))$ is symmetric and densely defined (see Lemma \ref{lem:sym}).
\end{remark}
%
%
%
%
\subsection{Main results}
%
%
Our main result is stated in \S \ref{ref:para-1} and concerns the question of self-\break adjointness of the operator $(D,\mathcal{D}(D))$ in a sector. When there are several self-adjoint extensions, we discuss in \S \ref{ref:para-2} which one should be chosen as the ``distinguished'' one. Finally in \S \ref{ref:para-3} we state results regarding polygonal domains and in \S \ref{ref:para-4} we give basic spectral properties of $D$.
%
%
\subsubsection{Self-adjointness in sectors}\label{ref:para-1}
%
%
In the following theorem, we give all self-adjoint extensions of the Dirac operator with infinite mass boundary conditions in sectors.
%
%
\begin{theorem}\label{thm:sa}
%
The following holds.
\begin{enumerate}[(i)]
%
\item \label{thm:cc} \textbf{[Convex sectors]}
If $\omega\in(0,\pi/2]$, $(D, \mathcal{D}(D))$ is self-adjoint.
\item\label{thm:ncc} \textbf{[Non-Convex sectors]}
If $\omega\in(\pi/2,\pi),$ $(D, \mathcal{D}(D))$ is symmetric and closed but not self-adjoint. The set of self-adjoint extensions of $D$ is the collection of operators
\[\{(D^\gamma, \mathcal{D}(D^\gamma))| \,\gamma\in \mathbb{C}, |\gamma| = 1\}\]
defined for $v\in\mathcal{D}(D) $ by
%
%
\[
\begin{split}
&\mathcal{D}(D^\gamma)
= \mathcal{D}(D)
+ \Span (v_++\gamma v_-),
\\
&D^\gamma v = D v,
\\
&D^\gamma(v_++\gamma v_-)
= i(v_+-\gamma v_-) + m\sigma_3(v_++\gamma v_-),
\end{split}
\]
%
and where
%
\[
\begin{split}
&v_+(r\cos(\theta),r\sin(\theta)) = K_{\nu_0}(r)u_0(\theta) - i K_{\nu_0+1}(r)u_{-1}(\theta),
\\
&v_-(r\cos(\theta),r\sin(\theta)) = K_{\nu_0}(r)u_0(\theta) + i K_{\nu_0+1}(r)u_{-1}(\theta),
\end{split}
%
\]
%
%
\[
u_{0}(\theta) :=\frac{1}{2\sqrt{\omega}}
\begin{pmatrix}
e^{i\theta\nu_0}\\
-ie^{-i\theta\nu_0}
\end{pmatrix},
u_{-1}(\theta) :=\frac{1}{2\sqrt{\omega}}
\begin{pmatrix}
e^{-i\theta(\nu_0+1)}\\
ie^{i\theta(\nu_0+1)}
\end{pmatrix}.
\]
%
Here, $r>0$, $\theta\in(-\omega,\omega)$,
%
%
$\nu_0 = \frac{\pi-2\omega}{4\omega}$ and $K_\nu$ denotes the modified Bessel function of the second kind of parameter $\nu$.
%
%
%
%
\end{enumerate}
%
\end{theorem}
%
%
\begin{remark} The distinction between convex an non-convex sectors in Theorem \ref{thm:sa} is not surprising: it is reminescent of \cite{MR0226187} where the so-called corner singularities for elliptic operators of even order are investigated. We also mention the books \cite{MR1173209,grisvard1980boundary} where the Laplacian in polygonal domains with various boundary conditions is studied.
\end{remark}
%
%
\begin{remark}\label{rem:turn}
%
For $\theta_0\in[0,2\pi]$, let us consider the rotated sector
\[
\Omega_{\omega,\theta_0} := \{r\big(\cos(\theta),\sin(\theta)\big)\in\mathbb{R}^2 : r>0, |\theta-\theta_0|<\omega\}.
\]
%
Remark that $e^{-i\sigma_2\theta_0}$ is a rotation matrix of angle $\theta_0$ and we have $\Omega_{\omega,\theta_0} = e^{-i\sigma_2\theta_0} \Omega_\omega$.
%
Let $\mathcal{U}_{\theta_0}$ be the unitary transformation defined by
\[
\begin{array}{lll}
\mathcal{U}_{\theta_0}:\,
& L^2(\Omega_{\omega,\theta_0},\, \mathbb{C}^2)&\longrightarrow L^2(\Omega_{\omega},\, \mathbb{C}^2)\\
& v&\longmapsto e^{i(\theta_0/2)\sigma_3} v(e^{-i\sigma_2\theta_0} \cdot ).
\end{array}
\]
It satisfies
\[
\begin{split}
%
&
\mathcal{U}_{\theta_0}^{-1}(-i\sigma\cdot \nabla +m\sigma_3)\mathcal{U}_{\theta_0} = -i\sigma\cdot \nabla +m\sigma_3,
%
\\
& \mathcal{U}_{\theta_0}^{-1}(\sigma\cdot \mathbf{n})\mathcal{U}_{\theta_0} = \sigma\cdot \left(e^{-i\sigma_2\theta_0}\mathbf{n}\right),
\end{split}
\]
for all unit vector $\mathbf{n}\in \mathbb{R}^2$ (see \cite[Sections 2 and 3]{Thaller1992}). This ensures that Theorem \ref{thm:sa} essentially covers every sectors.
%
\end{remark}
%
\begin{remark}\label{rem:besselasym}
%
Let $\nu\in \mathbb{R}$. For further use, we recall some properties of the modified Bessel functions $K_\nu$ of the second kind (see \cite[Chapter 7 Section 8 and Chapter 12 Section 1]{olver2014asymptotics} or \cite{abramowitz1964handbook}).
%
\begin{enumerate}[(i)]
%
\item The functions $r\in (0,+\infty)\mapsto K_\nu(r)\in\mathbb{R}$ are positive and decreasing
\item For $r>0$, we have
\[
K_{\nu}(r) = K_{-\nu}(r).
\]
\item For $r>0$, we have
%
\begin{equation}\label{eq:besselasym}
%
K_\nu(r) \sim_{r\to 0}
%
\begin{cases}
%
\frac{\Gamma(\nu)}{2}\left(\frac{r}{2}\right)^{-\nu}
&\mbox{ if $\nu>0$}
\\
-\log(r)
&\mbox{ if $\nu=0$}
\end{cases}
%
\end{equation}
and
\
%
K_\nu(r) \sim_{r\to +\infty}
%
\left(\frac{\pi}{2r}\right)^{1/2}e^{-r}.
%
\
\end{enumerate}
%
In particular, the domain of $D^\gamma$ (see Theorem \ref{thm:sa} \eqref{thm:ncc}) rewrites using\break
%
$
r^{-|\nu_0|}\chi(r)
$
and
$
r^{-(1-|\nu_0|)}\chi(r)
$,
%
instead of $K_{\nu_0}$ and $K_{\nu_0+1}$, respectively. Here $\chi:\mathbb{R}_+\mapsto [0,1]$ is a smooth function which equals $1$ in a neighborhood of $0$ and $0$ for $r$ large enough.
%
%
\end{remark}
%
%
\subsubsection{Physical remarks on the self-adjoint extensions in sectors}\label{ref:para-2}
%
%
For non-convex sectors, a natural question is to know whether some self-adjoint extensions given in Theorem \ref{thm:sa}\eqref{thm:ncc} are more relevant than others from the physical point of view. The following propositions try to shed some light on this question.
\vfill\eject
\par
\emph{Charge conjugation symmetry.}
The Dirac operator anticommutes with the charge conjugation operator $C$. It is defined for $u\in\mathbb{C}^2$ by
%
%
\begin{equation}\label{eqn:def_chargeconj}
C u = \sigma_1 \overline{u}.
\end{equation}
%
%
In particular, for all $\omega\in(0,\pi)$, the operator $C$ is an antiunitary transformation that leaves $\mathcal{D}(D)$ invariant, it satisfies $C^2 = 1_2$ and
%
%
\[
D C = -C D.
\]
%
This property is strongly related to the particle/antiparticle interpretation of the spectrum of the Dirac operator (see \cite[Section 1.4.6]{Thaller1992}).
%
The following proposition gives the extensions of $D$ that still satisfy these properties with respect to the charge conjugation operator $C$.
%
%
\begin{proposition}\label{prop:charge_conj} %
%
Let $\omega\in(\pi/2,\pi)$. The only self-adjoint extensions of\break $(D,\mathcal{D}(D))$ such that
%
\[
C\mathcal{D}(D^\gamma)= \mathcal{D}(D^\gamma)
\]
are the extensions $\big(D^\gamma,\mathcal{D}(D^\gamma)\big)$ for $\gamma=\pm 1$. In these cases, we have the anticommutation relation
\[
\{C,D^\gamma\} = CD^\gamma + D^\gamma C = 0.
\]
\end{proposition}
%
%
\par
\emph{Scale invariance.}
%
Since $\Omega_\omega$ is invariant by dilations, we immediately get that $\mathcal{D}(D)$ is stable under the action of the group of dilations. For non-convex sectors, we obtain the following proposition.
%
\begin{proposition}\label{prop:scaling}
Let $\omega\in(\pi/2,\pi)$. The only self-adjoint extensions of\break $(D,\mathcal{D}(D))$ such that for all $u\in \mathcal{D}(D^\gamma)$ and all $\alpha>0$ we have
%
\[
[\mathbf{x}\in\Omega_\omega\mapsto u(\alpha\mathbf{x})\in \mathbb{C}^2]\in \mathcal{D}(D^\gamma)
\]
%
are the extensions $\big(D^\gamma,\mathcal{D}(D^\gamma)\big)$ for $\gamma=\pm 1$. \end{proposition}
%
This proposition is essential in the proofs using Virial identities (see Remark \ref{rem:virial}).
%
%
\par
\emph{Kinetic energy.}
From a physical point of view, it is reasonable to impose the domain of the Dirac operator with infinite mass boundary conditions to be contained in the \emph{formal} form domain $H^{1/2}(\Omega_\omega)$. It turns out that only a single self-adjoint extension of $D$ satisfies this condition.
%
%
\begin{proposition}\label{prop:dist_ext} Let $\omega\in(\pi/2,\pi)$. The only self-adjoint extension of\break $(D,\mathcal{D}(D))$ satisfying $\mathcal{D}(D^\gamma)\subset H^{1/2}(\Omega_\omega)$ is $(D^{1},\mathcal{D}(D^{1}))$.
\end{proposition}
%
\begin{remark}\label{rem:regudom}
%
The proof of Proposition \ref{prop:dist_ext} shows a stronger statement. Indeed, if $\gamma=1$, we have $\mathcal{D}(D^\gamma)\subset H^{3/4-\varepsilon}(\Omega_\omega)$ for all $\varepsilon\in(0,1/4)$.
%
\end{remark}
%
%
%
%
\subsubsection{About polygonal domains}\label{ref:para-3}
%
Using Theorem \ref{thm:sa}\eqref{thm:cc}, Remark \ref{rem:turn} and partitions of unity, we obtain the following result.
%
%
\begin{corollary}
%
Let $\Omega\subset \mathbb{R}^2$ be a convex polygonal domain. The Dirac operator $(D_\Omega,\mathcal{D}(D_\Omega))$ defined by
%
\[
\begin{split}
%
&\mathcal{D}(D_\Omega) = \{u\in H^1(\Omega,\mathbb{C}^2),\ \mathcal{B}_\mathbf{n} u = u \mbox{ on }\partial \Omega\},
\\
&Du = -i\sigma\cdot\nabla u + m\sigma_3u\mbox{ for all } u\in \mathcal{D}(D_\Omega),
%
\end{split}
\]
%
is self-adjoint.
%
\end{corollary}
%
%
\begin{remark}
%
A similar statement can be formulated for non-convex polygonal domains using Theorem \ref{thm:sa}\eqref{thm:ncc}. We chose not to write it down here for the sake of readability.
%
\end{remark}
%
%
\subsubsection{Spectral properties in sectors}\label{ref:para-4}
%
%
Now, we investigate spectral properties of the self-adjoint operators in sectors. We restrict ourselves to the \emph{physical case} $\gamma = 1$ and, for the sake of readability, we introduce the following unified notation:
%
\[
D^{sa} = \begin{cases}
D&\mbox{ if }\omega\in(0,\pi/2],\\
D^1&\mbox{ if }\omega \in(\pi/2,\pi).
\end{cases}
\]
%
where $D$ and $D^1$ are defined in \eqref{def:DiracMIT} and Theorem \ref{thm:sa}\eqref{thm:ncc}, respectively. As defined $D^{sa}$ is self-adjoint. The following two propositions describe basic spectral properties of $D^{sa}$. The first one is about the structure of its essential spectrum and the second one deals with its point spectrum.
%
\begin{proposition}\label{prop:spec}
%
Let $\omega\in(0,\pi)$. We have
%
%
\
%
\Sp (D^{sa}) = \Sp_{ess} (D^{sa})=
%
\begin{cases}
%
\mathbb{R} &\mbox{ if }m\leq 0,
\\
\mathbb{R}\setminus (-m,m)&\mbox{ if }m\geq0.
%
\end{cases}
\]
\end{proposition}
%
%
\begin{proposition}\label{prop:pspec}
Let $\omega\in (0,\pi)$. $D^{sa}$ has no point spectrum in $\mathbb{R}\setminus (-|m|,|m|)$.
\end{proposition}
%
%
\begin{remark}\label{rem:virial}
%
The localization of the point spectrum is a consequence of the Virial identity (see in particular \cite[Section 4.7.2]{Thaller1992} and Section \ref{sec:virial}). Nevertheless, this identity gives no information on the existence of point spectrum in $(-|m|,|m|)$ for negative $m$.
%
\end{remark}
%
%
%
%
\subsection{Organization of the paper}
%
%
In Section \ref{sec:mainproofsa}, we prove Theorem \ref{thm:sa} and state the main lemmas that we need. Their proofs are gathered in Section \ref{sec:lemproofs}. In Section \ref{sec:physsa}, we discuss the physically relevant self-adjoint extensions. Finally, the spectral properties of Proposition \ref{prop:spec} and Proposition \ref{prop:pspec} are proved in Section \ref{sec:spectrum}.
%
%
\section{Self-adjoint extensions of $D$}\label{sec:mainproofsa}
%
%
%
%
In this section, we state the main lemmas on which rely the proofs of Theorem \ref{thm:sa}\eqref{thm:cc}-\eqref{thm:ncc}. Their proofs are detailed in Section \ref{sec:lemproofs}. Note that without loss of generality, we can assume that $m=0$ since $m\sigma_3$ is a bounded self-adjoint operator.
%
%
%
%
%
%
\subsection{The operator in polar coordinates}
%
%
%
%
Let us introduce the polar coordinates in $\Omega_\omega$
%
\begin{equation}\label{def:polarCoor}
%
\begin{split}
%
x(r,\theta) =
\begin{pmatrix}
x_1(r,\theta)\\
x_2(r,\theta))
\end{pmatrix}
=
\begin{pmatrix}
r\cos(\theta)\\
r\sin(\theta)
\end{pmatrix}= re_{rad}(\theta),
%
\end{split}
%
\end{equation}
%
for $r>0$ and $\theta\in(-\omega,\omega)$, where
%
\begin{equation}\label{eq:angBase}
e_{rad}(\theta) = \begin{pmatrix}
\cos(\theta)\\
\sin(\theta)
\end{pmatrix},
%
\quad
%
e_{ang}(\theta) = \frac{d}{d\theta}e_{rad}(\theta) = \begin{pmatrix}
-\sin(\theta)\\
\cos(\theta)
\end{pmatrix}.
\end{equation}
%
%
For further use, we recall the following basic relation
%
\begin{equation}\label{eqn:basiceqn}
i\sigma_3 \sigma\cdot e_{ang} = \sigma\cdot e_{rad}.
\end{equation}
%
For all $\Psi\in L^2(\Omega_\omega, \mathbb{C}^2)$, we get
%
\[
\psi(r,\theta) = \Psi(x(r,\theta))
\]
%
belongs to $L^2((0,+\infty),rdr)\otimes L^2((-\omega,\omega),\mathbb{C}^2)$.
%
In this system of coordinates, the Dirac operator rewrites
%
\begin{equation}\label{eq:DiracAng}
\begin{split}
%
D
&= -i\sigma\cdot e_{rad}\partial_r - \frac{i\sigma\cdot e_{ang}}{r}\partial_\theta
= -i\sigma\cdot e_{rad}\left(\partial_r +i \frac{\sigma_3}{r}\partial_\theta\right)
\\
&= -i\sigma\cdot e_{rad}\left(\partial_r +\frac{1_2 - K}{2r}\right)
\end{split}
\end{equation}
%
where
%
\begin{equation}\label{op:ang}
K = \sigma_3\left(
-2i\partial_\theta
\right) + 1_2.
\end{equation}
%
%
In what follows, we rely on properties of $K$ to build invariant subspaces of $D$.
%
%
\subsection{Study of the operator $K$}
%
%
Remark that for all $r>0$, the boundary matrices write
%
\begin{equation}\label{def:matpm}
%
\begin{split}
%
&
\mathcal{B}_{\mathbf{n}(re^{i\omega})} = -i\sigma_3 \sigma\cdot e_{ang}(\omega) =: \mathcal{B}_+
\\
&
\mathcal{B}_{\mathbf{n}(re^{-i\omega})} = i\sigma_3 \sigma\cdot e_{ang}(-\omega) =: \mathcal{B}_-
%
\end{split}
%
\end{equation}
%
where $\mathcal{B}_\mathbf{n}$ is defined in \eqref{def:BoundaryCond}.
%
%
%
%
%
Now, let us describe the spectral properties of $K$.
%
%
\begin{lemma}\label{prop:angOp}
The following holds.
\begin{enumerate}[(i)]
\item\label{prop:Ksa}
%
The operator $(K,\mathcal{D}(K))$ acting on $L^2((-\omega,\omega),\mathbb{C}^2)$ with $K$ defined in \eqref{op:ang} and
%
\[
\mathcal{D}(K) = \{
%
u\in H^1((-\omega,\omega),\mathbb{C}^2)\,: \mathcal{B}_+ u(\omega) = u(\omega) \mbox{ and } \mathcal{B}_- u(-\omega) = u(-\omega)
%
\}
\]
%
is self-adjoint and has compact resolvent.
\item \label{prop:KSp} Its spectrum is
%
\[
\Sp(K) = \left\{\lambda_\kappa\,, \kappa\in \mathbb{Z}\right\}
\]
%
with $\lambda_\kappa := \frac{\pi(1+2\kappa)}{2\omega}$. For $\kappa\in \mathbb{Z}$, we have $\ker\left(
K-\lambda_\kappa
\right) = \Span(u_{\kappa})$
%
where
%
\[
u_{\kappa} :=\frac{1}{2\sqrt{\omega}}
\begin{pmatrix}
e^{i\theta\frac{\lambda_\kappa-1}{2}}\\
(-1)^{\kappa+1}ie^{-i\theta\frac{\lambda_\kappa-1}{2}}
\end{pmatrix},
\]
and $(u_{\kappa})_{\kappa\in \mathbb{Z}}$ is an orthonormal basis of $L^2((-\omega,\omega),\mathbb{C}^2)$. %
\item \label{prop:Kcomm}
%
We have $\left(\sigma\cdot e_{rad} \right)\mathcal{D}(K)\subset \mathcal{D}(K)$,
%
$
\{K,\sigma\cdot e_{rad}\} = 0
$
%
and
%
\[
u_{-(\kappa+1)} = (-1)^\kappa i (\sigma\cdot e_{rad}) u_{\kappa}.
\]
%
\end{enumerate}
%
\end{lemma}
%
%
\begin{remark}
%
Thanks to Lemma \ref{prop:angOp}\eqref{prop:Kcomm}, we remark that $\Sp(K)$ is symmetric with respect to $0$.
%
\end{remark}
%
%
\begin{remark}\label{rem:commK}
%
%
The wave functions expansion in angular harmonics for the Dirac operator on $\mathbb{R}^2$ has been a major inspiration for this work. In this case, the operator acting in the angular variable is called the spin-orbit operator and is defined by
%
\[
\begin{split}
&
\tilde{K} = -2i\partial_\theta + \sigma_3 = \sigma_3K,
\\
&
\mathcal{D}(\tilde K) = H^1(\mathbb{R}/2\pi\mathbb{Z},\mathbb{C}^2).
\end{split}
\]
%
%
It is self-adjoint and commutes with the Dirac operator on $\mathbb{R}^2$ thus, the eigenspaces of $\tilde{K}$ yield invariant subspaces of the full operator.
%
We refer to \cite[Section 4.6]{Thaller1992} where the spherical symmetry in $\mathbb{R}^3$ is extensively studied.
In our case, $\tilde{K}$ does not behave well with respect to the infinite mass boundary conditions. Nevertheless, the slight change we have done overcome this difficulty. Below, we list properties of $K$ that motivates its introduction.
%
\begin{enumerate}[(a)]
\item It is a first order operator in the angular variable $\theta$.
\item Its domain takes into account the infinite mass boundary conditions and renders $K$ self-adjoint.
\item It has good anticommutation relations with $D$.
\end{enumerate}
%
\end{remark}
%
%
\subsection{Invariant subspaces of $D$}\label{subsec:stabsubspa}
%
%
Now, we introduce invariant subspaces of $D$ and study the resulting restricted operators. The following lemma is an adaptation of \cite[Theorem XIII.85]{reed-4} to our framework.
%
%
\begin{lemma}\label{lem:ssev}
%
We have
\[
L^2((0,+\infty),rdr)\otimes L^2((-\omega,\omega),\mathbb{C}^2) = \oplus_{\kappa\geq0} E_\kappa
\]
%
where $E_\kappa = L^2((0,+\infty),rdr)\otimes \Span(u_{\kappa},u_{-(\kappa+1)})$. Moreover, the following holds.
%
\begin{enumerate}[(i)]
%
\item \label{lempt:welldefssev} For all $\kappa\in \mathbb{N}$, the operator $(d^\kappa,\mathcal{D}(d^\kappa))$ defined by
%
\[
\begin{split}
&
\mathcal{D}(d^\kappa) = \mathcal{D}(D)\cap E_\kappa
\\
& d^\kappa = {D}_{\big| { E_\kappa}}
\end{split}
\]
%
is a well-defined unbounded operator on the Hilbert space $E_\kappa$.
%
%
\item \label{lempt:ssevunitequiv} For all $\kappa\in \mathbb{N}$, the operator $(d^\kappa,\mathcal{D}(d^\kappa))$ is unitarily equivalent to the operator $(\mathsf{ d}_\omega^\kappa,\mathcal{D}(\mathsf{d}_\omega^\kappa))$ defined by
%
\[
\begin{split}
&
\mathcal{D}(\mathsf{d}_\omega^\kappa)
%
=
\left\{
\begin{pmatrix}
a\\b
\end{pmatrix}
%
\in L^2((0,+\infty),\mathbb{C}^2,rdr)\, :\
\right.
\\
&
\left.
\quad\quad\quad\quad\quad
%
\int_0^\infty \left(|\dot a|^2 + |\dot{b}|^2 + \frac{|\lambda_\kappa-1|^2}{4r^2}|a|^2+ \frac{|\lambda_{\kappa}+1|^2}{4r^2}|b|^2\right)rdr <+\infty
%
\right\},
%
\\
& \mathsf{d}_\omega^\kappa
=
(-1)^\kappa\left(
i\sigma_2\left(
\partial_r + \frac{1}{2r}
\right)
+\sigma_1\frac{\lambda_\kappa}{2r}
\right) = (-1)^{\kappa} \begin{pmatrix} 0 & \partial_r + \frac{\lambda_\kappa + 1}{2r}\\ -\partial_r + \frac{\lambda_\kappa - 1}{2r} & 0\end{pmatrix}.
\end{split}
\]
%
%
\item \label{lempt:formequad} Let
%
\[
v = \sum_{\kappa\in \mathbb{Z}} a_{\kappa}u_\kappa
\]
%
be any element of $\mathcal{D}(D)$, we have
%
\[
\|D v\|^2_{L^2(\Omega_\omega,\mathbb{C}^2)} = \sum_{\kappa\in \mathbb{Z}}\int_0^\infty\left(|\dot{a_\kappa}|^2+|\lambda_\kappa-1|^2\frac{|a_\kappa(r)|^2}{4r^2}\right)rdr = \|\nabla v\|^2_{L^2(\Omega_\omega,\mathbb{C}^2)}.
\]
%
\item \label{lempt:closed}
For all $\kappa\in \mathbb{N}$, the operators $(D,\mathcal{D}(D))$ and $(\mathsf{ d}_\omega^\kappa,\mathcal{D}(\mathsf{d}_\omega^\kappa))$ are symmetric and closed.
%
\item \label{lempt:sa}
%
Let
$\{(\widetilde{\mathsf{ d}_\omega^\kappa},\mathcal{D}(\widetilde{\mathsf{ d}_\omega^\kappa}))| \ \kappa\geq0\}$ be a family of extensions of the operators $\{(\mathsf{ d}_\omega^\kappa,\mathcal{D}(\mathsf{d}_\omega^\kappa))|\ \kappa \geq0\}$. Denote by $(\widetilde D, \mathcal{D}(\widetilde D))$ the extension of $(D,\mathcal{D}(D))$ which satisfies
\[
\mathcal{D}(\widetilde D) = \oplus_{\kappa\geq0}\ \mathcal{D}(\widetilde{\mathsf{ d}_\omega^\kappa}).
\]
The operator $(\widetilde D, \mathcal{D}(\widetilde D))$ is self-adjoint if and only if the operators\break $(\widetilde{\mathsf{ d}_\omega^\kappa},\mathcal{D}(\widetilde{\mathsf{ d}_\omega^\kappa}))$ are self-adjoint. In this case, we have
\[
Sp(\widetilde D) = \bigcup_{\kappa\in\mathbb{Z}} Sp\big(\widetilde{\mathsf{ d}_\omega^\kappa}\big).
\]
%
\end{enumerate}
%
\end{lemma}
%
%
%
%
The following lemma concludes our study. Its proof relies on \cite[Theorem VIII.3]{reed-1}, \cite[Theorem X.2]{reed-2} and some properties of the modified Bessel functions (see for instance \cite[Chapter 12]{olver2014asymptotics} or \cite[Chapter 9]{abramowitz1964handbook}).
%
%
\begin{lemma}\label{lem:bessel}
%
The following holds.
%
%
\begin{enumerate}[(i)]
\item Let $\kappa \geq 1$ and $\omega\in(0,\pi)$. The operator $(\mathsf{ d}_\omega^\kappa,\mathcal{D}(\mathsf{d}_\omega^\kappa))$ is self-adjoint. When $\kappa = 0$, $(\mathsf{ d}_\omega^0,\mathcal{D}(\mathsf{d}_\omega^0))$ is self-adjoint as long as $\omega\in(0,\pi/2]$.
%
\item For all $\omega\in(\pi/2,\pi),$ $(\mathsf{ d}_\omega^0,\mathcal{D}(\mathsf{d}_\omega^0))$ is not self-adjoint but has infinitely many self-adjoint extensions $(\mathsf{ d}_\omega^{0,\gamma},\mathcal{D}(\mathsf{d}_\omega^{0,\gamma}))$ defined by
%
\[
\begin{split}
&\mathcal{D}(\mathsf{d}_\omega^{0,\gamma})
= \mathcal{D}(\mathsf{d}_\omega^{0})
+ \Span (a_++\gamma\sigma_3a_+),
\\
&\mathsf{ d}_\omega^{0,\gamma}(a + c_0(a_++\gamma\sigma_3a_+))
= \mathsf{ d}_\omega^{0}a + c_0i(a_+-\gamma\sigma_3a_+),
\end{split}
\]
%
with $a\in\mathcal{D}(\mathsf{d}_\omega^{0}) $, $c_0\in \mathbb{C}$,
%
\[
a_+ : r\mapsto \begin{pmatrix}
K_{\frac{\lambda_0-1}{2}}(r)
\\ -i K_{\frac{\lambda_0+1}{2}}(r)
\end{pmatrix}
\]
%
and $\gamma\in \mathbb{C}$ such that $|\gamma| = 1$.
%
\end{enumerate}
%
\end{lemma}
%
%
Theorem \ref{thm:sa}\eqref{thm:cc}-\eqref{thm:ncc} follow from Lemmas \ref{prop:angOp}, \ref{lem:ssev} and \ref{lem:bessel}.
%
%
\section{Proofs of Lemmas \ref{prop:angOp}, \ref{lem:ssev} and \ref{lem:bessel}}\label{sec:lemproofs}
In \S \ref{subsec:3.1}, we gather basic results that are necessary in what remains of this section. \S \ref{subsec:3.2}, \S \ref{subsec:3.3} and \S \ref{subsec:3.4} deal with the proofs of Lemmas \ref{prop:angOp}, \ref{lem:ssev} and \ref{lem:bessel}, respectively.
\subsection{Preliminary study}\label{subsec:3.1}
%
%
The following lemma is about basic spectral properties of the matrices $\mathcal{B}_\mathbf{v}$ defined in \eqref{def:BoundaryCond}.
%
%
\begin{lemma}\label{lem:PropB}
%
For all unit vector $\mathbf{v}\in \mathbb{R}^2$, the matrix $\mathcal{B}_\mathbf{v}$ satisfies
%
%
\[
\ker (\mathcal{B}_\mathbf{v}\pm1_2) = \sigma_3\ker (\mathcal{B}_\mathbf{v}\pm1_2)^\perp = \sigma\cdot \mathbf{v}\ker (\mathcal{B}_\mathbf{v}\pm1_2)^\perp.
\]
%
%
\end{lemma}
%
%
\begin{proof}
Since $\{\sigma_3,\mathcal{B}_\mathbf{v}\} = 0$, we have
%
\[
\sigma_3\ker (\mathcal{B}_\mathbf{v}\pm1_2)^\perp = \sigma_3 \ran (\mathcal{B}_\mathbf{v}\pm1_2) = \ran (\mathcal{B}_\mathbf{v}\mp1_2) = \ker (\mathcal{B}_\mathbf{v}\pm1_2).
\]
%
Moreover, as $\{\sigma\cdot \mathbf{v},\mathcal{B_\mathbf{v}}\} = 0$, the same proof yields the other equality.
\end{proof}
%
%
For the sake of completeness, we recall the following standard result on the symmetry of the Dirac operator with infinite mass boundary conditions.
%
\begin{lemma}\label{lem:sym}
%
The operator $(D, \mathcal{D}(D))$ is symmetric and densely defined.
%
\end{lemma}
%
%
%
%
\begin{proof}
%
Let $u,v\in\mathcal{D}(D)$. Since $\Omega_\omega$ is a Lipschitz domain, an integrations by parts yields
%
\[
\braket{D u,v}_{L^2(\Omega_\omega,\mathbb{C}^2)} - \braket{u,D v}_{L^2(\Omega_\omega,\mathbb{C}^2)}
%
= \braket{-i\sigma\cdot \mathbf{n} u,v}_{L^2(\partial\Omega_\omega,\mathbb{C}^2)},
\]
%
(see \cite[Section 3.1.2]{necas2011direct}).
%
Thanks to Lemma \ref{lem:PropB}, almost everywhere on the boundary we have
%
\[
-i(\sigma\cdot \mathbf{n}) u\in \sigma\cdot \mathbf{n}\ker (\mathcal{B}_\mathbf{n}-1_2) = \ker (\mathcal{B}_\mathbf{n}-1_2)^\perp.
\]
%
Thus $\braket{-i\sigma\cdot \mathbf{n} u,v}_{L^2(\partial\Omega_\omega,\mathbb{C}^2)} = 0$ and we obtain
%
\[
\braket{D u,v}_{L^2(\Omega_\omega,\mathbb{C}^2)} = \braket{u,D v}_{L^2(\Omega_\omega,\mathbb{C}^2)}.
\]
%
\end{proof}
%
%
%
\subsection{Study of the angular part : proof of Lemma \ref{prop:angOp}}\label{subsec:3.2}
%
%
The proof is divided into several steps.
%
\subsubsection*{Step 1: symmetry of $K$.}
Let $u,v\in \mathcal{D}(K)$, an integration by parts and Lemma \ref{lem:PropB} yield
%
\begin{align}\label{eq:intByPartsK}
%
&\braket{Ku,v}_{L^2((-\omega,\omega),\mathbb{C}^2)}-\braket{u,Kv}_{L^2((-\omega,\omega),\mathbb{C}^2)} \nonumber\\
&=
\int_{-\omega}^\omega\partial_\theta\big(\braket{-2i\sigma_3u,v}_{\mathbb{C}^2}\big)d\theta
\nonumber\\
&=\braket{-2i\sigma_3u(\omega),v(\omega)}_{\mathbb{C}^2}-\braket{-2i\sigma_3u(-\omega),v(-\omega)}_{\mathbb{C}^2} =0.
%
\end{align}
%
Hence, $K$ is symmetric.
%
\subsubsection*{Step 2: self-adjointness of $K$}
Let $u\in \mathcal{D}(K^*)$. Using test functions in\break $C^\infty_c((-\omega,\omega),\mathbb{C}^2)\subset \mathcal{D}(K)$, we remark that the distribution $Ku$ belongs to $L^2((-\omega,\omega),\mathbb{C}^2)$ which gives $u\in H^1((-\omega,\omega),\mathbb{C}^2)$. Performing again integration by parts \eqref{eq:intByPartsK}, we get
%
\[
\begin{split}
%
&\sigma_3u(\omega) \in \ker (\mathcal{B}_+-1_2)^\perp,\\
&\sigma_3u(-\omega) \in \ker (\mathcal{B}_--1_2)^\perp.
%
\end{split}
\]
%
Thanks to Lemma \ref{lem:PropB}, we obtain that $u$ belongs to $\mathcal{D}(K)$ and thus $K$ is self-adjoint. Finally, the compact Sobolev embedding
\[
H^1((-\omega,\omega),\mathbb{C}^2)\hookrightarrow L^2((-\omega,\omega),\mathbb{C}^2),
\]
implies that $K$ has compact resolvent. Hence, its spectrum is discrete and it concludes the proof of Lemma \ref{prop:angOp}\eqref{prop:Ksa}.
%
\subsubsection*{Step 3: study of $Sp(K)$}
%
Let $\lambda\in \mathbb{R}$, we look for solutions of
%
\begin{equation}\label{eq:Kspec}
Ku = \lambda u
\end{equation}
%
belonging to $\mathcal{D}(K)$. Remark that without taking the boundary conditions into account, the set of solutions of \eqref{eq:Kspec} is the vector space
%
\[
E_\lambda^1 := \Span \left(
\begin{pmatrix}
e^{i\theta\frac{\lambda-1}{2}}\\
0
\end{pmatrix},
\begin{pmatrix}
0\\
e^{-i\theta\frac{\lambda-1}{2}}
\end{pmatrix}
\right).
\]
%
Assume $u \in E^1_\lambda\cap \mathcal{D}(K)$. In particular, $u$ writes
%
\[
u =\begin{pmatrix}
ae^{i\theta\frac{\lambda-1}{2}}\\
be^{-i\theta\frac{\lambda-1}{2}}
\end{pmatrix}
\]
%
for some constants $a,b\in\mathbb{C}$. Using \eqref{eqn:basiceqn}, the boundary conditions read
%
{\small
\[
\begin{split}
&
u(\omega) = \mathcal{B}_+ u(\omega) = -\sigma\cdot e_{rad}(\omega)u(\omega) = -\begin{pmatrix}
0&e^{-i\omega}
\\
e^{i\omega}&0
\end{pmatrix}u(\omega)
=
\begin{pmatrix}
-be^{-i\omega\frac{\lambda+1}{2}}\\
-ae^{i\omega\frac{\lambda+1}{2}}
\end{pmatrix},
\\
&
u(-\omega) = \mathcal{B}_- u(-\omega) = \sigma\cdot e_{rad}(-\omega)u(-\omega) = \begin{pmatrix}
0&e^{i\omega}
\\
e^{-i\omega}&0
\end{pmatrix}u(-\omega)
=
\begin{pmatrix}
be^{i\omega\frac{\lambda+1}{2}}\\
ae^{-i\omega\frac{\lambda+1}{2}}
\end{pmatrix}.
\end{split}
\]
}
%
\noindent It yields
%
\[
a = be^{i\omega\lambda} = -be^{-i\omega\lambda}.
\]
%
Hence, there is a nontrivial solution of \eqref{eq:Kspec} belonging to $\mathcal{D}(K)$ if and only if $e^{2i\omega\lambda} = -1$. We deduce that the spectrum of $K$ is
%
\[
\Sp(K) = \left\{\frac{\pi(1+2\kappa)}{2\omega}\,,\ \kappa\in \mathbb{Z}\right\}.
\]
%
For $\kappa \in \mathbb{Z}$, define $\lambda_\kappa := \frac{\pi(1+2\kappa)}{2\omega}$. If $\kappa$ is even, we have $a = ib$ and
%
\[
\ker\left(
K-\lambda_\kappa
\right)
=
\Span
\begin{pmatrix}
e^{i\theta\frac{\lambda_\kappa-1}{2}}\\
-ie^{-i\theta\frac{\lambda_\kappa-1}{2}}
\end{pmatrix}
,
\]
%
if $\kappa$ is odd, we have $a = -ib$ and
%
\[
\ker\left(
K-\lambda_\kappa
\right)
=
\Span
\begin{pmatrix}
e^{i\theta\frac{\lambda_\kappa-1}{2}}\\
ie^{-i\theta\frac{\lambda_\kappa-1}{2}}
\end{pmatrix}
.
\]
This proves Lemma \ref{prop:angOp}\eqref{prop:KSp}.
%
\subsubsection*{Step 4: the commutation relation}
%
Since $\sigma\cdot e_{rad}$ commutes with $\mathcal{B}_\mathbf{n}$, we obtain
\[
(\sigma\cdot e_{rad})\mathcal{D}(K)\subset \mathcal{D}(K).
\]
%
We also have
%
\[
\begin{split}
K\sigma\cdot e_{rad} &= \sigma_3\left(\sigma\cdot e_{rad}(-2i\partial_\theta) -2i\sigma\cdot e_{ang}\right) + \sigma\cdot e_{rad}
\\
&
= -\sigma\cdot e_{rad}\sigma_3(-2i\partial_\theta)- \sigma\cdot e_{rad} = -\sigma\cdot e_{rad}K.
\end{split}
\]
This ends the proof of Lemma \ref{prop:angOp}\eqref{prop:Kcomm}.
%
%
%
%
\subsection{Invariant subspaces: proof of Lemma \ref{lem:ssev}}\label{subsec:3.3}
%
%
Let us remark that the direct sum decomposition is a direct consequence of Lemma \ref{prop:angOp}\eqref{prop:KSp}.
%
%
What remains of the proof is divided into several steps.
%
%
\subsubsection*{Proof of Points \eqref{lempt:welldefssev} and \eqref{lempt:ssevunitequiv}}
%
%
These points follow from identity \eqref{eq:DiracAng}
%
\[
D
= -i\sigma\cdot e_{rad}\left(\partial_r +\frac{1_2 - K}{2r}\right)
\]
%
and Lemma \ref{prop:angOp}\eqref{prop:Kcomm}.
%
Indeed, for all $\kappa\in \mathbb{N}$ and all $v\in E_\kappa$ there exist $a,b\in L^2((0,+\infty),rdr)$ such that for all $r>0$ and all $\theta\in(-\omega,\omega)$, $v$ writes
%
\[
v(r,\theta) = a(r)u_\kappa(\theta) +b(r)u_{-(\kappa+1)}(\theta).
\]
%
If $v\in H^1(\Omega_\omega,\mathbb{C}^2)$, since $-i\sigma_3\partial_\theta = \frac{K-1}{2}$, we have
\[
\|\nabla v\|_{L^2(\Omega_\omega,\mathbb{C}^2)}^2 = \int_0^\infty \left(|\dot a|^2 + |\dot{b}|^2 + \frac{|\lambda_\kappa-1|^2}{4r^2}|a|^2+ \frac{|\lambda_{-(\kappa+1)}-1|^2}{4r^2}|b|^2\right)rdr
\]
and
%
\[
D v = d^\kappa v = (-1)^{\kappa+1}u_{-(\kappa+1)}\left(\dot{a} + \frac{1-\lambda_\kappa}{2r}a\right) + (-1)^{\kappa} u_\kappa\left(\dot{b} + \frac{1+\lambda_\kappa}{2r}b\right).
\]
This ends this part of the proof.
%
%
%
\subsubsection*{Proof of Points \eqref{lempt:formequad} and \eqref{lempt:closed}}
%
%
Let $v\in\mathcal{D}(D)$. Decomposing $v$ in the orthonormal basis $(u_\kappa)_{\kappa\in\mathbb{Z}}$, it writes
%
\[
v = \sum_{\kappa\in \mathbb{Z}} a_{\kappa}u_\kappa.
\]
%
Using Lemma \ref{lem:radfun}, we have
%
\begin{align*}
\|D v\|^2_{L^2(\Omega_\omega,\mathbb{C}^2)}
&=
\sum_{\kappa\in \mathbb{Z}}\int_0^\infty\left|\dot{a_\kappa} + \frac{1-\lambda_\kappa}{2r}a_\kappa\right|^2rdr\\
&=
\sum_{\kappa\in \mathbb{Z}}\int_0^\infty\left(|\dot{a_\kappa}|^2+|\lambda_\kappa-1|^2\frac{|a_\kappa(r)|^2}{4r^2}\right)rdr
\\
& = \|\nabla v\|^2_{L^2(\Omega_\omega,\mathbb{C}^2)}.
\end{align*}
%
%
and Lemma \ref{lem:ssev}\eqref{lempt:closed} is proved.
%
\subsubsection*{Proof of Point \eqref{lempt:sa}}
%
This last point is proved as in \cite[Lemma 4.15]{Thaller1992} using \cite[Theorem VIII.3]{reed-1}.
%
%
%
\subsection{Proof of Lemma \ref{lem:bessel}}\label{subsec:3.4}
%
%
%
%
Let $\kappa\in \mathbb{N}$.
In this proof, we apply the basic criterion for self-adjointness \cite[Theorem VIII.3]{reed-1}. In particular, we have to study the vector spaces
\[
\ker(({\mathsf{d}_\omega^\kappa})^*\pm i1_2).
\]
%
Remark that
%
\[
\mathcal{D}(({\mathsf{d}_\omega^\kappa})^*)
%
\subset \left\{ \begin{pmatrix}
a\\b
\end{pmatrix}\in L^2((0,\infty),rdr)^2\,:\
{\mathsf{d}_\omega^\kappa}\begin{pmatrix}
a\\b
\end{pmatrix}\in L^2((0,\infty),rdr)^2
\right\}.
\]
%
Since $\{\mathsf{d}_\omega^\kappa,\sigma_3\}=0$, we obtain
%
\[
\ker(({\mathsf{d}_\omega^\kappa})^*-i1_2) = \sigma_3\ker(({\mathsf{d}_\omega^\kappa})^*+i1_2).
\]
%
Hence, it remains to look for $L^2((0,\infty),rdr)^2$-solutions of
\begin{equation}\label{eqn:soladj}
\left(\mathsf{d}_\omega^\kappa- i1_2\right) \begin{pmatrix}
a\\b
\end{pmatrix} = 0.
\end{equation}
%
It is well known that the set of solutions is a vector space of dimension $2$ and moreover, the solutions are smooth on $(0,\infty)$. Remark that
%
\[
\left(\mathsf{d}_\omega^\kappa+ i1_2\right)\left(\mathsf{d}_\omega^\kappa- i1_2\right) = (\mathsf{d}_\omega^\kappa)^2+1_2,
\]
%
which implies
%
{\small
\begin{multline*}
0 =\left((\mathsf{d}_\omega^\kappa)^2+ 1_2\right) \begin{pmatrix}
a\\b
\end{pmatrix}
%
\\= -\frac{1}{r^2}\begin{pmatrix}
r^2\partial_r^2 + r\partial_r - \left(r^2+\frac{(\lambda_\kappa-1)^2}{4}\right) & 0\\
0 & r^2\partial_r^2 + r\partial_r - \left(r^2+\frac{(\lambda_\kappa+1)^2}{4}\right)
\end{pmatrix} \begin{pmatrix}
a\\b
\end{pmatrix}.
\end{multline*}
}
\noindent Thus, $a$ and $b$ are modified Bessel functions (see \cite[Chapter 12, Section 1]{olver2014asymptotics} and \cite{abramowitz1964handbook}) of parameters $\frac{\lambda_\kappa-1}{2}$ and $\frac{\lambda_\kappa+1}{2}$, respectively.
%
The modified Bessel functions of the first kind do not belong to $L^2((1,\infty),rdr)$. Consequently, for $(a,b)$ to belong to $L^2((1,\infty),rdr)^2$ $a$ and $b$ necessarily write $a=a_0K_{\frac{\lambda_\kappa-1}{2}}$ and $b = b_0K_{\frac{\lambda_\kappa+1}{2}}$ with $a_0,b_0\in \mathbb{C}$. Recall that $K_{\nu}$ denotes the modified Bessel function of the second kind of parameter $\nu\in \mathbb{R}$. It is known that $K_{\nu} = K_{-\nu}$ and
\begin{equation}\label{eq:difBess}
\begin{split}
%
\dot{K_{|\nu|}}(r) + \frac{|\nu|}{r}K_{|\nu|}(r) = -K_{|\nu|-1}(r)
%
\end{split}
\end{equation}
%
for all $r>0$.
%
Thanks to Remark \ref{rem:besselasym}, for $b$ to belong to $L^2((0,1),rdr)$ one needs
$
\lambda_\kappa<1.
$
We have
%
%
\begin{enumerate}[(a)]
\item\label{cas1} $\lambda_\kappa\geq 3/2 \mbox{ for any }\kappa \geq 1 \mbox{ and any }\omega\in(0,\pi),$
\item\label{cas2}$\lambda_0\geq1 \mbox{ for any }\omega\in(0,\pi/2],$
\item\label{cas3}$\lambda_0<1 \mbox{ for any }\omega\in(\pi/2,\pi).$
\end{enumerate}
%
Hence, in Cases \eqref{cas1} and \eqref{cas2}, we get $b_0=0$. Taking into account \eqref{eqn:soladj} and \eqref{eq:difBess} we also get $a_0 = 0$ which implies
\[
\ker(({\mathsf{d}_\omega^\kappa})^*\pm i1_2) = \{0\}
\]
and \cite[Theorem VIII.3]{reed-1} ensures that $(\mathsf{d}_\omega^\kappa,\mathcal{D}(\mathsf{d}_\omega^\kappa))$ is a self-adjoint operator.
%
In Case \eqref{cas3}, we get
%
\[
\begin{pmatrix}
a\\b
\end{pmatrix}
\in \Span (a_+^0),\mbox{ with }a_+^0 =
\begin{pmatrix}
K_{\frac{\lambda_0-1}{2}}(r)
\\- i K_{\frac{\lambda_0+1}{2}}(r)
\end{pmatrix} .
\]
%
Actually, $a_+^0$ belongs to $\mathcal{D}(({\mathsf{d}_\omega^\kappa})^*)$ which yields
%
\[
\ker(({\mathsf{d}_\omega^\kappa})^*- i1_2) = \Span (a_+^0) \mbox{ and } \ker(({\mathsf{d}_\omega^\kappa})^*+ i1_2) = \Span (\sigma_3 a_+^0).
\]
We conclude thanks to \cite[Theorem X.2]{reed-2}.
%
%
\section{Distinguished self-adjoint extensions of $D$}\label{sec:physsa}
%
%
The goal of this section is to prove Propositions \ref{prop:charge_conj} and \ref{prop:dist_ext} about the distinguished extensions of $\big(D,\mathcal{D}(D)\big)$ when $\omega\in(\pi/2,\pi)$.
%
%
\subsection{Proof of Proposition \ref{prop:charge_conj}}
%
%
The anticommutation of $C$ with $D$ is straightforward. The only thing left to prove is the following lemma.
%
%
\begin{lemma}\label{lemma:}Let $\omega\in(\pi/2,\pi)$ and let $\gamma\in\mathbb{C}$ be such that $|\gamma| = 1$. The following statements are equivalent.
%
\begin{enumerate}[(a)]
\item\label{itm:a} $\gamma = \pm 1$.
\item\label{itm:b} $C \mathcal{D}(D^\gamma) \subset \mathcal{D}(D^\gamma)$ and $D^\gamma C = -C D^\gamma$.
\end{enumerate}
%
\end{lemma}
%
%
\begin{proof} Let $u\in \mathcal{D}(D^\gamma)$. Thanks to Theorem \ref{thm:sa}\eqref{thm:ncc}, there exist $v\in \mathcal{D}(D)$ and $c_0\in\mathbb{C}$ such that $u = v + c_0(v_+ + \gamma v_-)$. The following equalities hold:
%
%
\[
C v_+ = i v_+,\quad C v_- = i v_-.
\]
%
%
As $C u = C v + i\overline{c_0}(v_+ + \overline{\gamma}v_-)$ we have $Cu \in \mathcal{D}(D^\gamma)$ if and only if $\gamma\in\mathbb{R}$, that is $\gamma = \pm 1$. Now, let $\gamma=\pm1$, we have
%
%
\[
D^\gamma Cu = D Cv - \overline{c_0}(v_+ - \gamma v_-) = -C D v - \overline{c_0}(v_+ - \gamma v_-).
\]
%
%
As $D^\gamma(v_+ + \gamma v_-) = i(v_+ -\gamma v_-)$, we get $CD^\gamma(v_+ + \gamma v_-) = (v_+ - \gamma v_-)$ which yields $D^\gamma Cu = -CD^\gamma u$.
\end{proof}
%
%
\subsection{Proof of Proposition \ref{prop:scaling}}
%
%
Proposition \ref{prop:scaling} is a consequence of the following lemma.
%
\begin{lemma}\label{lem:scal}
%
Let $\alpha>0$ and $\gamma=e^{is}\in \mathbb{C}$ for $s\in [0,2\pi)$. The unitary map
%
\begin{equation}\label{eqn:transunitmap}
\begin{array}{lll}
%
\mathcal{V}_\alpha : &L^2(\Omega_\omega,\mathbb{C}^2)&\longrightarrow L^2(\Omega_\omega,\mathbb{C}^2)
\\
&u&\longmapsto \alpha u(\alpha\cdot)
%
\end{array}
\end{equation}
%
satisfies
%
\[
\begin{split}
%
&\mathcal{V}_\alpha^{-1}(-i\sigma\cdot \nabla)\mathcal{V}_\alpha = \alpha (-i\sigma\cdot \nabla),
\\
&\mathcal{V}_\alpha\mathcal{D}(D^\gamma) =
\begin{cases}
\mathcal{D}(D^\gamma) & \mbox{ if }\gamma = \pm 1 (\text{ i.e. if } s\in\{0,\pi\}),\\
\mathcal{D}(D^{\tilde \gamma}) & \mbox{ otherwise},
\end{cases}
%
\end{split}
%
\]
%
where $\tilde \gamma = e^{2i\arctan\left(\frac{\tan(s/2)}{\alpha^{\lambda_0}}\right)}$.
%
\end{lemma}
%
%
\begin{proof}
Let $\alpha>0$ and $\gamma = e^{is}\in \mathbb{C}$. As $\mathcal{D}(\mathsf{d}_\omega^\kappa)$ is scale-invariant for all $\kappa\geq 1$, using Lemma \ref{lem:ssev} we are reduced to investigate the operator $\mathsf{d}_\omega^{0,\gamma}$.
Thanks to Remark \ref{rem:besselasym} we have
\[
\mathcal{V}_\alpha\mathcal{D}(D^{\pm 1}) =
\mathcal{D}(D^{\pm 1})
\]
and for $\gamma\ne -1$ and $r>0$,
\[
(1+\gamma)\begin{pmatrix}
K_{\frac{1-\lambda_0}{2}}(\alpha r)
\\ -i\frac{1-\gamma}{1+\gamma} K_{\frac{1+\lambda_0}{2}}(\alpha r)
\end{pmatrix}
%
\sim_{r\rightarrow 0}
%
\frac{1+\gamma}{\alpha^{\frac{1-\lambda_0}{2}}}
\begin{pmatrix}
K_{\frac{1-\lambda_0}{2}}( r)
\\ -i\alpha^{-\lambda_0}\frac{1-\gamma}{1+\gamma} K_{\frac{1+\lambda_0}{2}}( r)
\end{pmatrix}
.
\]
%
We have
%
\[
-i\frac{1-\gamma}{1+\gamma}
=-\tan(s/2)
\]
%
which rewrites
%
\[
-i\alpha^{-\lambda_0}\frac{1-\gamma}{1+\gamma} = - \alpha^{-\lambda_0}\tan(s/2)=-i\frac{1-\tilde \gamma}{1+\tilde \gamma}
\]
%
for $\tilde \gamma = e^{2i\arctan\left(\frac{\tan(s/2)}{\alpha^{\lambda_0}}\right)}$. This ensures that $\mathcal{V}_\alpha\mathcal{D}(D^\gamma) = \mathcal{D}(D^{\tilde \gamma})$ and the result follows.
%
\end{proof}
%
%
\subsection{Proof of Proposition \ref{prop:dist_ext}}
%
%
To prove Proposition \ref{prop:dist_ext} it is enough to prove the following lemma.
%
%
\begin{lemma}\label{lem:sobinje} Let $\omega\in(\pi/2,\pi)$ and let $\nu_0$ be as defined in Theorem \ref{thm:sa}. The following holds true.
\begin{enumerate}[(i)]
\item\label{spaceok} The function
%
\[(r\cos(\theta),r\sin(\theta))\in\Omega_\omega \mapsto K_{\nu_0}(r)u_0(\theta)\]
%
belongs to $H^{1/2}(\Omega_\omega)$.
%
\item\label{spacenotok} The function
%
\[(r\cos(\theta),r\sin(\theta))\in\Omega_\omega \mapsto K_{\nu_0+1}(r)u_{-1}(\theta)\]
%
does not belong to $H^{1/2}(\Omega_\omega)$.
\end{enumerate}
\end{lemma}
%
%
\begin{proof}
%
Using \cite[Cor. 4.53.]{MR2440993}, we have $H^{1/2}(\Omega_\omega)\hookrightarrow L^4(\Omega_\omega)$ and thanks to Remark \ref{rem:besselasym} we get
%
\[
\left|
%
K_{\nu_0+1}(r)u_{-1}(\theta)
%
\right|^4r= \frac{1}{2^3\omega^2}
%
|K_{\nu_0+1}(r)|^4r \sim_{r\to0} \frac{C}{r^{4(\nu_0+1)-1}}
\]
%
for all $r>0$ and $\theta\in(-\omega,\omega)$.
%
Since
%
\[
\nu_0 = \frac{\pi-2\omega}{4\omega}>-1/2,
\]
this function does not belong to $L^4(\Omega_\omega)$
and Lemma \ref{lem:sobinje}\eqref{spacenotok} is proved.
%
%
%
Let us prove Lemma \ref{lem:sobinje}\eqref{spaceok}. Let $r>0$ and $\theta\in(-\omega,\omega)$, we have
%
\[
|\nabla K_{\nu_0}u_0|^2(r,\theta) = \frac{1}{4\omega}|\partial_r K_{\nu_0}(r)|^2 + \frac{|K_{\nu_0}(r)|^2}{4\omega r^2}2|\nu_0|^2.
\]
%
Thanks to \eqref{eq:difBess} and Remark \ref{rem:besselasym}, $K_{\nu_0}u_0$ belongs to $W^{1,p}(\Omega_\omega)$ as soon as
\[
1\leq p<\frac{2}{|\nu_0|+1}.
\]
%
Since we have
%
\[
\min_{\omega\in(\pi/2,\pi)}\,\frac{2}{|\nu_0|+1} = \frac{8}{5}>\frac{4}{3}
\]
%
and $W^{1,4/3}(\Omega_\omega)\hookrightarrow H^{1/2}(\Omega_\omega)$, we get Lemma \ref{lem:sobinje}\eqref{spaceok}.
%
%
\end{proof}
%
%
\section{Spectrum of $D^{sa}$}\label{sec:spectrum}
%
%
This section is devoted to the proofs of Propositions \ref{prop:spec} and \ref{prop:pspec}.
%
\subsection{Proof of Proposition \ref{prop:spec}}
The proof is divided into three steps. In Steps 1 and 2 we construct Weyl sequences for the Dirac operator with infinite mass boundary conditions on $\Omega_{\pi/2}$ and the free Dirac operator in $\mathbb{R}^2$ denoted by $D_0$, respectively. For a general $\omega\in(0,\pi)$, Weyl sequences for the Dirac operator with infinite mass boundary conditions in $\Omega_{\omega}$ can be obained using adequate cut-off functions. The last step ensures that the Weyl sequences actually yield the whole essential spectrum.
%
\emph{Step 1: Weyl sequences for $m\geq 0$ on $\mathbb{R}^2$.}
Let $\chi : [0,+\infty)\mapsto [0,1]$ be a $C^\infty$-smooth function such that
\begin{equation}\label{eqn:defWeylchi}
\chi(x) = \begin{cases}
1&\mbox{ if }x<1\\
0&\mbox{ if }x>2.
\end{cases}
\end{equation}
Let $\lambda>m$ and define
\[
u:\left\{\begin{array}{lcl}
\mathbb{R}^2&\longrightarrow&\mathbb{C}^2\\
(x_1,x_2)&\longmapsto& \begin{pmatrix}
\sqrt{\frac{\lambda+m}{\lambda-m}}\\1
\end{pmatrix}
e^{ix_1\sqrt{\lambda^2-m^2}}.
\end{array}\right.
\]
In particular, remark that
$
(-i\sigma\cdot\nabla+m\sigma_3-\lambda)u = 0.
$
Now, for $n>0$, we define the sequence of functions
\[
u_n(x_1,x_2) = u(x_1,u_2)\chi(|x_1|/n)\chi(|x_2|/n) \in H^1(\mathbb{R}^2,\mathbb{C}^2).
\]
We get
\[
\|u_n\|_{L^2(\mathbb{R}^2,\mathbb{C}^2)}^2 = 2 \Big(\frac{\lambda}{\lambda - m}\Big) n^2 \|\chi\|_{L^2(\mathbb{R})}^4
\]
and
\begin{align*}
&\|(-i\sigma\cdot\nabla +m\sigma_3 - \lambda)u_n\|_{L^2(\mathbb{R}^2,\mathbb{C}^2)}^2\\
&\leq \frac{4}{n^2} \Big(\frac{\lambda}{\lambda - m}\Big)\int_{\mathbb{R}^2} \big(\chi'(|x_1|/n)\chi(|x_2|/n)\big)^2 dx_1dx_2
\\ &= 4\Big(\frac{\lambda}{\lambda - m}\Big)\int_{\mathbb{R}^2} \big(\chi'(|x_1|)\chi(|x_2|)\big)^2 dx_1dx_2.
\end{align*}
Thus, we obtain
\[
\frac{\|(-i\sigma\cdot\nabla +m\sigma_3 - \lambda)u_n\|_{L^2(\mathbb{R}^2,\mathbb{C}^2)}^2}{\|u_n\|_{L^2(\mathbb{R}^2,\mathbb{C}^2)}^2} \rightarrow 0,\quad\text{when } n\rightarrow+\infty.
\]
In particular, $\lambda$ belongs to the spectrum of the free Dirac operator and thus $(m,+\infty)\subset Sp(D_0)$. As the spectrum of a self-adjoint operator is closed, the end-point $m$ also belongs to the spectrum. Recall that $C$ is the charge conjugation operator introduced in \eqref{eqn:def_chargeconj}. The same reasoning yield that the sequence $(Cu_n)_{n>0}$ is also a Weyl sequence but for the value $-\lambda$. In particular we obtain
\[
(-\infty,-m]\cup[m,+\infty) \subset Sp(D_0).
\]
As the set on the left-hand side is not discrete, actually we have proved that
\[
(-\infty,-m]\cup[m,+\infty) \subset Sp_{ess}(D_0).
\]
Finally, localizing the sequences $(u_n)_{n>0}$ and $(Cu_n)_{n>0}$ inside $\Omega_\omega$ and away from the boundary (with well chosen cut-off functions), we obtain
\[
(-\infty,-m]\cup[m,+\infty) \subset Sp_{ess}(D^{sa}).
\]
%
%
\emph{Step 2: Weyl sequences for $m<0$ on $\Omega_{\pi/2}$.} Let $\lambda\in \mathbb{R}$ and define
\[
u:\left\{\begin{array}{lcl}
\Omega_{\pi/2}&\longrightarrow&\mathbb{C}^2\\
(x_1,x_2)&\longmapsto&\begin{pmatrix}
1\\-i
\end{pmatrix}
e^{mx_1-i\lambda x_2}.
\end{array}\right.
\]
In particular, remark that we have $(-i\sigma\cdot\nabla +m\sigma_3 - \lambda)u = 0$ and $\mathcal{B}_{-e_1}u = u$ on $\partial \Omega_{\pi/2}$ where $e_1=(1,0)^T$. We define the sequence of functions $(u_n)_{n>0}$ by
\[
u_n(x_1,x_2) = u(x_1,x_2)\chi(|x_2|/n),\quad \text{ for } (x_1,x_2) \in \Omega_{\pi/2}
\]
and with $\chi$ defined in \eqref{eqn:defWeylchi}. Note that as constructed, $u_n \in \mathcal{D}(D^{sa})$. We get
\begin{align*}
\|u_n\|^2_{L^2(\Omega_{\pi/2},\mathbb{C}^2)} &= 2\left(\int_0^{+\infty}e^{2mx_1}dx_1\right)\left(\int_\mathbb{R} |\chi(|x_2|/n)|^2dx_2\right)\\
& = \frac{n}{|m|}\int_\mathbb{R} |\chi(|x_2|)|^2dx_2
\end{align*}
and
\begin{align*}
\|(D^{sa}-\lambda)u_n\|^2_{L^2(\Omega_{\pi/2},\mathbb{C}^2)}
&= \frac{2}{n^2}\||\partial_2 \chi(|x_2|/n)|e^{mx_1}\|^2_{L^2(\Omega_{\pi/2},\mathbb{C}^2)} \\
& = \frac{2}{|m|n}\int_\mathbb{R} |\partial_2\chi(|x_2|)|^2dx_2.
\end{align*}
This proves that $(u_n)_{n>0}$ is a Weyl sequence and $\lambda$ belongs to the spectrum of $D^{sa}$. We get
$\mathbb{R} \subset Sp(D^{sa})$ which actually read
\[
Sp(D^{sa}) = Sp_{ess}(D^{sa}) = \mathbb{R}.
\]
This proof can be adapted to the domain $\Omega_\omega$ with $\omega\in(0,\pi)$ and negative mass using Remark \ref{rem:turn} and adequate cut-off functions.
\emph{Step 3: Reverse inclusion for the essential spectrum.} The only thing left to investigate is the case $m>0$.
%
Thanks to Remark \ref{rem:regudom}, we have $\mathcal{D}(D^{sa})\subset H^{3/4-\varepsilon}(\Omega_\omega,\mathbb{C}^2)$ for all $\varepsilon\in(0,1/4)$. Hence, for all $u\in\mathcal{D}(D^{sa})$, an integration by parts yields
\[
\begin{split}
%
2\Re\braket{-i\sigma\cdot \nabla u,\sigma_3u}_{L^2(\Omega_\omega,\mathbb{C}^2)
%
%
%
= \|u\|^2_{L^2(\partial\Omega_\omega,\mathbb{C}^2)}.
%
\end{split}
\]
It gives
\[
\begin{split}
\|D^{sa}u\|^2_{L^2(\Omega_\omega,\mathbb{C}^2)}
&
= \|\sigma\cdot \nabla u\|^2_{L^2(\Omega_\omega,\mathbb{C}^2)} + m^2\|u\|^2_{L^2(\Omega_\omega,\mathbb{C}^2)} + m\|u\|^2_{L^2(\partial\Omega_\omega,\mathbb{C}^2)}.
\end{split}
\]
%
This ensures that whenever $m\geq 0$, the spectrum of $(D^{sa})^2$ is included in $[m^2,+\infty)$ and in particular, we have
\[
\Sp(D^{sa})\subset (-\infty,-m]\cup[m,+\infty).
\]
It concludes the proof of Proposition \ref{prop:spec}.
\subsection{Proof of Proposition \ref{prop:pspec}}\label{sec:virial}
%
Assume that $\lambda \in Sp(D^{sa})$ is an eigenvalue. Let $v$ be a normalized eigenfunction associated with $\lambda$. For $\alpha>0$, recall that $\mathcal{V}_\alpha$ is the unitary map introduced in \eqref{eqn:transunitmap}.
Thanks to Proposition \ref{prop:scaling}, $\mathcal{V}_\alpha\big(\mathcal{D}(D^{sa})\big) \subset \mathcal{D}(D^{sa})$. Moreover, we have
\[\begin{split}
0
&= \braket{(D^{sa} - \lambda)v, \mathcal{V}_\alpha v}_{L^2(\Omega_\omega,\mathbb{C}^2)}
= \braket{v,(D^{sa} - \lambda)\mathcal{V}_\alpha v}_{L^2(\Omega_\omega,\mathbb{C}^2)}
\\
&= \braket{v,\mathcal{V}_\alpha(\alpha (D^{sa}-\lambda) +(\alpha-1)(\lambda-m\sigma_3))v}_{L^2(\Omega_\omega,\mathbb{C}^2)}
\\
&= (\alpha-1)\braket{v,\mathcal{V}_\alpha(\lambda-m\sigma_3)v}_{L^2(\Omega_\omega,\mathbb{C}^2)}.
\end{split}\]
For $\alpha\neq 1$ we obtain
\[
0 = \braket{(\lambda 1_2 -m\sigma_3)v, \mathcal{V}_\alpha v}_{L^2(\Omega_\omega,\mathbb{C}^2)}.
\]
Since $\Big(\alpha\in(0,+\infty) \mapsto \mathcal{V}_\alpha v \in L^2(\Omega_\omega,\mathbb{C}^2)\Big)$ is continuous, taking the limit $\alpha\rightarrow 1$ gives
\begin{equation}\label{eqn:vir1}
\lambda\|v\|^2_{L^2(\Omega_\omega,\mathbb{C}^2)} = m\braket{\sigma_3v, v}_{L^2(\Omega_\omega,\mathbb{C}^2)}
\end{equation}
which yields
\[
|\lambda|\|v\|^2_{L^2(\Omega_\omega,\mathbb{C}^2)} = |m| \big|\braket{\sigma_3v, v}_{L^2(\Omega_\omega,\mathbb{C}^2)}\big|\leq |m|\|v\|^2_{L^2(\Omega_\omega,\mathbb{C}^2)}
\]
and necessarily $|\lambda|\leq |m|$. Assume by contradiction that $\lambda = m$. Writing $v = (v_1,v_2)$, \eqref{eqn:vir1} implies that $v_2$ vanishes on $\Omega_\omega$. Taking the infinite mass boundary conditions into account, we get $v_1|_{\partial\Omega_\omega} = 0$ which implies that $v_1 \in H_0^1(\Omega_\omega,\mathbb{C})$. Moreover, since $(v_1, 0)$ is an eigenfunction of $D^{sa}$, $v_1$ satisfies the Cauchy-Riemann equation
\[
-i(\partial_1+i\partial_2)v_1 = 0 \mbox{ in }\Omega_\omega.
\]
An integration by parts gives $\|\nabla v_1\|_{L^2(\Omega_\omega, \mathbb{C}^2)} = 0$ and $v_1$ has to vanish. This gives us the wanted contradiction and the case $\lambda = -m$ can be done similarly. It ends the proof of Proposition \ref{prop:pspec}
%
%
|
\section{Introduction}
\label{intro}
The difficulty of clustering a set of $n$ objects into $k$ disjoint clusters is one that is well known among researchers. Many methods have been proposed either to find the best partition according to a dissimilarity-based homogeneity criterion, or to fit a mixture model of multivariate distribution function. However, in some clustering problems, it is relevant to impose constraints on the set of allowable solutions. In the literature, a variety of different solutions have been suggested and applied in a number of fields, including earth science, image processing, social science, and - more recently - genetics. The most common type of constraints are contiguity constraints (in space or in time). Such restrictions occur when the objects in a cluster are required not only to be similar to one other, but also to comprise a contiguous set of objects. But what is a contiguous set of objects?
Consider first that the contiguity between each pair of objects is given by a matrix ${\bf C}=(c_{ij})_{n\times n}$, where $c_{ij}=1$ if the $i$th and the $j$th objects are regarded as contiguous, and 0 if they are not. A cluster $C$ is then considered to be contiguous if there is a path between every pair of objects in $C$ (the subgraph is connected). Several classical clustering algorithms have been modified to take this type of constraint into account (see e.g., Murtagh 1985a; Legendre and Legendre 2012; B\'ecue-Bertaut et al. 2014). Surveys of some of these methods can be found in Gordon (1996) and Murtagh (1985b). For instance, the standard hierarchical procedure based on Lance and Williams formula (1967) can be constrained by merging only contiguous clusters at each stage. But what defines ``contiguous" clusters? Usually, two clusters are regarded as contiguous if there are two objects, one from each cluster, which are linked in the contiguity matrix. But this can lead to reversals (i.e. inversions, upward branching in the tree) in the hierarchical classification. It was proven that only the complete link algorithm is guaranteed to produce no reversals when relational constraints are introduced in the ordinary hierarchical clustering procedure (Ferligoj and Batagelj 1982). Recent implementation of strict constrained clustering procedures are available in the R package \texttt{const.clust} (Legendre 2014) and in the Python library \texttt{clusterpy} (Duque et al. 2011). Hierarchical clustering of SNPs (Single Nucleotide Polymorphism) with strict adjacency constraint is also proposed in Dehman et al. (2015) and implemented in the R package \texttt{BALD} (www.math-evry.cnrs.fr/logiciels/bald). The recent R package \texttt{Xplortext} (B\'ecue-Bertaut et al 2017) implements also chronogically constrained agglomerative hierarchical clustering for the analysis of textual data.
The previous procedures which impose strict contiguity may separate objects which are very similar into different clusters, if they are spatially apart. Other non-strict constrained procedures have then been developed, including those referred to as soft contiguity or spatial constraints.
For example, Oliver and Webster (1989) and Bourgault et al. (1992) suggest running clustering algorithms on a modified dissimilarity matrix. This dissimilarity matrix is a combination of the matrix of geographical distances and the dissimilarity matrix computed from non-geographical variables. According to the weights given to the geographical dissimilarities in this combination, the solution will have more or less spatially contiguous clusters. However, this approach raises the problem of defining weight in an objective manner.
In image processing, there are many approaches for image segmentation including for instance usage of convolution and wavelet transforms. In this field non-strict spatially constrained clustering methods have been also developed. Objects are pixels and the most common choices for the neighborhood graph are the four and eight neighbors graphs. A contiguity matrix ${\bf C}$ is used (and not a geographical dissimilarity matrix as previously) but the clusters are not strictly contiguous, as a cluster of pixels does not necessarily represent a single region on the image. Ambroise et al. (1997, 1998) suggest a clustering algorithm for Markov random fields based on an EM (Expectation-Maximization) algorithm. This algorithm maximizes a penalized likelihood criterion and the regularization parameter gives more or less weight to the spatial homogeneity term (the penalty term). Recent implementations of spatially-located data clustering algorithms are available in SpaCEM3 (spacem3.gforge.inria.fr), dedicated to Spatial Clustering with EM and Markov Models. This software uses the model proposed in Vignes and Forbes (2009) for gene clustering via integrated Markov models. In a similar vein, Miele et al. (2014) proposed a model-based spatially constrained method for the clustering of ecological networks. This method embeds geographical information within an EM regularization framework by adding some constraints to the maximum likelihood estimation of parameters. The associated R package is available at http://lbbe.univ-lyon1.fr/geoclust. Note that all these methods are partitioning methods and that the constraints are neighborhood constraints.
In this paper, we propose a hierarchical clustering (and not partitioning) method including spatial constraints (not necessarily neighborhood constraints). This Ward-like algorithm uses two dissimilarity matrices $D_0$ and $D_1$ and a mixing parameter $\alpha \in [0,1]$. The dissimilarities are not necessarily Euclidean (or non Euclidean) distances and the weights of the observations can be non-uniform. The first matrix gives the dissimilarities in the `feature space' (socio-economic variables or grey levels for instance). The second matrix gives the dissimilarities in the `constraint space'. For instance, $D_1$ can be a matrix of geographical distances or a matrix built from the contiguity matrix ${\bf C}$. The mixing parameter $\alpha$ sets the importance of the constraint in the clustering procedure. The criterion minimized at each stage is a convex combination of the homogeneity criterion calculated with $D_0$ and the homogeneity criterion calculated with $D_1$. The parameter $\alpha$ (the weight of this convex combination) controls the weight of the constraint in the quality of the solutions. When $\alpha$ increases, the homogeneity calculated with $D_0$ decreases whereas the homogeneity calculated with $D_1$ increases. The idea is to determine a value of $\alpha$ which increases the spatial-contiguity without deteriorating too much the quality of the solution on the variables of interest. The R package \texttt{ClustGeo} (Chavent et al. 2017) implements this constrained hierarchical clustering algorithm and a procedure for the choice of $\alpha$.
The paper is organized as follows. After a short introduction (this section), Section \ref{sec:2} presents the criterion optimized when the Lance-Williams (1967) parameters are used in Ward's minimum variance method but dissimilarities are not necessarily Euclidean (or non-Euclidean) distances.
We also show how to implement this procedure with the package {\tt ClustGeo} (or the R function {\tt hclust}) when non-uniform weights are used. In Section \ref{sec:3} we present the constrained hierarchical clustering algorithm which optimizes a convex combination of this criterion calculated with two dissimilarity matrices. Then the procedure for the choice of the mixing parameter is presented as well as a description of the functions implemented in the package {\tt ClustGeo}. In Section \ref{sec:4} we illustrate the proposed hierarchical clustering process with geographical constraints using the package {\tt ClustGeo} before a brief discussion given in Section \ref{sec:5}.
Throughout the paper, a real dataset is used for illustration and reproducibility purposes. This dataset contains 303 French municipalities described based on four socio-economic variables. The matrix $D_0$ will contain the socio-economic distances between municipalities and the matrix $D_1$ will contain the geographical distances. The results will be easy to visualize on a map.
\section{Ward-like hierarchical clustering with dissimilarities and non-uniform weights}
\label{sec:2}
Let us consider a set of $n$ observations. Let $w_i$ be the weight of the $i$th observation for $i=1,\dots,n$.
Let $D=[d_{ij}]$ be a $n\times n$ dissimilarity matrix associated with the $n$ observations, where $d_{ij}$ is the dissimilarity measure between observations $i$ and $j$.
Let us recall that the considered dissimilarity matrix $D$ is not necessarily a matrix of Euclidean (or non-Euclidean) distances. When $D$ is not a matrix of Euclidean distances, the usual inertia criterion (also referred to as variance criterion) used in Ward (1963) hierarchical clustering approach is meaningless and the Ward algorithm implemented with the Lance and Williams (1967) formula has to be re-interpreted.
The Ward method has already been generalized to use with non-Euclidean distances, see e.g. Strauss and von Maltitz (2017) for $l_1$ norm or Manhattan distances. In this section the more general case of dissimilarities is studied.
We first present the homogeneity criterion which is optimized in that case and the underlying aggregation measure which leads to a Ward-like hierarchical clustering process. We then provide an illustration using the package {\tt ClustGeo} and the well-known R function {\tt hclust}.
\subsection{The Ward-like method}
\paragraph{Pseudo-inertia.}
Let us consider a partition $\mathcal{P}_K=(\mathcal{C}_1,\ldots,\mathcal{C}_K)$ in $K$ clusters. The pseudo-inertia of a cluster $\mathcal{C}_k$ generalizes the inertia to the case of dissimilarity data (Euclidean or not) in the following way :
\begin{equation}\label{inertia}
I(\mathcal{C}_k)=\sum_{i \in \mathcal{C}_k}\sum_{j \in \mathcal{C}_k}\frac{w_i w_j}{2\mu_k}d^2_{ij}
\end{equation}
where $\mu_k=\sum_{i \in \mathcal{C}_k}w_i$ is the weight of $\mathcal{C}_k$. The smaller the pseudo-inertia $I(\mathcal{C}_k)$ is, the more homogenous are the observations belonging to the cluster $\mathcal{C}_k$.
\noindent
The pseudo within-cluster inertia of the partition $\mathcal{P}_K$ is therefore:
$$W(\mathcal{P}_K)=\sum_{k=1}^K I(\mathcal{C}_k).$$
The smaller this pseudo within-inertia $W(\mathcal{P}_K)$ is, the more homogenous is the partition in $K$ clusters.
\smallskip
\paragraph{Spirit of the Ward hierarchical clustering.}
To obtain a new partition $\mathcal{P}_{K}$ in $K$ clusters from a given partition $\mathcal{P}_{K+1}$ in $K+1$ clusters,
the idea is to aggregate the two clusters $\mathcal{A}$ and $\mathcal{B}$ of $\mathcal{P}_{K+1}$ such that the new partition has minimum within-cluster inertia (heterogeneity, variance), that is:
\begin{equation}\label{opti1}
\arg\min_{\mathcal{A},\mathcal{B}\in \mathcal{P}_{K+1}} W(\mathcal{P}_{K}),
\end{equation}
where $\mathcal{P}_{K}=\mathcal{P}_{K+1} \backslash \{\mathcal{A},\mathcal{B}\} \cup \{\mathcal{A}\cup\mathcal{B}\}$ and
$$W(\mathcal{P}_{K})=W(\mathcal{P}_{K+1})-I(\mathcal{A})-I(\mathcal{B})+I(\mathcal{A} \cup \mathcal{B}).$$
Since $W(\mathcal{P}_{K+1})$ is fixed for a given partition $\mathcal{P}_{K+1}$, the optimization problem (\ref{opti1}) is equivalent to:
\begin{equation}\label{opti2}
\min_{\mathcal{A},\mathcal{B}\in \mathcal{P}_{K+1}} I(\mathcal{A} \cup \mathcal{B})-I(\mathcal{A})-I(\mathcal{B}).
\end{equation}
The optimization problem is therefore achieved by defining
$$\delta(\mathcal{A},\mathcal{B}):=I(\mathcal{A} \cup \mathcal{B})-I(\mathcal{A})-I(\mathcal{B})$$
as the aggregation measure between two clusters which is minimized at each step of the hierarchical clustering algorithm.
Note that $\delta(\mathcal{A},\mathcal{B})=W(\mathcal{P}_K)-W(\mathcal{P}_{K+1})$ can be seen as the increase of within-cluster inertia (loss of homogeneity).
\medskip
\paragraph{Ward-like hierarchical clustering process for non-Euclidean dissimilarities.}
The interpretation of the Ward hierarchical clustering process in the case of dissimilarity data is the following:
\begin{itemize}
\item Step $K=n$: initialization.
The initial partition $\mathcal{P}_n$ in $n$ clusters (i.e. each cluster only contains an observation) is unique.
\item Step $K=n-1,\dots,2$: obtaining the partition in $K$ clusters from the partition in $K+1$ clusters.
At each step $K$, the algorithm aggregates the two clusters $\mathcal{A}$ and $\mathcal{B}$ of $\mathcal{P}_{K+1}$ according to the optimization problem (\ref{opti2}) such that the increase of the pseudo within-cluster inertia is minimum for the selected partition over the other ones in $K$ clusters.
\item Step $K=1$: stop. The partition $\mathcal{P}_1$ in one cluster (containing the $n$ observations) is obtained.
\end{itemize}
The hierarchically-nested set of such partitions $\{\mathcal{P}_n,\dots,\mathcal{P}_K,\dots,\mathcal{P}_1\}$ is represented graphically by a tree (also called dendrogram) where the height of a cluster $\mathcal{C} = \mathcal{A} \cup \mathcal{B}$ is
$h(\mathcal{C}):=\delta(\mathcal{A},\mathcal{B}).$
\medskip
In practice, the aggregation measures between the new cluster $\mathcal{A} \cup \mathcal{B}$ and any cluster $\mathcal{D}$ of $\mathcal{P}_{K+1}$ are calculated at each step thanks to the well-known Lance and Williams (1967) equation:
\begin{equation}\label{LW-equation}
\begin{array}{lcl}
\delta(\mathcal{A} \cup \mathcal{B},\mathcal{D}) & = & \displaystyle \frac{\mu_\mathcal{A} + \mu_\mathcal{D}}{\mu_\mathcal{A} + \mu_\mathcal{B} + \mu_\mathcal{D}}\delta(\mathcal{A},\mathcal{D}) + \frac{\mu_\mathcal{B} + \mu_\mathcal{D}}{\mu_\mathcal{A} + \mu_\mathcal{B} + \mu_\mathcal{D}}\delta(\mathcal{B},\mathcal{D})\\
&&\\
&-&\displaystyle \frac{\mu_\mathcal{D}}{\mu_\mathcal{A} + \mu_\mathcal{B} + \mu_\mathcal{D}}\delta(\mathcal{A},\mathcal{B}).
\end{array}
\end{equation}
In the first step the partition is $\mathcal{P}_n$ and the aggregation measures between the singletons are calculated with
$$\delta_{ij}:=\delta(\{i\},\{j\})=\frac{w_i w_j}{w_i+w_j} d^2_{ij},$$
and stored in the $n \times n$ matrix $\Delta=[\delta_{ij}]$. For each subsequent step $K$, the Lance and Williams formula (\ref{LW-equation}) is used to build the corresponding $K\times K$ aggregation matrix.
The hierarchical clustering process described above is thus suited for non-Euclidean dissimilarities and then for non-numerical data. In this case, it optimises the pseudo within-cluster inertia criterion (\ref{opti2}).
\paragraph{Case when the dissimilarities are Euclidean distances.}
When the dissimilarities are Euclidean distances calculated from a numerical data matrix $X$ of dimension $n \times p$ for instance, the pseudo-inertia of a cluster $\mathcal{C}_k$ defined in (\ref{inertia}) is now equal to the inertia of the observations in $\mathcal{C}_k$:
$$I(\mathcal{C}_k)=\sum_{i \in \mathcal{C}_k} w_i d^2(x_i,g_k)$$
where $x_i\in \Re^p$ is the $i$th row of $X$ associated with the $i$th observation, and $g_k=\frac{1}{\mu_k}\sum_{i \in \mathcal{C}_k} w_i x_i \in R^p$ is the center of gravity of $\mathcal{C}_k$.
The aggregation measure $\delta(\mathcal{A},\mathcal{B})$ between two clusters is written then as:
$$\delta(\mathcal{A},\mathcal{B})=\frac{\mu_\mathcal{A} \mu_\mathcal{B}}{\mu_\mathcal{A}+\mu_\mathcal{B}}d^2(g_\mathcal{A},g_\mathcal{B}).$$
\subsection{Illustration using the package {\tt ClustGeo}}
Let us examine how to properly implement this procedure with R. The dataset is made up of $n=303$ French municipalities described based on $p=4$ quantitative variables and is available in the package {\tt ClustGeo}. A more complete description of the data is provided in Section \ref{sec:data}.
{\small
\begin{verbatim}
> library(ClustGeo)
> data(estuary)
> names(estuary)
[1] "dat" "D.geo" "map"
\end{verbatim}
}
\noindent
To carry out Ward hierarchical clustering, the user can use the function {\tt hclustgeo} implemented in the package {\tt ClustGeo} taking the dissimilarity matrix $D$
(which is an object of class {\tt dist}, i.e. an object obtained with the function {\tt dist} or a dissimilarity matrix transformed in an object of class {\tt dist} with the function {\tt as.dist})
and the weights $w=(w_1,\dots,w_n)$ of observations as arguments.
{\small
\begin{verbatim}
> D <- dist(estuary$dat)
> n <- nrow(estuary$dat)
> tree <- hclustgeo(D, wt=rep(1/n,n))
\end{verbatim}
}
\paragraph{Remarks.}
\begin{itemize}
\item The function {\tt hclustgeo} is a wrapper of the usual function {\tt hclust} with the following arguments:
\begin{itemize}
\item {\tt method = "ward.D"},
\item {\tt d =} $\Delta$,
\item {\tt members = w}.
\end{itemize}
For instance, when the observations are all weighted by $1/n$ , the argument {\tt d} must be the matrix $\Delta=\frac{D^2}{2n}$ and not the dissimilarity matrix $D$:
{\small
\begin{verbatim}
> tree <- hclust(D^2/(2*n), method="ward.D")
\end{verbatim}
}
\item As mentioned before, the user can check that the sum of the heights in the dendrogram is equal to the total pseudo-inertia of the dataset:
{\small
\begin{verbatim}
> inertdiss(D, wt=rep(1/n, n)) # the pseudo-inertia of the data
[1] 1232.769
> sum(tree$height)
[1] 1232.769
\end{verbatim}
}
\item When the weights are not uniform, the calculation of the matrix $\Delta$ takes a few lines of code and the use of the function {\tt hclustgeo} is clearly more convenient than {\tt hclust}:
{\small
\begin{verbatim}
> w <- estuary$map@data$POPULATION # non-uniform weights
> tree <- hclustgeo(D, wt=w)
> sum(tree$height)
[1] 1907989
\end{verbatim}
}
versus
{\small
\begin{verbatim}
> Delta <- D
> for (i in 1:(n-1)) {
for (j in (i+1):n) {
Delta[n*(i-1) - i*(i-1)/2 + j-i] <-
Delta[n*(i-1) - i*(i-1)/2 + j-i]^2*w[i]*w[j]/(w[i]+w[j])
}}
> tree <- hclust(Delta, method="ward.D", members=w)
> sum(tree$height)
[1] 1907989
\end{verbatim}
}
\end{itemize}
\section{Ward-like hierarchical clustering with two dissimilarity matrices}
\label{sec:3}
Let us consider again a set of $n$ observations, and let $w_i$ be the weight of the $i$th observation for $i=1,\dots,n$.
Let us now consider that two $n\times n$ dissimilarity matrices $D_0=[d_{0,ij}]$ and $D_1=[d_{1,ij}]$ are provided. For instance, let us assume that the $n$ observations are municipalities, $D_0$ can be based on a numerical data matrix of $p_0$ quantitative variables measured on the $n$ observations and $D_1$ can be a matrix containing the geographical distances between the $n$ observations.
\smallskip
\noindent
In this section, a Ward-like hierarchical clustering algorithm is proposed. A mixing parameter $\alpha\in [0,1]$ allows the user to set the importance of each dissimilarity matrix in the clustering procedure. More specifically, if $D_1$ gives the dissimilarities in the constraint space, the mixing parameter $\alpha$ sets the importance of the constraint in the clustering procedure and controls the weight of the constraint in the quality of the solutions.
\subsection{Hierarchical clustering algorithm with two dissimilarity matrices}
For a given value of $\alpha \in [0,1]$, the algorithm works as follows. Note that the partition in $K$ clusters will be hereafter indexed by $\alpha$ as follows: $\mathcal{P}_{K}^\alpha$.
\paragraph{Definitions.} The {\bf mixed} pseudo inertia of the cluster $C_k^\alpha$ (called mixed inertia hereafter for sake of simplicity) is defined as
\begin{equation}\label{heterogeneity-cluster}
I_\alpha(\mathcal{C}_k^\alpha) = (1-\alpha) \sum_{i \in \mathcal{C}_k^\alpha}\sum_{j \in \mathcal{C}^\alpha_k}\frac{w_i w_j}{2\mu_k^\alpha}d^2_{0,ij} + \alpha \sum_{i \in \mathcal{C}_k^\alpha}\sum_{j \in \mathcal{C}_k^\alpha}\frac{w_i w_j}{2\mu_k^\alpha}d^2_{1,ij},
\end{equation}
where $\mu_k^\alpha=\sum_{i \in \mathcal{C}_k^\alpha}w_i$ is the weight of $\mathcal{C}_k^\alpha$ and $d_{0,ij}$ (resp. $d_{1,ij}$) is the normalized dissimilarity between observations $i$ and $j$ in $D_0$ (resp. $D_1$).
\noindent
The {\bf mixed} pseudo within-cluster inertia (called mixed within-cluster inertia hereafter for sake of simplicity) of a partition $\mathcal{P}_K^\alpha=(\mathcal{C}_1^\alpha,\dots,\mathcal{C}_K^\alpha)$ is the sum of the mixed inertia of its clusters:
\begin{equation}\label{heterogeneity-partition}
W_\alpha(\mathcal{P}_K^\alpha)=\sum_{k=1}^K I_\alpha(\mathcal{C}_k^\alpha).
\end{equation}
\smallskip
\paragraph{Spirit of the Ward-like hierarchical clustering.}
As previously, in order to obtain a new partition $\mathcal{P}_{K}^\alpha$ in $K$ clusters from a given partition $\mathcal{P}_{K+1}^\alpha$ in $K+1$ clusters,
the idea is to aggregate the two clusters $\mathcal{A}$ and $\mathcal{B}$ of $\mathcal{P}_{K+1}$ such that the new partition has minimum {\bf mixed} within-cluster inertia.
The optimization problem can now be expressed as follows:
\begin{equation}\label{opti3}
\arg\min_{ \mathcal{A},\mathcal{B}\in \mathcal{P}_{K+1}^\alpha} I_\alpha(\mathcal{A} \cup \mathcal{B})-I_\alpha(\mathcal{A})-I_\alpha(\mathcal{B}).
\end{equation}
\paragraph{ Ward-like hierarchical clustering process.}
\begin{itemize}
\item Step $K=n$: initialization.
The dissimilarities can be re-scaled between 0 and 1 to obtain the same order of magnitude: for instance,
$$D_0 \leftarrow \frac{D_0}{\max(D_0)} ~~~\mbox{and}~~~
D_1 \leftarrow \frac{D_1}{\max(D_1)}.$$
Note that this normalization step can also be done in a different way.
The initial partition $\mathcal{P}_{n}^\alpha=:\mathcal{P}_n$ in $n$ clusters (i.e. each cluster only contains an observation) is unique and thus does not depend on $\alpha$.
\smallskip
\item Step $K=n-1,\dots,2$: obtaining the partition in $K$ clusters from the partition in $K+1$ clusters.
At each step $K$, the algorithm aggregates the two clusters $\mathcal{A}$ and $\mathcal{B}$ of $\mathcal{P}_{K+1}^\alpha$
according to the optimization problem (\ref{opti3})
such that the increase of the {\bf mixed} within-cluster inertia is minimum for the selected partition over the other ones in $K$ clusters.
More precisely, at step $K$, the algorithm aggregates the two clusters $\mathcal{A}$ and $\mathcal{B}$ such that the corresponding aggregation measure
$$\delta_\alpha(\mathcal{A},\mathcal{B}):=W_\alpha(\mathcal{P}_{K+1}^\alpha)-W_\alpha(\mathcal{P}_{K}^\alpha)=I_\alpha(\mathcal{A} \cup \mathcal{B})-I_\alpha(\mathcal{A})-I_\alpha(\mathcal{B}) $$
is minimum.
\smallskip
\item Step $K=1$: stop. The partition $\mathcal{P}_1^\alpha=:\mathcal{P}_1$ in one cluster is obtained.
Note that this partition is unique and thus does not depend on $\alpha$.
\end{itemize}
\noindent
In the dendrogram of the corresponding hierarchy, the value (height) of a cluster $\mathcal{A} \cup \mathcal{B}$ is given by the agglomerative cluster criterion value
$\delta_\alpha(\mathcal{A},\mathcal{B}).$
\medskip
In practice, the Lance and Williams equation (\ref{LW-equation}) remains true in this context (where $\delta$ must be replaced by $\delta_\alpha$).
The aggregation measure between two singletons are written now:
$$\delta_\alpha(\{i\},\{j\})=(1-\alpha)\frac{w_i w_j}{w_i+w_j} d^2_{0,ij} + \alpha\frac{w_i w_j}{w_i+w_j} d^2_{1,ij}. $$
The Lance and Williams equation is then applied to the matrix
$$\Delta_\alpha=(1-\alpha) \Delta_0 + \alpha \Delta_1.$$
where $\Delta_0$ (resp. $\Delta_1$) is the $n \times n$ matrix of the values $\delta_{0,ij}=\frac{w_i w_j}{w_i+w_j} d^2_{0,ij}$ (resp. $\delta_{1,ij}=\frac{w_i w_j}{w_i+w_j} d^2_{1,ij}$).
\paragraph{Remarks.}
\begin{itemize}
\item The proposed procedure is different from applying directly the Ward algorithm to the ``dissimilarity'' matrix obtained via the convex combination $D_\alpha=(1-\alpha) D_0 + \alpha D_1$. The main benefit of the proposed procedure is that the mixing parameter $\alpha$ clearly controls the part of pseudo-inertia due to $D_0$ and $D_1$ in (\ref{heterogeneity-cluster}). This is not the case when applying directly the Ward algorithm to $D_\alpha$ since it is based on a unique pseudo-inertia.
\medskip
\item When $\alpha=0$ (resp. $\alpha=1$), the hierarchical clustering is only based on the dissimilarity matrix $D_0$ (resp. $D_1$). A procedure to determine a suitable value for the mixing parameter $\alpha$ is proposed hereafter, see Section~\ref{criterion-alpha}.
\end{itemize}
\subsection{A procedure to determine a suitable value for the mixing parameter $\alpha$}
\label{criterion-alpha}
The key point is the choice of a suitable value for the mixing parameter $\alpha \in [0,1]$. This parameter logically depends on the number of clusters $K$ and this logical dependence is an issue when it comes to decide an optimal value for the parameter $\alpha$. In this paper a practical (but not globally optimal) solution to this issue is proposed: conditioning on $K$ and choosing $\alpha$ that best compromises between loss of socio-economic and loss of geographical homogeneity. Of course other solutions than conditioning on $K$ could be explored (conditioning on $\alpha$ or defining a global criterion) but these solutions seem to be more difficult to implement in a sensible procedure.
To illustrate the idea of the proposed procedure, let us assume that the dissimilarity matrix $D_1$ contains geographical distances between $n$ municipalities, whereas the dissimilarity matrix $D_0$ contains distances based on a $n\times p_0$ data matrix $X_0$ of $p_0$ socio-economic variables measured on these $n$ municipalities. An objective of the user could be to determine a value of $\alpha$ which increases the geographical homogeneity of a partition in $K$ clusters without adversely affecting socio-economic homogeneity. These homogeneities can be measured using the appropriate pseudo within-cluster inertias.
Let $\beta\in [0,1]$. Let us introduce the notion of proportion of the total {\bf mixed} (pseudo) inertia explained by the partition $\mathcal{P}_{K}^\alpha$ in $K$ clusters:
$$Q_\beta(\mathcal{P}_K^\alpha)=1-\frac{W_\beta(\mathcal{P}_K^\alpha)}{W_\beta(\mathcal{P}_1)} \in [0,1].$$
\paragraph{Some comments on this criterion.}
\begin{itemize}
\item When $\beta=0$, the denominator $W_0(\mathcal{P}_1)$ is the total (pseudo) inertia based on the dissimilarity matrix $D_0$ and the numerator is the (pseudo) within-cluster inertia $W_0(\mathcal{P}_{K}^\alpha)$ based on the dissimilarity matrix $D_0$, i.e. only from the socio-economic point of view in our illustration.
The higher the value of the criterion $Q_0(\mathcal{P}_K^\alpha)$, the more homogeneous the partition $\mathcal{P}_{K}^\alpha$ is from the socio-economic point of view (i.e. each cluster $\mathcal{C}_k^\alpha,~k=1,\dots,K$ has a low inertia $I_0(\mathcal{C}_k^\alpha)$ which means that individuals within the cluster are similar).
When the considered partition $\mathcal{P}_K^\alpha$ has been obtained with $\alpha=0$, the criterion $Q_0(\mathcal{P}_K^\alpha)$ is obviously maximal (since the partition $\mathcal{P}_K^0$ was obtained by using only the dissimilarity matrix $D_0$), and this criterion will naturally tend to decrease as $\alpha$ increases from 0 to 1.
\medskip
\item Similarly, when $\beta=1$, the denominator $W_1(\mathcal{P}_1)$ is the total (pseudo) inertia based on the dissimilarity matrix $D_1$ and the numerator is the (pseudo) within-cluster inertia $W_1(\mathcal{P}_{K}^\alpha)$ based on the dissimilarity matrix $D_1$, i.e. only from a geographical point of view in our illustration.
Therefore, the higher the value of the criterion $Q_1(\mathcal{P}_K^\alpha)$, the more homogeneous the partition $\mathcal{P}_{K}^\alpha$ from a geographical point of view.
When the considered partition $\mathcal{P}_K^\alpha$ has been obtained with $\alpha=1$, the criterion $Q_1(\mathcal{P}_K^\alpha)$ is obviously maximal (since the partition $\mathcal{P}_K^1$ was obtained by using only the dissimilarity matrix $D_1$), and this criterion will naturally tend to decrease as $\alpha$ decreases from 1 to 0.
\medskip
\item For a value of $\beta\in]0,1[$, the denominator $W_\beta(\mathcal{P}_1)$ is a total {\bf mixed} (pseudo) inertia which can not be easily interpreted in practice, and the numerator $W_\beta(\mathcal{P}_K^\alpha)$ is the {\bf mixed} (pseudo) within-cluster inertia.
Note that when the considered partition $\mathcal{P}_K^\alpha$ has been obtained with $\alpha=\beta$, the criterion $Q_\beta(\mathcal{P}_K^\alpha)$ is obviously maximal by construction, and it will tend to decrease as $\alpha$ moves away from $\beta$.
\medskip
\item Finally, note that this criterion $Q_\beta(\mathcal{P}_K^\alpha)$ is decreasing in $K$. Moreover, $\forall \beta\in [0,1]$, it is easy to see that $Q_\beta(\mathcal{P}_n)=1$
and $Q_\beta(\mathcal{P}_1)=0$. The more clusters there are in a partition, the more homogeneous these clusters are (i.e. with a low inertia).
Therefore this criterion can not be used to select an appropriate number $K$ of clusters.
\end{itemize}
\paragraph{How to use this criterion to select the mixing parameter $\alpha$.}
Let us focus on the above mentioned case where the user is interested in determining a value of $\alpha$ which increases the geographical homogeneity of a partition in $K$ clusters without deteriorating too much the socio-economic homogeneity.
For a given number $K$ of clusters (the choice of $K$ is discussed later), the idea is the following:
\begin{itemize}
\item Let us consider a given grid of $J$ values for $\alpha \in [0,1]$: $$\mathcal{G}=\{\alpha_1=0,\alpha_1,\dots,\alpha_J=1\}.$$
For each value $\alpha_j\in\mathcal{G}$, the corresponding partition $\mathcal{P}_K^{\alpha_j}$ in $K$ clusters is obtained using the proposed Ward-like hierarchical clustering algorithm.
\medskip
\item For the $J$ partitions $\{\mathcal{P}_K^{\alpha_j},~j=1,\dots,J\}$, the criterion $Q_0(\mathcal{P}_K^{\alpha_j})$ is evaluated. The plot of the points $\{(\alpha_j, Q_0(\mathcal{P}_K^{\alpha_j})),~j=1,\dots,J\}$ provides a visual way to observe the loss of socio-economic homogeneity of the partition $\mathcal{P}_K^{\alpha_j}$ (from the ``pure'' socio-economic partition $\mathcal{P}_K^{0}$) as $\alpha_j$ increases from 0 to 1.
\medskip
\item Similarly, for the $J$ partitions $\{\mathcal{P}_K^{\alpha_j},~j=1,\dots,J\}$, the criterion $Q_1(\mathcal{P}_K^{\alpha_j})$ is evaluated. The plot of the points $\{(\alpha_j, Q_1(\mathcal{P}_K^{\alpha_j})),~j=1,\dots,J\}$ provides a visual way to observe the loss of geographical homogeneity of the partition $\mathcal{P}_K^{\alpha_j}$ (from the ``pure'' geographical partition $\mathcal{P}_K^{1}$) as $\alpha_j$ decreases from 1 to 0.
\medskip
\item These two plots (superimposed in the same figure) allow the user to choose a suitable value for $\alpha\in \mathcal{G}$ which is a trade-off between the loss of socio-economic homogeneity and greater geographical cohesion (when viewed through increasing values of $\alpha$) .
\end{itemize}
\paragraph{Case where the two total (pseudo) inertias $W_0(\mathcal{P}_1)$ and $W_1(\mathcal{P}_1)$ used in $Q_0(\mathcal{P}_K^{\alpha})$ and $Q_1(\mathcal{P}_K^{\alpha})$ are very different.}
Let us consider for instance that the dissimilarity matrix $D_1$ is a ``neighborhood'' dissimilarity matrix, constructed from the corresponding adjacency matrix $A$: that is $D_1=\mathbf{1}_n-A$ with $\mathbf{1}_{n,ij}=1$ for all $(i,j)$, $a_{ij}$ equal to 1 if observations $i$ and $j$ are neighbors and 0 otherwise, and $a_{ii}=1$ by convention.
With this kind of local dissimilarity matrix $D_1$, the geographical cohesion for few clusters is often small: indeed, $W_1(\mathcal{P}_1)$ could be very small and thus the criterion $Q_1(\mathcal{P}_K^{\alpha})$ takes values generally much smaller than those obtained by the $Q_0(\mathcal{P}_K^{\alpha})$.
Consequently, it is not easy for the user to select easily and properly a suitable value for the mixing parameter $\alpha$ since the two plots are in two very different scales.
One way to remedy this problem is to consider a renormalization of the two plots.
Rather than reasoning in terms of absolute values of the criterion $Q_0(\mathcal{P}_K^{\alpha})$ (resp. $Q_1(\mathcal{P}_K^{\alpha})$) which is maximal in $\alpha=0$ (resp. $\alpha=1$), we will renormalize $Q_0(\mathcal{P}_K^{\alpha})$ and $Q_1(\mathcal{P}_K^{\alpha})$ as follows: $Q_0^*(\mathcal{P}_K^{\alpha})=Q_0(\mathcal{P}_K^{\alpha})/Q_0(\mathcal{P}_K^{0})$ and $Q_1^*(\mathcal{P}_K^{\alpha})=Q_1(\mathcal{P}_K^{\alpha})/Q_1(\mathcal{P}_K^{1})$ and we then reason in terms of proportions of these criteria.
Therefore the corresponding plot $\{(\alpha_j, Q_0^*(\mathcal{P}_K^{\alpha_j})),~j=1,\dots,J\}$ (resp. $\{(\alpha_j, Q_1^*(\mathcal{P}_K^{\alpha_j})),~j=1,\dots,J\}$)
starts from 100\% and decreases as $\alpha_j$ increases from 0 to 1 (resp. as $\alpha_j$ decreases from 1 to 0).
\paragraph{The choice of the number $K$ of clusters.}
The proposed procedure to select a suitable value for the mixing parameter $\alpha$ works for a given number $K$ of clusters. Thus, it is first necessary to select $K$.
One way of achieving this is to focus on the dendrogram of the hierarchically-nested set of such partitions $\{\mathcal{P}_n^0=\mathcal{P}_n,\dots,\mathcal{P}_K^0,\dots,\mathcal{P}_1^0=\mathcal{P}_1\}$
only based on the dissimilarity matrix $D_0$ (i.e. for $\alpha=0$, that is considering only the socio-economic point of view in our application).
According to the dendrogram, the user can select an appropriate number $K$ of clusters with their favorite rule.
\subsection{Description of the functions of the package {\tt ClustGeo}}
The previous Ward-like hierarchical clustering procedure is implemented in the function {\tt hclustgeo} with the following arguments:
\begin{center}
{\tt hclustgeo(D0, D1 = NULL, alpha = 0, scale = TRUE, wt = NULL)}
\end{center}
where:
\begin{itemize}
\item {\tt D0} is the dissimilarity matrix $D_0$ between $n$ observations. It must be an object of class {\tt dist}, i.e. an object obtained with the function {\tt dist}.
The function {\tt as.dist} can be used to transform object of class {\tt matrix} to object of class {\tt dist}.
\item {\tt D1} is the dissimilarity matrix $D_1$ between the same $n$ observations. It must be an object of class {\tt dist}. By default {\tt D1=NULL} and the clustering is performed using {\tt D0} only.
\item {\tt alpha} must be a real value between 0 and 1. The mixing parameter $\alpha$ gives the relative importance of $D_0$ compared to $D_1$. By default, this parameter is equal to 0 and only $D_0$ is used in the clustering process.
\item {\tt scale} must be a logical value. If {\tt TRUE} (by default), the dissimilarity matrices $D_0$ and $D_1$ are scaled between 0 and 1 (that is divided by their maximum value).
\item {\tt wt} must be a $n$-dimensional vector of the weights of the observations. By default, {\tt wt=NULL} corresponds to the case where all observations are weighted by $1/n$.
\end{itemize}
The function {\tt hclustgeo} returns an object of class {\tt hclust}.
\medskip
\noindent
The procedure to determine a suitable value for the mixing parameter $\alpha$ is applied through the function {\tt choicealpha} with the following arguments:
\begin{center}
{\tt choicealpha(D0, D1, range.alpha, K, wt = NULL, scale = TRUE, graph = TRUE)}
\end{center}
where:
\begin{itemize}
\item {\tt D0} is the dissimilarity matrix $D_0$ of class {\tt dist}, already defined above.
\item {\tt D1} is the dissimilarity matrix $D_1$ of class {\tt dist}, already defined above.
\item {\tt range.alpha} is the vector of the real values $\alpha_j$ (between 0 and 1) considered by the user in the grid $\mathcal{G}$ of size $J$.
\item {\tt K } is the number of clusters chosen by the user.
\item {\tt wt} is the vector of the weights of the $n$ observations, already defined above.
\item {\tt scale} is a logical value that allows the user to rescale the dissimilarity matrices $D_0$ and $D_1$, already defined above.
\item {\tt graph} is a logical value. If {\tt graph=TRUE}, the two graphics (proportion and normalized proportion of explained inertia) are drawn.
\end{itemize}
This function returns an object of class {\tt choicealpha} which contains
\begin{itemize}
\item {\tt Q} is a $J\times 2$ real matrix such that the $j$th row contains $Q_0(\mathcal{P}_K^{\alpha_j})$ and $Q_1(\mathcal{P}_K^{\alpha_j})$.
\item {\tt Qnorm} is a $J\times 2$ real matrix such that the $j$th row contains $Q_0^*(\mathcal{P}_K^{\alpha_j})$ and $Q_1^*(\mathcal{P}_K^{\alpha_j})$..
\item {\tt range.alpha} is the vector of the real values $\alpha_j$ considered in the $\mathcal{G}$.
\end{itemize}
A {\tt plot} method is associated with the class {\tt choicealpha}.
\section{An illustration of hierarchical clustering with geographical constraints using the package {\tt ClustGeo}}
\label{sec:4}
This section illustrates the procedure of hierarchical clustering with geographical constraints on a real dataset using the package {\tt ClustGeo}. The complete procedure and methodology for the choice of the mixing parameter $\alpha$ is provided with two types of spatial constraints (with geographical distances and with neighborhood contiguity). We have provided the R code of this case study so that readers can reproduce our methodology and obtain map representations from their own data.
\subsection{The data} \label{sec:data}
Data were taken from French population censuses conducted by the National Institute of Statistics and Economic Studies (INSEE). The dataset is an extraction of $p = 4$ quantitative socio-economic variables for a subsample of $n = 303$ French municipalities located on the Atlantic coast between Royan and Mimizan:
\begin{itemize}
\item {\tt employ.rate.city} is the employment rate of the municipality, that is the ratio of the number of individuals who have a job to the population of working age (generally defined, for the purposes of international comparison, as persons of between 15 and 64 years of age).
\item {\tt graduate.rate} refers to the level of education of the population, i.e. the highest qualification declared by the individual. It is defined here as the ratio for the whole population having completed a diploma equal to or greater than two years of higher education (DUT, BTS, DEUG, nursing and social training courses, la licence, ma\^itrise, DEA, DESS, doctorate, or Grande Ecole diploma).
\item {\tt housing.appart} is the ratio of apartment housing.
\item {\tt agri.land} is the part of agricultural area of the municipality.
\end{itemize}
\noindent We consider here two dissimilarity matrices:
\begin{itemize}
\item $D_0$ is the Euclidean distance matrix between the $n$ municipalities performed with the $p=4$ available socio-economic variables,
\item $D_1$ is a second dissimilarity matrix used to take the geographical proximity between the $n$ municipalities into account.
\end{itemize}
{\small
\begin{verbatim}
> library(ClustGeo)
> data(estuary) # list of 3 objects (dat, D.geo, map)
# where dat= socio-economic data (n*p data frame),
# D.geo = n*n data frame of geographical distances,
# map = object of class "SpatialPolygonsDataFrame"
# used to draw the map
> head(estuary$dat)
employ.rate.city graduate.rate housing.appart agri.land
17015 28.08 17.68 5.15 90.04438
17021 30.42 13.13 4.93 58.51182
17030 25.42 16.28 0.00 95.18404
17034 35.08 9.06 0.00 91.01975
17050 28.23 17.13 2.51 61.71171
17052 22.02 12.66 3.22 61.90798
> D0 <- dist(estuary$dat) # the socio-economic distances
> D1 <- as.dist(estuary$D.geo) # the geographic distances between the municipalities
\end{verbatim}
}
\subsection{Choice of the number $K$ of clusters}
To choose the suitable number $K$ of clusters, we focus on the Ward dendrogram based on the $p=4$ socio-economic variables, that is using $D_0$ only.
{\small
\begin{verbatim}
> tree <- hclustgeo(D0)
> plot(tree,hang=-1, label=FALSE, xlab="", sub="", main="")
> rect.hclust(tree, k=5, border=c(4, 5, 3, 2, 1))
> legend("topright", legend=paste("cluster", 1:5), fill=1:5, bty="n", border="white")
\end{verbatim}
}
\begin{figure}
\centering
\includegraphics[scale=0.4]{dendroD0.pdf}
\vspace{-1cm}
\caption{Dendrogram of the $n=303$ municipalities based on the $p=4$ socio-economic variables (that is using $D_0$ only).}
\label{dendroD0}
\end{figure}
\noindent
The visual inspection of the dendrogram in Figure~\ref{dendroD0} suggests to retain $K=5$ clusters.
We can use the map provided in the {\tt estuary} data to visualize the corresponding partition in five clusters, called {\tt P5} hereafter.
{\small
\begin{verbatim}
> P5 <- cutree(tree, 5) # cut the dendrogram to get the partition in 5 clusters
> sp::plot(estuary$map, border="grey", col=P5) # plot an object of class sp
> legend("topleft", legend=paste("cluster", 1:5), fill=1:5, bty="n", border="white")
\end{verbatim}
}
\begin{figure}
\centering
\includegraphics[scale=0.4]{map-P5.pdf}
\vspace{-1cm}
\caption{Map of the partition {\tt P5} in $K=5$ clusters only based on the socio-economic variables (that is using $D_0$ only).}
\label{map-P5}
\end{figure}
\noindent
Figure~\ref{map-P5} shows that municipalities of cluster 5 are geographically compact, corresponding to Bordeaux and the 15 municipalities of its suburban area and Arcachon. On the contrary, municipalities in cluster 3 are scattered over a wider geographical area from North to South of the study area. The composition of each cluster is easily obtained, as shown for cluster 5:
{\small
\begin{verbatim}
# list of the municipalities in cluster 5
> city_label <- as.vector(estuary$map$"NOM_COMM")
> city_label[which(P5 == 5)]
[1] "ARCACHON" "BASSENS" "BEGLES"
[4] "BORDEAUX" "LE BOUSCAT" "BRUGES"
[7] "CARBON-BLANC" "CENON" "EYSINES"
[10] "FLOIRAC" "GRADIGNAN" "LE HAILLAN"
[13] "LORMONT" "MERIGNAC" "PESSAC"
[16] "TALENCE" "VILLENAVE-D'ORNON"
\end{verbatim}
}
The interpretation of the clusters according to the initial socio-economic variables is interesting. Figure \ref{comparisons} shows the boxplots of the variables for each cluster of the partition (left column). Cluster~5 corresponds to urban municipalities, Bordeaux and its outskirts plus Arcachon, with a relatively high graduate rate but low employment rate. Agricultural land is scarce and municipalities have a high proportion of apartments. Cluster~2 corresponds to suburban municipalities (north of Royan; north of Bordeaux close to the Gironde estuary) with mean levels of employment and graduates, a low proportion of apartments, more detached properties, and very high proportions of farmland. Cluster~4 corresponds to municipalities located in the Landes forest. Both the graduate rate and the ratio of the number of individuals in employment are high (greater than the mean value of the study area). The number of apartments is quite low and the agricultural areas are higher to the mean value of the zone. Cluster~1 corresponds to municipalities on the banks of the Gironde estuary. The proportion of farmland is higher than in the other clusters. On the contrary, the number of apartments is the lowest. However this cluster also has both the lowest employment rate and the lowest graduate rate. Cluster~3 is geographically sparse. It has the highest employment rate of the study area, a graduate rate similar to that of cluster~2, and a collective housing rate equivalent to that of cluster 4. The agricultural area is low.
\subsection{Obtaining a partition taking into account the geographical constraints}
To obtain more geographically compact clusters, we can now introduce the matrix $D_1$ of geographical distances into {\tt hclustgeo}.
This requires a mixing parameter to be selected $\alpha$ to improve the geographical cohesion of the 5 clusters without adversely affecting socio-economic cohesion.
\paragraph{Choice of the mixing parameter $\alpha$.}
The mixing parameter $\alpha \in [0,1]$ sets the importance of $D_0$ and $D_1$ in the clustering process. When $\alpha=0$ the geographical dissimilarities are not taken into account and when $\alpha=1$ it is the socio-economic distances which are not taken into account and the clusters are obtained with the geographical distances only.
\noindent
The idea is to perform separate calculations for socio-economic homogeneity and the geographic cohesion of the partitions obtained for a range of different values of $\alpha$ and a given number of clusters $K$.
\noindent
To achieve this, we can plot the quality criterion $Q_0$ and $Q_1$ of the partitions $P_K^\alpha$ obtained with different values of $\alpha \in [0,1]$ and choose the value of $\alpha$ which is a trade-off between the lost of socio-economic homogeneity and the gain of geographic cohesion. We use the function {\tt choicealpha} for this purpose.
{\small
\begin{verbatim}
> cr <- choicealpha(D0, D1, range.alpha=seq(0, 1, 0.1), K=5, graph=TRUE)
> cr$Q # proportion of explained pseudo-inertia
Q0 Q1
alpha=0 0.8134914 0.4033353
alpha=0.1 0.8123718 0.3586957
alpha=0.2 0.7558058 0.7206956
alpha=0.3 0.7603870 0.6802037
alpha=0.4 0.7062677 0.7860465
alpha=0.5 0.6588582 0.8431391
alpha=0.6 0.6726921 0.8377236
alpha=0.7 0.6729165 0.8371600
alpha=0.8 0.6100119 0.8514754
alpha=0.9 0.5938617 0.8572188
alpha=1 0.5016793 0.8726302
> cr$Qnorm # normalized proportion of explained pseudo-inertias
Q0norm Q1norm
alpha=0 1.0000000 0.4622065
alpha=0.1 0.9986237 0.4110512
alpha=0.2 0.9290889 0.8258889
alpha=0.3 0.9347203 0.7794868
alpha=0.4 0.8681932 0.9007785
alpha=0.5 0.8099142 0.9662043
alpha=0.6 0.8269197 0.9599984
alpha=0.7 0.8271956 0.9593526
alpha=0.8 0.7498689 0.9757574
alpha=0.9 0.7300160 0.9823391
alpha=1 0.6166990 1.0000000
\end{verbatim}
}
\begin{figure}[h]
\centering
\includegraphics[scale=0.4]{plotQ-1.pdf}
\includegraphics[scale=0.4]{plotQnorm-1.pdf}
\caption{Choice of $\alpha$ for a partition in $K=5$ clusters when $D1$ is the geographical distances between municipalities. Top: proportion of explained pseudo-inertias $Q_0(\mathcal{P}_K^{\alpha})$ versus $\alpha$ (in black solid line) and $Q_1(\mathcal{P}_K^{\alpha})$ versus $\alpha$ (in dashed line). Bottom: normalized proportion of explained pseudo-inertias $Q_0^*(\mathcal{P}_K^{\alpha})$ versus $\alpha$ (in black solid line) and $Q_1^*(\mathcal{P}_K^{\alpha})$ versus $\alpha$ (in dashed line).}
\label{plotQ-1}
\end{figure}
\noindent
Figure~\ref{plotQ-1} gives the plot of the proportion of explained pseudo-inertia calculated with $D_0$ (the socio-economic distances) which is equal to 0.81 when $\alpha=0$ and decreases when $\alpha$ increases (black solid line). On the contrary, the proportion of explained pseudo-inertia calculated with $D_1$ (the geographical distances) is equal to 0.87 when $\alpha=1$ and decreases when $\alpha$ decreases (dashed line).
\noindent
Here, the plot would appear to suggest choosing $\alpha=0.2$ which corresponds to a loss of only 7\% of socio-economic homogeneity, and a 17\% increase in geographical homogeneity.
\paragraph{Final partition obtained with $\alpha=0.2$.}
We perform {\tt hclustgeo} with $D_0$ and $D_1$ and $\alpha=0.2$ and cut the tree to get the new partition in five clusters, called {\tt P5bis} hereafter.
{\small
\begin{verbatim}
> tree <- hclustgeo(D0, D1, alpha=0.2)
> P5bis <- cutree(tree, 5)
> sp::plot(estuary$map, border="grey", col=P5bis)
> legend("topleft", legend=paste("cluster", 1:5), fill=1:5, bty="n", border="white")
\end{verbatim}
}
\noindent
The increased geographical cohesion of this partition {\tt P5bis} can be seen in Figure~\ref{map-P5bis}. Figure~\ref{comparisons} shows the boxplots of the variables for each cluster of the partition {\tt P5bis} (middle column). Cluster~5 of {\tt P5bis} is identical to cluster~5 of {\tt P5} with the Blaye municipality added in. Cluster~1 keeps the same interpretation as in {\tt P5} but has gained spatial homogeneity. It is now clearly located on the banks of the Gironde estuary, especially on the north bank. The same applies for cluster~2 especially for municipalities between Bordeaux and the estuary. Both clusters 3 and 4 have changed significantly. Cluster~3 is now a spatially compact zone, located predominantly in the M\'edoc.
It would appear that these two clusters have been separated based on proportion of farmland, because the municipalities in cluster~3 have above-average proportions of this type of land, while cluster~4 has the lowest proportion of farmland of the whole partition. Cluster~4 is also different because of the increase in clarity both from a spatial and socio-economic point of view. In addition, it contains the southern half of the study area. The ranges of all variables are also lower in the corresponding boxplots.
\begin{figure}[h]
\centering
\includegraphics[scale=0.4]{map-P5bis.pdf}
\vspace{-1cm}
\caption{Map of the partition {\tt P5bis} in $K=5$ clusters based on the socio-economic distances $D_0$ and the geographical distances between the municipalities $D_1$ with $\alpha=0.2$.}
\label{map-P5bis}
\end{figure}
\subsection{Obtaining a partition taking into account the neighborhood constraints}
Let us construct a different type of matrix $D_1$ to take neighbouring municipalities into account when clustering the 303 municipalities.
Two regions with contiguous boundaries, that is sharing one or more boundary point, are considered as neighbors. Let us first build the adjacency matrix $A$.
{\small
\begin{verbatim}
> list.nb <- spdep::poly2nb(estuary$map,
row.names=rownames(estuary$dat)) #list of neighbors
\end{verbatim}
}
\noindent
It is possible to obtain the list of the neighbors of a specific city. For instance, the neighbors of Bordeaux (which is the 117th city in the R data table) is given by the script:
{\small
\begin{verbatim}
> city_label[list.nb[[117]]] # list of the neighbors of BORDEAUX
[1] "BASSENS" "BEGLES" "BLANQUEFORT" "LE BOUSCAT" "BRUGES"
[6] "CENON" "EYSINES" "FLOIRAC" "LORMONT" "MERIGNAC"
[11] "PESSAC" "TALENCE"
\end{verbatim}
}
\noindent
The dissimilarity matrix $D_1$ is constructed based on the adjacency matrix $A$ with $D_1=\mathbf{1}_n-A$.
{\small
\begin{verbatim}
> A <- spdep::nb2mat(list.nb, style="B") # build the adjacency matrix
> diag(A) <- 1
> colnames(A) <- rownames(A) <- city_label
> D1 <- 1-A
> D1[1:2, 1:5]
ARCES ARVERT BALANZAC BARZAN BOIS
ARCES 0 1 1 0 1
ARVERT 1 0 1 1 1
> D1 <- as.dist(D1)
\end{verbatim}
}
\paragraph{Choice of the mixing parameter $\alpha$.}
The same procedure for the choice of $\alpha$ is then used with this neighborhood dissimilarity matrix $D_1$.
{\small
\begin{verbatim}
> cr <- choicealpha(D0, D1, range.alpha=seq(0, 1, 0.1), K=5, graph=TRUE)
> cr$Q # proportion of explained pseudo-inertia
> cr$Qnorm # normalized proportion of explained pseudo-inertia
\end{verbatim}
}
\begin{figure}[h!]
\centering
\includegraphics[scale=0.4]{plotQ-2.pdf}
\includegraphics[scale=0.4]{plotQnorm-2.pdf}
\caption{Choice of $\alpha$ for a partition in $K=5$ clusters when $D1$ is the neighborhood dissimilarity matrix between municipalities. Top: proportion of explained pseudo-inertias $Q_0(\mathcal{P}_K^{\alpha})$ versus $\alpha$ (in black solid line) and $Q_1(\mathcal{P}_K^{\alpha})$ versus $\alpha$ (in dashed line). Bottom: normalized proportion of explained pseudo-inertias $Q_0^*(\mathcal{P}_K^{\alpha})$ versus $\alpha$ (in black solid line) and $Q_1^*(\mathcal{P}_K^{\alpha})$ versus $\alpha$ (in dashed line).}
\label{plotQ-2}
\end{figure}
\noindent
With these kinds of local dissimilarities in $D_1$, the neighborhood within-cluster cohesion is always very small. $Q_1(\mathcal{P}_K^{\alpha})$ takes small values: see the dashed line of $Q_1(\mathcal{P}_K^{\alpha})$ versus $\alpha$ at the top of Figure~\ref{plotQ-2}. To overcome this problem, the user can plot the normalized proportion of explained inertias (that is $Q_0^*(\mathcal{P}_K^{\alpha})$ and $Q_1^*(\mathcal{P}_K^{\alpha})$) instead of the proportion of explained inertias (that is $Q_0(\mathcal{P}_K^{\alpha})$ and $Q_1(\mathcal{P}_K^{\alpha})$). At the bottom of Figure~\ref{plotQ-2}, the plot of the normalized proportion of explained inertias suggests to retain $\alpha=0.2$ or $0.3$. The value $\alpha=0.2$ slightly favors the socio-economic homogeneity versus the geographical homogeneity. According to the priority given in this application to the socio-economic aspects, the final partition is obtained with $\alpha=0.2$.
\paragraph{Final partition obtained with $\alpha=0.2$.}
It remains only to determine this final partition for $K=5$ clusters and $\alpha = 0.2$, called {\tt P5ter} hereafter. The corresponding map is given in Figure~\ref{map-P5ter}.
{\small
\begin{verbatim}
> tree <- hclustgeo(D0, D1, alpha=0.2)
> P5ter <- cutree(tree, 5)
> sp::plot(estuary$map, border="grey", col=P5ter)
> legend("topleft", legend=paste("cluster", 1:5), fill=1:5, bty="n", border="white")
\end{verbatim}
}
\begin{figure}[h]
\centering
\includegraphics[scale=0.4]{map-P5ter.pdf}
\vspace{-1cm}
\caption{Map of the partition {\tt P5ter} in $K=5$ clusters based on the socio-economic distances $D_0$ and the ``neighborhood'' distances of the municipalities $D_1$ with $\alpha=0.2$.}
\label{map-P5ter}
\end{figure}
\noindent
Figure \ref{map-P5ter} shows that clusters of {\tt P5ter} are spatially more compact than that of {\tt P5bis}. This is not surprising since this approach builds dissimilarities from the adjacency matrix which gives more importance to neighborhoods. However since our approach is based on soft contiguity constraints, municipalities that are not neighbors are allowed to be in the same clusters. This is the case for instance for cluster~4 where some municipalities are located in the north of the estuary whereas most are located in the southern area (corresponding to forest areas). The quality of the partition {\tt P5ter} is slightly worse than that of partition {\tt P5ter} according to criterion $Q_0$ (72.69\% versus 75.58\%). However the boxplots corresponding to partition {\tt P5ter} given in Figure \ref{comparisons} (right column) are very similar to those of partition {\tt P5bis}. These two partitions have thus very close interpretations.
\begin{figure}
\centering
\includegraphics[scale=0.35]{comparisons.pdf}
\caption{Comparison of the final partitions {\tt P5}, {\tt P5bis} and {\tt P5ter} in terms of variables x1=employ.rate.city, x2=graduate.rate, x3=housing.appart and x4=agri.land.}
\label{comparisons}
\end{figure}
\section{Concluding remarks}
\label{sec:5}
In this paper, a Ward-like hierarchical clustering algorithm including soft spacial constraints has been introduced and illustrated on a real dataset.
The corresponding approach has been implemented in the R package {\tt ClustGeo} available on the CRAN. When the observations correspond to geographical units (such as a city or a region), it is then possible to represent the clustering obtained on a map regarding the considered spatial constraints. This Ward-like hierarchical clustering method can also be used in many other contexts where the observations do not correspond to geographical units. In that case, the dissimilarity matrix $D_1$ associated with the ``constraint space'' does not correspond to spatial constraints in its current form.
For instance, the user may have at his/her disposal a first set of data of $p_0$ variables (e.g. socio-economic items) measured on $n$ individuals on which he/she has made a clustering from the associated dissimilarity (or distance) matrix. This user also has a second data set of $p_1$ new variables (e.g. environmental items) measured on these same $n$ individuals, on which a dissimilarity matrix $D_1$ can be calculated. Using the {\tt ClusGeo} approach, it is possible to take this new information into account to refine the initial clustering without basically disrupting it.
\section*{References}
{\small
\begin{itemize}
\item[] \hspace{-0.8cm} Ambroise C, Dang M, Govaert G (1997) Clustering of spatial data by the EM algorithm. In: Soares A, G\`omez-Hernandez J, Froidevaux R (eds) geoENV~I~-Geostatistics for Environmental Applications. Springer, pp. 493-504
\item[] \hspace{-0.8cm} Ambroise C, Govaert G (1998) Convergence of an EM-type algorithm for spatial clustering. Pattern Recognition Letters 19(10): 919-927
\item[] \hspace{-0.8cm} B\'ecue-Bertaut M, Alvarez-Esteban R, S\`anchez-Espigares JA (2017) Xplortext: Statistical Analysis of Textual Data R package. \url{https://cran.r-project.org/package=Xplortext}. R package version 1.0
\item[] \hspace{-0.8cm} B\'ecue-Bertaut M, Kostov B, Morin A , Naro G (2014) Rhetorical strategy in forensic speeches: multidimensional statistics-based methodology. Journal of Classication 31(1): 85-106
\item[] \hspace{-0.8cm} Bourgault G, Marcotte D, Legendre P (1992) The Multivariate (co) Variogram as a Spatial Weighting Function in Classification Methods. Mathematical Geology 24(5): 463-478
\item[] \hspace{-0.8cm} Chavent M, Kuentz-Simonet V, Labenne A, Saracco J (2017) ClustGeo: Hierarchical Clustering with Spatial Constraints. \url{https://cran.r-
project.org/package=ClustGeo}. R package version 2.0
\item[] \hspace{-0.8cm} Dehman A, Ambroise C, Neuvial P (2015) Performance of a blockwise approach in variable selection using linkage disequilibrium information. BMC Bioinformatics 16:148.
\item[] \hspace{-0.8cm} Duque JC, Dev B, Betancourt A, Franco JL (2011) ClusterPy: Library of spatially constrained clustering algorithms, RiSE-group (Research in Spatial Economics). EAFIT University. \url{http://www.rise-group.org/risem/clusterpy/}. Version 0.9.9.
\item[] \hspace{-0.8cm} Ferligoj A, Batagelj V (1982) Clustering with relational constraint. Psychometrika 47(4):413-426
\item[] \hspace{-0.8cm} Gordon AD (1996) A survey of constrained classication. Computational Statistics \& Data Analysis 21:17-29
\item[] \hspace{-0.8cm} Lance GN, Williams WT (1967) A General Theory of Classicatory Sorting Strategies 1. Hierarchical Systems. The Computer Journal 9:373-380
\item[] \hspace{-0.8cm} Legendre P (2014) const.clust: Space-and Time-Constrained Clustering Package. \url{http://adn.biol.umontreal.ca/ numericalecology/Rcode/}
\item[] \hspace{-0.8cm} Legendre P, Legendre L (2012) Numerical Ecology, vol. 24. Elsevier
\item[] \hspace{-0.8cm} Miele V, Picard F, Dray S (2014) Spatially constrained clustering of ecological networks. Methods in Ecology and Evolution 5(8):771-779
\item[] \hspace{-0.8cm} Murtagh F (1985a) Multidimensional clustering algorithms. Compstat Lectures, Vienna: Physika Verlag
\item[] \hspace{-0.8cm} Murtagh F (1985b) A Survey of Algorithms for Contiguity-constrained Clustering and Related Problems. The Computer Journal 28:82-88
\item[] \hspace{-0.8cm} Oliver M, Webster R (1989) A Geostatistical Basis for Spatial Weighting in Multivariate Classication. Mathematical Geology 21(1):15-35
\item[] \hspace{-0.8cm} Strauss T, von Maltitz MJ (2017) Generalising Ward's Method for Use with Manhattan Distances. PloS ONE 12(1). \url{https://doi.org/10.1371/journal.pone.0168288}
\item[] \hspace{-0.8cm} Vignes M, Forbes F (2009) Gene Clustering via Integrated Markov Models Combining Individual and Pairwise Features. IEEE/ACM Transactions on Computational Biology and Bioinformatics (TCBB) 6(2):260-270
\item[] \hspace{-0.8cm} Ward Jr JH (1963) Hierarchical Grouping to Optimize an Objective Function. Journal of the American Statistical Association 58(301):236-244
\end{itemize}
}
\end{document}
|
\section{Introduction}
Aesthetics is the study of science behind the concept and perception of beauty. Although aesthetics of photograph is subjective, some aspect of its depends on the standard photography practices and general visual design rules. With the ever increasing volume of digital photographs, automatic aesthetic assessment is becoming increasingly useful for various applications, such as a personal photo assistant, photo manager, photo enhancement, image retrieval etc. Conventionally, automatic aesthetic assessment tasks have been modeled as either a regression problem (single aesthetic score) ~\cite{kao2015visual, wu2011learning, kao2016hierarchical} or as a classification problem (aesthetically good or bad photograph) ~\cite{datta2006studying,lu2014rapid,lu2015deep}.
\begin{figure*}
\includegraphics[width=\textwidth]{sampleImage.pdf}
\caption{Sample images taken from AADB for each attribute. top row: Highest rated images, bottom row: Lowest Rated Images. All images were padded to maintain aspect ratio for illustration purposes.}
\label{fig:sample}
\end{figure*}
Intensive data driven approaches have made substantial progress in this task, although it is a very subjective and context dependent task. Earlier approaches used custom designed features based on photography rules (e.g., focus, color harmony, contrast, lighting, rule of thirds) and semantic information (e.g., human profile, scene category) from low level image descriptors (\textit{e.g.} color histograms, wavelet analysis) \cite{datta2006studying,ke2006design,dhar2011high,sun2009photo,luo2008photo,luo2011content,nishiyama2011aesthetic,lo2012statistic,bhattacharya2010framework} and generic image descriptors ~\cite{marchesotti2011assessing}. With the evolution of deep learning based techniques, recent approaches have introduced deep convolution neural networks (DCNN) in aesthetic assessment tasks ~\cite{lu2014rapid,lu2015deep,tian2015query,kao2015visual,kong2016photo}.
Although these approaches give near-human performance in classifying whether a photograph is ``good" or ``bad", they do not give detailed insights or explanation for such claims. For example, if a photograph received a bad rating, one would not get any insights about the attributes (e.g., poor lighting, dull colors etc.) that led to that rating. We propose an approach in which we identify (eight) such attributes (such as Color Harmony, Depth of Field etc.) and report those along with the overall score. For this purpose, we propose a multi-task deep convolution network (\emph{DCNN}) which simultaneously learns the eight aesthetic attributes along with the overall aesthetic score. We train and test our model on the recently released aesthetics and attribute database (\emph{AADB}) ~\cite{kong2016photo}. Following are the eight attributes as mentioned in \cite{kong2016photo} (Figure ~\ref{fig:sample}):
\begin{enumerate}
\item \emph{Balancing Element} - Whether the image contains balanced elements.
\item \emph{Content} - Whether the image has good/interesting content.
\item \emph{Color Harmony} - Whether the overall color composition is harmonious.
\item \emph{Depth of Field} - Whether the image has shallow depth of field.
\item \emph{Light} - Whether the image has good/interesting lighting.
\item \emph{Object Emphasis} - Whether the image emphasizes foreground objects.
\item \emph{Rule of Thirds} - Whether the image follows rule of thirds principle. The rule of thirds involves dividing the photo into 9 parts with 2 vertical and 2 horizontal lines. The important elements and leading lines are placed on/near these these lines and intersections of these lines.
\item \emph{Vivid Color} - Whether the image has vivid colors, not necessarily harmonious colors.
\end{enumerate}
We also develop attribute activation maps (Figure ~\ref{fig:visualization}) for visualization of these attributes. These maps highlight the salient regions for the corresponding attribute, thus providing us insights about the representation of these attributes in our trained model.
\newline
In summary, followings are the main contributions of our paper:
\begin{enumerate}
\item We propose a novel deep learning based approach which simultaneously learns eight aesthetic attributes along with the overall score. These attributes enable us to provide more detailed feedback on automatic aesthetic assessment.
\item We also develop localized representation of these attributes from our learned model. We call these \emph{attribute activation maps} (Figure ~\ref{fig:visualization}). These maps provide us more insights about model's interpretability of the attributes.
\end{enumerate}
\section{Related Work}
Most of the earlier works have used low-level image features to design high level aesthetic attributes as mid-level features and trained aesthetic classifier over these features. Datta \textit{et al.} \cite{datta2006studying} proposed 56 visual features based on standard photography and visual design rules to encapsulate aesthetic attributes from low-level image features. Dhar \textit{et al.} divided aesthetic attributes into three categories Compositional (\textit{e.g.} Depth of field, Rule of thirds), Content(\textit{e.g.} faces, animals, scene types), Sky-Illumination (\textit{e.g.} clear sky, sunset sky). They trained individual classifiers for these attributes from low-level features (\textit{e.g.} color histograms, center surrounding wavelets, haar features) and used outputs of these classifiers as input features for the aesthetic classifier.
Marchesotti \textit{et al.} \cite{marchesotti2015discovering}, proposed to learn aesthetic attributes from textual comments on the photographs using generic image features. Despite increased performance, many of these textual attributes (good, looks great, nice try) do not map to well-defined visual characteristics. Lu \textit{et al.} ~\cite{lu2014rapid} proposed to learn several meaningful style attributes, and used these to fine-tune the training of aesthetics classification network. Kong \textit{et al.} \cite{kong2016photo} proposed attribute and content adaptive DCNN for aesthetic score prediction.
However, none of the previous works report the aesthetic attributes themselves. These attributes are used as features to predict the overall aesthetic score/class. In this paper, we learn aesthetic attributes along with the overall score, not just as intermediate features but as auxiliary information. Aesthetic assessment is relatively easier in images with evident high and low aesthetics than in ordinary images with marginal aesthetics (Figure ~\ref{fig:good_bad_ugly}). For these images, attributes information would greatly supplement the quality of feedback from an automatic aesthetic assessment system.
\begin{figure}
\includegraphics[width=\linewidth]{good_bad_ugly.pdf}
\caption{Sample images from AADB testing data. first column: Images rated high on aesthetic score, second Column: Images rated at mid-level, third Column: Images rated low.}
\label{fig:good_bad_ugly}
\end{figure}
Recently, deep learning techniques have shown significant performance gains in various computer vision tasks such as object classification, localization ~\cite{krizhevsky2012imagenet, Simonyan14c,szegedy2015going}. In deep learning, non-linear features are learned in a hierarchical fashion in increasing complexity
\small(e.g. colors, edges, objects\small). The aesthetic attributes can be learned as combinations of these features. Deep learning techniques have shown significant performance gains in comparison with traditional machine learning approaches for aesthetic assessment tasks ~\cite{lu2014rapid, lu2015deep, tian2015query, kao2015visual, kong2016photo}. Unlike traditional machine learning techniques, features are also learned during training in deep learning techniques. However these internal representations of DCNNs are still opaque. Various visualization techniques ~\cite{zeiler2014visualizing, zhou2014object, mahendran2015understanding, dosovitskiy2015inverting, zhou2016learning, selvaraju2016grad} have been proposed to visualize the internal representations of DCNNs in an attempt to have a better understanding of their working. However, these visualization techniques have not been applied in aesthetic assessment tasks. In this article, we apply the gradient based visualization technique proposed by Zhou \textit{et al.} ~\cite{zhou2016learning} to obtain attribute activation maps. These maps provide localized representation of these attributes. Additionally we also apply similar visualization technique \cite{selvaraju2016grad} to the model provided by Kong \textit{et al.} \cite{kong2016photo} to obtain similar maps for qualitative comparison of our results with the earlier approach.
\section{Method}
\subsection{Architecture}
We use the deep residual network (ResNet50)~\cite{he2016deep} to train all the
attributes along with the overall aesthetic score. ResNet50 has 50 layers which can be divided into 16 successive residual blocks. Each residual block contains 3 convolution layers followed by the batch normalization layer (Figure ~\ref{fig:visualization}). Each residual block is followed by a rectified linear activation layer (ReLU) \cite{nair2010rectified}. We take these rectified convolution maps from the ReLU output of all these 16 residual blocks, and pool features from each of these 16 blocks with a global average pooling (GAP) layer. GAP layer gives the spatial average of these rectified convolution maps. Then we concatenate all these pooled features and use this as a feature for a fully connected layer which produces the desired outputs (aesthetic attributes and the overall score) as shown in Figure ~\ref{fig:visualization}. We model the attribute and score prediction as a regression problem with mean squared error as loss function.
Due to this simple connectivity structure, we are able to identify the importance of image regions by projecting the weights of the output layer on to the rectified convolution maps, a technique we call \emph{attribute activation mapping}. This technique was first introduced by Zhou \textit{et al.}~\cite{zhou2016learning} to get class activation maps for different semantic classes in image classification task.
\subsection{Attribute Activation Mapping}
For a given image, let \textit{\emph{$f_k$\small($x, y$\small)}} represent the activation of unit \textit{$k$} in the rectified convolution map at spatial location \textit{\small(x, y\small)}. Then, for unit k, the result of performing global average pooling is $F$\textsuperscript{k} = $\sum\nolimits_{\textit{x,y}} \textit{\emph{$f_k$\small($x, y$\small)}}$. Thus, for a given attribute \textit{a}, the input to the regression layer, \textit{$R_a$}, is $\sum\nolimits_{\textit{x}} \textit{\emph{$w^a_k$\textit{$F_k$}}}$ where \textit{$w^a_k$} is the weight corresponding to attribute \textit{a} for unit k. Essentially, \textit{$w^a_k$} indicates the importance of \textit{$F_k$} for attribute \textit{a} as shown in Figure ~\ref{fig:visualization}.
We also synthesized similar attribute maps from the model proposed by Kong \textit{et al.} ~\cite{kong2016photo}. We did not have the final attribute and content adapted model from ~\cite{kong2016photo} due to patent rights but Kong \textit{et al.} shared the attribute adapted model with us. That model is based on alexnet architecture ~\cite{krizhevsky2012imagenet} consisting of fully connected layers along with convolution layers. In this architecture, outputs of convolution layers are separated from desired outputs by three stacked fully connected layers. The outputs from last FC layer are regression scores of attributes. In this architecture we compute weight of layer \emph{k} for attribute \emph{a} as summation of gradients (\emph{$g_k^a$}) of outputs with respect to $k^{th}$ convolution layer $w^a_k$ = $\sum\nolimits_{\textit{x,y}} \textit{\emph{$g_k^a$\small($x, y$\small)}}$.
This technique was first introduced by Selvaraju \textit{et al.}~\cite{selvaraju2016grad} to get class activation maps for different semantic classes and visual explanation (answers for questions).
\begin{figure}
\includegraphics[width=\linewidth]{visualization.pdf}
\caption{Our approach for generating attribute activation maps. The predicted score for a given attribute (object emphasis in the figure) is mapped back to the rectified convolution layers to generate the attribute activation maps. These maps highlight the attribute-specific discriminative regions as shown in the bottom section.}
\label{fig:visualization}
\end{figure}
\subsection{Implementation Details}
Out of $10000$ samples present in the AADB dataset, we have trained our model on $8500$ training samples. 500 and 1000 images have been set aside for validation and testing purposes, respectively. As the number of training samples \small($8500$\small) is not adequate for training of such a deep network \small(23,715,852 parameters\small) from scratch, we used a pre-trained ResNet50. It was trained on 1000-class Imagenet classification dataset ~\cite{deng2009imagenet} with approximately $1.2$ million images. We fixed the input image size to $299\times299$. We used \emph{horizontal flip} of the input images as a data augmentation technique. The last residual block gives convolution maps of size $10\times10$, so we reduce the sizes of the convolution maps from the previous Res-Blocks to the same size with appropriate sized average pooling. As ResNet50 has batch normalization layers, it is very sensitive to batch size. We fixed the batch size to 16 and trained it for 16 epochs. We report our model\textquotesingle s performance on test set ($1000$ images) provided in AADB. We have made our implementation publicly available~\footnote{\label{code_src}{\url{https://github.com/gautamMalu/Aesthetic_attributes_maps}}}.
\subsection{Dataset}
As mentioned earlier, we have used the aesthetics and attribute database (AADB) provided by Kong \textit{et al}.~\cite{kong2016photo}. AADB provides overall ratings for the photographs along with the ratings on the eleven aesthetic attributes as mentioned in ~\cite{kong2016photo} (Figure ~\ref{fig:sample}).
Users were asked to provide information about the effectiveness of these attributes on the overall aesthetic score. For example, if \emph{object emphasis} is positively contributing towards the overall aesthetics of a photograph, user will give a score of \emph{+1} for the attribute, if object is not emphasized adequately and this is contributing negatively towards the overall aesthetic score of the photograph, user will give a score of \emph{-1} for the attribute (See Fig~\ref{fig:interface}). The users also rated the overall aesthetic score on a scale of 1 to 5, with 5 being the most aesthetically pleasing score. Each image was rated by at least \emph{5} persons. The mean score was taken as the ground truth score for all attributes and the overall score.
\begin{figure}
\includegraphics[width=\linewidth]{attr_dist.pdf}
\caption{The distribution of all the eleven attributes in the training data of AADB. Most of the images are rated neutral for motion blur, repetition and symmetry.}
\label{fig:attrDist}
\end{figure}
If an attribute has enhanced the image quality, it was rated positively and if the attribute has degraded the image aesthetics it was rated negatively. The default zero (null) means the attribute does not affect the image aesthetics. For example, positive \emph{vivid color} means the vividness of the color presented in an image has a positive effect on the image aesthetics; while the negative \emph{vivid color} means the image has dull color composition. All the attributes except for \emph{Repetition} and \emph{Symmetry} are normalized to the range of [-1, 1] \emph{Repetition} and \emph{Symmetry} are normalized to the range of [0, 1], as negative values are not justifiable for these two attributes. The overall score is normalized to the range of [0, 1].Out of these eleven attributes, we omit \emph{Symmetry}, \emph{Repetition} and \emph{Motion blur} attributes from our experiment as most of the images rated null for these attributes (Figure ~\ref{fig:attrDist}). We model the other eight attributes along with the overall aesthetic score as a regression problem.
\begin{figure}[t]
\includegraphics[width=\linewidth]{interface.pdf}
\caption{Interface of data collection adopted by Kong \textit{et al.}\cite{kong2016photo}.}
\label{fig:interface}
\end{figure}
\section{Results \& Discussion}
To evaluate the aesthetic attribute scores predicted by our model, we report the Spearman\textquotesingle s ranking correlation coefficient \small($\rho$\small) between the estimated aesthetic attribute score and the corresponding ground truth score for the testing data. The ranking correlation coefficient \small($\rho$\small) evaluates the monotonic relationship between estimated scores and ground truth scores, hence there is no need of explicit calibration between them. The correlation coefficient lies in the range of [-1, 1], with greater values corresponding to higher correlation and vice-versa. For baseline comparison, we also train a model by fine tuning a pre-trained ResNet50 and label it as ResNet50-FT. Fine-tuning here refers to modifying the last layer of the pre-trained ResNet50 \cite{he2016deep} and training it for our aesthetic attribute prediction task.
Table ~\ref{tab:results} lists the performance on AADB using the two approaches. We also report the performance of the model shared by Kong \textit{et al.}\cite{kong2016photo}.
\begin {table
\caption{Spearman\textquotesingle s rank correlations for all the attributes. All correlation coefficients ($\rho$) are significant at $p<0.0001$. The coefficients marked with a * are best results for respective attributes.}
\label{tab:results}
\begin{center}
\begin{tabular} {|l|c|c|c|}
\hline
\emph{Attribute} & ResNet50-FT & Kong \textit{et al.}\cite{kong2016photo}& Our method\\
\hline
Balancing Elements & $0.184$ & \textbf{0.220*} & $0.186$ \\
\hline
Content & $0.572$ & $0.508$ & \textbf{0.584*} \\
\hline
Color Harmony & $0.452$ & $0.471$ & \textbf{0.475*} \\
\hline
Depth of Field & $0.450$ & $0.479$ & \textbf{0.495*} \\
\hline
Light & $0.379$ & \textbf{0.443*} & $0.399$ \\
\hline
Object Emphasis & $0.658$ & $0.602$ & \textbf{0.666*} \\
\hline
Rule of Thirds & $0.175$ & \textbf{0.225*} & $0.178$ \\
\hline
Vivid Colors & $0.661$ & $0.648$ & \textbf{0.681*} \\
\hline
Overall Aesthetic Score & $0.665$ & $0.654$\footnotemark$/0.678$ & \textbf{0.689*} \\
\hline
\end{tabular}
\end{center}
\end{table}
\footnotetext{The $\rho$ reported by Kong \textit{et al.} ~\cite{kong2016photo} for their final content and attribute adaptive model is $0.678$, here we are reporting the performance of the model shared by them.}
It should be noted that the spearman\textquotesingle s coefficient between the estimated overall aesthetic score and the corresponding ground truth reported by Kong \textit{et al}.\cite{kong2016photo} was 0.678. They did not report any metrics for the other aesthetic attributes. They used ranking loss along with mean squared error as loss functions. Their final approach was also content adaptive.
As can be seen from the results reported in Table~\ref{tab:results}, our model managed to outperform their approach in overall aesthetic score in-spite of only being trained with mean square error and without any content adaptive framework .
Our model significantly underperformed for \emph{Rule of Thirds} and \emph{Balancing elements} attributes. These attributes are location sensitive attributes. \emph{Rule of thirds} deals with positioning of the salient elements, \emph{Balancing Elements} deals with relative positioning of objects with each other and the frame. In our model, due to use of global average pooling \small(GAP\small) layers after activation layers we are losing location specificity. We selected GAP layer to reduce the number of parameters. The number of training samples \small($8500$\small) allows learning of only small parameter space. We also warp the input images to the fixed size input \small($299x299$\small), thus destroying the aspect ratio. These could be possible reasons for the under-performance of the model for these compositional and location sensitive attributes. Across all the attributes, our proposed method reports better results than ResNet50 fine-tuned model. Our model performs better than the model provided by Kong {et al.}~\cite{kong2016photo} for five-out-of-eight attributes.
Aspects of aesthetic judgments are very subjective in nature. To quantify this subjectivity. In AADB the ground-truth score is the mean score of ratings given by different individuals. To quantify the agreement between ratings, $\rho$ between each individual\textquotesingle s ratings and the ground-truth scores was calculated. The average of $\rho$ is reported in Table ~\ref{tab:human_results}. Our model actually outperforms the human consistently (as measured by $\rho$) averaged across all raters. However, when considering only the ``power raters'' who have annotated more images, human evaluators consistently outperform model\textquotesingle s results.
\begin{table}
\caption{Human performance on AADB. Our model actually outperforms the human consistently (as measured by $\rho$, last row) averaged across all raters (first row). However, when considering only the ``power raters'' who have annotated more images, human consistently outperform our model (second and third row).}
\label{tab:human_results}
\begin{center}
\begin{tabular} {|l|c|r|}
\hline
\emph{Number of Images rated} & Number of Raters & $\rho$ \\
\hline
$> 0$ & $195$& $0.6738$ \\
\hline
$> 100$ & $65$ & $0.7013$ \\
\hline
$> 200$ & $42$ & $0.7112$ \\
\hline
Our Approach & $-$ & $0.689$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure*}
\includegraphics[width=\textwidth]{Object_collage.pdf}
\caption{Object Emphasis activation maps. First row: Original Images \small(marked with ground truth score at the bottom right\small), second row: Activation Maps from Kong \textit{et al.}~\cite{kong2016photo} model~\small(marked with predicted score from their given model\small), third row: Activation Maps from our method~\small(marked with our predicted score \small). Color-bar indicates the color encoding of activation.}
\label{fig:object}
\end{figure*}
\begin{figure*}
\includegraphics[width=\textwidth]{Content_collage.pdf}
\caption{Content activation maps. First row: Original Images \small(marked with ground truth score at the bottom right\small), second row: Activation Maps from Kong \textit{et al.}~\cite{kong2016photo} model~\small(marked with predicted score from their given model\small), third row: Activation Maps from our method~\small(marked with our predicted score \small). Color-bar indicates the color encoding of activation.}
\label{fig:content}
\end{figure*}
\section{Visualization}
As mentioned above we generate attribute activation maps for different attributes, to get their localized representations. Here we omit the following attributes, namely, emph{balancing element} and the \emph{rule of thirds}, as our model\textquotesingle s performance is very low for these attributes as shown in Table~\ref{tab:results}. For each attribute, we have analyzed the activation maps and present the insights in this section. For illustration purposes, We have selected ten samples for each attribute. Out of these ten samples, first five are the highest rated by our model, and the next five are the lowest rated. We have selected these samples from test samples \small($1000$\small) and not from the train samples. We also have included the activation maps from model given by Kong \textit{et al.}~\cite{kong2016photo} \small(Kong's model\small). These activation maps highlight the most important regions for the given attributes. We define these activation maps as ``gaze" of the model.
\begin{figure*}
\includegraphics[width=\textwidth]{DoF_collage.pdf}
\caption{Depth of Field activation maps. First row: Original Images \small(marked with ground truth score at the bottom right\small), second row: Activation Maps from Kong \textit{et al.}~\cite{kong2016photo} model~\small(marked with predicted score from their given model\small), third row: Activation Maps from our method~\small(marked with our predicted score \small). Color-bar indicates the color encoding of activation.}
\label{fig:DoF}
\end{figure*}
\begin{figure*}
\includegraphics[width=\textwidth]{VividColor_collage.pdf}
\caption{Vivid Color activation maps. First row: Original Images \small(marked with ground truth score at the bottom right\small), second row: Activation Maps from Kong \textit{et al.}~\cite{kong2016photo} model~\small(marked with predicted score from their given model\small), third row: Activation Maps from our method~\small(marked with our predicted score \small). Color-bar indicates the color encoding of activation.}
\label{fig:VividColor}
\end{figure*}
\subsection{Object Emphasis}
By qualitative analysis of activation maps of object emphasis, it was observed that model gazes at the main object on the image. Even when the model predicts negative rating, i.e. object is not emphasized, the model searches for regions which contain objects Figure ~\ref{fig:object}. In comparison, activation maps from Kong's model are not always consistent as can be seen in the second row of activation maps in Figure~\ref{fig:object}. It showcases that our model has learned the object emphasis attribute as an attribute which is indeed related to objects.
\subsection{Content}
Interestingness of content is significantly subjective and is a context-dependent attribute. However, if a model is trained on this attribute, one would expect the model would have maximum activation at the content of the image while making this judgment. If there exists a well-defined object in an image, then that object is considered as the content of the image, for e.g., $2^{nd}$ and $3^{rd}$ columns of Figure ~\ref{fig:content}. Further, it can be observed in these columns that our proposed approach is better at identifying the content than Kong's model. Without the presence of explicit objects, the content of the image is difficult to localize, for e.g. $1^{st}$ and $5^{th}$ columns of Figure ~\ref{fig:content}. As shown in Figure~\ref{fig:content}, our model's activation maps are maximally active at the content of the image. In comparison activation maps from Kong's models are not consistent.
\subsection{Depth of Field}
On analyzing the representations of shallow depth of field, it was observed that model looks for blurry regions near the main object of the image while making the judgment as showcased in Figure ~\ref{fig:DoF}. Shallow depth of field technique is used to make the subject of the photograph stand out from its background. The model's interpretation of it is in that direction. The images for which model has predicted the negative score on this attribute, the activation maps are random. Activation maps from Kong's model also showcase a similar behavior, these maps are more active at the corner of the images.
\subsection{Vivid Color}
Vivid Color means the presence of bright and bold colors. The model's interpretation of this attribute seems to be along these lines. As shown in Figure ~\ref{fig:VividColor}, model gazes at vivid color areas while making the judgment about this attribute. For example, in $2^{nd}$ column of the Figure ~\ref{fig:VividColor} pink color of flowers and scarf, and in $3^{rd}$ column butterfly and flower were the most activated regions.
Authors couldn't find any pattern in activation maps from Kong's model.
\subsection{Light}
Good Lighting is quite a challenging concept to grasp. It does not merely depend on the light in the photograph, but rather how that light complements the whole composition. As shown in Figure ~\ref{fig:Light}, most of the time model seems to look at bright light, or source of the light in the photograph. Although model\textquotesingle s behavior is consistent, its understanding of this attribute is incomplete. This was also evident in the low correlation ratings of our proposed model for this attribute, as reported in Table~\ref{tab:results}.
\begin{figure*}
\includegraphics[width=\textwidth]{Light_collage.pdf}
\caption{Light activation maps. First row: Original Images \small(marked with ground truth score at the bottom right\small), second row: Activation Maps from Kong \textit{et al.}~\cite{kong2016photo} model~\small(marked with predicted score from their given model\small), third row: Activation Maps from our method~\small(marked with our predicted score \small). Color-bar indicates the color encoding of activation.}
\label{fig:Light}
\end{figure*}
\subsection{Color Harmony}
Although model\textquotesingle s performance is significant for this attribute, we could not find any consistent pattern in its activation maps. As color harmony is of many types, e.g., analogues, complementary, triadic; it is difficult to get a single representation pattern. For example, in the first example shown in Figure ~\ref{fig:ColorHarmony}, the green color of hills is in analogous harmony with blue color of water and sky; in the $3^{rd}$ example, brown sand color is in split complementary harmony with blue and green. The attribute activation maps for Color Harmony are shown in Figure ~\ref{fig:ColorHarmony}.
\begin{figure*}
\includegraphics[width=\textwidth]{ColorHarmony_collage.pdf}
\caption{Color Harmony activation maps. First row: Original Images \small(marked with ground truth score at the bottom right\small), second row: Activation Maps from Kong \textit{et al.}~\cite{kong2016photo} model~\small(marked with predicted score from their given model\small), third row: Activation Maps from our method~\small(marked with our predicted score \small). Color-bar indicates the color encoding of activation.}
\label{fig:ColorHarmony}
\end{figure*}
\section{Conclusion}
In this paper, we have proposed deep convolution neural network (DCNN) architecture to learn aesthetic attributes. Results show that estimated scores of five aesthetic attributes (Interestingness of Content, Object emphasis, shallow Depth of Field, Vivid Color, and Color Harmony) correlate significantly with their respective ground truth scores. Whereas in the case of attributes such as Balancing Elements, Light and Rule of Thirds, the correlation is inferior. The activation maps corresponding to the learned aesthetic attributes such as object emphasis, content, depth of field and vivid color indicate that the model has acquired internal representation suitable to highlight these attributes automatically. However, for color harmony and light, the visualization maps were not consistent.
Aesthetic judgment involves a degree of subjectivity. For example, in AADB the average correlation between the mean score and an individual's score for the overall aesthetic score is 0.67 ~\ref{tab:human_results}. Moreover, as reported by Kong \textit{et al}.~\cite{kong2016photo}, the model learned on a particular dataset might not work on a different dataset. Considering all these factors, empirical validity of aesthetic judgment models is still a challenge. We suggest that the visualization techniques presented in the current work is a step forward in that direction. Empirical validation could proceed by asking subjects to annotate the images (identifying the regions that correspond to different aesthetic attributes) and these empirical maps could in turn be compared with the predicted maps of the model. Such experiments need to be conducted in future to validate the current approach.
\begin{acks}
The authors would like to thank Ms. Shruti Naik, and Mr. Yashaswi Verma of International Institute of Information Technology, Hyderabad, India for their help in manuscript preparation.
\end{acks}
|
\section{Introduction}
In \cite{schiffmann2014indecomposable} Schiffmann computed the number of absolutely indecomposable vector bundles of rank $r$ and degree $d$ over a compete curve $C$ of genus $g$ over $\mathbb F_q$. Suppose the eigenvalues of the Frobenius acting on the first cohomology of $C$ are
$\alpha_1,\ldots,\alpha_{2g}$ with $\alpha_{i+g}=q\alpha_i^{-1}$ for $i=1,\ldots,g$. This means that for all $k\geq 1$ we have
\[
\# C(\mathbb F_{q^k}) = 1+q^k-\sum_{i=1}^{2g} \alpha_i^k.
\]
Schiffmann's result says that the number of absolutely indecomposable vector bundles of rank $r$ and degree $d$ on $C$ is given by a Laurent polynomial independent of $C$
\[
A_{g,r,d}(q,\alpha_1,\ldots,\alpha_g) \in \mathbb Z[q,\alpha_1^{\pm 1}, \ldots,\alpha_g^{\pm 1}],
\]
symmetric in $\alpha_i$ and invariant under $\alpha_i\to q\alpha_i^{-1}$.
Suppose $(r,d)=1$. Schiffmann showed that the number of stable Higgs bundles of rank $r$ and degree $d$ is given by $q^{1+(g-1)r^2} A_{g,r,d}$.
Let $C$ be a curve over $\mathbb C$. The moduli space of stable Higgs bundles $\mathcal M_{g,r,d}(C)$ is a quasi-projective variety and by a theorem of Katz (\cite{hausel2008mixed}) its $E$-polynomial defined as
\[
E_{g,r,d}(x,y) = \sum_{i,j,k} (-1)^k x^i y^j \dim \mathrm{Gr}_F^i \mathrm{Gr}^W_{i+j} H^k_c(\mathcal M_{g,r,d}(C), \mathbb C)
\]
is given by $(xy)^{1+(g-1)r^2} A_{g,r,d}(xy,x,\ldots,x)$. It is known (\cite{hausel2005mirror}) that this moduli space has pure cohomology.
In particular, the Poincar\'e polynomial
\[
P_{q,r,d}(q) = \sum_i (-1)^i q^{\frac{i}2} \dim H^i_c(\mathcal M_{g,r,d}(C))
\]
is the following specialization:
\[
P_{q,r,d}(q) = E_{g,r,d}(q^{\frac12},q^{\frac12}) = q^{1+(g-1)r^2} A_{g,r,d}(q,q^{\frac12},\ldots,q^{\frac12}).
\]
Since twisted character varieties are diffeomorphic to the moduli spaces of stable Higgs bundles (see \cite{hausel2008mixed}), their Poincar\'e polynomials coincide.
The formula of Schiffmann was difficult to work with. In particular, it was not clear that his formula is equivalent to a much simpler formula conjectured earlier by Hausel and Rodriguez-Villegas for Poincar\'e polynomials (\cite{hausel2008mixed}), and then extended by Mozgovoy for the polynomials $A_{g,r,d}$ (\cite{mozgovoy2012solutions}).
Here we study Schiffmann's formula from the combinatorial point of view and establish these conjectures. Our main result is:
\begin{thm}\label{thm: main intro}
Let $g\geq 1$. Let $\Omega_g$ denote the series
\[
\Omega_g = \sum_{\mu\in\mathcal P} T^{|\mu|} \prod_{\square\in\mu} \frac{\prod_{i=1}^g (z^{a(\square)+1} - \alpha_i q^{l(\square)}) (z^{a(\square)} - \alpha_i^{-1} q^{l(\square)+1})}{(z^{a(\square)+1} - q^{l(\square)}) (z^{a(\square)} - q^{l(\square)+1})},
\]
and let
\[
H_g = -(1-q)(1-z) \pLog \Omega_g,\qquad H_g = \sum_{n=1}^\infty {H_{g,r} T^r}.
\]
Then for all $r\geq 1$ $H_{g,r}$ is a Laurent polynomial in $q$, $z$ and $\alpha_1,\ldots,\alpha_g$, and for all $d$ $A_{g,r,d}$ is obtained by setting $z=1$ in $H_{g,r}$:
\[
A_{g,r,d}(q,\alpha_1,\ldots,\alpha_g) = H_{g,r}(q,1,\alpha_1,\ldots,\alpha_g).
\]
\end{thm}
As a corollary we obtain the $GL$-version of the conjecture of Hausel and Thaddeus (see Conjecture 3.2 in \cite{hausel2005mirror}):
\begin{cor}
For $r,d,d'$ satisfying $(r,d)=(r,d')=1$ the $E$-polynomials of $\mathcal M(g,r,d)$ and $\mathcal M(g,r,d')$ coincide.
\end{cor}
Davesh Maulik and Aaron Pixton announced an independent proof of Theorem \ref{thm: main intro}. Their approach is to make rigorous the physical considerations of \cite{chuang2015parabolic}. They claim that their work will settle the more general conjectures about Higgs bundles with parabolic structures. On the other hand, it would be interesting to extend Schiffmann's (\cite{schiffmann2014indecomposable}) and Schiffmann-Mozgovoy's (\cite{mozgovoy2014counting}) methods to the parabolic case and thus obtain another proof.
\section{Arms and legs}
We begin by stating an elementary formula which relates the generating series of arms and legs and the generating series of weights of partitions, proved in \cite{carlsson2012exts} (we follow notations from \cite{carlsson_vertex_2016}). For a partition $\lambda$ and any cell $\square$ we denote by $a_\lambda(\square)$ and $l_\lambda(\square)$ the arm and leg lengths of $\square$ with respect to $\lambda$. These numbers are non-negative when $\square\in\lambda$ and negative otherwise. For partitions $\mu, \nu$ define
\[
E_{\mu,\nu} = \sum_{\square\in\mu} q^{-a_\nu(\square)} t^{l_\mu(\square)+1} + \sum_{\square\in\nu} q^{a_\mu(\square)+1} t^{-l_\nu(\square)}.
\]
For any partition $\mu$ let
\[
B_\mu = \sum_{\square\in\mu} q^{c(\square)} t^{r(\square)},
\]
where $c(\square), r(\square)$ denote the column and row indices. For any $f$ let $f^*$ be obtained from $f$ by substitution $q\to q^{-1}$, $t\to t^{-1}$.
\begin{lem}\label{lem:munu}
For any partitions $\mu, \nu$ we have
\begin{equation}\label{eq:munu}
E_{\mu, \nu} = qt B_\mu + B_{\nu}^* - (q-1)(t-1) B_\mu B_\nu^*.
\end{equation}
\end{lem}
\begin{proof}
We prove by induction on the largest part $\mu_1$ of $\mu$ (defined to be $0$ if $\mu=\varnothing$). If $\mu=\varnothing$, we have $a_\mu(\square)=-1-c(\square)$. Therefore
\[
E_{\varnothing,\nu} = \sum_{\square\in\mu} q^{-c(\square)}t^{-l_\nu(\square)}.
\]
For each fixed value of $c(\square)$ the numbers $l_\nu(\square)$ go over the same range as the numbers $r(\square)$. Thus we obtain
\[
E_{\mu,\varnothing} = B_\nu^*.
\]
This establishes the case $\mu_1=0$.
For the induction step let $\mu'$ be obtained from $\mu$ by removing the first column, i.e. $\mu'=(\mu_1-1,\mu_2-1,\ldots)$. Splitting the sum according to whether $\square$ is in the first column we obtain
\[
\sum_{\square\in\mu} q^{-a_\nu(\square)} t^{l_\mu(\square)+1} =
q \sum_{\square\in\mu'} q^{-a_\nu(\square)} t^{l_\mu(\square)+1}
+ \sum_{i=1}^{l(\mu)} q^{1-\nu_i} t^{l(\mu)-i+1}.
\]
For any cell $\square$ we have
\[
a_\mu(\square) = \begin{cases} a_{\mu'}(\square)+1 & \text{if $r(\square)<l(\mu)$,}\\ -1-c(\square) & \text{otherwise.}\end{cases}
\]
This implies
\[
\sum_{\square\in\nu} q^{a_\mu(\square)+1} t^{-l_\nu(\square)} = q \sum_{\square\in\nu} q^{a_{\mu}'(\square)+1} t^{-l_\nu(\square)} + (1-q)\sum_{\square\in\nu\,:\, r(\square)\geq l(\mu)} q^{-c(\square)} t^{-l_\nu(\square)}.
\]
In the last sum for each fixed value of $c(\square)$ the numbers $l_\nu(\square)$ go over the same range as the numbers $r(\square)-l(\mu)$, so we have
\[
\sum_{\square\in\nu\,:\, r(\square)\geq l(\mu)} q^{-c(\square)} t^{-l_\nu(\square)} = \sum_{\square\in\nu\,:\, r(\square)\geq l(\mu)} q^{-c(\square)} t^{l(\mu)-r(\square)} = \sum_{i=l(\mu)+1}^\infty t^{l(\mu)-i+1} \frac{1-q^{-\nu_i}}{1-q^{-1}}.
\]
Putting things together we have
\[
E_{\mu,\nu} - q E_{\mu',\nu} = \sum_{i=1}^\infty t^{l(\mu)-i+1} (q^{1-\nu_i}-q) + q \sum_{i=1}^{l(\mu)} t^{l(\mu)-i+1}.
\]
The first sum reduces to
\[
\sum_{i=1}^\infty t^{l(\mu)-i+1} (q^{1-\nu_i}-q) = q t^{l(\mu)} (q^{-1}-1) B_\nu^* = (1-q) t^{l(\mu)} B_\nu^*.
\]
The second sum becomes
\[
q \sum_{i=1}^{l(\mu)} t^{l(\mu)-i+1} = qt \frac{t^{l(\mu)}-1}{t-1}.
\]
This implies
\[
E_{\mu,\nu} - q E_{\mu',\nu} = (1-q) t^{l(\mu)} B_\nu^* + qt \frac{t^{l(\mu)}-1}{t-1}
\]
On the other hand we have
\[
B_\mu - q B_{\mu'} = \sum_{i=1}^{l(\mu)} t^{i-1} = \frac{t^{l(\mu)}-1}{t-1}.
\]
Therefore if we denote the right hand side of \eqref{eq:munu} by $E'_{\mu,nu}$ we obtain
\[
E'_{\mu,\nu} - q E'_{\mu',\nu} = qt \frac{t^{l(\mu)}-1}{t-1} + (1-q) B_\nu^* - (q-1)(t-1)B_{\nu}^* \frac{t^{l(\mu)}-1}{t-1}
\]
\[
qt \frac{t^{l(\mu)}-1}{t-1} + (1-q)t^{l(\mu)} B_{\nu}^*.
\]
So $E_{\mu',\nu}=E'_{\mu',\nu}$ implies $E_{\mu,\nu}=E'_{\mu,\nu}$ and the induction step is established.
\end{proof}
For a partition $\mu$ we define $z_i(\mu)$ to match $z_i$ in \cite{schiffmann2014indecomposable}:
\[
z_i(\mu) = t^{-l(\mu)+i} q^{\mu_i}\qquad(i=1,2,\ldots,l(\mu)).
\]
Our notations match after the substitution $(q,z)\to (t,q)$. Note the following generating series identity:
\begin{equation}\label{eq:sumz}
\sum_{i=1}^{l(\mu)} z_i(\mu) = t^{-l(\mu)+1} \left((q-1)B_\mu + \frac{t^{l(\mu)}-1}{t-1}\right).
\end{equation}
What we will actually need is the following generating series:
\[
K_\mu := (1-t) \sum_{i<j} \frac{z_i(\mu)}{z_j(\mu)}.
\]
It can be obtained as follows. Note that the sum $K_\mu$ contains only terms with non-positive powers of $t$. So we can start with
\[
\tilde{K_\mu}:=(1-t) \sum_{i=1}^{l(\mu)} z_i(\mu) \sum_{i=1}^{l(\mu)} z_i(\mu)^{-1}=(1-t) \sum_{i,j=1}^{l(\mu)} \frac{z_i(\mu)}{z_j(\mu)},
\]
and take only non-positive powers of $t$. Let $L$ be the operator
\[
L(t^i q^j) = \begin{cases}t^i q^j & (i\leq 0)\\ 0 & (i>0)\end{cases}.
\]
Then
\[
K_\mu = L(\tilde{K_\mu}) - l(\mu).
\]
Note that we had to subtract $l(\mu)$ to cancel the contribution from the terms $i=j$ appearing in $\tilde K_\mu$. We can calculate $\tilde{K_\mu}$ using Lemma \ref{lem:munu} and \eqref{eq:sumz}:
\[
\tilde{K_\mu} = (1-t) \left((q-1)B_\mu + \frac{t^{l(\mu)}-1}{t-1}\right) \left((q^{-1}-1)B_\mu^* + \frac{t^{-l(\mu)}-1}{t^{-1}-1}\right)
\]
\[
=(q^{-1}-1)E_{\mu,\mu} - t^{l(\mu)} (q^{-1}-1) B_\mu^* + t^{1-l(\mu)}(q-1) B_\mu - (t^{l(\mu)}-1)\sum_{i=0}^{l(\mu)-1} t^{-i},
\]
from which it is clear that
\[
L(\tilde{K_\mu}) = (q^{-1}-1) L(E_{\mu,\mu}) + t^{1-l(\mu)} (q-1) B_\mu + \sum_{i=0}^{l(\mu)-1} t^{-i}.
\]
\[
= (q^{-1}-1) \sum_{\square\in\mu} q^{a_\mu(\square)+1} t^{-l_\mu(\square)} + \sum_{i=1}^{l(\mu)} z_i(\mu).
\]
The conclusion is the following
\begin{prop}
For any partition $\mu$ we have
\[
(1-t) \sum_{i<j} \frac{z_i(\mu)}{z_j(\mu)} = (q^{-1}-1) \sum_{\square\in\mu} q^{a_\mu(\square)+1} t^{-l_\mu(\square)} + \sum_{i=1}^{l(\mu)} (z_i(\mu)-1).
\]
\end{prop}
Converting additive generating functions to multiplicative with an extra variable $u$ we obtain
\begin{cor}\label{cor:product arms legs}
For any partition $\mu$ we have
\[
\prod_{i<j} \frac{1-t u \frac{z_i(\mu)}{z_j(\mu)}}{1-u \frac{z_i(\mu)}{z_j(\mu)}} = \prod_{\square\in\mu} \frac{1-u q^{a_\mu(\square)+1} t^{-l_\mu(\square)}}{1-u q^{a_\mu(\square)} t^{-l_\mu(\square)}} \prod_{i=1}^{l(\mu)} \frac{1-u}{1-u z_i(\mu)}.
\]
\end{cor}
Note that the left hand side contains ``non-symmetric'' ratios $\frac{z_i(\mu)}{z_j(\mu)}$ for $i<j$, while the right hand side contains ``simple terms'' $z_i(\mu)$ and $1$, ``correct arm-leg terms'' $q^{a_\mu(\square)+1} t^{-l_\mu(\square)}$ and ``incorrect arm-leg terms'' $q^{a_\mu(\square)} t^{-l_\mu(\square)}$. Our strategy is to trade incorrect arm-leg terms in Schiffmann's formula for non-symmetric ratios, which will complement or cancel other non-symmetric ratios so that the result contains only correct arm-leg terms and something symmetric.
\section{Schiffmann's terms}\label{sec:schiffmanns terms}
Let $X$ be a smooth projective curve over $\mathbb F_q$ of genus $g$ with zeta function
\[
\zeta_X(x) = \frac{\prod_{i=1}^{2g} (1-\alpha_i x)}{(1-x)(1-qx)}.
\]
Let us order $\alpha_i$ in such a way that $\alpha_{i+g}=\frac{q}{\alpha_i}$ holds. We will treat $\alpha_1, \alpha_2, \ldots, \alpha_g$ as formal variables and set $\alpha_{i+g}=\frac{q}{\alpha_i}$. An alternative way to think of the parameters $\alpha_i$ is to view them as the exponentials of the chern roots of the Hodge bundle on the moduli space of curves times $q^{\frac12}$. The expressions we will be writing will depend on $q,z,\alpha_1,\ldots,\alpha_g$. There is a correspondence between these variables and the variables from \cite{mellit2016integrality} given as follows:
\begin{equation}\label{eq:qzalpha to tqu}
q,z,\alpha_1,\ldots,\alpha_g \to t,q,u_1^{-1},\ldots,u_g^{-1}.
\end{equation}
The formula of Schiffmann (see \cite{schiffmann2014indecomposable}, \cite{mozgovoy2014counting}) involves a sum over partitions
\[
\Omega := \sum_{\mu} \Omega_\mu T^{|\mu|}.
\]
For each partition $\mu$ the corresponding coefficient is
\[
\Omega_\mu := q^{(g-1)\langle\mu,\mu\rangle} J_\mu H_\mu.
\]
Here $\langle\mu,\mu\rangle=\sum_i \mu_i'^2$ where $\mu'$ is the conjugate partition of $\mu$. We will proceed defining $J_\mu$ and $H_\mu$ and taking them apart in the process. We have
\[
J_\mu = \prod_{\square\in\mu} \frac{\prod_{i=1}^{2g} (1-\alpha_i q^{-1-l(\square)} z^{a(\square)})}{(1- q^{-1-l(\square)} z^{a(\square)})(1-q^{-l(\square)} z^{a(\square)})_{\neq0}}.
\]
The notation $(-)_{\neq0}$ means we omit the corresponding factor if it happens to be zero. This naturally splits as follows:
\[
J_\mu = \prod_{\square\in\mu} \frac{\prod_{i=1}^{g} (1-\alpha_i q^{-1-l(\square)} z^{a(\square)})}{1- q^{-1-l(\square)} z^{a(\square)}} \prod_{\square\in\mu} \frac{\prod_{i=1}^{g} (1-\alpha_i^{-1} q^{-l(\square)} z^{a(\square)})}{(1- q^{-l(\square)} z^{a(\square)})_{\neq0}}.
\]
Applying Corollary \ref{cor:product arms legs} we obtain
\[
J_\mu = \prod_{\square\in\mu} \frac{\prod_{i=1}^{g} (1-\alpha_i q^{-1-l(\square)} z^{a(\square)})}{1- q^{-1-l(\square)} z^{a(\square)}} \times \prod_{\square\in\mu} \frac{\prod_{i=1}^{g} (1-\alpha_i^{-1} q^{-l(\square)} z^{a(\square)+1})}{1- q^{-l(\square)} z^{a(\square)+1}}
\]
\[
\times \prod_{i<j} \prod_{k=1}^g \frac{1-\alpha_k^{-1} \frac{z_i(\mu)}{z_j(\mu)}}{1-q \alpha_k^{-1} \frac{z_i(\mu)}{z_j(\mu)}}
\times \prod_{i=1}^{l(\mu)} \prod_{k=1}^g \frac{1-\alpha_k^{-1}}{1-\alpha_k^{-1}z_i(\mu)}
\times \prod_{i<j} \frac{\left(1-q\frac{z_i(\mu)}{z_j(\mu)}\right)_{\neq0}}{1-\frac{z_i(\mu)}{z_j(\mu)}} \times \prod_{i=1}^{l(\mu)} (1-z_i(\mu)),
\]
where $z_i(\mu)=q^{-l(\mu)+i} z^{\mu_i}$ coincides with Schiffmann's $z_{n-i+1}$. Denote the four products above by $A,B,C,D$. Note that $\sum l(\square) + \sum (l(\square)+1) = \langle \mu,\mu\rangle$, so $q^{\langle\mu,\mu\rangle}$ together with the first two products produce
\[
\frac{\prod_{i=1}^g N_\mu(\alpha_i^{-1})}{N_\mu(1)},
\]
where $N_\mu$ is the arm-leg product as in \cite{mellit2016integrality}:
\begin{equation}\label{eq: Nmu definition}
N_\mu(u)=\prod_{\square\in\mu}(z^{a(\square)}-u q^{1+l(\square)})(z^{a(\square)+1}-u^{-1} q^{l(\square)}).
\end{equation}
So we have
\[
q^{\langle\mu,\mu\rangle} J_\mu = \frac{\prod_{i=1}^g N_\mu(\alpha_i^{-1})}{N_\mu(1)} ABCD,
\]
where
\[
A = \prod_{i<j} \prod_{k=1}^g \frac{1-\alpha_k^{-1} \frac{z_i(\mu)}{z_j(\mu)}}{1-q \alpha_k^{-1} \frac{z_i(\mu)}{z_j(\mu)}},\qquad
B=\prod_{i=1}^{l(\mu)} \prod_{k=1}^g \frac{1-\alpha_k^{-1}}{1-\alpha_k^{-1}z_i(\mu)}
\]
\[
C = \prod_{i<j} \frac{\left(1-q\frac{z_i(\mu)}{z_j(\mu)}\right)_{\neq0}}{1-\frac{z_i(\mu)}{z_j(\mu)}}
\qquad
D = \prod_{i=1}^{l(\mu)} (1-z_i(\mu)).
\]
We proceed by defining $H_\mu$. Let
\[
\tilde\zeta(x) = x^{1-g} \zeta(x) = \frac{\prod_{k=1}^{g} x^{-1} (1-\alpha_k x)(1-q \alpha_k^{-1} x)}{x^{-1}(1-x)(1-qx)}.
\]
Let $L(z_1,\ldots,z_{l(\mu)})$ be the rational function (note that we reversed the order of $z_i$)
\[
L(z_1,\ldots,z_{l(\mu)}) = \frac{1}{\prod_{i>j} \tilde\zeta\left(\frac{z_i}{z_j}\right)} \sum_{\sigma\in S_{l(\mu)}} \sigma \left\{\prod_{i<j} \tilde\zeta\left(\frac{z_i}{z_j}\right) \frac{1}{\prod_{i<l(\mu)} \left(1-q \frac{z_{i+1}}{z_{i}}\right)} \frac{1}{1-z_1}\right\}.
\]
Note that $\tilde\zeta$ appears in the numerator as many times as in the denominator, so it can be multiplied by a constant without changing $L$. So we replace $\tilde\zeta$ with something more resembling the other products we have seen:
\[
\tilde\zeta(x) = \frac{\prod_{k=1}^{g} (1-\alpha_k^{-1} x^{-1})(1-q \alpha_k^{-1} x)}{(1-x^{-1})(1-qx)}.
\]
$H_\mu$ is defined as the iterated residue (remember that our ordering of $z_i$ is the opposite of Schiffman's)
\[
H_\mu = \res_{z_i=z_i(\mu)} L(z_1,\ldots,z_{l(\mu)}) \prod_{i\,:\, \mu_i=\mu_{i+1}} \frac{d z_{i+1}}{z_{i+1}}.
\]
Note that the only poles $L$ can have at $z_i=z_i(\mu)$ are coming from factors of the form $1-q \frac{z_i}{z_{i+1}}$ for $i$ such that $\mu_i=\mu_{i+1}$. Each such factor can appear at most once in the denominator of $L$. We have
\[
\res_{z_{i+1}=q z_{i}} \frac{1}{1-q\frac{z_i}{z_{i+1}}} \frac{d z_{i+1}}{z_{i+1}} = 1.
\]
Thus we will obtain the same result if we multiply $L$ by the product of these factors and then evaluate at $z_i=z_i(\mu)$. Note that $C$ has precisely the same factors removed. Therefore we have
\[
C H_\mu = \left(\prod_{i<j} \frac{1-q\frac{z_i}{z_j}}{1-\frac{z_i}{z_j}} L\right)(z_1(\mu),\ldots,z_{l(\mu)}(\mu)).
\]
Putting in $A$ as well we obtain a nice expression:
\[
AC H_\mu = \left(\prod_{i\neq j} \frac{ 1-q\frac{z_i}{z_j}}{\prod_{k=1}^g 1-q \alpha_k^{-1} \frac{z_i}{z_j}} \sum_{\sigma\in S_{l(\mu)}} \sigma \left\{\cdots\right\} \right)(z_1(\mu),\ldots,z_{l(\mu)}(\mu)).
\]
We see that the product is symmetric in $z_i$, so it can be moved inside the summation. Since $B$ and $D$ are symmetric, they can also be moved inside the summation. After some cancellations we arrive at the following. Define for any $n$
\begin{multline}\label{eq:f definition}
f(z_1,\ldots,z_n) = \prod_i \prod_{k=1}^g \frac{1-\alpha_k^{-1}}{1- \alpha_k^{-1} z_i}
\\
\times \sum_{\sigma\in S_n}\sigma \left\{
\prod_{i>j}\left(\frac{1}{1-\frac{z_i}{z_j}} \prod_{k=1}^g \frac{1-\alpha_k^{-1} \frac{z_i}{z_j}}{1-q \alpha_k^{-1} \frac{z_i}{z_j}}\right) \prod_{i> j+1} (1-q \tfrac{z_i}{z_j}) \prod_{i\geq 2} (1-z_i)
\right\}.
\end{multline}
Then
\[
ABCD H_\mu = f(z_1(\mu),\ldots,z_{l(\mu)}(\mu)).
\]
Summarizing we obtain
\begin{prop}
For any partition $\mu$ the term $\Omega_\mu$ is given by
\[
\Omega_\mu = \frac{f_\mu \prod_{i=1}^g N_\mu(\alpha_i^{-1})}{N_\mu(1)},\qquad f_\mu=f(z_1(\mu),\ldots,z_{l(\mu)}(\mu)),
\]
where $z_i(\mu)=q^{-l(\mu)+i} z^{\mu_i}$, and $N$, $f$ are defined in \eqref{eq: Nmu definition}, \eqref{eq:f definition}.
\end{prop}
\begin{example}\label{ex:f small n}
Let us calculate $f$ in a few cases. It is convenient to set
\[
P(x) = \prod_{i=1}^g (1-\alpha_i^{-1} x).
\]
We have
\[
f(z_1) = \frac{P(1)}{P(z_1)}
\]
\[
f(z_1,z_2) = \frac{P(1)^2}{P(z_1)P(z_2)(z_1-z_2)}\left(z_1(1-z_2) \frac{P(\tfrac{z_2}{z_1})}{P(q\tfrac{z_2}{z_1})} - z_2(1-z_1) \frac{P(\tfrac{z_1}{z_2})}{P(q\tfrac{z_1}{z_2})}\right)
\]
Note that the denominator of this expression is $P(z_1) P(z_2) P(q\tfrac{z_1}{z_2})P(q\tfrac{z_2}{z_1})$ if no cancellations happen. If $z_2=q z_1$, the denominator reduces to $P(z_1) P(z_2) P(q^2)$, so it has only $3$ $P$-factors instead of $4$.
\end{example}
\section{Combinatorics of the function $f$}
\subsection{Bounding denominators}
First we analyse denominators of $f$ defined in \eqref{eq:f definition}. For generic values of $z_i$ the denominator of $f$ can be as bad as the full product
\[
\prod_{i} P(z_i) \prod_{i\neq j} P(q \tfrac{z_i}{z_j}),
\]
where $P(x)=\prod_{k=1}^g (1-\alpha_k^{-1} x)$. Pick numbers $r_1, r_2,\cdots$ such that $\sum_m r_m=n$. Split $z_1, z_2, \ldots, z_n$ into a union of subsequences of sizes $r_1$, $r_2$, \ldots. Let $j_m=1+\sum_{i<m} r_i$. For each $m$ the $m$-th subsequence looks like $z_{j_m}, z_{j_m+1}, \ldots, z_{j_m+r_m-1}$. Suppose each subsequence forms a geometric progression with quotient $q$:
\[
z_{j_m+i} = q^i z_{j_m}\qquad (i< r_m)
\]
Then $f$ can be viewed as a function of variables $z_{j_m}$. The denominator can be bounded as follows
\begin{prop}
The following expression is a Laurent polynomial:
\[
f \prod_{i=1}^n \left( P(z_i)\prod_{m:\, j_m+r_m> i} P(q^{r_m} \tfrac{z_{j_m}}{z_i}) \prod_{m:\, j_m> i} P(q \tfrac{z_i}{z_{j_m}})\right)
\]
\end{prop}
\begin{proof}
First write the definition of $f$ as follows:
\[
f = \prod_i \frac{P(1)}{P(z_i)} \sum_{\sigma\in S_n} \prod_{\sigma(i)>\sigma(j)}\frac{P(\frac{z_i}{z_j})}{(1-\frac{z_i}{z_j})P(q\frac{z_i}{z_j})}
\prod_{\sigma(i)> \sigma(j)+1} (1-q \tfrac{z_i}{z_j}) \prod_{\sigma(i)\geq 2} (1-z_i).
\]
Note that $1-\frac{z_i}{z_j}$ does not contribute to the denominator because of symmetrization. Next note that if $j=i+1$ and $j,i$ belong to the same subsequence, then $1-q\frac{z_i}{z_j}=0$. So all summands with $\sigma(i)>\sigma(j)+1$ vanish. So it is enough to sum only over those $\sigma$ which satisfy the condition
\begin{equation}\label{eq:sigma condition}
\sigma(i+1)\geq \sigma(i)-1 \quad \text{whenever $i, i+1$ are in the same subsequence.}
\end{equation}
So in each sequence $\sigma(j_m),\ldots,\sigma(j_m+r_m-1)$ if there is a drop, the size of the drop is $1$. Now for each such $\sigma$ we look at the product
\[
\prod_{\sigma(i)>\sigma(j)}\frac{P(\frac{z_i}{z_j})}{P(q\frac{z_i}{z_j})} = \prod_{i<j,\;\sigma(i)>\sigma(j)}\frac{P(\frac{z_i}{z_j})}{P(q\frac{z_i}{z_j})} \prod_{i<j,\;\sigma(i)<\sigma(j)}\frac{P(\frac{z_j}{z_i})}{P(q\frac{z_j}{z_i})}.
\]
It is enough to show that for each value of $i$ and each $\sigma$ the following expressions are Laurent polynomials:
\[
P(1) \prod_{m:\, j_m> i} P(q \tfrac{z_i}{z_{j_m}})\times \prod_{i<j,\;\sigma(i)>\sigma(j)}\frac{P(\frac{z_i}{z_j})}{P(q\frac{z_i}{z_j})},
\]
\[
\prod_{m:\, j_m+r_m> i} P(q^{r_m} \tfrac{z_{j_m}}{z_i}) \times \prod_{i<j,\;\sigma(i)<\sigma(j)}\frac{P(\frac{z_j}{z_i})}{P(q\frac{z_j}{z_i})}.
\]
Further, let us split the product over all $j>i$ into products over our subsequences. We only need to consider values of $m$ such that $j_m>i$ (when $j$ and $i$ are in different subsequences) or $j_m\leq i<j_m+r_m$ (when they are in the same subsequence). So it is enough to show that the following products are Laurent polynomials:
\begin{equation}\label{eq:case1}
P(q \tfrac{z_i}{z_{j_m}}) \prod_{k<r_m,\;\sigma(i)>\sigma(j_m+k)} \frac{P(\frac{z_i}{z_{j_m+k}})}{P(q\frac{z_i}{z_{j_m+k}})} \qquad(j_m>i),
\end{equation}
\begin{equation}\label{eq:case2}
P(q^{r_m} \tfrac{z_{j_m}}{z_i}) \prod_{k<r_m,\;\sigma(i)<\sigma(j_m+k)} \frac{P(\frac{z_{j_m+k}}{z_i})}{P(q\frac{z_{j_m+k}}{z_i})} \qquad(j_m>i),
\end{equation}
\begin{equation}\label{eq:case3}
P(1) \prod_{i-j_m<k<r_m,\;\sigma(i)>\sigma(j_m+k)} \frac{P(\frac{z_i}{z_{j_m+k}})}{P(q\frac{z_i}{z_{j_m+k}})} \qquad(j_m>i),
\end{equation}
\begin{equation}\label{eq:case4}
P(q^{r_m} \tfrac{z_{j_m}}{z_i}) \prod_{i-j_m<k<r_m,\;\sigma(i)<\sigma(j_m+k)} \frac{P(\frac{z_{j_m+k}}{z_i})}{P(q\frac{z_{j_m+k}}{z_i})} \qquad(j_m>i).
\end{equation}
Observe that because of the condition \eqref{eq:sigma condition} in each of the cases \eqref{eq:case1}---\eqref{eq:case4} the values of $k$ from a contiguous set $k_{min},\ldots,k_{max}$ (if non-empty). So the arguments to $P$ from a geometric progression with ratio $q$ or $q^{-1}$. Hence the product collapses and the only remaining denominator is $P(q\frac{z_i}{z_{j_m+k_{min}}})$ in cases \eqref{eq:case1} and \eqref{eq:case3}, and $P(q\frac{z_i}{z_{j_m+k_{max}}})$ in cases \eqref{eq:case2} and \eqref{eq:case4}. Further analysis leads to $k_{min}=0$ in \eqref{eq:case1}, $k_{max}=r_m-1$ in \eqref{eq:case2}, $k_{min}=i-j_m+1$ in \eqref{eq:case3} and $k_{max}=r_m-1$ in \eqref{eq:case4}.
\end{proof}
\begin{example}
In the situation of $n=1$ we obtain that $f P(z_1) P(q)$ is a Laurent polynomial. For $n=2$ and $z_2=q z_1$ we obtain $f P(z_1) P(z_2) P(q^2) P(q)$ is a Laurent polynomial. Comparing with Example \ref{ex:f small n} one can notice that our denominator bound is not optimal.
\end{example}
For the case when $z_i=z_i(\mu)=z^{\mu_i} q^{i-l(\mu)}$ for a partition $\mu$ we obtain
\begin{prop}
The following product is a Laurent polynomial for any partition $\mu$:
\[
f_\mu \prod_{\square\in\mu} P(z^{a(\square)+1} q^{-l(\square)}) P(z^{-a(\square)} q^{l(\square)+1}).
\]
\end{prop}
\begin{proof}
Recall that $f_\mu$ is a shorthand for $f(z_1(\mu),\ldots,z_{l(\mu)}(\mu))$ where $z_i(\mu) = z^{\mu_i} q^{i-l(\mu)}$.
It is enough to show that for each $i$ the product
\begin{equation}\label{eq:prod P}
P(z_i)\prod_{m:\, j_m+r_m> i} P(q^{r_m} \tfrac{z_{j_m}}{z_i}) \prod_{m:\, j_m> i} P(q \tfrac{z_i}{z_{j_m}})
\end{equation}
divides the corresponding arm-leg product over the cells of $\mu$ in the row $i$. Note that the our subsequences of geometric progressions in $z_i$ simply correspond to repeated parts of $\mu$. Let $\square$ be the cell in row $i$ and column $\mu_{j_m}$ ($j_m+r_m>i$). Then we have $a(\square)=\mu_i-\mu_{j_m}$, $l(\square)=j_m+r_m-1-i$. Therefore
\[
z^{-a(\square)} q^{l(\square)+1} = q^{r_m} \tfrac{z_{j_m}}{z_i}.
\]
Let $\square$ be the cell in row $i$ and column $\mu_{j_m}+1$ ($j_m>i$). Then $a(\square)=\mu_i-\mu_{j_m}-1$, $l(\square)=j_m-1-i$. Therefore
\[
z^{a(\square)+1} q^{-l(\square)} = q \tfrac{z_i}{z_{j_m}}.
\]
For the cell in column $1$ we have $a(\square)=\mu_i-1$, $l(\square)=l(\mu)-i$, so
\[
z^{a(\square)+1} q^{-l(\square)} = z_i.
\]
Thus the factors of \eqref{eq:prod P} form a sub-multiset of the factors of the arm-leg product, and the claim follows.
\end{proof}
\begin{cor}\label{cor:laurent}
For any partition $\mu$ the product $N_\mu(1)\Omega_\mu$ is a Laurent polynomial.
\end{cor}
\begin{proof}
We have
\[
N_\mu(1)\Omega_\mu = \prod_{i=1}^g N_\mu(\alpha_i^{-1}) f_\mu
\]
and
\[
\prod_{i=1}^g N_\mu(\alpha_i^{-1}) = \prod_{\square\in\mu} P(z^{a(\square)+1} q^{-l(\square)}) P(z^{-a(\square)} q^{l(\square)+1}) \times \pm\;\text{a monomial}.
\]
\end{proof}
\subsection{Interpolation}
Another nice property of the function $f$ is that substitution $z_n=1$ leads to essentially the same function in $n-1$ variables:
\begin{prop}\label{prop:regularity}
For any $n$ we have
\[
f(1,z_1,\ldots,z_n) = f(q z_1, \ldots, q z_n).
\]
\end{prop}
\begin{proof}
Note that because of the product $\prod_{i=2}^n (1-z_i)$ in the definition of $f(1, z_1, \ldots, z_n)$ only the terms with $\sigma(1)=1$ survive. So we can reduce the summation over $S_{n+1}$ to a summation over $S_n$. After cancellation of $\prod_i (1-z_i)$ we obtain
\begin{multline*}
f(1, z_1,\ldots,z_n) = \prod_i \frac{P(1)}{P(z_i)}
\\
\times \sum_{\sigma\in S_n}\sigma \left\{
\prod_{i>j}\frac{P(\frac{z_i}{z_j})}{(1-\frac{z_i}{z_j}) P(q \frac{z_i}{z_j})} \prod_{i> j+1} (1-q \tfrac{z_i}{z_j}) \prod_{i>1} (1-q z_i) \prod_{i}\frac{P(z_i)}{P(q z_i)}
\right\},
\end{multline*}
which coincides with $f(q z_1, \cdots, q z_n)$.
\end{proof}
\begin{cor}
Let $\mu$ be a partition and let $n\geq l(\mu)$. Define $z_{n,i}(\mu)=z^{\mu_i} q^{i-n}$ for $i=1,\ldots,n$. Then
\[
f(z_{n,1}(\mu), \ldots, z_{n,n}(\mu)) = f(z_1(\mu), \ldots, z_{l(\mu)}(\mu)).
\]
\end{cor}
Thus, instead of having a separate function for each value of $l(\mu)$ we can use the same function of $n$ arguments if $n$ is big enough.
\section{Polynomiality and the main result}
In this section we return to variables $q,t$ which correspond to Schiffmann's variables $z, q$ respectively. First we prove the following statement. The proof is straightforward using methods of \cite{mellit2016integrality}, but tedious. Let $R$ be a lambda ring containing $\mathbb Q(t)[q^{\pm 1}]$. We denote by $R^*$ the tensor product $R\otimes_{\mathbb Q(t)[q^{\pm 1}]} \mathbb Q(q,t)$ and assume $R\subset R^*$.
\begin{defn}
A regular function of $z_i$ is a sequence of Laurent polynomials
\[
f_n\in R[z_1^{\pm 1},\ldots,z_n^{\pm 1}] \qquad (n\geq 0)
\]
such that
\begin{enumerate}
\item $f_n$ is symmetric in $z_1, \ldots, z_n$,
\item $f_{n+1}(1,z_1,\ldots,z_n)=f_n(t z_1, \ldots, t z_n)$.
\end{enumerate}
\end{defn}
For a regular function $f$ and a partition $\mu$ we set
\[
f_\mu = f_{l(\mu)}(z_1(\mu), \ldots, z_{l(\mu)}(\mu)),\qquad z_i(\mu) = q^{\mu_i} t^{i-l(\mu)}.
\]
\begin{lem}\label{lem:deformation}
Let $f(u)=1+f^{(1)} u + f^{(2)} u^2+\cdots$ be a power series whose coefficients $f^{(i)}$ are regular functions in the above sense. Let
\[
\Omega[X] = \sum_{\mu\in \mathcal P} c_\mu \tilde{H}_\mu[X;q,t]
\]
be a series with $c_\mu\in R^*$, $c_{\varnothing}=1$ such that all coefficients of
\[
\mathbb H[X] = (q-1) \pLog \Omega[X]
\]
are in $R$. Let
\[
\Omega_f[X,u] = \sum_{\mu} c_\mu \tilde{H}_\mu[X;q,t] f_\mu(u),\qquad \mathbb H_f[X,u] = (q-1) \pLog \Omega_f[X,u].
\]
Consider the expansion
\[
\mathbb H_f[X,u]= \mathbb H[X] + u \mathbb H_{f,1}[X] + u^2 \mathbb H_{f,2}[X] + \cdots.
\]
Then all coefficients of $\mathbb H_{f,i}[X]$ for $i\geq 1$ are in $(q-1) R$. In other words, the specialization $q=1$ of $\mathbb H_f[X,u]$ is independent of $u$.
\end{lem}
\begin{proof}
Let $S=-(q-1)(t-1)$. Recall the notation $\int^S_X F[X,X^*]$ (see \cite{mellit2016integrality}). This is a linear operation such that
\[
\int^S_X G[X] F[X^*] = (G[X], F[X])_X^S =(G[X], F[SX])_X,
\]
and $(-,-)_X$ is the standard Hall scalar product,
\[
(s_\mu[X], s_{\lambda}[X])_X = \delta_{\mu,\lambda}.
\]
Recall that modified Macdonald polynomials are orthogonal with respect to $(-,-)_X^S$. In this proof we call an expression $F$ admissible if $(q-1)\pLog F$ has all of its coefficients in $R$. It was proved in \cite{mellit2016integrality} that $\int^S_X$ preserves admissibility. By the assumption $\Omega[X]$ is admissible. We will ``construct'' $\mathbb H_f[X,u]$ from admissible building parts.
Let $R[Z,Z^*]$ be the free lambda ring over $R$ with two generators $Z$ and $Z^*$. Fix a large integer $N$. For each $i\geq 1$ let $\tilde{f}^{(i)}\in R[Z,Z^*]$ be any element such that
\[
\tilde{f}^{(i)}\left[\sum_{i=1}^N z_i, \sum_{i=1}^N z_i^{-1}\right] = f^{(i)}_N(z_1,\ldots,z_N).
\]
One way to construct such an element is to find $m\geq 0$ such that $(z_1\cdots z_N)^m f^{(i)}_N(z_1,\ldots,z_N)=p(z_1,\ldots,z_n)$ does not contain negative powers of $z_i$, then lift $p$ to a symmetric function $\tilde{p}\in R[Z]$ and set
\[
\tilde{f}^{(i)}[Z,Z^*] = \tilde p[Z] e_N[Z^*]^m.
\]
Then set
\[
\tilde{f}(u) = 1 + \tilde{f}^{(1)}u + \tilde{f}^{(2)} u^2 +\cdots\in R[Z,Z^*][[u]].
\]
We can take plethystic logarithm:
\[
\pLog \tilde{f}(u) = g(u)= g^{(1)} u + g^{(2)} u^2 + \cdots \in u R[Z,Z^*][[u]].
\]
For any partition $\mu$ satisfying $l(\mu)\leq N$ by regularity of $f$ we have
\[
f_\mu = f_{l(\mu)}(q^{\mu_1} t^{1-l(\mu)}, q^{\mu_2} t^{2-l(\mu)}, \ldots, q^{\mu_{l(\mu)}}) = f_N(q^{\mu_1} t^{1-N}, q^{\mu_2} t^{2-N}, \ldots, q^{\mu_N}).
\]
Thus we can obtain $f_\mu$ from $\tilde f$ by specializing at
\[
Z = Z_\mu=\sum_{i=1}^N q^{\mu_i} t^{i-N} = t^{1-N} (q-1) B_\mu + \sum_{i=1}^N t^{i-N} = \frac{t^{1-N}}{1-t} S B_\mu + \frac{t^{-N}-1}{t^{-1}-1},
\]
and similarly for $Z^*$. Hence there exists a series
\[
g'(u)\in u R[Z,Z^*][[u]]
\]
such that for any partition $\mu$ with $l(\mu)\leq N$ we have
\[
f_\mu = \pExp[g'(u)[S B_\mu, S B_\mu^*]].
\]
Specialization can be replaced by scalar product using the identity
\[
F[S Y] = (F[X], \pExp[SXY])_X = (F[X], \pExp[XY])_X^*,
\]
and we obtain
\begin{equation}\label{eq:fmu integral}
f_\mu = \int^S_{Z,V} \pExp[g'(u)[Z, V]] \pExp[Z^* B_\mu + V^* B_\mu^*]\qquad (l(\mu)\leq N).
\end{equation}
Let us show that the sum
\begin{equation}\label{eq:omega4B}
\tilde{\Omega}[X,Z,V]=\sum_{\mu\in\mathcal P} c_\mu \tilde{H}_\mu[X;q,t] \pExp[Z B_\mu + V B_\mu^*]
\end{equation}
is admissible. Begin with the series
\begin{equation}\label{eq:omega4}
\sum_{\mu\in\mathcal P} \frac{\tilde{H}_\mu[X] \tilde{H}_\mu[Y] \tilde{H}_\mu[Z] \tilde{H}_\mu[V]}{(\tilde{H}_\mu, \tilde{H}_\mu)^S},
\end{equation}
which is admissible by the main theorem of \cite{mellit2016integrality}. Recall the nabla operator $\nabla$, the shift operator $\tau$ and the multiplication by $\pExp\left[\frac{X}{S}\right]$ operator $\tau^*$, and Tesler's identity
\[
\nabla \tau \tau^* \tilde{H}_\mu[X] = \pExp\left[\frac{D_\mu X}{S}\right],
\]
where $D_\mu = -1-S B_\mu$. This implies
\[
\tau^* \nabla \tau \tau^* \tilde{H}_\mu[X] = \pExp[-X B_\mu].
\]
All of the operators involved preserve admissibility (Corollary 6.3 from \cite{mellit2016integrality}. In particular, we see that the operator that sends $\tilde{H}_\mu[X]$ to $\pExp[X B_\mu]$ preserves admissibility. Let $\omega$ be the operator that sends $q,t,X$ to $q^{-1}, t^{-1}, -X$. Then using $\omega \nabla=\nabla^{-1}\omega$, $\omega \tilde{H}_\mu[X] = \frac{\tilde{H}_\mu[X]}{\tilde{H}_\mu[-1]}$ and the fact that $\nabla^{-1}$ preserves admissibility (Corollary 6.4 from \cite{mellit2016integrality}) we see that the operator that sends $\tilde{H}_\mu[X]$ to $\pExp[X B_\mu^*]$ preserves admissibility too. Applying these operators to \eqref{eq:omega4} in the variables $Z, V$ we obtain that the following series is admissible:
\[
\sum_{\mu\in\mathcal P} \frac{\tilde{H}_\mu[X] \tilde{H}_\mu[Y] \pExp[Z B_\mu + V B_\mu^*]}{(\tilde{H}_\mu, \tilde{H}_\mu)^S}.
\]
Finally, pairing this series with $\Omega[X]$ we obtain admissibility of \eqref{eq:omega4B}.
Because of \eqref{eq:fmu integral} we have
\[
\Omega_f(u) = \int^S_{Z,V} \pExp[g'(u)[Z, V]] \tilde{\Omega}[X, Z^*, V^*]\qquad \text{up to terms of degree $>N$ in $X$}.
\]
In what follows we ignore the terms of degree $>N$ in $X$. Since $N$ can be chosen as large as possible, this is enough. Notice that $\pExp[g'(u)[Z, V]]$ is ``more'' than admissible in the following sense. Introduce a new free (in the lambda ring sense) variable $W$. Then $\pExp[\tfrac{W}{S} g'(u)[Z, V]]$ is admissible. Therefore the following is admissible:
\[
\Omega_f[X,W,u] := \int^S_{Z,V} \pExp[\tfrac{W}{S} g'(u)[Z, V]] \tilde{\Omega}[X, Z^*, V^*].
\]
So we have
\[
\mathbb H_f[X,W,u] = (q-1)\pLog \Omega_f[X,W,u] = \sum_{i\geq 0} \mathbb H_{f,i}[X,W] u^i
\]
with $\mathbb H_{f,i}[X,W]\in R[X,W]$. Finally notice that
\[
\mathbb H_{f,i}[X] = \mathbb H_{f,i}[X,S]\equiv\mathbb H_{f,i}[X,0] \pmod{(q-1)R[X]},
\]
and
\[
\mathbb H_{f}[X,0,u] = (q-1)\pLog \Omega_f[X,0,u],
\]
\[
\Omega_f[X,0,u] = \int^S_{Z,V} \tilde{\Omega}[X, Z^*, V^*] = \tilde{\Omega}[X, 0, 0] = \Omega[X].
\]
This implies
\[
\mathbb H_{f,i}[X] \equiv 0 \pmod{(q-1)R[X]} \qquad (i\geq 1).
\]
\end{proof}
Then our main result is
\begin{thm}\label{thm:main tech}
For any $g\geq 0$ let
\[
\Omega(T, q, t,\alpha_1,\ldots,\alpha_g) = \sum_{\mu\in\mathcal P} \frac{\prod_{i=1}^g N_\mu(\alpha_i^{-1})}{N_\mu(1)} T^{|\mu|},
\]
where
\[
N_\mu(u)=\prod_{\square\in\mu}(q^{a(\square)}-u t^{1+l(\square)})(q^{a(\square)+1}-u^{-1} t^{l(\square)}).
\]
Let
\[
\Omega'(T, q, t,\alpha_1,\ldots,\alpha_g) = \sum_{\mu\in\mathcal P} \Omega_\mu(q,t,\alpha_1,\ldots,\alpha_g) T^{|\mu|},
\]
where $\Omega_\mu$ are the Schiffmann's terms defined in Section \ref{sec:schiffmanns terms}. Let
\[
\mathbb H(T, q, t,\alpha_1,\ldots,\alpha_g) = -(q-1)(t-1)\pLog \Omega(T, q, t,\alpha_1,\ldots,\alpha_g),
\]
\[
\mathbb H'(T, q, t,\alpha_1,\ldots,\alpha_g) = -(q-1)(t-1)\pLog \Omega'(T, q, t,\alpha_1,\ldots,\alpha_g),
\]
and let $\mathbb H(q, t,\alpha_1,\ldots,\alpha_g)_k$ denote the $k$-th coefficient of $\mathbb H(T, q, t,\alpha_1,\ldots,\alpha_g)$ viewed as a power series in $T$, and similarly for $\mathbb H'$. Then we have
\begin{enumerate}
\item $\mathbb H'(q, t,\alpha_1,\ldots,\alpha_g)_k \in \mathbb Q(t)[q^{\pm 1}, \alpha_1^{\pm 1},\ldots,\alpha_g^{\pm 1}]$,
\item $\mathbb H'(1, t,\alpha_1,\ldots,\alpha_g)_k=\mathbb H(1, t,\alpha_1,\ldots,\alpha_g)_k$.
\end{enumerate}
\end{thm}
\begin{proof}
By the main result of \cite{mellit2016integrality} we have
\[
\mathbb H(q, t,\alpha_1,\ldots,\alpha_g)_k \in \mathbb Q[t,q, \alpha_1^{\pm 1},\ldots,\alpha_g^{\pm 1}].
\]
By Corollary \ref{cor:laurent} we have
\[
\mathbb H'(q, t,\alpha_1,\ldots,\alpha_g)_k \in \mathbb Q(t,q)[\alpha_1^{\pm 1},\ldots,\alpha_g^{\pm 1}].
\]
So we can pass to the ring of Laurent series in $\alpha_1^{-1},\ldots,\alpha_g^{-1}$ and it is enough to prove the corresponding statements (i) and (ii) for the coefficients in front of monomials of the form $\prod_{i=1}^g \alpha_i^{m_i}$. Let us apply Lemma \ref{lem:deformation} for the ring $R=\mathbb Q(t)[q^{\pm 1}, \alpha_1^{\pm 1},\ldots,\alpha_g^{\pm 1}]$, series
\[
\Omega[X, q, t,\alpha_1,\ldots,\alpha_g] = \sum_{\mu\in\mathcal P} \frac{\prod_{i=1}^g N_\mu(\alpha_i^{-1}) \tilde{H}_\mu[X,q,t]}{N_\mu(1)},
\]
and the regular function $f(u)$ obtained from $f$ (see \eqref{eq:f definition} and Proposition \ref{prop:regularity}) by setting $u \alpha_i^{-1}$ in place of $\alpha_i^{-1}$, so that $f(u)$ becomes a power series in $u$ with coefficients in $R$.
To be able to apply Lemma \ref{lem:deformation} we need to show that the constant coefficient of $f(u)$ is $1$, in other words we need to check that
\[
\sum_{\sigma\in S_n}\sigma \left\{
\prod_{i>j}\left(\frac{1}{1-\frac{z_i}{z_j}}\right) \prod_{i> j+1} (1-q \tfrac{z_i}{z_j}) \prod_{i\geq 2} (1-z_i)
\right\}=1.
\]
We do this by induction. Denote the left hand side by $L_n$. Notice that $L_n$ is a polynomial. Suppose we know that $L_{n-1}=1$. Then by Proposition \ref{prop:regularity} we know that $L_{n}-1$ is divisible by $z_1-1$. Since it is a symmetric polynomial, it must be divisible by $\prod_{i=1}^n (z_i-1)$. On the other hand, the degree of $L_n$ is at most $n-1$, so necessarily $L_n-1=0$.
After applying Lemma \ref{lem:deformation} we can set $X=T$, where $T$ is the variable from the statement of the Theorem. In particular, $T$ is assumed to satisfy $p_k[T]=T^k$ and we can use the identity $\tilde{H}_\mu[T;q,t]=T^{|\mu|}$. Let
\[
\mathbb H'(T,q,t,u) = -(q-1)(t-1)\pLog\left[\sum_{\mu\in \mathcal P} \frac{\prod_{i=1}^g N_\mu(\alpha_i^{-1})}{N_\mu(1)} T^{|\mu|} f_\mu(u) \right].
\]
Lemma \ref{lem:deformation} says that
\[
\mathbb H'(T,q,t,u)-\mathbb H(T,q,t) \in (q-1) \mathbb Q(t)[q^{\pm 1}, \alpha_1^{\pm 1},\ldots,\alpha_g^{\pm 1}][[T,u]].
\]
On the other hand, the coefficient in front of any monomial in $\alpha_1, \ldots,\alpha_g, T$ has bounded degree in $u$, wo we can set $u=1$ and obtain a statement about Laurent series in $\alpha_i^{-1}$:
\[
\mathbb H'(T,q,t,1)-\mathbb H(T,q,t) \in (q-1) \mathbb Q(t)[q^{\pm 1}]((\alpha_1^{-1},\ldots,\alpha_g^{-1}))[[T]].
\]
Finally we remember that $\mathbb H'(T,q,t)=\mathbb H'(T,q,t,1)$ and remember that the coefficients of $\mathbb H'(T,q,t)$ are Laurent \emph{polynomials} in $\alpha_i$ to obtain
\[
\mathbb H'(T,q,t)-\mathbb H(T,q,t) \in (q-1) \mathbb Q(t)[q^{\pm 1}, \alpha_1^{\pm 1},\ldots,\alpha_g^{\pm 1}][[T]].
\]
\end{proof}
Theorem \ref{thm: main intro} is a direct corollary of Theorem \ref{thm:main tech} and \cite{schiffmann2014indecomposable}.
\section*{Acknowledgements}
Olivier Schiffmann drew my attention to his paper \cite{schiffmann2014indecomposable}, suggested to try to deduce the conjectures of Hausel-Rodriguez-Villegas and pointed out that the conjecture of Hausel-Thaddeus also follows. He also made several corrections in a preliminary version of this paper. Fernando Rodriguez-Villegas experimented with Schiffmann's formula on a computer, and I had interesting conversations with him on this subject. I would like to thank them and Tamas Hausel for encouragement.
I would like to thank the organizers of the workshops on Higgs bundles at EPFL Lausanne in 2016 and SISSA Trieste in 2017. I discovered some of the ideas leading to the result of this work during these workshops.
This work was performed during my stay at IST Austria, where I was supported by the Advanced Grant ``Arithmetic and Physics of Higgs moduli spaces'' No. 320593 of the European Research Council.
\printbibliography
\end{document}
|
\section{Introduction}
\section{Introduction}
Throughout, we write $\mathbb{Q}_n:=\mathbb{Q}(e^{2\pi i/n})$ to denote the extension field of $\mathbb{Q}$ obtained by adjoining $n$th roots of unity in $\mathbb{C}$. In particular, if a finite group $G$ has size $n$, then $\mathbb{Q}_n$ is a splitting field for $G$ and the group $\mathrm{Gal}(\mathbb{Q}_n/\mathbb{Q})$ acts on the set of irreducible ordinary characters, $\mathrm{Irr}(G)$, of $G$.
In \cite{navarro2004}, G. Navarro conjectured a refinement to the well-known McKay conjecture that incorporates this action of $\mathrm{Gal}(\mathbb{Q}_n/\mathbb{Q})$. Specifically, the ``Galois-McKay" conjecture posits that if $\ell$ is a prime, $\sigma\in\mathrm{Gal}(\mathbb{Q}_n/\mathbb{Q})$ sends every $\ell'$ root of unity to some $\ell$th power, and $P\in\mathrm{Syl}_\ell(G)$ is a Sylow $\ell$-subgroup of $G$, then the number of characters in $\mathrm{Irr}_{\ell'}(G)$ that are fixed by $\sigma$ is the same as the number of characters in $\mathrm{Irr}_{\ell'}(N_G(P))$ fixed by $\sigma$. Here for a finite group $X$, we write $\mathrm{Irr}_{\ell'}(X)=\{\chi\in\mathrm{Irr}(X)\mid\ell\nmid\chi(1)\}$.
In the same paper, Navarro shows that the validity of his Galois-McKay conjecture would imply the following statement in the case $\ell=2$ (as well as a corresponding statement for $\ell$ odd):
\begin{conjecture}[Navarro]\label{conj:mainprob}
Let $G$ be a finite group and let $\sigma\in\mathrm{Gal}(\mathbb{Q}_{|G|}/\mathbb{Q})$ fixing $2$-roots of unity and squaring $2'$-roots of unity. Then $G$ has a self-normalizing Sylow $2$-subgroup if and only if every irreducible complex character of $G$ with odd degree is fixed by $\sigma$.
\end{conjecture}
The ordinary McKay conjecture has been reduced to proving certain inductive statements for simple
groups in \cite{IsaacsMalleNavarroMcKayreduction}, and even recently proven for $\ell=2$ by G. Malle and B. Sp{\"a}th in \cite{MalleSpathMcKay2}. However, at the time of the writing of the current article, there is not yet such a reduction for the Galois-McKay conjecture. In fact, very few groups have even been shown to satisfy the Galois-McKay conjecture, and a reduction seems even more elusive in the case $\ell=2$ than for odd primes. We consider \prettyref{conj:mainprob} to be a weak form of the Galois-McKay refinement in this case, and hope that some of the observations made in the course of its proof will be useful in future work on the full Galois-McKay conjecture.
In \cite{NavarroTiepTurull2007}, Navarro, Tiep, and Turull proved the corresponding statement for $\ell$ odd. In \cite{SchaefferFrySN2S1}, the author proved a reduction theorem for \prettyref{conj:mainprob}, which reduced the problem to showing that slightly stronger inductive conditions hold for all finite nonabelian simple groups (groups satisfying these statements are called ``SN2S-Good"; see also \prettyref{sec:SimpleStatements} below for the statements), and proved that the sporadic, alternating, and several simple groups of Lie type satisfy these conditions. In \cite{SFTaylorTypeA}, the author, together with J. Taylor, extended the strategy from \cite{SchaefferFrySN2S1} to show that the conditions hold for every simple group of Lie type in characteristic $2$, as well as for the simple groups $PSL_n(q)$ and $PSU_n(q)$ for odd $q$.
The main result of this article is the completion of the proof of \prettyref{conj:mainprob}, which is achieved by showing that the remaining simple groups of Lie type satisfy the conditions to be SN2S-Good. That is, we prove the following:
\begin{theoremA}\label{thm:mainthm}
\prettyref{conj:mainprob} holds for every finite group.
\end{theoremA}
Theorem A has the following interesting consequence:
\begin{corollary*}
One can determine from the character table of a finite group $G$ whether a Sylow $2$-subgroup of $G$ is self-normalizing.
\end{corollary*}
We remark that several relevant results regarding the Galois automorphism $\sigma$ involved in \prettyref{conj:mainprob}, including a simplified version of the reduction of \prettyref{conj:mainprob}, have been obtained by G. Navarro and C. Vallejo in \cite{navarro-vallejo:2017:2-local-blocks-with-one-simple-module}. The latter appeared just before the submission of the current article, which is completely independent of the work in \cite{navarro-vallejo:2017:2-local-blocks-with-one-simple-module}. In particular, the conditions to be proved in the case of simple groups of Lie type of type $B, C, D,$ and $\tw{2}{D}$, which are the primary focus here, remain nearly the same under Navarro and Vallejo's version of the reduction. Specifically, our results here imply that simple groups of Lie type $B$, $C$, $D$, and $\tw{2}{D}$ satisfy \cite[Conjecture A]{navarro-vallejo:2017:2-local-blocks-with-one-simple-module}. Further, the strategy employed here in these cases, and in particular our analysis of the action of the Galois group on the Howlett--Lehrer parametrization of characters of groups of Lie type (see \prettyref{sec:gallact} below) may be of significant independent interest, since the effect of group actions on such parametrizations is an especially problematic component in a number of main problems regarding the representations of groups of Lie type.
We begin in \prettyref{sec:gal twist} by studying the actions of field automorphisms on modules in a very general setting. In \prettyref{sec:gallact}, we make use of this framework to describe the action of $\mathrm{Gal}(\mathbb{Q}_n/\mathbb{Q})$ on the Harish-Chandra parameterization of characters of groups with a $BN$-pair, which builds upon the strategy employed by G. Malle and B. Sp{\"a}th in \cite{MalleSpathMcKay2} for the case of the action of $\mathrm{Aut}(G)$. (See \prettyref{thm:GaloisAct}.) We hope that this will be useful beyond the scope of this article, especially in the context of proving the Galois-McKay conjecture (or the inductive conditions in an eventual reduction to simple groups) for groups of Lie type. In \prettyref{sec:SN2S}, we prove Theorem A by showing that the remaining simple groups of Lie type defined in odd characteristic are SN2S-Good. This is done in \prettyref{cor:princseries1mod8}, and Theorems \ref{thm:typeC}, \ref{thm:typeBD}, and \ref{thm:typeE6}.
\section{Galois Twists }\label{sec:gal twist}
Here we consider a general framework for studying the action of field automorphisms on modules. Let $\mathbb{F}$ be a field. Throughout, an $\mathbb{F}$-algebra $A$ will be taken to mean a finite-dimensional associative unital $\mathbb{F}$-algebra, and an $A$-module will be taken to mean a finite-dimensional left $A$-module.
\subsection{$\sigma$-Twists}
Let $\sigma\in\mathrm{Aut}(\mathbb{F})$ be an automorphism of $\mathbb{F}$. For $\alpha\in\mathbb{F}$, we write $\alpha^\sigma$ for the image of $\alpha$ under $\sigma$. Given a vector space $V$ over $\mathbb{F}$, we define the \emph{$\sigma$-twist}, $V^\sigma$, of $V$ to be the $\mathbb{F}$-vector space whose underlying abelian group is the same as $V$, but with scalar multiplication defined by
\[\alpha\star_\sigma v:=\alpha^{\sigma^{-1}} v\] for $\alpha\in\mathbb{F}$ and $v\in V$. Here the product on the right-hand side is just the ordinary scalar multiplication on $V$. For any $v\in V$, we denote by $v^\sigma$ the element $v$, but viewed in $V^\sigma$. Hence, the scalar multiplication may alternatively be written:
\[\alpha v^\sigma:=\alpha^{\sigma^{-1}} v.\]
With this in place, we may analogously define the $\sigma$-twist of an $\mathbb{F}$-algebra $A$ to be the $\mathbb{F}$-algebra $A^\sigma$ whose underlying ring structure is the same as that of $A$, but with $\sigma$-twisted vector space structure defined as above. For $a\in A$, we will denote by $a^\sigma$ the element $a$ but viewed in $A^\sigma$. Further, for an $A$-module $M$, we may define the $\sigma$-twist of $M$ to be the $A^\sigma$-module $M^\sigma$ whose underlying vector space structure is the $\sigma$-twisted vector space structure as above, and for $a\in A, m\in M$, we have \[a^\sigma m^\sigma=(am)^\sigma.\]
Given two vector spaces $V$ and $W$ over $\mathbb{F}$, we define a \emph{$\sigma$-semilinear} map to be a homomorphism of abelian groups $\phi\colon V\rightarrow W$ satisfying $\phi(\alpha v)=\alpha^\sigma\phi(v)$ for all $\alpha\in\mathbb{F}$ and $v\in V$.
Note then that for a vector space $V$ over $\mathbb{F}$ or $\mathbb{F}$-algebra $A$ the maps
\[-^\sigma\colon V\rightarrow V^\sigma \quad\hbox{ and }\quad -^\sigma\colon A\rightarrow A^\sigma\] defined by $v\mapsto v^\sigma$ and $a\mapsto a^\sigma$ are $\sigma$-semilinear isomorphisms of abelian groups and rings, respectively.
\subsection{$\sigma$-Twists and Representations}
Let $\mathrm{Irr}(A)$ denote the set of isomorphism classes of simple $A$-modules. Then we see that the map $-^\sigma\colon \mathrm{Irr}(A)\rightarrow \mathrm{Irr}(A^\sigma)$, given by $M\mapsto M^\sigma$, is naturally a bijection.
Let $M$ be an $A$-module and let $\rho\colon A\rightarrow \mathrm{End}_A(M)$ be the corresponding representation. That is, for $a\in A$, $\rho(a)$ is the endomorphism of $M$ given by $\rho(a)m:=am$ for $m\in M$. Fixing a basis $\mathfrak{B}$ for $M$, we write $[\rho(a)]_\mathfrak{B}\in \mathrm{Mat}_{\dim M}(\mathbb{F})$ for the image of $a\in A$ under the resulting matrix representation. The character of $A$ afforded by $M$ is the map $\eta\colon A\rightarrow\mathbb{F}$ defined by the trace $\eta(a):=\mathrm{Tr}\left([\rho(a)]_\mathfrak{B}\right).$
Similarly, we have a representation $\rho^\sigma\colon A^\sigma\rightarrow \mathrm{End}_{A^\sigma}(M^\sigma)$ which affords the $A^\sigma$-module $M^\sigma$. Here $\rho^\sigma(a^\sigma)m^\sigma:=a^\sigma m^\sigma=(am)^\sigma$ for $m\in M$, $a\in A$. Also, note that if $\mathfrak{B}$ is a basis for $M$, then the set $\mathfrak{B}^\sigma\subseteq M^\sigma$ is a basis for $M^\sigma$. If $\eta$ is the character of $A$ afforded by $M$, we will denote by $\eta^\sigma$ the character of $A^\sigma$ afforded by $M^\sigma$.
Given a matrix $X$ over $\mathbb{F}$, we denote by $X^\sigma$ the matrix obtained by applying $\sigma$ to each entry of $X$. That is, if $X=(x_{ij})\in\mathrm{Mat}_d(\mathbb{F})$ is a $d\times d$ matrix, the matrix $X^\sigma$ is $(x_{ij}^\sigma)\in\mathrm{Mat}_d(\mathbb{F})$. The following lemma tells us that the matrix corresponding to $\rho^\sigma(a^\sigma)$ is simply the matrix obtained in this way from $\rho(a)$.
\begin{lemma}\label{lem:sigmamatrixrep}
Let $M$ be an $A$-module affording the character $\eta$ of $A$ and let $\mathfrak{B}$ be a basis for $M$. Given $a\in A$, we have
\[\left[\rho^\sigma(a^\sigma)\right]_{\mathfrak{B}^\sigma}=\left[\rho(a)\right]_{\mathfrak{B}}^\sigma \in \mathrm{Mat}_{\dim M}(\mathbb{F})\quad\hbox{ and }\quad \eta^\sigma(a^\sigma)=\eta(a)^\sigma.\]
\end{lemma}
\begin{proof}
Let $a\in A$ and let $\mathfrak{b}\in\mathfrak{B}$ be an element of the basis. Then we may write
\[a\mathfrak{b}=\sum_{\mathfrak{c}\in\mathfrak{B}} \alpha_{\mathfrak{b}, \mathfrak{c}}\mathfrak{c}\] for some scalars $\alpha_{\mathfrak{b},\mathfrak{c}}\in\mathbb{F}$. Then
\[a^\sigma\mathfrak{b}^\sigma=(a\mathfrak{b})^\sigma=\left(\sum_{\mathfrak{c}\in\mathfrak{B}} \alpha_{\mathfrak{b}, \mathfrak{c}}\mathfrak{c}\right)^\sigma=\sum_{\mathfrak{c}\in\mathfrak{B}} \alpha_{\mathfrak{b}, \mathfrak{c}}^\sigma\mathfrak{c}^\sigma,\] where the last equality follows from the $\sigma$-semilinearity of $-^\sigma\colon M\rightarrow M^\sigma$. Hence $\left[\rho^\sigma(a^\sigma)\right]_{\mathfrak{B}^\sigma}=(\alpha_{\mathfrak{b},\mathfrak{c}}^\sigma)=\left[\rho(a)\right]_{\mathfrak{B}}^\sigma$, which proves the statement.
\end{proof}
\subsection{$\sigma$-Twists and Homomorphisms}
Let $C\subseteq A$ be a subalgebra and let $M$ and $N$ be $A$-modules. We denote by $\mathrm{Hom}_C(M,N)$ the vector space of all $\mathbb{F}$-linear maps $f\colon M\rightarrow N$ satisfying $f(cm)=cf(m)$ for every $c\in C$ and $m\in M$. In particular, $\mathrm{Hom}_A(M,N)$ is the set of all $A$-module homomorphisms. The following lemma and corollary, though straightforward, will be useful throughout the next several sections.
\begin{lemma}\label{lem:sigmahoms}
Given $A$-modules $M$ and $N$ and a subalgebra $C$ of $A$, the map
\[\phi\colon\mathrm{Hom}_C(M, N)^\sigma\rightarrow\mathrm{Hom}_{C^\sigma}(M^\sigma, N^\sigma),\] defined by $\phi(f^\sigma)(m^\sigma):=f(m)^\sigma$ for all $f\in \mathrm{Hom}_C(M,N)$ and $m\in M$, is an isomorphism of vector spaces over $\mathbb{F}$.
\end{lemma}
\begin{corollary}\label{cor:sigmaend}
Given an $A$-module $M$, the map
\[\phi\colon\mathrm{End}_A(M)^\sigma\rightarrow\mathrm{End}_{A^\sigma}(M^\sigma),\] defined by $\phi(f^\sigma)(m^\sigma):=f(m)^\sigma$ for all $f\in \mathrm{End}_A(M)$ and $m\in M$, is an isomorphism of $\mathbb{F}$-algebras.
\end{corollary}
We remark that given a representation $\rho\colon A\rightarrow\mathrm{End}_A(M)$ and $a\in A$, the endomorphisms $\rho^\sigma(a^\sigma)$ and $\phi(\rho(a)^\sigma)$ agree, where $\phi$ is the isomorphism in \prettyref{cor:sigmaend}.
\subsection{$\sigma$-Twists and Group Algebras}
We now turn our attention to the case that $A=\mathbb{F} G$ is a group algebra over $\mathbb{F}$ for a finite group $G$. Recall that members of $\mathbb{F} G$ are formal sums $\sum_{g\in G} \alpha_g g$, where $\alpha_g\in \mathbb{F}$.
We obtain a $\sigma^{-1}$-semilinear ring isomorphism $\iota_0\colon \mathbb{F} G\rightarrow \mathbb{F} G$ defined by
\[\iota_0\left(\sum_{g\in G} \alpha_g g\right):=\sum_{g\in G}\alpha_g^{\sigma^{-1}}g.\] Composing with the $\sigma$-semilinear ring isomorphism $-^\sigma\colon \mathbb{F} G\rightarrow (\mathbb{F} G)^\sigma$ yields an isomorphism of $\mathbb{F}$-algebras $\iota_0^\sigma\colon \mathbb{F} G\rightarrow (\mathbb{F} G)^\sigma$ defined by $\iota_0^\sigma(a):=\iota_0(a)^\sigma$ for $a\in \mathbb{F} G$.
If $M$ is an $\mathbb{F} G$-module, we may use this isomorphism to view the $(\mathbb{F} G)^\sigma$-module $M^\sigma$ as an $\mathbb{F} G$-module via
\[a\cdot m^\sigma:=\iota_0^\sigma(a)m^\sigma=\iota_0(a)^\sigma m^\sigma=(\iota_0(a)m)^\sigma\] for $a\in \mathbb{F} G$ and $m\in M$.
By an abuse of notation, if $\rho\colon \mathbb{F} G\rightarrow \mathrm{End}_{\mathbb{F} G}(M)$ is the representation afforded by $M$, we will denote by $\rho^\sigma$ the representation of $\mathbb{F} G$ afforded by $M^\sigma$. We remark that this can be viewed as $\iota_0^\sigma$ composed with the $(\mathbb{F} G)^\sigma$-representation $\rho^\sigma$.
Let $\mathbb{F}_0\subseteq\mathbb{F}$ denote the prime subfield of $\mathbb{F}$, and note that $\alpha^\sigma=\alpha$ for every $\alpha\in\mathbb{F}_0$. In this context, we may rephrase \prettyref{lem:sigmamatrixrep} as follows:
\begin{lemma}\label{lem:sigmamatrixrep2}
Let $M$ be an $\mathbb{F} G$-module affording the character $\chi$ of $\mathbb{F} G$ and let $\mathfrak{B}$ be a basis for $M$. Given $a\in \mathbb{F} G$, we have
\[\left[\rho^\sigma(a)\right]_{\mathfrak{B}^\sigma}=\left[\rho(\iota_0(a))\right]_{\mathfrak{B}}^\sigma \in \mathrm{Mat}_{\dim M}(\mathbb{F})\quad\hbox{ and }\quad \chi^\sigma(a)=\chi(\iota_0(a))^\sigma.\] In particular, $\chi^\sigma(a)=\chi(a)^\sigma$ for $a\in\mathbb{F}_0 G$.
\end{lemma}
\subsection{$\sigma$-Twists on (Co)-induced Modules}\label{sec:twistcoind}
Let $P\leq G$ be a subgroup, so that $\mathbb{F} P\subseteq \mathbb{F} G$ is a subalgebra. Note that $\iota_0^\sigma$ restricts to an isomorphism $\iota_0^\sigma\colon \mathbb{F} P\rightarrow(\mathbb{F} P)^\sigma$. Given an $\mathbb{F} P$-module $N$, we obtain the coinduced module $\mathfrak{F}_P^G(N):=\mathrm{Hom}_{\mathbb{F} P}(\mathbb{F} G, N)$, which is an $\mathbb{F} G$-module with action
\[(a f)(x):=f(xa),\] for $x, a\in\mathbb{F} G$. Since $G$ is finite, this module is naturally isomorphic to the induced module. Hence, $\mathfrak{F}_P^G(N)$ affords the induced character $\mathrm{Ind}_P^G(\lambda)$, where $\lambda$ is the character of $\mathbb{F} P$ afforded by $N$.
\begin{lemma}\label{lem:sigmacoinduced}
Let $P\leq G$ be a subgroup and let $N$ be an $\mathbb{F} P$-module. Then the map
\[\varphi\colon\mathfrak{F}_P^G(N)^\sigma\rightarrow\mathfrak{F}_P^G(N^\sigma),\] defined by $\varphi(f^\sigma)(a):=f(\iota_0(a))^\sigma$ for all $a\in \mathbb{F} G$ and $f\in\mathfrak{F}_P^G(N)$, is an isomorphism of $\mathbb{F} G$-modules.
\end{lemma}
\begin{proof}
From \prettyref{lem:sigmahoms}, we have an isomorphism of vector spaces
\[\phi\colon \mathfrak{F}_P^G(N)^\sigma=\mathrm{Hom}_{\mathbb{F} P}(\mathbb{F} G, N)^\sigma \rightarrow \mathrm{Hom}_{(\mathbb{F} P)^\sigma}((\mathbb{F} G)^\sigma, N^\sigma)= \mathrm{Hom}_{\mathbb{F} P} ((\mathbb{F} G)^\sigma, N^\sigma)\] such that for $x\in \mathbb{F} G$ and $f\in \mathfrak{F}_P^G(N)^\sigma$, $\phi(f^\sigma)(x^\sigma)=f(x)^\sigma$. Then for $a\in \mathbb{F} G$, we have
\[\phi(af^\sigma)(x^\sigma)=\phi((\iota_0(a)f)^\sigma)(x^\sigma)=(\iota_0(a)f(x))^\sigma=f(x\iota_0(a))^\sigma\] and
\[(a\phi(f^\sigma))(x^\sigma)=\phi(f^\sigma)(x^\sigma\iota_0^\sigma(a))=\phi(f^\sigma)((x\iota_0(a))^\sigma)=f(x\iota_0(a))^\sigma,\] so that $\phi$ is an isomorphism of $\mathbb{F} G$-modules.
Since $\iota_0^\sigma \colon\mathbb{F} G\rightarrow (\mathbb{F} G)^\sigma$ is an isomorphism of $\mathbb{F} G$-modules, we have an isomorphism of $\mathbb{F} G$-modules $\wt{\iota_0}^\sigma\colon \mathrm{End}_{\mathbb{F} P}((\mathbb{F} G)^\sigma, N^\sigma)\rightarrow \mathrm{End}_{\mathbb{F} P}(\mathbb{F} G, N^\sigma)$ defined by $\wt{\iota_0}^\sigma(f)(x):=f(\iota_0^\sigma(x))$. The composition $\wt{\iota_0}^\sigma\circ\phi$ gives the desired isomorphism.
\end{proof}
Considering the endomorphism algebras, we obtain the following corollary:
\begin{corollary}\label{cor:sigmacoinducedend}
Let $P\leq G$ be a subgroup and let $N$ be an $\mathbb{F} P$-module. There is an isomorphism of $\mathbb{F}$-algebras
\[\iota\colon\mathrm{End}_{\mathbb{F} G}(\mathfrak{F}_P^G(N))^\sigma\rightarrow \mathrm{End}_{\mathbb{F} G}(\mathfrak{F}_P^G(N^\sigma))\] given by $\iota(B^\sigma)(\varphi(f^\sigma))=\varphi((Bf)^\sigma)$ for $B\in\mathrm{End}_{\mathbb{F} G}(\mathfrak{F}_P^G(N))$ and $f\in \mathfrak{F}_P^G(N)$.
\end{corollary}
\begin{proof}
From \prettyref{cor:sigmaend} and \prettyref{lem:sigmacoinduced}, we have isomorphisms of $\mathbb{F}$-algebras $\mathrm{End}_{\mathbb{F} G}(\mathfrak{F}_P^G(N))^\sigma\cong \mathrm{End}_{(\mathbb{F} G)^\sigma}((\mathfrak{F}_P^G(N))^\sigma)\cong\mathrm{End}_{(\mathbb{F} G)^\sigma}(\mathfrak{F}_P^G(N^\sigma))=\mathrm{End}_{\mathbb{F} G}(\mathfrak{F}_P^G(N^\sigma))$. Considering the definitions of the maps there, we obtain the result.
\end{proof}
\subsection{$\sigma$-Twists and Endomorphism Algebras}
Keeping the notation of the previous section, we let $\mathcal{H}_G(P,N)$ denote the opposite algebra \[\mathcal{H}_G(P,N):=\mathrm{End}_{\mathbb{F} G}(\mathfrak{F}_P^G(N))^\mathrm{opp}.\]
Assume now that $\mathfrak{F}_P^G(N)$ is a semisimple $\mathbb{F} G$-module, so that $\mathcal{H}_G(P,N)$ is a semisimple $\mathbb{F}$-algebra. Then we denote by $\mathrm{Irr}(\mathbb{F} G|N)\subseteq\mathrm{Irr}(\mathbb{F} G)$ the isomorphism classes of simple submodules of $\mathfrak{F}_P^G(N)$. The map $\mathscr{H}_N\colon\mathrm{Irr}(\mathbb{F} G|N)\rightarrow\mathrm{Irr}(\mathcal{H}_G(P,N))$ given by \[\mathscr{H}_N(S):=\mathrm{Hom}_{\mathbb{F} G}(\mathfrak{F}_P^G(N),S)\] defines a bijection between the constituents of $\mathfrak{F}_P^G(N)$ and the irreducible $\mathcal{H}_G(P,N)$-modules. The $\mathcal{H}_G(P,N)$-module structure for $\mathscr{H}_N(S)$ is given by
\[Bf:=f\circ B\] for $B\in\mathcal{H}_G(P, N)$ and $f\in\mathscr{H}_N(S)$.
If $S$ is a simple $\mathbb{F} G$-submodule of $\mathfrak{F}_P^G(N)$, then $S^\sigma$ is a simple $\mathbb{F} G$-submodule of $\mathfrak{F}_P^G(N^\sigma)$, where we have made the identification $\mathfrak{F}_P^G(N)^\sigma\cong \mathfrak{F}_P^G(N^\sigma)$ from \prettyref{lem:sigmacoinduced}. In this way, we may view $\mathscr{H}_{N^\sigma}(S^\sigma)$ as an $\mathcal{H}_G(P,N^\sigma)$-module.
Further, through the isomorphism $\mathcal{H}_G(P,N)^\sigma\cong \mathcal{H}_G(P, N^\sigma)$ guaranteed by \prettyref{cor:sigmacoinducedend}, we see that $\mathscr{H}_N(S)^\sigma$ may also be viewed as an $\mathcal{H}_G(P, N^\sigma)$-module.
\begin{proposition}\label{prop:sigmasimplebij}
Keeping the notation above, let $S$ be a simple $\mathbb{F} G$-submodule of $\mathfrak{F}_P^G(N)$. Then there is an isomorphism
\[\phi\colon\mathscr{H}_N(S)^\sigma\rightarrow\mathscr{H}_{N^\sigma}(S^\sigma)\] of $\mathcal{H}_G(P,N^\sigma)$-modules.
\end{proposition}
\begin{proof}
From \prettyref{lem:sigmahoms}, there is an isomorphism $\phi\colon\mathscr{H}_N(S)^\sigma\rightarrow\mathscr{H}_{N^\sigma}(S^\sigma)$ as vector spaces over $\mathbb{F}$ given by $\phi(f^\sigma)(m^\sigma):=f(m)^\sigma$ for $f\in \mathscr{H}_N(S)$ and $m\in\mathfrak{F}_P^G(N)$. Note that, again, we have identified $\mathfrak{F}_P^G(N)^\sigma$ with $\mathfrak{F}_P^G(N^\sigma)$ through the isomorphism in \prettyref{lem:sigmacoinduced}. Let $B\in\mathcal{H}_G(P, N)$, $f\in \mathscr{H}_N(S)$, and $m\in\mathfrak{F}_P^G(N)$. Suppressing the notation for the isomorphisms $\varphi$ and $\iota$ in \prettyref{lem:sigmacoinduced} and \prettyref{cor:sigmacoinducedend} and simply identifying the relevant modules and algebras, we may write $B^\sigma (m^\sigma)=(B(m))^\sigma$.
Now, we have
\[\phi(B^\sigma f^\sigma)(m^\sigma) = \phi(f^\sigma\circ B^\sigma)(m^\sigma)=\phi((f\circ B)^\sigma)(m^\sigma)=((f\circ B)(m))^\sigma = f(B(m))^\sigma,\] but also
\[B^\sigma\phi(f^\sigma)(m^\sigma) = (\phi(f^\sigma)\circ B^\sigma)(m^\sigma)=\phi(f^\sigma)(B^\sigma(m^\sigma))=\phi(f^\sigma)(B(m)^\sigma)=f(B(m))^\sigma,\] so that $\phi$ is an isomorphism of $\mathcal{H}_G(P, N^\sigma)$-modules, as stated.
\end{proof}
\section{The Action of Galois Automorphisms on Harish-Chandra Series}\label{sec:gallact}
In \cite{MalleSpathMcKay2}, G. Malle and B. Sp{\"a}th analyze the action of $\mathrm{Aut}(G)$ on Harish-Chandra induced characters. In this section, our aim is to apply the framework introduced in \prettyref{sec:gal twist} to provide analogous statements in the case of $\mathrm{Gal}(\mathbb{Q}_{|G|}/\mathbb{Q})$ acting on characters lying in a given Harish-Chandra series. Throughout, we primarily follow the notation and treatment found in \cite{MalleSpathMcKay2} and \cite[Chapter 10]{carter2}.
For finite groups $Y\leq X$ and $\psi\in\mathrm{Irr}(Y)$ we write $\mathrm{Irr}(X|\psi)$ for the constituents of the induced character $\mathrm{Ind}_Y^X(\psi)$ and $X_\psi$ for the stabilizer in $X$ of $\psi$. Further, for $\chi\in\mathrm{Irr}(X)$, we write $\mathrm{Irr}(Y|\chi)$ for the constituents of the restricted character $\mathrm{Res}_Y^X(\chi):=\chi|_Y$.
Let $G$ be a finite group of Lie type, realized as the group of fixed points $\bG{G}^F$ of a connected reductive algebraic group $\bf{G}$ under a Frobenius endomorphism $F$. Let $\bf{P}$ be an $F$-stable parabolic subgroup of $\bf{G}$ with Levi decomposition $\bf{P}=\bf{L}\bf{U}$, where $\bf{L}$ is an $F$-stable Levi subgroup of $\bf{G}$. Write $L=\bG{L}^F$ and $P=\bG{P}^F$ for the corresponding Levi and parabolic subgroups, respectively, of $G$. For $\lambda\in\mathrm{Irr}(L)$ a cuspidal character, we obtain the Harish-Chandra induced character $R_L^G(\lambda)$ of $G$ via inflation of $\lambda$ to $P$ followed by inducing to $G$, and the irreducible constituents $\mathrm{Irr}\left(G|R_L^G(\lambda)\right)$ of this character make up the $(L,\lambda)$ Harish-Chandra series of $G$. In particular, when $\bG{P}=\bG{B}$ is an $F$-stable Borel subgroup and $L=T$ is a maximally split torus, the irreducible constituents of $R_T^G(\lambda)$ make up the so-called principal series of $G$, which we will be particularly interested in later due to results of \cite{MalleSpathMcKay2}.
It is known from the work of Howlett and Lehrer (see \cite{howlettlehrer80, howlettlehrer83}) that the members of $\mathrm{Irr}\left(G|R_L^G(\lambda)\right)$ are in bijection with characters $\eta\in\mathrm{Irr}(W(\lambda))$, where $W(\lambda)$ is the so-called {relative Weyl group with respect to $\lambda$}. (We note that in fact, the original bijection required a ``twist" by a certain $2$-cycle, but that it was shown later that one could take this $2$-cycle to be trivial - see \cite{geck93}.) Write $R_L^G(\lambda)_\eta$ for the constituent of $R_L^G(\lambda)$ corresponding to $\eta\in\mathrm{Irr}(W(\lambda))$ through this bijection. The goal of this section is to understand how $\sigma\in\mathrm{Gal}(\mathbb{Q}_{|G|}/\mathbb{Q})$ acts on these characters, in terms of this labeling. In what follows, we will utilize the construction in \cite[Chapter 10]{carter2} for the appropriate modules and bijections.
\subsection{The General Setting}
Here we introduce the notation and setting that will be used throughout the remainder of \prettyref{sec:gallact}. In particular, we adapt the more general setting as in \cite[Section 4]{MalleSpathMcKay2} of a finite group $G$ with a split $BN$-pair of characteristic $p$. As in there, we assume that the Weyl group $W=N/(N\cap B)$ is of crystallographic type; $\Phi$ and $\Delta$ denote the root system and simple roots of $W$, respectively; and for $\alpha\in\Phi$, $s_{\alpha}\in W$ is the corresponding reflection.
We fix a standard parabolic subgroup $P=UL$ of $G$ with Levi subgroup $L$ and unipotent radical $U$ and denote by $N(L)$ the group $N(L)=(N_G(L)\cap N)L$. Further, $W(L)=N(L)/L$ is the relative Weyl group of $L$ in $G$. We also fix a cuspidal character $\lambda\in\mathrm{Irr}(L)$, afforded by a representation $\rho$ for $L$, and write $W(\lambda):=N(L)_\lambda/L$. By an abuse of notation, we also denote by $\lambda$ and $\rho$ the irreducible character and representation, respectively, of $P$ obtained by inflation.
Let $\mathbb{F}$ be a splitting field for $G$ which is a finite Galois extension of $\mathbb{Q}$ containing $\sqrt{p}$, and let $\sigma\in\mathrm{Aut}(\mathbb{F})$ be any automorphism of $\mathbb{F}$. Let $M$ be an irreducible left $\mathbb{F} P$-module affording $\lambda$ and let $\rho$ denote the corresponding representation over $\mathbb{F}$. We can construct a module affording $R_L^G(\lambda)=\mathrm{Ind}_P^G(\lambda)$ by taking $\mathfrak{F}(\lambda):=\mathrm{Hom}_{\mathbb{F} P}(\mathbb{F} G, M)$ as in \prettyref{sec:twistcoind}. Then $R_L^G(\lambda^\sigma)$ is afforded by $\mathfrak{F}(\lambda^\sigma)=\mathrm{Hom}_{\mathbb{F} P}(\mathbb{F} G, M^\sigma)$.
Finally, as in \cite[Section 4]{MalleSpathMcKay2}, $\Lambda$ will represent a fixed $N(L)$-equivariant extension map for $L\lhd N(L)$, which exists by \cite{geck93} and \cite[Theorem 8.6]{Lusztig84}. In particular, this means that for each $\lambda\in\mathrm{Irr}(L)$, $\Lambda(\lambda)=:\Lambda_{\lambda}$ is an irreducible character of $N(L)_\lambda$ that extends $\lambda$.
We remark that by definition of the action of $\sigma$ on $\lambda$, we have $N(L)_\lambda=N(L)_{\lambda^\sigma}$. For each $w\in W(\lambda):=N(L)_{\lambda}/L$, we fix a preimage $\dot{w}\in N(L)$.
Now, notice that $(\Lambda_\lambda)^\sigma$ and $\Lambda_{\lambda^\sigma}$ are two extensions of $\lambda^\sigma$ to $N(L)_{\lambda}$ that do not necessarily agree on $N(L)_{\lambda}\setminus L$, so that by Gallagher's theorem (see, for example, \cite[Corollary 6.17]{isaacs}), there is some well-defined linear character $\delta_{\lambda, \sigma}\in\mathrm{Irr}(W(\lambda))$ such that
\[(\Lambda_\lambda)^\sigma=\delta_{\lambda, \sigma}\Lambda_{\lambda^\sigma}.\]
Let $\widetilde{\rho}$ be an extension of $\rho$ affording $\Lambda_\lambda$. Note that $\widetilde{\rho}':=\delta_{\lambda,\sigma}^{-1}\widetilde{\rho}^\sigma$ is then an extension of $\rho^\sigma$ affording $\Lambda_{\lambda^\sigma}$.
\subsection{Action on Basis Elements }
As in \cite[10.1.4-5]{carter2}, there is a basis for $\mathrm{End}_{\mathbb{F} G}(\fl{\lambda})$ comprised of elements $B_{w,\lambda}$ for $w\in W(\lambda)$, where $B_{w,\lambda}\in\e{\lambda}$ is defined by
\[(B_{w,\lambda}f)(x):=\widetilde{\rho}(\dot{w}) f(\dot{w}^{-1} e_U x)\] for all $f\in \fl{\lambda}$ and $x\in \mathbb{F} G$, where \[e_U:=\frac{1}{|U|}\sum_{u\in U} u\] and $U$ is the unipotent radical of $P$.
Similarly, $\e{\lambda^\sigma}$ has a basis comprised of $B_{w,\lambda^\sigma}$ with \[(B_{w,\lambda^\sigma}f)(x):=\widetilde{\rho}'(\dot{w}) f(\dot{w}^{-1} e_U x)\] for all $f\in \fl{\lambda^\sigma}$ and $x\in \mathbb{F} G$.
Recall from \prettyref{sec:gal twist} that we have a $\sigma$-semilinear isomorphism of rings \[-^\sigma\colon\e{\lambda}\rightarrow\e{\lambda^\sigma},\] where we identify $\fl{\lambda}^\sigma$ with $\fl{\lambda^\sigma}$ and $\e{\lambda}^\sigma$ with $\e{\lambda^\sigma}$ according to \prettyref{lem:sigmacoinduced} and \prettyref{cor:sigmacoinducedend}. Throughout, we will continue to make these identifications and to suppress the notation of the isomorphisms $\varphi$ and $\iota$ introduced there. We observe in the next lemma how the basis elements $B_{w,\lambda}$ and $B_{w,\lambda^\sigma}$ are related under this map.
\begin{lemma}\label{lem:iotaBw}
For each $w\in W(\lambda)$, we have
\[B_{w,\lambda}^\sigma=\delta_{\lambda,\sigma}(w)B_{w,\lambda^\sigma}.\]
\end{lemma}
\begin{proof}
Let $f\in\fl{\lambda}$ and $a\in\mathbb{F} G$. Using the descriptions above, \prettyref{cor:sigmacoinducedend}, and \prettyref{lem:sigmacoinduced}, we see \[B_{w,\lambda}^\sigma f^\sigma(a)=(B_{w,\lambda}f)^\sigma(a)=B_{w, \lambda} f(\iota_0(a))^\sigma=\left(\wt{\rho}(\dot{w})f(\dot{w}^{-1}e_U\iota_0(a))\right)^\sigma.\] Further, since $\dot{w}^{-1}e_U\in\mathbb{Q} G$, and hence is fixed by $\iota_0$, this means
\[B_{w,\lambda}^\sigma f^\sigma(a)=\delta_{\lambda,\sigma}({w})\wt{\rho}'(\dot{w})f(\iota_0(\dot{w}^{-1}e_Ua))^\sigma=\delta_{\lambda,\sigma}({w})\wt{\rho}'(\dot{w})f^\sigma(\dot{w}^{-1}e_Ua)=\delta_{\lambda,\sigma}({w})(B_{w,\lambda^\sigma}f^\sigma)(a),\] which proves the statement.
\end{proof}
We define $\Omega, p_{\alpha,\lambda}, \Phi_{\lambda}, \Delta_{\lambda}, R(\lambda),$ and $C(\lambda)$ as in \cite[4.B]{MalleSpathMcKay2}, so that $W(\lambda)=C(\lambda)\ltimes R(\lambda)$ and $R(\lambda)=\langle s_{\alpha} \mid \alpha\in \Phi_\lambda\rangle$ is a Weyl group with simple system $\Delta_{\lambda}$.
To be more precise, $\Omega$ is the set
\[\Omega=\{\alpha\in\Phi\setminus\Phi_L\mid w(\Delta_L\cup\{\alpha\})\subseteq\Delta\hbox{ for some $w\in W$ and }(w_0^Lw_0^{\alpha})^2=1\}.\] Here $w_0^L, w_0^\alpha$ are the longest elements in $W(L)$ and $\langle W(L), s_\alpha\rangle$, respectively, and $\Phi_L\subseteq\Phi$ is the root system of $W(L)$ with simple system $\Delta_L\subseteq\Delta$. Then for $\alpha\in\Omega$, letting $L_\alpha$ denote the standard Levi subgroup of $G$ with simple system $\Delta_L\cup\{\alpha\}$, $L$ is a standard Levi subgroup of $L_\alpha$ and $p_{\alpha,\lambda}\geq1$ is defined to be the ratio between the degrees of the two constituents of $R_L^{L_\alpha}(\lambda)$. Then $\Phi_{\lambda}$ is the subset of $\Omega$ consisting of $\alpha$ such that $s_\alpha\in W(\lambda)$ and $p_{\alpha,\lambda}\neq1$. This is a root system with simple roots $\Delta_\lambda\subseteq\Phi_\lambda\cap\Phi^+$. Then $R(\lambda)=\langle s_{\alpha} \mid \alpha\in \Phi_\lambda\rangle$ is a Weyl group with simple system $\Delta_{\lambda}$, $C(\lambda)$ is the stabilizer of $\Delta_\lambda$ in $W(\lambda)$, and $W(\lambda)=C(\lambda)\ltimes R(\lambda)$ .
We see immediately from the definitions that:
\begin{lemma}
For $\alpha\in\Omega$, $p_{\alpha,\lambda}=p_{\alpha,\lambda^\sigma}$. Hence, we have $\Phi_{\lambda}=\Phi_{\lambda^\sigma}$ and $\Delta_{\lambda}=\Delta_{\lambda^\sigma}$. Further, $R(\lambda^\sigma)=R(\lambda)$ and $C(\lambda^\sigma)=C(\lambda)$.
\end{lemma}
Now, for each $\alpha\in\Delta_{\lambda}$, let $\epsilon_{\alpha,\lambda}\in\{\pm1\}$ be as in \cite[(4.3)]{MalleSpathMcKay2}. Namely, we have
\[B_{s_\alpha,\lambda}^2=\mathrm{Ind}(s_\alpha)\cdot \mathrm{id}+\epsilon_{\alpha,\lambda}\frac{p_{\alpha,\lambda}-1}{\sqrt{\mathrm{Ind}(s_{\alpha})p_{\alpha,\lambda}}}B_{s_{\alpha},\lambda},\] and similarly
\[B_{s_\alpha,\lambda^\sigma}^2=\mathrm{Ind}(s_\alpha)\cdot \mathrm{id}+\epsilon_{\alpha,\lambda^\sigma}\frac{p_{\alpha,\lambda}-1}{\sqrt{\mathrm{Ind}(s_{\alpha})p_{\alpha,\lambda}}}B_{s_{\alpha},\lambda^\sigma}.\]
Here, for any $w\in W$, we write $\mathrm{Ind}(w):=|U_0\cap U_0^{w_0w}|$ where $U_0$ is the unipotent radical of the subgroup $B$ and $w_0$ is the longest element of $W$. We remark that since $p_{\alpha,\lambda}$ and $\mathrm{Ind}(w)$ are powers of $p$ (see \cite[Theorem 10.5.5]{carter2}), the square root $\sqrt{\mathrm{Ind}(s_{\alpha})p_{\alpha,\lambda}}$ lies in $\mathbb{F}$ by our assumption $\sqrt{p}\in\mathbb{F}$.
Using this, we may further relate $\epsilon_{\alpha,\lambda}$ to $\epsilon_{\alpha,\lambda^\sigma}$:
\begin{lemma}\label{lem:epsilons}
For each $\alpha\in\Delta_{\lambda}$, \[\displaystyle{\frac{\epsilon_{\alpha,\lambda}}{\epsilon_{\alpha,\lambda^\sigma}}=\delta_{\lambda,\sigma}(s_\alpha)\frac{\left(\sqrt{\mathrm{Ind}(s_\alpha)p_{\alpha,\lambda}}\right)^\sigma}{\sqrt{\mathrm{Ind}(s_\alpha)p_{\alpha,\lambda}}}}.\]
\end{lemma}
\begin{proof}
Notice that from \prettyref{lem:iotaBw} and the $\sigma$-semilinearity of $-^\sigma\colon\e{\lambda}\rightarrow\e{\lambda^\sigma}$, we see
\[(B_{s_{\alpha},\lambda}^2)^\sigma=\mathrm{Ind}(s_{\alpha})\cdot\mathrm{id}+\epsilon_{\alpha,\lambda}\frac{p_{\alpha,\lambda}-1}{\left(\sqrt{\mathrm{Ind}(s_{\alpha})p_{\alpha,\lambda}}\right)^\sigma}B_{s_\alpha,\lambda}^\sigma\]
\[=\mathrm{Ind}(s_{\alpha})\cdot\mathrm{id}+\epsilon_{\alpha,\lambda}\frac{p_{\alpha,\lambda}-1}{\left(\sqrt{\mathrm{Ind}(s_{\alpha})p_{\alpha,\lambda}}\right)^\sigma}\delta_{\lambda,\sigma}(s_{\alpha})B_{s_\alpha,\lambda^\sigma}\] and this is the same as
\[(\delta_{\lambda,\sigma}(s_{\alpha}))^2B_{s_\alpha,\lambda^\sigma}^2=(\delta_{\lambda,\sigma}(s_{\alpha}))^2\mathrm{Ind}(s_\alpha)\cdot \mathrm{id}+(\delta_{\lambda,\sigma}(s_{\alpha}))^2\epsilon_{\alpha,\lambda^\sigma}\frac{p_{\alpha,\lambda}-1}{\sqrt{\mathrm{Ind}(s_{\alpha})p_{\alpha,\lambda}}}B_{s_{\alpha},\lambda^\sigma}.\]
Since $s_{\alpha}^2=1$ and $\delta_{\lambda,\sigma}$ is linear, this means $\epsilon_{\alpha,\lambda}\frac{p_{\alpha,\lambda}-1}{\left(\sqrt{\mathrm{Ind}(s_{\alpha})p_{\alpha,\lambda}}\right)^\sigma}\delta_{\lambda,\sigma}(s_\alpha)=\epsilon_{\alpha,\lambda^\sigma}\frac{p_{\alpha,\lambda}-1}{\sqrt{\mathrm{Ind}(s_{\alpha})p_{\alpha,\lambda}}}$, completing the proof.
\end{proof}
We now wish to determine the image under $\sigma$ of the alternate basis $T_{w,\lambda}$ for $\e{\lambda}$. For $\alpha\in\Delta_{\lambda}$, the element $T_{s_\alpha,\lambda}$ is defined by $T_{s_\alpha,\lambda}:=\epsilon_{\alpha,\lambda}\sqrt{\mathrm{Ind}(s_\alpha)p_{\alpha,\lambda}}B_{s_\alpha,\lambda}$. For $w\in C(\lambda)$, $T_{w,\lambda}:=\sqrt{\mathrm{Ind}(w)}B_{w,\lambda}$, and for $w=w_1s_{\alpha_1}\cdots s_{\alpha_r}$ where $w_1\in C(\lambda)$ and $s_{\alpha_1}\cdots s_{\alpha_r}$ is a reduced expression in $R(\lambda)$, $T_{w,\lambda}$ is defined by
\begin{equation}\label{eq:Tw}
T_{w,\lambda}:=T_{w_1,\lambda}T_{s_{\alpha_1},\lambda}\cdots T_{s_{\alpha_r},\lambda}.\end{equation} We remark that this definition is independent of choice of reduced expression. (See \cite[Proposition 10.8.2]{carter2}.)
This yields the following:
\begin{proposition}\label{prop:iotaTw}
Let $w\in W(\lambda)$ have the form $w=w_1w_2$ for $w_1\in C(\lambda)$ and $w_2\in R(\lambda)$. Then
\[T_{w,\lambda}^\sigma=\frac{\sqrt{\mathrm{Ind}(w_1)}^\sigma}{\sqrt{\mathrm{Ind}(w_1)}}\delta_{\lambda,\sigma}(w_1)T_{w,\lambda^\sigma}.\]
\end{proposition}
\begin{proof}
We may write $w_2\in R(\lambda)$ as a reduced expression $w_2=s_{\alpha_1}\cdots s_{\alpha_r}$ as above. First note that by \prettyref{lem:iotaBw} and $\sigma$-semilinearity,
\[T_{w_1,\lambda}^\sigma = (\sqrt{\mathrm{Ind}(w_1)}B_{w_1,\lambda})^\sigma = \sqrt{\mathrm{Ind}(w_1)}^\sigma B_{w_1,\lambda}^\sigma = \sqrt{\mathrm{Ind}(w_1)}^\sigma\delta_{\lambda,\sigma}(w_1)B_{w_1,\lambda^\sigma},\] and for each $i$,
\[T_{s_{\alpha_i},\lambda}^\sigma= \left(\epsilon_{\alpha_i,\lambda}\sqrt{\mathrm{Ind}(s_{\alpha_i})p_{\alpha_i,\lambda}}B_{s_{\alpha_i},\lambda}\right)^\sigma =\epsilon_{\alpha_i,\lambda}\left(\sqrt{\mathrm{Ind}(s_{\alpha_i})p_{\alpha_i,\lambda}}\right)^\sigma B_{s_{\alpha_i},\lambda}^\sigma \]\[=\epsilon_{\alpha_i,\lambda}\left(\sqrt{\mathrm{Ind}(s_{\alpha_i})p_{\alpha_i,\lambda}}\right)^\sigma \delta_{\lambda,\sigma}(s_{\alpha_i})B_{s_{\alpha_i},\lambda^\sigma} . \]
But notice $\delta_{\lambda,\sigma}(w_1)T_{w_1,\lambda^\sigma}=\delta_{\lambda,\sigma}(w_1)\sqrt{\mathrm{Ind}(w_1)}B_{w_1,\lambda^\sigma},$ so that \[T_{w_1,\lambda}^\sigma = \frac{\sqrt{\mathrm{Ind}(w_1)}^\sigma}{\sqrt{\mathrm{Ind}(w_1)}}\delta_{\lambda,\sigma}(w_1)T_{w_1,\lambda^\sigma}. \]
Moreover,
$\delta_{\lambda,\sigma}(s_{\alpha_i})T_{s_{\alpha_i},\lambda^\sigma}=\epsilon_{\alpha_i,\lambda^\sigma}\sqrt{\mathrm{Ind}(s_{\alpha_i})p_{\alpha_i,\lambda}}\delta_{\lambda,\sigma}(s_{\alpha_i})B_{s_{\alpha_i},\lambda^\sigma},$ and hence
\[T_{s_{\alpha_i},\lambda}^\sigma=\frac{\epsilon_{\alpha_i,\lambda}\left(\sqrt{\mathrm{Ind}(s_{\alpha_i})p_{\alpha_i,\lambda}}\right)^\sigma}{\epsilon_{\alpha_i,\lambda^\sigma}\sqrt{\mathrm{Ind}(s_{\alpha_i})p_{\alpha_i,\lambda}}}\delta_{\lambda,\sigma}(s_{\alpha_i})T_{s_{\alpha_i},\lambda^\sigma}=T_{s_{\alpha_i},\lambda^\sigma},\] where the last equality is by \prettyref{lem:epsilons}. Finally, since $\sigma$ is an isomorphism of rings, the statement follows from \eqref{eq:Tw}.
\end{proof}
For $w=w_1w_2\in W(\lambda)$ with $w_1\in C(\lambda)$ and $w_2\in R(\lambda)$, we define $r_\sigma(w):=\frac{\sqrt{\mathrm{Ind}(w_1)}^\sigma}{\sqrt{\mathrm{Ind}(w_1)}}$. As may be inferred from \prettyref{prop:iotaTw}, the term $r_\sigma(w)$ will play an important role. Notice that $r_\sigma(w)\in\{\pm1\}$. Indeed, for $z\in\mathbb{Q}$, $\sqrt{z}^\sigma$ must be a solution to the polynomial $t^2-z=0$, and hence $\sqrt{z}^\sigma\in\{\pm\sqrt{z}\}$.
In some cases, it is clear that $r_\sigma\colon W(\lambda)\rightarrow\{\pm1\}$ defines a homomorphism, and hence a linear character of $W(\lambda)$. When this is the case, note that $r_\sigma(w)=r_\sigma(w)^{-1}=r_\sigma({w^{-1}})$ and that $r_\sigma$ is simply a character of $C(\lambda)\cong W(\lambda)/R(\lambda)$ inflated to $W(\lambda)$. In particular, this holds in the case that $G$ is a group of Lie type, by \cite[Section 2.9]{carter2} (see \prettyref{lem:rw} below and the discussion preceding it).
We remark that in particular, if $R(\lambda)\leq \ker(\delta_{\lambda,\sigma})$, then
$T_{w,\lambda}^\sigma=r_\sigma(w)\delta_{\lambda,\sigma}(w)T_{w,\lambda^\sigma}.$
However, more generally, we may inflate the linear character $\delta_{\lambda,\sigma}|_{C(\lambda)}$ of $C(\lambda)\cong W(\lambda)/R(\lambda)$ to another linear character $\delta_{\lambda,\sigma}'$ of $W(\lambda)$ that is trivial on $R(\lambda)$. This gives the following corollary:
\begin{corollary}\label{cor:iotaTw}
Let $w\in W(\lambda)$. Then
\[T_{w,\lambda}^\sigma=r_\sigma(w)\delta'_{\lambda,\sigma}(w)T_{w,\lambda^\sigma}.\]
\end{corollary}
\subsection{Action on Modules}
We now turn our attention to $\e{\lambda}$-modules and prove an analogue to \cite[Proposition 4.5]{MalleSpathMcKay2}.
\begin{proposition}\label{prop:moduleact}
Let $\eta$ be an irreducible character of $\e{\lambda}$. Then the irreducible character $\eta^\sigma$ of $\e{\lambda^\sigma}$ satisfies
\[\eta^\sigma(T_{w,\lambda^\sigma})={r}_{\sigma}(w)\delta'_{\lambda,\sigma}(w^{-1})\eta(T_{w,\lambda})^\sigma\]
for every $w\in W(\lambda)$.
\end{proposition}
\begin{proof}
From \prettyref{lem:sigmamatrixrep} and \prettyref{cor:iotaTw}, we have
\[\eta(T_{w,\lambda})^\sigma=\eta^\sigma(T_{w,\lambda}^\sigma)=\eta^\sigma(r_\sigma(w)\delta'_{\lambda,\sigma}(w)T_{w,\lambda^{\sigma}}).\] Since $r_\sigma(w)\in\{\pm1\}$ and $\delta'_{\lambda,\sigma}$ is a linear character, this yields the statement.
\end{proof}
\subsection{The Generic Algebra}
Here we discuss and set the notation for the generic algebras and specializations that yield the bijection $\mathfrak{f}\colon\mathrm{Irr}(\e{\lambda})\rightarrow\mathrm{Irr}(W(\lambda))$ in Howlett--Lehrer theory. Like before, we follow largely the notation and discussion in \cite[4.D]{MalleSpathMcKay2}. Let $\textbf{u}=(u_\alpha\colon\alpha\in\Delta_\lambda)$ be indeterminates such that $u_\alpha=u_\beta$ if and only if $\alpha$ is $W(\lambda)$-conjugate to $\beta$. Let $A_0:=\mathbb{C}[\bf{u},\bf{u}^{-1}]$ and let $K$ be an algebraic closure of the quotient field of $A_0$ and $A$ the integral closure of $A_0$ in $K$. We define the generic algebra $\mathcal{H}$ to be a free $A$-module with basis $\{a_w\colon w\in W(\lambda)\}$ equipped with the $A$-bilinear associative multiplication satisfying the properties as in \cite[5.D]{MalleSpathMcKay2}.
We denote by $\mathcal{H}^K$ the $K$-algebra $K\otimes_A \mathcal{H}$, and for any ring homomorphism $f\colon A\rightarrow\mathbb{C}$, we obtain the $\mathbb{C}$-algebra $\mathcal{H}^f:=\mathbb{C}\otimes_A\mathcal{H}$ with basis $\{1\otimes a_w\colon w\in W(\lambda)\}$. By \cite[Proposition 4.7]{howlettlehrer83}, when $\mathcal{H}^f$ is semisimple, this yields a bijection $\eta\mapsto \eta^f$ between $K$-characters of simple $\mathcal{H}^K$-modules and characters of simple $\mathcal{H}^f$-modules, where $\eta^f(1\otimes a_w)=f(\eta(a_w))$.
In particular, the morphisms $f$ and $g \colon A\rightarrow \mathbb{C}$ will denote extensions of $f_0, g_0\colon A_0\rightarrow\mathbb{C}$ given by $f_0(u_\alpha)=p_{\alpha,\lambda}$ and $g_0(u_\alpha)=1$, respectively. By \cite[Lemma 4.2]{howlettlehrer83}, these extensions exist and yield isomorphisms $\mathcal{H}^f\cong \mathrm{End}_G(\fl\lambda)$ and $\mathcal{H}^g\cong \mathbb{C} W(\lambda)$ via $1\otimes a_w\mapsto T_{w,\lambda}$ and $1\otimes a_w\mapsto w,$ respectively.
\subsubsection{Further Remarks on the Generic Algebra in the Case of the Principal Series}\label{sec:genericsubalg}
For this section only, we assume that $L=T$ , so that $\lambda\in\mathrm{Irr}(T)$ is a linear character and we are in the case of the principal series. We fix here some notation for later use.
Let $\mathcal{H}_0$ denote the subalgebra of $\mathcal{H}$ generated by $\{a_w\mid w\in R(\lambda)\}$. Then $\mathcal{H}_0$ is a generic algebra corresponding to the Coxeter group $R(\lambda)$. In later sections, we will be interested in understanding the analogy between the Clifford theory from $R(\lambda)$ to $W(\lambda)$ and the characters of $\mathcal{H}_0$ in relation to $\mathcal{H}$. Here we record some of the essentials.
First, by \cite[Theorem 3.7]{HowlettKilmoyer}, any irreducible character $\psi$ of $\mathcal{H}_0^K$ extends to an irreducible character $\tau$ of $C(\lambda)_{\psi}\mathcal{H}_0^K$, where $C(\lambda)_{\psi}\mathcal{H}_0^K$ is the tensor product $KC(\lambda)_\psi\otimes\mathcal{H}_0^K$ as defined in \cite{HowlettKilmoyer}. Further, by \cite[(3.11) and Lemma 3.12]{HowlettKilmoyer}, every irreducible character of $\mathcal{H}^K$ is of the form $\tau^{\mathcal{H}}$, where $\tau$ is an extension of some $\psi\in\mathrm{Irr}(\mathcal{H}_0^K)$ to $C(\lambda)_{\psi}\mathcal{H}_0^K$ and
\[\tau^\mathcal{H}(a_w):=|C(\lambda)_\psi|^{-1}\sum \tau(a_{cwc^{-1}})\] where the sum is over $c\in C(\lambda)$ satisfying $cwc^{-1}\in C(\lambda)_\psi R(\lambda)$.
\subsection{Alternate Twists and Action on the Harish-Chandra Series}
Recall that the groups $W(\lambda)$ and $W(\lambda^\sigma)$ are the same, as $N(L)_\lambda=N(L)_{\lambda^\sigma}$. We also see that $\e{\lambda}$ and $\e{\lambda^\sigma}$ are isomorphic as vector spaces, simply by mapping each basis element $T_{w,\lambda}$ to the corresponding basis element $T_{w,\lambda^\sigma}$ and extending linearly. However, recalling that the structure constants for $\e{\lambda}$ and $\e{\lambda^\sigma}$ are the same (depending only on the numbers $p_{\alpha,\lambda}=p_{\alpha,\lambda^\sigma}$), we see that in fact, this yields an isomorphism of $\mathbb{F}$-algebras.
\begin{lemma}
The map $\e{\lambda}\rightarrow \e{\lambda^\sigma}$, defined by $T_{w,\lambda}\mapsto T_{w,\lambda^\sigma}$ and extending linearly, is an isomorphism of $\mathbb{F}$-algebras.
\end{lemma}
With this in place, given a matrix representation $\mathfrak{X}\colon \e{\lambda}\rightarrow \mathrm{Mat}_d(\mathbb{F})$ for $\e{\lambda}$, we immediately obtain a corresponding representation \[\overline{\mathfrak{X}}\colon\e{\lambda^\sigma}\rightarrow \mathrm{Mat}_d(\mathbb{F})\] via $\overline{\mathfrak{X}}(T_{w,\lambda^\sigma})=\mathfrak{X}(T_{w,\lambda})$. Further, since the structure constants are rational, we obtain a representation
\[\mathfrak{X}^{(\sigma)}\colon\e{\lambda}\rightarrow \mathrm{Mat}_d(\mathbb{F})\] defined by $\mathfrak{X}^{(\sigma)}(T_{w,\lambda})=\mathfrak{X}(T_{w,\lambda})^\sigma$, and extending linearly. Together, this also yields a representation $\overline{\mathfrak{X}}^{(\sigma)}$ of $\e{\lambda^\sigma}$ satisfying that $\overline{\mathfrak{X}}^{(\sigma)}(T_{w,\lambda^\sigma})=\mathfrak{X}(T_{w,\lambda})^\sigma$.
If $\eta$ is a character of $\e{\lambda}$ afforded by $\mathfrak{X}$, we denote by $\overline{\eta}, \eta^{(\sigma)},$ and $\overline{\eta}^{(\sigma)}$ the corresponding characters afforded by $\overline{\mathfrak{X}}, \mathfrak{X}^{(\sigma)},$ and $\overline{\mathfrak{X}}^{(\sigma)}$, respectively. Note then that by \prettyref{prop:moduleact},
\[\overline{\eta}^{(\sigma)}(T_{w,\lambda^\sigma}) = \eta(T_{w,\lambda})^\sigma={r}_{\sigma}(w)\delta'_{\lambda,\sigma}(w)\eta^\sigma(T_{w,\lambda^\sigma}).\]
Now let $\mathfrak{f}\colon\mathrm{Irr}(\e{\lambda})\rightarrow\mathrm{Irr}(W(\lambda))$ denote the bijection induced by the standard specializations of the generic algebra discussed above. Taking into consideration \prettyref{prop:moduleact}, we are interested in how this bijection behaves under the action of $\sigma$ on the values of $\eta\in\mathrm{Irr}(\e{\lambda})$. It is clear that this bijection does not preserve fields of values, for example from the fact that the field of values for a Weyl group of type $E_8$ is $\mathbb{Q}$ but for the Hecke algebra $\mathrm{End}_G(1_B^G)$ is $\mathbb{Q}(\sqrt{q})$.
Hence we define a non-standard bijection $\mathrm{Irr}(W(\lambda))\rightarrow\mathrm{Irr}(W(\lambda^\sigma))$ induced by $\sigma$ as follows. We denote by $\gamma^{(\sigma)}$ the irreducible character of $W(\lambda^\sigma)$ such that $\gamma^{(\sigma)}=\mathfrak{f}\left(\overline{\eta}^{(\sigma)}\right)$, where $\eta\in\mathrm{Irr}(\e{\lambda})$ is the character such that $\gamma=\mathfrak{f}(\eta)$. That is, this new action of $\sigma$ on $\mathrm{Irr}(W(\lambda))$ is such that
\[\left(\mathfrak{f}(\eta)\right)^{(\sigma)}=\mathfrak{f}(\overline{\eta}^{(\sigma)}).\]
\prettyref{prop:moduleact}, together with the above discussion, yields a description of how the Galois group acts on constituents of $R_L^G(\lambda)$, which is the Galois action analogue to \cite[Theorem 4.6]{MalleSpathMcKay2}. (Note that since $\mathbb{Q}\subseteq\mathbb{F}\subseteq\mathbb{C}$ is a splitting field, the complex characters of $W(\lambda), \mathrm{End}_{\mathbb{C} G}(\fl{\lambda}\otimes_{\mathbb{F}}\mathbb{C}),$ and $G$ can be considered naturally as $\mathbb{F}$-characters of $W(\lambda), \mathrm{End}_{\mathbb{F} G}(\fl{\lambda}),$ and $G$.) We write $R_L^G(\lambda)_\gamma$ to denote the constituent of $R_L^G(\lambda)$ afforded by a module $S$ such that $\mathrm{Hom}_{\mathbb{F} G}(\mathfrak{F}(\lambda), S)$ affords $\eta$ and $\mathfrak{f}(\eta)=\gamma$.
\begin{theorem}\label{thm:GaloisAct}
Let $\sigma\in\mathrm{Aut}(\mathbb{F})$ and let $\gamma\in \mathrm{Irr}(W(\lambda))$. Then
\[\left(R_L^G(\lambda)_\gamma\right)^\sigma=R_L^G(\lambda^\sigma)_{\gamma'},\] where $\gamma'\in\mathrm{Irr}(W(\lambda))=\mathrm{Irr}(W(\lambda^\sigma))$ is defined by $\gamma'(w)=r_\sigma(w)\delta'_{\lambda,\sigma}(w^{-1})\gamma^{(\sigma)}(w)$ for each $w\in W(\lambda)$.
\end{theorem}
We remark that when $R(\lambda)\leq\ker(\delta_{\lambda,\sigma})$, $\gamma'$ is just the character $\gamma'=r_\sigma\delta_{\lambda,\sigma}^{-1}\gamma^{(\sigma)}$.
\section{The Proof of Theorem A
}\label{sec:SN2S}
As an application of our analysis in \prettyref{sec:gallact}, we complete the proof of Navarro's self-normalizing Sylow $2$-subgroup conjecture (see \prettyref{conj:mainprob})
by proving that the remaining simple groups are SN2S-Good, in the sense described in \cite{SchaefferFrySN2S1} (see also below). We now finally specialize our choice of $\sigma$ to be the Galois automorphism relevant for \prettyref{conj:mainprob}.
\begin{center}
\fbox{
\parbox{6in}{\begin{center}
For the remainder of the article, $\sigma$ will always denote the Galois automorphism in \prettyref{conj:mainprob}, and $\mathbb{K}$ will always denote the fixed field of $\sigma$.\end{center}}
}
\end{center}
\subsection{The Conditions for Simple Groups}\label{sec:SimpleStatements}
Here we recall the definitions and main reduction theorem from \cite{SchaefferFrySN2S1}.
\begin{condition}\label{cond:conjIF}
Let $G$ be a finite quasisimple group with center $Z:=Z(G)$ and $Q$ a finite $2$-group acting on $G$ as automorphisms. Assume $P/Z\in \mathrm{Syl}_2(G/Z)$ is $Q$-invariant and $C_{N_G(P)/P}(Q)=1$. Then for any $Q$-invariant, $\sigma$-fixed $\lambda\in\mathrm{Irr}(Z)$, we have $\chi^\sigma=\chi$ for any $Q$-invariant $\chi\in\mathrm{Irr}_{2'}(G|\lambda)$.
\end{condition}
We note that the condition $C_{N_G(P)/P}(Q)=1$ is equivalent to $GQ/Z$ having a self-normalizing Sylow $2$-subgroup (see, for example, \cite[Lemma 2.1 (ii)]{NavarroTiepTurull2007}) and that to prove \prettyref{cond:conjIF}, it suffices to prove the statement for the Schur cover of $G/Z$.
\begin{condition}\label{cond:conjFI}
Let $S$ be a finite nonabelian simple group, $Q$ a finite $2$-group acting on $S$ as automorphisms, and $P\in \mathrm{Syl}_2(S)$ be $Q$-invariant. If every $Q$-invariant $\chi\in\mathrm{Irr}_{2'}(S)$ is fixed by $\sigma$, then $C_{N_S(P)/P}(Q)=1$.
\end{condition}
\begin{definition}{\cite[Definition 1]{SchaefferFrySN2S1}}\label{def:Goodness}
Let $S$ be a finite nonabelian simple group. We will say $S$ is ``SN2S-Good" if $S$ satisfies \prettyref{cond:conjFI} and $G$ satisfies \prettyref{cond:conjIF} whenever $G$ is a quasisimple group with $G/Z(G)\cong S$.
\end{definition}
\begin{theorem}{\cite[Theorem 3.7]{SchaefferFrySN2S1}}\label{thm:NTT62anal}
Let $G$ be a finite group and $P\in\mathrm{Syl}_2(G)$. Assume that every finite nonabelian simple group involved in $G$ is SN2S-Good (see \prettyref{def:Goodness}). Then $P=N_G(P)$ if and only if every $\chi\in\mathrm{Irr}_{2'}(G)$ is fixed by $\sigma$.
\end{theorem}
The results of \cite{SchaefferFrySN2S1} and \cite{SFTaylorTypeA} imply that to prove \prettyref{conj:mainprob}, it suffices to show that when $q$ is odd, the simple groups $E_6^\pm(q)$, $E_8(q)$, $P\Omega_n^\pm(q)$ for $n\geq 7$, and $PSp_{2n}(q)$ for $n\geq2$ are SN2S-Good.
Here $E_6^+(q)$ means the untwisted simple group $E_6(q)$, $E_6^-(q)$ denotes the twisted simple group $\tw{2}{E}_6(q),$ and $P\Omega_n^\pm(q)$ is taken to mean $P\Omega_{n}(q)$ in the case $n$ is odd.
Sylow $2$-normalizers for finite simple groups are considered in \cite{kondratiev2005}. From this, we see that if $S$ is one of the simple groups $PSp_{2n}(q)$ with $q\equiv\pm1\pmod8$, $P\Omega_n^\pm(q)$ with $q$ odd, or $E_8(q)$ with $q$ odd, then $S$ has a self-normalizing Sylow 2-subgroup, and hence this will hold for the Schur cover $G$ as well. In these cases, to prove SN2S-Goodness, it therefore suffices to show that every irreducible character of $G$ of odd degree is fixed by $\sigma$.
If $S$ is $PSp_{2n}(q)$ with $q\equiv\pm3\mod8$, there is not a self-normalizing Sylow $2$-subgroup. Further, in this case $q$ must be an odd power of an odd prime, so a $2$-group of automorphisms $Q$ must be contained in $PCSp_{2n}(q)$, which does have a self-normalizing Sylow 2-subgroup (see, for example, \prettyref{lem:CSpSN2S} below). Further, \cite[Proposition 4.8]{Malle08} and the discussion preceding it yields that the $PCSp_{2n}(q)$-invariant characters with odd degree are exactly the unipotent characters of odd degree and that since the Schur multiplier is size $2$, odd-degree characters of the Schur cover can be considered as characters of $S$. Thus in this case it suffices to show that every unipotent character of odd degree of $S$ is fixed by $\sigma$ and that there exists some non-unipotent $\chi\in\mathrm{Irr}_{2'}(S)$ which is not fixed by $\sigma$. The first of these conditions is verified by the following proposition.
\begin{proposition}\label{prop:unipsfixed}
Let $G$ be a finite group of Lie type with odd defining characteristic and no component of Suzuki or Ree type. Then every unipotent character of odd degree is realizable over $\mathbb{Q}$.
\end{proposition}
\begin{proof}
By \cite[Proposition 7.4]{MalleSpathMcKay2}, every unipotent character of odd degree lies in the principal series of $G$. But according to \cite[Theorem 2.9]{BensonCurtis} (see also \cite{BensonCurtisCorrection}), every member of $\mathrm{Irr}_{2'}(G|R_T^G(1))$ can be realized over $\mathbb{Q}$.
\end{proof}
\subsection{Toward SN2S-Goodness}\label{sec:oddchars}
We collect here some additional observations that will be useful for determining SN2S-Goodness. In what follows, we let $\bG{G}$ be a connected reductive algebraic group and $G=\textbf{G}^F$ be the group of fixed points of $\textbf{G}$ under a Frobenius endomorphism $F$ defined over $\mathbb{F}_q$. We will at times do computations in $G$ using the Chevalley generators and relations, as in \cite[Theorem 1.12.1]{gorensteinlyonssolomonIII}. In particular, $x_\alpha(t), n_\alpha(t),$ and $h_\alpha(t)$ are as defined there.
The dual of $G$ is $G^\ast=\textbf{G}^{\ast F^\ast}$, where $\bG{G}^\ast$ is dual to $\bG{G}$ as in, for example, \cite[Section 4.2]{carter2}. Let $\textbf{T}$ and $\textbf{B}$ be an $F$-stable maximal torus and Borel subgroup of $\textbf{G}$, respectively, with $\textbf{T}\leq\textbf{B}$ fixed as in \cite[Section 2.B]{MalleSpathMcKay2} and let $\bG{T}^\ast$ be an $F^\ast$-stable maximal torus of $\bG{G}^\ast$ dual to $\bG{T}$. Write $T=\textbf{T}^F$ and $T^\ast=(\bG{T}^\ast)^{F^\ast}$. We write $W=\bg{W}^F$, where $\bg{W}=N_\bg{G}(\bg{T})/\bg{T}$, and similarly for $W^\ast$.
This duality induces an isomorphism $\mathrm{Irr}(T)\rightarrow T^\ast$. We begin by reconciling our understanding of $W(\lambda)$ in the case $\lambda\in\mathrm{Irr}(T)$ with a Weyl group, using this isomorphism. We remark that given $s\in T^\ast$, the only nontrivial action of $F^\ast$ on the Weyl group $\bg{W}(s)$ of $C_{\bG{G}^\ast}(s)$ is given by the automorphism induced by that of $F^\ast$ on $\bg{W}^\ast$, since $s$ is contained in a maximally split torus.
\begin{lemma}\label{lem:W(s)}
Let $\lambda\in\mathrm{Irr}(T)$ and let $s\in T^\ast$ correspond to $\lambda$ in the sense of \cite[Proposition 4.4.1]{carter2}. Denote by $W(s)$ and $W^\circ(s)$ the fixed points of the Weyl groups of $C_{\bG{G}^\ast}(s)$ and $C_{\bG{G}^\ast}^\circ(s)$, respectively, under $F^\ast$. Then
\begin{enumerate}[label={(\arabic*)}]
\item $W(\lambda)$ is isomorphic to $W(s)$.
\item If $\bG{G}$ is simple of simply connected type, not of type $A_n$, then there is an isomorphism $\kappa\colon W(\lambda)\rightarrow W(s)$ such that $\kappa(R(\lambda))=W^\circ(s)$. In particular, in this case $W(\lambda)/R(\lambda)$ is isomorphic to $(C_{\bG{G}^\ast}(s)/C_{\bG{G}^\ast}^\circ(s))^{F^\ast}$.
\end{enumerate}
\end{lemma}
\begin{proof}
Let $W^\ast$ denote the Weyl group of $\bG{G}^\ast$ with respect to $\bG{T}^\ast$ and write $(X(\bG{T}), \Phi, Y(\bG{T}), \Phi^\ast)$ and $(X(\bG{T}^\ast), \Phi^\ast, Y(\bG{T}^\ast), \Phi)$ for the root data for $\bG{G}$ and $\bG{G}^\ast$, respectively. By the duality of $\bG{G}$ and $\bG{G}^\ast$, we have an isomorphism $X(\bG{T})\rightarrow Y(\bG{T}^\ast)$, which induces an isomorphism $W\rightarrow W^\ast$. Statement (1) follows from the descriptions of this isomorphism and the isomorphism $\mathrm{Irr}(T)\cong T^\ast$ (see, for example, \cite[Propositions 4.2.3, 4.4.1]{carter2}), together with the identification of $W(s)$ as the subgroup of $W^\ast$ consisting of the elements which fix $s$ and that of $W(\lambda)$ as the subgroup of $W$ of elements which fix $\lambda$.
Now let $\bG{G}$ be simple of simply connected type, not of type $A_n$. Then by \cite[Lemma 5.1]{MalleSpathMcKay2} and its proof, we have $R(\lambda)$ is generated by the $s_\alpha$ for $\alpha\in\Phi$ such that $\lambda(h_{\alpha}(t))=1$ for each $t\in\mathbb{F}_q^\times$. But we may identify $W^\circ(s)$ as the subgroup of $W^\ast$ generated by $s_{\alpha^\ast}$, $\alpha^\ast\in\Phi^\ast$ such that $\alpha^\ast(s)=1$. Hence statement (2) again follows from the definitions of the isomorphisms and \cite[Remark 2.4]{dignemichel}.
\end{proof}
It is clear that an understanding of odd-degree characters will be imperative for proving SN2S-Goodness. For this reason, we next recall some key statements regarding odd-degree characters from \cite{MalleSpathMcKay2}.
\begin{theorem}{\cite[Theorem 7.7]{MalleSpathMcKay2}}\label{thm:MS7.7}
Let $\textbf{G}$ be simple, of simply connected type, not of type ${A}_n$. Let $\chi \in\mathrm{Irr}_{2'}(G)$. Then either $\chi$ lies in the principal series of
$G$, or $q\equiv 3\pmod{4}$, $G=Sp_{2n}(q)$ with $n\geq 1$ odd, $\chi\in\mathcal{E}(G,s)$ with $C_{G^\ast}(s)=B_{2k}(q)\cdot\tw{2}{D}_{n-2k}(q).2$, where $0\leq k\leq (n-3)/2$ and $\chi$ lies in the Harish-Chandra series of a cuspidal character of degree $(q-1)/2$ of a Levi subgroup $Sp_2(q)\times (q-1)^{n-1}$.\end{theorem}
\begin{lemma}{\cite[extension of Lemma 7.9]{MalleSpathMcKay2}}\label{lem:MS7.9}
Let $\textbf{G}$ be simple, of simply connected type, not of type ${A}_n$. Let $\chi\in\mathrm{Irr}_{2'}(G)$. Then $\chi=R_T^G(\lambda)_\gamma$, where $T$ is a maximally split torus of $G$, $\lambda\in\mathrm{Irr}(T)$ is such that $2\nmid[W:W(\lambda)]$, and $\gamma\in\mathrm{Irr}_{2'}(W(\lambda))$, except possibly in the case $\textbf{G}$ is type ${C}_n$ and $q\equiv3\pmod{4}$. In the latter case, $\chi$ may also be of the form $\chi=R_L^G(\lambda)_{\gamma}$ with $(L,\lambda)$ as in \prettyref{thm:MS7.7}, $2\nmid[W(L):W(\lambda)]$ and $\gamma\in\mathrm{Irr}_{2'}(W(\lambda))$.
\end{lemma}
The following lemma will be useful for verifying that a rational Lusztig series of $G$ containing an odd-degree character is fixed by $\sigma$.
\begin{lemma}\label{lem:lambdafixseriesfix}
Let $\chi$ be a constituent of $R_T^G(\lambda)$ where $T$ is a maximally split torus of $G$ and $\lambda\in \mathrm{Irr}(T)$. If $\lambda^\sigma=\lambda$, then $\chi$ lies in a rational Lusztig series $\mathcal{E}(G,s)$ indexed by some $s\in G^\ast$ such that $\mathcal{E}(G,s)^\sigma=\mathcal{E}(G,s)$.
\end{lemma}
\begin{proof}
Let $\chi\in\mathcal{E}(G,s)$. By \cite[11.10]{bonnafe06}, we know $\mathcal{E}(G,s)$ is a disjoint union of Harish-Chandra series $R_L^G(\lambda)$ with $\lambda\in\mathcal{E}(L,s)$. If $\chi$ is a member of $R_T^G(\lambda)$ with $\lambda^\sigma=\lambda$, this forces $\mathcal{E}(T,s)^\sigma=\mathcal{E}(T,s)$. Writing $s=s_2s_{2'}$ where $s_2$ is a $2$-element and $s_{2'}$ is a $2'$-element, this means that $s$ is conjugate to $s_2s_{2'}^2$ in $T^\ast$, and hence $G^\ast$, by \cite[Lemma 3.4]{SFTaylorTypeA}. This yields that $\mathcal{E}(G,s)^\sigma=\mathcal{E}(G,s)$, again by \cite[Lemma 3.4]{SFTaylorTypeA}.
\end{proof}
\subsubsection{On the Term $r_\sigma$}
Recall that the term $r_\sigma(w)$ appears in \prettyref{thm:GaloisAct}, where writing $w=w_1w_2\in W(\lambda)$ for $w_1\in C(\lambda)$ and $w_2\in R(\lambda)$, we have $r_\sigma(w):={\sqrt{\mathrm{Ind}(w_1)}^\sigma}/{\sqrt{\mathrm{Ind}(w_1)}}$. For the purpose of the next lemma only, we relax our assumptions on $G$ and return to the more general setting of \prettyref{sec:gallact}, in the sense that $G$ is not necessarily a group of Lie type. (However, $\sigma$ is still the Galois automorphism as in \prettyref{conj:mainprob}.)
Using Gauss sums, ${\sqrt{p}=\sum_{n=1}^{p-1}{\legendre{n}{p}}\zeta_p^{n}}$ or ${-\sqrt{-1}\sum_{n=1}^{p-1}{\legendre{n}{p}}\zeta_p^{n}}$ for an odd prime $p$, where ${\legendre{n}{p}}$ denotes the Legendre symbol and $\zeta_p$ is a primitive $p$th root of unity. Then ${\sqrt{p}^\sigma=\sum_{n=1}^{p-1}{\legendre{n}{p}}\zeta_p^{2n}}$ or ${-\sqrt{-1}\sum_{n=1}^{p-1}{\legendre{n}{p}}\zeta_p^{2n}}$. But note that ${\legendre{2n}{p}}={\legendre{2}{p}}{\legendre{n}{p}}$ and ${\legendre{2}{p}}=1$ if $p\equiv\pm1\mod 8$ and $-1$ if $p\equiv \pm3\mod 8$. This proves the following lemma:
\begin{lemma}\label{lem:rwgen}
Let $w\in W(\lambda)$ and let $r_\sigma$ be defined as above. Then
\begin{itemize}
\item If $\mathrm{Ind}{(w_1)}\equiv\pm1\mod8$, then $r_\sigma(w)=1$.
\item If $\mathrm{Ind}(w_1)\equiv\pm3\mod8$, then $r_\sigma(w)=-1$.
\end{itemize}
\end{lemma}
Now, and for the remainder of the article, we assume $G=\textbf{G}^F$ is a group of Lie type defined over $\mathbb{F}_q$. By \cite[Section 2.9]{carter2}, $\mathrm{Ind}(w_1)=q^{\ell(w_1)}$, where $\ell(w_1)$ is the length of $w_1$ in the Weyl group of $\textbf{G}$. This yields that $r_\sigma$ is a character in this case, and that we have the following reformulation of \prettyref{lem:rwgen} in the case of groups of Lie type:
\begin{lemma}\label{lem:rw}
Let $q$ be a power of an odd prime and let $G=\textbf{G}^F$ be a group of Lie type defined over $\mathbb{F}_q$. For $w\in W(\lambda)$, \begin{itemize}
\item If $q\equiv\pm1\mod8$, then $r_\sigma(w)=1$.
\item If $q\equiv\pm3\mod8$, then $r_\sigma(w)=(-1)^{\ell(w_1)}$, where $\ell(w_1)$ is the length of $w_1$ in the Weyl group of $\textbf{G}$.
\end{itemize}
\end{lemma}
We next address the case of the values of $r_\sigma$ when $G=\textbf{G}^F$ is a group of Lie type $B_n$.
\begin{lemma}\label{lem:ClatypeB}
Let $G=\textbf{G}^F$ be of simply connected type, where $\textbf{G}$ is of type ${B}_n$ ($n\geq 3$). Let $\chi\in\mathrm{Irr}_{2'}(G)$, and further write $\chi=R_T^G(\lambda)_\gamma$ as in \prettyref{lem:MS7.9}. Then every member of $C(\lambda)$ has even length in the Weyl group of $\textbf{G}$.
\end{lemma}
\begin{proof}
Recall that by \prettyref{lem:MS7.9}, the index $[{W}:W(\lambda)]$ is odd, so that $\lambda$ is fixed by a Sylow $2$-subgroup of ${W}$. Here $W$ can be viewed as a wreath product $C_2\wr\mathfrak{S}_n=\overline{K}\rtimes \mathfrak{S}_n$ with $\overline{K}\cong C_2^n$. In particular, note that this means $\overline{K}$ is contained in $W(\lambda)$.
Further, we fix a regular embedding $\bG{G}\hookrightarrow \wt{\bG{G}}$ as in \cite[Section 15.1]{cabanesenguehard}, so that $\wt{\bG{G}}$ has connected center, and write $\wt{G}=\wt{\bG{G}}^F$. Recall that this induces a surjection $\wt{\bG{G}}^\ast\rightarrow\bG{G}^\ast$. Let $\wt{\bG{T}}$ be an $F$-stable maximal torus of $\wt{\bG{G}}$ containing $\bG{T}$, let $\wt{\bG{T}}^\ast$ be dual to $\wt{\bG{T}}$, and denote by $\wt{T}=\wt{\bG{T}}^F$ and $\wt{T}^\ast=(\wt{\bG{T}}^\ast)^{F^\ast}$ the corresponding tori of $\wt{G}$ and $\wt{G}^\ast$.
Then we may extend $\lambda$ to a character $\wt{\lambda}$ of $\wt{T}$. Let $s$ and $\wt{s}$ be semisimple elements in $T^\ast$ and $\wt{T}^\ast$, respectively, corresponding to $\lambda$ and $\wt{\lambda}$. Then by \prettyref{lem:W(s)}, $R(\lambda)\cong W^\circ(s)\cong W(\wt{s})\cong W(\wt{\lambda})$ and $[W:R(\lambda)]_2\leq 2$. This yields that $R(\lambda)$ must be a direct product of the form $C_2\wr \mathfrak{S}_{n_1}\times C_2\wr \mathfrak{S}_{n_2} \times \mathfrak{S}_{n_3}$, where $n=n_1+n_2+n_3$ and $n_3\leq 1$. We also note that $C(\lambda)$ induces a graph automorphism on $R(\lambda)$, so $n_1=n_2$ if $C(\lambda)$ is nontrivial.
If $n_3=1$, we must have $[\mathfrak{S}_n:\mathfrak{S}_{n_1}\times \mathfrak{S}_{n_2}]$ is odd and $n_1+n_2=n-1$. But this cannot occur if $n_1=n_2$, and hence in this case $R(\lambda)=W(\lambda)$. However, note that this contradicts the assumption that $W(\lambda)$ contains $\overline{K}$.
Hence we see that $R(\lambda)$ is of the form $C_2\wr \mathfrak{S}_{n_1}\times C_2\wr \mathfrak{S}_{n_2}$ with $n_1+n_2=n$. The fact that $[\mathfrak{S}_n:\mathfrak{S}_{n_1}\times \mathfrak{S}_{n_2}]_2\leq 2$ forces either $R(\lambda)=W(\lambda)$ or $n_1=n_2$ is a $2$-power. (Recall that we exclude the case $n=2$.) But this implies that the elements mapping $\mathfrak{S}_{n_1}$ to $\mathfrak{S}_{n_2}$ have even length, as desired.
\end{proof}
Next, we consider the case that $G$ is of type $D_n^\pm$. For this, we recall the embedding of the simply connected group of type ${D}_n$ into ${B}_n$ discussed in \cite[2.C]{MalleSpathMcKay2} and \cite[Sections 10-12]{Spath10}.
Write $\overline{\bf{G}}$ for the simple algebraic group of simply connected type over $\overline{\mathbb{F}}_q$ with root system $\overline{\Phi}=\{\pm e_i\pm e_j\}\cup \{\pm e_i\}$ of type $B_n$, where $\{e_1,...,e_n\}$ is an orthonormal basis for a Euclidean space. The root system $\overline{\Phi}$ has simple roots $\overline{\Delta} = \{\overline{\alpha_1}, \alpha_2,...,\alpha_n\}$ with $\overline{\alpha}_1=e_1$ and $\alpha_i=e_i-e_{i-1}$ for $i>1$. Then there is a root subsystem $\Phi\subseteq\overline{\Phi}$ of type $D_n$ comprised of all the long roots, with simple roots $\Delta = \{\alpha_1,...,\alpha_n\}$, where $\alpha_1=2\overline{\alpha}_1+\alpha_2$. We write $\textbf{G}$ for the simply connected group corresponding to $\Phi$. Letting $\textbf{T}$ and $\overline{\textbf{T}}$ be the corresponding maximally split tori and $\textbf{N}$ and $\overline{\textbf{N}}$ their normalizers in $\textbf{G}$ and $\overline{\textbf{G}}$, respectively, we have $\textbf{T}=\overline{\textbf{T}}$ and $\textbf{N}=\overline{\textbf{N}}\cap\textbf{G}$. Note that then $T=\textbf{T}^F$ and $\overline{T}=\overline{\textbf{T}}^F$ coincide. We write $\bG{W}$ and $\overline{\bG{W}}$ for the corresponding Weyl groups and $W=\bG{W}^F, \overline{W}=\overline{\bG{W}}^F$ for the Weyl groups of $G$ and $\overline{G}$, so that $\bG{W}\leq \overline{\bG{W}}$ has index $2$.
Note that $W\cong\bG{W}$ and $\overline{W}\cong\overline{\bG{W}}$ if $F$ is a split Frobenius endomorphism. If $F$ is twisted, then $W\cong \overline{\bG{W}}'$, where $\overline{\bG{W}}'$ is a Weyl group of a root system of type $B_{n-1}$. For $\lambda\in\mathrm{Irr}(T)$, we write $W(\lambda)=C(\lambda)\ltimes R(\lambda)$ and $\overline{W}(\lambda)=\overline{C}(\lambda)\ltimes \overline{R}(\lambda)$ for the decompositions of the corresponding relative inertia groups of $\lambda$ in $W$ and $\overline{W}$, respectively, as in \prettyref{sec:gallact}.
Now, note that we may realize $\overline{\bG{W}}$ as the wreath product $C_2\wr \mathfrak{S}_n=\overline{K}\rtimes \mathfrak{S}_n$ with $\overline{K}=\overline{K}_1\times\cdots\times \overline{K}_n$ for $\overline{K}_i\cong C_2$. Then ${\bG{W}}=K\rtimes \mathfrak{S}_n$, where $K$ is the subgroup of $\overline{K}$ of index two consisting of elements $(\epsilon_1,\ldots,\epsilon_n)$ for $\epsilon_i\in \{\pm1\}$ satisfying $\prod_{i=1}^n \epsilon_i=1$. We write $k_i$ for the generator of $\overline{K}_i$, which can be seen to be induced by the element $n_{e_i}(1)$ of $\overline{\bf{N}}$, or $s_{e_i}$ in $\overline{\bG{W}}$.
\begin{proposition}\label{prop:Clambdaeven}
Let $G=\textbf{G}^F$ be of simply connected type, where $\textbf{G}$ is of type ${B}_n$ ($n\geq 3$) or ${D}_n$ ($n\geq 4$) and $G\neq \tw{3}{D_4}(q)$. Let $\chi\in\mathrm{Irr}_{2'}(G)$, and further write $\chi=R_T^G(\lambda)_\gamma$ as in \prettyref{lem:MS7.9}. Then every member of $C(\lambda)$ has even length in the Weyl group of $\textbf{G}$.
\end{proposition}
\begin{proof}
For type $B_n$, this is just \prettyref{lem:ClatypeB}, so we assume $\textbf{G}$ is of type ${D}_n$.
(1) Suppose first that $F$ is a split Frobenius endomorphism, so that $G$ is untwisted. Note that since $T$ is maximally split, we may apply \cite[Lemma 11.3]{Spath10} in the case $\nu=1=\nu'$ to see that $W(\lambda)=\overline{W}(\lambda)\cap W$, and hence $[\overline{W}(\lambda):W(\lambda)]$ divides $2$. Further, by \cite[Lemma 11.5]{Spath10}, $\overline{W}(\lambda)$ contains a normal subgroup of index dividing $2$ of the form $(A_1\times\cdots\times A_n)\rtimes U$ where $A_i\leq \overline{K}_i$ for each $i$ and $U$ is a product of symmetric groups.
If $W(\lambda)=\overline{W}(\lambda)$, then since $[W:W(\lambda)]=[W:\overline{W}(\lambda)]$ is odd, we know $\overline{W}(\lambda)$ contains $K$. Then $\overline{W}(\lambda)$ does not contain $k_i$ for any $i$, since then $\overline{W}(\lambda)$ would contain $\overline{K}$, contradicting $\overline{W}(\lambda)\leq W$. It follows that each $A_i$ is trivial, so $\overline{W}(\lambda)$ contains $U$ with index at most $2$, contradicting $K\leq \overline{W}(\lambda)$.
Hence we see $W(\lambda)$ has index $2$ in $\overline{W}(\lambda)$, so $\overline{W}=\overline{W}(\lambda)W$ and $[\overline{W}:\overline{W}(\lambda)]=[W:W(\lambda)]$ is odd, so $\overline{K}\leq \overline{W}(\lambda)$. Further, using \cite[Lemma 5.1]{MalleSpathMcKay2}, we see that $\Phi_\lambda=\overline{\Phi}_\lambda\cap\Phi$. This yields that $R(\lambda)=\overline{R}(\lambda)\cap W$, so $[\overline{R}(\lambda):R(\lambda)]$ divides $2$.
Now, the proof of \prettyref{lem:ClatypeB} yields that $\overline{R}(\lambda)=\overline{W}(\lambda)$ or $\overline{R}(\lambda)\cong (C_2\wr\mathfrak{S}_{2^a})^2 $ where $n=2^{a+1}$ and $[\overline{W}(\lambda):\overline{R}(\lambda)]=2$. This yields that $R(\lambda)=W(\lambda)$, unless $\overline{R}(\lambda)$ is as in the second case and $[\overline{R}(\lambda):R(\lambda)]=2$. In this case, notice that $\overline{R}(\lambda)$ contains $\overline{K}$, so $R(\lambda)$ contains $\overline{K}\cap W=K$. Then $R(\lambda)=K\rtimes (\mathfrak{S}_{2^a}\times\mathfrak{S}_{2^a})$, and $C(\lambda)=\overline{C}(\lambda)$ again consists of elements of even length.
(2) Finally, suppose that $F$ is twisted, so $W\cong \overline{\bG{W}}'$. It will be useful to identify $W$ in two ways. First, as a subgroup of $\bG{W}$, $W$ is generated by $s_2':=s_{\alpha_1}s_{\alpha_2}$ and $s_{\alpha_i}$ for $i\geq 3$. Note that the generator $s_2'$ has even length in $\bG{W}$ and that it induces the graph automorphism on a subgroup of $W$ isomorphic to a Weyl group of type $D_{n-1}$. Now, the map $s_2'\mapsto s_{e_2}; s_{\alpha_i}\mapsto s_{\alpha_i}, i\geq 3$ defines an isomorphism between $W$ and a Weyl group $\overline{\bG{W}}'$ of type $B_{n-1}$ with simple roots $\{\frac{1}{2}(\alpha_1+\alpha_2), \alpha_3,\ldots,\alpha_n\}=\{e_2,\alpha_3,\ldots,\alpha_n\}$. We will write $\overline{\Phi}'$ for the corresponding root system, which is a subsystem of $\overline{\Phi}$. Since $\overline{\bG{G}}$ is simply connected, we may apply \cite[Theorem 1.12.5]{gorensteinlyonssolomonIII} to $\overline{\bG{G}}':=\langle x_{\alpha}(t)\mid\alpha\in\overline{\Phi}'; t\in\overline{\mathbb{F}}_q\rangle$ and $\overline{\bG{T}}':=\langle h_\alpha(t)\mid\alpha\in\overline{\Phi}'; t\in\overline{\mathbb{F}}_q^\times\rangle$ to see that $\overline{\bG{G}}'$ is simply connected of type ${B}_{n-1}$.
Now, note that $T=\overline{\bG{T}}^F\cong \mathbb{F}_{q^2}^\times \times (\mathbb{F}_q^\times)^{n-2}$ is the set of elements of $\overline{\bG{T}}$ of the form $\prod_{i=1}^n h_{\alpha_i}(t_i)$ where $t_1, t_2\in\mathbb{F}_{q^2}^\times$ and $t_i\in\mathbb{F}_q^\times$ for $i\geq 3$, and $t_2=t_1^q$. The subgroup $T'=\overline{\bG{T}}'^F\cong (\mathbb{F}_q^\times)^{n-1}$ of $T$ is the set of elements of $\overline{\bG{T}}$ of the form $\prod_{i=1}^n h_{\alpha_i}(t_i)$ where $t_i\in\mathbb{F}_q^\times$ for $i\geq 1$, and $t_2=t_1$. Further, writing $N'=\overline{\bG{N}}'^F$ for $\overline{\bG{N}}'=\langle n_{\alpha}(t)\mid\alpha\in\overline{\Phi}'; t\in\overline{\mathbb{F}}_q^\times\rangle$, we have $N'T=N$ and $N'\cap T=T'$. Then letting $\lambda'\in\mathrm{Irr}(T'|\lambda)$, we have $W(\lambda)=N_{\lambda}/T$ is isomorphic to a subgroup of $\overline{W'} (\lambda')=N_{\lambda'}'/T'$, with odd index since $[W:W(\lambda)]$ is odd. Writing $\overline{W}'(\lambda')=\overline{R}'(\lambda')\rtimes \overline{C}'(\lambda')$ for the corresponding decomposition, since $[\overline{\bG{W}}':\overline{W'}(\lambda')]$ is odd, we may again apply the proof of \prettyref{lem:ClatypeB} to see that $\overline{R'}(\lambda')$ is of the form $C_2\wr \mathfrak{S}_{n_1}\times C_2\wr \mathfrak{S}_{n_2}$ with $n_1+n_2=n-1$. In particular, if we write $\overline{\bG{W}}$ as the wreath product $C_2\wr \mathfrak{S}_{n-1}=\overline{K}'\rtimes \mathfrak{S}_{n-1}$ in analogy to $\overline{\bG{W}}$, we see that $\overline{K}\leq \overline{R}'(\lambda')$ and $\overline{C}'(\lambda')$ is comprised of elements of even length.
Recall that $\overline{R}'(\lambda')=\langle s_{\alpha}\mid\alpha\in\overline{\Phi}'_{\lambda'}\rangle$, and viewing $W(\lambda)$ as a subgroup of $\overline{W}'(\lambda')$, we have $R(\lambda)=\langle s_{\alpha}\mid\alpha\in\overline{\Phi}'_{\lambda}\rangle$. Note that by \cite[Lemma 5.1]{MalleSpathMcKay2}, $\overline{\Phi}'_\lambda=\{\alpha\in\overline{\Phi}'\mid\lambda(T\cap \langle X_\alpha, X_{-\alpha}\rangle) = 1\}$ and $\overline{\Phi}'_{\lambda'}=\{\alpha\in\overline{\Phi}'\mid\lambda'(T'\cap \langle X_\alpha, X_{-\alpha}\rangle) = 1\}$, so that $R(\lambda)$ is a subgroup of $\overline{R}'(\lambda')$.
Since $s_{e_2}\in \overline{R}'(\lambda')$, we see that either $s_{e_2}\in R(\lambda)$, in which case $\lambda$ is trivial on the $\mathbb{F}_{q^2}^\times$ factor of $T$, so $R(\lambda)=\overline{R}'(\lambda')$, or $s_{e_2}$ induces the quotient $\overline{R}'(\lambda')/R(\lambda)$. Recalling that as a member of $\bG{W}$, $s_{e_2}$ corresponds to the element $s_{\alpha_1}s_{\alpha_2}$, we therefore see that the members of $C(\lambda)$ have even length in $\bG{W}$.
\end{proof}
\subsubsection{On the Character $\gamma^{(\sigma)}$}
\prettyref{thm:GaloisAct} and \prettyref{lem:MS7.9} also suggest that another important ingredient for SN2S-Goodness will be to understand the behavior of $\mathrm{Irr}_{2'}(W(\lambda))$ under the action of $\sigma$.
\begin{proposition}\label{lem:W(la)}
Let ${G}$ be a group of Lie type with odd defining characteristic and let $\lambda\in\mathrm{Irr}(T)$ such that $C(\lambda)$ is a $2$-group. Then every $\gamma\in\mathrm{Irr}_{2'}(W(\lambda))$ is fixed by $\sigma$.
\end{proposition}
\begin{proof}
Let $\gamma\in\mathrm{Irr}_{2'}(W(\lambda))$. Recall that $W(\lambda)=R(\lambda)\rtimes C(\lambda)$ and that $R(\lambda)$ is a Weyl group. By Clifford theory, $\gamma|_{R(\lambda)}=e\sum_{i=1}^t\theta^{g_i}$ for some $\theta\in\mathrm{Irr}(R(\lambda))$ with $e, t$ dividing $|C(\lambda)|$. Then since $\gamma$ has odd degree and $C(\lambda)$ is a $2$-group, we see that $e=t=1$ and $\gamma|_{R(\lambda)}=\theta\in\mathrm{Irr}_{2'}(R(\lambda))$. Now, since $R(\lambda)$ is a Weyl group, $\theta$ takes values in the rational numbers (see, for example, \cite[Theorems 5.3.8, 5.4.5, 5.5.6, and Corollary 5.6.4]{GeckPfeiffer}), so recalling that $C(\lambda)$ is abelian, we see by \cite[Lemma 3.4]{SchaefferFrySN2S1} that $\gamma$ is fixed by $\sigma$.
\end{proof}
Due to the nature of the action of $\sigma$ on the parametrization $R_T^G(\lambda)_\gamma$ given in \prettyref{thm:GaloisAct}, what we really hope for is a corresponding statement to \prettyref{lem:W(la)} but for the character $\eta\in\mathrm{End}_G(\fl{\lambda})$ satisfying $\mathfrak{f}(\eta)=\gamma$. We make progress toward this end with the next two observations. Here we keep the notation of \prettyref{sec:genericsubalg}.
\begin{proposition}\label{prop:fixHf}
Let $\lambda\in\mathrm{Irr}(T)$ such that $C(\lambda)$ is a $2$-group. Let $\mathbb{K}$ denote the fixed field of $\sigma$, and suppose that every simple $\mathcal{H}_0^f$-module is afforded over $\mathbb{K}$. Then $\eta^{(\sigma)}=\eta$ for every irreducible character $\eta$ of $\mathrm{End}_G(\fl{\lambda})$ of odd degree. In particular, this is true for $\eta$ satisfying $\mathfrak{f}(\eta)=\gamma$ for $\gamma\in\mathrm{Irr}_{2'}(W(\lambda))$.
\end{proposition}
\begin{proof}
Keeping the notation of \prettyref{sec:genericsubalg}, we recall that every irreducible character of $\mathcal{H}^f\cong\mathrm{End}_G(\fl{\lambda})$ is of the form $(\tau^{\mathcal{H}})_f$, where $\tau$ is an extension to $C(\lambda)_\psi\mathcal{H}_0^K$ of some irreducible character $\psi$ of the generic algebra $\mathcal{H}_0^K$ corresponding to $R(\lambda)$. Observing the description of values of $\tau^{\mathcal{H}}$, notice that $\tau^{\mathcal{H}}(a_1)=[C(\lambda):C(\lambda)_{\psi}] \tau(a_1)$. Hence, if $(\tau^{\mathcal{H}})_f$ has odd degree, then since $C(\lambda)$ is a $2$-group, we see that $C(\lambda)=C(\lambda)_{\psi}$, $\mathcal{H}^K=C(\lambda)\mathcal{H}_0^K=C(\lambda)_\psi\mathcal{H}_0^K$ and $\tau^\mathcal{H}=\tau$. It therefore suffices to show that if $\psi$ has odd degree, then $\tau_f$ takes its values in $\mathbb{K}$.
Recall that the characters of the form $\beta\tau$ for $\beta\in\mathrm{Irr}(C(\lambda))$ form the complete set of extensions of $\psi$ to $\mathcal{H}^K$. Let $\varrho$ be a representation of $\mathcal{H}_0^K$ affording $\psi$ and $\widetilde{\varrho}$ a representation of $\mathcal{H}^K$ affording $\tau$ and extending $\varrho$. Notice that $\mu_0:=\det\circ \varrho$ is also a representation of $\mathcal{H}_0^K$. As such, we see that $\mu_0$ also extends to a representation $\mu$ of $\mathcal{H}^K$, and the complete set of such extensions is given by $\beta\mu$ for $\beta\in\mathrm{Irr}(C(\lambda))$. In particular, we may take $\mu=\det\circ\widetilde{\varrho}$.
Now, from the proof of \cite[Theorem 3.8]{HowlettKilmoyer}, we see that $\mu(d\otimes b)=\mu(d)\mu(b)=\mu(d)\mu_0(b)$ for $d\in C(\lambda)$ and $b\in \mathcal{H}_0^K$. Since $\mu$ is degree $1$, we may consider it as a character. Then taking the specializations, we see that $\mu_f(d\otimes b)=\mu_f(d)(\mu_0)_f(b)$. By assumption, $(\mu_0)_f(b)$ takes values in $\mathbb{K}$, and all characters of the $2$-group $C(\lambda)$ are $\sigma$-fixed. Hence we see that in fact, $\mu_f$ takes values in $\mathbb{K}$.
Now, since $\psi_f$ has values in $\mathbb{K}$, we see that $(\tau_f)^{(\sigma)}$ must also be an extension of $\psi_f$. Then $(\tau_f)^{(\sigma)}=\beta\tau_f$ for some $\beta\in\mathrm{Irr}(C(\lambda))$, and hence $(\det\circ\widetilde{\varrho})_f=\mu_f=(\mu_f)^{(\sigma)}=(\det\circ\widetilde{\varrho}^{(\sigma)})_f=\beta^{\psi(1)}\mu_f$. Since $\beta$ is a character of a $2$-group and $\psi(1)$ is odd, we see that $\beta=1$ and hence $\tau_f$ has values in $\mathbb{K}$, as desired.
\end{proof}
Armed with \prettyref{prop:fixHf}, we are interested in determining when the characters of $\mathcal{H}_0^f$ are afforded over $\mathbb{K}$. The next proposition gives us a partial answer to this question.
\begin{lemma}\label{lem:fixH0f}
Let $\lambda\in\mathrm{Irr}(T)$ and suppose that the Weyl group $R(\lambda)$ has no component of type $G_2, E_7,$ or $E_8$. Then every simple $\mathcal{H}_0^f$ - module is realizable over $\mathbb{Q}$.
\end{lemma}
\begin{proof}
Let $R(\lambda)=W_1\times\cdots\times W_t$ be the decomposition of $R(\lambda)$ into irreducible Weyl groups. Then we can write $\mathcal{H}_0=\mathcal{H}_1\otimes\cdots\otimes\mathcal{H}_t$ where $\mathcal{H}_i$ is the generic subalgebra generated by $\{a_w\mid w\in W_i\}$. Suppose that $M$ is an irreducible $\mathcal{H}_0^f$-module which is not realizable over $\mathbb{Q}$. Then we may write $M=M_1\otimes\cdots\otimes M_t$ where each $M_i$ is a simple $\mathcal{H}_i^f$-module. Since $M$ is not realizable over $\mathbb{Q}$, there must be some $i$ such that $M_i$ is not realizable over $\mathbb{Q}$. But this contradicts our assumption that $W_i$ is not of type $G_2,$ $E_7,$ or $E_8$, by \cite[9.3.4]{GeckPfeiffer}.
\end{proof}
We remark that the definition of $\mathfrak{f}$ and the fact that $W(\lambda)=W(\lambda^\sigma)$ yield that $\mathfrak{f}(\eta)=\mathfrak{f}(\overline{\eta})$. Hence, if $\eta^{(\sigma)}=\eta$, then $\gamma^{(\sigma)}=\gamma$. Then \prettyref{lem:fixH0f} and \prettyref{prop:fixHf} give the following:
\begin{corollary}\label{cor:newsigmaaction}
Let $\lambda\in\mathrm{Irr}(T)$ such that the Weyl group $R(\lambda)$ has no component of type $G_2, E_7,$ or $E_8$ and the complement $C(\lambda)$ is a $2$-group. Then for $\gamma\in\mathrm{Irr}_{2'}(W(\lambda))$, $\eta^{(\sigma)}=\eta$, where $\eta $ is the character of $\mathrm{End}_G(\fl\lambda)$ satisfying $\mathfrak{f}(\eta)=\gamma$. Hence in this case, $\gamma^{(\sigma)}=\gamma$.
\end{corollary}
\begin{comment}
\begin{proposition}
Let $G$ be a group of Lie type defined over a field with $q$ elements, where $q$ is odd. Let $\lambda\in\mathrm{Irr}(T)$ and suppose that the Weyl group $R(\lambda)$ has no component of type $E_7$ or $E_8$. Then every simple $\mathcal{H}_0^f$-module is realizable over $\mathbb{Q}(\sqrt{q})$. In particular, when $q\equiv\pm1\mod 8$, we see that every simple $\mathcal{H}_0^f$-module is realizable over $\mathbb{K}$.
\end{proposition}
\begin{proof}
First, notice that by \cite[5.3.6, 5.4.5, 5.5.6, 5.6.4, and 6.3.8]{GeckPfeiffer}, $\mathbb{Q}$ is a splitting field for $R(\lambda)$. Then since $\mathcal{H}_0$ is a generic Iwahori Hecke algebra corresponding to the Coxeter group $R(\lambda)$, we see by \cite[9.3.5]{GeckPfeiffer} that every irreducible $\mathcal{H}_0^K$ -module is realizable over $\mathbb{Q}(\sqrt{u_\alpha} \colon \alpha\in\Delta_\lambda)$. Hence we see that every $\mathcal{H}_0^f$-module is realizable over $\mathbb{Q}(\sqrt{q})$, since the $p_{\alpha,\lambda}$ are powers of $q$.
\end{proof}
\end{comment}
\subsection{SN2S-Goodness for $\bG{G}$ of Types ${B}_n, {C}_n, {D}_n,$ and ${E}_8$}
In this section, we show that for $q$ odd, the simple groups $PSp_{2n}(q)$ for $n\geq 2$, $P\Omega^\pm_{n}(q)$ for $n\geq 7$, and $E_8(q)$ are SN2S-Good.
We begin with an observation about $\delta_{\lambda,\sigma}$ in the case $\lambda^\sigma=\lambda$ and $\lambda$ is linear, which follows from careful consideration of the extension maps constructed in \cite{Spath09} and \cite{MalleSpathMcKay2}.
\begin{proposition}\label{prop:extn}
Let $\bG{G}$ be simple, of simply connected type, not of type $A_n$, with odd defining characteristic. Let $\lambda\in\mathrm{Irr}_{2'}(L)$ be linear and fixed by $\sigma$. Then there exists an extension $\Lambda_\lambda$ of $\lambda$ to $N(L)_{\lambda}$ which is $\sigma$-invariant.
\end{proposition}
\begin{proof}
Let $H$ and $V$ be defined as in \cite[Section 2]{Spath09}. Then the proof of \cite[Lemma 4.2]{Spath09} yields that it suffices to note that there is a $\sigma$-fixed extension of $\lambda|_H$ to its stabilizer $V_\lambda$ in $V$, since $H$ is a $2$-group and $\lambda$ is linear.
\end{proof}
\begin{theorem}\label{thm:princseries1mod8}
Let $G$ be a finite group of Lie type of simply connected type $B_n$ ($n\geq 3$), $C_n$ ($n\geq 2$), or $D_n^{\pm}$ ($n\geq4$) over a field with $q$ elements, where $q\equiv\pm1\mod 8$, or let $G$ be the simple group $E_8(q)$ for $q$ odd. Then every $\chi\in\mathrm{Irr}_{2'}(G)$ lying in the principal series is fixed by $\sigma$.
\end{theorem}
\begin{proof}
Let $\chi\in\mathrm{Irr}_{2'}(G)$ lie in the principal series, so that there is $\lambda\in\mathrm{Irr}(T)$ such that $\chi\in\mathrm{Irr}_{2'}(G|R_T^G(\lambda))$. By \prettyref{lem:MS7.9}, we may write $\chi=(R_T^G(\lambda))_\gamma$, where $\gamma\in\mathrm{Irr}_{2'}(W(\lambda))$.
First assume that $G$ is type $B_n$, $C_n$, or $D^{\pm}_n$. Then by \cite[Lemma 7.5]{MalleSpathMcKay2}, we see that $\lambda^2=1$, and hence $\lambda^\sigma=\lambda$. Further, by \prettyref{prop:extn}, there is an extension $\Lambda_\lambda$ of $\lambda$ to $N_G(T)_{\lambda}$ which is $\sigma$-fixed, so that $\delta_{\lambda,\sigma}=1$ and certainly $R(\lambda)\leq\ker\delta_{\lambda,\sigma}$. Further, note that $C(\lambda)$ embeds into $Z(\bG{G})/Z(\bG{G})^\circ$, yielding that $C(\lambda)$ is a $2$-group. Then by \prettyref{thm:GaloisAct}, \prettyref{lem:rw}, and \prettyref{cor:newsigmaaction}, we have $\chi^\sigma=\chi$, using our assumption that $q\equiv\pm1\mod8$.
Next assume that $G=E_8(q)$. Then $G$ is self-dual and has a self-normalizing Sylow $2$-subgroup. Hence any semisimple $s\in G^\ast$ centralizing a Sylow $2$-subgroup must have $2$-power order. Then if $\chi\in\mathcal{E}(G,s)$, it follows that $\lambda$ also has $2$-power order so is fixed by $\sigma$, and by \cite[Theorem 3.2]{SchaefferFrySN2S1}, $\lambda$ is realizable over the field $\mathbb{K}$ of fixed points under $\sigma$. Note that since $\gamma$ has odd degree, the corresponding character of $\e{\lambda}$ also has odd degree. Then since $G$ is of adjoint type, \cite[Proposition 5.5]{Geck03} and its proof yield that $\chi$ can be realized over $\mathbb{K}$ as well. (Indeed, note that the excluded characters in \cite[Proposition 5.5]{Geck03} correspond to characters of $\e{\lambda}$ of even degree.)
\end{proof}
When combined with \cite[Theorem 7.7]{MalleSpathMcKay2}, \prettyref{thm:princseries1mod8} yields the following immediate consequence.
\begin{corollary}\label{cor:princseries1mod8}
The simple groups $E_8(q)$ are SN2S-Good for all $q$. The simple groups $P\Omega_n^\pm(q)$ are SN2S-Good for $n\geq 7$ when $q\equiv\pm1\mod8$. Further, when $q\equiv1\mod8$, the simple groups $PSp_{2n}(q)$ for $n\geq 2$ are also SN2S-Good.
\end{corollary}
We next address the case that $S=PSp_{2n}(q)$. The following observation yields that $PCSp_{2n}(q)$ has a self-normalizing Sylow $2$-subgroup. Note that $PCSp_{2n}(q)$ is the group $\mathrm{InnDiag}(S)\leqslant\mathrm{Aut}(S)$, as defined in \cite[2.5.10(d)]{gorensteinlyonssolomonIII}.
\begin{lemma}\label{lem:CSpSN2S}
Let $S=PSp_{2n}(q)$ with $q\equiv\pm3\pmod 8$. Then $\mathrm{InnDiag}(S)\cong PCSp_{2n}(q)$ has a self-normalizing Sylow $2$-subgroup.
\end{lemma}
\begin{proof}
Let $\wt{S}=PCSp_{2n}(q)$. Note that $|\wt{S}/S|=2$, and in fact we may write $\wt{S}\cong S\rtimes \langle\delta\rangle$ for some diagonal automorphism $\delta$. Now, by \cite{carterfong}, we see that a Sylow $2$-subgroup of $G=Sp_{2n}(q)$ is the direct product $P=P_1\times\cdots\times P_t$, where $2n=2^{r_1}+\cdots +2^{r_t}$ is the 2-adic expansion of $2n$ and $P_i$ is a Sylow $2$-subgroup of $Sp_{2^{r_i}}(q)$. Further, writing $\overline{P}=PZ(G)/Z(G)$ for a Sylow $2$-subgroup of $S$, we see that $N_G(P)/P\cong N_S(\overline{P})/\overline{P}$ is elementary abelian of order $3^t$, where the $i$th copy of the cyclic group $C_3$ corresponds to $N_{Sp_{2^{r_i}}(q)}(P_i)/P_i$. Hence it suffices to show that $C_{V_i}(\delta)=1$, where we write $N_{Sp_{2^{r_i}}(q)}(P_i)=P_i\rtimes V_i$ with $V_i\cong C_3$.
Using the results of \cite{carterfong}, we can further realize $P_i$ as the wreath product $P'\wr T_{r_i-1}$, where $P'$ is a Sylow $2$-subgroup of $Sp_{2}(q)$ and $T_{r_i-1}$ is a Sylow $2$-subgroup of the symmetric group $\mathfrak{S}_{2^{r_i}-1}$. As such, $V_i$ acts on $P_i$ via the action of $N_{Sp_2(q)}(P')/P'\cong C_3$ on $P'$. We also see that $\delta$ can be taken to act on $P'$ via the action of $PGL_2(q)/PSL_2(q)$, so that $C_{V_i}(\delta)=1$ since $PGL_2(q)$ has a self-normalizing Sylow $2$-subgroup, using \cite[Lemma 3]{carterfong}.
\end{proof}
\begin{theorem}\label{thm:typeC}
Let $S=PSp_{2n}(q)$ for $n\geq 2$, where $q$ is odd. Then $S$ is SN2S-Good.
\end{theorem}
\begin{proof}
Let $G=Sp_{2n}(q)$. Note that by the results of \cite{SchaefferFrySN2S1}, we may assume that $S$ does not have an exceptional Schur multiplier, and if $q\equiv 1\mod 8$, we are done by \prettyref{cor:princseries1mod8}.
First suppose $q\equiv 3\mod 4$, so that \cite[Theorem 7.7]{MalleSpathMcKay2} yields that there are odd degree characters lying in a series $R_L^G(\lambda)$ where $L\cong Sp_2(q)\times T_1$ for $T_1\cong (q-1)^{n-1}$ and $\lambda=\psi\times\lambda_1$ with $\psi\in\mathrm{Irr}(Sp_2(q))$ of degree $\frac{q-1}{2}$. If $q\equiv 3\mod 8$, then $S$ does not have a self-normalizing Sylow $2$-subgroup. The character values of $Sp_2(q)$ are well-known, and it is clear that $\psi^\sigma\neq\psi$, since $\sqrt{q}^\sigma\neq\sqrt{q}$, so $\lambda\in\mathrm{Irr}(L)$ is not fixed by $\sigma$. Hence $R_L^G(\lambda)$ is not fixed by $\sigma$, so there exist characters of odd degree not fixed by $\sigma$, and we are done in this case by \prettyref{prop:unipsfixed} and the discussion preceding it. If $q\equiv 7\mod 8$, then $G$ has a self-normalizing Sylow $2$-subgroup and it suffices to show that every $\chi\in\mathrm{Irr}_{2'}(G)$ is fixed by $\sigma$. In particular, \prettyref{thm:princseries1mod8} yields that it suffices to show that $\chi^\sigma=\chi$ for $\chi=R_L^G(\lambda)_\gamma$ in the non-principal series mentioned above. Notice that $r_\sigma(w)=1$ by \prettyref{lem:rw}. Further, since $\Lambda_\lambda$ is of the form $\psi\times\Lambda_1(\lambda_1)$ for an extension $\Lambda_1(\lambda_1)$ of $\lambda_1$ to $N(T_1)_{\lambda_1}$, we have $\delta_{\lambda,\sigma}=1$ by \prettyref{prop:extn}. Since $W(\lambda)\cong W(\lambda_1)$, \prettyref{cor:newsigmaaction} still yields $\gamma^{(\sigma)}=\gamma$ in this situation, proving the statement by \prettyref{thm:GaloisAct}.
Finally, let $q\equiv 5\mod 8$. As in the case $q\equiv 3\mod 8$, $S$ does not have a self-normalizing Sylow $2$-subgroup, and it suffices to show that there is a character of odd degree not fixed by $\sigma$. Let $s\in G^\ast$ with $C_{G^\ast}(s)$ as in the first two lines of \cite[Table 1]{MalleSpathMcKay2} with $k$ the largest power of $2$ smaller than $n$. Let $\{\alpha_1,...,\alpha_n\}$ denote the simple roots determined by $\bG{T}$ and $\bG{B}$, with $\alpha_1=e_1$ and $\alpha_i=e_i-e_{i-1}$ as in \cite[1.8.8]{gorensteinlyonssolomonIII}. Let $\lambda$ be trivial on $ h_{e_i}(t)$ for $i> k$ and $t\in \mathbb{F}_{q}^\times$ and have order $2$ on the subgroups $\langle h_{e_i}(t)\colon t\in\mathbb{F}_q^\times\rangle$ for $1\leq i\leq k$. Then $R(\lambda)$ is a reflection group of type $D_k\times B_{n-k}$, $W(\lambda)$ is of type $B_k\times B_{n-k}$, and $C(\lambda)$ induces the graph automorphism on the $D_k$ component. Hence a nontrivial element in $C(\lambda)$ is conjugate to a simple reflection, meaning it has odd length in $W$, yielding that $r_\sigma(w)=-1$ by \prettyref{lem:rw}. Then letting $\chi=R_T^G(\lambda)_\gamma$, we see that since $\lambda^\sigma=\lambda$ by construction, $\delta_{\lambda,\sigma}=1$ by \prettyref{prop:extn} and $\gamma^{(\sigma)}=\gamma$ by \prettyref{cor:newsigmaaction}, it must be that $\chi^\sigma\neq\chi$ by \prettyref{thm:GaloisAct}.
\end{proof}
\begin{theorem}\label{thm:typeBD}
Let $S=P\Omega_{n}^{\pm}(q)$ for $n\geq 7$, where $q$ is odd. Then $S$ is SN2S-Good.
\end{theorem}
\begin{proof}
Note that $S$ has a self-normalizing Sylow $2$-subgroup and that by the results of \cite{SchaefferFrySN2S1}, we may assume that $S$ does not have an exceptional Schur multiplier. Let $G=\mathrm{Spin}_{n}^\pm(q)$. Then by \prettyref{thm:GaloisAct}, \prettyref{prop:extn}, \prettyref{lem:rw}, \prettyref{cor:newsigmaaction}, and \prettyref{cor:princseries1mod8}, it suffices to note that by \prettyref{prop:Clambdaeven}, in the case $q\equiv\pm3\mod 8$, $\ell(w_1)$ is even for each $w_1\in C(\lambda)$ when $\chi=R_T^G(\lambda)_\gamma\in\mathrm{Irr}_{2'}(G)$.
\end{proof}
\subsection{SN2S-Goodness for $\bG{G}$ of Type ${E}_6$}
For this section, we keep the notation of the beginning of \prettyref{sec:oddchars}. Let $q$ be a power of an odd prime $p$ and let $\bG{G}$ be simply connected of type ${E}_6$, so that $G=\bG{G}^F=E_6^\epsilon(q)_{sc}$, and let $S$ be the simple group $E_6^\epsilon(q)=G/Z$ with $Z:=Z(G)$. Here we let $\epsilon=1$ (or $+$) if $F$ is split and $\epsilon = -1$ (or $-$) if $F$ is twisted. Then the dual group $G^\ast\cong E_6^\epsilon(q)_{ad}$ satisfies $[G^\ast,G^\ast]\cong S$, and we make this identification. Further, we fix a regular embedding $\bG{G}\hookrightarrow \wt{\bG{G}}$ and write $\wt{G}=\wt{\bG{G}}^F$ so that $G=[\wt{G}, \wt{G}]$ and $G^\ast\cong \wt{G}/Z(\wt{G})\cong \wt{G^\ast}/Z(\wt{G}^\ast)$. Recall that this induces a surjection $\wt{\bG{G}}^\ast\rightarrow\bG{G}^\ast$.
We will write $\phi$ for a fixed nontrivial graph automorphism of $G$ normalizing a maximally split torus $T=\bG{T}^F$ and let $F_p$ denote the field automorphism induced by the map $x\mapsto x^p$ on $\overline{\mathbb{F}}_p$, so that $\mathrm{Aut}(G)\cong G^\ast\rtimes \langle \phi, F_p\rangle$. We begin by describing when the group $GQ/Z$ has a self-normalizing Sylow $2$-subgroup, for some $2$-group $Q\in\mathrm{Aut}(G)$. First, we describe the situation when $Q=1$, which can be found in \cite[Theorem 6]{kondratievmazurov}.
\begin{lemma}[Kondrat'ev-Mazurov]\label{lem:whenSN2Ssimple}
A simple group $E_6^\epsilon(q)$ has a self-normalizing Sylow $2$-subgroup if and only if $(q-\epsilon)_{2'} = (3,q-\epsilon)_{2'}$.
\end{lemma}
In fact, from \cite[Lemma 1.3 and proof of Theorem 6]{kondratievmazurov} and \cite[Lemma 4.13]{navarro-tiep:2015:irreducible-representations-of-odd-degree}, we see that for $P$ a Sylow $2$-subgroup of $G$, we have \[N_{G/Z}(PZ/Z)=PZ/Z \times \overline{C}\quad\hbox{ and }\quad N_{G}(P)=P \times {C},\] where $C\cong C_{(q-\epsilon)_{2'}}$ and $\overline{C}\cong C_{(q-\epsilon)_{2'}/(3,q-\epsilon)}$ is a subgroup of $C_{G^\ast}(t)$, with $t$ the unique involution in the center of $PZ/Z$. By \cite[Proofs of Lemmas 4.25, 4.26]{harris}, we see that in fact $\overline{C}$ is comprised of elements $h(\chi)$ such that $\chi(\alpha_i)=1$ for $2\leq i\leq 5$ and $\chi(\alpha_6)=\chi(\alpha_1)^{-1}$, where we write $\{\alpha_1,...,\alpha_6\}$ for the simple roots numbered as in \cite[13.3.3]{Carter1} and $h(\chi)$ is as in \cite[Section 7.1]{Carter1}. This, together with the definition of the action of $\phi$ and $F_p$ on the Chevalley generators yields the following:
\begin{lemma}\label{lem:autactC}
The graph automorphism $\phi$ acts on $\overline{C}$ and $C$ by inversion. Further, the field automorphism $F_p$ acts on $\overline{C}$ and $C$ by $z\mapsto z^p$.
\end{lemma}
\begin{lemma}\label{lem:whenSN2S}
Write $p^a$ for $q$ or $q^2$ in the case $\epsilon =1$ or $\epsilon =-1$, respectively, and let $Q \leqslant \mathrm{Aut}(G)$ be a $2$-group. The quotient $GQ/Z$ has a self-normalizing Sylow $2$-subgroup if and only if at least one of the following is satisfied:
\begin{enumerate}
\item $G/Z$ has a self-normalizing Sylow $2$-subgroup;
\item $Q$ contains a graph automorphism in case $\epsilon = 1$ or an involutary field automorphism in case $\epsilon=-1$, either of which we may identify as the map $\phi$, up to inner and diagonal automorphisms;
\item $\epsilon=1$, $Q$ contains a field or graph-field automorphism $\varphi$ of order $a/m$ for some $m$ dividing $a$ (which we identify with $\phi^{\delta'} F_{ p}^m$, up to inner and diagonal automorphisms for $\delta'\in\{0,1\}$), and $\bG{G}^\varphi/Z(\bG{G}^\varphi)$ has a self-normalizing Sylow $2$-subgroup. That is, by \prettyref{lem:whenSN2Ssimple}, one of the following holds, where we write $\delta=(-1)^{\delta'}$:
\begin{itemize}
\item[(3a)] $a$ is a $2$-power and $(p^m-\delta)_{2'}=1$, or
\item[(3b)] $(p^m-\delta)_{2'}=\gcd(3,p^m-\delta)_{2'}$;
\end{itemize}
\item $\epsilon=1$, $Q$ contains a field automorphism $\varphi_1$ of order $a/m_1$ and a graph-field automorphism $\varphi_2$ of order $a/m_2$ for some $m_1,m_2$ dividing $a$ (which we identify with $F_p^{m_1}$ and $\phi F_{ p}^{m_2}$, up to inner and diagonal automorphisms), and $\gcd(p^{m_1}-1, p^{m_2}+1)_{2'}=1$.
\end{enumerate}
\end{lemma}
We remark that \prettyref{lem:whenSN2S} is quite similar to the corresponding statement for type $A^\epsilon_n(q)$ in \cite{SFTaylorTypeA}.
\begin{proof}[Proof of \prettyref{lem:whenSN2S}]
Armed with \prettyref{lem:autactC}, the proof is analogous to a simplified version of that of \cite[Lemma 8.5]{SFTaylorTypeA}, but we include it for completeness. Let $P$ be a Sylow $2$-subgroup of $G$ stabilized by $Q$.
First suppose that one of (1), (2), (3), or (4) holds. Note that in case (1), the statement is certainly true, since then $N_G(P)=PZ$, so $C_{N_G(P)/PZ}(Q)=1$. Hence we may assume that $G/Z$ does not have a self-normalizing Sylow $2$-subgroup and that $Q$ contains an outer automorphism. Specifically, either $Q$ contains a graph automorphism (in case $\epsilon =1$) or involutary field automorphism (in case $\epsilon = -1$), which we identify with $\phi$ on $G$, up to conjugation in $\widetilde{G}$; or $\epsilon=1$ and $Q$ contains a field or graph-field automorphism, which we identify as $\phi^{\delta'}F_{p^m}$ on $G$, up to conjugation in $\widetilde{G}$, for some $m\geq 1$. Write $\varphi$ for the corresponding graph, field, or graph-field automorphism, respectively. We will show that $C_{N_G(P)/PZ}(\varphi)=1$ in cases (2) and (3) and that $C_{N_G(P)/PZ}(\varphi_1, \varphi_2)=1$ in case (4). Write $\overline{N}:=N_G(P)/PZ$ and let $\overline{g}$ denote the image of an element $g\in N_G(P)$ in $\overline{N}$. Suppose $g\in N_G(P)$ satisfies that $\overline{g}\in C_{\overline{N}}(\varphi)$. That is, $\overline{g}$ is fixed by $\varphi$ (resp. $\varphi_1$ and $\varphi_2$).
Now, we have $g=xz$ for some $x\in P$ and $z\in C$. Then observing the action of $\varphi$ on the $2'$-part of $g$, we see $\varphi(z)=zy$ for some $y\in Z$ of odd order in case (2) or (3). Hence in cases (2) or (3a), there is some integer $c\geq 1$ such that $z^{2^c}=y$. Since $z$ has odd order, we therefore see that $g\in PZ$, so that $\overline{g}=1$, yielding that in cases (2) and (3a), $C_{\overline{N}}(\varphi)=1$. Similarly, in case (4), for $i=1,2$, we have $\varphi_i(z)=zy_i$ for some $y_i\in Z$ of odd order, so since $\gcd(p^{m_1}-1, p^{m_2}+1)_{2'}=1$, we also see in case (4) that $g\in PZ$ and $C_{\overline{N}}(\varphi_1, \varphi_2)=1$.
Now assume condition (3b) holds, so that $\epsilon=1$, $(p^m-\delta)_{2'}=\gcd(3, p^m-\delta)_{2'}$, and $E_6^\delta(p^m)$ and $E_6^\delta(p^m)_{ad}$ have self-normalizing Sylow $2$-subgroups. In this case, $z^{p^m-\delta}$ is an element of $Z$ of odd order. In particular, the image of $z$ in $G/Z$ may be viewed as a member of $E_6^\delta(p^m)$ which centralizes a Sylow $2$-subgroup $P_m$ of $E_6^\delta(p^m)$ contained in $PZ/Z$. But since $E_6^\delta(p^m)$ has a self-normalizing Sylow $2$-subgroup, we see that the image of $z$ in $G/Z$ is trivial, so that $\overline{g}=1$, and $C_{\overline{N}}(\varphi)=1$ in case (3b) as well.
Now, assume that none of (1) to (4) hold. Then $(q-\epsilon)_{2'}>(3, q-\epsilon)_{2'}$ by \prettyref{lem:whenSN2Ssimple} and $\phi$ is not contained in $Q$ up to conjugation in $\wt{G}$. Further, since the diagonal automorphisms of $S$ are odd, it follows that $Q$ does not contain a diagonal automorphism.
It suffices to exhibit a $z\in C$ such that $\varphi(z)=z$ for each field or graph-field automorphism $\varphi$ contained in $Q$. If $Q$ does not contain a field or graph-field automorphism, then we may choose such a $z$ to be a generator of $\overline{C}$. In particular, this is the case if $\epsilon=-1$, so we assume $\epsilon=1$. Write $q=p^a$. If $GQ/G$ is cyclic containing a field or graph-field automorphism, we may identify the generator of the subgroup of $Q$ consisting of field or graph-field automorphisms as $\phi^{\delta'}F_p^m$ for some $m|a$ and $\delta'\in\{0,1\}$. Similarly, if $Q$ contains both a field automorphism and graph-field automorphism, we may assume the corresponding subgroups are generated by $\varphi_1:=F_p^{m_1}$ and $\varphi_2:=\phi F_p^{m_2}$ for some $m_1, m_2|a$. Without loss, it suffices to assume that these generators generate the largest $2$-group of automorphisms possible without inducing conditions (3) and (4). In the cyclic case, we may choose $z$ to be an element of $\overline{C}\cap E^\delta_6(p^m)$. Otherwise, we may take $z$ to be a generator of the subgroup of $C$ of size $\gcd(p^{m_1}-1, p^{m_2}+1)_{2'}$, which can be viewed as a member of $\overline{C}$ since (3) does not hold. Hence $z$ is fixed by $Q$ and has the required form.
\end{proof}
\begin{theorem}\label{thm:typeE6}
The simple groups $E_6^\epsilon(q)$ are SN2S-Good for odd $q$.
\end{theorem}
\begin{proof}
We begin by showing \prettyref{cond:conjFI} holds for the simple groups $E_6^\epsilon(q)$ for $q$ odd. Let $A=SQ$ be an almost simple group obtained by adjoining a $2$-group $Q$ of automorphisms to $S=E_6^\epsilon(q)$. Suppose that $A$ does not contain a self-normalizing Sylow $2$-subgroup, so that $Q$ is not as in (1)-(4) of \prettyref{lem:whenSN2S}.
By \cite[Lemma 4.13]{navarro-tiep:2015:irreducible-representations-of-odd-degree}, we know that every odd-degree character of $G^\ast$ not lying over a unipotent character restricts irreducibly to $S$, and that every non-unipotent odd-degree character of $S$ can be obtained in this way. Hence it suffices to exhibit a $Q$-invariant $\wt{\chi}\in\mathrm{Irr}_{2'}(G^\ast)$ which is not fixed by $\sigma$. Letting $s$ be the pre-image in $C$ of the element $z$ of $\overline{C}$ obtained in the last paragraph of the proof of \prettyref{lem:whenSN2S}, we see that the semisimple character $\chi_s$ in the rational Lusztig series $\mathcal{E}(G^\ast, s)$ has odd degree, is $Q$-invariant, and is not fixed by $\sigma$ (see \cite[Corollary 2.4]{NavarroTiepTurullCyclo} and \cite[Lemma 3.4]{SFTaylorTypeA}).
Next we prove that \prettyref{cond:conjIF} holds for the simple groups $E_6^\epsilon(q)$ for $q$ odd. Let $Q$ be as in \prettyref{lem:whenSN2S}, so that $GQ/Z$ has a self-normalizing Sylow $2$-subgroup. Let $\chi\in\mathrm{Irr}_{2'}(G)$ be non-unipotent and fixed by $Q$, and suppose that the rational Lusztig series $\mathcal{E}(G,s)$ contains $\chi$. By \cite[Lemma 4.13]{navarro-tiep:2015:irreducible-representations-of-odd-degree}, $\chi$ extends to an irreducible character of $\wt{G}$. We claim that it suffices to show that the series $\mathcal{E}(G,s)$ is stabilized by $\sigma$. Indeed, in this case, note that $\wt{\chi}^\sigma\in\mathrm{Irr}(\wt{G}|\chi^\sigma)$ if $\wt{\chi}\in\mathrm{Irr}(\wt{G}|\chi)$. Note that $\wt{\chi}\in\mathcal{E}(\wt{G},\wt{s})$ for some pre-image $\wt{s}\in\wt{G}^\ast$ of $s$ by \cite[Corollaire 9.5]{bonnafe06}. Similarly, $\wt{\chi}^\sigma\in\mathcal{E}(\wt{G},\wt{s}t)$ for some $t\in Z(\wt{G}^\ast)$. So by \cite[Proposition 13.30]{dignemichel}, $\wt{\chi}^\sigma=\wt{\chi}' \otimes\wh{t}$ for some linear character $\wh{t}$ of $\wt{G}^\ast$ and $\wt{\chi}'\in\mathcal{E}(\wt{G},\wt{s})$. But the proof of \cite[Lemma 4.13]{navarro-tiep:2015:irreducible-representations-of-odd-degree} shows that the character degrees amongst members of $\mathcal{E}(\wt{G},\wt{s})$ lying above odd-degree characters of $G$ are distinct, so that $\wt{\chi}=\wt{\chi}'$ and $\chi^\sigma=\wt{\chi}^\sigma|_G=\wt{\chi}|_G=\chi$, proving the claim.
In particular, note that $\mathcal{E}(G,s)$ is stabilized by $\sigma$ in the case that $s$ has $2$-power order, by \cite[Lemma 3.4]{SFTaylorTypeA}. If case (1) of \prettyref{lem:whenSN2S} holds, then we are done since $C_{G^\ast}(s)$ must contain a Sylow $2$-subgroup of $G^\ast$, which is self-normalizing and hence self-centralizing. In the remaining cases, we may write $R_T^G(\lambda)$ for the Harish-Chandra series containing $\chi$, where $T=\bG{T}^F$ is a maximally split torus, by \cite[Lemma 7.9]{MalleSpathMcKay2}. By \prettyref{lem:lambdafixseriesfix}, it suffices to show that $\lambda^\sigma=\lambda$. However, since $\chi$ is fixed by $Q$, it is clear that $R_T^G(\lambda)$, and hence $\lambda$, is fixed by $Q$, since $\langle \phi, F_p\rangle$ stabilizes $T$. But notice that $\phi$ acts on $T$ via inversion and $F_p$ acts on $T$ via $t\mapsto t^{p}$, so that in cases (2), (3a), or (4), we see that $\lambda$ must have $2$-power order, so is fixed by $\sigma$. In case (3b), we may view $\lambda$ as a character of the maximally split torus $\bG{T}^{\varphi}$ of $\bG{G}^\varphi$, so that $\lambda\in\mathcal{E}(\bG{T}^\varphi, s)$ and $s$ has $2$-power order since $s\in (\bG{T}^\varphi)^\ast\leq (\bG{G}^\varphi)^\ast$, which has a self-normalizing Sylow $2$-subgroup.
\end{proof}
\section{Acknowledgements}
The author would first and foremost like to express her gratitude to Britta Sp{\"a}th, who suggested the main strategy of analyzing the action on Harish-Chandra series, provided a number of extremely useful conversations, and without whose help the original manuscript likely would not have existed. She is also extremely grateful to the anonymous referee, whose careful reading lead to detailed comments and suggestions that vastly improved and clarified the exposition, particularly in Section 2.
The author was supported in part by a grant through MSU Denver's Faculty Scholars Program, a grant from the Simons Foundation (Award \#351233), and an NSF-AWM Mentoring Travel Grant. She would like to thank G.\ Malle and B. Sp{\"a}th for their hospitality and helpful discussions during the research visit to TU Kaiserslautern supported by the latter grant, during which the majority of the work for this article was accomplished. She would also like to thank G. Malle, B. Sp{\"a}th, and J. Taylor for their extremely helpful comments on an earlier version of the manuscript.
Part of this was completed while the author was in residence at the Mathematical
Sciences Research Institute in Berkeley, California during the Spring 2018 semester program
on Group Representation Theory and Applications, supported by the National Science Foundation
under Grant No. DMS-1440140. She thanks the institute and the organizers of the program for
making her stay possible and providing a collaborative and productive work environment.
In addition, the author would like to thank P. H. Tiep and G. Navarro for suggesting this problem and their kindness and many helpful conversations throughout the three articles detailing her work on \prettyref{conj:mainprob}.
\bibliographystyle{alpha}
|
\section{Introduction}\label{sec:intro}
In two recent studies the subject of {\em semi-inclusive} charge-changing neutrino (CC$\nu$) reactions with nuclei \cite{semi} and application to the special case of deuterium \cite{deut} were presented. Analogous to the semi-inclusive reaction $(e,e'x)$ where one assumes that the scattered electron and some particle $x$ are detected in coincidence, in the weak interaction case one considers reactions of the type $(\nu_\ell, \ell^- x)$ and $(\bar{\nu}_\ell, \ell^+ x)$. These involve incident neutrinos or anti-neutrinos of specific flavor ($\ell = e$, $\mu$ or $\tau$) together with coincident detection of the corresponding charged leptons and some particle $x$. In the present work we shall focus on nucleons ejected from the nucleus, and hence $x=N$, where $N=p$ or $n$. Note that in the nuclear case the ``natural'' type of nucleon may not be the one of interest, whereas for a single-nucleon target and when no other particle is produced other that the final-state nucleon ({\it i.e.,} no pion production, kaon production, {\it etc.}) charge conservation forces the final-state nucleon to be only of one type. Namely, in this latter case one only has reactions of the type $\nu_\ell + n\rightarrow \ell^- + p$ and $\bar{\nu}_\ell + p \rightarrow \ell^+ + n$. In the present work we shall specialize still further and consider only incident neutrinos, final-state negative leptons and emission of protons ($x=p$). For completeness in defining the terminology commonly being used, we note that reactions where only the final-state leptons are detected, such as $(e,e')$, $(\nu_\ell, \ell^-)$ or $(\bar{\nu}_\ell, \ell^+)$, are called {\em inclusive} reactions.
As has become quite clear in recent years, the typical high-energy neutrino beams used in studies of neutrino oscillations, typically at neutrino energies of around a GeV to tens of GeV, $E_\nu$, have rather broad spreads in energy, which introduces model dependence in the specification of the distance over energy ratio $L/E_\nu$ that enters in the standard oscillation expressions. However, as discussed in \cite{deut}, deuterium provides, at least in principle, an exception to the typical case of heavier nuclei. Namely, once so-called ``no-pion'' events are isolated, all that can occur for the case of incident neutrinos is the reaction $\nu_\ell + {}^2$H$ \rightarrow \ell^- + p + p$. Upon detecting two of the three particles in the final state and knowing the direction of the incident neutrino the neutrino's energy can be reconstructed using nothing beyond the kinematics of the reaction. In \cite{deut} a specific relativistic model for the deuterium ground state and final $NN$ scattering state was employed to model this reaction; in the present study we use the same model for the $A=2$ states and the required electroweak current matrix elements.
This said, there are still practical issues of which to be aware. Namely, making very large target/detectors of hydrogen or deuterium is problematical because of the safety issues involved and the difficulty of providing very large amounts of these nuclei. Using target/detectors of something involving large fractions of deuterium together with other light nuclei, such as heavy water (D$_2$O) or deuterated methane (CD$_4$), might alleviate the safety issue and could provide practical amounts of deuterium, although having other nuclei such as oxygen or carbon present potentially can bring in new considerations. In this study we have focused on a specific case to explore how such mixed nuclear cases behave; specifically, here we consider the case of ${}^2$H$_2 {}^{16}$O. The goal is to take what we have already done for deuterium, add model results for CC$\nu$ semi-inclusive reactions on ${}^{16}$O and determine the degree to which events from the two nuclear species can be separated. One expects the deuteron events to be very peaked and to occur in a different part of the kinematic space involved from the oxygen events, and, as well, the oxygen events to be much more spread in the appropriate kinematic variables so that the ratio of deuterium to oxygen becomes quite favorable. Indeed, we shall show that this is the case.
We will be drawing on our previous study of semi-inclusive CC$\nu$ reactions in \cite{semi} to highlight and quantify the differences of deuterium and a more typical nucleus such as oxygen (here the nucleus could be chosen to be carbon or any other relatively light nucleus). As a specific model for the oxygen case we employ the spectral function approach of \cite{Benhar:1994hw,Benhar:2005dj}. The goal will be to optimize the selection of semi-inclusive events for the case of deuterium and then see what emerges for the ``background'' from the oxygen events.
The paper is organized as follows. In Sect.~\ref{sec:semi} we summarize the necessary formalism for the semi-inclusive CC$\nu$ reaction, taking as a basis the previous study reported in \cite{semi}, and include some of the relevant formalism needed to inter-relate the experimental ``lab frame'' to the so-called ``$q$-frame''. In Sect.~\ref{sec:deut} we specialize the results of the previous section to the case of deuterium to make very clear the advantage provided by this particular nucleus. We do not repeat the discussion of the formalism for the dynamics and currents involved in the deuterium case, since these have been reported in \cite{deut}. For the case of oxygen we present the required formalism in the context of the spectral function in Sect.~\ref{sec:spect}, following which we employ the two models discussed above to obtain typical results for heavy water and present these in Sect.~\ref{sec:results}. In Sect.~\ref{sec:concl} we offer our conclusions, while in the Appendix we collect expressions for the off-shell single-nucleon response functions employed for the oxygen spectral function case.
\section{Semi-Inclusive Cross Section}\label{sec:semi}
Semi-inclusive CC$\nu$ scattering is represented by the Feynman diagram shown in Fig.~\ref{fig:semi}, where $Q^{\mu}$ is the four-momentum of the W-boson,
\begin{figure}
\centerline{\includegraphics[height=2in]{semi.pdf}}
\caption{(color online) Feynman diagram for semi-inclusive charge-changing neutrino reactions involving a target nucleus with nucleon number $A$ with emission and detection of a nucleon with four-momentum $P_N^{\mu}$ together with detection of a final-state charged lepton with four-momentum $K'^{\mu}$ }
\label{fig:semi}
\end{figure}
\begin{equation}
K^\mu=(\varepsilon,\bm{k})
\end{equation}
is the incident lepton four-momentum and
\begin{equation}
{K'}^\mu=(\varepsilon',\bm{k}')
\end{equation}
is the four-momentum of the lepton in the final state,
where $\varepsilon=\sqrt{k^2+m^2}$ and $\varepsilon'=\sqrt{{k'}^2+{m'}^2}$ are the energies of the incident and final leptons with respective masses $m$ and $m'$. Then the four-momentum transfer is
\begin{equation}
Q^\mu=K^\mu-{K'}^\mu=(\varepsilon-\varepsilon',\bm{k}-\bm{k}')=(\omega,\bm{q})\,.
\end{equation}
The four-momentum of the target nucleus with nucleon number $A$ can be written in its rest frame as
\begin{equation}
P_A^\mu=(M_A,\bm{0})\,.
\end{equation}
The four-momentum of the detected nucleon is
\begin{equation}
P_N^\mu=(E_N,\bm{p}_N)\,,
\end{equation}
where $m_N$ is the nucleon mass, $E_N \equiv \sqrt{\bm{p}_N^2+m_N^2}$ and
the four-momentum of the residual $A-1$ system is
\begin{equation}
P_{A-1}^\mu=(\sqrt{\bm{p}_m^2+W_{A-1}^2},\bm{p}_m)
\end{equation}
with the invariant mass $W_{A-1}$.
The energy of an incoming neutrino can be determined by measuring the three-momenta of the outgoing charged lepton, which we take to be a muon in what follows (although clearly the $e$ or $\tau$ cases can also be considered), and nucleon, corresponding to kinematics B of \cite{semi}. In this case the four-fold differential cross section in the laboratory frame is then
\begin{align}
\frac{d\sigma}{dk'd\Omega_{k'}dp_Nd\Omega_N^L}=&\frac{G^2\cos^2\theta_cm_N{k'}^2p_N^2W_{A-1}}{2(2\pi)^5k\varepsilon'E_N}\int \frac{d^3p_m}{\sqrt{p_m^2+W_{A-1}^2}}\eta_{\mu\nu}W^{\mu\nu}\nonumber\\
&\times\delta^4(K+P_A-K'-P_N-P_{A-1})\nonumber\\
=&\frac{G^2\cos^2\theta_cm_N{k'}^2p_N^2W_{A-1}}{2(2\pi)^5k\varepsilon'E_N}\int \frac{d^3p_m}{\sqrt{p_m^2+W_{A-1}^2}}\eta_{\mu\nu}W^{\mu\nu}\nonumber\\
&\times\delta(\varepsilon+M_A-\varepsilon'-E_N-\sqrt{p_m^2+W_{A-1}^2})\delta(\bm{k}-\bm{k}'-\bm{p}_N-\bm{p}_m) \, ,\label{eq:dsig0}
\end{align}
where $G$ is the weak interaction coupling constant and $\theta_c$ is the Cabibbo mixing angle. Defining
\begin{equation}
E_B=\varepsilon'+E_N-M_A
\end{equation}
and
\begin{equation}
\bm{p}_B=\bm{k}'+\bm{p}_N \, ,
\end{equation}
the cross section becomes
\begin{align}
\frac{d\sigma}{dk'd\Omega_{k'}dp_Nd\Omega_N^L}
=&\frac{G^2\cos^2\theta_cm_N{k'}^2p_N^2W_{A-1}}{2(2\pi)^5k\varepsilon'E_N}\int \frac{d^3p_m}{\sqrt{p_m^2+W_{A-1}^2}}\eta_{\mu\nu}W^{\mu\nu}\nonumber\\
&\times\delta(\varepsilon-E_B-\sqrt{p_m^2+W_{A-1}^2})\delta(\bm{k}-\bm{p}_B+\bm{p}_m)\nonumber\\
=&\frac{G^2\cos^2\theta_cm_N{k'}^2p_N^2W_{A-1}}{2(2\pi)^5k\varepsilon'E_N} \frac{1}{\sqrt{(\bm{p}_B-\bm{k})^2+W_{A-1}^2}}\eta_{\mu\nu}W^{\mu\nu}\nonumber\\
&\times\delta(\varepsilon-E_B-\sqrt{(\bm{p}_B-\bm{k})^2+W_{A-1}^2}) \, .
\end{align}
Using the remaining $\delta$-function, the incident neutrino momentum and energy are given by
\begin{equation}
k_0=\frac{1}{a_B}\left(X_Bp_B\cos\theta_B+E_B\sqrt{X_B^2+m^2a_B}\right) \label{Xk0}
\end{equation}
and
\begin{equation}
\varepsilon_0=\frac{1}{a_B}\left(E_BX_B+p_B\cos\theta_B\sqrt{X_B^2+m^2a_B}\right)\,, \label{Xe0}
\end{equation}
where
\begin{equation}
X_B=\frac{1}{2}\left(p_B^2-E_B^2+W_{A-1}^2-m^2\right)
\end{equation}
and
\begin{equation}
a_B=p_B^2\cos^2\theta_B-E_B^2\,.
\end{equation}
The energy-conserving $\delta$-function can be rewritten as
\begin{equation}
\delta(\varepsilon-E_B-\sqrt{(\bm{p}_B-\bm{k})^2+W_{A-1}^2})=\frac{\varepsilon_0\sqrt{(\bm{p}_B-\bm{k})^2+W_{A-1}^2}}{\sqrt{X_B^2+m^2a_B}}\delta(k-k_0)\,.
\end{equation}
The cross section then becomes
\begin{equation}
\frac{d\sigma}{dk'd\Omega_{k'}dp_N^2d\Omega_N^L}
=\frac{G^2\cos^2\theta_cm_N{k'}^2\varepsilon\, p_N^2W_{A-1}v_0}{2(2\pi)^5k\varepsilon'E_N\sqrt{X_B^2+m^2a_B}} \mathcal{F}^2_\chi
\delta(k-k_0) \, ,\label{eq:sigma_lab}
\end{equation}
where $\mathcal{F}^2_\chi \equiv \eta_{\mu\nu}W^{\mu\nu}/v_0$ with $v_0 \equiv (\varepsilon + \varepsilon')^2 - q^2$. The resulting response may be written
\begin{align}
{\cal F}_\chi^2=&\hat{V}_{CC}(w_{CC}^{VV(I)}+w_{CC}^{AA(I)})
+2\hat{V}_{CL}(w_{CL}^{VV(I)}+w_{CL}^{AA(I)})
+\hat{V}_{LL}(w_{LL}^{VV(I)}+w_{LL}^{AA(I)})\nonumber\\
&+\hat{V}_T(w_{T}^{VV(I)}+w_{T}^{AA(I)})\nonumber\\
&+\hat{V}_{TT}\left[(w_{TT}^{VV(I)}+w_{TT}^{AA(I)})\cos 2\phi_N+(w_{TT}^{VV(II)}+w_{TT}^{AA(II)})\sin 2\phi_N\right]\nonumber\\
&+\hat{V}_{TC}\left[(w_{TC}^{VV(I)}+w_{TC}^{AA(I)})\cos\phi_N
+(w_{TC}^{VV(II)}+w_{TC}^{AA(II)})\sin\phi_N)\right]\nonumber\\
&+\hat{V}_{TL}\left[(w_{TL}^{VV(I)}+w_{TL}^{AA(I)})\cos\phi_N
+(w_{TL}^{VV(II)}+w_{TL}^{AA(II)})\sin\phi_N\right]
\nonumber\\
&+\chi\left[\hat{V}_{T'}w^{VA(I)}_{T'}+\hat{V}_{TC'}(w^{VA(I)}_{TC'}\sin\phi_N
+w_{TC'}^{VA(II)}\cos\phi_N)\right.\nonumber\\
&\left.+\hat{V}_{TL'}(w^{VA(I)}_{TL'}\sin\phi_N
+w^{VA(II)}_{TL'}\cos\phi_N)\right]\label{eq:cal_F2_chi}
\end{align}
with
\begin{equation}
\chi=\left\{
\begin{array}{rl}
-1 &\mathrm{for\ neutrinos}\\
1 &\mathrm{for\ antineutrinos}
\end{array}\right.\,.
\end{equation}
The kinematic functions $V_a$ and response functions $w^i_j$ are as defined in \cite{semi} with the explicit dependence on azimuthal angle $\phi_N$ defined in the $q$-fixed frame. Response functions labeled by the superscript $(II)$ vanish in the plane-wave limit.
If the neutrino momentum distribution normalized to unity is designated as $P(k)$,
the cross section weighted by this distribution is then given by
\begin{align}
\left<\frac{d\sigma}{dk'd\Omega_{k'}dp_Nd\Omega_N^L}\right>
=&\int_0^\infty dk \frac{G^2\cos^2\theta_cm_N{k'}^2\varepsilon\, p_N^2W_{A-1}}{2(2\pi)^5k\varepsilon'E_N\sqrt{X_B^2+m^2a_B}} v_0 \mathcal{F}^2_\chi
\delta(k-k_0)P(k)\nonumber\\
=&\frac{G^2\cos^2\theta_cm_N{k'}^2\varepsilon_0\, p_N^2W_{A-1}v_0}{2(2\pi)^5k_0\varepsilon'E_N\sqrt{X_B^2+m^2a_B}} \mathcal{F}^2_\chi
P(k_0)\, .\label{eq:sigma_lab}
\end{align}
Next it is useful to inter-relate the variables in the laboratory frame shown in Fig.~\ref{fig:kinelab} to those in the so-called $q$-system shown in Fig.~\ref{fig:kineq}.
\begin{figure}
\centerline{\includegraphics[height=3in]{lab_system_color.pdf}}
\caption{(color online) Semi-inclusive $(\nu_{\mu},\mu^-p)$ CC$\nu$ reaction in the laboratory frame. Here the incident neutrino with three-momentum ${\bm k}$ is along the $3'$ direction, the neutrino and the final-state muon with three-momentum ${\bm k}'$ lie in the $1'$--$3'$ plane and the normal to the plane defines the $2'$ direction. The outgoing nucleon (here a proton) has three-momentum ${\bm p}_N$ and is traveling in the direction characterized by polar angle $\theta_N^L$ and azimuthal angle $\phi_N^L$ in the lab system, as shown.}
\label{fig:kinelab}
\end{figure}
\begin{figure}
\centerline{\includegraphics[height=3in]{q_system_color.pdf}}
\caption{(color online) Semi-inclusive $(\nu_{\mu},\mu^-p)$ CC$\nu$ reaction in the $q$-system. Here the three-momentum transfer ${\bm q}$ defines the $3$ direction, the neutrino and the final-state muon lie in the $1$--$3$ plane and the normal to the plane defines the $2$ direction. The outgoing nucleon (here a proton) has three-momentum ${\bm p}_N$ and is traveling in the direction characterized by polar angle $\theta_N$ and azimuthal angle $\phi_N$ in the $q$-system, as shown.}
\label{fig:kineq}
\end{figure}
We have the following identities relating the angles in the two systems:
\begin{align}
\cos\theta_N=&\cos\theta_N^L \cos\theta_{q}-\cos\phi^L_N \sin\theta_N^L \sin\theta_{q}\label{eq:cos_theta_N}\\
\sin\theta_N=&\sqrt{1-\cos^2\theta_N}\label{eq:sin_theta_N}\\
\cos\phi_N=&\frac{\cos\phi^L_N\sin\theta_N^L\cos\theta_{q}+\cos\theta_N^L\sin\theta_{q}}
{\sin\theta_N}\label{eq:cos_phi_N}\\
\sin\phi_N=&\frac{\sin\phi_N^L\sin\theta_N^L}{\sin\theta_N}\label{eq:sin_phi_N}
\end{align}
and the inverse relations are given by
\begin{align}
\cos\theta_N^L=&\cos\theta_N \cos\theta_{q}+\cos\phi_N \sin\theta_N \sin\theta_{q}\label{eq:cos_theta_NL}\\
\sin\theta_N^L=&\sqrt{1-\cos^2\theta_N^L}\label{eq:sin_theta_NL}\\
\cos\phi^L_N=&\frac{\cos\phi_N\sin\theta_N\cos\theta_{q}+\cos\theta_N\sin\theta_{q}}{\sin\theta_N^L}\label{eq:cos_phi_NL}\\
\sin\phi^L_N=&\frac{\sin\phi_N\sin\theta_N}{\sin\theta_N^L}\label{eq:sin_phi_NL}\,.
\end{align}
Note that as the neutrino energy changes, even for fixed directions for the outgoing muon and nucleon, the direction of the momentum transfer also changes, and, therefore, through these relationships, the polar and azimuthal angles in the $q$-system also change. The lab system is relevant when experimental issues are being considered; however, the $q$-system with the $3$-direction along the momentum of the exchanged boson has special symmetries that are masked in the lab system.
Here, we want to express the cross section in lab frame. This can be done by using Eqs.~(\ref{eq:cos_phi_N}) and (\ref{eq:sin_phi_N}) to replace the azimuthal angular dependence in Eq.~(\ref{eq:cal_F2_chi}) and by defining the three-momenta
\begin{equation}
\bm{k}=k\hat{\bm{u}}_{3'}\,,
\end{equation}
\begin{equation}
\bm{k}'=k'\left(\sin\theta_l\hat{\bm{u}}_{1'}+\cos\theta_l\hat{\bm{u}}_{3'}\right)
\end{equation}
and
\begin{equation}
\bm{p}_N=p_N\left(\cos\phi_N^L\sin\theta_N^L\hat{\bm{u}}_{1'}+\sin\phi_N^L\sin\theta_N^L\hat{\bm{u}}_{2'}+\cos\theta_N^L\hat{\bm{u}}_{3'}\right)\,,
\end{equation}
where $\theta_l$ is the lepton scattering angle. The unit vectors in the lab frame are $\hat{\bm{u}}_{1'}$, $\hat{\bm{u}}_{2'}$ and $\hat{\bm{u}}_{3'}$, as shown in Fig.~\ref{fig:kinelab}. The three-momentum transfer is
\begin{equation}
\bm{q}=\bm{k}-\bm{k}'
\end{equation}
and its square is
\begin{equation}
q^2=k^2+{k'}^2-2kk'\cos\theta_l.
\end{equation}
The angle between $\bm{k}$ and $\bm{q}$ can be obtained from
\begin{equation}
\bm{k}\cdot\bm{q}=kq\cos\theta_{q}=\bm{k}\cdot\bm{k}-\bm{k}\cdot\bm{k}'=k^2-kk'\cos\theta_l \, ,
\end{equation}
which can be solved to yield
\begin{equation}
\cos\theta_{q}=\frac{k-k'\cos\theta_l}{q}\,.
\end{equation}
Similarly we can use
\begin{equation}
\bm{k}\cdot\bm{p}_B=kp_B\cos\theta_B=\bm{k}\cdot(\bm{k}'+\bm{p}_N)=(kk'\cos\theta_l+kp_N\cos\theta_N^L)
\end{equation}
to obtain
\begin{equation}
p_B\cos\theta_B=k'\cos\theta_l+p_N\cos\theta_N^L\,.
\end{equation}
The remaining expressions needed to obtain the cross section in the lab frame are
\begin{equation}
p_B^2={k'}^2+p_N^2+2\bm{k}'\cdot\bm{p}_N={k'}^2+p_N^2+2k'p_N\left(\cos\phi_N^L\sin\theta_N^L\sin\theta_l+\cos\theta_N^L\cos\theta_l\right)
\end{equation}
and
\begin{equation}
p_m^2=k^2+p_B^2-2\bm{k}\cdot\bm{p}_B=k^2+p_B^2-2kp_B\cos\theta_B\,.
\end{equation}
As noted in \cite{deut}, Eq.~(\ref{eq:sigma_lab}) applies also to the case of exclusive scattering from the deuteron by making the substitutions $M_A\rightarrow M_d$ and $W_{A-1}\rightarrow m_N$.
\section{Deuterium}\label{sec:deut}
For the purpose of determining whether the deuterium cross section can be separated from that of oxygen, we wish to choose kinematics which are optimal for the deuteron and then use the values $\bm{k}'$ and $\bm{p}_N$ determined from the deuteron in calculating the semi-inclusive scattering from oxygen.
To obtain the optimal kinematics for scattering from the deuteron we start with Mandelstam $s$ for the virtual $W$ and the deuteron. This is given by
\begin{equation}
s=(P_d+Q)^2=(M_d+\omega)^2-q^2\,.
\end{equation}
The scaling variables \cite{Day:1990mf}
\begin{equation}
y=\frac{(M_d+\omega)\sqrt{s(s-4m_N^2)}}{2s}-\frac{q}{2}
\end{equation}
and
\begin{equation}
Y=y+q
\end{equation}
can be used to obtain limiting values for the magnitude of the missing momentum $p_m$ as
\begin{equation}
|y|\leq p_m\leq Y\,.
\end{equation}
Since the deuteron cross section behaves roughly as the deuteron momentum distribution $n(p_m)$, which peaks at $p_m=0$, the cross section can be optimized by choosing kinematics such that $y=0$. Solving this for the incident neutrino energy yields
\begin{align}
\varepsilon_0=&\frac{1}{2
\left[(\varepsilon'-M_d+m_N)^2-{k'}^2 \cos^2(\theta_l)\right]}\left\{\zeta k' \cos(\theta_l)\left[ -2
{\varepsilon'}^2 \left(m^2-2
(M_d-m_N)^2\right)\right.\right.\nonumber\\
&\left.\left.-4\varepsilon' (M_d-m_N)\left(-m^2+M_d^2-2 M_d m_N+{m'}^2\right)+2
{k'}^2 m^2 \cos (2\theta_l)+m^4-2 m^2 M_d^2\right.\right.\nonumber\\
&\left.\left.+4 m^2 M_d
m_N-4 m^2
m_N^2+M_d^4-4 M_d^3
m_N+4 M_d^2 m_N^2+2
M_d^2 {m'}^2-4 M_d
m_N
{m'}^2+{m'}^4\right]^{\frac{1}{2}}\right.\nonumber\\
&\left.-2
{\varepsilon'}^2 M_d+2
{\varepsilon'}^2 m_N+\varepsilon'
m^2+3 \varepsilon' M_d^2-6
\varepsilon' M_d m_N+2
\varepsilon' m_N^2+\varepsilon'
{m'}^2-m^2 M_d+m^2
m_N\right.\nonumber\\
&\left.-M_d^3+3 M_d^2
m_N-2 M_d
m_N^2-M_d
{m'}^2+m_N {m'}^2\right\} \, ,
\end{align}
where
\begin{equation}
\zeta=\left\{
\begin{array}{rr}
-1 & \mathrm{for}\ \theta_l\leq \frac{\pi}{2}\\
1 & \mathrm{for}\ \theta_l> \frac{\pi}{2}
\end{array}\right.
\end{equation}
Four-momentum conservation for the deuteron requires that
\begin{align}
0=&M_d+\omega-\sqrt{p_N^2+m_N^2}-\sqrt{p_m^2+m_N^2}\label{eqn:deuteron_energy}\\
\bm{0}=&\bm{q}-\bm{p}_N+\bm{p}_m \, .\label{eqn:deuteron_momentum}
\end{align}
Using Eq.~(\ref{eqn:deuteron_energy}) the square of the detected nucleon momentum is
\begin{equation}
p_N^2=\left(M_d+\omega-\sqrt{p_m^2+m_N^2}\right)^2-m_N^2\,.\label{eq:p_N_squared}
\end{equation}
Using Eq.~(\ref{eqn:deuteron_momentum}),
\begin{equation}
\bm{p}_m=\bm{q}-\bm{p}_N\,,
\end{equation}
yields
\begin{equation}
p_{m}^2=q^2+p_N^2-2p_Nq\cos\theta_N\,.
\end{equation}
Solving this for $\cos\theta_N$ gives
\begin{equation}
\cos\theta_N=\frac{q^2+p_N^2-p_m^2}{2p_Nq}\,.\label{eq:cos_theta_N_2}
\end{equation}
By specifying $k'$, $\theta_l$, $\phi_N$ and using Eqs.~(\ref{eq:p_N_squared}) and (\ref{eq:cos_theta_N_2}), the lab frame angles are then given by Eqs.~(\ref{eq:cos_theta_NL}), (\ref{eq:sin_theta_NL}), (\ref{eq:cos_phi_NL}) and (\ref{eq:sin_phi_NL}). This provides a complete set of input variables to evaluate the deuteron and oxygen cross sections. Note that Eq.~(\ref{eq:cos_theta_N_2}) results in a correlation of the values of $p_N$ and $\cos\theta^L_N$.
All of the conditions required by these constrained kinematics can only be satisfied by limiting
\begin{equation}
0\leq\theta_l\leq\left\{
\begin{array}{ll}
\cos^{-1}\left(\frac{\varepsilon'-M_d+m_N}{k'}\right)& \mathrm{for}\ -k'<\varepsilon'-M_d+m_N\leq k'\\
\pi & \mathrm{for}\ \varepsilon'-M_d+m_N\leq -k'
\end{array}
\right.\,.\label{eq:theta_l_max}
\end{equation}
The deuterium matrix elements needed to construct the cross section are described in \cite{deut}.
\section{Spectral Function}\label{sec:spect}
\begin{figure}
\centerline{\includegraphics[height=2in]{PWIA_semi.pdf}}
\caption{(color online) Feynman diagram for a factorized approximation to the semi-inclusive charge-changing neutrino reaction illustrated for the general case in Fig.~1. }
\label{fig:PWIA_semi}
\end{figure}
For this work we estimate the oxygen semi-inclusive cross sections using a factorized spectral function model. The current matrix element for this model can be written as
\begin{equation}
\left<\bm{p}_N,s_N;\bm{P}_{A-1},s_{A-1}\left|J^\mu(q)\right|\bm{P}_A,s_A\right>=\bar{u}(\bm{p}_N,s_N)_a J^\mu(q)_{ab}\Psi(P_{A-1},s_{A-1};P_A,s_A)_{bc} \, ,
\end{equation}
where $s_N$, $s_A$ and $s_{A-1}$ are the spins of the ejected proton, target nucleus and residual system, respectively, and $\Psi(P_{A-1},s_{A-1};P_A,s_A)$ represents a three-point function with the $A$ line truncated. The Dirac indices are explicitly indicated. The nuclear response tensor is then given by
\begin{align}
W^{\mu\nu}=&\sum_{s_N}\sum_{s_A}\sum_{s_{A-1}}
\bar{u}(\bm{p}_N,s_N)_a J^\nu(q)_{ab}\Psi(P_{A-1},s_{A-1};P_A,s_A)_{bc}\nonumber\\
&\times\bar{\Psi}(P_{A-1},s_{A-1};P_A,s_A)_{cd} J^\mu(-q)_{de}u(\bm{p}_N,s_N)_e\nonumber\\
=&\sum_{s_N}
\bar{u}(\bm{p}_N,s_N)_a J^\nu(q)_{ab}\frac{1}{8\pi}\Lambda^+(\bm{p}_m)_{bd}S(p_m,E_m) J^\mu(-q)_{de}u(\bm{p}_N,s_N)_e\nonumber\\
=&\frac{1}{8\pi} \mathrm{Tr}\left[J^\mu(-q)\Lambda^+(\bm{p}_N)J^\nu(q)\Lambda^+(\bm{p}_m)\right]S(p_m,E_m)\nonumber\\
=&\frac{1}{8\pi} w^{\mu\nu}(P_{A}-P_{A-1},Q)S(p_m,E_m)\,,
\end{align}
where $w^{\mu\nu}(P_{A}-P_{A-1},Q)$ is an off-shell single-nucleon response tensor and $S(p_m,E_m)$ is the spectral function.
The missing energy is approximated by
\begin{equation}
E_m\cong E_s+\mathcal{E} \, ,
\end{equation}
where $E_s$ is the separation energy,
\begin{equation}
\mathcal{E}=\sqrt{p_m^2+W_{A-1}^2}-\sqrt{p_m^2+{W_{A-1}^0}^2}\,,
\end{equation}
and $W_{A-1}^0$ is the invariant mass of the lowest state of the residual system. Energy conservation requires that
\begin{align}
0=&M_A+\omega-\sqrt{p_N^2+m_N^2}-\sqrt{p_m^2+W_{A-1}^2}\nonumber\\
=&M_A+\omega-\sqrt{p_N^2+m_N^2}-\sqrt{p_m^2+W_{A-1}^2}+\sqrt{p_m^2+{W_{A-1}^0}^2}-\sqrt{p_m^2+{W_{A-1}^0}^2}\nonumber\\
=&M_A+\omega-\sqrt{p_N^2+m_N^2}-\mathcal{E}-\sqrt{p_m^2+{W_{A-1}^0}^2}\,.
\end{align}
So $\mathcal{E}$ can also be written as
\begin{equation}
\mathcal{E}=M_A+\omega-\sqrt{p_N^2+m_N^2}-\sqrt{p_m^2+{W_{A-1}^0}^2}\,.
\end{equation}
From momentum conservation $\bm{p}_N=\bm{q}-\bm{p}_m$, and therefore
\begin{equation}
\mathcal{E}=M_A+\omega-\sqrt{(\bm{q}-\bm{p}_m)^2+m_N^2}-\sqrt{p_m^2+{W_{A-1}^0}^2}\,.
\end{equation}
The range of $\mathcal{E}$ is then limited by
\begin{equation}
\mathcal{E}_+\leq\mathcal{E}\leq\mathcal{E}_-\,,
\end{equation}
where
\begin{equation}
\mathcal{E}_-=M_A+\omega-\sqrt{(p_m-q)^2+m_N^2}-\sqrt{p_m^2+{W_{A-1}^0}^2}
\end{equation}
and
\begin{equation}
\mathcal{E}_+=\max(M_A+\omega-\sqrt{(p_m+q)^2+m_N^2}-\sqrt{p_m^2+{W_{A-1}^0}^2},0)\,.
\end{equation}
The normalization of the spectral function $S(p_m,E_m)$ is defined here such that
\begin{equation}
\int_0^\infty dE_m S(p_m,E_m)=n(p_m)
\end{equation}
is the momentum distribution and
\begin{equation}
\frac{1}{(2\pi)^3}\int_0^\infty dp_m p_m^2 n(p_m)=A-Z\,.
\end{equation}
Expressing the four-momentum of the struck nucleon as
\begin{equation}
P_A^\mu-P_{A-1}^\mu=(M_A-\sqrt{p_m^2+W_{A-1}^2},-\bm{p}_m)\,,
\end{equation}
defining
\begin{equation}
\bm{p}=-\bm{p}_m
\end{equation}
and using energy conservation
\begin{equation}
M_A-\sqrt{p^2+W_{A-1}^2}=\sqrt{p_N^2+m_N^2}-\omega\,,
\end{equation}
one has
\begin{align}
P_A^\mu-P_{A-1}^\mu=&(\sqrt{p_N^2+m_N^2}-\omega,\bm{p})\nonumber\\
=&(\sqrt{p_N^2+m_N^2}-\omega-\sqrt{p^2+m_N^2}+\sqrt{p^2+m_N^2},\bm{p})\nonumber\\
=&(\sqrt{p_N^2+m_N^2}-\omega-\sqrt{p^2+m_N^2},\bm{0})+(\sqrt{p^2+m_N^2},\bm{p})\nonumber\\
=&(\delta,\bm{0})+(\sqrt{p^2+m_N^2},\bm{p})=\Delta^\mu+P^\mu \, ,
\end{align}
where
\begin{equation}
P^{\mu}=(\sqrt{p^2+m_N^2},\bm{p})
\end{equation}
is an on-shell four-vector and
\begin{equation}
\Delta^\mu=(\delta,\bm{0})
\end{equation}
is off-shell with
\begin{equation}
\delta=\sqrt{p_N^2+m_N^2}-\sqrt{p^2+m_N^2}-\omega\,.
\end{equation}
The quantity $\mathcal{F}^2_\chi$ in Eq.~(\ref{eq:sigma_lab}) is then given by
\begin{equation}
\mathcal{F}^2_\chi\cong\frac{1}{8\pi}\widetilde{\mathcal{F}}^2_\chi S(p_m,E_m) \, ,
\end{equation}
where
\begin{align}
\widetilde{\mathcal{F}}^2_\chi=& \widehat{V}_{CC}\left(\widetilde{w}^{VV(I)}_{CC}+\widetilde{w}^{AA(I)}_{CC}\right)+2\widehat{V}_{CL}\left(\widetilde{w}^{VV(I)}_{CL}+\widetilde{w}^{AA(I)}_{CL}\right)+\widehat{V}_{LL}\left(\widetilde{w}^{VV(I)}_{LL}+\widetilde{w}^{AA(I)}_{LL}\right)\nonumber\\
&+\widehat{V}_{T}\left(\widetilde{w}^{VV(I)}_{T}+\widetilde{w}^{AA(I)}_{T}\right)
+\widehat{V}_{TT}\left(\widetilde{w}^{VV(I)}_{TT}+\widetilde{w}^{AA(I)}_{TT}\right)\cos 2\phi_N \nonumber\\
&+\widehat{V}_{TC}\left(\widetilde{w}^{VV(I)}_{TC}+\widetilde{w}^{AA(I)}_{TC}\right)\cos\phi_N+\widehat{V}_{TL}\left(\widetilde{w}^{VV(I)}_{TL}+\widetilde{w}^{AA(I)}_{TL}\right)\cos\phi_N \nonumber\\
&
+\chi \left[ \widehat{V}_{T^{\prime }}\widetilde{w}^{VA(I)}_{T^{\prime }}+\widehat{V}_{TC^{\prime }}\widetilde{w}^{VA(I)}_{TC^{\prime }}\sin\phi_N+\widehat{V}_{TL^{\prime }}\widetilde{w}^{VA(I)}_{TL^{\prime }}\sin\phi_N\right] \, .
\end{align}
The off-shell single-nucleon response functions $\widetilde{w}^i_j$ are listed in the Appendix.
Since the invariant mass of the residual $A-1$ system is not measured, it is necessary that the semi-inclusive cross section be integrated over all possible values of $W_{A-1}$ to give
\begin{align}
\left<\frac{d\sigma}{dk'd\Omega_{k'}^Ldp_Nd\Omega_N^L}\right>
=&\int_{W^0_{A-1}}^{\infty}dW_{A-1}\frac{G^2\cos^2\theta_cm_N{k'}^2\varepsilon_0 p_N^2W_{A-1}v_0}{8(2\pi)^6k_0\varepsilon'E_N\sqrt{X_B^2+m^2a_B}} \widetilde{\mathcal{F}}^2_\chi S(p_m,E_m)
P(k_0) \, ,
\end{align}
where $W^0_{A-1}$ is the lowest possible mass for the residual system which in some cases may not be a bound state. For the specific case considered in the present study this corresponds to the ground-state mass of $^{15}$O.
Note that the integral over the invariant mass requires that $k_0$ and $\varepsilon_0$ in Eqs.~(\ref{Xk0}) and (\ref{Xe0}) must take on a range of values rather than being fixed as in the case of the deuteron.
\section{Results}\label{sec:results}
For the purposes of this paper, we have chosen to weight the cross sections using the flux momentum distribution for the DUNE experiment \cite{Alion:2016uaj} normalized to unit area, represented $P(k)$ as shown in Fig. \ref{fig:DUNE_flux}. The spectral function for oxygen is from \cite{Benhar:1994hw,Benhar:2005dj} renormalized according to the units and conventions used here.
\begin{figure}
\centerline{\includegraphics[height=3in]{DUNE_probability.pdf}} \caption{(color online) DUNE flux converted to a probability density as a function of $k$ in GeV. }\label{fig:DUNE_flux}
\end{figure}
Figures \ref{fig:k_prime_1}, \ref{fig:k_prime_2} and \ref{fig:k_prime_3} show cross sections for $^2\mathrm{H}$ and $^{16}\mathrm{O}$ for $k'=1,2\ \mathrm{and}\ 3\,\mathrm{GeV}$ respectively, as a function of the polar angle of the detected proton $\theta_N^L$, for a variety of lepton scattering angles subject to the $y=0$ constraint (see Sect.~\ref{sec:deut}) and the restriction required by Eq.~(\ref{eq:theta_l_max}). For each scattering angle, the values of the incident neutrino energy $k$ and the momentum transfer $q$ are given for the deuteron. For oxygen these quantities cover a range of values due to their dependence on the invariant mass $W_{A-1}$ which is integrated over to the semi-inclusive cross section. For completeness, each figure contains the momentum of the detected proton $p_N$ as a function of $\theta_N^L$ with values given by the right-hand scale. Since for $^2\mathrm{H}_2{}^{16}\mathrm{O}$ there are two deuterium nuclei for each oxygen nucleus, the cross sections for deuterium are multiplied by a factor of 2.
\begin{figure}
\centerline{\includegraphics[height=2in]{sigma_1_5.pdf}\includegraphics[height=2in]{sigma_1_15.pdf}}
\centerline{\includegraphics[height=2in]{sigma_1_25.pdf}\includegraphics[height=2in]{sigma_1_35.pdf}}
\centerline{\includegraphics[height=2in]{sigma_1_45.pdf}\includegraphics[height=2in]{sigma_1_55.pdf}}
\centerline{\includegraphics[height=2in]{sigma_1_65.pdf}\includegraphics[height=2in]{sigma_1_75.pdf}}
\caption{(color online) Probability weighted cross sections for $k'=1$ GeV for various scattering angles $\theta_l$. The solid lines represent twice the deuteron cross section and the dashed lines are for the oxygen cross section versus $\theta^L_N$. The value of $p_N$ is represented by the dotted lines.}\label{fig:k_prime_1}
\end{figure}
\begin{figure}
\centerline{\includegraphics[height=2in]{sigma_2_5.pdf}\includegraphics[height=2in]{sigma_2_15.pdf}}
\centerline{\includegraphics[height=2in]{sigma_2_25.pdf}\includegraphics[height=2in]{sigma_2_35.pdf}}
\centerline{\includegraphics[height=2in]{sigma_2_45.pdf}}
\caption{(color online) As for Fig. \ref{fig:k_prime_1} but now for $k'=2$ GeV. }\label{fig:k_prime_2}
\end{figure}
\begin{figure}
\centerline{\includegraphics[height=2in]{sigma_3_5.pdf}\includegraphics[height=2in]{sigma_3_15.pdf}}
\centerline{\includegraphics[height=2in]{sigma_3_25.pdf}}
\caption{(color online) As for Fig. \ref{fig:k_prime_1} but now for $k'=3$ GeV. }\label{fig:k_prime_3}
\end{figure}
In all cases the maximum value of the oxygen cross section is at most one tenth of the deuterium cross section at its maximum value with the relative size decreasing for increased muon energy and scattering angle. It should be remembered, however, that these cross sections are evaluated and kinematics chosen to maximize the contribution of deuterium.
The size of the deuterium cross sections relative to those of oxygen may still seem rather startling. The explanation for this is straightforward. The semi-inclusive cross sections are roughly proportional to the neutron momentum distributions for the two nuclei as shown in Fig \ref{fig:momentum_dist}. Note that the maximum value of the deuterium momentum distribution is roughly five times as large as that for oxygen. Given that there are two deuterium nuclei for each oxygen nucleus, this difference in the peak values of the deuterium and oxygen momentum distributions explains the difference in the size of the cross section shown above.
\begin{figure}
\centerline{\includegraphics[height=3in]{momentum_distributions.pdf}}
\caption{(color online) Neutron momentum distributions for $^2$H (solid line) and $^{16}$O (dashed line).}\label{fig:momentum_dist}
\end{figure}
Figure \ref{fig:inclusive} shows the inclusive cross sections for deuterium and oxygen as a function of the incident neutrino momentum. This shows that integrating over all possible values of proton three-momentum results in a much larger and broader quasielastic peak for oxygen than for deuterium, as should be expected. This indicates that the unconstrained semi-inclusive cross section is distributed over a much larger region of phase space than that for deuterium.
\begin{figure}
\centerline{\includegraphics[height=3in]{inclusive.pdf}}
\caption{(color online) Inclusive CC$\nu$ cross sections for $^2$H (solid line) and $^{16}$O (dashed line).}\label{fig:inclusive}
\end{figure}
\section{Conclusions}\label{sec:concl}
The study presented in this paper of the semi-inclusive charge-changing neutrino reaction $(\nu_{\mu},\mu^- p)$ on a target of heavy water (D$_2$O) indicates that by careful choice of muon and proton three-momenta it is theoretically possible to separate deuterium events from those for oxygen. While naive considerations such as simply counting the number of neutrons provided by the two nuclei, namely, two for the two deuterium nuclei versus eight for the oxygen might lead one to expect that the latter will constitute a large background when the goal is to focus on events from the former, this proves not to be the case for events selected to favor the deuterium. As discussed in the previous section where results are given, this expectation is not necessarily the case: the spectral function for deuterium is sharply peaked at small values of the missing momentum, whereas that for oxygen peaks at larger missing momenta where contributions from the 1p-shell are dominant and at low missing momenta but at higher missing energies where the 1s-shell contributions occur. Furthermore, these contributions to the oxygen spectral function are spread much more widely in missing momentum than the corresponding sharply peaked ones for deuterium, roughly by the factor of four obtained by forming the ratio of the Fermi momenta for the two nuclei, namely 55 MeV/c for deuterium and 230 MeV/c for oxygen.
In passing we note that the high missing energy/missing momentum region, while contributing perhaps 20\% to the inclusive cross section, is essentially irrelevant for the semi-inclusive cross section as the strength there is very broadly distributed and little is picked up when only limited parts of the ``nuclear landscape'' of the spectral function for a nucleus like oxygen are involved.
In summary, from this theoretical study it appears that targets such as heavy water or deuterated methane containing significant amounts of deuterium together with light nuclei such as oxygen or carbon have the potential to provide unique information for studies of charge-changing neutrino reactions. Upon isolating the deuterium events using semi-inclusive reactions the kinematics alone will yield the incident neutrino energy on an event-by-event basis. Moreover, the cross section for such reactions on deuterium are arguably the best known throughout the periodic table even at quite high energies where relativistic modeling of the type used in the present work is undertaken. This being the case, such measurements hold the promise of determining the incident neutrino flux, thereby providing a very high-quality calibration of other existing or planned near detectors for neutrino oscillation experiments. Additionally, the fact that the nuclear structure issues are so well under control for the case of deuterium means that measurements of this type could serve in determining other aspects of the reaction, for instance, yielding new insights into the nature of the isovector axial-vector form factor of the nucleon. The issue now is an experimental one: can a practical target/detector of heavy water be realized? How are the protons in the final state to be detected? Can layers of (normal, un-deuterated) scintillator be used, as some have suggested, or are there other techniques to employ? Also: what is the optimal oscillation experiment using heavy water? While a near detector of heavy water appears worth contemplating, a far detector would be more challenging. Perhaps this last issue should be viewed in reverse, starting with the largest practical heavy water detector, then using the cross section to find how far from the neutrino source it could be placed, and then, finally, determining from the ``sweet spot'' for oscillation studies what beam energy is appropriate.
|
\section{Introduction}
Entropies play a fundamental role in quantum information theory as characterizations of the optimal rates of information theoretic tasks, and as measures of uncertainty. The mathematical properties of entropic functions therefore have important physical implications. The von Neumann entropy $S$, for instance, as a function of $d$-dimensional quantum states, is strictly concave, continuous, and is bounded by $\log d$. As the von Neumann entropy characterizes the optimal rate of data compression for a memoryless quantum information source \cite{Schumacher}, continuity of the von Neumann entropy, for example, implies that the quantum data compression limit is continuous in the source state. The $\alpha$-R\'enyi entropies $S_\alpha$ are parametrized by $\alpha \in (0,1)\cup (1,\infty)$, and are a generalization of the von Neumann entropy in the sense that $\lim_{\alpha\to 1} S_\alpha = S$. The $\alpha$-R\'enyi entropy has been used to bound the quantum communication complexity of distributed information-theoretic tasks \cite{vDH02}, can be interpreted in terms of the free energy of a quantum or classical system \cite{Baez11}, and is the fundamental quantity defining the entanglement $\alpha$-R\'enyi entropy \cite{WMVF16}.
In fact, the $\alpha$-R\'enyi entropies are members of a large family of entropies called the $(h,\phi)$-entropies, which are parametrized by two functions $h,\phi$ on $\mathbb{R}$ subject to certain constraints (see \Cref{sec:notation_and_def}). This family includes the Tsallis entropies \cite{Tsallis1988} and the unified entropies (considered by Rastegin in \cite{Ras2011}). Note that the $(h,\phi)$-entropy of a quantum state is the classical $(h,\phi)$-entropy of its eigenvalues, and therefore the results here apply equally well to probability distributions on finite sets.
Continuity is a useful property of entropic functions, particularly when cast in the form of a \emph{uniform continuity bound}: given two $d$-dimensional states which are at a trace distance of at most $\varepsilon \in (0,1)$, this provides a bound on their entropy difference entirely in terms of $\varepsilon$ and $d$. Fannes first proved a uniform continuity bound for the von Neumann entropy \cite{Fannes1973}. This bound was improved to a tight form by Audenaert \cite{Audenaert07} and is often called the called the Audenaert-Fannes bound (see also \cite[Theorem 3.8]{PetzQITbook}). Rastegin proved similar continuity bounds for the unified entropies, which include the $\alpha$-R\'enyi entropies and Tsallis entropies, but the resulting bounds are not known to be tight \cite{Ras2011}. Recently, Chen et al proved continuity bounds for the $\alpha$-R\'enyi entropy for $\alpha \in (0,1)\cup(1,\infty)$ using techniques similar to Audenaert's proof of the Audenaert-Fannes bound \cite{Renyi-CMNF}, but the resulting bounds are known to be not tight \cite{comm-CMNF}.
In \cite{HD17}, we considered {\em{local continuity bounds}}. Given a $d$-dimensional quantum state $\sigma$, a \emph{local} continuity bound of an entropic function $H$ at $\sigma$ is a bound on the entropy difference $|H(\omega) - H(\sigma)|$ for any $\omega$ in an $\varepsilon$-ball around $\sigma$, which depends not only on $\varepsilon$ and $d$ but also on the state $\sigma$ itself. These local bounds hence incorporate additional information about the state $\sigma$, for example, its spectrum, to yield a bound which is tighter than a uniform continuity bound. By finding maximizers and minimizers of the \emph{majorization order} on $d$-dimensional quantum states over the $\varepsilon$-ball around $\sigma$, local bounds were obtained for any $(h,\phi)$-entropy, in fact, for any Schur concave entropic function in \cite{HD17}.
Given a quantum state $\sigma$ and $\varepsilon\in(0,1]$, we denote the $\varepsilon$-ball in trace distance around $\sigma$ by $B_\varepsilon(\sigma)$ (defined by \cref{eq:eps-ball} below). For a given $\sigma$ and $\varepsilon$, there exist two quantum states $\sigma^*_\sigma,\sigma_{*,\varepsilon} \in B_\varepsilon(\sigma)$ such that for any $\omega\in B_\varepsilon(\sigma)$ centered at $\sigma$,
\[
\sigma^*_\varepsilon \prec \omega \prec \sigma_{*,\varepsilon}
\]
where $\prec$ denotes the majorization order (defined in \Cref{sec:notation_and_def}). In \cite{HD17}, this fact was proved by explicit construction of these states, using the notation $\rho^*_\varepsilon(\sigma)$ for $\sigma^*_\varepsilon$
and $\rho_{*,\varepsilon}(\sigma)$ for $\sigma_{*,\varepsilon}$. These states were also independently found by Horodecki, Oppenheim, and Sparaciari \cite{HO17approxmaj}, and considered in the context of thermal majorization \cite{Remco, MNW17_thermal}. In \cite{HD17} we also established that the minimal state $\rho^*_\varepsilon(\sigma)\equiv \sigma^*$ in the majorization order, satisfied a semigroup property: $\rho^*_{\varepsilon_1+\varepsilon_2}(\sigma) = \rho^*_{\varepsilon_1}( \rho^*_{\varepsilon_2}(\sigma))$. This property plays a key role in the proof of the main results of this paper.
In \Cref{sec:notation_and_def} we introduce the basic notation and definitions and in \Cref{sec:main_results} we state our main results. The proof strategy is described in \Cref{sec:proof-strat} and in \Cref{sec:Lambda-eps} the construction of the minimal state (in the majorization order), $\sigma_\varepsilon^*$, which we use in our proof, is formulated. \Cref{sec:proof-Delta-eps-Schur-convex} consists of a proof of the main technical result \Cref{thm:Delta_eps_Schur_convex} and employs certain lemmas which are proved in \Cref{sec:proof_lemmas}. In \Cref{sec:elem-prop-concave-fun}, we recall an elementary property of concave functions.
\section{Notation and definitions \label{sec:notation_and_def}}
Let $\mathcal{H}$ denote a finite-dimensional Hilbert space, with $\dim \H = d$, $\mathcal{B}(\mathcal{H})$ the set of (bounded) linear operators on $\mathcal{H}$, and $\mathcal{B}_\text{sa}(\mathcal{H})$ the set of self-adjoint linear operators on $\mathcal{H}$. A quantum state (or density matrix) is a positive semidefinite element of $\mathcal{B}(\mathcal{H})$ with trace one. Let $\mathcal{D}(\mathcal{H})$ be the set quantum states on $\mathcal{H}$. We denote the completely mixed state by $\tau := \frac{\mathds{1}}{d}$. A pure state is a rank-1 density matrix; we denote the set of pure states by $\D_\text{pure}(\mathcal{H})$. For two quantum states $\rho,\sigma \in \mathcal{D}(\mathcal{H})$, the \emph{trace distance} between them
is given by
\[
T(\rho,\sigma) = \frac{1}{2}\|\rho-\sigma\|_1.
\]
We define the $\varepsilon$-ball around $\sigma\in \mathcal{D}(\mathcal{H})$ as the set
\begin{equation}
B_\varepsilon(\sigma) = \{ \omega\in \mathcal{D}(\mathcal{H}): T(\omega,\sigma) \leq \varepsilon \}. \label{eq:eps-ball}
\end{equation}
For any $A\in \mathcal{B}_\text{sa}(\mathcal{H})$, let $\lambda_+ (A)$ and $\lambda_-(A)$ denote the maximum and minimum eigenvalue of $A$, respectively, and $k_+(A)$ and $k_-(A)$ denote their multiplicities. Let $\lambda_j(A)$ denote the $j$th largest eigenvalue, counting multiplicity; that is, the $j$th element of the ordering
\[
\lambda_1(A)\geq \lambda_2(A) \geq \dotsm \geq \lambda_d(A).
\]
We set $\vec \lambda (A) := (\lambda_i(A))_{i=1}^d \in\mathbb{R}^d$ and denote the set of eigenvalues of $A\in \mathcal{B}_\text{sa}(\mathcal{H})$ by $\spec A \subset \mathbb{R}$.
Given $x\in \mathbb{R}^d$, write $x^\downarrow = (x^\downarrow_j)_{j=1}^d$ for the permutation of $x$ such that $x^\downarrow_1 \geq x^\downarrow_2 \geq \dotsm \geq x^\downarrow_d$. For $x,y\in \mathbb{R}^d$, we say $x$ \emph{majorizes} $y$, written $x \succ y$, if
\begin{equation} \label{def:majorize}
\sum_{j=1}^k x^\downarrow_j \geq \sum_{j=1}^k y^\downarrow_j \quad \forall k=1,\dotsc,d-1, \quad \text{and}\quad \sum_{j=1}^d x^\downarrow_j = \sum_{j=1}^d y^\downarrow_j.
\end{equation}
Given two states $\rho,\sigma\in \mathcal{D}$, we say $\sigma$ majorizes $\rho$, written $\rho\prec \sigma$ if $\vec\lambda(\rho) \prec \vec\lambda(\sigma)$. We say that $\varphi: \mathcal{D} \to \mathbb{R}$ is \emph{Schur convex} if $\varphi(\rho)\leq \varphi(\sigma)$ for any $\rho,\sigma\in \mathcal{D}$ with $\rho \prec \sigma$. If $\varphi(\rho) < \varphi(\sigma)$ for any $\rho,\sigma\in \mathcal{D}$ such that $\rho \prec \sigma$, and $\rho$ is not unitarily equivalent to $\sigma$, then $\varphi$ is \emph{strictly Schur convex}. We say $\varphi$ is Schur concave (resp.~strictly Schur concave) if $(-\varphi)$ is Schur convex (resp.~strictly Schur convex).
Let $h: \mathbb{R}\to \mathbb{R}$ and $\phi: [0,1] \to \mathbb{R}$ with $\phi(0) = 0$ and $h(\phi(1)) = 0$, such that either $h$ is strictly increasing and $\phi$ strictly concave, or $h$ strictly decreasing and $\phi$ strictly convex. Then the $(h,\phi)$-entropy, $H_{(h,\phi)}$, is defined by
\begin{equation} \label{eq:def_h-phi-entropy}
H_{(h,\phi)}(\rho) := h(\tr[\phi(\rho)])
\end{equation}
where $\phi$ is defined on $\mathcal{D}(\mathcal{H})$ by functional calculus, i.e.~given the eigen-decomposition $\rho = \sum_i \lambda_i(\rho) \pi_i$, we have $\phi(\rho) = \sum_i \phi(\lambda_i(\rho)) \pi_i$. Every $(h,\phi)$-entropy is strictly Schur concave and unitarily invariant; moreover, if $h$ is concave, then $H_{(h,\phi)}$ is concave \cite{Bosyk2016}. Here, we are most interested in the following three examples of $(h,\phi)$ entropies:
\begin{itemize}
\item The von Neumann entropy
\[
S(\rho) = -\tr (\rho \log \rho).
\]
$S$ is the $(h,\phi)$ entropy with $h=\id$, i.e., $h(x) = x$ for $x\in \mathbb{R}$, and with $\phi(x) = - x \log x$ for $x\in [0,1]$. The von Neumann entropy satisfies the following tight continuity bound known as the Audenaert-Fannes bound \cite{Audenaert07} (see also \cite[Theorem 3.8]{PetzQITbook}). Given $\varepsilon\in(0,1]$ and $\rho,\sigma\in \mathcal{D}(\mathcal{H})$ with $T(\rho,\sigma)\leq \varepsilon$,
\begin{equation}
|S(\rho) - S(\sigma) | \leq \begin{cases}
\epsilon \log (d-1) + h(\epsilon) & \text{if } \epsilon < 1 - \tfrac{1}{d} \\
\log d & \text{if } \epsilon \geq 1 - \tfrac{1}{d}
\end{cases} \label{eq:Audenaert-Fannes_bound}
\end{equation}
where $h(\varepsilon) := - \varepsilon \log \varepsilon - (1-\varepsilon) \log (1-\varepsilon)$ denotes the binary entropy.
\item The $q$-Tsallis entropy for $q\in(0,1)\cup(1,\infty)$,
\[
T_q(\rho) = \frac{1}{1-q}[\tr(\rho^q)-1].
\]
$T_q$ can be written as the $(h,\phi)$-entropy with $h(x) = x - \frac{1}{1-q}$ for $x\in \mathbb{R}$ and $\phi(x) = \frac{1}{1-q}x^q$. With these choices, $h$ is strictly increasing and affine (and therefore concave) and $\phi$ is strictly concave, for all $q\in (0,1)\cup(1,\infty)$.
\item The $\alpha$-R\'enyi entropy for $\alpha \in (0,1)\cup(1,\infty)$,
\[
S_\alpha(\rho) = \frac{1}{1-\alpha} \log \left(\tr \rho^\alpha\right).
\]
$S_\alpha$ is the $(h,\phi)$-entropy with $h(x) = \frac{1}{1-\alpha}\log x$ for $x\in \mathbb{R}$ and $\phi(x) = x^\alpha$ for $x\in [0,1]$. For $\alpha\in (0,1)$, $h$ is concave and strictly increasing and $\phi$ is strictly concave. For $\alpha >1$, $h$ is convex and strictly decreasing, and $\phi$ is strictly convex. It is known that $\lim_{\alpha \to 1}S_\alpha(\rho) = S(\rho)$.
\end{itemize}
In the above, all logarithms are taken to base $2$.
\section{Main results \label{sec:main_results}}
\begin{theorem}[Uniform continuity bounds] Let $H_{(h,\phi)}$ be an $(h,\phi)$-entropy, defined through \eqref{eq:def_h-phi-entropy}) with $h$ concave and $\phi$ strictly concave.
For $\varepsilon \in (0,1]$ and any states $\rho,\sigma\in \mathcal{D}(\mathcal{H})$ such that $\frac{1}{2}\|\rho-\sigma\|_1\leq \varepsilon$, we have
\begin{equation}
| H_{(h,\phi)}(\rho) - H_{(h,\phi)}(\sigma) | \leq \begin{cases}
h( \phi(1-\varepsilon) + (d-1) \phi( \frac{\varepsilon}{d-1})) & \varepsilon < 1-\frac{1}{d}\\
h(d\phi(\frac{1}{d})) & \varepsilon \geq 1 - \frac{1}{d}
\end{cases} \label{eq:hphi-uniform-bound}
\end{equation}
and in particular, for $\alpha \in (0,1)$,
\begin{equation}
| S_\alpha(\rho) - S_\alpha(\sigma) | \leq \begin{cases}
\frac{1}{1-\alpha} \log ( (1-\varepsilon)^\alpha + (d-1)^{1-\alpha} \varepsilon^\alpha) & \varepsilon < 1-\frac{1}{d}\\
\log d& \varepsilon \geq 1 - \frac{1}{d}
\end{cases} \label{eq:uniform_Renyi_bound}
\end{equation}
and for $q \in (0,1)\cup(1,\infty)$,
\begin{equation}
| T_q(\rho) - T_q(\sigma) | \leq \begin{cases}
\frac{1}{1-q} ( (1-\varepsilon)^q + (d-1)^{1-q} \varepsilon^q - 1) & \varepsilon < 1-\frac{1}{d}\\
\frac{ d^{1-q}-1}{1-q}& \varepsilon \geq 1 - \frac{1}{d},
\end{cases} \label{eq:uniform_Tsallis_bound}
\end{equation}
where $d =\dim \mathcal{H}$. Moreover, equality in \eqref{eq:hphi-uniform-bound}, \eqref{eq:uniform_Renyi_bound}, or \eqref{eq:uniform_Tsallis_bound} occurs if and only if one of the two states (say, $\sigma$) is pure, and either
\begin{enumerate}
\item $\varepsilon < 1 - \frac{1}{d}$ and $\vec \lambda(\rho) = ( 1- \varepsilon, \frac{\varepsilon}{d-1},\dotsc, \dotsc, \frac{\varepsilon}{d-1})$, or
\item $\varepsilon \geq 1- \frac{1}{d}$, and $\rho = \tau := \frac{\mathds{1}}{d}$.
\end{enumerate}\label{thm:hphi-GCB}
\end{theorem}
\begin{remark}
~\begin{itemize}
\item When \eqref{eq:hphi-uniform-bound} is applied to the von Neumann entropy $S$, one recovers the Audenaert-Fannes bound, \eqref{eq:Audenaert-Fannes_bound}, with equality conditions. The sufficiency of these equality conditions were shown in \cite{Audenaert07}, and their necessity was recently derived in \cite{HD17} by an analysis of the proof of the bound presented in \cite[Thm. 3.8]{PetzQITbook} and \cite{Winter16}, which involves a coupling argument. We establish that these necessary and sufficient conditions are the same for every $(h,\phi)$-entropy satisfying the conditions of the theorem.
\item The inequality \eqref{eq:uniform_Renyi_bound} reduces to the Audenaert-Fannes bound \eqref{eq:Audenaert-Fannes_bound} when the limit $\alpha \to 1$ is taken on both sides of it.
\item The bound \eqref{eq:uniform_Tsallis_bound} appeared in \cite{Renyi-CMNF} as Lemma 1.2, and was derived with a different method. However, the equality conditions were not established.
\item See \Cref{fig:bound_comparison} for a comparison of our uniform continuity bound for the $\alpha$-R\'enyi entropy, \eqref{eq:uniform_Renyi_bound}, for $\alpha = \frac{1}{2}$, with those obtained in \cite{Ras2011} and \cite{Renyi-CMNF}.
\end{itemize}
\end{remark}
\begin{figure}[ht]
\centering
\includegraphics[width = .75\textwidth]{plots/boundsplot.tikz}
\caption{In dimensions $d=5$, for $\alpha = \frac{1}{2}$, the bound given by the right-hand side of \eqref{eq:uniform_Renyi_bound} is compared to the bounds given by Equation (7) of \cite{Renyi-CMNF} and by Equation (27) of \cite{Ras2011}. We also include the trivial bound $\log d$, as well as 500 points corresponding the local bounds found in \cite{HD17} computed at (uniformly) randomly chosen $\sigma\in \mathcal{D}(\mathcal{H})$.} \label{fig:bound_comparison}
\end{figure}
\section{Proof strategy} \label{sec:proof-strat}
Given a state $\sigma\in \mathcal{D}(\mathcal{H})$ and $\varepsilon\in (0,1]$, one can construct two states $\sigma_\varepsilon^*,\sigma_{*,\varepsilon}\in B_\varepsilon(\sigma)$ such that
\begin{equation} \label{eq:maj-order}
\sigma_\varepsilon^* \prec \omega \prec \sigma_{*,\varepsilon}
\end{equation}
for any $\omega\in B_\varepsilon(\sigma)$. This was done in \cite{HD17}, with the notation $\rho_\varepsilon^*(\sigma)$ (resp.~$\rho_{*,\varepsilon}(\sigma)$) to denote $\sigma_\varepsilon^*$ (resp.~$\sigma_{*,\varepsilon}$). These states were also independently found in \cite{HO17approxmaj}, and considered in the context of thermal majorization in \cite{Remco,MNW17_thermal}. The proof of our main result relies on the form of $\sigma_\varepsilon^*$ and its properties. An explicit construction of $\sigma_\varepsilon^*$ is given in \Cref{sec:Lambda-eps}, and its properties are described in \Cref{prop:properties_of_Lambda_eps}.
Consider an $(h,\phi)$ entropy $H_{(h,\phi)}$, and let $\epsilon\in (0,1]$, and $\rho,\sigma\in \mathcal{D}(\mathcal{H})$ with $T(\rho,\sigma)\leq \epsilon$. If $H_{(h,\phi)}(\rho) \geq H_{(h,\phi)}(\sigma)$, then since $\rho\in B_\varepsilon(\sigma)$,
\begin{equation}
|H(\rho) - H(\sigma)| = H(\rho) - H(\sigma) \leq \max_{\omega\in B_\varepsilon(\sigma)} H(\omega) - H(\sigma) = H(\sigma_\varepsilon^*) - H(\sigma) \label{eq:Hrho-Hsig}
\end{equation}
where the last equality follows from the first majorization relation in \cref{eq:maj-order} and the strict Schur concavity of $H_{(h,\phi)}$. Similarly, if $H_{(h,\phi)}(\sigma) \geq H_{(h,\phi)}(\rho)$, \cref{eq:Hrho-Hsig} holds with $\sigma$ (resp.~$\sigma_\varepsilon^*$) replaced by $\rho$ (resp.~$\rho_\varepsilon^*$). Hence, in general,
\begin{equation} \label{eq:UBHrho-Hsig_by_max}
|H_{(h,\phi)}(\rho) - H_{(h,\phi)}(\sigma)| \leq \max \{\Delta_\varepsilon(\rho),\Delta_\varepsilon(\sigma)\} \leq \max_{\omega \in \mathcal{D}(\mathcal{H})} \Delta_\varepsilon(\omega),
\end{equation}
where
\begin{equation} \label{eq:def_Delta-eps}
\begin{aligned}
\Delta_\varepsilon : \qquad \mathcal{D}(\mathcal{H}) &\to \mathbb{R}_{\geq 0}\\
\omega &\mapsto H_{(h,\phi)}\circ \mmm_\varepsilon(\omega) - H_{(h,\phi)}(\omega),
\end{aligned}
\end{equation}
and $\mmm_\varepsilon$ is the \emph{majorization-minimizer map},
\begin{equation} \label{eq:def_mmm-eps}
\begin{aligned}
\mmm_\varepsilon : \qquad \mathcal{D}(\mathcal{H}) &\to \mathcal{D}(\mathcal{H})\\
\omega &\mapsto \omega_\varepsilon^*.
\end{aligned}
\end{equation}
This map is defined explicitly by \cref{def:Lambdaeps} in \Cref{sec:Lambda-eps}.
Note that $\Delta_\varepsilon(\omega)\geq 0$ for $\omega\in \mathcal{D}(\mathcal{H})$ follows from the Schur concavity of the $(h,\phi)$-entropy. To prove \Cref{thm:hphi-GCB}, it remains to maximize $\Delta_\varepsilon$ over $\mathcal{D}(\mathcal{H})$.
We show that for $(h,\phi)$-entropies for which $h$ is concave and $\phi$ (strictly) convex, $\Delta_\varepsilon$ is a Schur convex function on $\mathcal{D}(\mathcal{H})$, which is our main technical result. We u defer its proof to \Cref{sec:proof-Delta-eps-Schur-convex}.
\begin{theorem} \label{thm:Delta_eps_Schur_convex}
Assume $h$ is concave and $\phi$ is strictly concave. Let $\varepsilon\in (0,1]$. Then $\Delta_\varepsilon: \mathcal{D}(\mathcal{H})\to \mathbb{R}_{\geq 0}$ is Schur convex. That is, if $\rho\prec \sigma$,
\[
\Delta_\varepsilon(\rho) \leq \Delta_\varepsilon(\sigma).
\]
Moreover, $\Delta_\varepsilon(\rho) = \Delta_\varepsilon(\sigma)
$ implies $\lambda_+(\rho) = \lambda_+(\sigma)$. Lastly, if $h$ is strictly concave, then $\Delta_\varepsilon$ is strictly Schur convex.
\end{theorem}
Note that if $h$ is not strictly concave, $\Delta_\varepsilon$ need not be strictly Schur convex. In fact, for the von Neumann entropy we can find a counterexample to strict Schur convexity of $\Delta_\varepsilon$. Setting $\rho = \diag(0.1,0.2,0.2,0.5)$ and $\sigma = \diag(0.1, 0.15, 0.25, 0.5)$ yields $\rho \prec \sigma$ and that $\rho$ and $\sigma$ are not unitarily equivalent. However, for $\varepsilon \leq 0.05$, we have $\Delta_\varepsilon(\rho) = \Delta_\varepsilon(\sigma)$.
\begin{corollary}If $h$ is concave, $\phi$ strictly concave, and $\varepsilon\in (0,1]$, then $\Delta_\varepsilon$ achieves a maximum on $\mathcal{D}(\mathcal{H})$, and moreover $\argmax \Delta_\varepsilon = \D_\text{pure}(\mathcal{H}) $. \label{cor:Delta-eps-max-at-pure}
\end{corollary}
\begin{proof}
Since any pure state $\psi$ satisfies $\rho\prec \psi$ for every $\rho\in \mathcal{D}(\mathcal{H})$, we have $\Delta_\varepsilon(\psi) \geq \Delta_\varepsilon(\rho)$ for every $\rho\in \mathcal{D}(\mathcal{H})$. Therefore, $\D_\text{pure}(\mathcal{H}) \subset \argmax \Delta_\varepsilon$. On the other hand, if $\omega\in \mathcal{D}(\mathcal{H})$ has $\omega \in \argmax \Delta_\varepsilon$, then
\[
\Delta_\varepsilon(\omega) = \Delta_\varepsilon(\psi)
\]
for a pure state $\psi$. Therefore, $\lambda_+(\omega) = \lambda_+(\psi)=1$, and $\omega$ must be a pure state.
\end{proof}
\begin{figure}[ht]
\centering
\includegraphics[width=.408\textwidth]{plots/Delta__alpha=0p5__epsilon=0p1__res=1000_nobar.png}
\includegraphics[width=.492\textwidth]{plots/Delta__alpha=1p0__epsilon=0p1__res=1000.png}
\includegraphics[width=.408\textwidth]{plots/Delta__alpha=1p5__epsilon=0p1__res=1000_nobar.png}
\includegraphics[width=.492\textwidth]{plots/Delta__alpha=2p0__epsilon=0p1__res=1000.png}
\caption{In dimensions $d=3$, we parametrize $\sigma = \diag(x,y,1-x-y)$, and plot $(x,y) \mapsto \Delta_\varepsilon(\sigma)$ for $\varepsilon = 0.1$, with $H_{(h,\phi)} = S_\alpha$, the R\'enyi entropy. That is, above each $(x,y)$ in the $xy$-plane, the value of $\Delta_\varepsilon(\diag(x,y,1-x-y))$ is plotted. Top row, left: $\alpha = \frac{1}{2}$, right: $\alpha = 1$. Bottom row, left: $\alpha = 1.5$, right: $\alpha = 2$. The three points $(0,0)$, $(0,1)$, $(1,0)$ in the $xy$-plane correspond to the pure states $\diag(0,0,1)$, $\diag(0,1,0)$, and $\diag(1,0,0)$, respectively. The central point $(\frac{1}{3},\frac{1}{3})$ corresponds to the completely mixed state $\tau = \frac{1}{3}\mathds{1}$. We observe for $\alpha =\frac{1}{2}$ and $\alpha=1$ the maximum of $\Delta_\varepsilon$ appears to occur at the pure states. On the other hand, for $\alpha=1.5$, the maximum is along the boundary (i.e.~for a state $\sigma$ with exactly one zero eigenvalue), and for $\alpha = 2$, the maximum occurs at states without any zero eigenvalues.} \label{fig:Delta_eps_for_Renyi}
\end{figure}
Using these results, the proof of \Cref{thm:hphi-GCB} is completed as follows. Let $\psi$ be any pure state, $\psi\in \D_\text{pure}(\mathcal{H})$. Then for any $\omega\in \mathcal{D}(\mathcal{H})$, we have $\omega \prec \psi$. Therefore, by \Cref{thm:Delta_eps_Schur_convex}, we have $\Delta_\varepsilon(\omega)\leq \Delta_\varepsilon(\psi)$, for any $\omega\in \mathcal{D}(\mathcal{H})$, and in particular for $\omega\in \{\rho,\sigma\}$. Therefore, by \eqref{eq:UBHrho-Hsig_by_max} we have
\[
|H_{(h,\phi)} (\rho) - H_{(h,\phi)}(\sigma)| \leq \Delta_\varepsilon(\psi).
\]
By computing $\Delta_\varepsilon(\psi)$ using the form given in \Cref{prop:properties_of_Lambda_eps}\ref{item:Lambda-eps-on-psi}, we obtain the right-hand side of \cref{eq:hphi-uniform-bound}.
It remains to check under which conditions equality occurs in \eqref{eq:hphi-uniform-bound}. Assume without loss of generality that $H_{(h,\phi)}(\rho) \geq H_{(h,\phi)}(\sigma)$. Equality in \eqref{eq:UBHrho-Hsig_by_max} is equivalent to $\sigma\in \D_\text{pure}(\mathcal{H})$ by \Cref{cor:Delta-eps-max-at-pure}. Next, since the $(h,\phi)$-entropy is strictly Schur concave and $\sigma_\varepsilon^* \prec \rho$, equality in \eqref{eq:Hrho-Hsig} is equivalent to the fact that $\rho$ is unitarily equivalent to $\sigma_\varepsilon^*$. The expression for $\sigma_\varepsilon^*$ when $\sigma\in \D_\text{pure}(\mathcal{H})$ is given in \Cref{prop:properties_of_Lambda_eps}\ref{item:Lambda-eps-on-psi}. This completes the proof.\hfill\proofSymbol
\Cref{thm:Delta_eps_Schur_convex} does not extend to the $\alpha$-R\'enyi entropy for $\alpha > 1$, in which case $h$ is convex and $\phi$ strictly convex. This is discussed in the remark following \Cref{lem:reduce-to-id}, and is illustrated in \Cref{fig:Delta_eps_for_Renyi}.
\section{The majorization-minimizer map $\mmm_\varepsilon$} \label{sec:Lambda-eps}
In order to prove \Cref{thm:hphi-GCB}, we need to use properties of the majorization-minimizer map $\mmm_\varepsilon$ introduced in \eqref{eq:def_mmm-eps}. Let $\sigma\in \mathcal{D}(\mathcal{H})$ and $\varepsilon\in (0,1]$. We formulate the definition of $\mmm_\varepsilon$ by constructing $\sigma_\varepsilon^*$. Note that the following is a reformulation of Lemma 4.1 of \cite{HD17}. For notational simplicity, we often suppress dependence on $\sigma$ and $\varepsilon$ in this section, and write $\lambda_j = \lambda_j(\sigma)$ so that the eigenvalues of $\sigma$ are $\lambda_1 \geq \lambda_2 \geq \dotsm \geq \lambda_d$.
We first define a quantity $\gamma_+^{(m)}\equiv \gamma_+^{(m)}(\sigma,\varepsilon)$, for $m\in \{0,1,\dotsc,d-1\}$, as follows
\[
\gamma_+^{(m)} := \begin{cases}
\frac{1}{m}\left(\sum_{i=1}^{m} \lambda_i - \varepsilon\right) & \text{if }T(\sigma,\tau) > \varepsilon \text{ and } m\neq 0\\
\frac{1}{d} & \text{else}.
\end{cases}
\]
Similarly, a quantity $\gamma_-^{(m)} \equiv \gamma_-^{(m)}(\sigma,\varepsilon) $ is defined by
\[
\gamma_-^{(m)}:= \begin{cases}\frac{1}{m}\left(\sum_{i=d-m+1}^d \lambda_i + \varepsilon\right) & \text{if } T(\sigma,\tau) > \varepsilon \text{ and } m\neq 0\\
\frac{1}{d} & \text{else}.
\end{cases}
\]
Then for $\sigma \neq \tau$, we define $m_+ = m_+(\sigma,\varepsilon)$ as the unique solution to the following inequalities:
\begin{equation}\label{eq:m_is_sol_to_this}
\lambda_{m+1} \leq \gamma_+^{(m)} < \lambda_m, \qquad m\in \{1,\dotsc,d-1\}
\end{equation}
and we set $m_+(\tau,\varepsilon) = 0$.
Similarly, for $\sigma\neq \tau$, we define $m_- = m_-(\sigma,\varepsilon)$ as the unique solution to the inequalities:
\begin{equation} \label{eq:n_is_sol_to_this}
\lambda_{d-m+1}< \gamma_-^{(m)}\leq \lambda_{d-m}, \qquad m\in \{1,\dotsc,d-1\}
\end{equation}
and set $m_-(\tau,\varepsilon) = 0$. Finally, we set $\gamma_+ = \gamma_+(\sigma,\varepsilon) := \gamma_+^{(m_+)}$ and $\gamma_-=\gamma_-(\sigma,\varepsilon) := \gamma_-^{(m_-)}$.
Given the eigen-decomposition $\sigma = \sum_{i=1}^d \lambda_i \ketbra{i}{i}$, we define
\begin{equation} \label{def:Lambdaeps}
\mmm_\varepsilon(\sigma) := \sigma_\varepsilon^* := \sum_{i=1}^{m_+} \gamma_+\ketbra{i}{i} + \sum_{i=m_++1}^{d-m_-} \lambda_i\ketbra{i}{i} + \sum_{i=d-m_-+1}^d \gamma_-\ketbra{i}{i}.
\end{equation}
To summarize, we construct $\sigma_\varepsilon^*$ as follows: we decrease the $m_+$ largest eigenvalues of $\sigma$ by setting them to $\gamma_+$ (where $m_+$ and $\gamma_+$ are related by \cref{eq:m_is_sol_to_this}), increase the $m_-$ smallest eigenvalues of $\sigma$ by setting them to $\gamma_-$ (where $m_-$ and $\gamma_-$ are related by \cref{eq:n_is_sol_to_this}), and we keep the other eigenvalues of $\sigma$ unchanged. This is illustrated in \Cref{fig:sig-levels}, for a state $\sigma\in \mathcal{D}(\mathcal{H})$ with $\varepsilon = 0.07$ and $d=12$.
\begin{figure}[ht]
\centering
\includegraphics{plots/levels_fig.pdf}
\caption{We choose $d=12$, a state $\sigma \in \mathcal{D}(\mathcal{H})$, and $\varepsilon=0.07$, for which $m_+=2$ and $m_-=4$. Left: the eigenvalues $\lambda_1 \geq \lambda_2\geq \dotsc \geq \lambda_d$ of $\sigma$ are plotted. Center: the smallest four eigenvalues of $\sigma$ are increased to $\gamma_- = \frac{1}{4}[\lambda_1 + \dotsm + \lambda_4+ \varepsilon]$, and the largest two eigenvalues of $\sigma$ decreased to $\gamma_+ = \frac{1}{2}[\lambda_1+\lambda_2 - \varepsilon]$. Right: the eigenvalues of $\sigma_\varepsilon^*$ are $\gamma_+$ with multiplicity two, $\lambda_3,\lambda_4,\dotsc,\lambda_{d-4}$, and $\gamma_-$ with multiplicity four. \label{fig:sig-levels}}
\end{figure}
Considered as a map on $\mathcal{D}(\mathcal{H})$, $\mmm_\varepsilon$ has several useful properties which are presented in the following proposition. It should be noted, however, that $\mmm_\varepsilon$ is \emph{not} a linear map.
\begin{proposition}[Properties of $\mmm_\varepsilon$] \label{prop:properties_of_Lambda_eps} Let $\sigma\in \mathcal{D}(\mathcal{H})$. We have the following properties of $\mmm_\varepsilon$, for any $\varepsilon\in (0,1]$.
\begin{enumerate}[label*=\alph*.,ref=(\alph*)]
\item \label{item:Lambda_epsstates-to-states} Maps states to states: $\mmm_\varepsilon : \mathcal{D}(\mathcal{H}) \to \mathcal{D}(\mathcal{H})$.
\item \label{item:Lambda_eps_min_maj_order} Minimal in majorization order: $\mmm_\varepsilon(\sigma)\in B_\varepsilon(\sigma)$ and for any $\omega\in B_\varepsilon(\sigma)$, we have $\mmm_\varepsilon(\sigma) \prec \omega$.
\item \label{item:Lambda_eps-semigroup} Semi-group property:
if $\varepsilon_1,\varepsilon_2\in (0,1]$ with $\varepsilon_1+\varepsilon_2 \leq 1$, we have $\mmm_{\varepsilon_1+\varepsilon_2}(\sigma) = \mmm_{\varepsilon_1} \circ \mmm_{\varepsilon_2}(\sigma)$.
\item \label{item:Lambda_eps-maj-preserving} Majorization-preserving: let $\rho\in \mathcal{D}(\mathcal{H})$ such that $\rho \prec \sigma$. Then $\mmm_\varepsilon(\rho) \prec \mmm_\varepsilon(\sigma)$.
\item \label{item:Lambda_eps_fixed_point} $\tau = \frac{\mathds{1}}{d}$ is the unique fixed point of $\mmm_\varepsilon$, i.e.~the unique solution to $\sigma = \mmm_\varepsilon(\sigma)$ for $\sigma \in \mathcal{D}(\mathcal{H})$. Moreover, for any $\sigma \neq \tau$, $\mmm_\varepsilon(\sigma)$ is not unitarily equivalent to $\sigma$.
\item \label{item:Lambda-eps-near-tau}For any state $\sigma \in B_\varepsilon(\tau)$, we have $\mmm_\varepsilon(\sigma) = \tau$.
\item \label{item:Lambda-eps-on-psi}For any pure state $\psi\in \D_\text{pure}(\mathcal{H})$, the state $\mmm_\varepsilon(\psi)$ has the form
\begin{equation}
\mmm_\varepsilon(\psi) = \begin{cases}
\diag(1- \varepsilon, \frac{\varepsilon}{d-1},\dotsc \frac{\varepsilon}{d-1}) & \varepsilon < 1 - \frac{1}{d}\\
\tau := \frac{\mathds{1}}{d} & \varepsilon \geq 1 - \frac{1}{d}.
\end{cases} \label{eq:Lambda-eps-pure}
\end{equation}
\end{enumerate}
\end{proposition}
The proof of properties \ref{item:Lambda_epsstates-to-states} and \ref{item:Lambda_eps_min_maj_order} can be found in \cite{HD17,HO17approxmaj}; the property \ref{item:Lambda_eps-semigroup} was proved in in \cite{HD17}, property \ref{item:Lambda_eps-maj-preserving} can be found in Lemma 2 of \cite{HO17approxmaj}. The property \ref{item:Lambda_eps_fixed_point} can be shown as follows. $\mmm_\varepsilon(\rho)$ is not unitarily equivalent to $\rho$ for $\rho\neq\tau$ follows from the construction presented above, in particular, the fact that the eigenvalues of $\mmm_\varepsilon(\rho)$ differ from $\rho$. One immediately has that $\tau$ is a fixed point of $\mmm_\varepsilon$, and uniqueness follows from the fact that $\mmm_\varepsilon(\sigma)$ is not unitarily equivalent to $\sigma$ for $\sigma\neq \tau$. Lastly, the properties \ref{item:Lambda-eps-near-tau} and \ref{item:Lambda-eps-on-psi} follow from the construction given above.
\section{Proof of \Cref{thm:Delta_eps_Schur_convex} \label{sec:proof-Delta-eps-Schur-convex}}
\subsection{Reducing to $h=\id$}
Our first task is to reduce to the case when $h=\id$, i.e. $h(x) = x$ for all $x\in \mathbb{R}$. Fix $\varepsilon \in (0,1]$ and $\rho,\sigma \in \mathcal{D}(\mathcal{H})$ such that $\rho\prec \sigma$ and $\rho$ and $\sigma$ are not unitarily equivalent. Let us define four variables
\begin{gather*}
a := H_{(\id,\phi)} \circ \mmm_{ \varepsilon} (\rho) ,\qquad b := H_{(\id,\phi)}(\rho), \qquad
c := H_{(\id,\phi)} \circ \mmm_{\varepsilon} (\sigma) ,\qquad d := H_{(\id,\phi)}(\sigma)
\end{gather*}
which are non-negative real numbers.
\Cref{thm:Delta_eps_Schur_convex} is the statement that
\begin{equation} \label{eq:h-abcd}
h(a) - h(b) \leq h(c) - h(d).
\end{equation}
\begin{lemma}Let $h$ be concave, and $\phi$ strictly concave.
If $a-b \leq c-d$, then \eqref{eq:h-abcd} holds. Moreover, if $h$ is strictly concave, then \eqref{eq:h-abcd} holds with strict inequality.
\label{lem:reduce-to-id}
\end{lemma}
\begin{proof}
By the strict Schur concavity of the $(\id,\phi)$-entropy, we have $b<a$ and $d<c$, and by \Cref{prop:properties_of_Lambda_eps} \ref{item:Lambda_eps_min_maj_order}, we have $b> d$ and $a> c$. Therefore, since $h$ is concave, we apply \Cref{cor:main_concave_ieq} to obtain
\[
\frac{h(b) - h(d)}{b-d} \geq \frac{h(a) - h(c)}{a-c}.
\]
That is,
\[
[h(b) - h(d)] \frac{a-c}{b-d} \geq h(a) -h(c)
\]
Since we have $a-c \leq b - d$ using the assumption, then $\frac{a-c}{b-d} \leq 1$, and therefore
\[
h(b) - h(d) \geq h(a) -h(c)
\]
and adding $h(c) - h(b)$ to each side yields \eqref{eq:h-abcd}.
\end{proof}
Therefore, it remains to establish $a-b\leq c-d$, which is \Cref{thm:Delta_eps_Schur_convex} when $h=\id$.
\begin{remark}An extension of \Cref{thm:hphi-GCB} to treat the $\alpha$-R\'enyi entropy for $\alpha > 1$ would need to address the case in which $h$ is convex and strictly decreasing, and $\phi$ is strictly convex. In this case, $\rho \mapsto \tr \phi(\rho)$ is Schur convex, and we have $a<b$, $c<d$, $b<d$, and $a<c$. The analog to \Cref{lem:reduce-to-id} would be to show that $a-b \geq c-d$ implies \eqref{eq:h-abcd}. However, repeating the proof of \Cref{lem:reduce-to-id} in this case yields e.g.
\[
[h(b)- h(d)] \frac{c-a}{d-b} \leq h(a)-h(c)
\]
which is inconclusive in showing \eqref{eq:h-abcd} when $a-b \geq c-d$. This is the technical reason this proof does not extend to the $\alpha$-R\'enyi entropy for $\alpha > 1$.
In fact, the associated quantity $\Delta_\varepsilon$ for an $\alpha$-R\'enyi entropy with $\alpha>1$ is not Schur convex. For the example stated after \Cref{thm:Delta_eps_Schur_convex}, it can be shown that choosing $H_{(h,\phi)} = S_\alpha$ for $\alpha > 1$ yields $\Delta_\varepsilon(\rho) > \Delta_\varepsilon(\sigma)$.
\end{remark}
\subsection{The case $h=\id$}
We prove \Cref{thm:Delta_eps_Schur_convex} in several steps. First, we use the semigroup property of $\mmm_\varepsilon$ to decompose $\Delta_{\varepsilon}$ for $\varepsilon=\varepsilon_1+\varepsilon_2$ in terms of $\varepsilon_1$ and $\varepsilon_2$ in \Cref{lem:Delta_eps1_plus_eps2}. Then we define a quantity $\delta(\rho,\sigma)$ in \Cref{def:delta-rho-sigma} such that for $\varepsilon \leq \delta(\rho,\sigma)$, we can show that $\Delta_\varepsilon(\rho) \leq \Delta_\varepsilon(\sigma)$ if $\rho\prec \sigma$ (\Cref{lem:Delta_eps_Schur_convex_deltarho-sig}), using properties of $\delta(\rho,\sigma)$ presented in \Cref{lem:delta_rho}. Finally, we show that for arbitrary $\varepsilon\in (0,1]$, we can use \Cref{lem:Delta_eps1_plus_eps2} finitely many times to prove \Cref{thm:Delta_eps_Schur_convex}. We state the lemmas here but defer their proofs to \Cref{sec:proof_lemmas}.
\begin{lemma} \label{lem:Delta_eps1_plus_eps2}
Let $\rho\in \mathcal{D}(\mathcal{H})$, and $\varepsilon_1,\varepsilon_2\in (0,1]$ with $\varepsilon_1+\varepsilon_2 \leq 1$. Then
\[
\Delta_{\varepsilon_1+\varepsilon_2}(\rho) = \Delta_{\varepsilon_1} \circ \mmm_{\varepsilon_2}(\rho) + \Delta_{\varepsilon_2}(\rho).
\]
\end{lemma}
\begin{definition}[$\delta(\rho,\sigma)$] \label{def:delta-rho-sigma}Let $\rho\in \mathcal{D}(\mathcal{H})$ for $\rho \neq \tau$.
Let $\mu_1 > \mu_2 > \dotsm > \mu_\ell$ denote the distinct ordered eigenvalues of $\rho$, and define
\begin{equation}
\delta(\rho) =\min \{k_+(\rho) (\mu_1-\mu_2), k_-(\rho)(\mu_{\ell-1}-\mu_\ell) \}. \label{eq:delta_rho}
\end{equation}
For $\rho,\sigma\in\mathcal{D}(\mathcal{H})$ with $\rho\neq\tau\neq\sigma$, define
\begin{equation}
\delta(\rho,\sigma) = \min \{ \delta(\rho), \delta(\sigma) \}.
\label{eq:def-delta-rho-sigma}
\end{equation}
\end{definition}
For any $\varepsilon \leq \delta(\rho,\sigma)$, the map $\mmm_\varepsilon$ only ``moves'' the largest and smallest eigenvalue of $\rho$ and of $\sigma$, as shown by the following result and illustrated through an example in \Cref{fig:Lambda-eps}.
\begin{figure}[ht]
\centering
\includegraphics[width=.75\textwidth]{plots/M-eps.pdf}
\caption{For the 5-dimensional state $\sigma = \diag(0.32, 0.26, 0.19, 0.13, 0.10)$, the spectrum of $\sigma_\varepsilon^* = \mmm_\varepsilon(\sigma)$ is plotted as a function of $\varepsilon$. This plot is a continuous (in $\varepsilon$) analog to the type of plot shown in \Cref{fig:sig-levels}, which shows the spectrum of $\sigma_\varepsilon^*$ at two discrete points, $\varepsilon = 0$ and $\varepsilon= 0.07$, in a different example.
Here, at $\varepsilon=0$, the five lines correspond to the five eigenvalues of $\sigma$, each with multiplicity one. For $\varepsilon \leq 0.03$, $\sigma_\varepsilon^* = \diag(0.32 -\varepsilon, 0.26, 0.19, 0.13, 0.10+\varepsilon)$ and differs from $\sigma$ only in the smallest and largest eigenvalue. When $\varepsilon$ reaches $0.03$, the multiplicity of the smallest eigenvalue of $\sigma_\varepsilon^*$ increases to 2. Between $\varepsilon= 0.03$ and $\varepsilon = 0.06$, again only the smallest and largest eigenvalues change, but the smallest eigenvalue has multiplicity 2. This process continues until every eigenvalue reaches $\frac{1}{d}=0.2$ at $T(\sigma,\tau) = 0.18$.} \label{fig:Lambda-eps}
\end{figure}
\begin{lemma} \label{lem:delta_rho}
Let $\rho \neq \tau$. For any $\varepsilon \leq \delta(\rho)$, we have
\[
m_+(\rho,\varepsilon) = k_+(\rho), \quad\text{and}\quad m_-(\rho,\varepsilon) = k_-(\rho).
\]
Moreover, if $\varepsilon = \delta(\rho)$ then either $k_+ (\mmm_{\varepsilon}(\rho)) > k_+ (\rho)$ or $k_- (\mmm_{\varepsilon}(\rho)) > k_- (\rho)$.
\end{lemma}
Using this result, we can prove the Schur convexity of $\Delta_\varepsilon$ for $\varepsilon$ small enough (depending on $\rho$ and $\sigma$).
\begin{lemma}\label{lem:Delta_eps_Schur_convex_deltarho-sig}
Let $\rho,\sigma \in \mathcal{D}(\mathcal{H})$ with $\rho \prec \sigma$. Let $\varepsilon \leq \delta(\rho,\sigma)$, defined by \eqref{eq:def-delta-rho-sigma}. Then
\begin{equation} \label{eq:Delta_schur_convex_eps_small}
\Delta_\varepsilon(\rho) \leq \Delta_\varepsilon(\sigma).
\end{equation}
Moreover, equality in \eqref{eq:Delta_schur_convex_eps_small} implies that $\lambda_{\pm}(\rho) = \lambda_\pm(\sigma)$.
\end{lemma}
We can iterate this result using \Cref{lem:Delta_eps1_plus_eps2} to prove \Cref{thm:Delta_eps_Schur_convex} in general.
\paragraph{Proof of \Cref{thm:Delta_eps_Schur_convex}}
Let $\rho,\sigma\in \mathcal{D}(\mathcal{H})$ and $\varepsilon\in(0,1]$. Note that if $\sigma = \tau$, then $\rho \prec \sigma$ implies $\rho = \tau$, and hence $\Delta_\varepsilon(\rho) = 0 = \Delta_\varepsilon(\sigma)$. If $\rho = \tau \neq \sigma$, then $\Delta_\varepsilon(\rho) = 0 < \Delta_\varepsilon(\sigma)$ by the strict Schur concavity of the $H_{(h,\phi)}$ entropy. Therefore, we can assume $\rho \neq \tau\neq \sigma$.
\begin{enumerate}[{Step }1.]
\item Set $\rho_1=\rho$ and $\sigma_1=\sigma$. If $\varepsilon \leq \delta_1:=\delta(\rho_1,\sigma_1)$, we conclude via \Cref{lem:Delta_eps_Schur_convex_deltarho-sig}. Otherwise, set $\varepsilon_1 = \varepsilon- \delta_1$. Then by \Cref{lem:Delta_eps1_plus_eps2},
\[
\Delta_{\varepsilon_1+\delta_1}(\rho_1) = \Delta_{\varepsilon_1} \circ \mmm_{\delta_1}(\rho_1) + \Delta_{\delta_1}(\rho_1)
\]
and
\[
\Delta_{\varepsilon_1+\delta_1}(\sigma_1) = \Delta_{\varepsilon_1} \circ \mmm_{\delta_1}(\sigma_1) + \Delta_{\delta_1}(\sigma_1).
\]
We invoke \Cref{lem:Delta_eps_Schur_convex_deltarho-sig} to find $\Delta_{\delta_1}(\rho_1) \leq \Delta_{\delta_1}(\sigma_1)$; it remains to show
\[
\Delta_{\varepsilon_1} ( \mmm_{\delta_1}(\rho_1))\leq \Delta_{\varepsilon_1} ( \mmm_{\delta_1}(\sigma_1)).
\]
\item Set $\rho_2 = \mmm_{\delta_1}(\rho_1)$ and $\sigma_2 = \mmm_{\delta_1}(\sigma_1)$. If either $\rho_2 = \tau$ or $\sigma_2 = \tau$ we conclude by the argument presented at the start of the proof. Otherwise, we set $\delta_2 := \delta(\rho_2,\sigma_2)$ and proceed.
If $\varepsilon_1 \leq \delta_2$, we conclude by \Cref{lem:Delta_eps_Schur_convex_deltarho-sig}. Otherwise, set $\varepsilon_2 = \varepsilon_1 - \delta_2$, expand e.g.
\[
\Delta_{\varepsilon_2+\delta_2}(\rho_2) = \Delta_{\varepsilon_2} \circ \mmm_{\delta_2}(\rho_2) + \Delta_{\delta_2}(\rho_2), \qquad
\Delta_{\varepsilon_2+\delta_2}(\sigma_2) = \Delta_{\varepsilon_2} \circ \mmm_{\delta_2}(\sigma_2) + \Delta_{\delta_2}(\sigma_2), \qquad
\]
and conclude $\Delta_{\delta_2}(\rho_2) \leq \Delta_{\delta_2}(\sigma_2)$ by \Cref{lem:Delta_eps_Schur_convex_deltarho-sig}. It remains to show
\[
\Delta_{\varepsilon_2} \circ \mmm_{\delta_2}(\rho_2) \leq \Delta_{\varepsilon_2} \circ \mmm_{\delta_2}(\sigma_2).
\]
\item[Step $k$.] We continue recursively: for $k\geq 3$, we define $\rho_k =\mmm_{\delta_{k-1}}(\rho_{k-1})$, $\sigma_k = \mmm_{\delta_{k-1}}(\sigma_{k-1})$. If either $\rho_k = \tau$ or $\sigma_k = \tau$, we conclude as before; otherwise, set $\delta_k = \delta(\rho_k,\sigma_k)$. If $\varepsilon_{k-1} \leq \delta_k$, we conclude by \Cref{lem:Delta_eps_Schur_convex_deltarho-sig}; otherwise, define $\varepsilon_k = \varepsilon_{k-1} - \delta_k$, expand by \Cref{lem:Delta_eps1_plus_eps2} to find
\[
\Delta_{\varepsilon_k+\delta_k}(\rho_k) = \Delta_{\varepsilon_k} \circ \mmm_{\delta_k}(\rho_k) + \Delta_{\delta_k}(\rho_k), \qquad
\Delta_{\varepsilon_k+\delta_k}(\sigma_k) = \Delta_{\varepsilon_k} \circ \mmm_{\delta_k}(\sigma_k) + \Delta_{\delta_k}(\sigma_k),
\]
and conclude
$\Delta_{\delta_k}(\rho_k) \leq \Delta_{\delta_k}(\sigma_k)$ by \Cref{lem:Delta_eps_Schur_convex_deltarho-sig}. At the end of step $k$, it remains to show that
\[
\Delta_{\varepsilon_k} \circ \mmm_{\delta_k}(\rho_k) \leq \Delta_{\varepsilon_k} \circ \mmm_{\delta_k}(\sigma_k).
\]
\end{enumerate}
This process must terminate in less than $4d$ steps, as follows. At each step $k$ for which the process does not conclude, we have either $\delta(\rho_k,\sigma_k) = \delta(\rho_k)$ and therefore $k_+(\rho_k) > k_+(\rho_{k-1})$ or $k_-(\rho_k) > k_-(\rho_{k-1})$ or else $\delta(\rho_k,\sigma_k) = \delta(\sigma_k)$ and therefore $k_+(\sigma_k) > k_+(\sigma_{k-1})$ or $k_-(\sigma_k) > k_-(\sigma_{k-1})$, by \Cref{lem:delta_rho}. Since $k_\pm(\omega)\leq d$ for $\omega\in \mathcal{D}(\mathcal{H})$ and one of each of the four integers $k_\pm(\rho), k_\pm(\sigma)$ increases at each step, there cannot be more than $4d$ steps in total. Note that $\Delta_\varepsilon(\rho) = \Delta_\varepsilon(\sigma)$ implies equality in the use of \Cref{lem:Delta_eps_Schur_convex_deltarho-sig} in Step 1, which requires $\lambda_+(\rho) = \lambda_+(\sigma)$.\hfill\proofSymbol
\section{Proof of lemmas \label{sec:proof_lemmas}}
\paragraph{Proof of \Cref{lem:Delta_eps1_plus_eps2}}
We expand
\[
\Delta_{\varepsilon_1+\varepsilon_2}(\rho) = H_{(h,\phi)} \circ \mmm_{\varepsilon_1+\varepsilon_2}(\rho) - H_{(h,\phi)}(\rho).
\]
By \Cref{prop:properties_of_Lambda_eps}\ref{item:Lambda_eps-semigroup}, we have $\mmm_{\varepsilon_1+\varepsilon_2}(\rho) = \mmm_{\varepsilon_1}\circ \mmm_{\varepsilon_2}(\rho)$. Thus,
\begin{align*}
\Delta_{\varepsilon_1+\varepsilon_2}(\rho) &= H_{(h,\phi)} \circ\mmm_{\varepsilon_1}( \mmm_{\varepsilon_2}(\rho)) - H_{(h,\phi)}( \mmm_{\varepsilon_2}(\rho)) + H_{(h,\phi)}( \mmm_{\varepsilon_2}(\rho)) - H_{(h,\phi)}(\rho)\\
&= \Delta_{\varepsilon_1} ( \mmm_{\varepsilon_2}(\rho)) + \Delta_{\varepsilon_2}(\rho). \tag*{\proofSymbol}
\end{align*}
\paragraph{Proof of \Cref{lem:delta_rho}}
We use the notation of \Cref{def:delta-rho-sigma}. We check that the choice $m=k_+(\rho)$ satisfies the definition of $m_+(\rho,\varepsilon)$, namely that the choice $m=k_+(\rho)$ solves \eqref{eq:m_is_sol_to_this}.
\begin{itemize}
\item If $T(\rho,\tau) > \varepsilon$, then $\gamma_+^{(m)}(\rho,\varepsilon) = \frac{1}{m}\left(\sum_{i=1}^{m} \lambda_i(\rho) - \varepsilon\right)$.
And indeed, taking $m = k_+(\rho)$ we find
\[
\lambda_{k_++1}(\rho) = \mu_2 \leq \frac{1}{k_+}\left(\sum_{i=1}^{k_+} \lambda_i(\rho) - \varepsilon\right) = \mu_1 - \frac{\varepsilon}{k_+}
\]
since $\frac{\varepsilon}{k_+} \leq \frac{1}{k_+}\delta(\rho,\sigma) \leq \mu_1-\mu_2$. Additionally, $ \mu_1 - \frac{\varepsilon}{k_+} < \mu_1 = \lambda_{k_+}(\rho)$. Therefore, $m = k_+(\rho)$ solves \eqref{eq:m_is_sol_to_this}, hence $m_+(\rho,\varepsilon) = k_+(\rho)$.
\item In the case $0 < T(\rho,\tau) \leq \varepsilon$. Then $\gamma_+(\rho,\varepsilon) = \frac{1}{d}$. Since $\rho \neq \tau$, we have $\lambda_{k_+(\rho)}(\rho) = \mu_1 > \frac{1}{d}$. Moreover,
\[
k_+(\rho) \big(\mu_1 - \tfrac{1}{d}\big) \leq \tr [(\rho - \tau)_+ ] = T(\rho,\tau) \leq \varepsilon \leq k_+(\rho) (\mu_1 - \mu_2)
\]
and therefore $\mu_1 - \frac{1}{d}\leq \mu_1 - \mu_2$, yielding $\mu_2 \leq \frac{1}{d}$. Thus, $m_+(\rho,\varepsilon) = k_+(\rho)$.
\end{itemize}
Proving that $m_-(\rho,\varepsilon) = k_-(\rho)$ is analogous.
Next, consider $\varepsilon = \delta(\rho)$. If $0 < T(\rho,\tau) \leq \varepsilon$, then $\mmm_\varepsilon(\rho) = \tau$ (by \Cref{prop:properties_of_Lambda_eps}\ref{item:Lambda-eps-near-tau}) and $d = k_+(\mmm_\varepsilon(\rho)) > k_+(\rho)$, by the assumption that $\rho\neq \tau$.
Otherwise, without loss of generality, assume $\delta(\rho,\sigma) = k_+(\rho) (\mu_1-\mu_2)$. We show that $k_+(\mmm_\varepsilon(\rho)) > k_+(\rho)$. By the above, $m_+(\rho,\varepsilon) = k_+(\rho)$, and therefore
\[
\gamma_+(\rho,\varepsilon)= \mu_1 + \frac{\varepsilon}{k_+(\rho)} = \mu_1 +(\mu_1-\mu_2) = \mu_2.
\]
As
\[
\mmm_\varepsilon(\rho) =
\displaystyle \sum_{i=1}^{m_+(\rho,\varepsilon)} \gamma_+(\rho,\varepsilon)\ketbra{i}{i} + \sum_{i=m+1}^{d-m_-(\rho,\varepsilon)} \lambda_i(\rho)\ketbra{i}{i} + \sum_{i=d-m_-(\rho,\varepsilon)+1}^d \gamma_-(\rho,\varepsilon)\ketbra{i}{i}
\]
by \cref{def:Lambdaeps} and $\lambda_{m_+(\rho,\varepsilon)+1}(\rho) = \mu_2$, we have that $k_+(\mmm_\varepsilon(\rho))$, the multiplicity of $\mu_2$ for $\mmm_\varepsilon(\rho)$, is strictly larger than $k_+(\rho) =m_+(\rho,\varepsilon)$. \hfill\proofSymbol
\paragraph{Proof of \Cref{lem:Delta_eps_Schur_convex_deltarho-sig}}
As in the proof of \Cref{thm:Delta_eps_Schur_convex}, if $\sigma = \tau$, then $\rho \prec \sigma$ implies $\rho = \tau$, and hence $\Delta_\varepsilon(\rho) = 0 = \Delta_\varepsilon(\sigma)$. If $\rho = \tau \neq \sigma$, then $\Delta_\varepsilon(\rho) = 0 < \Delta_\varepsilon(\sigma)$ by the strict Schur concavity of the $H_{\id,\phi}$ entropy.
Now, assume $\rho \neq \tau \neq \sigma$. We aim to show
\begin{equation}
H_{(\id,\phi)} \circ \mmm_\varepsilon(\rho) - H_{(\id,\phi)} (\rho) \leq H_{(\id,\phi)} \circ \mmm_\varepsilon(\sigma) - H_{(\id,\phi)} (\sigma) .\label{eq:WTS_diffHs}
\end{equation}
By two applications of \Cref{lem:delta_rho}, we have $m_+(\rho,\varepsilon) = k_+(\rho)$, $m_-(\rho,\varepsilon) = k_-(\rho)$, $m_+(\sigma,\varepsilon) = k_+(\sigma)$, and $m_-(\sigma,\varepsilon) = k_-(\sigma)$. Therefore, by \eqref{eq:def_h-phi-entropy} and \eqref{def:Lambdaeps},
\[
H_{(\id,\phi)} (\mmm_\varepsilon(\rho)) = k_+(\rho) \phi( \gamma_+(\rho)) + \sum_{i = k_+(\rho)+1}^{d- k_-(\rho)} \phi(\lambda_i(\rho)) + k_-(\rho) \phi(\gamma_-(\rho))
\]
since $h=\id$. The $\phi(\lambda_i(\rho))$ terms for $i=k_+(\rho)+1,\dotsc,d- k_-(\rho)$ therefore cancel in $\Delta_\varepsilon(\rho)$ yielding
\begin{equation} \label{eq:Hidphirho}
H_{(\id,\phi)} \circ \mmm_\varepsilon(\rho) - H_{(\id,\phi)} (\rho) = k_+(\rho) [\phi( \gamma_+(\rho,\varepsilon)) -\phi(\lambda_+(\rho))]+ k_-(\rho)[ \phi(\gamma_-(\rho,\varepsilon)) -\phi(\lambda_-(\rho))]
\end{equation}
and similarly
\begin{equation}\label{eq:Hidphisigma}
H_{(\id,\phi)} \circ \mmm_\varepsilon(\sigma) - H_{(\id,\phi)} (\sigma) = k_+(\sigma) [\phi( \gamma_+(\sigma,\varepsilon)) - \phi(\lambda_+(\sigma))] + k_-(\sigma) [\phi(\gamma_-(\sigma,\varepsilon)) - \phi(\lambda_-(\sigma))].
\end{equation}
We conclude by invoking \Cref{lem:ineq} below, which bounds the first term (resp.~second term) of \eqref{eq:Hidphirho} by the first term (resp.~second term) of \eqref{eq:Hidphisigma}.\hfill\proofSymbol
\begin{lemma} \label{lem:ineq} For $\rho\prec \sigma$ with $\rho\neq\tau\neq\sigma$ and $0<\varepsilon \leq \delta(\rho,\sigma)$, we have
\begin{gather} \label{eq:alpha2ieq-proof}
k_\pm(\rho)[ \phi(\gamma_\pm(\rho,\varepsilon)) -\phi(\lambda_\pm(\rho)) ] \leq k_\pm(\sigma)[ \phi(\gamma_\pm(\sigma,\varepsilon))-\phi(\lambda_\pm(\sigma))]
\end{gather}
and that equality in \eqref{eq:alpha2ieq-proof} implies $\lambda_\pm(\rho) = \lambda_\pm(\sigma)$.
\end{lemma}
To prove this result, we first recall a simple consequence of the majorization order $\rho \prec \sigma$.
\begin{lemma}
If $\rho \prec \sigma$, then $T(\rho,\tau) \leq T(\sigma,\tau)$.
\end{lemma}
\begin{proof}
If $\rho \prec \sigma$, then by Theorem 2-2 (b) of \cite{alberti-uhlmann1982}, we have $\rho = \Phi(\sigma)$ for a map $\Phi(\cdot) = \sum_i p_i U_i \cdot U_i^*$ where $p_i$ is a finite probability distribution and each $U_i$ is unitary. $\Phi$ is completely positive and trace-preserving (CPTP) as well as unital. Since $\Phi(\tau)=\tau$,
\[
T(\rho,\tau) = T(\rho,\Phi(\tau)) = T(\Phi(\sigma),\Phi(\tau)) \leq T(\sigma,\tau)
\]
where the inequality follows from the monotonicity of the trace distance under CPTP maps.
\end{proof}
\paragraph{Proof of \Cref{lem:ineq}}
We prove the case $+$ in \cref{eq:alpha2ieq-proof}; the case $-$ is proved analogously. First, we have $\gamma_+(\rho,\varepsilon)\leq \gamma_+(\sigma,\varepsilon)$ and $\lambda_+(\rho) \leq \lambda_+(\sigma)$, using that $\rho \prec \sigma$ and $\mmm_\varepsilon \rho \prec \mmm_\varepsilon \sigma$ by \Cref{prop:properties_of_Lambda_eps}. Moreover, by definition, $\gamma_+(\rho,\varepsilon) < \lambda_+(\rho)$ and $\gamma_+(\sigma,\varepsilon) < \lambda_+(\sigma)$. Therefore, by applying \Cref{cor:main_concave_ieq} and multiplying by minus one, we have
\begin{equation}\label{eq:phi-ieq-proof-minus-1}
\frac{ \phi(\gamma_+(\rho,\varepsilon)) -\phi(\lambda_+(\rho)) }{\lambda_+(\rho)-\gamma_+(\rho,\varepsilon)} \leq \frac{ \phi(\gamma_+(\sigma,\varepsilon)) - \phi(\lambda_+(\sigma))}{\lambda_+(\sigma) - \gamma_+(\sigma,\varepsilon)}.
\end{equation}
and that equality requires $\lambda_+(\rho) = \lambda_+(\sigma)$.
Now, we complete the proof of \eqref{eq:alpha2ieq-proof} in three cases.
\begin{enumerate}
\item[Case 1:] $T(\sigma,\tau) \leq \varepsilon$. Then $T(\rho,\tau)\leq \varepsilon$ as well, and
\[
\lambda_+(\rho) - \gamma_+(\rho,\varepsilon) = \lambda_+(\rho) - \frac{1}{d}.
\]
As shown in the proof of \Cref{lem:delta_rho}, the second-largest eigenvalue of $\rho$ is less or equal to $\frac{1}{d}$. Therefore,
\[
T(\rho,\tau) = \tr[ (\rho -\tau)_+] = k_+(\rho)\big(\lambda_+(\rho) - \tfrac{1}{d}\big),
\]
and hence,
\begin{equation}\label{eq:rho-diff-is-T_proof}
\lambda_+(\rho) - \gamma_+(\rho,\varepsilon) = \frac{1}{k_+(\rho)}T(\rho,\tau).
\end{equation}
As $T(\sigma,\tau) \leq \varepsilon$, we likewise have $\lambda_+(\sigma) - \gamma_+(\sigma,\varepsilon) = \frac{1}{k_+(\sigma)}T(\sigma,\tau)$.
Then \eqref{eq:phi-ieq-proof-minus-1} yields
\[
\frac{k_+(\rho)[ \phi(\gamma_+(\rho,\varepsilon)) -\phi(\lambda_+(\rho)) ]}{T(\rho,\tau)} \leq \frac{ k_+(\sigma)[ \phi(\gamma_+(\sigma,\varepsilon))-\phi(\lambda_+(\sigma))]}{T(\sigma,\tau)}.
\]
Since $ T(\sigma,\tau)\geq T(\rho,\tau) $, we may bound the right-hand side by $\frac{ k_+(\sigma)[ \phi(\gamma_+(\sigma,\varepsilon))-\phi(\lambda_+(\sigma))]}{T(\rho,\tau)}$. Then multiplying by $T(\rho,\tau)$ yields \eqref{eq:alpha2ieq-proof}.
\item[Case 2:] $T(\rho,\tau) \leq \varepsilon < T(\sigma,\tau)$. In this case, \eqref{eq:rho-diff-is-T_proof} holds, and $\gamma_+(\sigma,\varepsilon) = \lambda_+(\sigma) - \frac{\varepsilon}{k_+(\sigma)}$.
Therefore, \eqref{eq:phi-ieq-proof-minus-1} yields
\[
\frac{k_+(\rho) [\phi(\gamma_+(\rho,\varepsilon)) -\phi(\lambda_+(\rho)) ]}{T(\rho,\tau)} \leq \frac{ k_+(\sigma)[\phi(\gamma_+(\sigma,\varepsilon)) - \phi(\lambda_+(\sigma))]}{\varepsilon}.
\]
Similarly to the previous case, the inequality $\varepsilon\geq T(\rho,\tau) $ bounds the right-hand side by $\frac{ k_+(\sigma)[\phi(\gamma_+(\sigma,\varepsilon)) - \phi(\lambda_+(\sigma))]}{T(\rho,\tau)}$, and multiplying by $T(\rho,\tau)$ yields \eqref{eq:alpha2ieq-proof}.
\item[Case 3:] $T(\rho,\tau) > \varepsilon$. Then $\gamma_+(\rho,\varepsilon) = \lambda_+(\rho) - \frac{\varepsilon}{k_+(\rho)}$, and $\gamma_+(\sigma,\varepsilon) = \lambda_+(\sigma) - \frac{\varepsilon}{k_+(\sigma)}$. Therefore, \eqref{eq:phi-ieq-proof-minus-1} yields
\[
\frac{k_+(\rho)[ \phi(\gamma_+(\rho,\varepsilon)) -\phi(\lambda_+(\rho)) ]}{\varepsilon} \leq \frac{ k_+(\sigma)[ \phi(\gamma_+(\sigma,\varepsilon))-\phi(\lambda_+(\sigma))]}{\varepsilon}
\]
and multiplying by $\varepsilon$ yields \eqref{eq:alpha2ieq-proof}.
\end{enumerate}
Note that in each case, equality in \eqref{eq:alpha2ieq-proof} requires equality in \eqref{eq:phi-ieq-proof-minus-1}.
\hfill\proofSymbol
|
\section{Introduction}
The enormous development in the field of ultracold atoms in the last 20 years is not only due to the advanced cooling methods \cite{Chu1998,Phillips1998,Cohen1998}, but also to the imaging techniques. A diagnosis of a Bose-Einstein condensate is always based on atom-light interactions. The multitude of possibilities created by this phenomenon allows researchers to deeply examine various aspects of quantum systems at extremely low temperatures \cite{Anderson1995,Davis1995, Ketterle2002, Cornell2002}.\\
\indent One of the most widespread experimental method of investigation of ultracold gases is the absorption imaging. The key role in this technique plays a fact that the absorption rate is proportional to the column density of atoms, or so is tacitly assumed. Sending a resonant probe light pulse through a sample and observing the "shadow" left by the atomic absorption with a CCD camera opens access to many relevant properties like density profiles of atomic clouds and higher order correlation functions \cite{grondalski1999,altman2004,folling2005,greiner2005,esteve2006}. Recently it was even possible to image a single atom \cite{streed2012}. Typically a weak pulse is used to avoid multiple scatterings and therefore enhance control over the measurement.\\
\indent Although the general principles of absorption imaging are well understood, there remain some open questions waiting to be addressed. Recently a process of splitting of the Bose-Einstein condensate was analysed \cite{gorski2015}. Within the classical field approximation, see for instance \cite{witkowska2009,brewczyk2013}, it was shown that the statistical properties of the condensate depend on the observation time. Whereas it is possible that these findings are not exactly related to a real quantum measurement with light, a role of spatial and temporal properties of a light pulse in the absorption imaging, to the best of our knowledge, was not analysed in detail. Well controlled experiments with just a few trapped atoms are now possible \cite{Serwane2011,Baier2016}. Such a simple system offers the unique opportunity to scrutinize the absorption imaging method at the microscopic level.\\
\indent To shed more light on this problem we present an oversimplified example of only two atoms located in a spherical harmonic trap. Then, a diagnosis of the system with a light pulse is done through the absorption. Identifying one-photon and a two-photon absorption probabilities as a source of a one-particle density distribution and two-body distribution respectively we study the influence of pulse properties on the results.
\section{Model}
Our method of diagnosing the two-body wave function is based on the absorption of sufficiently well collimated light pulses by atoms - both bosons and fermions. The probabilities of one or two photons being absorbed should be measured for different positions of light beams. Out of the estimated likelihoods we may find one - particle density distribution and two - body distribution. In our studies we are going to ignore spontaneous emission.
The total Hamiltonian of the two contact interacting ultracold particles of equal masses ${m_1} = {m_2}$ absorbing photons from a light beam reads:
\begin{equation}
H = {H_{FS}} + {H_{AF}}.
\end{equation}
The term $H_{FS}$ stands for the free system Hamiltonian. It can be written as:
\begin{eqnarray}
{H_{FS}} &=& \underbrace{\sum\limits_{i = 1}^2 {\left( {{T_i}({{\bf{r}}}_i) + {V_t}({{\bf{r}}}_i)} \right)} + {V_I}({{\bf{r}}}_1 - {{\bf{r}}}_2)}_{H_{S}} \nonumber \\
&\ & + \frac{1}{2}\hbar {\omega _0}\sum\limits_{i = 1}^2{\left( {\left. {\left| e \right.} \right\rangle_i \left\langle {\left. e \right| - } \right.\left. {\left| g \right.} \right\rangle_i \left\langle {\left. g \right|} \right.} \right)},\label{hamfree}
\end{eqnarray}
where ${{T_i}({{\bf{r}}}_i)}=-\frac{\hbar}{2m_i}\nabla^2_{{\bf {r}}_i}$ is the kinetic energy of an atom and ${V_t}({\bf {r}}_i) = \frac{1}{2}m_i\omega {\bf {r}}_i$ a spherical harmonic trapping potential. The short range interaction term for ultracold bosons is expressed by ${V_I}({\bf {r}}) = g\delta_{ps}({\bf {r}})$ with $\delta_{ps}({\bf {r}})$ standing for the pseudo potential which depends on dimensionality, see \cite{Busch1998}. The interactions strength $g$ can be either positive or negative. Its dependence on the scattering length for different dimensions can be found in \cite{Busch1998}. Note that for two ultra cold fermions in the same spin state ${V_I}({\bf {r}}) =0$. We denote the spatial part of $H_{FS}$ as $H_S$. The last sum in Eq. \eqref{hamfree} describes the internal structure of atoms, which is in a form of a simple two-level model \cite{knight2005}. Here $\left. {\left| g \right.} \right\rangle$ indicates the ground state, $\left. {\left| e \right.} \right\rangle$ stands for the excited state and $\hbar\omega_0$ is the energy difference between two states.\\
\indent The interaction of atoms with a beam is represented by $H_{AF}$ term
\begin{eqnarray}
{H_{AF}} &=& \hbar \lambda \sum\limits_{i = 1}^2 {\left( {\sigma _ + ^i + \sigma _ - ^i} \right) \times} \nonumber \\ &\ & \times \left( {\epsilon({{\bf{r}}_i},t){e^{i({{\bf{k}}_L} \cdot {{\bf{r}}_i} - {\omega _L}t)}} + c.c.} \right) \\
&\approx& \hbar \lambda \sum\limits_{i = 1}^2 {\left( {\sigma _ + ^i{e^{ i({{\bf{k}}_L} \cdot {{\bf{r}}_i} - {\omega _L}t)}} + \sigma _ - ^i{e^{-i({{\bf{k}}_L} \cdot {{\bf{r}}_i} - {\omega _L}t)}}} \right)} \nonumber, \label{hamatfieldfinal}
\end{eqnarray}
where ${\sigma _ \pm }$ are the ladder operators defined as ${\sigma _ + } = \left. {\left| e \right.} \right\rangle \left\langle {\left. g \right|,\quad } \right.{\sigma _ - } = \left. {\left| g \right.} \right\rangle \left\langle {\left. e \right|} \right.$. The expression in the last line of Eq. \eqref{hamatfieldfinal} is obtained by using the Rotating Wave Approximation (RWA) \cite{knight2005}. We assume a weak classical monochromatic beam with an electric field given by ${\bf{E}}({\bf{r}},t) =E_0\left( {\epsilon({\bf{r}},t)e^{i({{\bf{k}}_L} \cdot {\bf{r}} - {\omega _L}t)}} + {\epsilon^{*}({\bf{r}},t)e^{ - i({{\bf{k}}_L} \cdot {\bf{r}} - {\omega _L}t)}}\right)$, where $E_0$ indicates a real-valued magnitude of an amplitude of the electric field, ${\bf k}_L$ a wave vector and $\omega_L$ an angular frequency of the light. An envelope of a light pulse denoted by $\epsilon({\bf{r}},t)$ is spatially and temporally dependent. We assume a weak intensity of the pulse so that a parameter $\lambda=\frac{dE_0}{\hbar}$ is small as compared to the other terms in the total Hamiltonian. Here $d$ stands for a transition dipole moment of the atom.\\
\indent A state vector ${\bf \Psi} ({{\bf r}_1},{{\bf r}_2},t)$ of two atoms within our model can be written in a general form as:
\begin{equation}
{\bf \Psi} ({{\bf r}_1},{{\bf r}_2},t)= \begin{pmatrix} \phi ({{\bf r}_1},{{\bf r}_2},t)\left| {gg} \right\rangle \\ \frac{1}{\sqrt{2}} \left({\chi _1}({{\bf r}_1},{{\bf r}_2},t)\left| {eg} \right\rangle \pm {\chi _2}({{\bf r}_1},{{\bf r}_2},t)\left| {ge} \right\rangle \right) \\ \eta ({{\bf r}_1},{{\bf r}_2},t)\left| {ee} \right\rangle \end{pmatrix}
\end{equation}
where $+$ sign in a second line of the state vector corresponds to bosons, whereas $-$ sign to fermions. The time-dependent Shr\"{o}dinger equation $i\hbar \frac{{\partial {\bf \Psi}}}{{\partial t}} = H{\bf \Psi}$ can be expressed as a system of equations for unknown functions $ \phi$, $ \chi_1$, $ \chi_2 $ and $ \eta$ by:
\begin{equation}
\left\{
\begin{aligned}
i\hbar {\mathop \phi \limits^.} &= \left( {{H_S} - \hbar {\omega _0}} \right) \phi + \frac{\hbar \lambda}{{\sqrt 2 }} \left( {\epsilon^*}({{\bf{r}}_1},t){e^{ - i({{\bf{k}}_{\bf{L}}} \cdot {{\bf{r}}_{\bf{1}}} - {\omega _L}t)}}{\chi _1} \right. \\
& \qquad \qquad \qquad \qquad \pm \left. {\epsilon^*}({{\bf{r}}_2},t){e^{ - i({{\bf{k}}_{\bf{L}}} \cdot {{\bf{r}}_2} - {\omega _L}t)}}{\chi _2} \right) \\
i\hbar {\mathop \chi \limits^.}_1 &= {H_S}{\chi _1} + \sqrt 2 \hbar \lambda \left( \epsilon({{\bf{r}}_1},t){e^{i({{\bf{k}}_{\bf{L}}} \cdot {{\bf{r}}_1} - {\omega _L}t)}}\phi \right. \\ & \qquad \qquad \qquad \qquad \left. + {\epsilon^*}({{\bf{r}}_2},t){e^{ - i({{\bf{k}}_{\bf{L}}} \cdot {{\bf{r}}_2} - {\omega _L}t)}}\eta \right) \\
i\hbar {\mathop \chi \limits^.}_2 &= {H_S}{\chi _2} \pm \sqrt 2 \hbar \lambda \left( \epsilon({{\bf{r}}_2},t){e^{i({{\bf{k}}_{\bf{L}}} \cdot {{\bf{r}}_2} - {\omega _L}t)}}\phi \right. \\ & \qquad \qquad \qquad \qquad \left. + {\epsilon^*}({{\bf{r}}_1},t){e^{ - i({{\bf{k}}_{\bf{L}}} \cdot {{\bf{r}}_1} - {\omega _L}t)}}\eta \right) \\
i\hbar \mathop \eta \limits^. &= \left( {{H_S} + \hbar {\omega _0}} \right)\eta + \frac{\hbar \lambda }{{\sqrt 2 }}\left( \epsilon({{\bf{r}}_2},t){e^{i({{\bf{k}}_{\bf{L}}} \cdot {{\bf{r}}_2} - {\omega _L}t)}}{\chi _1} \right. \\ & \qquad \qquad \qquad \qquad \left. \pm \epsilon({{\bf{r}}_1},t){e^{i({{\bf{k}}_{\bf{L}}} \cdot {{\bf{r}}_1} - {\omega _L}t)}}{\chi _2} \right)
\end{aligned}\label{eqs}
\right.
\end{equation}
The above system of equations may be solved approximately in the following way. As it was mentioned before we consider a very weak driving to neglect a depletion of the initial state. We also assume that initially two atoms are in the internal ground states, namely that for $t=0$
a state vector ${\bf \Psi} ({{\bf r}_1},{{\bf r}_2},0)=\phi ({{\bf r}_1},{{\bf r}_2},0)\left| {gg} \right\rangle$. Therefore we assume that during the interaction between the system and the light the state vector remains almost unchanged, that is to say $\left| \phi \right| \gg \left| {{\chi _1}} \right|,\left| {{\chi _2}} \right| \gg \left| \eta \right|$ for the duration of the pulse. Then, introducing the interaction picture by following substitutions $\phi \to \quad {e^{i{\omega _0}t}}\phi$, $\eta \to \quad {e^{ - i{H_S}t/\hbar }}{e^{ - i{\omega _0}t}}\eta$ and ${\chi _{1(2)}} \to \quad {e^{ - i{H_S}t/\hbar }}{\chi _{1(2)}}$ we obtain the final equations
\begin{equation}
\left\{
\begin{aligned}
\phi &= {e^{ - i{H_S}t/\hbar }} \phi({\bf {r}}_1,{\bf {r}}_2,0) \\
{\mathop \chi \limits^.}_1&= -i\sqrt 2 \lambda {e^{ i{H_S}t/\hbar }} \epsilon({{\bf{r}}_1},t){e^{i{{\bf{k}}_{\bf{L}}} \cdot {{\bf{r}}_1}}}{e^{ - i\Delta t}} \phi \\
{\mathop \chi \limits^.}_2 &= \mp i \sqrt 2 \lambda {e^{ i{H_S}t/\hbar }} \epsilon({{\bf{r}}_2},t){e^{i{{\bf{k}}_{\bf{L}}} \cdot {{\bf{r}}_2}}} {e^{ - i\Delta t}}\phi \\
\mathop \eta \limits^. &= -i \frac{1}{{\sqrt 2 }} \lambda \left( {e^{ i{H_S}t/\hbar }} \epsilon({{\bf{r}}_2},t){e^{i{{\bf{k}}_{\bf{L}}} \cdot {{\bf{r}}_2}}}{e^{ - i\Delta t}}{e^{ - i{H_S}t/\hbar }} {\chi _1} \right. \\ & \left. \qquad \pm{e^{ i{H_S}t/\hbar }} \epsilon({{\bf{r}}_1},t){e^{i{{\bf{k}}_{\bf{L}}} \cdot {{\bf{r}}_1}}}{e^{ - i\Delta t}}{e^{ - i{H_S}t/\hbar }} {\chi _2} \right)
\end{aligned}
\right.\label{eqsfinal}
\end{equation}
where we define a detuning by $\Delta=\omega_L-\omega_0$. As we may note the final form of the above system of equations comes directly from the fact that in a general case $\left[ {\epsilon({\bf {r}}_i){e^{i{{\bf{k}}_{\bf{L}}} \cdot {{\bf{r}}_i}}},{H_S}} \right] \ne 0$.
\section{Solutions}
The analytical solutions of the spatial Hamiltonian $H_{S}$ are well known both for two ideal bosons or fermions (${V_I}({\bf {r}}) = 0$) and for two interacting ultra cold bosons \cite{Busch1998}. We assume a rectangle pulse envelope. When the light is on $ \epsilon({{\bf{r}}},t)=\epsilon({{\bf{r}}})$. Thus it is possible to solve Eq. \eqref{eqsfinal} analytically. Without loss of generality we choose the initial state as an eigenvector of $H_S$, namely that $\phi ({{\bf r}_1},{{\bf r}_2},0)=\phi_n({{\bf r}_1},{{\bf r}_2})$ with an index $n$ indicating the nth eigenvector in a chosen basis. Then using Dirac notation and a formula ${e^{ - i{H_S}t/\hbar }} = \sum\limits_i {{e^{ - i{E_i}t/\hbar }}\left| {{\phi _i}} \right\rangle \left\langle {{\phi _i}} \right|} $ with $E_i$ standing for the i-th eigenvalue we find a general solution for $ \chi_1$, $ \chi_2 $ and $ \eta$ as
\begin{equation}
\left\{
\begin{split}
\phi &= {e^{ - i{E_n}t/\hbar }}\left| {{\phi _n}} \right\rangle\\
\chi_1&= -\sqrt 2 \lambda \sum\limits_i {\epsilon_{in}\frac{\left({e^{ i{\tilde{\Delta}_{in}}t}}-1\right)}{\tilde{\Delta}_{in}} \left| {{\phi _i}} \right\rangle} \\
\chi_2&= \mp\sqrt 2 \lambda \sum\limits_i {\tilde{\epsilon}_{in}\frac{\left({e^{ i{\tilde{\Delta}_{in}}t}}-1\right)}{\tilde{\Delta}_{in}} \left| {{\phi _i}} \right\rangle} \\
\eta &= -\lambda^2 \sum\limits_{i,k} \frac{\tilde{\epsilon}_{ik}{\epsilon}_{kn}+{\epsilon}_{ik}\tilde{\epsilon}_{kn}}{\tilde{\Delta}_{kn}}\times \\ & \quad \left(\frac{{e^{ i{\tilde{\Delta}_{ik}}t}}-1}{\tilde{\Delta}_{ik}}-\frac{{e^{ i\left({\tilde{\Delta}_{ik}}+{\tilde{\Delta}_{kn}}\right)t}}-1}{\tilde{\Delta}_{ik}+\tilde{\Delta}_{kn}}\right) \left| {{\phi _i}} \right\rangle
\end{split}
\right.\label{solution}
\end{equation}
where $\epsilon_{ij}=\left \langle {{\phi _i}} \right|\epsilon({\bf{r}}_1){e^{i{{\bf{k}}_{\bf{L}}} \cdot {{\bf{r}}_1}}}\left| {{\phi _j}} \right\rangle$, $\tilde{\epsilon}_{ij}=\left \langle {{\phi _i}} \right|\epsilon({\bf{r}}_2){e^{i{{\bf{k}}_{\bf{L}}} \cdot {{\bf{r}}_2}}}\left| {{\phi _j}} \right\rangle$ and obviously $\epsilon_{ij}=\tilde{\epsilon}_{ij}$. Here we define a scalar product by $ \langle {{\phi _i}} \left| {{\phi _j}} \right\rangle=\int{d{\bf{r}}_1d{\bf{r}}_2{\phi^*_i}({\bf{r}}_1,{\bf{r}}_2){\phi_j}({\bf{r}}_1,{\bf{r}}_2)}$. The generalized energy difference between i-th and j-th states reads $\tilde{\Delta}_{ij}=\Delta_{ij}-\Delta=\frac{E_i-E_j}{\hbar}-\Delta$. Note that for resonant terms i.e. $\tilde{\Delta}_{ij}=0$ in any sum of Eq. \eqref{solution} its proper element has to be evaluated by taking a limit $\tilde{\Delta}_{ij}\rightarrow 0$. A resonant term behaves as $t$ for $\chi_1$, $\chi_2$ and as $t^2$ for $\eta$. For the clarity of our argumentation hereafter we will take a resonant case with $\Delta=0$, but our conclusions will hold also for $\Delta \neq 0$.\\
\indent The probabilities of having one or two photons absorbed by atoms are easily defined by $P_1(t)={\left| {\chi _1}\left| {eg} \right\rangle \pm {\chi _2} \left| {ge} \right\rangle \right|}^2={\left| {\chi _1} \right|}^2+{\left| {\chi _2} \right|}^2 $ and $P_2(t)={\left| {\eta}\left| {ee} \right\rangle \right|}^2={\left| {\eta} \right|}^2$. After some straightforward calculations they can be expressed as:
\begin{align}
P_1(t) &=2{\lambda ^2}{\left| {{\epsilon _{nn}}} \right|^2}{t^2} + 4{\lambda ^2}\sum\limits_{i \ne n} {\frac{{{{\left| {{\epsilon _{in}}} \right|}^2}}}{{{\Delta ^2}_{in}}}\left( {1 - \cos ({\Delta _{in}}t)} \right)}\label{p1} \\
P_2(t) &={\lambda ^4}{\left| {{\epsilon _{nn}}} \right|^4}{t^4} + 4\lambda^4\sum\limits_{\scriptstyle i \ne n \hfill \atop
\scriptstyle k,k' \hfill} {\epsilon^*_{ik'}\epsilon^*_{k'n}\epsilon_{ik}\epsilon_{kn}}p_{ikk'n}(t)\label{p2},
\end{align}
where $p_{ikk'n}(t)$ is of order $o(t^4)$. Its definition can be find in the footnote \footnote{The full expression for $p_{ikk'n}(t)$ is given by:
\begin{equation}\label{eqapp}
\begin{split}
p_{ikk'n}(t)=&\frac{1}{\Delta_{k'n}\Delta_{kn}}\times \\
&\left(\frac{{e^{ i{\Delta_{ik}}t}}-1}{{\Delta}_{ik}}-\frac{{e^{ i\left({{\Delta}_{ik}}+{{\Delta}_{kn}}\right)t}}-1}{{\Delta}_{ik}+{\Delta}_{kn}}\right) \times \\ & \left(\frac{{e^{ -i{\Delta_{ik'}}t}}-1}{{\Delta}_{ik'}}-\frac{{e^{ -i\left({{\Delta}_{ik'}}+{{\Delta}_{k'n}}\right)t}}-1}{{\Delta}_{ik'}+{\Delta}_{k'n}}\right)
\end{split}
\end{equation}
Note that for resonant terms i.e. ${\Delta}_{ij}=0$ in any sum of Eq. \eqref{eqapp} its proper element has to be evaluated by taking a limit ${\Delta}_{ij}\rightarrow 0$. By using Eq. \eqref{eqapp} it is easy to check that the sum in Eq. \eqref{p2} is a real function of time.}. A short time characteristic of the probabilities, when $t\ll \Delta_{in}^{-1}$ with $i$ corresponding to the nearest eigenvalue to $n$, reads:
\begin{align}
P_1(t) &\approx 2{\lambda ^2}\left \langle {{\phi _n}} \right|\left|\epsilon({\bf{r}}_1)\right|^2\left| {{\phi _n}} \right\rangle{t^2}, \quad t\rightarrow 0 \label{p1s} \\
P_2(t) &\approx{\lambda ^4}\left \langle {{\phi _n}} \right|\left|\epsilon({\bf{r}}_1)\right|^2\left|\epsilon({\bf{r}}_2)\right|^2\left| {{\phi _n}} \right\rangle{t^4}, \quad t\rightarrow 0 \label{p2s}.
\end{align}
The analysis of Eq. \eqref{p1}, \eqref{p2}, \eqref{p1s} and \eqref{p2s} reveals an intriguing discrepancy between the short time and the long time behaviour of the probabilities. First of all, the long time probabilities depend on couplings between an actual state of the system and different eigenstates that occur because, for the experimental relevance, beam width must be narrower than a characteristic system width. This fact automatically leads to the conclusion that for longer pulses the information about the actual state of the system is blurred. Secondly, although the dominant terms are of the same order in both situations, the coefficients determining their magnitude are not. For the short time the probabilities coefficients are related to the intensity $\left|\epsilon({\bf{r}})\right|^2 $. Note that for $P_1(t), \quad t \rightarrow 0$ the coefficient in front of $t^2$ can be rewritten as $\int{d{\bf{r}}{\rho}({\bf{r}})\left|\epsilon({\bf{r}})\right|^2}$ with a one-particle density ${\rho}({\bf{r}})=\int{d{\bf{r}}'\left|\phi_n({\bf{r}},{\bf{r}}') \right|^2 }$ which is a very intuitive result. On the other hand the coefficients for the long time depends on the amplitude of the pulse proportional to $\epsilon({\bf{r}}){e^{i{{\bf{k}}_{\bf{L}}} \cdot {{\bf{r}}}}}$ rather than to its intensity alone. In the next section we are going to show the most striking examples of the above differences.
\section{Results}
In this section we present results obtained within our model which are mimicking an experiment diagnosing a quantum state of two ultra cold atoms. We restrict our findings to a quasi-1D system which captures all essentials features of our model and provides with a clear picture. In a real experiment it corresponds to cigar-shaped traps with a very strong transverse confinement. We send a probing light pulses along transverse direction z which is related to an electric field ${\bf{E}}({\bf{r}},t) =E_0\left( {\epsilon(x)e^{i(k_L z - {\omega _L}t)}} + {\epsilon^{*}(x)e^{ - i(k_L z - {\omega _L}t)}}\right)$ with $k_L=\left| {\bf{k}}_L \right|$. We may also assume that $1 /k_L$ is much bigger than a typical transverse length of a probe so that a driving term $e^{\pm i(k_L z )}$ may be neglected. As the initial state $\phi(x_1,x_2,0)$ we select the ground state both for ideal fermions or bosons and interacting bosons.
\subsection{One-particle density function}
In order to find a one-particle density function we use a single pulse with $\epsilon (x;{x_0},\sigma ) = \frac{1}{{\sigma \sqrt \pi }}{e^{ - {{\left( {x - {x_0}} \right)}^2}/{\sigma ^2}}}$. Then using a short-time characteristic of $P_1(t)$ expressed by Eq. \eqref{p1s} we evaluate a probability of one-photon absorption as a function of a position of the pulse center $x_0$. We compile our results for two ideal or interacting bosons and two fermions in Fig. \ref{fig:fro}. As we may note by comparing with the well known analytical solutions of $H_S$ for the ground state the one-photon absorption diagnosis gives a direct access to the one-particle density distribution $\rho(x)$ defined in the preceding section. A clear difference between interacting and non-interacting case is seen as well as between bosons and fermions. A repulsive system distribution is wider than that of an ideal gas, while an attractive system is narrower than the ideal one.\\
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{fig1}
\caption{\label{fig:fro}}One-photon absorption probability $P_1(x_0; t_0, \sigma)/\lambda^2t_0^2$ as a function of a beam center $x_0$. A time $t_0$ was chosen so that Eq. \eqref{p1} and Eq. \eqref{p1s} agree with each other i.e. $t_0=0.0001\Delta_{10}^{-1}$, where $\Delta_{10}$ is related to the energy difference between the ground and the first excited state. The straight blue line corresponds to two ideal bosons, the straight red line comes from two interacting repulsively bosons, the black dashed line is related to two interacting attractively bosons and the green dashdotted line to two fermions. The oscillatory units are used.
\end{figure}
\indent The analysis of Eq. \eqref{p1} shows that using a pulse that is too long may affect a measured one-particle density profile. For pulses with complicated wave fronts the coefficients in the Eq. \eqref{p1} would differ significantly from these of the Eq. \eqref{p1s}. To illustrate the unwanted field amplitude dependence of the result for long pulses we choose an extreme example of a pulse with cross-section given by $\epsilon (x;{x_0},\sigma ) = \frac{1}{{\sigma \sqrt \pi }}{e^{ - {{\left( {x - {x_0}} \right)}^2}/{\sigma ^2}}}{\rm sgn(x)}$ with ${\rm sgn(x)}$ staying for the sign function. A comparison between the one-particle density profiles for the ideal bosons for two different pulse durations $t_0=0.0001\Delta_{10}^{-1}$ and $t_0=\Delta_{10}^{-1}$, where $\Delta_{10}$ is related to the energy difference between the ground and the first excited state, is presented in Fig. \ref{fig:fcompro}. For the result based on Eq. \eqref{p1} we truncate the sum at $i=20$ ensuring that adding another eigenstate would not change the result up to 1\% accuracy. A striking difference can be observed. The density profile obtained after a measurement with a long pulse has nothing in common with the actual one-particle density. It is a clear indication that a diagnosis of a few-body quantum state can be highly biased for longer pulses.
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{fig2}
\caption{\label{fig:fcompro}}One-photon absorption probability $P_1(x_0; t_0, \sigma)/\lambda^2t_0^2$ as a function of a beam center $x_0$ for two ideal bosons. The straight blue line corresponds to time $t_0=0.0001\Delta_{10}^{-1}$, whereas the red dashed to time $t_0=\Delta_{10}^{-1}$. The oscillatory units are used.
\end{figure}
\subsection{Two-body wave function}
Measuring a two-photon absorption probability a two-body wave function can be diagnosed. To achieve that one has to use a pulse with a double-focused envelope, namely with $\epsilon (x;{x_1},{x_2},\sigma ) = \frac{1}{{\sigma \sqrt \pi }}\left( {{e^{{{\left( {x - {x_1}} \right)}^2}/{\sigma ^2}}} + {e^{{{\left( {x - {x_2}} \right)}^2}/{\sigma ^2}}}} \right)$. Then a coefficient in Eq. \eqref{p2s} reads
\begin{equation*}
\begin{split}
& \left \langle {{\phi _G}} \right|\left|\epsilon(x; x_1, x_2, \sigma)\right|^2\left|\epsilon(y; x_1, x_2, \sigma)\right|^2\left| {{\phi _G}} \right\rangle \approx \\
& \qquad \left \langle {{\phi _G}} | 2\left|\epsilon(x; x_1, \sigma)\right|^2\left|\epsilon(y; x_2, \sigma)\right|^2 + \left|\epsilon(x; x_1, \sigma)\right|^2 \right. \times \\
&\qquad \qquad \quad \left. \left|\epsilon(y; x_1, \sigma)\right|^2 +\left|\epsilon(x; x_2, \sigma)\right|^2\left|\epsilon(y; x_2, \sigma)\right|^2 | {{\phi _G}} \right\rangle
\end{split}
\end{equation*}
with $\left| {{\phi _G}} \right\rangle$ denoting the ground state of $H_{S}$. Here we neglect the interference terms like ${e^{{{\left( {x - {x_1}} \right)}^2}/{\sigma ^2}}}{e^{{{\left( {x - {x_2}} \right)}^2}/{\sigma ^2}}}$ that in experiment can be realized either by ensuring $\left| x_1-x_2\right| > 3\sigma$ or by introducing a phase difference between two pulses and averaging over many measurements. The last two terms of a sum in the above equation are corresponding to processes where two photons were absorbed at the same space point. The probability of such a process should be measured independently and then subtracted from the total result of the two-photon absorption. The easiest way to notice that is by assumption that $\left|\epsilon(x; x_1, \sigma)\right|^2\approx \delta (x-x_1)$ which makes
$\left \langle {{\phi _G}} \right|\left|\epsilon(x; x_1, x_2, \sigma)\right|^2\left|\epsilon(y; x_1, x_2, \sigma)\right|^2\left| {{\phi _G}} \right\rangle \approx 2\left|\phi_G( x_1, x_2)\right|^2+\left|\phi_G( x_1, x_1)\right|^2+\left|\phi_G( x_2, x_2)\right|^2$.\\
\indent We sum up our considerations with an example of two repulsive bosons with interaction strength $g=6.$. The total two-photon absorption probability $P_2(t)$ was found for $t_0=0.0001\Delta_{10}^{-1}$ and for several beam positions $x_1$ and $x_2$. Then we subtract from it the probabilities of a single pulse two-photon absorption. Finally we compare our findings with the analytical solution for $\phi_G(x,y)$ which is plotted in Fig. \ref{fig:wavfun}. It is a straightforward observation that we reconstructed the actual two-body wave function density within our model.
\begin{figure*}
\centering
\includegraphics[width=0.75\textwidth]{fig3.eps}
\caption{\label{fig:wavfun}} Two-photon absorption probability $P_2(x_1, x_2; t_0, \sigma)/\lambda^4t_0^4$ for two interacting bosons with $g=6$ as a function of a beam positions $x_1$ and $x_2$ for $t_0=0.0001\Delta_{10}^{-1}$ after subtracting a single beam two-photon absorption (right). The oscillatory units are used. Left plot shows the analytical solution of $\left| \phi_G(x_1, x_2)\right|^2$ for $g=6$.
\end{figure*}
Analogously to the previous subsection the results for a long pulse in a time domain when Eq. \eqref{p2} holds may lead to a wrong two-body wave function. One more time we use the example of the highly modified wave front with
$\epsilon (x;{x_1},{x_2},\sigma )=\frac{1}{{\sigma \sqrt \pi }}\left( {{e^{{{\left( {x - {x_1}} \right)}^2}/{\sigma ^2}}} + {e^{{{\left( {x - {x_2}} \right)}^2}/{\sigma ^2}}}} \right) {\rm sgn(x)}$. The resulting two-photon probability absorption $P_2(t)/\lambda^4t^4$ for $t_0=0.0001\Delta_{10}^{-1}$ and $t_0=\Delta_{10}^{-1}$ for two interacting bosons with $g=6.$ as a function of beam positions can be found in Fig. \ref{fig:bias}. For the result based on Eq. \eqref{p2} we truncate the sum at $i=20$ ensuring that adding another eigenstate would not change the result up to 1\%. accuracy. Our findings stress the fact that a pulse duration in an experimental diagnosis should be chosen very carefully.
\begin{figure*}
\centering
\includegraphics[width=0.75\textwidth]{fig4.eps}
\caption{\label{fig:bias}}Two-photon absorption probability $P_2(x_1, x_2; t_0, \sigma)/\lambda^4t_0^4$ for two interacting bosons with $g=6$ as a function of a beam positions $x_1$ and $x_2$ for $t_0=0.0001\Delta_{10}^{-1}$ (left) and $t_0=\Delta_{10}^{-1}$ (right) after subtracting a single beam two-photon absorption. The oscillatory units are used.
\end{figure*}
\section{Conclusions}
In conclusions, we studied a simple model of diagnosing a two-body state with light for interacting or ideal bosons and fermions. We demonstrated that results of an experiment based on our theory would crucially depend on a pulse duration. For a sufficiently short pulses we correctly estimate with our measurement the actual one-particle density function and the two-body wave function. For longer pulses a hypothetical experimental findings would be highly biased. The main reason for that is that the calculated probabilities of one-photon or two-photons absorptions are related to the intensity of the light beam for sufficiently short time, whereas for longer time they depend on the amplitude of a pulse. The structure of Eq. \eqref{p1} and Eq. \eqref{p2} can be understood that the probability of the absorption in a space point is strongly blurred by free evolution of the initial state. Our results can be generalized to systems containing more particles which allows to investigate an experimental procedure of diagnosing higher-order correlation functions from the many-body perspective.
\acknowledgements
RO and KRz acknowledge the support of (Polish) National Science Center grant no. 2015/19/B/ST2/02820
\bibliographystyle{eplbib.bst}
|
\section{\label{sec:level1}Introduction}
String theory provides us a valuable tool to investigate strongly coupled gauge theories using AdS/CFT duality(holography)\cite{Maldacena:1997re,Gubser:1998bc,Witten:1998qj}. The duality establishes a connection between strongly coupled gauge theory in d-dimensions on the boundary and its weakly coupled gravity dual in (d+1)-dimensional bulk spacetime. Many important phenomenon of nuclear physics, condensed matter physics and high $T_c$ superconductors are being explored using this duality \cite{Son:2002sd, Erlich:2005qh,DaRold:2005zs,Karch:2006pv,Hartnoll:2008kx,Jain:2009bi}.
Recently, considerable interest is seen to study realistic condensed matter systems using holography. The investigation of different phases of strongly coupled systems requires new holographic models. Significant results have been obtained in this area after including momentum dissipation term in holographic models. Realistic examples of strongly coupled systems have finite DC conductivity either due to the presence of impurity or as a result of broken translational invariance. One can introduce momentum relaxation in holographic systems through various ways; by introducing impurity in holographic set-up \cite{Hartnoll:2008hs, Andrade:2013gsa} or by introducing spatial source field which breaks translational symmetry \cite{Donos:2014yya, Amoretti:2014ola}. Also the breaking of diffeomorphism invariance using massive gravity term in the bulk theory, results in finite DC conductivity \cite{Davison:2013jba,Blake:2013bqa,Vegh:2013sk,Amoretti:2014mma,Baggioli:2014roa}.
Further, efforts are being made to study strongly coupled condensed matter systems near quantum critical points using holography. These critical points are realized in the boundary system by opting for anisotropic scaling between temporal and spatial directions in the gravity set-up. Although the introduction of anisotropy results in breaking of Lorentz invariance, the metric remains scale invariant. The system with this anisotropy while possessing scaling symmetry is known as Lifshitz-like geometry,
\begin{equation}
\bar{x} \rightarrow \lambda \bar{ x} \qquad t =\lambda^z t.
\end{equation}
where $\bar{x}$ is spatial coordinate, t is temporal coordinate and $z >1$ is known as the dynamical critical exponent.
Several efforts are being made to realize this type of geometry (Lifshitz-like) in gravity set-up. The most common way is working with Einstein-Maxwell-Dilaton (EMD) theories in the gravity system. This geometry was first introduced by \cite {Kachru:2008yh} and found wide range of application in analyzing thermodynamical and hydrodynamical aspects of strongly coupled systems \cite{Liu:2009dm,Bertoldi:2010ca,Berglund:2011cp,Iizuka:2011hg,Gouteraux:2011ce,Alishahiha:2012cm,Dey:2012tg,Alishahiha:2012qu,Gouteraux:2014hca}.
Generalized system with warped metric is also used for the detailed studies of realistic condensed matter system \cite{Cremonini:2016avj,Charmousis:2010zz,Ammon:2012je,Kuang:2014pna,Karch:2014mba,Kuang:2015mlf,Wu:2015ajt,Li:2016rcv}.
\begin{equation}
ds^2_{d+2}=r^{\frac{2\theta}{d}}\left(-r^{2z}dt^2+\frac{dr^2}{r^2}+r^2\sum_{i=1}^d dx_i^2\right),
\end{equation}
where $\theta$ is known as hyperscaling violating parameter in d-dimensions.
In this work, we investigate the DC conductivity and (thermo)electric conductivity of (2+1) dimensional systems with the hyperscaling violating term for Lifshitz-like geometry. Hence, we consider two different gauge fields in gravity set-up, one field will introduce Lifshitz-like geometry while other introduces finite charge density. Linear axionic fields have been introduced in the system to break translational invariance and obtain finite DC conductivity. Introducing an external magnetic field \cite{Donos:2015bxe, Kim:2015wba, Blake:2015ina, Amoretti:2015gna} has given us an edge for detailed study of the holographic system and is the main motivation of our present work. We discuss the dependence of transport on the dynamical exponents in the presence of external magnetic field and multiple gauge fields. To read transport coefficients for boundary theory using correlation function, one can use various approaches \cite{Papadimitriou:2004rz,Iqbal:2008by,Kuperstein:2011fn,Heemskerk:2010hk,Faulkner:2010jy,Kuperstein:2013hqa,Amoretti:2014zha,Ge:2016lyn,Tian:2017vfk,Hartnoll:2007ai,Hartnoll:2007ih, Blake:2014yla,Ge:2016sel,Amoretti:2017xto,Cremonini:2017qwq,Amoretti:2016cad}.
In this work, we have simplified the calculation while using Wilsonian RG(renormalization group) flow approach\cite{Matsuo:2011fk}. The advantage of this approach is that second order coupled differential equations reduces to first order ordinary differential equations \cite{Kuperstein:2011fn, Kuperstein:2013hqa, Tian:2017vfk}. We have studied these RG flow equations and extract the transport coefficients of the boundary theory.
The paper is organized as follows. We introduce basic holographic set-up of Lifshitz-like geometry with hyperscaling violation term in the next section. We use Wilsonian RG flow approach to study DC conductivity and (thermo)electric conductivity for the system in the presence of the external magnetic field in the following section. We discuss the dependence of transport coefficients on magnetic field and strength of momentum relaxation by various plots. The temperature dependence of various counductivities is also investigated. The concluding section contains the detailed discussion and summary of the work done.
\section{EMD system}
Let us consider the EMD system with two gauge fields and axionic fields. The gravity action is given by,
\begin{equation}
S=\int d^4 x \sqrt{-g}\left(R-\frac{1}{2}(\partial \phi)^2+V(\phi)-\frac{1}{4}\sum_{i=1}^{2}e^{\lambda_i \phi}F^2_i -e^{\lambda_3 \phi}\sum_{i=1}^{2}\partial \chi_i^2\right), \label{eq:action}
\end{equation}
where $F_1$ and $ F_2$ are two $U(1)$ gauge fields, the role of first gauge field is to break Lorentz-invariance and introduce Lifshitz-like geometry while second gauge field introduced the charge in the gravitational set-up. $V(\phi)$ is dilaton potential and $\chi_i$ are the axionic fields. In this work, we take the potential of form, $V(\phi)=-2 \Lambda e^{\lambda_0 \phi}$ for further calculation.
The Einstein equation from the above action is given by,
\begin{equation}
R_{\mu \nu}=\frac{1}{2}\partial _{\mu}\phi \partial_{\nu}\phi +\Lambda e^{\lambda_0 \phi}g_{\mu \nu}+\frac{1}{2}e^{\lambda_3 \phi}(\partial_\mu \chi_i \partial_\nu \chi_i)+\sum_{i=1}^2 \frac{1}{2}e^{\lambda_i \phi}(F_{i\mu \rho}F^\rho_{i ~\nu}-\frac{1}{4}F_i^2g_{\mu\nu}).
\end{equation}
The matter fields equations of motion are obtained as,
\begin{eqnarray}
\square\phi= \frac{1}{2}\lambda_3\sum_{i=1}^2(\partial \chi_i)+\frac{1}{4} \sum_{i=1}^2 \lambda_i e^{\lambda_i \phi}F_i^{2} + 2\lambda_0 \Lambda e^{\lambda_0 \phi}, \\
0=\nabla_\mu\left(e^{\lambda_1 \phi} F^{\mu \nu}_1 \right),\quad
0=\nabla_\mu\left(e^{\lambda_2 \phi} F^{\mu \nu}_2 \right), \label{eq:pa1}\\
0=\nabla_\mu\left(e^{\lambda_3 \phi} \nabla^\mu \chi_i\right).
\end{eqnarray}
The metric ansatz for the above action (\ref{eq:action}) is given by Lifshitz-like, hyperscaling violating black-brane solution as in \cite{Gouteraux:2014hca,Alishahiha:2012qu,Li:2016rcv,Cremonini:2016avj},
\begin{equation}
ds^2=r^{\theta}\left(-r^{2z}f(r) dt^2 +\frac{dr^2}{r^2 f(r)}+r^2(dx^2+dy^2)\right).
\end{equation}
The external magnetic field is introduced in the set-up in the given form,
\begin{equation}
A_{2}=A_2(r)dt+Bxdy.
\end{equation}
The axion fields are linear in spatial direction given by, $\chi_1=\alpha x$ and $\chi_2=\alpha y$ where $\alpha$ is considered as the strength of the momentum dissipation.\\
Using the gravity solution, the parameters of given model are related by, \cite{Cremonini:2016avj}
\begin{eqnarray}
\gamma=\sqrt{(\theta+2)(\theta+2 z-2)}, \qquad \lambda_0=\frac{-\theta}{\gamma},\qquad \lambda_1=\frac{-(4+\theta)}{\gamma}, \\
\lambda_2=\frac{(\theta+2 z-2)}{\gamma}, \qquad \lambda_3=\frac{-\gamma}{\theta+2}, \qquad q_1=\sqrt{2(z-1)(\theta+z+2)},
\end{eqnarray}
with $\Lambda=\frac{-1}{2}(\theta+z+1)(\theta+z+2)$ and $\phi =\gamma \log r$.
From the temporal component of the gauge field equation (\ref{eq:pa1}) we obtain,
\begin{equation}
J_i^t=q_i= r^{-z+3}e^{\lambda_i \phi}(A_i)'_t,
\end{equation}
considering $q_i$ as the charges of two gauge fields. Here the role of $q_1$ is the used to introduce Lifshitz-scaling whereas $q_2$ is interpreted as the black hole charge.\\
The black hole factor with an external magnetic field (B), mass (m) and charge $(q_2)$ along with axionic strength($\alpha$) is given by,
\begin{equation}
f(r)=1-\frac{m}{r^{\theta+z+2}}+\frac{q_2^2+B^2}{2(\theta+2)(\theta+z)r^{2(\theta +z+1)}}+\frac{\alpha^2}{(\theta+2)(z-2)r^{\theta+2z}}. \label{eq:fs}
\end{equation}
The Hawking temperature of black hole is obtained from the expression given below,
\begin{equation}
T=\frac{r_h^{z+1}f'(r_h)}{4 \pi}, \label{eq:HT}.
\end{equation}
Thus,
\begin{equation}
T=\frac{z+2+\theta}{4 \pi}r_h^z -\frac{q_2^2}{8 \pi (2+\theta)}\frac{1}{r_h^{z+2+2 \theta}}-\frac{\alpha^2}{4 \pi (2+ \theta)}\frac{1}{r_h^{z+\theta}}. \label{eq:T}
\end{equation}
The constraint from the gravity solution is that every point in space-time follows the null energy condition (NEC) i.e.,
$
T_{\mu \nu}V^{\mu}V^{\nu} \geq 0,
$
where $V^{\mu}$ is the light like vector. Thus, the allowed values of `z' and `$\theta$' consistent with the gravity solution are \cite{Cremonini:2016avj},
\begin{eqnarray}
(2+\theta)(2z-2+\theta) \geq 0, \\
(z-1)(2+z+\theta) \geq 0.
\end{eqnarray}
Later, we shall see that the consistency of coupled equations demand that $\theta=z-1$ and both the conditions reqire $z\geq 1$. Further at $z=2$, the solution of black hole factor ($\ref{eq:fs}$) is not valid as the last term changes sign. So we shall consider the range $1\leq z <2$, which corresponds to $0\leq\theta<1$.
To study the response of system, we introduce the following perturbations,
\begin{eqnarray}
\delta A_{1i}(t, r)&=& \int_{-\infty}^{\infty} \frac{ d\omega}{2 \pi}a_{1i}(r)e^{-i \omega t}, \\
\delta A_{2i}(t, r)&=& \int_{-\infty}^{\infty} \frac{ d\omega}{2 \pi}a_{2i}(r)e^{-i \omega t},\\
\delta \chi_i(t, r)&=&\int_{-\infty}^{\infty} \frac{ d\omega}{2 \pi}b_i(r)e^{-i\omega t},\\
\delta g_{ti}(t, r)&=&\int_{-\infty}^{\infty} \frac{ d\omega}{2 \pi}r^{\theta+2} h_{ti}(r)e^{-i \omega t}.
\end{eqnarray}
The coupled linearized equations of motion for the fields are obtained as,
\begin{eqnarray}
\left.\begin{aligned}
0&=(r^{z-3-\theta}f a'_{1i})'+\frac{\omega^2 a_{1i}}{r^{5+z+\theta}f}+q_1h'_{ti}, \label{eq:gauge1}\\
0&=(r^{3z-1+\theta}fa'_{2i})'+ \frac{\omega^2 a_{2i}}{r^{3-z-\theta}f}+q_2h'_{ti}+\epsilon_{ij}\frac{i\omega B h_{tj}}{f r^{3-z-\theta}}\label{eq:eq2} \label{eq:gauge2}, \\
0&=(r^{5-z}fb'_i)'+\frac{\omega^2b_i}{f r^{3z-3}}-\frac{i\omega \alpha h_{ti}}{f r^{3z-3}}. \\
\end{aligned}\right.
\end{eqnarray}
where $\epsilon_{ij}$ is the Levi-Civita tensor. The constraint equation of the set-up is given by,
\begin{equation}
0=\omega r^{5-z+\theta}h'_{ti}+\omega q_1 a_{1i}+\omega q_2 a_{2i}+i \alpha r^{5-z}f b_i'-q_2 B h_{ti}-Bfa'_{z_i}r^{3z-1+\theta}. \label{eq:cons}
\end{equation}
whereas metric perturbation equation is obtained in the given form,
\begin{eqnarray}
0=(r^{5-z+\theta} h'_{ti})'-q_1 r^{z-1-\theta}a'_{1i}-q_2 r^{z-1-\theta}a'_{2i} +\frac{(\alpha^2r^{-\theta-2z+2}+B^2 r^{2z-4})h_{ti} }{f} \nonumber \\+\frac{i \alpha \omega r^{-\theta-2z+2}b_i}{f} +\epsilon_{ij}\frac{i\omega B r^{2z-4}a_{2i}}{f}. \label{eq:metric}
\end{eqnarray}
\section{Holographic Approach}
To study the transport properties of (2+1) dimensional boundary system, one has to solve coupled equations of motion (\ref{eq:gauge1}), (\ref{eq:cons}) and (\ref{eq:metric}) using the standard procedure as mentioned in \cite{Blake:2014yla,Kim:2015wba}. However in this work we follow the approach introduced by \cite{Matsuo:2011fk} and developed through various studies\cite{Kuperstein:2011fn,Kuperstein:2013hqa,Tian:2017vfk}. Here, DC conductivities are obtained from first order RG flow equations in the near horizon limit. To study the conductivity of a system we simply apply Ohm's Law as given below.
\begin{equation}
J=\sigma E.
\end{equation}
Applying holographic techniques to obtain transport coefficients we use the Onsager relation ($J=\tau X$) in the matrix form as shown,
\begin{equation}
\begin{pmatrix} J_{1i} & J_{1j} \\ J_{2i} & J_{2j}\end{pmatrix} =\tau \begin{pmatrix}X_{1i}& X_{1j}\\ X_{2i} & X_{2j} \end{pmatrix},
\end{equation}
where `$X_i$' are the linear independent sources and `$J_i$' are the responses of system. $\tau$ matrix are the coefficients evaluated in the near horizon limit. The detailed discussion and application of the formalism is presented in \cite{Matsuo:2011fk,Ge:2016sel,Tian:2017vfk}. Also, we could use the following notation
\begin{equation}
\begin{Vmatrix}
J_1\\J_2
\end{Vmatrix}=\tau \begin{Vmatrix}X_1 \\X_2 \end{Vmatrix}
\end{equation}
And $\tau =J X^{-1}$ can be expressed as,
\begin{equation}
\tau= \begin{Vmatrix} J_1\\ J_1 \end{Vmatrix} \begin{Vmatrix}X_1 \\X_2 \end{Vmatrix}^{-1}
\end{equation}
The linearized equations of motion (\ref{eq:gauge2}) and (\ref{eq:metric}) can be put in the matrix form as,
\begin{equation}
\tau =
\begin{Vmatrix}
-r^{z-3-\theta} f a_{1i'}\\-r^{3z-1+\theta} fa_{2i}' \\ -r^{5-z}f b_i'\\-r^{5-z+\theta}h_{ti}' \end{Vmatrix} \begin{Vmatrix}
i \omega a_{1i}\\ i \omega a_{2i}\\ i \omega b_{i}\\ i \omega h_{ti}
\end{Vmatrix}^{-1} \label{eq:tau}
\end{equation}
Now, taking the radial derivative and substituting the equations of motion we get,
\begin{eqnarray}
\tau'=
\begin{Vmatrix}
\frac{\omega^2 a_{1i}}{r^{5+z+\theta}f}+q_1 h'_{ti}\\
\frac{\omega^2 a_{2i}}{r^{3-z-\theta}f}+q_2 h'_{ti}+\epsilon_{ij}\frac{\omega B h_{tj}}{f r^{3-z-\theta}}\\
\frac{\omega^2b_i}{f r^{3z-3}}-\frac{i\omega \alpha h_{ti}}{f r^{3z-3}}\\
q_1 r^{z-1-\theta}a'_{1i}+q_2 r^{z-1-\theta}a'_{2i} +\frac{\omega B r^{2z-4} \epsilon_{ij}a_{2j}}{f}-\frac{i \alpha \omega r^{-\theta-2z+2}b_i}{f}-\frac{(\alpha^2r^{-\theta-2z+2}+B^2 r^{2z-4})h_{ti}}{f}
\end{Vmatrix}
\begin{Vmatrix}
i \omega a_{1i}\\ i \omega a_{2i}\\ i \omega b_{i}\\ i \omega h_{ti}
\end{Vmatrix}^{-1}\nonumber \\ -\tau
\begin{pmatrix}
\frac{-i \omega}{r^{z-3-\theta}f}&0&0&0\\
0&\frac{-i \omega}{r^{3z-1+\theta}f}&0&0\\
0&0&\frac{-i \omega}{r^{5-z}f}&0\\
0&0&0&\frac{-i \omega}{r^{5-z+\theta}}
\end{pmatrix}
\tau \nonumber
\end{eqnarray}
Further simplifying we get,
\begin{eqnarray}
\tau'=
\begin{pmatrix}
\frac{-i \omega}{r^{z+5+\theta}f} &0&0&0\\
0&\frac{-i \omega}{r^{3-z-\theta}f} &0&\frac{i B}{ r^{3-z-\theta}f}\\
0&0&\frac{-i\omega}{ r^{3z-3}f}&-\frac{\alpha}{ r^{3z-3}f} \nonumber \\
0&\frac{-i B r^{2z-4}}{f}&\frac{-\alpha r^{-\theta-2z+2}}{f}&-\frac{(B^2 r^{2z-4}+\alpha^2r^{-\theta-2z+2})}{i \omega f}\\
\end{pmatrix}\nonumber \\+
\begin{pmatrix}
\frac{-q_1}{r^{5-z+\theta}} &0&0&0\\
0&0&0&\frac{-q_2}{r^{5-z+\theta}}\\
0& 0& 0& 0 \\
\frac{-q_1}{r^{-2}f}&\frac{-q_2}{r^{2z+2\theta}f}& 0 & 0\\
\end{pmatrix}
\tau -\tau
\begin{pmatrix}
\frac{-i \omega}{r^{z-3-\theta}f}&0&0&0\\
0&\frac{-i \omega}{r^{3z-1+\theta}f}&0&0\\
0&0&\frac{-i \omega}{r^{5-z}f}&0\\
0&0&0&\frac{-i \omega}{r^{5-z+\theta}}\\
\end{pmatrix}
\tau
\end{eqnarray}
Multiplying the above metric by black hole factor `f(r)' and taking its near horizon limit, where $f(r_h)=0$ we obtain $\tau_h$.\\
Considering the constraint equation (\ref{eq:cons}) in the matrix form as,
\begin{equation}
\begin{pmatrix} 0& 0&0 &0\\ 0& i B & \alpha & -i \omega \\ 0& 0&0 &0\\ 0& 0&0 &0 \end{pmatrix} \tau= \begin{pmatrix} 0& 0&0 & 0 \\-q_1 & -q_2 & 0 & \frac{-q_2 B}{\omega} \\ 0& 0&0 &0\\ 0& 0&0 &0 \end{pmatrix}
\end{equation}
we obtain,
\begin{eqnarray}
iB \tau_{11}+\alpha \tau_{21}-i \omega\tau_{31}&=&-q_1, \\
iB \tau_{12}+\alpha \tau_{22}-i \omega\tau_{32}&=&-q_2, \\
iB \tau_{13}+\alpha \tau_{23}-i \omega\tau_{33}&=&0, \\
iB \tau_{14}+\alpha \tau_{24}-i \omega\tau_{34}&=&-\frac{q_2 B}{\omega}.
\end{eqnarray}
To maintain the consistency of the equations, we fixed $\theta =z-1$ and the $\tau_h$ matrix is given as,
\begin{equation}
\tau_h =
\begin{pmatrix}
r_h^{-3-z}&0&0&0\\
0&r_h^{-3+3z}& 0& \frac{-B r_h^{-3+3z}}{\omega}\\
0& 0& r_h^{-2z+4} & \frac{\alpha r_h^{-2z+4}}{i \omega}\\
\frac{q_1}{i \omega}&\frac{i Br_h^{-3+2z}+q_2}{i \omega}&\frac{r_h^{4-2z}}{i \omega}&-\frac{B^2 r_h^{3z-3}+\alpha^2 r_h^{4-2z}-iq_2B}{\omega^2}\\
\end{pmatrix}
\end{equation}
By substituting $\tau_h$ in equation (\ref{eq:tau}), we find the flow equation in the near horizon limit,
\begin{eqnarray}
\left.\begin{aligned}
(-r_h^{-2}fa_{1i})'&=r_h^{-3-z}i \omega a_{1i}\label{eq:f1},\\
(-r_h^{4z-2}fa_{2i})'&=r_h^{-3+3z}i \omega a_{2i}+\epsilon_{ij}iBr_h^{-3+3z}h_{tj},\\
(-r_h^{5-z}f b_i)'&=r_h^{-2z+4}i\omega b_i+\alpha r_h^{-2z+4}h_{ti} \label{eq:f1},\\
(-r_h^{4}h_{ti})'&=(iBr_h^{-3+3z}+q_2)a_{2i}+r_h^{4-2z}\alpha b_i+q_1a_{1i}-\frac{(iB^2 r_h^{3z-3}+\alpha^2r_h^{4-2z}+q_2B)h_{tj}}{\omega}.
\end{aligned}\right.
\end{eqnarray}
Near horizon limit of equation (\ref{eq:metric}) is given by,
\begin{equation}
(B^2 r_h^{2z-4}+\alpha^2 r_h^{-3z+3})h_{ti}=q_1a_{1i}' +q_2 a_{2i}'+\epsilon_{ij}\omega B a_{2j}r_h^{2z-4}-i \alpha \omega r_h^{-3z+3}b_i.
\end{equation}
Simplifying the above expression while using the flow equations (\ref{eq:f1}) we obtain the expression for the metric perturbation as,
\begin{equation}
h_{ti}|_{r=r_h}=-i \omega\frac{(q_1 r_h^{-z-1})a_{1i} + \epsilon_{ij}(q_2r_h^{-z-1}+i B r_h^{2z-4})a_{2j}}{(B^2 r_h^{2z-4}+\alpha^2 r_h^{-3z+3})-i q_2 Br_h^{-z-1}} -\frac{i \alpha \omega r_h^{-3z+3}b_i}{(B^2 r_h^{2z-4}+\alpha^2r_h^{-3z+3})-i q_2 Br_h^{-z-1}}. \label{eq:mi}
\end{equation}
\subsection{DC conductivity}
Using gauge field equation (\ref{eq:gauge1}), we get the expression for conserved currents for first gauge field as,
\begin{equation}\label{eq:cur1}
J_{1i}=-r^{z-3-\theta} f a_{1i}'-q_1h_{ti}.
\end{equation}
On substituting the equation (\ref{eq:f1}) the expression takes the form,
\begin{equation}
J_{1i}=r_h^{-3-z}i \omega a_{1i}-q_1 h_{ti}.
\end{equation}
Now from equation (\ref{eq:mi}) neglecting the second term (i.e. axion perturbation part) we obtain the DC conductivity using,
\begin{equation}
\sigma_{ij} =\frac{\partial J_i}{\partial E_j},\qquad {\text{where}} \quad E_j=i \omega a_j.
\end{equation}
Thus we obtain,
\begin{equation}
\sigma_{xx}^{11}=\sigma_{yy}^{11}= r_h^{-3-z}+\frac{q_1^2 \left(B^2 r_h^{z-5}+\alpha^2 r_h^{-4z+2}\right)}{(B^2 r_h^{2z-4}+\alpha^2 r_h^{-3z+3})^2+B^2q_2^2 r_h^{-2z-2}}, \label{eq:si1}
\end{equation}
Also, we have some mixed terms for DC conductivity, where charges of both the fields effect the conductivity as shown,
\begin{eqnarray}
\sigma_{xx}^{12}&=&\sigma_{yy}^{12}=\frac{q_1 q_2 \alpha^2 r_h^{-4z+2}}{(B^2 r_h^{2z-4}+\alpha^2 r_h^{-3z+3})^2+B^2q_2^2 r_h^{-2z-2}},\\
\sigma_{xy}^{11}&=&-\sigma_{yx}^{11}=\frac{q_1^2 q_2 B r_h^{-2}}{(B^2 r_h^{2z-4}+\alpha^2 r_h^{-3z+3})^2+B^2q_2^2 r_h^{-2z-2}},\\
\sigma_{xy}^{12}&=&-\sigma_{yx}^{12}=q_1 B\frac{B^2 r_h^{4z-8}+\alpha^2 r_h^{-z-1}+q_2^2 r_h^{-2-2z}}{(B^2 r_h^{2z-4}+\alpha^2 r_h^{-3z+3})^2+B^2q_2^2 r_h^{-2z-2}}.
\end{eqnarray}
Since these expression are quite complex to analyse we numerically study the dependence of conductivities on magnetic field and momentum dissipation term for two different values of the dynamical exponent, $z=1$ and $z=4/3$ in Fig.1 to Fig.4.
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{1_1.jpg}
\includegraphics[width=.45\textwidth]{1_2.jpg}
\caption{Variation of $\sigma_{xx}^{11}$ with B and $\alpha$ for $z=1$(left) and $z=4/3$(right) }
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{A1.jpg}
\includegraphics[width=.45\textwidth]{A2.jpg}
\caption{Variation of $\sigma_{xy}^{11}$ with B and $\alpha$ for $z=1$(left) and $z=4/3$(right) }
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{2_1.jpg}
\includegraphics[width=.45\textwidth]{2_2.jpg}
\caption{Variation of $\sigma_{xx}^{12}$ with B and $\alpha$ for $z=1$(left) and $z=4/3$(right) }
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{3_1.jpg}
\includegraphics[width=.45\textwidth]{3_2.jpg}
\caption{Variation of $\sigma_{xy}^{12}$ with B and $\alpha$ for $z=1$(left) and $z=4/3$(right) }
\end{figure}
Keeping $z = 1$ reduces the geometry to RN-AdS and we observe trivial DC conductivity flow. However, for $z \neq 1$ non-trivial dependence of DC conductivity on magnetic field and momentum relaxation strength is seen. We also observe a discontinuity while considering $z=1$ ($\theta=0$) case, as $\sigma_{xx}^{11}=r_h^{-4}$. On the other hand for RN-AdS black hole the conductivity is given is a constant term given by, $\sigma_{xx}^{11}=1$.
Similarly, for the second gauge field we can evaluate the conductivity accordingly. The conserved current for the boundary theory is obtained using equation (\ref{eq:gauge1}),
\begin{equation}
J_{2i}=-r^{3z-1+\theta} f a'_{2i}-q_2h_{ti},
\end{equation}
On substituting equation (\ref{eq:f1}) we obtain the modified expression for $J_{2i}$ as,
\begin{equation}
J_{2i}=r_h^{-3+3z}i \omega a_{2i}+ \epsilon_{ij}iBr_h^{-3+3z}h_{tj}-q_2h_{ti}.
\end{equation}
Then the DC conductivity is evaluated concerning the second gauge field.
\begin{eqnarray}
\sigma_{xx}^{22}&=&\sigma_{yy}^{22}=\frac{\alpha^2 r_h^{2z-4}[B^2+r_h^{-5z+7}(q_2^2r_h^{-z-1}+\alpha^2)]}{ B^2q_2^2r_h^{-2z-2}+(B^2r_h^{2z-4}+\alpha^2 r_h^{-3z+3})^2}, \label{eq:dc2}\\
\sigma_{xy}^{22}&=&-\sigma_{yx}^{22}=\frac{ q_2 B r_h^{4 z-8} \left[B^2+r_h^{-5 z+7} \left(q_2^2 r_h^{-z-1}+2 \alpha^2\right)\right]}{B^2q_2^2 r_h^{-2 z-2}+\left(B^2 r_h^{2 z-4}+\alpha ^2 r_h^{3-3 z}\right)^2}, \label{eq:dc3}
\end{eqnarray}
The above expressions are complicated to interpret. So we study the dependence of these conductivities on magnetic field and $\alpha$ using plots as shown in Fig.5 to Fig.6.
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{4_1.jpg}
\includegraphics[width=.45\textwidth]{4_2.jpg}
\caption{Variation of $\sigma^{22}_{xx}$ with B and $\alpha$ for $z=1$(left) and $z=4/3$(right) }
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{5_1.jpg}
\includegraphics[width=.45\textwidth]{5_2.jpg}
\caption{Variation of $\sigma^{22}_{xy}$ with B and $\alpha$ for $z=1$(left) and $z=4/3$(right) }
\end{figure}
Also from the plots of different DC conductivities (for both the fields) we observe at fixed momentum relaxation strength conductivity $\sigma^{ij}_{xx}$ shows a monotonic dependence on the external magnetic field whereas in $\sigma^{ij}_{xy}$ this behavior is not seen. \\
Let us consider the limit $B\rightarrow 0$.
\begin{equation}
\sigma^{22}_{xx}= r_h^{3z-3}+r_h^{2z-4}\frac{q_2^2}{\alpha^2},\qquad \sigma^{22}_{xy}=0\label{eq:d}
\end{equation}
From the above expression it is noted that electric conductivity obeys inverse Matthiessen's rule given by,
\begin{equation}
\sigma_{DC}=\sigma_{Q}+\sigma_{D}
\end{equation}
where $\sigma_{Q}$ is the charge conjugation symmetric part and $\sigma_{D}$ is the momentum dissipation part.
The temperature dependence of conductivity is governed by equation ($\ref{eq:T}$) and $T \sim r_h^z$. Here we observe the following scaling in the DC conductivity,
\begin{itemize}
\item[i] For $z=1$, \quad $\sigma^{22}_{xx} \sim 1+\frac{q_2^2}{\alpha^2 T^2}$
\item[ii] For $z=4/3$,\quad $\sigma^{22}_{xx} \sim T^{3/4}+\frac{q_2^2}{T\alpha^2}$
\item[ii] For $z \rightarrow 2$, \quad $\sigma^{22}_{xx} \sim T^{3/2}+\frac{q_2^2}{\alpha^2}$
\end{itemize}
In our holographic model, we are able to capture the low temperature behavior of the DC conductivity obeying Fermi-Liquid law ( $\sigma_{DC} \sim \frac{1}{T^2}$ ) for $z=1$ along with a constant term. This behavior changes to unconventional metallic behavior ($ \sigma_{DC} \sim \frac{1}{T}$) as we increase the Lifshitz scaling to $z=4/3$ and becomes constant in the limiting case $z\rightarrow 2$. Thus, there is non-trivial dependence of conductivity on temperature for hyperscaling range ($1<z<2$).
\subsection{Halls Angle}
We can obtain the expression for Hall angle using equations (\ref{eq:dc2}) and (\ref{eq:dc3})\cite{Kim:2015wba,Ge:2016sel,Amoretti:2017xto}.
Thus,
\begin{equation}
\tan{ \theta_H}=\frac{\sigma^{22}_{xy}}{\sigma^{22}_{xx}}
\end{equation}
\begin{equation}
\theta_H=\frac{B q_2 r_h^{2 z-4} \left[B^2+ r_h^{7-5 z} \left(2 \alpha ^2+q_2^2 r_h^{-z-1}\right)\right]}{\alpha ^2 \left[B^2+r_h^{7-5 z} \left(\alpha ^2+q_2^2 r_h^{-z-1}\right)\right]} \label{eq:ha1}
\end{equation}
From the above expression it is observed that, $\theta_H \propto \frac{B q_2 r_h^{2z-4}}{\alpha^2}$ as the terms in the bracket is consider as a geometric quantity \cite{Blake:2014yla}.
Comparing the result with that of DC conductivity Hall angle consists of only dissipation part $(\sigma_{D})$, unlike DC conductivity which is the combination of two different terms (shown in equation (\ref{eq:d})). This is responsible for the presence of different scaling in strange metals\cite{Andri:1997}. We plot the dependence of Hall angle on the magnetic field applied and the strength of momentum relaxation in Fig.7 for fixed $q_2$.
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{11_1.jpg}
\includegraphics[width=.45\textwidth]{11_2.jpg}
\caption{Variation of $\theta_H$ with B and $\alpha$ for $z=1$(left) and $z=4/3$(right) }
\end{figure}
Let us consider the temperature dependence of Hall angle,
\begin{itemize}
\item[i] For $z=1$, \quad $\theta_H \sim 1/T^2$
\item[ii] For $z=4/3$,\quad $\theta_H \sim 1/T$
\item[ii] For $z \rightarrow 2$, \quad $\theta_H \sim 1/T^0$
\end{itemize}
For $z=1$, we observe the temperature dependence is same as measured in cuprates \cite{Hussey:2008}. However the behavior changes to $\theta_H \sim 1/T$ with the non-trivial scaling $z \neq 4/3$. Further $\theta_H$ reduces to a constant for $z \rightarrow 2$. We show the temperature and magnetic field dependence on the Hall angle from plots given in Fig.8 and Fig.9.
\begin{figure}[h!]
\centering
\includegraphics[width=.38\textwidth]{h1.jpg}
\hspace{1cm}
\includegraphics[width=.38\textwidth]{h2.jpg}
\caption{$\theta$ vs T at $\alpha$= 0.1(blue),0.5(red),1(green) for $z=1$(left) and $z=4/3$(right)}
\end{figure}
\begin{figure}[h!]
\centering{
\includegraphics[width=.38\textwidth]{ha4.jpg}
\hspace{1cm}
\includegraphics[width=.38\textwidth]{h5.jpg}
\caption{$\theta$ vs B at $T=$ 0.5(blue),1(red),1.5(green) for $z=1$(left) and $z=4/3$(right) }}
\end{figure}
\subsection{Thermoelectric conductivity}
Next, we evaluate thermoelectric conductivity using heat current expression.
\begin {eqnarray}
Q_i=-4\pi T g_{xx}h_{ti}, \quad {\text{and}} \quad \alpha_{ij} =\frac{\partial Q_i}{T \partial E_j}
\end{eqnarray}
We get the following expression for thermoelectric conductivity depending on external magnetic field and dynamical exponent for our model as shown in Fig. 10 and Fig. 11.
\begin{eqnarray}
\alpha^{22}_{xx}=\alpha^{22}_{yy}=\frac{4 \pi r_h^{-3z+3}q_2 \alpha^2}{(B^2r_h^{2z-4}+\alpha^2 r_h^{-3z+3})^2+B^2q_2^2r_h^{-2z-2}},
\end{eqnarray}
\begin{eqnarray}
\alpha^{22}_{xy}=-\alpha^{22}_{yx}=\frac{4 \pi B r_h^{z+1}(B^2 r_h^{4z-8}+\alpha^2 r_h^{-z-1}+q_2^2 r_h^{-2-2z})}{(B^2r_h^{2z-4}+\alpha^2 r_h^{-3z+3})^2+B^2q_2^2r_h^{-2z-2}}.
\end{eqnarray}
In the limiting case $B\rightarrow 0$, we obtain,
\begin{equation}
\alpha^{22}_{xx}=\frac{4 \pi q_2 r_h^{3z-3}}{\alpha^2}, \qquad \alpha^{22}_{xy}=0 \label{eq:c1}
\end{equation}
Thus, the thermoelectric conductivity also shows non-trivial dependence on hyperscaling and unconventional temperature dependence.
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{6_1.jpg}
\includegraphics[width=.45\textwidth]{6_2.jpg}
\caption{Variation of $\alpha^{22}_{xx}$ with B and $\alpha$ for $z=1$(left) and $z=4/3$(right) }
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{7_1.jpg}
\includegraphics[width=.45\textwidth]{7_2.jpg}
\caption{Variation of $\alpha^{22}_{xy}$ with B and $\alpha$ for $z=1$(left) and $z=4/3$(right) }
\end{figure}
\subsection{Seebeck Coefficient}
The generation of transverse electric field in the system is given by the thermoelectric power (Seebeck coefficient). Using the results of the conductivity we obtain,
\begin{equation}
S=\frac{\alpha^{22}_{xx}}{\sigma^{22}_{xx}}=\frac{4 \pi r_h^{-5z+7}q_2}{[B^2 + r_h ^{-5z+7}(q_2^2 r_h^{-z-1}+\alpha^2)]} \cdot
\end{equation}
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{12_1.jpg}
\includegraphics[width=.45\textwidth]{12_2.jpg}
\caption{Variation of S with B and $\alpha$ for $z=1$(left) and $z=4/3$(right) }
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=.38\textwidth]{y1.jpg}
\hspace{1cm}
\includegraphics[width=.38\textwidth]{y2.jpg}
\caption{S vs T at $\alpha$=0.5(blue),1(red),1.5(green) for $z=1$(left) and $z=4/3$(right) }
\end{figure}
\begin{figure}[h!]
\centering
{\includegraphics[width=.38\textwidth]{y4.jpg}
\hspace{1cm}
\includegraphics[width=.38\textwidth]{y5.jpg}}
\caption{S vs B at $T$=0.5(blue),1(red),1.5(green) for $z=1$(left) and $z=4/3$(right) }
\end{figure}
The variation of Seebeck coefficient with applied magnetic field and momentum relaxation strength is shown in Fig. 12. The result from experiments suggest that at high temperature Seebeck coefficient remains constant. The dependence of Seebeck coefficient on model parameters is shown in Fig.13 and Fig.14. We observe at different temperature the behavior of Seebeck coefficients does not change appreciably for non-trivial scaling. The temperature scaling of the coefficient is still unclear from experimental results\cite{Kim:2004}.
\subsection{Thermal conductivity}
Using the results for thermoelectric and DC conductivity, we can also obtain the thermal conductivity \cite{Tian:2017vfk}.
The thermal conductivity for non-zero magnetic field can be obtained using the relation given below,
\begin{equation}
\begin{pmatrix} <J_{i}> \\ <Q_{i}> \end{pmatrix} = \begin{pmatrix}\sigma_{ij}& \alpha_{ij}T\\ \bar{\alpha}_{ii}T & \bar{\kappa}_{ij}T \end{pmatrix} \begin{pmatrix}E_j \\ -(\nabla_j T)/T\end{pmatrix}
\end{equation}
Considering the thermal current $Q_x=0$ and $E_y=0$ we obtain the expression for the thermal conductivity using,
\begin{equation}
\bar{\kappa}^{22}_{xx}=\frac{T (\alpha^{22}_{xx})^2}{\sigma^{22}_{xx}-\sigma^{22}_{xx}(0)}, \quad {\text{and}} \quad \bar{\kappa}^{22}_{xy}=\frac{T \alpha^{22}_{xx} \alpha^{22}_{xy}}{\sigma^{22}_{xx}}
\end{equation}
where $\sigma^{22}_{xx}(0)$ is the electric conductivity for $Q_x=0$ (vanishing heat currents).
Thus, we obtain,
\begin{equation}
\bar{\kappa}^{22}_{xx}=\bar{\kappa}^{22}_{yy}=\frac{16 \pi^2 r_h^{5 + 4 z} T (B^2 r_h^{5 z}+ r_h^7 \alpha^2)}{
B^4 r_h^{10 z} + r_h^{14} \alpha^4 +
B^2 r_h^{6 + 4 z} (q_2^2 + 2 r_h^{1 + z}\alpha^2)}
\end{equation}
\begin{equation}
\bar{\kappa}^{22}_{xy}=-\bar{\kappa}^{22}_{yx}=\frac{16 \pi ^2 T B q_2 r_h^{8+6z}}{B^4 r_h^{10z}+r_h^{14}\alpha^2 +B^2 r_h^{6+4z}(q_2^2+2 r_h^{1+z}\alpha^2)}
\end{equation}
Variation of thermal conductivity with different model parameters is shown in Fig.15 and Fig.16
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{8_1.jpg}
\includegraphics[width=.45\textwidth]{8_2.jpg}
\caption{Variation of $\bar{\kappa}^{22}_{xx}$ with B and $\alpha$ for $z=1$(left) and $z=4/3$(right) }
\end{figure}
which also shows non-trivial dependence on hyperscaling.
Let us discuss the limiting case $B\rightarrow 0$,
\begin{equation}
\kappa^{22}_{xx}=\frac{16 \pi ^2 T r_h^{4 z-2}}{\alpha ^2}, \qquad \kappa^{22}_{xy}=0,
\end{equation}
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{9_1.jpg}
\includegraphics[width=.45\textwidth]{9_2.jpg}
\caption{Variation of $\bar{\kappa}^{22}_{xy}$ with B and $\alpha$ for $z=1$(left) and $z=4/3$(right) }
\end{figure}
\subsection{Lorenz ratio}
To complete the discussion, we check the expression for Lorenz ratio.
First we obtain the Hall Lorentz ratio using \cite{Ge:2016sel},
\begin{equation}\label{eq:L1}
L=\frac{ \bar{\kappa}_{xy}}{T \sigma_{xy}}=\frac{16 \pi ^2 r_h^{2 z+8}}{B^2 r_h^{6 z}+q_2^2 r_h^6+2 \alpha ^2 r_h^{z+7}},
\end{equation}
The expression for the Lorenz ratio is,
\begin{equation}
\bar{L}=\frac{ \bar{\kappa}_{xx}}{T \sigma_{xx}}=\frac{16 \pi ^2 r_h^{2 z+1} \left(B^2 r_h^{5 z}+\alpha ^2 r_h^7\right)}{\alpha ^2 \left(B^2 r_h^{6 z}+q_2^2 r_h^6+\alpha ^2 r_h^{z+7}\right)}.
\end{equation}
The explicit dependence of Lorenz ratio on the dynamical scaling and momentum relaxation strength is shown in Fig.17.
\begin{figure}[h!]
\includegraphics[width=.45\textwidth]{10_1.jpg}
\includegraphics[width=.45\textwidth]{10_2.jpg}
\caption{Variation of L with B and $\alpha$ for $z=1$(left) and $z=4/3$(right) }
\end{figure}
For $B \rightarrow 0$,
\begin{equation}
\bar{L}=\frac{16 \pi ^2 r_h^{2 z+2}}{q_2^2+\alpha ^2 r_h^{z+1}}.
\end{equation}
We obtain the Lorenz ratio ratio at zero temperature for vanishing magnetic field keeping $z=1$ as,
\begin{equation}\label{eq:L0}
\bar{L}=\frac{\bar{\kappa}_{xx}}{T \sigma_{xx}}|_{T, B \rightarrow 0}= \frac{4}{3} \pi ^2 \left(1+\frac{\alpha ^2}{\sqrt{\alpha ^4+12 q_2^2}}\right)
\end{equation}
The given expression indicates, at $B=0$ the WF law is valid and we obtain Fermi-liquid type ground state for $z=1$.
The temperature dependence of Lorenz ratio is shown in Fig.18 for different momentum relaxation strength. According to Wiedemann-Franz (WF) law the Lorenz ratio is constant for normal metals. Since our results show explicit dependence on temperature and also on Lifshitz scaling, the WF Law is violated in this model. The temperature dependence can not be extracted in a simple manner because of the interplay of different parameters in the system. We also plot the variation of the Lorenz ratio with the magnetic field in Fig.19.
\begin{figure}[h!]
\centering
\includegraphics[width=.38\textwidth]{m1.jpg}
\hspace{1cm}
\includegraphics[width=.38\textwidth]{m2.jpg}
\caption{L vs T at $\alpha$=0.1(blue),0.5(red),1(green) for $z=1$(left) and $z=4/3$(right) }
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=.38\textwidth]{m4.jpg}
\hspace{1cm}
\includegraphics[width=.38\textwidth]{m5.jpg}
\caption{L vs B at $T=$0.5(blue),1(red),1.5(green) for $z=1$(left) and $z=4/3$(right) }
\end{figure}
\section{Conclusions and Summary}
In this paper, we have investigated the DC transport of holographic systems with Lifshitz-like geometry and hyperscaling violation. The geometry is dual to non-relativistic ($z \neq 1$) condensed matter systems under the applied external magnetic field. We considered the near horizon limit of linearized equations of motion and calculated DC conductivity, thermoelectric and thermal conductivity analytically.
We introduced the perturbations in both the gauge fields and obtained the expressions for the transport coefficients. The behavior of transport coefficients is depicted by numerical plots for different values, $z=1$ and $z=4/3$, of the dynamical exponent. While we have non-trivial scaling for $z=4/3$, the geometry reduces to RN-AdS black hole for $z=1$. Dependence of DC transport on magnetic field and momentum relaxation strength is also studied while considering different Lifshitz scaling. The pattern of different type of conductivities ( $\sigma^{ij}_{xx} , \alpha^{ij}_{xx}$ and $\bar{\kappa}^{ij}_{xx}$) are quite similar showing monotonic dependence on the magnetic field but this behavior is absent in $\sigma^{ij}_{xy}$ etc .
We also obtained the Hall angle, Seebeck coefficient and Lorenz ratio for the system and plotted them as a function of temperature and magnetic field. Hall angle shows $1/T^2$ in temperature behavior as given in equation (\ref{eq:ha1}) for $z=1$ but it changes to $1/T$ for $z=4/3$. The Wiedemann-Franz law is violated for our holographic model depicting unconventional metallic behavior. However, at zero temperature and $B \rightarrow 0$ the Fermi-liquid behavior is obtained for $z=1$. Seebeck coefficient and Lorenz ratio also showed non-trivial dependence on the hyperscaling parameter. This unconventional dependence of transport coefficient on temperature can be useful to study the strange metal phenomenon \cite{Blake:2014yla,Amoretti:2016cad}.
|
\section{Introduction}
\renewcommand\[{\begin{equation}}
\renewcommand\]{\end{equation}}
In modern evidence-based medicine, decisions on a diagnosis or personalized
treatment plan are often guided by risk scores generated from prognostic
models \cite{baskin2007recipient,hernandez2009novel,lorent2016mortality}.
Such prognostic risk scores can be either a single risk factor, such as a biomarker, or a risk probability calculated from multiple
risk factors. For a risk score to be utilized in clinical practice,
its predictive accuracy is often assessed through two types of metrics:
(1) the discrimination metric, which measures how well the risk score
can distinguish subjects with and without the disease condition,
and (2) the calibration metric, which measures how well the predicted
risk matches the observed risk in the target population. Motivated
by the prediction of end-stage renal disease (ESRD) among a cohort of patients with chronic kidney disease, the goal
of this paper is to propose a framework to estimate the predictive
accuracy of a risk score from a prognostic model, accounting
for right censoring and competing events.
For a continuous time-to-event outcome, the presence and absence
of a disease condition at any time point $\tau$ can be viewed as a binary outcome. To study the relationship between a continuous risk
score and this binary outcome at any prespecified time point $\tau$, the
time-dependent receiver operating characteristics (ROC) curve is widely used for assessing discrimination, i.e., the separation of subjects with and without a given disease at time $\tau$ by the risk score \cite{heagerty2000time}. For example, the risk score is the $\tau$-year (e.g., $\tau=5$) survival probability
calculated based on the characteristics of a cancer patient at initial
diagnosis, and the disease presence or absence is defined by whether
the patient died of cancer within $\tau$ years after the initial
diagnosis. For such a risk score, the area under the ROC curve (AUC)
presents the probability that a subject with the disease at time $\tau$ has a higher predicted risk score than a subject without the disease. A challenge of estimating such time-dependent ROC curve is that the disease
status at $\tau$ is unknown among subjects who are censored prior to $\tau$. A number of methods
have been developed
to address this issue, including the nearest neighboring estimator (NNE)
\cite{heagerty2000time} and inverse probability censoring weighting
(IPCW) \cite{blanche2013review,chiang2010non,uno2007evaluating}.
In addition to the metrics for discrimination, metrics for calibration \cite{graf1999assessment} quantify the absolute
deviance of the risk score from the observed outcome, known as the prediction error. Time-dependent prediction error
metrics for survival outcomes have been proposed \cite{graf1999assessment,gerds2006consistent,korn1990measures,schemper2000predictive}.
The prediction error can be constructed through a class of
loss functions that link the risk score and the binary disease outcome
at $\tau$ \cite{graf1999assessment}. Among those, the quadratic loss, known as the Brier score
\cite{brier1950verification}, is a popular choice \cite{blanche2015quantifying,cortese2013comparing,parast2012landmark}. Censoring remains a challenge when estimating the Brier score, and an IPCW method
was proposed to deal with it \cite{graf1999assessment,gerds2006consistent}.
Competing risks are common in clinical research that involves time-to-event data. For example, in a cardiovascular study, one may be interested in the time to the first myocardial infarction after cardiovascular surgery, but patients may die before experiencing the event of interest. Limited statistical methodology is available to estimate the predictive accuracy metrics in the context of competing risks. To estimate the time-dependent
ROC, Saha \& Heagerty \cite{saha2010time} extended the NNE method \cite{heagerty2000time}
to the competing risk context. Zheng et al. \cite{zheng2012evaluating} further
extended the method of Saha \& Heagerty \cite{saha2010time} to covariate-adjusted time-dependent
ROC. Blanche et al. \cite{blanche2013estimating} studied the use of IPCW in estimating the time-dependent ROC with competing risk data.
For the estimation of the Brier score with competing risk data, the available published methods are based on the IPCW \cite{blanche2015quantifying,liu2016robust,schoop2011quantifying}, with the censoring distribution estimated either by the Kaplan-Meier (KM) method without conditioning on the risk score \cite{graf1999assessment} or by the Cox proportional hazards model conditional on the risk score \cite{gerds2006consistent}.
This paper focuses on the time-dependent discrimination and calibration
estimation in the context of competing risk outcomes. We propose a novel
nonparametric kernel-weighted estimation framework for both time-dependent discrimination and calibration measures. The proposed method first estimates the conditional probability of experiencing an event of interest at $\tau$
given the observed data of the subjects. This is done through nonparametric
kernel regression for the cumulative incidence function. Then the
time-dependent predictive accuracy metrics, such as sensitivity, specificity,
and Brier score, are estimated by weighting each subject with their
own conditional probabilities.
The proposed method has some attractive properties. First, it is fully
nonparametric, without any distributional or modeling assumptions.
This is desirable for estimating predictive accuracy metrics since
it reduces the bias from the estimation procedure itself. Second, the
proposed method, unlike other nonparametric methods such as NNE
\cite{heagerty2000time}, is insensitive to the bandwidth choice. This
is shown in this paper with both numerical and methodological
justifications. Third, the method automatically accommodates correlation between
the censoring time and the risk score. Furthermore, the proposed method can be invariant
to monotone transformation of the risk score when the tuning parameter
is specified by the span, the proportion of subjects included in the kernel estimation.
Also, the estimated sensitivity, specificity, and ROC curve are monotone in
the cut-off point $c$.
Our simulation shows that the proposed method has competitive performance
in terms of bias and the mean squared error (MSE) when compared with other published
methods. Section 2 presents the notations and definitions for the time-dependent
ROC and time-dependent prediction error. Section 3 describes the proposed
estimators for the predictive accuracy metrics. Then the finite sample
performance is evaluated by simulations in Section 4. In Section
5, we illustrate the method with data from the African American Study of Kidney
Disease and Hypertension (AASK) in evaluating the prediction
of ESRD. Section 6 concludes the paper by discussing the findings and
providing some perspective.
\section{Predictive Accuracy for Time-to-Event Data with Competing Risks}
\subsection{Notation}
Let $T$ denote the event time, $C$ the censoring time, $\delta$
the event type, and $\Delta=1(T\le C)$ the censoring indicator, where
$1(\cdot)$ is the indicator function. We observe independent and
identically distributed (i.i.d.) samples of $\{(\tilde{T}_{i},U_{i},\tilde{\delta}_{i}),i=1,2,\ldots n\}$ in a validation data set,
where $\tilde{T}_{i}=\textrm{min}(T_{i},C_{i})$ is the observed
time to the event or censoring, whichever comes first. The observed status
$\tilde{\delta}_{i}=\Delta_{i}\delta_{i}$, which equals zero for censored
subjects and equals one of the $K$ possible causes, $\delta_{i}\in\{1,2,\ldots K\}$,
for uncensored subjects. Without loss of generality, we present our
methodology with $K=2$ to match the data application in Section
5. The methodology still applies with other choices of $K\ (K>2)$. For clarity,
suppose that we are interested in assessing the predictive accuracy
of event type $\delta=1$. Let $U_{i}$ denote the risk
score for subject $i$, with higher values of $U_{i}$ indicating higher
risk of the event. For example, $U_{i}$
can be the predicted cumulative incidence probability from a competing
risk regression model that we want to evaluate, i.e., $U_{i}=\pi_{1}(\tau\vert\boldsymbol{Z}_{i})=P(T_{i}\le\tau,\delta_{i}=1\vert\boldsymbol{Z}_{i})$, where $\boldsymbol{Z}$ denotes the predictor and $\tau$ is the predictive horizon. The predictive model is often developed from a training data set that is different from the validation data set.
This paper focuses on estimating the predictive accuracy metrics in a validation data set. We do not study how the model for the risk score $U$ is estimated or whether the model is correctly estimated. We assume that this model has already been developed, needs to be evaluated, and the risk score $U$ has the interpretation of being the subject-specific predicted cumulative incidence probability at horizon $\tau$.
\subsection{Definitions of the time-dependent ROC curve and AUC\label{subsec:tdROC_comprsk}}
In the presence of competing events, the definition of cases is straightforward.
The cases at time $\tau$ for event type $k$ are defined as subjects
who undergo event $\delta=k$ before time $\tau$, i.e., $Case_{k}=\{i:T_{i}\le\tau,\delta_{i}=k\}$.
At a given threshold $c$, the cause-specific sensitivity at time
$\tau$ is defined as
\begin{equation}
Se(c,\tau)=P(U>c\vert T\le\tau,\delta=k).\label{eq:def_sen}
\end{equation}
This is the definition of $cumulative/dynamic$ sensitivity
\cite{heagerty2000time}. When $U$ is higher than the threshold value
$c$, the patient is predicted to experience event $k$ within the time
window $(0,\tau]$.
We consider two definitions of controls that lead
to two different definitions of time-dependent specificity. Saha \& Heagerty \cite{saha2010time} originally
defined the control group at time $\tau$ as the event-free subjects,
i.e., $\{i:T_{i}>\tau\}$. According to this definition, subjects who
experienced competing events other than $k$ are neither cases nor
controls. Therefore, Zheng et al. \cite{zheng2012evaluating} introduced an alternative
definition of the control group $\{i:T_{i}>t\}\cup\{i:T_{i}\le t,\delta_{i}\ne k\}$,
which includes both event-free subjects and subjects who experience other competing
events. We study the estimation under both definitions:
\begin{lyxlist}{00.00.0000}
\item [{\textbf{Definition}}] \textbf{A. }Case $k$: $T\le\tau,\delta=k$;
Control$_{A}$: $(T>\tau)\cup(T\le\tau\cap\delta\ne k).$
\end{lyxlist}
\begin{lyxlist}{00.00.0000}
\item [{\textbf{Definition}}] \textbf{B.} Case $k$: $T\le\tau,\delta=k$;
Control$_{B}$: $T>t$ .
\end{lyxlist}
The specificity at time $\tau$ with respect to the two types of definitions
is
\begin{align}
Sp_{A}(c,\tau) & =P(U\le c\vert\{T>\tau\}\cup\{T\le\tau,\delta\ne k\})\nonumber \\
Sp_{B}(c,\tau) & =P(U\le c\vert T>\tau).\label{eq:def_sp}
\end{align}
Two different time-dependent ROC curves can be obtained by plotting
$Se(c,\tau)$ versus either $1-Sp_{A}(c,\tau)$ or $1-Sp_{B}(c,\tau)$,
i.e., $ROC_{A}(x,\tau)=Se(Sp_{A}^{-1}(1-x,\tau),\tau)$ and $ROC_{B}(x,\tau)=Se(Sp_{B}^{-1}(1-x,\tau),\tau)$
for $x\in[0,1]$. The corresponding $AUC$s can be defined as $AUC(\tau)=\int_{0}^{1}ROC(x,\tau)dx$ or
as the proportion of concordance pairs among the population \cite{blanche2013estimating}:
\begin{align}
AUC_{A}(\tau) & =P(U_{i}>U_{j}\vert T_{i}\le\tau,\delta_{i}=k,\{T_{j}>\tau\}\cup\{T_{j}\le\tau,\delta_{j}\ne k\})\nonumber \\
AUC_{B}(\tau) & =P(U_{i}>U_{j}\vert T_{i}\le\tau,\delta_{i}=k,T_{j}>\tau),\label{eq:AUC def}
\end{align}
where $i$ and $j$ indicate two independent subjects under comparison.
The subjects who experienced the competing events before $\tau$ contribute
to $AUC_{A}(\tau)$ but not $AUC_{B}(\tau)$. The justification for
both definitions is related to the clinical interpretation \cite{zheng2012evaluating}.
\subsection{Definitions of the time-dependent prediction error }
The time-dependent prediction error in the competing risk framework
is defined as the distance between the event-specific status $1(T\le\tau,\delta=k)$
and the subject-specific predicted cumulative incidence function at horizon $\tau$, $\pi_{k}(\tau\vert \boldsymbol{Z})=P(T\le\tau,\delta=k\vert\boldsymbol{Z})$.
Suppose we are interested in evaluating the prediction for event type
1, three types of prediction error measurements can be defined as
follows \cite{van2011dynamic}:
\[
AbsErr(\tau)=E\Big|1\{T\le\tau,\delta=1\}-\pi_{1}(\tau\vert\mathbf{Z})\Big|
\]
\[
Brier(\tau)=E\Big[1\{T\le\tau,\delta=1\}-\pi_{1}(\tau\vert\mathbf{Z})\Big]^{2}
\]
\[
KL(\tau)=-E\Big[1\{T\le\tau,\delta=1\}\cdot\textrm{ln}\pi_{1}(\tau\vert\mathbf{Z})+1\{(T>\tau)\cup(T\le\tau,\delta\ne1)\}\cdot\textrm{ln(}1-\pi_{1}(\tau\vert\mathbf{Z}))\Big].
\]
Among the three measures, $AbsErr(\tau)$ is not \textquotedblleft proper\textquotedblright{}
in the sense that it is not minimized by the predicted cumulative incidence function (CIF) from the true model \cite{graf1999assessment}.
$Brier(\tau)$ is not only \textquotedblleft proper\textquotedblright,
but has the attractive property that it can be decomposed into a term related to the
bias of the predictive survival probability and a term related to the variance of disease status \cite{schoop2011quantifying}.
The Kullback-Leibler score, $KL(\tau)$, has a close connection to the
likelihood ratio test and the Akaike information criteria (AIC), but its
disadvantage is that $KL(\tau)$ goes to infinity when $\pi_{1}(\tau\vert\mathbf{Z})=0$
and $\{T\le\tau,\delta=1\}$, or when $\pi_{1}(\tau\vert\mathbf{Z})=1$
and $\{T>\tau$ or $T\le\tau,\delta\ne1\}$ \cite{van2011dynamic}. The Brier score is more widely used than the other two, and we will
focus on the Brier score for the rest of this paper, even though our methodology also applies to the other two metrics.
\section{The Proposed Nonparametric Weighting Estimators}
Without censoring, sensitivity and specificity can be estimated
empirically as the fraction of true positives and true negatives.
However, when subjects are censored before $\tau$, the true disease
status at $\tau$ is unknown. The empirical fractions can no longer
be used and proper adjustment for censoring is needed. In the context
of right-censored data without competing events, Li et al. \cite{li2016simple} proposed to weigh each
subject by their respective conditional probability of having the
disease at $\tau$ given all the observed data for that subject. The
conditional probability equals 0 if a subject survives beyond $\tau$
without the disease or 1 if the subject acquires the disease prior to $\tau$.
If a subject is censored prior to $\tau$, the conditional probability
is estimated through a nonparametric kernel regression. In this paper,
we extend that approach to the context of competing risk data. The weight
is defined as the conditional probability of being a case prior to
time $\tau$ given the observed time to the event, event status and prognostic
risk score:
\begin{align}
W_{1i} & =P(T_{i}\le\tau,\delta_{i}=1\vert\tilde{T}_{i},\tilde{\delta}_{i},U_{i})\nonumber \\
& =\Big\{1(\tilde{\delta}_{i}=0)\cdot\frac{F_{1}(\tau\vert U_{i})-F_{1}(\tilde{T}_{i}\vert U_{i})}{S(\tilde{T}_{i}\vert U_{i})}+1(\tilde{\delta}_{i}=1)\Big\}\cdot1(\tilde{T}_{i}\le\tau),\label{eq:wt_cmprsk}
\end{align}
where $F_{1}(t\vert U_{i})=P(T_{i}\le t,\delta_{i}=1\vert U_{i})$ is the conditional cumulative incidence
function for event 1, and $S(t\vert U_{i})=P(T_{i}>t \vert U_{i})$ is the conditional overall survival probability.
According to equation (\ref{eq:wt_cmprsk}), we have $W_{1i}=1$ for
subjects with observed event 1 before $\tau$: $\{i:\tilde{T}_{i}\le\tau,\tilde{\delta}_{i}=1\}$;
$W_{1i}=0$ for subjects without any events before $\tau$ or with
competing events before $\tau$: $\Big\{ i:\{\tilde{T}_{i}>\tau\}\cup\{\tilde{T}_{i}\le\tau,\tilde{\delta}_{i}\notin\{0,1\}\}\Big\}$
; and $W_{1i}=\dfrac{F_{1}(\tau\vert U_{i})-F_{1}(\tilde{T}_{i}\vert U_{i})}{S(\tilde{T}_{i}\vert U_{i})}$
for subjects censored before $\tau$: $\{i:\tilde{T}_{i}\le\tau,\tilde{\delta}_{i}=0\}$.
This weighting approach uses the observed status for uncensored subjects and only
imputes the unknown status for censored subjects with a probability.
A heuristic justification is that the case group includes not
only those who are known to have experienced event 1 but also fractions
of those whose status is unknown due to censoring. Similar justification applies to the controls. This differs from
the IPCW method \cite{blanche2013estimating,schoop2011quantifying}, which uses only uncensored subjects and reweights them to account for
censoring. The IPCW weight is defined as $W_{i}^{IPCW}(\tau)=\dfrac{1(T_{i}\le\tau,\tilde{\delta}_{i}\ne0)}{G_{n}(\tilde{T_{i}}\vert \cdot)}+\dfrac{1(T_{i}>\tau)}{G_{n}(\tau\vert \cdot)}$. It is
the inverse of the probability of being censored, where $G(t\vert \cdot)$
is the censoring distribution that can be estimated by the Kaplan-Meier
estimator or conditionally given covariates.
Estimation of the proposed weight (\ref{eq:wt_cmprsk}) includes estimation
of two quantities: the conditional CIF $F_{1}(\cdot\vert U_{i})$
and the conditional overall survival probability $S(\cdot\vert U_{i})$. We propose
to use a nonparametric kernel-weighted Kaplan-Meier estimator \cite{li2016simple}:
\begin{equation}
\widehat{S}_{T}(t\vert U_{i})=\prod_{\zeta\in\Omega,\zeta\le t}\Big\{1-\frac{\sum_{j}K_{h}(U_{j},U_{i})\cdot1(\tilde{T}_{j}=\zeta,\tilde{\delta}_{j}\ne0)}{\sum_{j}K_{h}(U_{j},U_{i})\cdot1(\tilde{T}_{j}\ge\zeta)}\Big\},\label{eq:kernel_surv}
\end{equation}
and the kernel-weighted CIF \cite{kalbfleisch2011statistical}:
\begin{equation}
\widehat{F}_{1}(t\vert U_{i})=\sum_{\zeta\in\Omega,\zeta\le t}\frac{\sum_{j}K_{h}(U_{j},U_{i})1(\tilde{T_{j}}=\zeta,\tilde{\delta_{j}}=1)}{\sum_{j}K_{h}(U_{j},U_{i})1(\tilde{T}_{j}\ge\zeta)}\cdot\widehat{S}_{T}(\zeta-\vert U_{i}).\label{eq:kernel_cif}
\end{equation}
$\Omega$ is the set of distinct $\tilde{T}_{i}$'s for $\tilde{\delta_{j}}\ne0$
; and $K_{h}(x,x_{0})=\frac{1}{h}K(\frac{x-x_{0}}{h})$ is the kernel
weight with kernel function $K(\cdot)$ and bandwidth $h$. Alternatively,
we can specify a $span$ instead of a fixed bandwidth. A span is the
proportion of subjects around the neighborhood involved in the kernel
estimation with a uniform kernel function. In implementation, the CIF in (\ref{eq:kernel_cif}) can be estimated as a Kaplan-Meier type product-limit estimator, with the hazard function being replaced by the sub-distribution hazard. The at-risk set in the sub-distribution hazard is obtained by reweighting the individuals who had competing events. This process can be achieved by reformatting the competing risk data into a counting
process with \texttt{crprep()} function from the \texttt{mstate} package, and using \texttt{survfit()} in the \texttt{survival}
package by specifying a time-dependent $weight$ in \texttt{R} \cite{geskus2011cause}.
\subsection{The proposed weighting estimators for the time-dependent ROC curve and
AUC }
The estimated weight $\widehat{W}_{1i}$ can be obtained by replacing the CIF and survival functions in (\ref{eq:wt_cmprsk}) with their estimators given by (\ref{eq:kernel_cif}) and (\ref{eq:kernel_surv}). The
$Se(c,\tau)$, $Sp_{A}(c,\tau)$ and $Sp_{B}(c,\tau)$ can be estimated by
\begin{align}
\widehat{Se}(c,\tau) & =\frac{\sum_{i=1}^{n}\widehat{W}_{1i}\cdot1(U_{i}>c)}{\sum_{i=1}^{n}\widehat{W}_{1i}}\nonumber \\
\widehat{Sp}_{A}(c,\tau) & =\frac{\sum_{i=1}^{n}(1-\widehat{W}_{1i})\cdot1(U_{i}\le c)}{\sum_{i=1}^{n}(1-\widehat{W}_{1i})}\label{eq:sen_est}\\
\widehat{Sp}_{B}(c,\tau) & =\frac{\sum_{i=1}^{n}(1-\sum_{k=1}^{K}\widehat{W}_{ki})\cdot1(U_{i}\le c)}{\sum_{i=1}^{n}(1-\sum_{k=1}^{K}\widehat{W}_{ki}).}\nonumber
\end{align}
The estimator of sensitivity can be justified theoretically
as
\begin{align*}
Se(c,\tau) & =P(U>c\vert T\le\tau,\delta=1)\\
& =\frac{E\big(1\{U>c\}\times1\{T\le\tau,\delta=1\}\big)}{E\big(1\{T\le\tau,\delta=1\}\big)}\\
& =\frac{E\Big\{1\{U>c\}\times E\big(1\{T\le\tau,\delta=1\}\vert\tilde{T},\tilde{\delta},U\big)\Big\}}{E\Big\{ E\big(1\{T\le\tau,\delta=1\}\vert\tilde{T},\tilde{\delta},U\big)\Big\}}\\
& =\textrm{lim}_{n\rightarrow\infty}\frac{\sum_{i=1}^{n}1(U_{i}>c)\cdot P(T_{i}\le\tau,\delta_{i}=1\vert\tilde{T}_{i},\tilde{\delta}_{i},U_{i})}{\sum_{i=1}^{n}P(T_{i}\le\tau,\delta_{i}=1\vert\tilde{T}_{i},\tilde{\delta}_{i},U_{i}).}
\end{align*}
The justification for the specificity estimator is similar. The time-dependent
ROC curve is an increasing function obtained by plotting the time-dependent
sensitivity and 1-specificity over a range of threshold $c$'s. By definition,
the AUC can be calculated by trapezoidal integration: $\int_{0}^{1}\widehat{ROC}_{A}(x,\tau)dx=\int_{0}^{1}\widehat{Se}(\widehat{Sp}_{A}^{-1}(1-x,\tau),\tau)dx$
and $\int_{0}^{1}\widehat{ROC}_{B}(x,\tau)dx=\int_{0}^{1}\widehat{Se}(\widehat{Sp}_{B}^{-1}(1-x,\tau),\tau)dx$.
Alternatively, it can be estimated by the empirical estimator of the proportion
of concordance pairs, with the proposed weight estimator $\widehat{W}_{1i}$:
\begin{align}
\widehat{AUC}_{A}(\tau) & =\frac{\sum_{i}\sum_{j}\widehat{W}_{1i}(1-\widehat{W}_{1i})\cdot1(U_{i}>U_{j})}{\sum_{i}\sum_{j}\widehat{W}_{1i}(1-\widehat{W}_{1i})}\nonumber \\
\widehat{AUC}_{B}(\tau) & =\frac{\sum_{i}\sum_{j}\widehat{W}_{1i}(1-\sum_{k=1}^{K}\widehat{W}_{ki})\cdot1(U_{i}>U_{j})}{\sum_{i}\sum_{j}\widehat{W}_{1i}(1-\sum_{k=1}^{K}\widehat{W}_{ki}).}\label{eq:AUC_est}
\end{align}
In practice, we can add $0.5\times1(U_{i}=U_{j})$ to the group of
$1(U_{i}>U_{j})$ to account for ties between the $U$'s. The theoretical
justification for the AUC estimators above is as follows.
\begin{align}
AUC_{A}(\tau) & =P(U_{i}>U_{j}\vert T_{i}\le\tau,\delta_{i}=1,\{T_{j}>\tau\}\cup\{T_{j}\le\tau,\delta_{j}\ne1\})\nonumber \\
& =\frac{E\big(1(T_{i}\le\tau,\delta_{i}=1)\times1(\{T_{j}>\tau\}\cup\{T_{j}\le\tau,\delta_{j}\ne1\})\times1(U_{i}>U_{j})\big)}{E\big(1(T_{i}\le\tau,\delta_{i}=1)\times1(\{T_{j}>\tau\}\cup\{T_{j}\le\tau,\delta_{j}\ne1\})\big)}\nonumber \\
& =\frac{E\Big\{1(U_{i}>U_{j})\cdot E\big(1(T_{i}\le\tau,\delta_{i}=1)\cdot1(\{T_{j}>\tau\}\cup\{T_{j}\le\tau,\delta_{j}\ne1\})\vert\tilde{T}_{i},\tilde{\delta_{i}},U_{i},\tilde{T_{j}},\tilde{\delta_{j}},U_{j}\big)\Big\}}{E\Big\{ E\big(1(T_{i}\le\tau,\delta_{i}=1)\cdot1(\{T_{j}>\tau\}\cup\{T_{j}\le\tau,\delta_{j}\ne1\})\vert\tilde{T}_{i},\tilde{\delta_{i}},U_{i},\tilde{T_{j}},\tilde{\delta_{j}},U_{j}\big)\Big\}}\nonumber \\
& =\textrm{lim}_{n\rightarrow\infty}\frac{\sum_{i}\sum_{j}1(U_{i}>U_{j})\cdot P(T_{i}\le\tau,\delta_{i}=1\vert\tilde{T}_{i},\tilde{\delta}_{i},U_{i})\cdot\big(1-P(T_{i}\le\tau,\delta_{i}=1\vert\tilde{T}_{i},\tilde{\delta}_{i},U_{i})\big)}{\sum_{i}\sum_{j}P(T_{i}\le\tau,\delta_{i}=1\vert\tilde{T}_{i},\tilde{\delta}_{i},U_{i})\cdot\big(1-P(T_{i}\le\tau,\delta_{i}=1\vert\tilde{T}_{i},\tilde{\delta}_{i},U_{i})\big)}\nonumber \\
& =\textrm{lim}_{n\rightarrow\infty}\frac{\sum_{i}\sum_{j}1(U_{i}>U_{j})\times W_{1i}\times\big(1-W_{1i}\big)}{\sum_{i}\sum_{j}W_{1i}\times\big(1-W_{1i}\big)}\label{eq:AUC_justify-1}
\end{align}
A similar justification for $AUC_{B}(\tau)$ is obtained by replacing $\big(1-W_{1i}\big)$ in the
formula (\ref{eq:AUC_justify-1}) with $(1-\sum_{k=1}^{K}W_{ki})$ for the control definition B.
In our numerical studies, the estimator in (\ref{eq:AUC_est}) is
almost identical (up to four digits after the decimal) to the AUC estimator obtained by trapezoidal integration.
The confidence intervals for sensitivity, specificity and AUC can be
estimated numerically by bootstrapping.
\subsection{The Proposed Weighting Estimators for the Brier Score}
By definition, the Brier score is the expected quadratic loss function between the
true disease status $1(T_{i}\le\tau,\delta_{i}=1)$ and the risk score for
event 1, $U_{i}=\pi_{1}(\tau\vert\mathbf{Z}_{i})$, calculated from a prognostic
model to be evaluated. We propose the following estimator for the Brier score, weighting observations according to their probability of having the event of interest:
\[
\widehat{Brier}(\tau)=\dfrac{1}{n}\sum_{i=1}^{n}\Big(\widehat{W}_{1i}\cdot(1-U_{i})^{2}+(1-\widehat{W}_{1i})\cdot(0-U_{i})^{2}\Big).
\]
The justification for consistency of the above estimator is
\begin{align*}
Brier(\tau) & =E\Big\{1(T_{i}\le\tau,\delta_{i}=1)-U_{i}\Big\}^{2}\\
& =E\Big\{ E\Big(\Big[1(T_{i}\le\tau,\delta_{i}=1)-U_{i}\Big]{}^{2}\vert\tilde{T}_{i},\tilde{\delta}_{i},U_{i}\Big)\Big\}\\
& =E\Big\{ P(T_{i}\le\tau,\delta_{i}=1\vert\tilde{T}_{i},\tilde{\delta}_{i},U_{i})\cdot(1-U_{i})^{2}+(1-P(T_{i}\le\tau,\delta_{i}=1\vert\tilde{T}_{i},\tilde{\delta}_{i},U_{i}))\cdot(0-U_{i})^{2}\Big\}\\
& =\textrm{lim}_{n\rightarrow\infty}\frac{1}{n}\sum_{i=1}^{n}\Big(W_{1i}\cdot(1-U_{i})^{2}+(1-W_{1i})\cdot(0-U_{i})^{2}\Big).
\end{align*}
Similarly, the $AbsErr(\tau)$ and $KL(\tau)$ can be estimated with
the proposed conditional probability weight:
\[
\widehat{KL}(\tau)=-\frac{1}{n}\sum_{i=1}^{n}\Big(\widehat{W}_{1i}\text{\ensuremath{\cdot}log}U_{i}+(1-\widehat{W}_{1i})\text{\ensuremath{\cdot}log}(1-U_{i})\Big)
\]
and
\[
\widehat{AbsErr}(\tau)=\frac{1}{n}\sum_{i=1}^{n}\Big(\widehat{W}_{1i}\cdot(1-U_{i})+(1-\widehat{W}_{1i})\cdot U_{i}\Big).
\]
\\
\\
To summarize, the proposed method is a nonparametric
method for estimating the time-dependent predictive accuracy for competing
risk data. It extends the methodology in Li et al. \cite{li2016simple} for a single right-censored time-to-event outcome to competing risk outcomes and to time-dependent calibration metrics. The proposed methodology has a connection to some existing methods. In the context of semi-competing risks with interval censoring, Jacqmin-Gadda et al. \cite{jacqmin2016receiver} proposed an imputation estimator that weights the data with a similar conditional probability of observing an
event in the presence of interval censoring. But their estimator of the conditional probability
is calculated from a parametric illness-death model using the survival and
marker. Schemper \& Henderson \cite{schemper2000predictive} also proposed an imputation method with a Cox model-based estimator for $AbsErr(\tau)$. But this method was shown to be biased when the
prognostic model was misspecified, and an alternative IPCW estimator was proposed in that situation \cite{schmid2011robust}. In contrast, our method is nonparametric, without modeling assumptions, and is applicable to both time-dependent discrimination and calibration metrics. We demonstrated the robustness of the nonparametric method to the selection of tuning parameters in Section 4.3.
\section{Simulation}
In this section, we present simulation studies to evaluate the performance
of the proposed method in estimating both the time-dependent ROC and time-dependent
Brier score in the context of competing risk data. The performance of the proposed
method is compared with those of NNE \cite{saha2010time,zheng2012evaluating}
and IPCW \cite{blanche2013estimating,schoop2011quantifying} methods
from the published literature.
\subsection{Simulation design}
We generate two independent baseline covariates $\boldsymbol{Z}_{i}=(Z_{i1},Z_{i2})$,
where $Z_{i1}$ is a biomarker variable of standard normal distribution,
and $Z_{i2}$ is a baseline characteristic (e.g., gender) of Bernoulli distribution with probability 0.5. The event times are generated according to a Fine-Gray model by using the procedure described in Fine \& Gray \cite{fine1999proportional} with a baseline
sub-distribution hazard (SDH) function and additive covariate effects
on the log SDH. The baseline SDH of event 1 follows a mixture
of Weibull distribution with scale $\lambda_{1}$ and shape $\alpha_{1}$, and a point mass with probability $1-p$ at $\infty$. The log SDH
ratios for covariates $Z_{i1}$ and $Z_{i2}$ are denoted by $\boldsymbol{\beta}=(\beta_{1},\beta_{2})'$ for event 1 and $\boldsymbol{\gamma}=(\gamma_{1},\gamma_{2})'$ for event 2. In our simulations, we set $\boldsymbol{\beta}= (-0.6,0.5)'$,
and $\boldsymbol{\gamma}= (-0.1,-0.2)'$. The event indicator is generated from a Bernoulli distribution with the probability of event 1 being $P_{1}=F_{1}(\infty\vert \boldsymbol{Z})=1-(1-p)^{\textrm{exp}(\boldsymbol{Z \beta}))}$.
The values of $p$ are set to be $(0.22,0.42,0.61)$ to achieve 30\%,
50\% and 70\% of event 1 given the covariate effects. Unless otherwise specified, the random censoring times are generated from a mixture
of uniform distributions on the intervals of $(0,3]\cup(3,6]\cup(6,9]\cup(9,12]\cup(12,15]\cup(15,18]$.
We adjust the probability of falling into each interval to control
the censoring rate. Each simulated data set consists of i.i.d. samples
of $\{(\tilde{T}_{i},U_{i},\tilde{\delta}_{i}),i=1,2,\ldots n\}$:
the observed event time $\tilde{T_{i}}$ is the true event time or censoring
time, whichever comes first; the prognostic score $U_{i}$ is the probability
of experiencing event 1 prior to $\tau$; and the event indicator $\tilde{\delta}_{i}$
takes values of 0, 1, or 2. We use the simulated data sets as validation
data sets to evaluate the predictive accuracy of prognostic score
$U_{i}$ at horizon $\tau$.
We organize the simulation scenarios into a $3\times2\times2$ factorial
design. We consider three proportions for event type 1 (70\%, 50\%
and 30\%), two levels of censoring rates (medium: 25\%-30\% and high:
45\%-50\%) and two sample sizes ($300$ and $600$). The predictive
accuracy is estimated at a time horizon $\tau$, which is approximately
at the 65\% quantile of the observed event time distribution for each
scenario. We compute the true values of $AUC(\tau)$ and $Brier(\tau)$
by a Monte Carlo method using 20,000 independent data sets without
censoring. The prognostic score $U_{i}$ is computed from the true
CIF at $\tau$: $F_{1}(\tau;\boldsymbol{Z})=P(T\le\tau,\delta=1\vert\boldsymbol{Z})=1-\{1-p(1-e^{-\lambda_{1}\tau^{\alpha_{1}}})\}^{\textrm{exp}(\boldsymbol{Z \beta})}$.
In each setting, $500$ Monte Carlo repetitions are performed and
the results are aggregated to compute the bias percentage (bias\%) and MSE in estimating $AUC(\tau)$ and $Brier(\tau)$.
The results are presented in Sections 4.2, 4.3 and 4.4. In Section
4.2, we compare the finite sample performance of the proposed method
with those of some existing methods. For the estimation of the time-dependent
ROC, we compare the proposed estimator with those of the
NNE \cite{saha2010time,zheng2012evaluating} and IPCW methods \cite{blanche2013estimating}.
The NNE method is available in the \texttt{R} package \texttt{CompRisksROC} \cite{saha2010time} for
Definition B of Section \ref{subsec:tdROC_comprsk}, and package
\texttt{SurvCompetingRisk} \cite{zheng2012evaluating} for Definition A. The IPCW method is available
in the R package \texttt{timeROC} \cite{blanche2013estimating}.
For the estimation of the Brier score, the proposed estimator is compared
with that of the IPCW method \cite{schoop2011quantifying}. Since the proposed
method is nonparametric with a tuning parameter (bandwidth or span),
we study the sensitivity of the results to the tuning parameter selection in Section 4.3 and compare the performance with that of another
nonparametric method (NNE) that also uses a bandwidth. In Section 4.4,
we take a closer examination of the relative performance of the proposed
method and IPCW when the censoring time is correlated with the risk score. We consider two
versions of IPCW methods that have been reported in the literature. The first one is the IPCW.KM method \cite{blanche2013estimating,graf1999assessment}, where the censoring distribution in the weight function is estimated by the Kaplan-Meier estimator without conditioning on the risk score:
\begin{equation}
\widehat{W}_{i}^{IPCW.KM}(\tau)=\frac{1(\tilde{T_{i}}\le\tau,\tilde{\delta_{i}}\ne0)}{\widehat{G}(\tilde{T_{i}})}+\frac{1(\tilde{T_{i}}>\tau)}{\widehat{G}(\tau)}.\label{eq:IPCW.KM}
\end{equation}
The second one is the IPCW.Cox method \cite{gerds2006consistent,schoop2011quantifying}, where the censoring distribution in the weight function is estimated from a Cox proportional hazard model, conditioning on the risk score
\begin{equation}
\widehat{W}_{i}^{IPCW.Cox}(\tau)=\frac{1(\tilde{T_{i}}\le\tau,\tilde{\delta_{i}}\ne0)}{\widehat{G}(\tilde{T_{i}}\vert U)}+\frac{1(\tilde{T_{i}}>\tau)}{\widehat{G}(\tau\vert U)}.\label{eq:IPCW.Cox}
\end{equation}
The sensitivity, specificity and Brier score based on the IPCW weight
$\widehat{W}^{IPCW}(t)=1/\widehat{G}(t\vert \cdot)$ from the equations above are estimated as
\begin{align*}
\widehat{Se}^{IPCW}(c,\tau) & =\dfrac{\sum_{i=1}^{n}1(U_{i}>c)\times1(\tilde{T}_{i}\le\tau,\tilde{\delta}_{i}=1)\times\widehat{W}_{i}^{IPCW}(\tilde{T}_{i})}{\sum_{i=1}^{n}1(\tilde{T}_{i}\le\tau,\tilde{\delta}_{i}=1)\times\widehat{W}_{i}^{IPCW}(\tilde{T}_{i})}\\\\
\widehat{Sp}_{A}^{IPCW}(c,\tau) & = \dfrac{\sum_{i=1}^{n}1(U_{i}\le c)\times1(\tilde{T}_{i}>\tau)\times\widehat{W}_{i}^{IPCW}(\tau)}{\sum_{i=1}^{n}1(\tilde{T}_{i}>\tau)\times\widehat{W}_{i}^{IPCW}(\tau)}\\\\
\widehat{Sp}_{B}^{IPCW}(c,\tau) & = \dfrac{\sum_{i=1}^{n}1(U_{i}\le c)\times\Big(1(\tilde{T}_{i}>\tau)\cdot\widehat{W}_{i}^{IPCW}(\tau)+1(\tilde{T}_{i}\le\tau,\tilde{\delta}_{i}\notin\{0,1\}\cdot\widehat{W}_{i}^{IPCW}(\tilde{T}_{i})\Big)}{\sum_{i=1}^{n}\Big\{1(\tilde{T}_{i}>\tau)\cdot\widehat{W}_{i}^{IPCW}(\tau)+1(\tilde{T}_{i}\le\tau,\tilde{\delta}_{i}\notin\{0,1\}\cdot\widehat{W}_{i}^{IPCW}(\tilde{T}_{i})\Big\}}\\\\
\widehat{Brier}^{IPCW}(\tau) & =\frac{1}{n}\sum_{i=1}^{n}\Big(1(\tilde{T}_{i}\le\tau,\tilde{\delta}_{i}=1)-\pi(\tau\vert \boldsymbol{Z})\Big)^{2}\times\widehat{W_{i}}^{IPCW}.
\end{align*}
\subsection{Simulation results on the finite sample performance of the proposed method.}
Table 1 shows the performances of the proposed method, IPCW
and NNE for estimating $\widehat{AUC}_{A}(\tau)$ and $\widehat{AUC}_{B}(\tau)$
under 12 simulation scenarios. For IPCW,
we use the estimator with the weight calculated by (\ref{eq:IPCW.KM}).
In general, the proposed method has smaller bias than the IPCW, and
the magnitude of the bias is negligible ($<1\% $ in most settings).
The NNE method has notably larger bias, especially for $\widehat{AUC}(\tau)$.
The MSE for the proposed method is also the smallest among the three methods
studied. Table 2 shows the performance of the proposed estimators
and IPCW estimators for estimating the Brier score. The bias percentages
of the proposed estimator are less than 1.5\% in all settings and are
in general smaller than those from the IPCW method. The MSEs of the
proposed estimators are also similar to or smaller than those from the
IPCW method. The NNE method was proposed in the literature only for estimating the AUC and hence was not included in the simulation about the Brier score. We conclude
that the proposed method performs similarly or better than the IPCW
method, and both methods are substantially better than the NNE method.
\subsection{Simulation results on the sensitivity to tuning parameter selection.}
One advantage of the proposed method is that it is nonparametric,
which prevents the predictive accuracy from being affected
by the modeling assumptions involved in calculating the predictive accuracy metrics themselves. However, it
does involve a tuning parameter, which is the bandwidth or span that
is used in the kernel weight calculation. Therefore, it is important
to study whether this estimator is sensitive to the tuning parameter
selection. Since the NNE method also uses the tuning parameter, and
to our knowledge no previous work has studied its sensitivity
to the tuning parameter selection, we include that method in the comparison.
Table 3 presents the performance of the proposed and NNE
methods in estimating the AUC under different spans. This table
only includes the results with 70\% of event
1; the results under other scenarios lead to the same general conclusion
and are hence omitted for brevity. When the $span$ varies from 0.05
to 0.5, the proposed method is quite stable and the bias remains under
1.5\% in all scenarios. Slightly larger biases are observed under
two scenarios: small sample size ($n=300$) with small span $(span=0.05$),
and large sample size $(n=600$) with unrealistically large span ($span=0.5$).
When both the sample size and span are small, there is not enough data
for estimation; and when both the sample size and span are
large, bias may be introduced. In contrast, the NNE estimator
is very sensitive to the span and can result in a large bias when the
span is not chosen properly. We speculate that this led to the relatively
large bias shown in Table 1. A similar performance is observed in Table 4
when the Brier score is estimated. A heuristic explanation of the robustness
of the proposed method to the tuning parameter selection is as follows. First,
the tuning parameter only affects subjects who are censored prior
to time $\tau$ because their disease status at $\tau$ is unknown.
This is a smaller proportion than the overall censoring proportion
of the data. Second, the probability weight $W_{1i}=\dfrac{F_{1}(\tau\vert U_{i})-F_{1}(\tilde{T}_{i}\vert U_{i})}{S(\tilde{T}_{i}\vert U_{i})}$
is defined as the ratio of two conditional probabilities for subjects
censored before $\tau$.
The numerator of $W_{1i}$ can be expressed as the cause-specific
survival probability between $\tilde{T}_{i}$ and $\tau$: $S_{1}(\tilde{T}_{i}\vert U_{i})-S_{1}(\tau\vert U_{i})=Pr(\tilde{T}_{i}<T\le\tau,\delta_{i}=1\vert U_{i})$;
and the denumerator is the overall survival probability beyond $\tilde{T}_{i}$. The asymptotic bias of two conditional survival probabilities as
a function of bandwidth are in the same direction \cite{bordes2011uniform}.
Therefore, the bias of their ratio can be canceled out to some extent,
particularly when $\tilde{T}_{i}$ and $\tau$ are close.
\subsection{Simulation results for the performance of the proposed method under dependent
censoring. }
In this section, we compare the proposed method and IPCW under a
dependent censoring scenario where the event time $T$ and censoring
time $C$ are marginally dependent but are conditionally independent
given the risk score $U$. In practice, the censoring time is often correlated with baseline covariates. Since $U$ is a function of these covariates,
$C$ and $U$ may also be correlated. Literature on the time-dependent
ROC and time-dependent Brier score describes estimation under dependent censoring of this kind using the
IPCW approach, where a Cox model is used to estimate the censoring distribution, conditioning on the risk score \cite{gerds2006consistent,schoop2011quantifying}. In contrast,
our proposed method does not model the censoring distribution, which is a nuisance for scientific purposes. We directly estimate the conditional survival and CIF nonparametrically. In this simulation, we consider two settings. In setting $(a)$, we generate censoring
time $C_{i}$ from a Weibull$(\lambda_{c},\alpha_{c})$ distribution
with the mean $\mu_{C}=\frac{\Gamma(1+1/\alpha_{c})}{\lambda_{c}^{1/\alpha_{c}}}=a*1\{(\zeta >0.4)\cup(\zeta <-0.6)\}+b*1\{-0.6\le \zeta \le0.4\}$, where $\zeta=\boldsymbol{Z \beta}$ is a monotone transformation of $U$.
Different values of $(a,b)$ and $\alpha_{c}$ are chosen to achieve
a medium or high censoring rate. The dependency between the censoring
distribution and $U$ is not monotone and cannot be correctly estimated
by a proportional hazard model. We use setting (a) to study the robustness of the methods to model misspecification. In setting $(b)$, we generate
the censoring time from a Cox model on $\zeta$, so that the censoring time is correctly modeled by the IPCW. For both
settings, we compare the performance of the proposed method and IPCW methods with both weight estimators (\ref{eq:IPCW.KM}) and (\ref{eq:IPCW.Cox}).
Tables 5 and 7 compare the performance of the proposed method with that of the IPCW
in estimating $AUC(\tau)$. All bias percentages for the proposed
method are under 1.5\% and 1\% for settings $(a)$ and $(b)$, respectively.
In contrast, the IPCW.KM method, which ignores the dependent censoring,
produces results with a large bias under both mechanisms. Compared to
IPCW.KM, the IPCW.Cox estimator in setting $(a)$ alleviates the bias
by accounting for the dependence but still has larger bias and MSE
than the proposed method, especially when the type 1 event rate is
low (e.g., 30\%). When the censoring times are generated from the
Cox model in setting $(b)$, the bias from the IPCW.Cox method is controlled
under 1.5\% but is still slightly larger than that from the proposed method
in general. This indicates that the proposed method is more robust
than the IPCW methods under different dependence structures of $C$ and
$U$.
Tables 6 and 8 present similar comparisons between the proposed method
and IPCW in estimating the Brier score. The overall performance
is similar to that of $\widehat{AUC}(\tau)$. However, we notice that when the IPCW.Cox method is used under a misspecified censoring mechanism in setting $(a)$, it produces a larger bias in the estimation of the Brier score than the AUC. In contrast, the performance of IPCW.Cox under
setting $(b)$ is similar in both estimands, with the biases well
controlled under 1.5\%. The results indicate that estimation of $\widehat{Brier}(\tau)$
appears to be more sensitive to misspecification than that of $\widehat{AUC}(\tau)$.
We speculate that this is because $AUC(\tau)$ is based on the rankings
of the data, whereas $Brier(\tau)$ measures the actual deviation
from the true status in quantity and therefore is more sensitive to the
misspecification of the estimation procedure.
The results above suggest that our nonparametric method does not suffer from
bias caused by model dependence. The rationale for developing a nonparametric
estimation method is that the estimator of a predictive accuracy metric
should be an objective reflection of the model under evaluation, without
introducing another source of bias due to the modeling assumption of
the estimation method. In this spirit, one can extend
the IPCW method by using a nonparametric estimator
for the conditional distribution of the censoring time given the risk score. But from a clinical perspective, this conditional distribution is less intuitive than directly
modeling the conditional survival distribution, which offers additional
insight into the relationship between the risk score and disease development. In addition, the relationship between the risk score and the survival time is expected to be monotone by the definition of the ROC, but this is not necessarily the case for the relationship between the risk score and the censoring time. The nonparametric smoothing literature suggests that the nonparametric regression result is less sensitive to the tuning parameters when the relationship between the outcome and covariate is monotone \cite{meyer2008inference}.
In summary, the simulation results from Table 1 to Table 8 demonstrate
that the proposed method has similar or better performance than other
published methods. While the NNE method only estimates the time-dependent
ROC, the proposed method works with the time-dependent ROC, time-dependent
Brier score and other predictive accuracy metrics, with notably smaller
bias and MSE. Unlike the NNE, the proposed method is
robust to tuning parameter selection, which makes it easy to use in
practice. As a nonparametric method, the proposed method outperforms
the IPCW under dependent censoring, particularly in light of the
possibility that IPCW may use a misspecified model for the censoring
distribution.
\section{Application}
We illustrate the proposed method with a data set from AASK, a randomized
clinical trial for 1,094 patients with chronic kidney disease, whose
baseline estimated glomerular filtration rates (eGFRs) were between $20-65\ \textrm{mL/min/1.73m}^{2}$
\cite{wright2002effect}. The patients were followed for 6.5 years
during the trial period. Among them, 179 developed ESRD and 85 died before developing ESRD. We evaluate the predictive
accuracy of a prognostic risk score developed from a proportional
sub-distributional hazard model with five baseline covariates: the
eGFR, urine protein creatinine
ratio, age, gender, the randomized blood pressure group (low and medium) and the randomized anti-hypertensive therapy (ramipril, metoprolol, amlodipine). The prognostic
score is the predicted CIF for ESRD at prespecified horizons.
Figure 1 compares the time-dependent ROC curves estimated from the
proposed method (red), IPCW.KM (black), IPCW.Cox (blue) and NNE (green)
at three predictive horizons: 3, 4 and 5 years from baseline. The span used in the proposed and NNE methods is 0.05, which includes 5\% of the neighborhood data. The two rows in the panel present the estimated ROC curves based on the two definitions (Section 2.2).
Definition A discriminates patients with ESRD within $\tau$ years from ESRD-free patients, which include patients who are event-free and who die by year $\tau$.
Definition B discriminates patients with ESRD within $\tau$ years from those who are event-free at year $\tau$.
The ROC curves from the two IPCW methods, IPCW.KM and IPCW.Cox, are
almost identical. The curves by IPCW and the proposed method are also very
close, and the differences between the $\widehat{AUC}(\tau)$ are within
5\%. The estimated $\widehat{AUC}^{A}(\tau)$ and $\widehat{AUC}^{B}(\tau)$ are also very close within the different estimation methods except
for NNE. This indicates that the sub-distribution hazard
model we used can discriminate well between ESRD patients and ESRD-free or event-free patients. A possible explanation is that the patients who died in the study period are a relatively small population and may have died from causes unrelated to kidney disease. Therefore, adding these patients to the control group may not substantially change the discrimination
of the risk score, which primarily consists of risk factors for ESRD. There
is some discussion of how to use different definitions of controls in the ROC estimation \cite{zheng2012evaluating}; the choice is related to the clinical context and here we provide estimation methods
for both.
In Figure 2, we show further results of our study of the proposed and NNE methods with varying spans of
0.05, 0.1, and 0.3. The proposed
method produces stable $\widehat{AUC}(\tau)$ around 0.88 while the NNE
method is very sensitive to the span specification. This result is consistent
with the simulation results in Table 4. Such robustness to the tuning parameter
selection is a very attractive feature for our nonparametric estimator.
The Brier scores over all the predictive horizons are plotted
in Figure 3, along with the percentages of ESRD and censoring at each predictive horizon. The prediction error increases with the predictive horizon. This result implies that the predictive accuracy decreases as the predictive horizon moves away from the time of prediction. Overall the estimated Brier scores are small, between 0 and 0.11. Prior to year 3.5, when there is little censoring,
the three estimation methods produce almost identical results.
When the percentage of censoring increases beyond 3.5 years, the results from the three
methods begin to diverge but the absolute differences among them remain
small.
\section{Discussion }
In this paper, we propose an analytical framework for estimating
time-dependent predictive accuracy metrics with competing risk data
that are subject to right censoring. The method is illustrated with
the time-dependent ROC and time-dependent Brier score. The
proposed framework first computes a nonparametric estimator of the
conditional probability of the true event status given the observed
data and then uses it to weigh the data in an empirical calculation
of the time-dependent metrics. This is a unified
approach to estimating
the time-dependent ROC, time-dependent Brier score, and time-dependent
metrics constructed from other loss functions. The proposed method requires
no parametric assumptions about the marginal, conditional or joint
distribution of the risk score and time to the event of interest. It can be applied to evaluate the discrimination for a
single biomarker or a risk score constructed from a prognostic model with multiple biomarkers, and to evaluate the
calibration of the prognostic model.
The method is applicable when the censoring time and the risk score are correlated. It is also insensitive to the tuning parameter specification. Such robustness to the tuning
parameter specification has not been studied in nonparametric estimations of time-dependent predictive accuracy metrics \cite{heagerty2000time,saha2010time,zheng2012evaluating}
and no guidelines are yet available for practical users. When compared with competing methods in simulations, our proposed method demonstrates better
overall performance and robustness to tuning parameters, particularly when the censoring is correlated with the risk score. The R code that implements the proposed methodology is available upon request and will be added to the \texttt{tdROC} package in \texttt{R}.
One limitation with the proposed method is that, like many other nonparametric
methods, it works better with larger sample sizes. When the sample size is very
small, there may not be enough subjects with events for calculating $\widehat{F}_{1}(t\vert U_{i})$ and $\widehat{S}_{T}(t\vert U_{i})$
within some local neighborhoods defined by the kernel. In such case, the bandwidth may need to be increased for those neighborhoods.
\section{Acknowledgment }
The authors declare no potential conflicts of interest with respect
to the research, authorship and publication of this article. This
research was supported by the U.S. National Institutes of Health (grants
5P30CA016672 and 5U01DK103225).
|
\section{Introduction}
Spatial filtering has been central in the development of large eddy
simulation reduced order models
(LES-ROMs)~\cite{wang2012proper,xie2017data,xie2017approximate} and
regularized reduced order models
(Reg-ROMs)~\cite{sabetghadam2012alpha,wells2017evolve} for efficient
and relatively accurate numerical simulation of convection-dominated
fluid flows. In this paper, we perform a numerical investigation of
spatial filtering. To this end, we consider one of the simplest
Reg-ROMs, the Leray ROM
(L-ROM)~\cite{sabetghadam2012alpha,wells2017evolve}, which uses ROM
spatial filtering to smooth the flow variables and decrease the amount
of energy aliased to the lower index ROM basis functions. We also
propose a new form of ROM differential
filter~\cite{sabetghadam2012alpha,wells2017evolve} and use it as a
spatial filter for the L-ROM. We investigate the performance of this
new form of ROM differential filter in the numerical simulation of a
flow past a circular cylinder at a Reynolds number $Re=760$.
\section{Reduced Order Modeling}
For the Navier-Stokes equations (NSE), the standard reduced order
model (ROM) is constructed as follows: (i) choose modes $\{ \podvec_1,
\ldots, \podvec_d \}$, which represent the recurrent spatial
structures of the given flow; (ii) choose the dominant modes $\{
\podvec_1, \ldots, \podvec_r \}$, $r \leq d$, as basis functions for
the ROM; (iii) use a Galerkin truncation $\uu_r = \sum_{j=1}^{r} a_j
\, \podvec_j$; (iv) replace $\uu$ with $\uu_r$ in the NSE; (iii) use a
Galerkin projection of NSE($\uu_r$) onto the ROM space $X^r :=
\mbox{span} \{ \podvec_1, \ldots, \podvec_r \}$ to obtain a
low-dimensional dynamical system, which represents the ROM:
\begin{eqnarray}
\dot{\va}
= A \, \va
+ \va^{\top} \, B \, \va \, ,
\label{eqn:g-rom}
\end{eqnarray}
where $\va$ is the vector of unknown ROM coefficients and $A, B$ are
ROM operators; (iv) in an offline stage, compute the ROM operators;
and (v) in an online stage, repeatedly use the ROM (for various
parameter settings and/or longer time intervals).
\section{ROM Differential Filter}
\label{sec:filtering:pod-differential-filter}
The ROM differential filter is based on the classic Helmholtz filter
that has been used to great success in LES for turbulent
flows~\cite{germano1986differential}. Let \(\delta\) be the radius of
the differential filter. Then, for a given velocity field \(\uu_r \in
X^r\), the filtered flow field \(\FF(\uu_r) \in X^f\), where \(X^f\)
is a yet to be specified space of filtered ROM functions, is defined
as the solution to the Helmholtz problem
\begin{equation}
{\rm Find}\ \FF(\uu_r) \in X^f \ {\rm such\ that}\
\label{eq:filtering:pod-differential-filter}
\left((I - \delta^2 \Delta) \FF(\uu_r), \vv\right)
= (\uu_r, \vv),
\ {\rm for\ all}\ \vv \in X^f.
\end{equation}
We consider two different versions for the choice of the range of the
ROM differential filter \(X^f\):
\smallskip
\noindent\textbf{The FE Version.} This version corresponds to \(X^f =
X^h\), where $X^h$ is the finite element (FE) space: we seek the FE
representation of \(\FF(\uu)\) and work in the full discrete space
when calculating the filtered ROM vectors. The FE
representation of \(\FF(\uu)\) suffices in applications because we use
it to assemble the components of the ROM before time evolution: put
another way, since filtering is a linear procedure, it only has to be
done once and not in every ROM time step, \eg for FE mass
and stiffness matrices \(M\) and \(S\) we have that, modulo boundary
condition terms,
\begin{equation}
a_j (M + \delta^2 S) \FF\left(\podvec_j\right)
= a_j M \podvec_j
\Rightarrow
(M + \delta^2 S) \sum_{j = 1}^r a_j \FF(\podvec_j)
= M \sum_{j = 1}^r a_j \podvec_j.
\label{eq:filtering:discrete-pod-differential-filter}
\end{equation}
Hence, applying the differential filter to each proper orthogonal
decomposition (POD) basis vector \(\podvec_j\), results in
\(\FF(\podvec_j) \notin X^r\).
Due to the properties of the differential filter (see
Fig.~\ref{figure:filtering:filtered-pod-vectors}), these new ROM
functions will correspond to longer length scales and contain less
energy. \smallskip
\noindent\textbf{The ROM Version.} Alternatively, we can pick \(X^f =
X^r\), \ie the ROM differential filter simply corresponds to an \(r
\times r\) Helmholtz problem.
\begin{equation}
(M_r + \delta^2 S_r) \FF\left(\va\right)
= M_r \va \, ,
\label{eq:rom-df}
\end{equation}
where $M_r$ and $S_r$ and the ROM mass and stiffness matrices,
respectively, and $\va$ and $\FF\left(\va\right)$ are the POD
coefficient vectors of \(\podvec_j\) and \(\FF(\podvec_j)\),
respectively.
Here, unlike in the FE version, the range of the Helmholtz filter is
\(X^r\), so filtered solutions retain the weakly divergence free
property. \smallskip
\noindent\textbf{Properties.}
\begin{figure}[t]
\centering
\includegraphics[width=1.0\textwidth
{fig1}
\caption{Contour plots of \(y\) velocity of the first and fifth POD
vectors from a 3D flow past a cylinder computation. We use the FE
version of the ROM differential filter with $\delta = 0.5$. The
unfiltered POD vectors are on top and the filtered are on the
bottom. The choice $\delta = 0.5$ is too large for practical
purposes, but demonstrates that filtering both removes kinetic
energy (the isosurfaces are smaller) and enlarges the scales of
motion (\eg the first POD vector goes from twelve structures to just
nine.)}
\label{figure:filtering:filtered-pod-vectors}
\end{figure}
Both versions of the ROM differential
filter~(\ref{eq:filtering:pod-differential-filter}) share several
appealing properties~\cite{BIL05}. They act as spatial filters, since
they eliminate the small scales (\ie high frequencies) from the input.
Indeed, the ROM differential
filter~(\ref{eq:filtering:pod-differential-filter}) uses an elliptic
operator to smooth the input variable. They also have a low
computational overhead. For efficiency, the algorithmic complexity of
any additional filters should be dominated by the \(\OO(r^3)\) cost in
evaluating the nonlinearity. The ROM version is equivalent to solving
an \(r \times r\) linear system; since the matrix only depends on the
POD basis, it may be factorized and repeatedly solved for a cost of
\(\OO(r^2)\), which is also dominated by the cost of the nonlinearity.
The FE version requires solving large FE linear systems, but these
linear systems are solved in the offline stage; thus, the online
computational cost of the FE version is negligible. Finally, we
emphasize that the ROM differential filter uses an \emph{explicit}
length scale \(\delta\) to filter the ROM solution vector. This is
contrast to other types of spatial filtering, e.g., the ROM
projection, which do not employ an explicit length scale.
\section{Leray ROM}
\label{sec:regrom:leray-model}
Jean Leray attempted to solve the NSEs in his landmark 1934
paper~\cite{leray1934sur}. He was able to prove the existence of
solutions for the modified problem
\begin{equation}
\label{eq:regrom:leray_regularized_ns_pde}
\ww_t = \frac{1}{Re} \Delta \ww-\FF(\ww) \cdot \nabla \ww-\nabla p,
\end{equation}
where $\nabla \cdot \ww =0$, and \(\FF(\ww)\) is a convolution with a
compact support mollifier with filter radius \(\delta\), or
\begin{equation}
\label{eq:regrom:gaussian_convolution}
\FF(\ww) = g_\delta \star \ww.
\end{equation}
For additional discussion on the properties of different filters
see~\cite{BIL05,layton2012approximate, Sag06}.
We approximate the convolution with the differential filter
\begin{equation}
\label{eq:regrom:differential_filter}
\FF(\ww) = (\delta^2 \Delta + 1)^{-1} \ww.
\end{equation}
In turbulence modeling, Leray's model is the basis for a class of
stabilization methods called the
Leray-\(\alpha\) regularization models~\cite{layton2012approximate}. Leray's
key observation was that the nonlinear term is the most problematic as
it serves to transfer energy from resolved to unresolved
scales.
The Leray model has been recently extended to the ROM
setting~\cite{sabetghadam2012alpha,wells2017evolve}. The resulting
\emph{Leray-ROM (L-ROM)} can be written as
\begin{equation}
\label{eq:regrom:filtered_centered_pde}
(\ww_r)_t = \frac{1}{Re} \Delta \ww_r-\FF(\ww_r) \cdot \nabla \ww_r-\nabla p,
\end{equation}
which is the same as the Galerkin ROM up to the filtering of the
advective term in the nonlinearity.
\section{Numerical Results}
We consider the flow past a cylinder problem with parabolic Dirichlet
inflow conditions, no-slip boundary conditions on the walls of the
domain, and zero tangential flow at the outflow. %
We compute snapshots by running the \texttt{deal.II}
\cite{bangerth2016deal} step-35 tutorial program for \(t \in [0, 500]\).
We use a kinematic viscosity value of $1/100$, a circular
cylinder with diameter of $1$, and parabolic inflow boundary
conditions with a maximum velocity of $7.6$; this results in a
Reynolds number $Re=760$.
We calibrate the filter radius \(\delta\) by choosing a value for
\(\delta\) that gives the L-ROM the same mean kinetic energy as the
original numerical simulation. Calibrating the ROM to this filter
radius also improves accuracy in some structural properties: this
amount of filtering removes enough kinetic energy that the phase
portrait connecting the coefficients in the ROM on the first and
second POD basis functions are close to the values obtained by
projecting the snapshots onto the POD basis over the same time
interval.
Fig.~\ref{fig:regrom:ns-leray-df-l2-norms} displays the time evolution
of the \(L^2\) norm of the solutions of the L-ROM and DNS for \(r =
6\) and \(r = 20\). Fig.~\ref{fig:regrom:ns-leray-df-l2-norms} shows
that, for the optimal \(\delta\) value, the L-ROM-DF accurately
reproduces the average, but not the amplitude of the time evolution of
the \(L^2\) norm of the DNS results for both \(r = 6\) and \(r = 20\).
\begin{figure}[!htb]
\label{fig:regrom:ns-leray-df-l2-norms}
\centering
\includegraphics[width=1.0\textwidth
{fig2}
\caption{3D flow past a cylinder, L-ROM (green) and DNS
(blue). Mean (left column) and time evolution (right column)
of the $L^2$ norm of the solution; $r = 6$ (top row) and $r
= 20$ (bottom row). The time evolution of the $L^2$ norm of
the solution (right column) is plotted for the optimal mean
$L^2$ norm of the solution (left column): $\delta = 0.33$
for $r = 6$ (top row) and $\delta = 0.18$ for $r = 20$
(bottom row). }
\end{figure}
Fig.~\ref{fig:regrom:ns-leray-df-phase-0-1} displays the phase
portraits for the first and second POD coefficients of the
L-ROM-DF and POD projection of DNS data for \(r = 6\) and \(r =
20\). Fig.~\ref{fig:regrom:ns-leray-df-phase-0-1} shows that, for
the optimal \(\delta\) value, the L-ROM-DF yields moderately
accurate results for \(r = 6\) and accurate results for \(r =20\).
\begin{figure}[!htb]
\label{fig:regrom:ns-leray-df-phase-0-1}
\centering
\includegraphics[width=1.0\textwidth
{fig3}
\caption{3D flow past a cylinder, L-ROM-DF with optimal
$\delta$ value (green) and POD projection of DNS data
(blue). Phase portraits for $a_1$ and $a_2$; $r = 6$ (left)
and $r = 20$ (right).}
\end{figure}
\section{Conclusions}
In this paper, we proposed a new type of ROM differential filter. We
used this new filter with the L-ROM, which is one of the simplest
Reg-ROMs. We tested this filter/ROM combination in the numerical
simulation of a flow past a circular cylinder at Reynolds number
$Re=760$ for $r=6$ and $r=20$. The new type of ROM differential
filter yielded encouraging numerical results, which were comparable to
those for the standard type of ROM differential filter and better than
those for the ROM projection~\cite{wells2017evolve}. We emphasize
that a major advantage of the new type of ROM differential filter over
the standard ROM differential filter is its low computational
overhead. Indeed, since the filtering operation in the new type of
ROM differential filter is performed at a FE level (as opposed to the
ROM level, as it is generally done), the new filter is applied to each
ROM basis function in the offline stage. In the online stage, the
computational overhead of the new type of ROM differential filter is
practically zero, since it simply amounts to using the filtered ROM
basis functions computed and stored in the offline stage.
The first results for the new type of ROM differential filter are
encouraging. We plan to perform a thorough investigation of the new
filter, including a comparison with the standard form of the ROM
differential filter and the ROM projection, in the numerical
simulation of realistic
flows~\cite{wells2017evolve,xie2017approximate}.
|
\section{Introduction}
The standard cosmological model is based on the postulate that the universe is homogeneous and
isotropic on large distance scales. However, there exist many observations which suggest that
this postulate is violated. The first indication of statistical anisotropy came from the analysis
of radio polarization from distant radio galaxies \citep{Birch1982, Jain1999, Jain2004}. Those
authors found that the polarization offsets, after eliminating the effect of Faraday rotation
show a dipole pattern on the sky and the dipole axis points towards the Virgo cluster of galaxies,
very close to the direction of the CMB dipole \citep{Jain1999}. The optical polarizations of
distant quasars shows alignment over very large distance scales \citep{Hutsemekers1998, Hutsemekers2001}.
The distance scale of alignment is found to be of the order of 1 Gpc \citep{Jain2004}, and the
axis again points close to the CMB dipole axis.
The CMB quadrupole and octopole also indicates a preferred direction pointing towards the Virgo
cluster \citep{Copi2004, Ralston2004, Costa2004, Schwarz2004, Bielewicz2005, Samal2008}.
This phenomenon where several axes from various data sets broadly indicate the same direction
has been called \emph{Virgo alignment puzzle} \citep{Jain2004}.
Besides the Virgo alignment, there also exists statistically significant signals of anisotropy
in CMB temperature data viz., hemispherical power asymmetry \citep{Eriksen2004a, Prunet2005,
Hansen2009, Hanson2009, Hoftuft2009, Rath2013, Akrami2014, Aiola2015, Pranati2015},
parity asymmetry \citep{Land2005, Kim2010, Kim2011, Aluri2012, Zhao2014, Aluri2017} and a region
of significant temperature decrement known as cold spot \citep{Vielva2004, Cruz2005, Cruz2006,
Cruz2008, Zhao2013, Nadathur2014, Aluri2016}.
Large angle CMB anisotropies have been a subject of several studies in the cosmology literature
- see for example \citet{wmap7yranom,Planck2013iso,Planck2015iso} for an evaluation of some
of the prominent large angle anomalies by WMAP and PLANCK collaborations.
General methods have been developed to test any violation of statistical isotropy in the CMB data
\citep{Hajian2003, Hajian2006, Copi2004, Copi2006, Samal2008, Samal2009}.
In \citet{Samal2008, Samal2009}, a symmetry based method for testing the isotropy of CMB
temperature data called \emph{Power tensor} was proposed.
The method is based on identifying invariants corresponding to each multipole using the
Power tensor matrix defined as
\begin{equation}
A_{ij} (l) = \frac{1}{l(l+1)(2l+1)}\sum_{m,m'} a^*_{lm}(J_{i} J_{j})_{mm'} a_{lm'}\,.
\label{eq:pt}
\end{equation}
Here $J_i (i = 1, 2, 3)$ are the angular momentum operators in spin$-l$ representation.
The sum of the three eigenvalues reproduce the usual angular power spectrum, $C_l$.
Statistical nature of CMB anisotropies lead to fluctuations in the eigenvalues of Power tensor
about their expected value of $C_l/3$ in a given realization.
The significance of any deviation from isotropy is measured using an invariant combination
of normalized eigenvalues of the Power tensor called \emph{Power entropy}
\citep{Samal2008,Samal2009,Rath2015}.
In this paper, we probe statistical anisotropy, if any, in the PLANCK CMB polarization
data using our symmetry based method, and its possible relation to the existing anomalies
of CMB temperature sky. A violation of statistical isotropy can arise due to a variety
of sources, for example, from residual contamination due to foregrounds, beam systematics,
inhomogeneous noise, etc. Here we do not make any attempt to find the exact cause of
breakdown of isotropy, if found eventually in our analysis.
The paper is organized as follows. In section~\ref{sec:revpol}, we briefly review CMB
polarization convention. Then the statistics used to analyze the data are defined in
section~\ref{sec:stat}. The data used in the present work is described in section~\ref{sec:data},
and our results are discussed in section~\ref{sec:results}. There, we first present
the results from analyzing full sky PLANCK CMB polarization maps. Then we go on to discuss
the effect of foreground residuals if any on our results by employing different galactic
masks, also provided by PLANCK team. We also present results from our tests for any
coherent alignments among various multipoles. Finally, the work is summarized in
section~\ref{sec:con}.
\section{CMB polarization maps}
\label{sec:revpol}
The full sky CMB signal is described by Stokes parameters $I$, $Q$, $U$ and $V$. The Stokes
parameter $I$ represents the temperature field, and the Stokes parameters $(Q,U)$ represent
the linear polarization field. In the standard model of cosmology, the CMB temperature
fluctuations are expected to be an isotropic Gaussian random field, and are conventionally
expanded in terms of the spherical harmonics as
\begin{equation}
T(\hat n) = \sum_{lm} a_{lm}Y_{lm}(\hat n)\,,
\end{equation}
where $a_{lm}$ are the coefficients of expansion. The CMB polarization is induced by Thomson
scattering of CMB photons at the last scattering surface. The WMAP and PLANCK teams provided
maps of CMB polarization in terms of the Stokes parameters $Q$ and $U$, though with low signal
to noise ratio (SNR). The Stokes parameter $V$, which describes circular polarization is
ignored as it cannot be generated through Thomson scattering.
From now onwards we use the notation of \citet{Zaldarriaga1997} for the polarization fields.
Instead of the Stokes parameters $Q$ and $U$, it is useful to employ special combinations of
these parameters as $X_\pm=Q \pm i U$, which transform as spin-$2$ fields under coordinate
rotation. Under a rotation by an angle $\phi$ of the co-ordinate frame in which the polarization
vector is defined, the combinations $X_\pm$ behave as spin $\pm 2$ fields viz.,
\begin{equation}
(Q\pm i U)'(\hat{n}) = e^{\mp 2 i \phi}( Q\pm i U)(\hat{n})\,.
\end{equation}
Analogous to the expansion of temperature field in terms of the spherical harmonics,
$Y_{lm}(\hat n)$, there also exists a set of spin-$s$ spherical harmonics $_{s}Y_{lm}(\hat n)$,
in terms of which one can expand a spin-$s$ function on a sphere. We can, therefore,
expand $X_\pm(\hat n)=( Q\pm i U)(\hat{n})$ in terms of the spin-$2$ spherical harmonics as
\begin{equation}
X_\pm(\hat{n})=\sum_{lm} a_{\pm 2,lm} \ _{\pm 2}Y_{lm}(\hat{n})\,.
\label{spin2}
\end{equation}
Owing to the real nature of $Q$ and $U$ parameters, the expansion coefficients satisfy
the condition : $a^*_{-2,lm} = a_{2,l-m}$. Using the following identities
\begin{eqnarray}
\eth \ _{s}Y_{lm} &=& \sqrt{(l-s)(l+s+1)} \ _{s+1}Y_{lm}\,,\\
\bar{\eth}\ _{s}Y_{lm} &=& -\sqrt{(l+s)(l-s+1)} \ _{s-1}Y_{lm}\,.
\end{eqnarray}
for the spin raising and lowering operators : $\eth$ and $\bar{\eth}$, spin-$0$ objects
can be constructed from spin-$2$ fields as
\begin{eqnarray}
\bar{\eth}^2(Q+ i U)(\hat{n})&=&\sum_{lm}\sqrt{\frac{(l+2)!}{(l-2)!}} \ a_{2,lm}\ Y_{lm}(\hat{n})\,,\\
\eth^2(Q- i U)(\hat{n})&=&\sum_{lm}\sqrt{\frac{(l+2)!}{(l-2)!}} \ a_{-2,lm}\ Y_{lm}(\hat{n})\,.
\end{eqnarray}
Finally, the rotationally invariant polarization fields are defined as
\begin{eqnarray}
\tilde{E}(\hat{n})& = &-\frac{1}{2}\left[\bar{\eth}^2(Q+ i U)+\eth^2(Q- i U)\right]\nonumber\\
& = &\sum_{lm} \tilde{a}^E_{lm}Y_{lm}(\hat{n})\,,
\end{eqnarray}
\begin{eqnarray}
\tilde{B}(\hat{n}) & = &-\frac{1}{2i}\left[\bar{\eth}^2(Q+ i U)-\eth^2(Q- i U)\right]\nonumber\\
& = &\sum_{lm} \tilde{a}^B_{lm}Y_{lm}(\hat{n})\,
\end{eqnarray}
where $\tilde{a}^{E/B}_{lm}=\sqrt{(l+2)!/(l-2)!} \ a^{E/B}_{lm}$, and $a^E_{lm}$ and
$a^B_{lm}$ are the spherical harmonic coefficients of ``$E$-mode'' and ``$B$-mode''
polarization fields. Note that often $E/B$ harmonic coefficients (and consequently
the $E/B$ fields) are defined without the extra $\sqrt{(l+2)!/(l-2)!}$ factor.
These $E/B$ spherical harmonic coefficients are given by linear combinations of
spin-$2$ spherical harmonic coefficients of Stokes $Q/U$ polarization fields as
\begin{equation}
a^E_{lm}=-\frac{1}{2}\left( a_{2,lm}+ a_{-2,lm}\right)\,,
\label{eq:Epol}
\end{equation}
\begin{equation}
a^B_{lm}=-\frac{1}{2i}\left( a_{2,lm}- a_{-2,lm}\right)\,.
\label{eq:Bpol}
\end{equation}
Here we restrict our attention to the $E(\hat n)$ field (ie., $a^E_{lm}$ coefficients),
to study statistical isotropy of PLANCK polarization maps.
\section{Statistics}
\label{sec:stat}
The angular orientation of each mode is given by a unique orthonormal frame $e_k^{\alpha}(l)$
and rotationally invariant singular values $\Lambda_{\alpha}(l)$ of the Power tensor, $A(l)$,
defined in Eq.~(\ref{eq:pt}). Here $k=1,2,3$ denote the components of the frame vector
${\bf e}^\alpha(l)$ and $\alpha=1,2,3$ stands for the singular value index.
In terms of these quantities, the Power tensor matrix for each multipole `$l$' can be
expressed as
\begin{equation}
A_{ij}(l) = \sum_{\alpha} e_{i}^{\alpha}(\Lambda^{\alpha})^2 e_j^{\alpha*}\,.
\end{equation}
We do not explicitly display the index $l$ when it is obvious.
We refer to the eigenvector corresponding to the largest eigenvalue of the Power tensor
as \emph{principal eigenvector} (PEV), and is taken to be the anisotropy axis of that
multipole. The preferred direction represented by a PEV of any multipole is quantified by
parametrizing the dispersion of eigenvalues using \emph{Power entropy} that is defined as
\begin{equation}
S_{p} (l) = -\sum_{\alpha} \lambda^{\alpha} \log(\lambda^{\alpha}) \,
\label{eq:powent}
\end{equation}
where $\lambda^{\alpha} = (\Lambda^{\alpha})^2/\sum_{\beta}(\Lambda^{\beta})^2$. In the ideal
case of isotropy, where all the three eigenvalues are degenerate and equal to ${C_l}/3$, we
have maximum Power entropy, $S_p \rightarrow \log(3)$. In the case of a \emph{pure state},
where one of the eigenvalues contains the total power and other two vanishes, it leads to
vanishing Power entropy, $S_p \rightarrow 0$. So for our observational data, the range of
Power entropy is $0\le S_p \le \log(3)$. Hence low Power entropy in data, compared to consistently
generated concordance model simulations, is a measure of isotropy violation in the data.
Now, a common alignment vector using PEVs for a set of multipoles or range of multipoles
can be calculated using what is called an \emph{Alignment entropy}, defined by
\begin{equation}
S_{X} = -Tr(\rho_X \log(\rho_X))\,,
\label{eq:alignent}
\end{equation}
where $\rho_X = {X}/{Tr(X)}$ is the normalized $3\times 3$ matrix `$X$' that is referred
to as \emph{Alignment tensor}. It is given by
\begin{equation}
X_{ij} = \sum_{l_{min}}^{l_{max}} e^i(l)e^j(l) \,,
\label{eq:alignten}
\end{equation}
where ${\bf e}(l)$ is the PEV of a multipole, $l$. $Tr(X)$ denotes trace of $X$.
An unusually low value of $S_{X}$ compared to $\log(3)$ confirms violation of isotropy
over a wider multipole range.
We note that the Power entropy and the Alignment entropy are independent of each other.
The significance of statistical anisotropy is determined by comparing the data statistic
value with that of simulations and the significance is quoted by the $P$-value.
A $P$-value is defined as the probability that a random realization may yield a statistic
smaller than that seen in data.
The effective probability for a collection of PEVs with respective $P$-values less than
a reference probability `$\mathbb{P}$', is estimated using the binomial distribution of
\emph{pass} and \emph{fail} outcomes. The probability to encounter $k$ instances of
passing defined by probability $\mathbb{P}$ in $n$ trials is
\begin{equation}
f(k|n,\mathbb{P})= \frac{\mathbb{P}^{k}(1-\mathbb{P})^{n-k}n \,!}{(n-k)\,! k\,!}
\label{eq:prob}
\end{equation}
In assessing many $P$-values, we report the cumulative binomial probabilities as
\begin{equation}
f(k \ge k_{*}|n,\mathbb{P})= \sum_{k=k_{*}}^{n} f(k|n,\mathbb{P})
\label{eq:cumprob}
\end{equation}
which is the probability to see $k_*$ or more instances of passing among $n$ trials
defined by a threshold probability $\mathbb{P}$.
\section{Data Used}
\label{sec:data}
The PLANCK team has provided four foreground reduced CMB polarization maps referred to as
COMMNDER, NILC, SMICA and SEVEM \citep{Planck2015cmb} maps, named after the component
separation procedure used. Out of these four, we study only the COMMNDER and NILC as they
are full sky maps which are suitable for our study.
The other clean CMB maps viz., SMICA and SEVEM polarization solutions have a portion of
the sky removed, particularly in the galactic plane. Hence we will not use them in our
analysis. All these cleaned polarization CMB maps were estimated using all the frequency
channels aboard PLANCK that are sensitive to polarization (from $30$ to $353$ GHz).
Due to the presence of significant noise in the polarization maps and also due to the
possible residual foregrounds that may still be present even after cleaning, any signature
of large scale isotropy breakdown has to be interpreted with care. Here we use half-ring
half-difference (HRHD) maps \citep{Planck2015cmb} as noise proxy in our analysis.
The publicly available polarization maps don't include low-$l$ up to $l = 40$ owing to
systematics \citep{Planck2015cmb}.
Consequently, we only analyze multipoles $l \geq 40$. As noise contribute dominantly to
the polarization maps from PLANCK at high-$l$, we restrict our analysis up to $l=150$.
To start with, we extract the $E$-mode polarization map from the full sky Stokes $Q$ and $U$
maps from PLANCK available at a HEALPix \citep{Gorski2005} resolution of $N_{side}=1024$
and have a beam resolution given by a Gaussian beam of $FWHM=10'$ (arcmin).
We first analyze the full sky COMMNDER and NILC foreground cleaned PLANCK $E$-mode
polarization maps thus obtained.
Subsequently we generate foreground-residual minimized full sky CMB $E$-mode data maps by
following the procedure described in \citet{Samal2010,Pranati2013,Pranati2015}.
For the purposes of masking, we use the common polarization mask $UPB77$, as well as the
polarization masks specific to each component separation method employed by PLANCK \citep{Planck2015cmb}.
These masks are shown in Fig.~(\ref{fig:mask-pol}). The common polarization mask, $UPB77$
cover $\approx 77\%$ of the sky whereas the COMMANDER and NILC polarization masks cover
$\approx83\%$ and $96\%$ of the sky respectively. From now on we abbreviate the COMMANDER
polarization mask as $PMCMDR$ and the NILC polarization mask as $PMNILC$.
\begin{figure*}
\centering
\includegraphics[width=0.32\textwidth]{newplot/Common_UPB77.pdf}
~
\includegraphics[width=0.32\textwidth]{newplot/commanderPmask.pdf}
~
\includegraphics[width=0.32\textwidth]{newplot/nilcPmask.pdf}
\caption{\emph{Left} : $UPB77$ common polarization analysis mask, \emph{Middle} : COMMANDER polarization map
($PMCMDR$), \emph{Right} : NILC polarization mask ($PMNILC$). All these masks
are available at $N_{side} = 1024$.}
\label{fig:mask-pol}
\end{figure*}
A full sky CMB $E$-mode data map with minimized residual foregrounds is generated following
the steps listed below :
\begin{enumerate}
\item We first generate a full sky CMB $IQU$ random realization using the best fit theoretical
angular power spectrum ($C^{th}_l$). The best fit $C^{th}_l$ are generated using PLANCK 2015
cosmological parameters \citep{Planckpara2016a,Planckpara2016b} as input to CAMB
software \citep{Lewis2000, Howlett2012}. The values of cosmological parameters from PLANCK 2015
results that we used are baryon matter density $\Omega_{b}h^2 = 0.0222$, cold dark matter
density $\Omega_{c}h^2 = 0.1203$, neutrino energy density $\Omega_{\nu}h^2 = 0.00064$,
cosmological constant density fraction $\Omega_{\Lambda}= 0.6823$, Hubble parameter $H_{0}$
with $h=0.6712$, scalar spectral index of the primordial power spectrum $n_{s} = 0.96$,
amplitude of primordial power spectrum $A_{s} = 2.09\times 10^{-9}$, and reionization optical
depth $\tau = 0.078$. The best fit theoretical angular power spectum, $C^{th}_l$, thus obtained
from CAMB using these parameters, is employed to generate random realizations of CMB sky using
HEALPix. A CMB $IQU$ map is generated with a Gaussian beam of $FWHM=10'$ (arcmin)
at $N_{side}=1024$.
\item A cosine filter is then applied on the simulated CMB $IQU$ random realization obtained
in step~(i), to remove the large angular scales following \citet{Planck2015cmb}. The cosine
filter is defined as
\begin{equation}
w_l =
\begin{cases}
0, & \text{}\ l<l_1 \\
\frac{1}{2}\left[1-\cos\left(\pi\frac{l-l_1}{l_2-l_1}\right) \right], & \text{}\ l_1 \le l \le l_2 \\
1, & \text{}\ l_2 < l,
\end{cases}
\label{eq:cosbeam}
\end{equation}
where $l_1=20$ and $l_2=40$.
\item The $IQU$ noise proxies viz., the HRHD maps corresponding to COMMANDER/NILC component
separation method are added to the filtered CMB $IQU$ realization obtained in step~(ii).
Note that the cosine filter given by Eq.~(\ref{eq:cosbeam}) is already applied to the HRHD
polarization data maps that are made publicly available. This results in a random CMB
realization with noise levels similar to data.
\item Now, the inverse of polarization masks shown in Fig.~(\ref{fig:mask-pol}) are applied
to the random CMB $IQU$ realization added with an HRHD map.
By doing so, the resultant map will have signal only in the galactic region (and few other
regions), with rest of the sky set to zero.
Next, we apply the polarization masks of Fig.~(\ref{fig:mask-pol}), as they are, on the
CMB data $IQU$ map to remove the potentially contaminated galactic region.
Finally the two pieces that have complementary regions masked are added together to construct
a full sky CMB $IQU$ data map at $N_{side}=1024$.
\item Subsequently, we derive the $E$-mode polarization map from this composite data $IQU$
map at $N_{side}=1024$.
\end{enumerate}
Thus we have effectively minimized the residual contamination in the data CMB polarization map.
Due to random filling of the masked regions, one will get a (slightly) different value for
the statistic, compared to the \emph{true} sky. Hence our data statistic is taken as average
value of the same quantity derived from $100$ such \emph{filled} full sky COMMANDER and
NILC $IQU$ maps.
The significance of isotropy violation is estimated by comparing the (average) data statistic
value with $4000$ random CMB realizations (added with HRHD noise maps). Since we have different
HRHD maps for COMAMNDER and NILC maps, we generate two sets of $4000$ random realizations.
Simulations of CMB $IQU$ maps with noise levels similar to data are generated following step
(i) - (iii) described above. We then extract the $E$-mode polarization map from these realistic
$IQU$ maps. We note that the temperature realizations generated here are only a by-product and
have no use for us. So any operation like masking or filtering performed on $I$ map, together
with $Q$ and $U$ maps, has no relevance to our study.
\section{Results}\label{sec:results}
First we analyze the \emph{full sky} PLANCK $E$-mode polarization maps as obtained from cleaned
CMB data $IQU$ map. Later we use the \emph{filled} data maps to understand the effect of residual
contamination that may still be present in the recovered CMB sky. Recall that in order to minimize
likely foreground bias on our results, we filled the masked regions shown in Fig.~(\ref{fig:mask-pol}),
with a filtered random isotropic realization added with an HRHD map.
\subsection{Power entropy vis-a-vis Axiality of multipoles}
\label{subsec:poweren}
\subsubsection{Full sky analysis}
Here we study statistical anisotropy of the multipoles in PLANCK $E$-mode polarization
map as obtained from full sky Stokes $Q/U$ CMB maps, derived using COMMANDER and NILC
cleaning procedures.
The Power entropy, $S_p$, is computed for each multipole in the chosen multipole range
$l=40-150$ from the $E$-mode polarization maps using Eq.~(\ref{eq:powent}).
The statistical significance of the Power entropy values thus computed from data are
studied using $4000$ isotropic random CMB $E$-mode polarization maps that are appropriately
filtered (see Eq.~(\ref{eq:cosbeam}))
and added with the noise proxy of the data i.e., half-ring half-difference (HRHD) maps of
respective component separation methods.
Fig.~(\ref{fig:hist}) shows the null distribution of Power entropy for the multipoles in
the range $l = 40-150$, but at an interval of $10$ multipoles i.e., for $l=40,50,60\cdots150$
for brevity. The two histograms in each plot corresponding to the two data sets (component
separation maps) used in the analysis and the two vertical lines indicate the respective
data values.
In Fig.~(\ref{fig:counterplot1}), we show the Power entropy, $S_p$, values from data
for all the multipoles in the range $l=40-150$. The data values are denoted by
\emph{red} and \emph{blue} points, where the blue ones correspond to those multipoles
whose Power entropy has a probability of $P\leq 5\%$.
Also plotted are $90\%$, $95\%$ and $99\%$ confidence contours of distribution of Power entropy,
as obtained using 4000 simulations, with a \emph{magenta} line, \emph{gold} and \emph{cyan} colour
bands respectively. The rugged nature of the distribution can be understood from the fact
that we used the same HRHD map in our simulations to mimic the data noise levels.
\begin{figure*}
\centering
\includegraphics[width=0.98\textwidth]{newplot/histogram_sim_power_entropy.pdf}
\caption{Empirical distribution of Power entropy `$S_p$', obtained from $4000$ simulated
$E$-maps with appropriate noise, for the multipole range $40\le l \le 150$, but
shown at intervals of $10$ multipoles. The \emph{blue} and \emph{red} color
histograms in each plot correspond to COMMANDER and NILC simulations. Similarly
the vertical lines in respective colors indicate the data value from full sky
$E$-mode polarization maps obtained from COMMANDER and NILC Stokes $Q/U$ maps
as provided.}
\label{fig:hist}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{newplot/power_entropy_data_fullsky.pdf}
\caption{Power entropy ($S_p$) values from full sky COMMANDER (\emph{left}) and
NILC (\emph{right}) $E$-polarization maps over the range $l=40-150$ are
shown here.
The data $S_p$ values in each plot are shown as \emph{red} and \emph{blue}
dots, where the \emph{blue} points highlight those multioles whose Power
entropy has a $P$-value less than $5\%$.
The magenta line, gold and cyan bands represent $90\%$, $95\%$ and $99\%$
confidence contours estimated from $4000$ simulations.}
\label{fig:counterplot1}
\end{figure*}
The list of multipoles which are found to be statistically anisotropic are listed
in second column of Table~(\ref{tab:tab1}) corresponding to each PLANCK component
separated polarization map we used.
We find that there are $k_{data}=9$ and $21$ number of multipoles which have $P \le 0.05$
in the range we analyzed in COMMANDER and NILC $E$-mode polarization maps respectively.
The total number of independent trials for the range $40 \le l \le 150$ is $111$.
Following Eq.~(\ref{eq:cumprob}), we can compute the cumulative probability of finding
$k_{data}$ or more instances of statistically anisotropic modes among all the modes analyzed.
Note that we defined the criteria of pass or fail with a reference probability of $\mathbb{P}=0.05$.
Thus, from the binomial distribution, the {\it cumulative} probability, $f(k \geq k_{data}|111, \mathbb{P}=0.05)$
are found to be $0.104$ and $1.390 \times 10^{-7}$, respectively for COMMANDER
and NILC maps. These are also listed in Table~(\ref{tab:tab1}), in the \emph{third} column.
As is obvious, we find that the cumulative probability for the observed Power entropy in NILC
$E$-mode polarization map is very small compared to that of COMMANDER map. This is so owing
to more number of multipoles being axial in NILC map at the level of $2\sigma$ or more
than in COMMANDER map.
\begin{table}
\scriptsize
\centering
\begin{tabular}{c l c}
\hline
\hline
Map & Multipoles & Cumulative \\
& & Probability\\
\hline
\hline
& \\
COMMANDER & 52, 57, 60, 65, 76, 92, 95, & $ 0.104 $ \\
& 97, 146 & \\
& & \\
NILC & 44, 45, 51, 54, 56, 60, 66, & $1.390 \times 10^{-7}$ \\
& 78, 81, 90, 92, 95, 97, 98, & \\
& 100, 102, 104, 108, 122, & \\
& 143, 147 & \\
& & \\
\hline
\hline
\end{tabular}
\caption{List of multipoles with probability $P \le 0.05$ for Power entropy from PLANCK's
full sky cleaned $E$-mode polarization maps as indicated are given in \emph{second} column.
The cumulative probability for finding the observed number of statistically anisotropic modes
in individual cleaned maps with $P\leq0.05$ are furnished in the \emph{third} column.}
\label{tab:tab1}
\end{table}
From Fig.~(\ref{fig:counterplot1}) or Table~(\ref{tab:tab1}) we see that there is a significant
indication of violation of statistical isotropy in the PLANCK polarization maps we studied.
One may argue that this may be arising due to residual foregrounds that are potentially present
in the cleaned maps. In the next section we will try to minimize the effect of this residual
contamination following the procedure described in section~\ref{sec:data} and re-evaluate the significances
reported here.
\subsubsection{Understanding the effect of galactic residuals}
In this section, we present the result obtained from the full sky $E$-mode data polarization
maps constructed as discussed in section~\ref{sec:data}.
We first apply the common $UPB77$ mask on the cleaned PLANCK $IQU$ polarization maps and
construct the full sky CMB data $IQU$ maps by filling the masked region with a filtered
random CMB realization (Eq.~(\ref{eq:cosbeam}) added with HRHD noise map. We then extract
full sky $E$-mode polarization maps from the filled $Q/U$ maps, and use these to study the
statistical isotropy of multipoles in the range $l=40-150$.
We also use component separation specific polarization masks $PMCMDR$ and $PMNILC$ corresponding
to COMMANDER and NILC procedures respectively to construct full sky $E$-mode polarization maps
following the same procedure.
The exercise is then repeated with these maps to reassess the statistics presented before.
As explained in section~\ref{sec:data}, the data statistic is taken as average value of
the statistic derived from $100$ such randomly filled full sky COMMANDER and NILC $IQU$ maps.
We note that such a filling procedure is expected to reduce any signal of statistical anisotropy
if present, because of the randomizing effect of the filling.
The list of multipoles having $P \le 0.05$ in various cases considered are given in
\emph{second} column of Table~(\ref{tab:tab2}).
Now, using the common UPB77 mask for filling, we find that there are only $5$ and $12$ multipoles
having a probability of $P \leq 5\%$ compared to that of simulations for COMMANDER and NILC
maps respectively.
Similarly, by filling the masked region with the individual masks available with each
component separation map (PMCMDR and PMNILC), we find that there are $7$ and $12$ multipoles
with $P \le 0.05$ for COMMANDER and NILC $E$-mode polarization maps respectively.
Thus applying various masks reveals the (in)stability of anomalous multipoles against
galactic cuts and potential foreground contamination.
The corresponding cumulative probabilities in various cases are listed in the \emph{third} column
of Table~(\ref{tab:tab2}).
\begin{table}
\scriptsize
\centering
\begin{tabular}{c l c}
\hline
\hline
Map/Mask & Multipoles & Cumulative \\
& & Probability \\
\hline
\hline
& & \\
COMMANDER/ & 58, 76, 98, 108, 146 & $ 0.656 $ \\
UPB77 & & \\
& & \\
COMMANDER/ & 57, 58, 76, 95, 98, 108, & $ 0.320 $ \\
PMCMDR & 146 & \\
& & \\
NILC/ & 45, 51, 54, 56, 60, 78, & $9.726 \times 10^{-3}$ \\
UPB77 & 81, 84, 92, 95, 98, 108 & \\
& & \\
NILC/ & 56, 60, 66, 78, 84, 90, & $9.726 \times 10^{-3}$ \\
PMNILC & 92, 95, 104, 108, 122, & \\
& 149 & \\
& & \\
\hline
\hline
\end{tabular}
\caption{ Same as Table~(\ref{tab:tab1}), but for full sky $E$-polarization maps constructed by filling
the potentially contaminated galactic region as defined by various masks (see section~\ref{sec:data}
for more details).
The two component separated polarization maps considered here, were studied using both the
individual polarization masks - the PMCMDR and PMNILC masks, as well as the common UPB77 mask
shown in Fig.~(\ref{fig:mask-pol}).
The \emph{second} column lists the anomalous multipoles that are outside the $2\sigma$ confidence
level, and the \emph{third} column lists the cumulative probabilities for finding the observed
number or more of the statistically anisotropic multipoles in the range $l=40-150$, that have
a $P$-value $\leq 5\%$.}
\label{tab:tab2}
\end{table}
From Table~(\ref{tab:tab2}), one notices that the cumulative probabilities are larger in the
\emph{filled} sky $E$-mode polarization maps than cleaned \emph{full} sky maps, evidently
owing to the decrease in number of anomalous multipoles in the later case.
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{newplot/power_entropy_data_filledsky.pdf}
\caption{Plotted here are the Power entropy ($S_p$) values as obtained from full sky
PLANCK's COMMANDER and NILC $E$-maps constructed after filling the masked
sky defined by various polarization masks shown in Fig.~(\ref{fig:mask-pol}).
The \emph{first} and \emph{second} columns correspond to the Power entropy
for the range $l=40-150$ as obtained from COMMANDER and NILC $E$-polarization
maps. The \emph{top} and \emph{bottom} figures in each column correspond to
the use of UPB77 common polarization mask and the component separation
specific PMCMDR/PMNILC mask on respective cleaned maps.
The color coding of data points and confidence contours is same as in
Fig.~(\ref{fig:counterplot1}).}
\label{fig:counterplot2}
\end{figure*}
In Fig.~(\ref{fig:counterplot2}), we show the Power entropy values from data in comparison
to those derived from simulations. As mentioned earlier in section~\ref{sec:data}, the observed
$S_p$ values for each multipole are obtained as average value of the statistic over $100$ random
fillings of the data to construct full sky $E$-mode polarization maps.
The contours of $90\%$, $95\%$ and $99\%$ confidence levels from simulations are shown as
\emph{magenta} line, \emph{gold} band, and \emph{cyan} band respectively. The data points
are shown as \emph{red} dots, while those which are outside the $2\sigma$ contour are denoted by
\emph{blue} points. This contour plot neatly highlights various multipoles which are inconsistent
with the isotropic predictions.
So, we may now say that presence of residual foregrounds indeed had an effect on our isotropy
test when full sky polarization maps are used as provided.
In both COMMANDER and NILC maps, the number of anomalous multipoles nearly reduced by half
when full sky maps are constructed using UPB77 mask. However since noise is dominant in the
PLANCK polarization maps, the stability of these modes can only be validated in the future.
\subsection{Alignments across multipoles}
\label{subsec:AlignEntropy}
In this section, we discuss alignments among mutlipoles in the chosen multipole range
using Alignment entropy, $S_X$, as defined in Eq.~(\ref{eq:alignent}).
We divide the chosen range $l=40-150$ into $10$ multipole bins having $11$ multipoles
per bin. Thus the multipole bins we analyze are $l=40-50$, $51-61$, $62-72$, $73-83$,
$84-94$, $95-105$, $106-116$, $117-127$, $128-138$ and $139-149$.
The statistic, $S_X$, is computed for each multipole bin and its significance is
estimated using simulations that are generated as discussed in section~\ref{sec:data}.
\subsubsection{Full sky analysis}
Here we probe for any coherent alignments across multipoles in $E$-mode polarization
maps as obtained using PLANCK's full sky Stokes $Q/U$ CMB maps, derived using COMMANDER
and NILC cleaning procedures. In each multipole bin of the data maps, Alignment entropy
is calculated. The statistical significance of $S_X$ are obtained by comparing the
data statistic with $4000$ isotropic random CMB $E$-mode filtered polarization maps
added with the noise proxy (HRHD map) of respective component separation methods.
The list of multipole bins and the significance of $S_X$ for these bins are listed
in \emph{second} and \emph{third} column of Table~(\ref{tab:tab5}) respectively.
As we can see from that table,
the multipole bins $40-50$, $84-94$, $95-105$, $106-116$, $117-127$, and $139-149$
have a $P$-value $\le 0.05$ for Alignment entropy. We extract the common alignment
vector from PEVs for a bundle of multipoles using the Alignment tensor, $X$, defined in
Eq.~(\ref{eq:alignten}). The common alignment vector is taken as the eigenvector
corresponding to the largest eigenvalue of the Alignment tensor matrix.
The common alignment vector direction of those bins in galactic co-ordinates, $(l,b)$,
having $P$-value $\leq 5\%$ for $S_X$ are also listed in Table~(\ref{tab:tab5}),
in the \emph{fourth} column.
We notice that these vectors almost lie in the galactic region. The effect of
the presence of residual contamination on these vectors will be assessed in the
next section.
For ease of comparison, we also show the observed significances of $S_X$
in various multipole bins from full sky COMMANDER and NILC $E$-maps in
Fig.~(\ref{fig:algn-entrp-fullsky-E}).
The Alignment entropy from data corresponding to some of the multipole bins was always
smaller than the simulations. Therefore we denoted those histogram bars with a triangle
at the top to indicate that the significance of data statistic is $<1/4000$.
\begin{table}
\scriptsize
\centering
\begin{tabular}{|c c c c}
\hline
\hline
Map & Multipole Range & $P$-value & $(l,b)$ \\
\hline
\hline
COMMANDER & $40-50$ & $<1/4000$ & $(87^\circ,27^\circ)$\\
& $51-61$ & $0.19$ & - \\
& $62-72$ & $0.23$ & - \\
& $73-83$ & $0.15$ & - \\
& $84-94$ & $<1/4000$ & $(95^\circ,19^\circ)$\\
& $95-105$ & $<1/4000$ & $(96^\circ,30^\circ)$\\
& $106-116$ & $0.013$ & $(75^\circ,18^\circ)$ \\
& $117-127$ & $0.002$ & $(103^\circ,33^\circ)$\\
& $128-138$ & $0.424$ & -\\
& $139-149$ & $0.021$ & $(90^\circ,26^\circ)$\\
\hline
NILC & $40-50$ & $0.008$ & $(87^\circ,24^\circ)$\\
& $51-61$ & $0.092$ & -\\
& $62-72$ & $0.083$ & -\\
& $73-83$ & $0.001$ & $(99^\circ,29^\circ)$\\
& $84-94$ & $<1/4000$ & $(93^\circ,31^\circ)$\\
& $95-105$ & $<1/4000$ & $(95^\circ,29^\circ)$\\
& $106-116$ & $<1/4000$ & $(82^\circ,21^\circ)$\\
& $117-127$ & $0.02$ & $(96^\circ,36^\circ)$\\
& $128-138$ & $0.12$ & -\\
& $139-149$ & $0.002$ & $(98^\circ,46^\circ)$\\
\hline
\hline
\end{tabular}
\caption{Significance of Alignment entropy ($S_X$) and direction of common alignment axis
obtained using Alignment tensor ($X$) corresponding to various multipole bins
of the cleaned $E$-mode polarization maps, derived from full sky Stokes $Q/U$
COMMANDER and NILC maps, are listed here.
The range $l=40-150$ is divided into $10$ multipole bins. The $P$-value of $S_X$
for various bins are given in the \emph{third} column.
Where none of the simulations yield a value less than the data statistic,
it's significance is listed as $<1/4000$.
The direction of the common alignment axis for those bins
whose $P$-value is $\leq 0.05$ are given in the \emph{fourth} column.}
\label{tab:tab5}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{newplot/hist_pval_alignentrp_fullsky_new.pdf}
\caption{$P$-values of Alignment entropy, $S_X$, corresponding to various multipole bins in the
range $l=40-150$ (divided into 10 bin) are plotted here.
The two bars at each mutlipole bin as indicated correspond to the significances
of Alignment entropy for that bin from CMB $E$-mode polarization maps derived
from full sky Stokes $Q/U$ COMMANDER and NILC maps.
Those histogram bars which have a triangle at the top, denote a significance
$<1/4000$.}
\label{fig:algn-entrp-fullsky-E}
\end{figure}
\subsubsection{Filled sky analysis}
In order to understand the likely effect of residual contamination in the PLANCK full sky
CMB polarization maps on our collective alignments' study, here we use filled $E$-mode
data polarization map, generated as discussed in section~\ref{sec:data}.
The common $UPB77$ mask is first applied on the cleaned PLANCK's COMMANDER and NILC $IQU$
maps and the statistic $S_X$ is obtained as mean value of the same quantity from filled
$IQU$ maps constructed using 100 random realizations of CMB with HRHD noise proxy
to fill the masked portion in the data maps.
The filling is also performed using the other two polarization masks shown in Fig.~(\ref{fig:mask-pol}),
viz., the component specific $PMCMDR$ and $PMNILC$ masks corresponding to COMMANDER and NILC
foreground cleaning schemes.
The statistical significance of the observed value of Alignment entropy is studied using
$4000$ CMB $E$-mode realizations with noise.
\begin{table}
\scriptsize
\centering
\begin{tabular}{c c c c}
\hline
\hline
Map/Mask & Multipole bin & $P$-value & $(l,b)$ \\
\hline
\hline
COMMANDER/ & $40-50$ & $0.0005$ & $(85^\circ,24^\circ)$ \\
UPB77 & $51-61$ & $0.106$ & - \\
& $62-72$ & $0.273$ & - \\
& $73-83$ & $0.056$ & - \\
& $84-94$ & $0.002$ & $(91^\circ,22^\circ)$ \\
& $95-105$ & $0.025$ & $(98^\circ,47^\circ)$ \\
& $106-116$ & $0.014$ & $(72^\circ,29^\circ)$ \\
& $117-127$ & $0.064$ & - \\
& $128-138$ & $0.123$ & - \\
& $139-149$ & $0.012$ & $(88^\circ,34^\circ)$\\
\hline
COMMANDER/ & $40-50$ & $0.0005$ & $(85^\circ,31^\circ)$\\
PMCMDR & $51-61$ & $0.055$ & - \\
& $62-72$ & $0.13$ & - \\
& $73-83$ & $0.58$ & - \\
& $84-94$ & $0.0013$ & $(89^\circ,27^\circ)$ \\
& $95-105$ & $0.007$ & $(98^\circ,40^\circ)$ \\
& $106-116$ & $0.0008$ & $(72^\circ,28^\circ)$ \\
& $117-127$ & $0.05$ & - \\
& $128-138$ & $0.156$ & - \\
& $139-149$ & $0.0234$ & $(91^\circ,33^\circ)$\\
\hline
NILC/ & $40-50$ & $<1/4000$ & $(93^\circ,26^\circ)$ \\
UPB77 & $51-61$ & $0.007$ & $(101^\circ,35^\circ)$\\
& $62-72$ & $0.342$ & - \\
& $73-83$ & $0.137$ & - \\
& $84-94$ & $<1/4000$ & $(92^\circ,16^\circ)$ \\
& $95-105$ & $0.0003$ & $(100^\circ,36^\circ)$ \\
& $106-116$ & $<1/4000$ & $(84^\circ,27^\circ)$ \\
& $117-127$ & $0.002$ & $(99^\circ,25^\circ)$ \\
& $128-138$ & $0.002$ & $(92^\circ,38^\circ)$ \\
& $139-149$ & $0.002$ & $(90^\circ,34^\circ)$ \\
\hline
NILC/ & $40-50$ & $0.003$ & $(90^\circ,26^\circ)$\\
PMNILC & $51-61$ & $0.002$ & $(102^\circ,38^\circ)$ \\
& $62-72$ & $0.47$ & - \\
& $73-83$ & $0.03$ & $(105^\circ,28^\circ)$\\
& $84-94$ & $<1/4000$ & $(97^\circ,29^\circ)$ \\
& $95-105$ & $<1/4000$ & $(100^\circ,31^\circ)$ \\
& $106-116$ & $<1/4000$ & $(80^\circ,20^\circ)$ \\
& $117-127$ & $0.001$ & $(98^\circ,27^\circ)$ \\
& $128-138$ & $0.08$ & - \\
& $139-149$ & $0.002$ & $(99^\circ,43^\circ)$ \\
\hline
\hline
\end{tabular}
\caption{Same as Table~(\ref{tab:tab5}), but the significances of $S_X$ and direction
of common alignment axis for various bins correspond to filled sky polarization
maps as indicated (see text for detials).
The $P$-values are listed in \emph{third} column and the common alignment vectors
for only those bins whose Alignment entropy, $S_X$, is found to be anomalous at
$2\sigma$ level are listed in \emph{fourth} column.}
\label{tab:tab7}
\end{table}
The list of multipole bins and the
significance of $S_X$ from these bins are given in Table~(\ref{tab:tab7}). Now, we find
that the multipole bins, $l=40-50$, $84-94$, $95-105$, $106-116$, and $139-149$ have
$P\leq 0.05$ using the UPB77 mask, and $l=40-50$, $84-94$, $95-105$, $106-116$, $117-127$,
and $139-149$ multipole bins have $P$-value $\leq 0.05$ with PMCMDR mask on COMMANDER
polarization map. Similarly for NILC polarization map, the multipole bins that are anomalous
at $2\sigma$ level are $l=40-50$, $51-61$, $84-94$, $95-105$, $106-116$, $117-127$, $128-138$,
and $139-149$ when UPB77 mask was used, and $l=40-50$, $51-61$, $73-83$, $84-94$, $95-105$,
$106-116$, $117-127$, $128-138$, and $139-149$ bins when PMNILC mask was used.
The collective alignment vector direction for various bins obtained using Alignment tensor, $X$,
(Eq.~(\ref{eq:alignten})) which are found to have a $P$-value $\leq 5\%$ for $S_X$ are
tabulated in the last column of Table~(\ref{tab:tab7}). Even in this case, we see that
these collective alignment vectors lie closer to the galactic region. Hence these modes
found to be anomalous may still be affected by residual foreground bias.
In Fig.~(\ref{fig:algn-entrp-filledky-E}), the $P$-values of $S_X$ found in the filled
sky case for various multipole bins and masks used are shown as histograms for ready comparison.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{newplot/hist_pval_alignentrp_filledsky_new.pdf}
\caption{Same as Fig.~(\ref{fig:algn-entrp-fullsky-E}), but for the filled sky
CMB polarization maps viz, COMMANDER $E$-map by filling the masked region
defined by the common UPB77 and PMCMDR masks, and the NILC $E$-map where
the inside of the masking portion defined by UPB77 and PMNILC
masks is filled, using a random CMB realization and HRHD noise proxy.
The data statistic is computed as average of the same quantity extracted
from 100 such random fillings of the data, which is then used to compute significances.
Here also, those multipole bins for a given component separation map/mask combination
where the significance was found to be $<1/4000$, their histogram bar is plotted with
a triangle at the top.}
\label{fig:algn-entrp-filledky-E}
\end{figure}
\section{Conclusion}
\label{sec:con}
In the present work, we studied violation of isotropy of various modes that are available
in the cleaned CMB polarization maps from PLANCK full mission data release. Specifically
we scrutinized the multipole range $l=40-150$ of the $E$-mode CMB maps derived using the
COMMANDER and NILC $Q/U$ polarization maps. We applied our symmetry based \emph{Power tensor}
method to test statistical (an)isotropy of these maps.
A set of 4000 simulations were generated using the theoretical angular power spectrum
obtained using the best fit cosmological parameters from PLANCK 2015 release in the CAMB
software package. These realizations were smoothed with appropriate beam window function,
and filtered using a high-pass cosine filter to retain only those modes in simulations,
that are currently made available in the data.
The half-ring half-difference (HRHD) maps of respective component separation methods were
taken as noise proxy, and added with the smoothed, filtered $E$-mode random CMB realizations
to generate mock observed maps.
In order to understand the effect of potential bias due to residual foregrounds in the full sky
cleaned polarization maps on our isotropy studies, we generated another set of full sky
maps where we filled part of the sky that is omitted by the UPB77 polarization mask with
100 appropriate random CMB maps with noise. All the statistics computed in the filled sky
case are taken as average value of the same quantity over these 100 \emph{filled} full sky
$E$-mode data maps. We also used the individual polarization masks from COMMANDER and NILC
foreground removal methods. We reiterate that the filling procedure we employed is only expected
to lower any signal of statistical anisotropy if present, because of the randomizing nature
of the filling process.
We note the following observations. The number of anomalous multipoles that indicate isotropy
violation at the level of $2\sigma$ in the full sky COMMNADER and NILC polarization maps
are found to be $9$ and $21$ respectively. However when this same range $l=40-150$ is analyzed
using filled sky maps constructed using the conservative UPB77 polarization mask, the number of
anomalous modes have reduced to nearly half. The number remains albeit the same when component
specific polarization masks are used. Thus we may say that the galactic residuals indeed have an
effect on our test of isotropy of various multipoles. It is interesting to note that the
number of anomalous multipoles with $P$-value $\leq 0.05$ are more in NILC CMB $E$-map than
in the COMMANDER map. This observation is particularly interesting given that the recovered
CMB signal using NILC procedure is supposed to be very reliable over a much larger
fraction of the sky than the COMMANDER map.
The respective masking fractions of COMMANDER and NILC polarization masks are $\approx 83\%$
and $96\%$. The effective probability of finding the observed number of statistically anisotropic
modes in the range $l=40-150$ in NILC map using various masks is correspondingly low.
We then studied alignments across multipoles using \emph{Alignment entropy} over the
chosen range $l=40-150$, divided into 10 blocks with 11 multipole per bin.
Here also we analyzed full sky as well as filled sky maps for coherent alignments
across multipoles using 4000 simulations. All three polarization masks considered
in the preceding analysis were applied to data $E$-maps to understand any foreground biases.
We find a tentative evidence for collective alignment in some of the multipole blocks.
However, we also found that the direction of these common alignment axes lie closer to the
galactic plane in both full sky and filled sky cases. Hence the effect of galactic bias,
even after filling the potentially contaminated regions, may still be significant. In these
alignment tests as well, we find that more number of multipole bins are anomalous in NILC map
compared to COMMANDER CMB $E$-map.
Thus, the modes currently available in the polarization maps we analyzed appear to be sensitive
to various galactic cuts. Further, the number and actual modes that are anomalous change, by applying
different masks with few common multipoles surviving among them. These
multipoles may also be effected by noise, given that the noise levels are significant in PLANCK
polarization measurements. In light of these, an analysis of improved polarization maps from PLANCK
in a future data release can only confirm our findings. In this work, however, we didn't
make any attempt to identify the exact cause of the observed isotropy violation in polarization maps.
\section*{Acknowledgements}
We thank the anonymous referee for a careful reading of our manuscript and
the suggestions made, which greatly helped in improving clarity and
overall presentation of our work.
Some of the results in the current work were derived using the publicly available
HEALPix package\footnote{\url{https://healpix.jpl.nasa.gov/}}.
We also acknowledge the use of CAMB\footnote{\url{http://camb.info/}}, a freely
available Boltzmann solver for CMB anisotropies.
Part of the results presented here are based on observations obtained with
PLANCK\footnote{\url{http://www.esa.int/Planck}}, an ESA science mission with instruments
and contributions directly funded by ESA Member States, NASA, and Canada.
|
\section{Introduction}
Heterogeneous networks (HetNets), consisting of macro base stations (MBSs) and ultra-densely deployed small cell base stations (SBSs), are envisioned as the dominant theme to meet the 1000$\times$ capacity enhancement in 5G networks and beyond {{ \cite{Ge16_5g_ultra_dense_WC, ultra_dense_SC_2015}.}}
With network further densified, deploying ideal backhaul with unconstrained capacity for each small cell may be impractical, due the unacceptably high costs of deployment and operation {{ \cite{Yu16_backhaul_cache_GC, Ding12_predoding_backhaul}. }}
Thus, one of the key problems towards 5G is to reduce the required backhaul capacity while keeping the system capacity.
Mobile edge caching provides a promising solution to address the problem, by exploiting the content information \cite{GreenDelivery_SZhou_2015, Yang16_catalyzing_cloud_fog_SDN}.
As the requested content of mobile users, e.g., video, may show high similarity, caching popular contents at base stations can effectively alleviate the backhaul pressure and enhance network service capability \cite{Qiao_proactive_mmWave}.
Meanwhile, the delay performance can be significantly improved, with service demands accommodated locally.
Since the study on mobile edge caching is still nascent, many research issues need to be addressed, such as architecture design \cite{Tandon16_fog_mag}, content placement \cite{Li16_cache_min_weighted_load_ICC, Bharath16_learning_caching_TC} and update \cite{Akon13_content_update}.
However, the caching deployment is overlooked in the existing literature.
Specifically, the fundamental problem of cache deployment is to optimize the cache sizes of different BSs in HetNets, so as to minimize network deployment and operational costs while guaranteeing quality of service (QoS) performance.
The basic tradeoff for cache deployment exists between caching efficiency and spectrum efficiency.
{{On one hand, the contents cached at MBSs can serve more users due to the large cell coverage, providing high caching efficiency.
On the other hand, the densely deployed SBS tier is more likely to be backhaul-constrained, as extensive spatial spectrum reuse introduces substantial access traffic.
As a result, deploying more cache instances at SBSs can narrow the gap between backhaul and radio access capacities, and thus improve spectrum efficiency systematically.
In this regard, cache instances should be deployed appropriately, such that network resources can be balanced and fully utilized \cite{Liu16_EE_cache_JSAC, Liu15_MIMO_cache_TSP}.}}
However, the cache deployment problem is challenging, as different cache size also influences the traffic load distributions across the network.
For example, more traffic needs to be served by MBSs when the MBS cache size increases, changing the loads of both radio access and backhaul of MBS and SBS tiers.
Therefore, load balancing should be also considered to avoid problems like service outage and resource under-utilization.
To this end, traffic steering can be leveraged to tune load distribution, and jointly optimized with cache deployment \cite{NZhang16_LTEU}.
In this paper, the cache deployment problem is investigated in two-tier HetNets, where each SBS caches the most popular files while each MBS caches the less popular ones (i.e., hierarchical caching).
If cached at the associated MBSs or SBSs, the requested contents will be directly delivered to mobile users through radio access, i.e., content hit.
Otherwise, the requested contents will be delivered through remote file fetching via backhaul connections, i.e., content miss.
For a given cache deployment budget, we maximize network capacity while guaranteeing the average file transmission rates, by jointly optimizing the MBS/SBS cache sizes and the inter-tier traffic steering ratio of content miss users.
However, the problem is of great challenge due to the transmission rate requirements.
Specifically, file transmission rates depend on both radio and backhaul access conditions, which should account for multi-randomness of traffic load, user location, channel fading and network topology.
Through stochastic geometry analysis, the lower bound of average file transmission rates are derived in closed form, based on which the cache deployment problem is simplified and numerical results can be obtained.
To offer insights into practical network design, we then focus on the scenario when the MBSs have sufficiently large backhaul capacity while the SBS tier is backhaul constrained.
The optimal cache deployment is obtained, which is threshold-based with respect to the network cache budget.
When the cache budget is smaller than certain threshold, all the cache instances should be deployed at SBSs to maximize network capacity.
When the cache budget exceeds the threshold, the cache deployment problem has multiple optimal solutions to achieve maximal network capacity, and we find the one which can simultaneously maximize content hit rate.
In fact, cache budget threshold can be interpreted as the deficiency of SBS backhaul, i.e., the minimal cache budget required to match the backhaul and radio resources.
Moreover, the threshold characterizes the trading relationship between backhaul and cache capacities, which can be applied to cost-effective network deployment.
The contributions of this paper are summarized as follows:
\begin{enumerate}
\item The average file transmission rates in large-scale cache-enabled HetNets are analyzed theoretically, considering the constraints of both backhaul capacities and radio resources;
\item The cache deployment is optimized in HetNets, which maximizes QoS-guaranteed network capacity with the given cache budget;
\item The inter-tier traffic steering is jointly optimized to balance the loads of MBS and SBS tiers, considering the influence of cache deployment on traffic distributions;
\item The proposed method can provide the cost-optimal combination of backhaul and radio resource provisioning, which can be applied to practical cache-enabled HetNet deployment.
\end{enumerate}
The remaining of this paper is organized as follows. Firstly, related work on mobile edge caching is reviewed in Section~\ref{sec_review}.
Then, the system model is presented in Section~\ref{sec_system_model}, and the cache deployment problem is formulated in Section~\ref{sec_formulation}. In Section~\ref{sec_capacity_analysis}, the QoS-constrained network capacity is obtained, based on which the optimal cache deployment is analyzed in Section~\ref{sec_solution}. The analytical results are validated through extensive simulations in Section~\ref{sec_simulation}, followed by the cost-effective network deployment illustrations with numerical results. Finally, Section~\ref{sec_conclusions} summarizes the work and discusses future research topics.
\section{Literature Review}
\label{sec_review}
Content caching at mobile edge networks is considered as a promising solution to cope with the mismatch between explosive mobile video traffic and limited backhaul/wireless capacity, which has drawn increasing attention recently.
Cache-enabled 5G network architectures have been designed in \cite{Wang14_cache_framework_wireless_mag, Bastug14_cache_framework_BS_D2D_mag}, which were shown to have a great potential to reduce mobile traffic through trace-driven simulations.
The performance of cache-enabled networks has also been analyzed theoretically, which was demonstrated to be more spectrum-efficient compared with the conventional HetNets in backhaul-constrained cases \cite{Liu16_EE_cache_JSAC}.
{{Meanwhile, effective cache placement schemes have been devised with respect to different optimization objectives, such as maximizing content hit rate \cite{Akon12_cache_BBCR, Serbetci16_multi_PPP_hit_WCNC, Chen16_probabilistic_cache_analysis}, reducing file downloading delay \cite{Chang16_ICC_cache_matching, Cui16_cache_place_delay, Liu16_cache_delay_distributed_algorithm_ICC, bacstuug2016delay}, enhancing user quality of experience (QoE) \cite{Sun_15_cache_QoE_vehicular_streaming}, improving mobility support \cite{Wang16_cache_mobility_mag, Qiao15_video_buffer}, and minimizing specific cost functions \cite{Gregori16_D2D_caching_JSAC, Tao16_TWC_caching_beamforming}.}}
Although the existing cache placement schemes were designed based on the predefined cache size for each BS, studies on cache deployment were quite limited.
In the very recent work \cite{Ghoreishi16_cache_multi_layer_video_provisioning_ICC}, the storage costs of different network entities (like remote servers, gateways, and BSs) have been considered, and a multi-layered cache deployment scheme was proposed to maximize the ratio of content hit rate to storage cost.
The performances of BS-caching and gateway-caching have been compared in \cite{Han16_cache_compare_core_BS}, based on which the cache deployment was optimized to achieve Pareto optimal spectrum efficiency and content hit rate.
The BS cache sizes are optimized to maximize the minimal user success probability, under the constraints of backhaul capacity and cache deployment budget \cite{Peng16_cache_size_ICC}.
Insightful as it is, the algorithm in \cite{Peng16_cache_size_ICC} mainly focused on small-scale networks.
Different from existing work, this paper investigates the cache deployment problem in large-scale HetNets for the first time, aiming at maximizing network capacity while meeting the QoS requirements in terms of file transmission rate.
Meanwhile, the cache sizes of different BSs are jointly optimized with inter-tier traffic steering.
The analytical results have taken into account the multi-randomness of network topology, traffic distribution and channel fading, which can provide a guideline for practical network design with mobile edge caching.
\section{System Model}
\label{sec_system_model}
In this section, we present system model and the hierarchical caching framework, with important notations summarized in Table~\ref{tab_notation}.
\subsection{Cache-Enabled Heterogeneous Network Architecture}
In 5G HetNets, MBSs are responsible for network coverage with control signaling, whereas SBSs are expected to be densely deployed to boost network capacity in a ``plug-and-play'' manner.
The topology of MBSs are modeled as regular hexagonal cells with density $\rho_\mathrm{m}$, while the distribution of SBSs are modeled as Poisson Point Process (PPP) of density $\rho_\mathrm{s}$.
MBSs and SBSs use orthogonal spectrum bands to avoid inter-tier interference, and the spectrum reuse factor within each tier is set to be 1.
Denote by $W_\mathrm{m}$ and $W_\mathrm{s}$ the bandwidths available to each MBS and SBS, respectively.
Both MBSs and SBSs are connected with core network through wired backhauls, with capacities denoted as $U_\mathrm{MBH}$ and $U_\mathrm{SBH}$, respectively.
The distribution of active users is modeled as a PPP of density $\lambda$, independent of the location of MBSs and SBSs.
The service process is illustrated as Fig.~\ref{fig_scenario}.
Each user keeps dual connectivity with an MBS and an SBS \cite{LTE_standard, Ismail13_multihoming_twc, Song12_multi_radio_TWC}, where the MBS (and SBS) which provides the highest average intra-tier signal to interference ratio (SINR) is selected for association.
The coverage area of each MBS is a hexagonal cell with side length of $D_\mathrm{m}$ ($\frac{3\sqrt{3}}{2} {D_\mathrm{m}}^2 = \frac{1}{\rho_\mathrm{m}}$), and small cells form the Voronoi tessellation, as shown in Fig.~\ref{fig_scenario}.
As for the service process, mobile users can be directly served through radio access networks if the required files are cached at the MBS or SBS tiers (i.e., content-hit users).
Instead, content-miss users will randomly choose the associated SBS or MBS with probability $\varphi$ and $1-\varphi$, and the chosen SBS/MBS needs to fetch the required file from remote servers via backhaul.
Define $\varphi$ as the inter-tier \emph{traffic steering ratio}, which influences the load of MBS and SBS tiers.
\begin{figure}[!t]
\centering
\includegraphics[width=3.2in]{scenario_separation}
\caption{Cache-enabled heterogeneous networks.}
\label{fig_scenario}
\end{figure}
\subsection{Hierarchical Caching}
\begin{table}[!t]
\caption{Important notations}
\label{tab_notation}
\centering
\begin{tabular}{cc|cc}
\hline
\hline
$\rho_\mathrm{m}$ & MBS density & $\rho_\mathrm{s}$ & SBS density \\
$W_\mathrm{m}$ & MBS bandwidth & $W_\mathrm{s}$ & SBS bandwidth \\
$U_\mathrm{MBH}$ & MBS backhaul capacity & $U_\mathrm{SBH}$ & SBS backhaul capacity \\
$F$ & number of files & $q_f$ & popularity of file-$f$ \\
$C_\mathrm{m}$ & MBS cache size & $C_\mathrm{s}$ & SBS cache size \\
$P_\mathrm{Hit}^{(\mathrm{m})}$ & MBS content hit rate & $P_\mathrm{Hit}^{(\mathrm{s})}$ & SBS content hit rate \\
$C$ & network cache budget & $P_\mathrm{Hit}$ & aggregated content hit rate \\
$\lambda$ & traffic density & $\sigma^2$ & noise power\\
$\varphi$ & traffic steering ratio & $\xi_c$ & cache deployment ratio \\
$\check{R}_\mathrm{RAN}$ & required rate at RAN & $\check{R}_\mathrm{BH}$ & required rate at backhaul\\
\hline
\hline
\end{tabular}
\end{table}
Denote by $\mathcal{F}=\left\{1,2,...,f,...F\right\}$ the set of files that may be requested, and denote by $\mathcal{Q} = \left\{ q_1, q_2,...,q_f,...,q_F \right\}$ the popularity distribution ($\sum_{f=1}^{F} q_f=1$, $q_f>0$ for $f=1,2,...,F$).
Without loss of generality, we assume the files are sorted with descending popularity ($q_f \geq q_{f+1}$ for $f = 1,2,...,F-1$) and have the same size of $L$\footnote{If files have different sizes, they can be divided into the same size to conduct analysis.}.
A hierarchical content caching framework is adopted, where the SBS tier caches the most popular files while the MBS tier caches the less popular ones to increase content diversity.
Denote by $C_\mathrm{m}$ and $C_\mathrm{s}$ the cache sizes of each MBS and SBS, respectively.
Thus, files $\left\{ 1,2,...,C_\mathrm{s} \right\}$ are cached at each SBS, and files $\left\{ C_{\mathrm{s}+1}, C_{\mathrm{s}+2}, ..., C_{\mathrm{s}} + C_{\mathrm{m}} \right\}$ are cached at each MBS.
Then, the content hit rates of MBS and SBS tiers, $P_\mathrm{Hit}^{(\mathrm{m})}$ and $P_\mathrm{Hit}^{(\mathrm{s})}$, are given by
\begin{equation}
\label{eq_P_hit_s_m}
P_\mathrm{Hit}^{(\mathrm{m})} = \sum\limits_{C_\mathrm{s}+1}\limits^{C_\mathrm{m}+C_\mathrm{s}} q_f,~~~~P_\mathrm{Hit}^{(\mathrm{s})} = \sum\limits_{f=1}\limits^{C_\mathrm{s}} q_f,
\end{equation}
and total content hit rate is
\begin{equation}
\label{eq_P_hit}
P_\mathrm{Hit} = P_\mathrm{Hit}^{(\mathrm{m})} + P_\mathrm{Hit}^{(\mathrm{s})}= \sum\limits_{f=1}\limits^{C_\mathrm{m}+C_\mathrm{s}} q_f.
\end{equation}
With the dual connectivity, the equivalent cache size for each mobile user is $C_\mathrm{m}+C_\mathrm{s}$ according to Eq.~(\ref{eq_P_hit})\footnote{``Eq.'' is short for Equation, and ``Eqn.'' is short for inequation}.
{{Thus, caching the most popular $C_\mathrm{m}+C_\mathrm{s}$ files can maximize content hit rate, since each mobile user can be served only by the associated MBS or SBS with no intra-tier BS cooperation.
In addition, caching more contents at SBSs instead of MBSs can steer more users to the SBS-tier from MBSs.
As SBSs are more densely deployed than MBSs in practical networks, steering traffic to the SBS tier can fully utilize rich radio resources with inter-tier load balancing.}}
Define $C$ the network caching budget, i.e., the number of files cached per unit area:
\begin{equation}
\label{eq_cache_size}
C = \rho_\mathrm{m} C_\mathrm{m} + \rho_\mathrm{s} C_\mathrm{s}.
\end{equation}
Cache deployment determines $C_\mathrm{m}$ and $C_\mathrm{s}$ to optimize network performance, for the given network caching budget $C$.
\subsection{File Transmission Rate}
With hierarchical caching, users can be classified into four types: (1) MHU (MBS-hit-users), served by the MBS tier with cached contents; (2) SHU (SBS-hit-users), served by the SBS tier with cached contents; (3) MMU (MBS-missed-users), served by the MBS tier through backhaul file fetching; and (4) SMU (SBS-missed-users), served by the SBS tier through backhaul file fetching.
Based on the properties of PPP, the four types of users also follow independent PPPs, with densities of $P_\mathrm{Hit}^{(\mathrm{m})} \lambda$, $P_\mathrm{Hit}^{(\mathrm{s})} \lambda$, $ (1-P_\mathrm{Hit}) (1-\varphi) \lambda$, and $(1-P_\mathrm{Hit}) \varphi \lambda$, respectively.
The file transmission rates of MHUs and SHUs only depend on the radio access (i.e., wireless part), whereas the rates of MMUs and SMUs are also constrained by the limited backhaul capacities.
Consider a typical mobile user-$u$.
If user-$u$ is served by the MBS tier, the achievable rate for radio access is given by
\begin{equation}
\label{eq_R_mr_define}
R_\mathrm{MR} = \frac{W_\mathrm{m}}{N_\mathrm{MR}+1} \log_2 \left( 1 + \gamma_\mathrm{m} \right),
\end{equation}
where $N_\mathrm{MR}$ denotes the number of residual users being served by the associated MBS except user-$u$ (both MHUs and MMUs included), $\gamma_\mathrm{m}$ is the received SINR given by
\begin{equation}
\label{eq_R_mr_define_1}
\gamma_\mathrm{m} = \min \left( \gamma_{\max}, \frac{P_\mathrm{TM} h_\mathrm{m} {d_\mathrm{m}}^{-\alpha_\mathrm{m}}}{\sigma^2 + I_\mathrm{m} } \right),
\end{equation}
$\gamma_{\max}$ is the maximal received SINR, $P_\mathrm{TM}$ is the MBS transmit power, $h_\mathrm{m}$ is an exponential random variable with mean 1 incorporating the effect of Rayleigh fading, $\alpha_\mathrm{m}$ is the path loss exponent of the MBS-tier, $d_\mathrm{m}$ denotes the distance from user-$u$ to the associated MBS, $\sigma^2$ is the addictive noise power, and $I_\mathrm{m}$ represents inter-cell interference from other MBSs.
In practical systems, $N_\mathrm{MR}$ varies randomly with the dynamic arrival and departure of file transmission demands, and $d_\mathrm{m}$ is also uncertain from the network perspective.
If user-$u$ is served by the SBS tier, the achievable rate for radio access can be given by
\begin{equation}
\label{eq_R_sr_define}
R_\mathrm{SR} = \frac{W_\mathrm{s}}{N_\mathrm{SR}+1} \log_2 \left( 1 + \gamma_\mathrm{s} \right),
\end{equation}
where $N_\mathrm{SR}$ denotes the number of residual users being served by the associated SBS except user-$u$ (including both SHUs and SMUs),
\begin{equation}
\label{eq_R_sr_define_1}
\gamma_\mathrm{s} = \min \left( \gamma_{\max}, \frac{P_\mathrm{TS} h_\mathrm{s} {d_\mathrm{s}}^{-\alpha_\mathrm{s}}}{\sigma^2 + I_\mathrm{s} } \right),
\end{equation}
$P_\mathrm{TS}$ is the SBS transmit power, $h_\mathrm{s}$ is an exponential random variable with mean 1 incorporating the effect of Rayleigh fading, $\alpha_\mathrm{s}$ is the path loss exponent of the SBS-tier, $d_\mathrm{s}$ denotes the distance from user-$u$ to the associated SBS, and $I_\mathrm{s}$ represents the inter-cell interference from other SBSs.
Similarly, $N_\mathrm{SR}$ and $d_\mathrm{s}$ are also random variables.
In addition, the probability distribution functions (PDFs) of $N_\mathrm{SR}$ and $d_\mathrm{s}$ can be more complex due to the uncertain small cell sizes.
The MBS and SBS backhaul transmission rates only depend on the corresponding traffic loads and capacities:
\begin{equation}
\label{eq_r_BH}
R_\mathrm{MBH} = \frac{U_\mathrm{MBH}}{N_\mathrm{MBH}+1}, ~~~ R_\mathrm{SBH} = \frac{U_\mathrm{SBH}}{N_\mathrm{SBH}+1},
\end{equation}
where $N_\mathrm{MBH}$ (or $N_\mathrm{SBH}$) represents the number of residual MMUs (or SMUs) sharing the MBS (or SBS) backhaul expect the considered user-$u$.
\section{Capacity-Optimal Caching Formulation}
\label{sec_formulation}
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{analytical_framework}
\caption{Influence of cache deployment and traffic steering on QoS performance.}
\label{fig_analytical_framework}
\end{figure}
To meet QoS requirements, the transmission rates should to be guaranteed for successful file delivery, which depend on the transmission rates of each network part (i.e., MBS radio access, MBS backhaul, SBS radio access and SBS backhaul), as shown in Fig.~\ref{fig_analytical_framework}.
In addition, cache deployment (i.e., cache sizes $[C_\mathrm{m}, C_\mathrm{s}]$) determines the traffic distributions across the network together with traffic steering ratio $\varphi$, thus influencing the transmission rates.
Therefore, cache deployment should be jointly optimized with traffic steering, which can be formulated as follows:
\begin{subequations}
\label{eq_P2}
\begin{align}
\max\limits_{C_\mathrm{s},\varphi}~~~& \mu(C_\mathrm{s},\varphi) \\
(\mbox{P1})~~\mbox{s.t.}~~~~& \mathds{E}\left[ R_\mathrm{MR} \right] \geq \check{R}_\mathrm{RAN},\\
& \mathds{E}\left[ R_\mathrm{MBH} \right] \geq \check{R}_\mathrm{BH},\\
& \mathds{E}\left[ R_\mathrm{SR} \right] \geq \check{R}_\mathrm{RAN},\\
& \mathds{E}\left[ R_\mathrm{SBH} \right] \geq \check{R}_\mathrm{BH}, \\
& 0 \leq C_\mathrm{s} \leq C/\rho_\mathrm{s}, ~~ 0 \leq \varphi \leq 1,
\end{align}
\end{subequations}
where the objective function $\mu(C_\mathrm{s},\varphi) = \underset{\lambda}{\max}{\left(\lambda |_{\{C_\mathrm{s},\varphi\}}\right)}$ is the network capacity for the given SBS cache size $C_\mathrm{s}$ and traffic steering ratio $\varphi$ (i.e., the maximal traffic density that can be catered), $\check{R}_\mathrm{RAN}$ and $\check{R}_\mathrm{BH}$ denote the per user rate requirements for radio access and backhaul transmissions\footnote{{{In practical systems, the backhaul rate requirement $\check{R}_\mathrm{BH}$ is generally much higher than that of the radio access part $\check{R}_\mathrm{RAN}$, considering the end-to-end delay and the serial transmission structure.}}}, respectively.
{{The cache size of MBSs $C_\mathrm{m}$ can be determined with $C_\mathrm{s}$ according to Eq.~(\ref{eq_cache_size}).}}
Constraint (\ref{eq_P2}b) and (\ref{eq_P2}d) guarantee the QoS requirements of content hit users, while (\ref{eq_P2}c) and (\ref{eq_P2}e) further account for the file fetching delay requirements of content miss users.
{{The average transmission rate is adopted for QoS guarantee as the services suitable for pro-active caching are mostly elastic in practical networks, such as popular video streaming.
Furthermore, when the average per user transmission rate is guaranteed, the objective function $\mu(C_\mathrm{s},\varphi)$ can also reflect the network goodput.}}
According to the properties of PPP, the traffic distributions of each network part also follow PPP.
Denote by $\lambda_\mathrm{MR}$, $\lambda_\mathrm{MBH}$, $\lambda_\mathrm{SR}$ and $\lambda_\mathrm{SBH}$ the equivalent user density for MBS radio access, MBS backhaul, SBS radio access, and SBS backhaul:
\begin{subequations}
\label{eq_lambda_define}
\begin{align}
\lambda_\mathrm{MR} & = \left[ P_\mathrm{Hit}^{(\mathrm{m})} + (1-P_\mathrm{Hit}) (1-\varphi) \right] \lambda, \\
\lambda_\mathrm{MBH} & = (1-P_\mathrm{Hit}) (1-\varphi) \lambda, \\
\lambda_\mathrm{SR} & = \left[ P_\mathrm{Hit}^{(\mathrm{s})} + (1-P_\mathrm{Hit}) \varphi \right] \lambda, \\
\lambda_\mathrm{SBH} & = (1-P_\mathrm{Hit}) \varphi \lambda.
\end{align}
\end{subequations}
The constraints (\ref{eq_P2}b)-(\ref{eq_P2}e) provide the maximal value of $\lambda_\mathrm{MR}$, $\lambda_\mathrm{MBH}$, $\lambda_\mathrm{SR}$, and $\lambda_\mathrm{SBH}$, denoted by $\hat{\lambda}_\mathrm{MR}$, $\hat{\lambda}_\mathrm{MBH}$, $\hat{\lambda}_\mathrm{SR}$, and $\hat{\lambda}_\mathrm{SBH}$, respectively.
In addition, $\hat{\lambda}_\mathrm{MR}$, $\hat{\lambda}_\mathrm{MBH}$, $\hat{\lambda}_\mathrm{SR}$, and $\hat{\lambda}_\mathrm{SBH}$ further constrains the traffic arrival rate $\lambda$ with Eq.~(\ref{eq_lambda_define}), for the given $C_\mathrm{s}$ and $\varphi$.
Thus, the network capacity depends on the bottleneck:
\begin{equation}
\begin{split}
& \mu(C_\mathrm{s},\varphi) = \min\left( \frac{\hat{\lambda}_\mathrm{MR}}{P_\mathrm{Hit}^{(\mathrm{m})} + (1-P_\mathrm{Hit}) (1-\varphi)}, \right.\\
& \frac{\hat{\lambda}_\mathrm{MBH}}{(1-P_\mathrm{Hit}) (1-\varphi)}, \left.\frac{\hat{\lambda}_\mathrm{SR}}{P_\mathrm{Hit}^{(\mathrm{s})} + (1-P_\mathrm{Hit}) \varphi}, \frac{\hat{\lambda}_\mathrm{SBH}}{(1-P_\mathrm{Hit}) \varphi}\right).
\end{split}
\end{equation}
The key issue is the transmission rates analysis, which will be addressed in the next section.
\section{QoS-Constrained Capacity Analysis}
\label{sec_capacity_analysis}
In this section, the file transmission rates of different networks parts are analyzed respectively, based on which the constraints (\ref{eq_P2}b)-(\ref{eq_P2}e) can be simplified with respect to $\lambda$.
\subsection{MBS backhaul}
$N_\mathrm{MBH}$ follows Poisson distribution of mean $\lambda_\mathrm{MBH}/\rho_\mathrm{m}$, according to Slivnyak-Mecke theorem \cite{Slivnyak_theorem}.
Thus, based on Eq.~(\ref{eq_r_BH}), the average file transmission rate of MBS backhaul can be derived:
\begin{equation}
\label{eq_rate_MBH}
\begin{split}
& \mathds{E} [ R_\mathrm{MBH} ] = \sum\limits_{n=0}^{\infty} \frac{U_\mathrm{MBH}}{n+1} \Pr\left\{ N_\mathrm{MBH} = n \right\} \\
& = \sum\limits_{n=0}^{\infty} \frac{U_\mathrm{MBH}}{n+1} \frac{\left(\frac{\lambda_\mathrm{MBH}}{\rho_\mathrm{m}}\right)^n}{n!} e^{-\frac{\lambda_\mathrm{MBH}}{\rho_\mathrm{m}}} = \frac{U_\mathrm{MBH} \rho_\mathrm{m}}{\lambda_\mathrm{MBH}} \left( 1-e^{-\frac{\lambda_\mathrm{MBH}}{\rho_\mathrm{m}}} \right).
\end{split}
\end{equation}
Combining Eq.~(\ref{eq_rate_MBH}) with Eq.~(\ref{eq_lambda_define}b), the SBS backhaul constraint Eqn.~(\ref{eq_P2}c) can be simplified with respect to traffic density $\lambda$.
Denote by $\hat{\lambda}_\mathrm{MBH} = \max\left\{ \lambda_\mathrm{MBH} | \mathds{E} [ R_\mathrm{MBH} ] \geq \check{R}_\mathrm{BH} \right\}$, the maximal traffic load on MBS backhaul.
As the average rate $\mathds{E} [ R_\mathrm{MBH} ]$ decreases with $\lambda_\mathrm{MBH}$, $\hat{\lambda}_\mathrm{MBH}$ satisfies
\begin{equation}
\label{eq_lambda_MBH_max}
\frac{\rho_\mathrm{m}}{\hat{\lambda}_\mathrm{MBH}} \left( 1-e^{-\frac{\hat{\lambda}_\mathrm{MBH}}{\rho_\mathrm{m}}} \right) = \frac{\check{R}_\mathrm{BH}}{U_\mathrm{MBH}},
\end{equation}
according to Eq.~(\ref{eq_rate_MBH}).
\subsection{SBS backhaul}
Compared with MBS backhaul, the transmission rate of SBS backhaul is more complex due to the random small cell size.
Denote by $A_\mathrm{s}$ the cell area size, which follows Gamma distribution with shape $\kappa=3.575$ and scale $1/\kappa \rho_\mathrm{s}$ \cite{Cao13_optimal_density_TWC}.
Thus, the PDF of $A_\mathrm{s}$ is given by
\begin{equation}
\label{eq_pdf_A_s}
f_{A_\mathrm{s}} (A) = A^{\kappa-1} e^{-\kappa \rho_\mathrm{s} A} \frac{(\kappa\rho_\mathrm{s})^\kappa}{\Gamma(\kappa)},
\end{equation}
where $\Gamma(\cdot)$ is the gamma function.
Furthermore, the number of SMUs served through SBS backhaul follows Poisson distribution of mean $\lambda_\mathrm{SBH} A$ given the cell size $A_\mathrm{s}=A$.
Thus, based on Eq.~(\ref{eq_r_BH}), the average transmission rate can be derived:
\begin{equation}
\label{eq_rate_SBH_derived}
\begin{split}
& \mathds{E}[R_\mathrm{SBH}] = \int\limits_{A=0}\limits^{\infty} \left(\sum_{n=0}^{\infty} \frac{U_\mathrm{SBH}}{n+1} \Pr\left\{N_\mathrm{SBH}=n | A \right\}\right) f_{A_\mathrm{s}}(A) \mbox{d} A \\
& = \int\limits_{A=0}\limits^{\infty} \left( \sum_{n=0}^{\infty} \frac{U_\mathrm{SBH}}{n+1} \frac{(\lambda_\mathrm{SBH}A)^{n}}{n!} e^{-\lambda_\mathrm{SBH}A} \right) f_{A_\mathrm{s}}(A) \mbox{d} A \\
& = \int\limits_{A=0}\limits^{\infty} \frac{U_\mathrm{SBH}}{\lambda_\mathrm{SBH}A}\left(1-e^{-\lambda_\mathrm{SBH}A}\right) A^{\kappa-1} e^{-\kappa \rho_\mathrm{s} A} \frac{(\kappa\rho_\mathrm{s})^\kappa}{\Gamma(\kappa)} \mbox{d} A \\
& = \frac{U_\mathrm{SBH}}{\lambda_\mathrm{SBH}} \left\{\int\limits_{A=0}\limits^{\infty} A^{\kappa-2} e^{-\kappa \rho_\mathrm{s} A} \frac{(\kappa\rho_\mathrm{s})^\kappa}{\Gamma(\kappa)} \mbox{d} A \right.\\
& \left. - \int\limits_{A=0}\limits^{\infty} A^{\kappa-2} e^{-(\kappa \rho_\mathrm{s}+\lambda_\mathrm{SBH}) A} \frac{(\kappa\rho_\mathrm{s})^\kappa}{\Gamma(\kappa)} \mbox{d} A \right\}\\
& = \frac{U_\mathrm{SBH}\kappa\rho_\mathrm{s}}{\lambda_\mathrm{SBH}} \frac{\Gamma(\kappa-1)}{\Gamma(\kappa)} \left( 1 - \frac{1}{\left(1+\frac{\lambda_\mathrm{SBH}}{\kappa \rho_\mathrm{s}}\right)^{\kappa-1}} \right).
\end{split}
\end{equation}
Combining Eq.~(\ref{eq_rate_SBH_derived}) with Eq.~(\ref{eq_lambda_define}d), the SBS backhaul constraint Eqn.~(\ref{eq_P2}e) can be simplified.
According to Eq.~(\ref{eq_rate_SBH_derived}), the maximal traffic load on SBS backhaul $\hat{\lambda}_\mathrm{SBH}$ can be given by
\begin{equation}
\label{eq_lambda_SBH_max}
\frac{\kappa\rho_\mathrm{s}}{\hat{\lambda}_\mathrm{SBH}} \frac{\Gamma(\kappa-1)}{\Gamma(\kappa)} \left( 1 - \frac{1}{\left(1+\frac{\hat{\lambda}_\mathrm{SBH}}{\kappa \rho_\mathrm{s}}\right)^{\kappa-1}} \right) = \frac{\check{R}_\mathrm{BH}}{U_\mathrm{SBH}}.
\end{equation}
\subsection{MBS Radio Access}
According to Eq.~(\ref{eq_R_mr_define}), the transmission rate of MBS radio access can be given by
\begin{equation}
\label{eq_rate_MR_define}
\begin{split}
& \mathds{E}[R_\mathrm{MR}] = \underset{ \{ N_\mathrm{MR}, d_\mathrm{m} \}} {\mathds{E}} \left[\frac{W_\mathrm{m}}{1+N_\mathrm{MR}}\log_2(1+\gamma_\mathrm{m})\right]
\end{split}
\end{equation}
where the user number $N_\mathrm{MR}$ follows Poisson distribution of mean $\lambda_\mathrm{MR}/\rho_\mathrm{m}$:
\begin{equation}
\label{eq_p_N_MR}
p_{N_\mathrm{MR}}(n) = \frac{(\frac{\lambda_\mathrm{MR}}{\rho_\mathrm{m}})^{n}}{n!} e^{-\frac{\lambda_\mathrm{MR}}{\rho_\mathrm{m}}},
\end{equation}
and the communication distance $d_\mathrm{m}$ can be considered to follow:
\begin{equation}
f_{d_\mathrm{m}}(d) = \frac{2 d}{ {D_\mathrm{m}}^2},
\end{equation}
by approximating MBS coverage as a circle of radius $D_\mathrm{m}$.
Then, the lower bound of average transmission rate for MBS radio access can be obtained by approximating the random inter-cell interference with the average value, which can be quite accurate under the condition of high signal-to-noise ratio (SNR) \cite{mine_TWC_SCoff}.
\textbf{Lemma~1.} The lower bound of average transmission rate of MBS radio access is given by:
\begin{equation}
\label{e_R_MR_aver_lemma_1}
\mathds{E} [R_\mathrm{MR}] \geq \frac{{\tau}_\mathrm{m} W_\mathrm{m}}{\bar{N}_\mathrm{MR}},
\end{equation}
where
\begin{equation}
\begin{split}
& {\tau}_\mathrm{m} = \log_2\frac{P_\mathrm{TM} D_\mathrm{m}^{-\alpha_\mathrm{m}} }{(1+\theta_\mathrm{m})\sigma^2}+\frac{\alpha_\mathrm{m}}{2 \ln 2}\left(1-\frac{D_\mathrm{min}^2}{D_\mathrm{m}^2}\right), \\
& \bar{N}_\mathrm{MR} = \frac{\lambda_\mathrm{MR}} {\rho_\mathrm{m}} \left(1-e^{-\frac{\lambda_\mathrm{MR}}{\rho_\mathrm{m}}}\right)^{-1},
\end{split}
\end{equation}
$D_\mathrm{min}$ is the transmission distance corresponding to the maximal received SINR (i.e., $\frac{P_\mathrm{TM}D_\mathrm{min}^{-\alpha_\mathrm{m}}}{(1+\theta_\mathrm{m})\sigma^2} = \gamma_\mathrm{max}$), and $\theta_\mathrm{m}$ denotes the ratio of average inter-cell interference to noise (i.e., $\theta_\mathrm{m} \sigma^2 = \mathds{E}[I_\mathrm{m}] $).
The equality of Eqn.~(\ref{e_R_MR_aver_lemma_1}) holds when $\frac{\sigma^2}{P_\mathrm{TM}} \rightarrow 0$.
\emph{Proof:}~Please refer to Appendix~\ref{appendix_rate_MR}.
\hfill \rule{4pt}{8pt}
\emph{Remark:}~The physical meaning of ${\tau}_\mathrm{m}$ is the average spectrum efficiency of MBS tier, and $\bar{N}_\mathrm{MR}$ reflects the average number of users accessing each MBS. In practical cellular networks, the received SINR is usually guaranteed to be high enough for reliable communications, through methods like inter-cell interference control. Therefore, Lemma~1 can be applied to approximate average data rate, and constraint Eqn.~(\ref{eq_P2}b) can be simplified with respect to traffic load $\lambda$ based on Eq.~(\ref{eq_lambda_define}a).
In addition, the maximal traffic load on MBS radio access $\hat{\lambda}_\mathrm{MR}$ can be given by:
\begin{equation}
\label{eq_lambda_MR_max}
\frac{\rho_\mathrm{m}}{\hat{\lambda}_\mathrm{MR}}\left(1-e^{-\frac{\hat{\lambda}_\mathrm{MR}}{\rho_\mathrm{m}}}\right) = \frac{\check{R}_\mathrm{RAN}}{{\tau}_\mathrm{m}W_\mathrm{m}}.
\end{equation}
\subsection{SBS Radio Access}
According to Eq.~(\ref{eq_R_sr_define}), the average transmission rate for SBS radio access is given by
\begin{equation}
\label{eq_rate_MR_define}
\begin{split}
& \mathds{E}[R_\mathrm{SR}] = \underset{ \{ A_\mathrm{s}, N_\mathrm{SR}, d_\mathrm{s} \}} {\mathds{E}} \left[\frac{W_\mathrm{s}}{1+N_\mathrm{SR}}\log_2(1+\gamma_\mathrm{s})\right].
\end{split}
\end{equation}
The accurate average transmission rate cannot be derived in closed form, due to the random SBS topology and user location.
Similarly, the lower bound of average transmission rate can be obtained by approximating the random inter-cell interference by the average value, given by Lemma~2.
\textbf{Lemma~2.}~The lower bound of average transmission rate of SBS radio access is given by:
\begin{equation}
\label{e_R_SR_aver_lemma_2}
\mathds{E} [R_\mathrm{SR}] \geq \frac{\tau_\mathrm{s}W_\mathrm{s}}{\bar{N}_\mathrm{SR}},
\end{equation}
where
\begin{equation}
\begin{split}
& \tau_\mathrm{s} = \log_2 \frac{P_\mathrm{TS} (\pi \rho_\mathrm{s})^{\frac{\alpha_\mathrm{s}}{2}}}{(1+\theta_\mathrm{s})\sigma^2} + \frac{\alpha_\mathrm{s}}{2 \ln 2} \gamma,\\
& \bar{N}_\mathrm{SR} = \frac{\lambda_\mathrm{SR}}{\kappa \rho_\mathrm{s}} \frac{\Gamma(\kappa)}{\Gamma(\kappa-1)} \left[1-\frac{1}{\left(1+\frac{\lambda_\mathrm{SR}}{\kappa \rho_\mathrm{s}}\right)^{\kappa-1}}\right]^{-1},
\end{split}
\end{equation}
$\gamma \approx 0.577$ is Euler-Mascheroni constant, and $\theta_\mathrm{s}$ denotes the ratio of average inter-cell interference to noise at SBS tier.
The equality of Eqn.~(\ref{e_R_SR_aver_lemma_2}) holds when $\frac{\sigma^2}{P_\mathrm{TS}} \rightarrow 0$.
\emph{Proof}: Please refer to Appendix~\ref{appendix_rate_SR}.
\hfill \rule{4pt}{8pt}
\emph{Remark:}~${\tau}_\mathrm{s}$ can be interpreted as the average spectrum efficiency of SBS tier, and $\bar{N}_\mathrm{SR}$ reflects the average number of users accessing each SBS.
The constraint Eqn.~(\ref{eq_P2}d) can be simplified by combing Lemma~2 with Eq.~(\ref{eq_lambda_define}c).
In addition, the maximal traffic load on SBS radio access $\hat{\lambda}_\mathrm{SR}$ can be given by
\begin{equation}
\label{eq_lambda_SR_max}
\frac{\kappa \rho_\mathrm{s}}{\hat{\lambda}_\mathrm{SR}} \frac{\Gamma(\kappa-1)}{\Gamma(\kappa)} \left[1-\frac{1}{\left(1+\frac{\hat{\lambda}_\mathrm{SR}}{\kappa \rho_\mathrm{s}}\right)^{\kappa-1}}\right] = \frac{\check{R}_\mathrm{RAN}}{\tau_\mathrm{s}W_\mathrm{s}} ,
\end{equation}
\section{Capacity-Optimal Hierarchical Caching}
\label{sec_solution}
Based on the transmission rates analysis, problem (P1) can be simplified as follows:
\begin{subequations}
\label{eq_P3}
\begin{align}
\max\limits_{C_\mathrm{s},\varphi}~~~& \mu(C_\mathrm{s},\varphi) \\
(\mbox{P2})~~\mbox{s.t.}~~~~& \left[P_\mathrm{Hit}^{(\mathrm{m})}+(1-P_\mathrm{Hit})(1-\varphi)\right] \lambda \leq \hat{\lambda}_\mathrm{MR},\\
& (1-P_\mathrm{Hit})(1-\varphi) \lambda \leq \hat{\lambda}_\mathrm{MBH},\\
& \left[P_\mathrm{Hit}^{(\mathrm{s})}+(1-P_\mathrm{Hit})\varphi\right] \lambda \leq \hat{\lambda}_\mathrm{SR},\\
& (1-P_\mathrm{Hit})\varphi \lambda \leq \hat{\lambda}_\mathrm{SBH},\\
& 0 \leq C_\mathrm{s} \leq C/\rho_\mathrm{s}, ~~ 0 \leq \varphi \leq 1,
\end{align}
\end{subequations}
where $\hat{\lambda}_\mathrm{MR}$, $\hat{\lambda}_\mathrm{MBH}$, $\hat{\lambda}_\mathrm{SR}$, and $\hat{\lambda}_\mathrm{SBH}$ are given by Eqs.~(\ref{eq_lambda_MR_max}, \ref{eq_lambda_MBH_max}, \ref{eq_lambda_SR_max}, and \ref{eq_lambda_SBH_max}), while the content hit rate $P_\mathrm{Hit}^{(\mathrm{m})}$, $P_\mathrm{Hit}^{(\mathrm{s})}$ and $P_\mathrm{Hit}$ can be derived by Eqs.~(\ref{eq_P_hit_s_m}) and (\ref{eq_P_hit}) with respect to different caching deployment $[C_\mathrm{s},C_\mathrm{m}]$.
Although different cache deployments can influence the traffic load distribution, problem (P2) differs significantly from the conventional load balancing problems.
The total traffic load remains constant in load balancing problems, where the traffic load is shifted from one part to another.
Instead, different cache deployments may change backhaul loads, as the content hit rate varies with cache sizes.
\subsection{Problem Analysis and Solutions}
Denote by
\begin{subequations}
\label{eq_capacity_part}
\begin{align}
\mu_\mathrm{MR} & = \frac{\hat{\lambda}_\mathrm{MR}}{ P_\mathrm{Hit}^{(\mathrm{m})} + (1-P_\mathrm{Hit}) (1 - \varphi) }, \\
\mu_\mathrm{MBH} & = \frac{\hat{\lambda}_\mathrm{MBH}}{(1-P_\mathrm{Hit})(1-\varphi)}, \\
\mu_\mathrm{SR} & = \frac{\hat{\lambda}_\mathrm{SR}}{ P_\mathrm{Hit}^{(\mathrm{s})} + (1-P_\mathrm{Hit}) \varphi}, \\
\mu_\mathrm{SBH} & = \frac{\hat{\lambda}_\mathrm{SBH}}{ (1-P_\mathrm{Hit})\varphi },
\end{align}
\end{subequations}
the maximal network traffic density constrained by the corresponding network part, and the network capacity is given by
\begin{equation}
\label{eq_P3_opt_simple}
\mu(C_\mathrm{s},\varphi) = \min\left\{ \mu_\mathrm{MR}, \mu_\mathrm{MBH}, \mu_\mathrm{SR}, \mu_\mathrm{SBH} \right\}.
\end{equation}
Numerical results of problem (P2) can be obtained by exhaustive search based on Eqs.~(\ref{eq_lambda_MBH_max}, \ref{eq_lambda_SBH_max}, \ref{eq_lambda_MR_max}, \ref{eq_lambda_SR_max}, \ref{eq_capacity_part} and \ref{eq_P3_opt_simple}).
Furthermore, low-complexity heuristic algorithms can be designed.
To maximize network capacity, the traffic loads of each network part should be balanced according to the corresponding service capabilities.
Notice that the traffic load distribution can be manipulated by adjusting cache deployment strategy or traffic steering ratio.
Specifically, Table~\ref{tab_variation} gives the variations of traffic load distribution with respect to cache size and traffic steering ratio, with proof provided in Appendix~\ref{appendix_variation}.
Table~\ref{tab_variation} can provide a guideline to enhance network capacity in practical networks.
For example, when the MBS radio access is the performance bottleneck (i.e., $\mu_\mathrm{MR}<\mu_\mathrm{MBH}$, $\mu_\mathrm{MR}<\mu_\mathrm{SR}$, $\mu_\mathrm{MR}<\mu_\mathrm{SBH}$), we can either reduce cache size at MBSs, or increase the traffic steering ratio.
Instead, when the SBS backhaul is limited, we can either reduce the SBS cache size or lower the traffic steering ratio.
\begin{table}[!t]
\caption{Capacity variations by increasing cache size or steering ratio}
\label{tab_variation}
\centering
\begin{tabular}{ccccc}
\hline
\hline
& $\mu_\mathrm{MR}$ & $\mu_\mathrm{MBH}$ & $\mu_\mathrm{SR}$ & $\mu_\mathrm{SBH}$ \\
\hline
$C_\mathrm{s}$ & Increase & Decrease & Decrease & Decrease \\
$\varphi$ & Increase & Increase & Decrease & Decrease \\
\hline
\hline
\end{tabular}
\end{table}
\subsection{SBS-Backhaul-Constrained HetNets}
In practical systems, MBSs are expected to be equipped with optical fiber backhauls which can provide sufficiently large bandwidth, whereas the capacity for radio access can be much smaller due to spectrum resource scarcity.
In this case, problem (P2) can be further simplified by removing constraint (\ref{eq_P3}c).
The condition of ideal MBS backhaul is $\hat{\lambda}_\mathrm{MR} \leq \hat{\lambda}_\mathrm{MBH}$, whereby constraint (\ref{eq_P3}c) always holds if (\ref{eq_P3}b) is satisfied.
In what follows, we focus on HetNets with ideal MBS backhaul, and find the analytical solutions to problem (P2) with different network settings.
To begin with, we consider a simple case when both the MBS and SBS tiers have unconstrained backhaul capacity, i.e., $\hat{\lambda}_\mathrm{MR} \leq \hat{\lambda}_\mathrm{MBH}$ and $\hat{\lambda}_\mathrm{SR} \leq \hat{\lambda}_\mathrm{SBH}$.
In this case, constraints (\ref{eq_P3}c) and (\ref{eq_P3}e) can be both neglected, and the network capacity cannot be improved by deploying cache.
This case corresponds to conventional network deployment, where the backhaul capacity is sufficiently reserved while radio resources are limited.
Problem (P2) degenerates to the conventional inter-tier load balancing problem which can be easily solved.
By adding constraints (\ref{eq_P3}b) and (\ref{eq_P3}d), we have $\lambda \leq \hat{\lambda}_\mathrm{MR} + \hat{\lambda}_\mathrm{SR}$.
Thus, the maximal network capacity is $\mu(C_\mathrm{s},\varphi)^* = \hat{\lambda}_\mathrm{MR} + \hat{\lambda}_\mathrm{SR}$, and the optimal traffic steering is given by $\varphi^* = \frac{\hat{\lambda}_\mathrm{SR}}{\hat{\lambda}_\mathrm{MR} + \hat{\lambda}_\mathrm{SR}}$ without mobile edge caching.
As SBSs further densifies, the network capacity for radio access can scale almost linearly with SBS density, whereas densely deploying high speed fiber backhaul for each SBS is not practical considering the high cost.
In addition, the traffic of multiple SBSs can be geographically aggregated and transmitted through a shared backhaul (e.g., the cloud-RAN architecture), which further limits the backhaul capacity of each SBS \cite{CRAN_white_paper}.
Thus, the SBS tier can be backhaul-constrained, i.e., $\hat{\lambda}_\mathrm{SR} > \hat{\lambda}_\mathrm{SBH}$.
In this case, deploying caching at SBSs can improve network capacity by reducing backhaul traffic load, which is equivalent to increasing backhaul capacity.
Furthermore, the optimal solutions to (P2) is threshold-based.
Denote by $C_\mathrm{min}$ and $C_\mathrm{max}$ the two thresholds of cache budgets, which are given by
\begin{equation}
\label{eq_C_min_max}
\begin{split}
\sum\limits_{f=1}\limits^{C_\mathrm{min}/\rho_\mathrm{s}} q_f = \frac{\hat{\lambda}_\mathrm{SR}-\hat{\lambda}_\mathrm{SBH}}{\hat{\lambda}_\mathrm{MR}+ \hat{\lambda}_\mathrm{SR}}, & \\
\sum\limits_{f=C_\mathrm{min}/\rho_\mathrm{s}+1}\limits^{C_\mathrm{min}/\rho_\mathrm{s} + (C_\mathrm{max}-C_\mathrm{min}) / \rho_\mathrm{m} } q_f = & \frac{\hat{\lambda}_\mathrm{MR} }{ \hat{\lambda}_\mathrm{MR} + \hat{\lambda}_\mathrm{SR} } .
\end{split}
\end{equation}
The optimal solution to (P2) is summarized in Propositions~2-4, under different cache budgets.
\textbf{Proposition~2.} If $C < C_\mathrm{min} $, the optimal solution to (P2) is given by $C_\mathrm{s} = C/ \rho_\mathrm{s}$, $\varphi = \hat{\lambda}_\mathrm{SBH}/(\hat{\lambda}_\mathrm{MR} + \hat{\lambda}_\mathrm{SBH})$.
\emph{Proof}: Please refer to Appendix~\ref{appendix_insufficient_cache}.
\hfill \rule{4pt}{8pt}
\emph{Remark:} As deploying caching is equivalent to increasing backhaul capacity, cache instances need to be deployed at the backhaul-constrained SBSs for compensation.
$C_\mathrm{min}$ can be interpreted as the deficiency of SBS backhaul, and $C_\mathrm{min}/\rho_\mathrm{min}$ is the minimal SBS cache size needed to match with radio access resources.
When the cache budget is smaller than $C_\mathrm{min}$, the SBS tier is still backhaul-constrained even when all the cache instances are deployed at SBSs, and the network capacity increases with the cache budget.
Furthermore, the SBS radio resources are always redundant compared with SBS backhaul, and thus the performance bottleneck exists at either the SBS backhaul or the MBS radio access.
Accordingly, the load of SBS backhaul and MBS radio access should be balanced, by steering the content miss users to the two tiers appropriately.
When the cache budget increases to $C_\mathrm{min}$, the SBS backhaul deficiency can be completely compensated, and the network capacity achieves $\hat{\lambda}_\mathrm{MR}+\hat{\lambda}_\mathrm{SR}$.
As cache budget further increases, the network performance will be constrained by radio access instead of SBS backhaul, and the network capacity no longer increases.
If $C > C_\mathrm{min}$, there exist solutions to achieving the maximal network capacity $\mu(C_\mathrm{s},\varphi)^* = \hat{\lambda}_\mathrm{MR}+\hat{\lambda}_\mathrm{SR}$, as long as the SBS cache size is large enough to compensate backhaul deficiency, i.e., $C_\mathrm{s} \geq C_\mathrm{min}/\rho_\mathrm{s}$.
Among these capacity-optimal solutions, those with higher content hit rates can further improve user experience by reducing content fetching delay.
Thus, we aim to find the solution $\left[C_\mathrm{s}^*, \varphi^*\right]$, which can maximize content hit rate while guaranteeing network capacity $\hat{\lambda}_\mathrm{MR}+\hat{\lambda}_\mathrm{SR}$.
Although increasing the MBS cache size can improve content hit rate, larger MBS cache size results in heaver traffic load at MBSs, which can degrade the transmission rate at MBS radio access, especially when the cache budget is large.
Thus, the optimal cache deployment and traffic steering depend on the cache budget, given by Proposition~3 and 4.
\textbf{Proposition~3.} If $ C_\mathrm{min} \leq C < C_\mathrm{max}$, the solution $[C_\mathrm{s}^*, \varphi^*]$ satisfying
\begin{subequations}
\label{eq_P3_opt_sufficient_C}
\begin{align}
C_\mathrm{s}^* & = C_\mathrm{min}/\rho_\mathrm{s}, \\
(1-P_\mathrm{Hit}^*) \varphi^* & = \sum\limits_{f=C_\mathrm{s}^*+C_\mathrm{m}^*+1}\limits^{F} q_f \varphi^* = \frac{\hat{\lambda}_\mathrm{SBH}}{\hat{\lambda}_\mathrm{MR} +\hat{\lambda}_\mathrm{SR}} ,
\end{align}
\end{subequations}
can maximize content hit rate while maximizing network capacity, where $C_\mathrm{m}^*=(C - \rho_\mathrm{s} C_\mathrm{s}^*) / \rho_\mathrm{m}$ and $P_\mathrm{Hit}^*$ is the corresponding aggregated content hit rate.
\emph{Proof}: Please refer to Appendix~\ref{appendix_sufficient_cache}.
\hfill \rule{4pt}{8pt}
\textbf{Proposition~4.} If $ C \geq C_\mathrm{max} $, the solution $\left[C_\mathrm{s}^*, \varphi^* \right]$ satisfying
\begin{subequations}
\label{eq_P3_opt_oversupplied_C}
\begin{align}
\varphi^* & = 1 \\
{P_\mathrm{Hit}^{(\mathrm{m})}}^* = \sum\limits_{f=C_\mathrm{s}^*+1}\limits^{C_\mathrm{s}^*+C_\mathrm{m}^*} & = \frac{\hat{\lambda}_\mathrm{MR} }{\hat{\lambda}_\mathrm{MR} +\hat{\lambda}_\mathrm{SR}},
\end{align}
\end{subequations}
can maximize content hit rate while maximizing network capacity, where $C_\mathrm{m}^*=(C - \rho_\mathrm{s} C_\mathrm{s}^*) / \rho_\mathrm{m}$ and $P_\mathrm{Hit}^{(\mathrm{m})}$ is the corresponding content hit rate at MBSs.
\emph{Proof}: Substituting Eq.~(\ref{eq_P3_opt_oversupplied_C}) into (\ref{eq_capacity_part}) and (\ref{eq_P3_opt_simple}), the network capacity can achieve $\hat{\lambda}_\mathrm{MR}+\hat{\lambda}_\mathrm{SR}$. As $\varphi^*=1$, $\mu_\mathrm{MR}$ will decrease as $C_\mathrm{m}$ increases, degrading network capacity. Therefore, $\left[C_\mathrm{s}^*, \varphi^* \right]$ achieves maximal content hit rate among the capacity-optimal schemes.
\hfill \rule{4pt}{8pt}\\
Based on the above analysis, we summarize the propose capacity-optimal cache deployment scheme for SBS-backhaul-constrained HetNets as follows:
\begin{itemize}
\item \textbf{Case-1:} If $C \leq C_\mathrm{min} $, all cache instances should be deployed at the SBS tier;
\item \textbf{Case-2:} If $C_\mathrm{min} < C \leq C_\mathrm{max} $, the cache size of each SBS is $C_\mathrm{min}/\rho_\mathrm{s}$, and the remaining cache budget should be deployed at the MBS tier (i.e., $C_\mathrm{m} = (C- C_\mathrm{min})/\rho_\mathrm{m}$);
\item \textbf{Case-3:} If $C > C_\mathrm{max} $, the optimal cache deployment should guarantee that MBS-tier content hit rate satisfies Eq.~(\ref{eq_P3_opt_oversupplied_C}b).
\end{itemize}
Meanwhile, traffic steering ratio should be adjusted with caching deployment, to balance inter-tier traffic load.
In addition, the analytical results of thresholds $C_\mathrm{min}$ and $C_\mathrm{max}$ are derived as Eq.~(\ref{eq_C_min_max}), which depend on backhaul and radio resource provisions.
\section{Simulation and Numerical Results}
\label{sec_simulation}
In this section, simulations are conducted to validate the obtained analytical results, and numerical results are provided to offer insights into practical network deployment.
The file popularity is considered to follow Zipf distribution \cite{video_popularity_2009}:
\begin{equation}
q_f = \frac{1/f^\nu}{\sum_{h=1}^{F} 1/h^\nu},
\end{equation}
where $\nu\geq 0$ indicates the skewness of popularity distribution.
In this simulation, $\nu$ is set as $0.56$, featuring video streaming services \cite{video_popularity_2009}.
Important parameters are listed in Table~\ref{tab_parameter} \cite{Liu16_EE_cache_JSAC}.
\begin{table}[!t]
\caption{Simulation parameters}
\label{tab_parameter}
\centering
\begin{tabular}{cccc}
\hline
\hline
Parameter & Value & Parameter & Value \\
\hline
$D_\mathrm{m}$ & 500 m & $\rho_\mathrm{s}$ & 50 /km$^2$ \\
$P_\mathrm{TM}$ & 10 W & $P_\mathrm{TS}$ & 2 W \\
$W_\mathrm{M}$ & 100 MHz & $W_\mathrm{S}$ & 10 MHz \\
$\alpha_\mathrm{m}$ & 3.5 & $\alpha_\mathrm{s}$ & 4 \\
$\sigma^2$ & -105 dBm/MHz & $U_\mathrm{MBH}$ & 100 Gbps\\
$\theta_\mathrm{m}$ & 1000 & $\theta_\mathrm{s}$ & 1000\\
$\check{R}_\mathrm{RAN}$ & 5 Mbps & $\check{R}_\mathrm{BH}$ & 50 Mbps \\
$F$ & 1000 & $\nu$ & 0.56 \\
\hline
\hline
\end{tabular}
\end{table}
\subsection{Analytical Results Evaluation}
{{The analytical results of file transmission rates for MBS/SBS radio access are validated in Fig.~\ref{fig_evaluation}, where 15\% traffic is served by the MBS tier and the remaining is steered to the SBS tier.}}
Monte Carlo method is applied in simulation, with SBS topology, user location and channel fading generated according to the corresponding PDFs.
The simulation results is averaged over 10000 samples.
The analytical results are calculated based on Lemmas~1 and 2.
As the analytical and simulation results are shown to be very close, Lemmas~1 and 2 can be applied to approximate transmission rate analysis for radio access.
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{evaluation_rate}
\caption{{{Evaluation of derived file tranmsission rate at radio access parts.}}}
\label{fig_evaluation}
\end{figure}
\begin{figure*}[!t]
\centering
\subfloat[$C$=50 files/km$^2$] {\includegraphics[width=2.2in]{optimal_planning_insufficient_cache}}
\subfloat[$C$=900 files/km$^2$]{\includegraphics[width=2.2in]{optimal_planning_medium_cache}}
\subfloat[$C$=1000 files/km$^2$]{\includegraphics[width=2.2in]{optimal_planning_sufficient_cache}}
\caption{Optimal hierarchical cache.}
\label{fig_optimal_cache_3D}
\end{figure*}
\subsection{Optimal Hierarchical Caching}
To validate the theoretical analysis, Fig.~\ref{fig_optimal_cache_3D} shows network capacity with respect to different cache deployment and traffic steering ratios, where the analytical results obtained by Propositions~2-4 are marked as the star points.
The per user rates for radio access and backhaul are set as $\check{R}_\mathrm{RAN}$=5 Mbps and $\check{R}_\mathrm{BH}$=50 Mbps, respectively.
$\xi_\mathrm{c}$ is the ratio of cache budget deployed at SBSs, i.e., $\xi_\mathrm{c} = C_\mathrm{s} \rho_\mathrm{s}/C$.
According to Eq.~(\ref{eq_C_min_max}), $C_\mathrm{min}$=870 files/km$^2$, $C_\mathrm{max}$=930 files/km$^2$.
Thus, the three subfigures correspond to the cases of Propositions~2-4, respectively.
In Fig.~\ref{fig_optimal_cache_3D}(a), the star point is shown to achieve the maximal network capacity, validating the analysis of Proposition~2.
In Figs.~\ref{fig_optimal_cache_3D}(b) and (c), there are multiple solutions that can achieve the maximal network capacity, including the star points.
Furthermore, the star points also minimize the SBS cache size (i.e., minimal $\xi_\mathrm{c}$) among all the capacity-optimal schemes, indicating high content hit rate.
When $C$=900 files/km$^2$, the SBS backhaul will become the bottleneck when $\xi_\mathrm{c}$ is lower than $\xi_\mathrm{c}^*$, degrading network capacity as shown in Fig.~\ref{fig_optimal_cache_3D}(b).
Instead, in Fig.~\ref{fig_optimal_cache_3D}(c), the performance bottleneck is due to the MBS radio access, and thus the network capacity will decrease when $\xi_\mathrm{c}$ is lower than $\xi_\mathrm{c}^*$.
The numerical results of Figs.~\ref{fig_optimal_cache_3D}(a)-(c) are consistent with Propositions~2-4, validating the theoretical analysis.
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{optimal_planning_cache_budget}
\caption{Optimal cache spliting with respect to cache budget.}
\label{fig_optimal_planning_cache_budget}
\end{figure}
Fig.~\ref{fig_optimal_planning_cache_budget} further demonstrates the relationship between the cache budget and the optimal cache deployment, obtained by exhaustive search.
As shown by the dash line, the optimal cache deployment can be divided into three cases.
Firstly, all cache budget should be allocated to the SBS tier when the cache budget is insufficient, i.e., $C<C_\mathrm{min}$.
As the cache budget achieves $C_\mathrm{min}$ and is lower than $C_\mathrm{max}$, it is shown that the ratio of cache budget allocated to the SBS tier begins to decrease.
Furthermore, the ratio of cache budget allocated to the SBS tier increases again when the cache budget exceeds $C_\mathrm{max}$.
\subsection{Cache-Backhaul Trading}
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{capacity_cache_backhaul}
\caption{Cache-enabled network capacity.}
\label{fig_capacity_cache_backhaul}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{cache_backhaul_trading}
\caption{Cache budget demand on different backhaul capacity.}
\label{fig_cache_backhaul_trading}
\end{figure}
The solid line of Fig.~\ref{fig_optimal_planning_cache_budget} presents the network capacity with respect to cache budget, which is normalized by the capacity without cache.
As the cache budget increases, the network capacity firstly increases and then levels off as a constant.
The reason is that the SBS backhaul is no longer the bottleneck when the cache budget achieves $C_\mathrm{min}$, and the network performance is constrained by the radio resources.
Fig.~\ref{fig_capacity_cache_backhaul} further illustrates the cache-enabled network capacity gain under different SBS backhaul capacities.
Similarly, the network capacities firstly increase and then level off, and the turning points $C_\mathrm{min}$ depend on cache budgets.
Furthermore, a larger backhaul capacity results in a smaller turning point, as shown in Fig.~\ref{fig_capacity_cache_backhaul}.
Specifically, no cache budget is needed when the backhaul capacity is $U_\mathrm{SBH}=1.2$ Gbps, since such backhaul capacity is sufficiently large compared with SBS radio access resources.
Fig.~\ref{fig_capacity_cache_backhaul} reveals the trading relationship between backhaul capacity and cache budget demands.
Specifically, networks with insufficient backhaul capacity can be compensated by deploying cache, and the backhaul deficiency determines the amount of cache budget needed.
The relationship between required cache budget and backhaul capacity is illustrated in Fig.~\ref{fig_cache_backhaul_trading}, where the backhaul capacity is normalized by the capacity of radio access.
It is shown that less cache budget is needed as the backhaul capacity increases, and denser networks demand higher cache budget.
The trading relationship between backhaul capacity and cache budget demand can be applied to cost-effective network deployment, which determines the optimal combination of backhaul capacity and cache budget.
\subsection{Case Studies on Cost-Effective Network Deployment}
\begin{figure*}[!t]
\centering
\subfloat[Cache-enabled small cells] {\includegraphics[width=2.5in]{efficient_network_deployment}}
\hfil
\subfloat[Caching station]{\includegraphics[width=2.5in]{cache_station_deployment_no_backhaul}}
\caption{Cost-effective network deployment.}
\label{fig_network_deployment}
\end{figure*}
Finally, we provide case studies on cost-effective network deployment, by applying the results of Fig.~\ref{fig_cache_backhaul_trading}.
Fig.~\ref{fig_network_deployment}(a) illustrates the cost-optimal SBS deployment for the given backhaul deployment cost, under different caching budget $C$.
The backhaul deployment cost is considered to increase with SBS density as well as backhaul capacity: $\rho_\mathrm{s}(1+K_\mathrm{BH} U_\mathrm{SBH}^{\zeta_\mathrm{BH}})$\footnote{$K_\mathrm{BH}$ denotes the ratio of backhaul deployment cost to the SBS equipment cost, and $\zeta_\mathrm{BH}$ reflects how the backhaul deployment cost scales with capacity.}.
For illustration, $K_\mathrm{BH}=0.001$ and $\zeta_\mathrm{BH}=0.5$\footnote{With this setting, the backhaul deployment cost is comparable to the SBS equipment cost when backhaul capacity is 1 Gbps.}.
When budget of the backhaul cost is 100 /km$^2$, the network capacity with respect to different SBS density is shown as Fig.~\ref{fig_network_deployment}(a).
The network capacity is demonstrated to firstly increase and then decrease with SBS densities, falling into two regions.
{{On one hand, the capacity of radio access increases with SBS density.
On the other hand, the backhaul capacity per SBS decreases with SBS density due to the constrained deployment cost.
Accordingly, the performance of SBS tier is constrained by the radio access resources when the SBS density is low, and becomes backhaul-constrained when the SBS density exceeds some threshold.
In fact, the optimal SBS density achieves the best match between radio and backhaul resource settings.}}
Fig.~\ref{fig_network_deployment}(b) further demonstrates the cost-optimal deployment of caching stations, which is a special case when the SBSs have no backhaul and all content miss users are served by MBSs.
The deployment cost is considered to increase with SBS density as well as SBS cache size, i.e., $\rho_\mathrm{s} (1+K_\mathrm{c} {C_\mathrm{s}}^{\zeta_\mathrm{C}})$\footnote{$K_\mathrm{c}$ denotes the ratio of storage cost to the cost of other modules, and $\zeta_\mathrm{C}$ reflects how storage cost scales with cache size.}.
$K_\mathrm{C}$ and $\zeta_\mathrm{C}$ are system parameters, set as $K_\mathrm{C}=0.1$ and $\zeta_\mathrm{\textsc{C}}=0.5$ for illustration\footnote{With this setting, the storage cost is comparable to the other modules when cache size is 100 files, i.e., 10\% of all contents}.
The network capacity with respect to SBS density is demonstrated in Fig.~\ref{fig_network_deployment}(b).
{{The cost-optimal SBS density reflects the tradeoff between radio access capacity and content hit rate.
The capacity for radio access increases with SBS density, whereas the cache size decreases due to the deployment budget.
With low density, the SBSs are overloaded due to the constrained radio resources.
However, the high-dense SBSs can only serve few users due to low content hit rate, causing radio resource underutilized.
Thus, the optimal SBS density should balance the capacity of SBS radio access and content hit rate, to maximize network capacity.
The results of Fig.~\ref{fig_network_deployment} can provide insightful design criteria for cost-effective network deployment.
}}
\section{Conclusions and Future Work}
\label{sec_conclusions}
In this paper, the cost-effective cache deployment problem has been investigated for a large-scale two-tier HetNet, aiming at maximizing network capacity while meeting file transmission rate requirements.
By conducting stochastic geometry analysis, the capacity-optimal cache sizes have been derived, which is threshold-based with respect to cache budget under the SBS-backhaul-constrained case.
The analytical results of cache budget threshold have been obtained, which characterize the backhaul deficiency and the cache-backhaul trading relationship.
{{The proposed cache deployment schemes can be applied to practical network upgrades as well as capacity enhancement.
When the existing networks upgrade with storage units for edge caching, the optimal cache sizes of different BSs can be directly determined with the obtained cache budget threshold, based on system parameters such as base station density, radio resources, backhaul capacity, and content popularity.
When more cache-enabled MBSs or SBSs are deployed for capacity enhancement, the proposed method can be applied to determine the optimal cache sizes and simplify the optimization of other system parameters.}}
{{For future work, we will optimize cache deployment based on cooperative caching scheme, where multiple SBSs can cooperate to serve users.}}
\appendices{}
\section{Proof of Lemma~1}
\label{appendix_rate_MR}
The average transmission rate of MBS radio access is given by:
\begin{equation}
\label{eq_rate_MR_1}
\begin{split}
& \mathds{E}[R_\mathrm{MR}] = W_\mathrm{m} {\mathds{E}} \left[\frac{1}{1+N_\mathrm{MR}}\right] {\mathds{E}} \left[\log_2(1+\gamma_\mathrm{m})\right] \\ &
= \frac{W_\mathrm{m}}{\ln 2} {\mathds{E}} \left[\frac{1}{1+N_\mathrm{MR}}\right] {\mathds{E}} [\ln(1+\gamma_\mathrm{m})]
\end{split}
\end{equation}
As $N_\mathrm{MR}$ follows Poisson distribution of mean $\lambda_\mathrm{MR}/\rho_\mathrm{m}$,
\begin{equation}
\label{eq_N_MR_aver_appendix}
{\mathds{E}} \left[\frac{1}{1+N_\mathrm{MR}}\right] = \frac{ \rho_\mathrm{m}}{\lambda_\mathrm{MR}} \left( 1-e^{-\frac{\lambda_\mathrm{MR}}{\rho_\mathrm{m}}} \right),
\end{equation}
which can be derived in the same way as Eq.~(\ref{eq_rate_MBH}) by replacing $\lambda_\mathrm{MBH}$ by $\lambda_\mathrm{MR}$.
Furthermore,
\begin{subequations}
\label{eq_SE_MR}
\begin{align}
& {\mathds{E}} \left[\ln(1+\gamma_\mathrm{m})\right] = \frac{D_\mathrm{min}^2}{D_\mathrm{m}^2} \ln \left( 1+\frac{P_\mathrm{TM} D_\mathrm{min}^{-\alpha_\mathrm{m}}}{(1+\theta_\mathrm{m})\sigma^2} \right) \nonumber \\
& + \int_{D_\mathrm{min}}^{D_\mathrm{m}} \frac{2d}{D_\mathrm{m}^2} \ln \left( 1+\frac{P_\mathrm{TM}d^{-\alpha_\mathrm{m}}}{(1+\theta_\mathrm{m})\sigma^2} \right) \mbox{d} d \nonumber\\
& \geq \frac{D_\mathrm{min}^2}{D_\mathrm{m}^2} \ln \left(\frac{P_\mathrm{TM} D_\mathrm{min}^{-\alpha_\mathrm{m}}}{(1+\theta_\mathrm{m})\sigma^2} \right) - \frac{2 \alpha_\mathrm{m} }{D_\mathrm{m}^2} \int\limits_{D_\mathrm{min}}\limits^{D_\mathrm{m}} d \ln d \mbox{d} d \nonumber \\
& + \left(1-\frac{D_\mathrm{min}^2}{D_\mathrm{m}^2}\right) \ln \left( \frac{P_\mathrm{TM}}{(1+\theta_\mathrm{m})\sigma^2} \right) \\
& = \ln\frac{P_\mathrm{TM} D_\mathrm{m}^{-\alpha_\mathrm{m}} }{(1+\theta_\mathrm{m})\sigma^2}+\frac{\alpha_\mathrm{m}}{2}\left(1-\frac{D_\mathrm{min}^2}{D_\mathrm{m}^2}\right), \nonumber
\end{align}
\end{subequations}
where $\theta_\mathrm{m} \sigma^2 = \mathds{E}[I_\mathrm{m}]$, and the equality of (\ref{eq_SE_MR}a) holds when $\frac{\sigma^2}{P_\mathrm{TM}} \rightarrow 0$.
Substituting Eqs.~(\ref{eq_N_MR_aver_appendix}) and (\ref{eq_SE_MR}) into (\ref{eq_rate_MR_1}), Lemma~1 can be proved.
\section{Proof of Lemma~2}
\label{appendix_rate_SR}
The average transmission rate of SBS radio access is given by:
\begin{equation}
\label{eq_rate_SR_1}
\mathds{E}\left[R_\mathrm{SR}\right] = \frac{W_\mathrm{m}}{\ln 2} {\mathds{E}} \left[\frac{1}{1+N_\mathrm{SR}}\right] {\mathds{E}} \left[\ln(1+\gamma_\mathrm{s})\right].
\end{equation}
Similar to Eq.~(\ref{eq_rate_SBH_derived}),
\begin{equation}
\label{eq_N_SR_aver_appendix}
{\mathds{E}} \left[\frac{1}{1+N_\mathrm{SR}}\right] = \frac{\kappa\rho_\mathrm{s}}{\lambda_\mathrm{SR}} \frac{\Gamma(\kappa-1)}{\Gamma(\kappa)} \left( 1 - \frac{1}{\left(1+\frac{\lambda_\mathrm{SR}}{\kappa \rho_\mathrm{s}}\right)^{\kappa-1}} \right).
\end{equation}
As SBSs follows PPP of density $\rho_\mathrm{s}$, the PDF of transmission distance $d_\mathrm{s}$ follows
\begin{equation}
f_{d_\mathrm{s}}(d) = \frac{\mbox{d}}{\mbox{d} d} \left( 1-e^{-\pi \rho_\mathrm{s} d^2 } \right).
\end{equation}
Thus,
\begin{subequations}
\label{eq_SE_SR}
\begin{align}
& {\mathds{E}} \left[\ln(1+\gamma_\mathrm{s})\right] = \int_{0}^{\infty} \ln \left( 1+\frac{P_\mathrm{TS}d^{-\alpha_\mathrm{s}}}{(1+\theta_\mathrm{s})\sigma^2} \right) f_{d_\mathrm{s}}(d) \mbox{d} d \nonumber\\
& \geq \ln \frac{P_\mathrm{TS}}{(1+\theta_\mathrm{s})\sigma^2} - \alpha_\mathrm{s} \int_{0}^{\infty} 2 \pi \rho_\mathrm{s} d e^{-\pi \rho_\mathrm{s} d^2} \ln d \mbox{d} d \\
& = \ln \frac{P_\mathrm{TS}(\pi \rho_\mathrm{s})^{\frac{\alpha_\mathrm{s}}{2}}}{(1+\theta_\mathrm{s})\sigma^2} - \alpha_\mathrm{s} \int_{0}^{\infty} e^{-x} \ln x \mbox{d} x \nonumber\\
& = \ln \frac{P_\mathrm{TS}(\pi \rho_\mathrm{s})^{\frac{\alpha_\mathrm{s}}{2}}}{(1+\theta_\mathrm{s})\sigma^2} + \frac{1}{2} \alpha_\mathrm{s} \gamma,
\end{align}
\end{subequations}
where $\theta_\mathrm{s} \sigma^2 = \mathds{E}\left[I_\mathrm{s}\right]$, and the equality of (\ref{eq_SE_SR}b) holds when $\frac{\sigma^2}{P_\mathrm{TS}} \rightarrow 0$.
Substituting Eqs.~(\ref{eq_N_SR_aver_appendix}) and (\ref{eq_SE_SR}) into (\ref{eq_rate_SR_1}), Lemma~2 can be proved.
\section{Proof of Table~\ref{tab_variation}}
\label{appendix_variation}
When $\varphi$ increases, the ratio of traffic load steered to MBS backhaul and radio access both decrease according to Eq.~(\ref{eq_capacity_part}a) and (\ref{eq_capacity_part}b), hence increasing $\mu_\mathrm{MR}$ and $\mu_\mathrm{MBH}$.
On the contrary, $\mu_\mathrm{SR}$ and $\mu_\mathrm{SBH}$ both decrease, according to Eq.~(\ref{eq_capacity_part}c) and (\ref{eq_capacity_part}d).
Suppose the SBS cache size $C_\mathrm{s}$ increases to $C'_\mathrm{s} = C_\mathrm{s} + \Delta_\mathrm{s}$, and the MBS cache size $C_\mathrm{m}$ decreases to $C'_\mathrm{m}=C_\mathrm{m} + \Delta_\mathrm{m}$, where $\rho_\mathrm{s} \Delta_\mathrm{s} + \rho_\mathrm{m} \Delta_\mathrm{m}=0$.
As $\rho_\mathrm{s}>\rho_\mathrm{m}$ in practical networks, $\Delta_\mathrm{s} + \Delta_\mathrm{m} < 0$.
Denote by $\Delta P_\mathrm{Hit}^{(\mathrm{m})}$ and $\Delta P_\mathrm{Hit}^{(\mathrm{s})}$ the corresponding variations of MBS-tier and SBS-tier content hit rates, respectively.
Apparently, $\Delta P_\mathrm{Hit}^{(\mathrm{m})}<0$ and $\Delta P_\mathrm{Hit}^{(\mathrm{s})}>0$, and $\Delta P_\mathrm{Hit}^{(\mathrm{m})} + \Delta P_\mathrm{Hit}^{(\mathrm{s})} < 0$ since $\Delta_\mathrm{s} + \Delta_\mathrm{m} < 0$.
Thus, the total content hit rate decreases.
Therefore, $\mu_\mathrm{MBH}$ and $\mu_\mathrm{SBH}$ both decrease, according to Eqs.~(\ref{eq_capacity_part}b) and (\ref{eq_capacity_part}d).
In addition, $\mu_\mathrm{SR}$ also decreases as $P_\mathrm{Hit}^{(\mathrm{s})}$ increases, according to Eq.~(\ref{eq_capacity_part}c).
For $\mu_\mathrm{MR}$,
\begin{equation}
\begin{split}
& \Delta P_\mathrm{Hit}^{(\mathrm{m})} - ( \Delta P_\mathrm{Hit}^{(\mathrm{m})} + \Delta P_\mathrm{Hit}^{(\mathrm{s})}) (1-\varphi)\\
& = -\Delta P_\mathrm{Hit}^{(\mathrm{s})} + (\Delta P_\mathrm{Hit}^{(\mathrm{m})} + \Delta P_\mathrm{Hit}^{(\mathrm{s})}) \varphi < 0
\end{split}
\end{equation}
and thus $\mu_\mathrm{MR}$ increases with $C_\mathrm{s}$, according to Eq.~(\ref{eq_capacity_part}a).
\section{Proof of Proposition~2}
\label{appendix_insufficient_cache}
Set $C_\mathrm{s} = C/\rho_\mathrm{s}$, $C_\mathrm{m}=0$, $\varphi=\frac{\hat{\lambda}_\mathrm{SBH}} {\hat{\lambda}_\mathrm{MR} + \hat{\lambda}_\mathrm{SBH}}$.
Accordingly, $P_\mathrm{Hit}^{(\mathrm{s})} = \sum_{f=1}^{C/\rho_\mathrm{s}} q_f$, $P_\mathrm{Hit}^{(\mathrm{m})} = 0$, and $P_\mathrm{Hit} = \sum_{f=1}^{C/\rho_\mathrm{s}} q_f$.
Substituting $P_\mathrm{Hit}^\mathrm{(m)}$ and $\varphi$ into Eqs.~(\ref{eq_capacity_part}a) and (\ref{eq_capacity_part}d), we have $\mu_\mathrm{MR} = \mu_\mathrm{SBH} = \frac{\hat{\lambda}_\mathrm{MR}+\hat{\lambda}_\mathrm{SBH}}{1-P_\mathrm{Hit}}$, and $\mu_\mathrm{SR} = \frac{\hat{\lambda}_\mathrm{SR}}{1-\frac{\hat{\lambda}_\mathrm{MR}}{\hat{\lambda}_\mathrm{MR} + \hat{\lambda}_\mathrm{SBH}} (1-P_\mathrm{Hit}) }$.
As $C < C_\mathrm{min}$,
\begin{equation}
\begin{split}
& P_\mathrm{Hit} = \sum_{f=1}^{C/\rho_\mathrm{s}} q_f < \frac{\hat{\lambda}_\mathrm{SR} - \hat{\lambda}_\mathrm{SBH}}{ \hat{\lambda}_\mathrm{MR} + \hat{\lambda}_\mathrm{SR}} \\
\Longleftrightarrow & P_\mathrm{Hit} (\hat{\lambda}_\mathrm{MR} + \hat{\lambda}_\mathrm{SR}) < \hat{\lambda}_\mathrm{SR} - \hat{\lambda}_\mathrm{SBH} \\
\Longleftrightarrow & P_\mathrm{Hit} \hat{\lambda}_\mathrm{MR} + \hat{\lambda}_\mathrm{SBH} < (1-P_\mathrm{Hit}) \hat{\lambda}_\mathrm{SR}
\end{split}
\end{equation}
Notice that
\begin{equation}
\begin{split}
& \frac{\hat{\lambda}_\mathrm{SR}}{1-\frac{\hat{\lambda}_\mathrm{MR}}{\hat{\lambda}_\mathrm{MR} + \hat{\lambda}_\mathrm{SBH}} (1-P_\mathrm{Hit}) } - \frac{\hat{\lambda}_\mathrm{MR}+\hat{\lambda}_\mathrm{SBH}}{1-P_\mathrm{Hit}} \\
& = \frac{\hat{\lambda}_\mathrm{MR}+\hat{\lambda}_\mathrm{SBH}}{1-P_\mathrm{Hit}} \left[ \frac{\hat{\lambda}_\mathrm{SR} (1-P_\mathrm{Hit}) }{ \hat{\lambda}_\mathrm{SBH} + P_\mathrm{Hit} \hat{\lambda}_\mathrm{MR}} -1 \right] >0.
\end{split}
\end{equation}
Thus, $\mu_\mathrm{SR} > \mu_\mathrm{MR} = \mu_\mathrm{SBH}$.
According to Table~\ref{tab_variation}, $\mu_\mathrm{MR}$ and $\mu_\mathrm{SBH}$ cannot be simultaneously improved.
Therefore, $C_\mathrm{s}^* = C/\rho_\mathrm{s}, \varphi^* = \frac{\hat{\lambda}_\mathrm{SBH}} {\hat{\lambda}_\mathrm{MR} + \hat{\lambda}_\mathrm{SBH}}$ is the optimal solution to problem (P2).
\section{Proof of Propositions~3}
\label{appendix_sufficient_cache}
Firstly, we prove that $\left[C_\mathrm{s}^*, \varphi^* \right]$ satisfying Eq.~(\ref{eq_P3_opt_sufficient_C}) is feasible to constraint (\ref{eq_P3}f) in problem (P2).
As $C \geq C_\mathrm{min} $, $C_\mathrm{s}^* = C_\mathrm{min}/\rho_\mathrm{s}$ is feasible to (\ref{eq_P3}f).
When $C_\mathrm{s}^* = C_\mathrm{min}/\rho_\mathrm{s}$, the SBS-tier content hit rate is ${P_\mathrm{Hit}^{(\mathrm{s})}}^* = \frac{\hat{\lambda}_\mathrm{SR} - \hat{\lambda}_\mathrm{SBH} }{\hat{\lambda}_\mathrm{MR} +\hat{\lambda}_\mathrm{SR}}$, according to the definition of $C_\mathrm{min}$ in Eq.~(\ref{eq_C_min_max}).
As $C < C_\mathrm{max}/\rho_\mathrm{s}$, ${P_\mathrm{Hit}^{(\mathrm{m})}}^* < \frac{\hat{\lambda}_\mathrm{MR} }{\hat{\lambda}_\mathrm{MR} +\hat{\lambda}_\mathrm{SR}}$.
Thus, ${1-P_\mathrm{Hit}}^* > \frac{\hat{\lambda}_\mathrm{SBH} }{\hat{\lambda}_\mathrm{MR} +\hat{\lambda}_\mathrm{SR}}$, and $\varphi \in (0,1)$ is feasible to (\ref{eq_P3}f) in (P2).
Then, we prove that the network capacity achieves the maximum under $\left[C_\mathrm{s}^*, \varphi^* \right]$.
Substituting Eq.~(\ref{eq_P3_opt_sufficient_C}) into (\ref{eq_capacity_part}), we have $\mu^*_\mathrm{MR} = \mu^*_\mathrm{SR} = \mu^*_\mathrm{SBH} = \hat{\lambda}_\mathrm{MR}+\hat{\lambda}_\mathrm{SR}$.
Thus, the network capacity is $\hat{\lambda}_\mathrm{MR}+\hat{\lambda}_\mathrm{SR}$.
Adding constraints (\ref{eq_P3}b) and (\ref{eq_P3}d), we can prove that the maximal network capacity cannot exceed $\hat{\lambda}_\mathrm{MR}+\hat{\lambda}_\mathrm{SR}$.
Therefore, Eq.~(\ref{eq_P3_opt_sufficient_C}) guarantees the optimality of capacity.
In addition, we prove that the network capacity is smaller than $\hat{\lambda}_\mathrm{MR} +\hat{\lambda}_\mathrm{SR}$ if $C_\mathrm{s} \leq C_\mathrm{s}^*$, by contradiction.
Assume there exist a solution $[C_\mathrm{s}', \varphi']$ with network capacity of $\hat{\lambda}_\mathrm{MR} +\hat{\lambda}_\mathrm{SR}$, where $C_\mathrm{s}'\leq C_\mathrm{min}/\rho_\mathrm{s}$.
According to Eqs.~(\ref{eq_P3_opt_sufficient_C}a) and (\ref{eq_P3_opt_sufficient_C}d), we have
\begin{subequations}
\label{eq_app_P3_opt_proof_1}
\begin{align}
& {P_\mathrm{Hit}^{(\mathrm{M})}}' + (1-P'_\mathrm{Hit}) (1 - \varphi') \leq \frac{\hat{\lambda}_\mathrm{MR}}{ \hat{\lambda}_\mathrm{MR}+ \hat{\lambda}_\mathrm{SR}}, \\
& (1-P'_\mathrm{Hit}) \varphi' \leq \frac{\hat{\lambda}_\mathrm{SBH}}{ \hat{\lambda}_\mathrm{MR}+ \hat{\lambda}_\mathrm{SR}}.
\end{align}
\end{subequations}
In addition
\begin{subequations}
\label{eq_app_P3_opt_proof_2}
\begin{align}
& {P_\mathrm{Hit}^{(\mathrm{M})}}' + (1-P'_\mathrm{Hit}) (1 - \varphi') \nonumber \\
& \geq 1-{P_\mathrm{Hit}^{(\mathrm{s})}}' - \frac{\hat{\lambda}_\mathrm{SBH}}{ \hat{\lambda}_\mathrm{MR}+ \hat{\lambda}_\mathrm{SR}} \\
& > 1 - {P_\mathrm{Hit}^{(\mathrm{s})}}^* - (1-P^*_\mathrm{Hit})\varphi^* \\
& = \frac{\hat{\lambda}_\mathrm{MR}}{ \hat{\lambda}_\mathrm{MR}+ \hat{\lambda}_\mathrm{SR}},
\end{align}
\end{subequations}
where (\ref{eq_app_P3_opt_proof_2}a) is based on (\ref{eq_app_P3_opt_proof_1}b), (\ref{eq_app_P3_opt_proof_2}b) is due to condition (\ref{eq_P3_opt_sufficient_C}b), (\ref{eq_app_P3_opt_proof_2}c) holds as $P_\mathrm{Hit}^{(\mathrm{s})}$ increases with $C_\mathrm{s}$, and (\ref{eq_app_P3_opt_proof_2}d) comes from (\ref{eq_P3_opt_sufficient_C}).
As (\ref{eq_app_P3_opt_proof_2}) is contradictory with (\ref{eq_app_P3_opt_proof_1}a), there exists no $C_\mathrm{s}'\leq C_\mathrm{min}/\rho_\mathrm{s}$ to achieve the maximal network capacity.
As content hit rate decreases with SBS cache size, $[C_\mathrm{s}^*,\varphi^*]$ achieves the maximal content hit rate among the capacity-optimal solutions.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
A constrained percolation model is a probability measure on subgraphs of a lattice, satisfying certain local constraints. Each subgraph is called a configuration. These models are abstract mathematical models for ubiquitous phenomena in nature, and have been interesting topics in mathematical and scientific research for long.
Examples of constrained percolation models include the dimer model (see \cite{RK09}), the 1-2 model (see \cite{GrLReview}), the six-vertex model (or 6V model, see \cite{Bax08,BCG16,KMSW16}), and general vertex models (see \cite{Val,SB,ZLlv}). The study of these models may give deep insights to human's understanding of many natural phenomena, such as structure of matter, phase transition, limit shape, and critical behavior.
We are interested in the classical percolation problem in a constrained model: under which probability measure does there exist an infinite connected set (infinite cluster) in which every vertex is present in the random configuration, or equivalently, included in the randomly-chosen subgraph? Such a question has been studied extensively in the unconstrained case - in particular the i.i.d Bernoulli percolation - see, for instance, \cite{ha60,he80,HS94,blps99,fh15,GrPc}. The major difference between the constrained percolation and the unconstrained percolation lies in the fact that imposing local constraints usually makes stochastic monotonicity, which is a crucial property when studying the unconstrained model, invalid. Therefore new techniques need to be developed to study constrained percolation models.
Some constrained percolation models, including the 1-2 model, the periodic plane dimer model, certain 6V models, may be exactly solvable; see \cite{KOS06,ZLejp,ZL12,GrL15,BCG16}. The integrability properties of these models make it possible to compute the local correlations. When the parameters associated to the probability measure vary, different behaviors of the local correlations imply a phase transition. If we consider phase transitions from a macroscopic, or geometric point of view, different approaches may be applied to study the existence of infinite clusters for a large class of constrained percolation models.
In \cite{HL16}, we studied a constrained percolation model on the $\ZZ^2$ lattice, and showed that if the underlying probability measure satisfies mild assumptions like symmetry, ergodicity and translation-invariance, then with probability 0 the number of infinite clusters is nonzero and finite. The technique makes use of the planarity and amenability of the 2D square grid $\ZZ^2$. As an application, we obtained percolation properties for the XOR Ising model (a random spin configuration on a graph in which each spin is the product of two spins from two i.i.d Ising models, see \cite{DBW11}) on $\ZZ^2$, with the help of the combinatorial correspondence between the XOR Ising model and the dimer model proved in \cite{Dub,bd14}. In this paper, we further develop the technique to study constrained percolation models on a number of planar lattices, which may be amenable or non-amenable, including the $[m,4,n,4]$ lattice and the square tilings of the hyperbolic plane; see \cite{CFKP97} for an introduction to hyperbolic geometry. The general results we obtain can be used to prove further results concerning percolation properties of the XOR Ising model on the hexagonal and the triangular lattices, as well as on regular tilings of the hyperbolic plane.
The specific geometric properties of non-amenable graphs make it an interesting problem to study percolation models on such graphs; and a set of techniques have been developed in the past few decades; see \cite{BS96,blps,blps99,HP,LS99,Sch99,HPS,PSN,Wu00,RS01,HJL02,NP12,LP} for an incomplete list.
In this paper, we also study the general automorphism-invariant percolation models on transitive planar graphs. Our results may be related to the following two conjectures:
\begin{conjecture}(Conjecture 7 of \cite{BS96})\label{cj1} Suppose that $G$ is a planar, connected graph, and the minimal vertex degree in $G$ is at least 7. In an i.i.d Bernoulli site percolation on $G$, the vertices are open (resp.\ closed) with probability $p$ (resp.\ $1-p$) independently, where $p\in[0,1]$. The critical probability $p_c$ is the supreme of $p$'s such that almost surely there are no infinite open clusters. Then at every $p$ in the range $(p_c,1-p_c)$, there are infinitely many infinite open clusters in the i.i.d Bernoulli site percolation on $G$. Moreover, we conjecture that $p_c<\frac{1}{2}$, and the above interval is nonempty.
\end{conjecture}
In \Cref{exl2}, we explain why \Cref{cj1} is true for the i.i.d Bernoulli site percolation on vertex-transitive triangular tilings of the hyperbolic plane where each vertex has degree $n\geq 7$.
\begin{conjecture}(Conjecture 8 of \cite{BS96})\label{cj2} Let $G$ be a planar, connected graph. Let $p=\frac{1}{2}$ be the probability that a vertex is open and assume that a.s. percolation occurs in the site percolation on $G$. Then almost surely there are infinitely many infinite clusters.
\end{conjecture}
Our \Cref{p118} implies that \Cref{cj2} is true for automorphism-invariant site percolation (not necessarily independent, or insertion tolerant) on vertex-transitive triangular tilings of the hyperbolic plane where each vertex has degree $n\geq 7$ if the underlying measure is ergodic and invariant under switching state-1 vertices and state-0 vertices.
We then apply our results concerning the general automorphism-invariant percolation models on transitive planar graphs to study the infinite ``$+$''-clusters and ``$-$''-clusters for the Ising model on vertex-transitive triangular tilings of the hyperbolic plane where each vertex has degree $n\geq 7$, and describe the behaviors of such clusters with respect to varying coupling constants under the free boundary condition and the wired boundary condition. A surprising result we obtain is that it is possible that the random cluster representation of the Ising model has no infinite open clusters, while the Ising model has infinitely many infinite ``$+$''-clusters and infinitely many infinite ``$-$''-clusters - in contrast with the Ising percolation and its random cluster representation on the 2d square grid $\ZZ^2$ (see \cite{crpr76,Hig93,GrGrc}) where the Ising model has an infinite ``$+$'' or ``$-$''-cluster if and only if its random cluster representation has an infinite open cluster.
The organization of the paper is as follows.
In \Cref{xorh}, we introduce the $[m,4,n,4]$ lattice and state the result concerning constrained percolation models on the $[m,4,n,4]$ lattice. In \Cref{rcq}, we review the multiple phase transitions of random cluster models with $q\geq 1$ on some non-amenable planar graphs, as well as the coupling of a random cluster model with $q=2$ and an Ising model. In \Cref{Is}, we state the main results concerning infinite clusters in the Ising model on regular triangular tilings of the hyperbolic plane, and, in particular, provide a description of the numbers of infinite ``$+$'' and ``$-$'' clusters of the ferromagnetic Ising model with the free boundary condition, the ``$+$" boundary condition or the ``$-$" boundary condition on the triangular tilings of the hyperbolic plane for different values of coupling constants. In \Cref{xIs}, we state the main results concerning infinite clusters in the XOR Ising model on regular triangular tilings of the hyperbolic plane and its dual graph. In \Cref{db}, we introduce the $[4,6,12]$ lattice and state the main result concerning the dimer model on the $[4,6,12]$ lattice. In Section \ref{xori}, we state the result proved in this paper concerning the percolation properties of the XOR Ising model on the hexagonal lattice and the triangular lattice. In \Cref{sthp}, we introduce the square tilings of the hyperbolic plane, state and prove the main result concerning constrained percolation models on square tilings of the hyperbolic plane.
The remaining sections are devoted to prove the theorems stated in preceding sections.
In \Cref{ctcl}, we prove combinatorial results concerning contours and clusters in preparation to prove \Cref{m23,m21,m22}. In \Cref{pnc}, we discuss known results about percolation on non-amenable graphs that will be used to prove \Cref{m23,m21,m22}. In \Cref{p23}, we prove \Cref{m23}. In \Cref{p212}, we prove \Cref{m21}. In \Cref{p211}, we prove \Cref{m22}. In \Cref{peh}, we discuss the applications of the techniques developed in the proof of \Cref{m23}
to prove results concerning unconstrained site percolation on vertex-transitive, triangular tilings of the hyperbolic plane in preparation of proving \Cref{ipl,coii,xorc}. In \Cref{pipl}, we prove \Cref{ipl}. In \Cref{pxorc}, we prove \Cref{coii} and \Cref{xorc}.
In \Cref{pm31}, we prove \Cref{m31}. In \Cref{p412}, we prove \Cref{chi,lth}.
\section{Constrained percolation on the $[m,4,n,4]$ lattice }\label{xorh}
In this section, we state the main result proved in this paper for the constrained percolation models on the $[m,4,n,4]$ lattice. We shall start with a formal definition of the $[m,4,n,4]$ lattice.
Let $m,n$ be positive integers satisfying
\begin{eqnarray}
&& m\geq 3,\qquad n\geq 3\label{cmn1}\\
&&\frac{1}{m}+\frac{1}{n}\leq \frac{1}{2}.\label{cmn2}
\end{eqnarray}
The $[m,4,n,4]$ lattice is a vertex-transitive graph which can be embedded into the Euclidean plane or the hyperbolic plane such that each vertex is incident to 4 faces with degrees $m,4,n,4$ in cyclic order. When $\frac{1}{m}+\frac{1}{n}=\frac{1}{2}$, the graph is amenable and can be embedded into the Euclidean plane. When $\frac{1}{m}+\frac{1}{n}<\frac{1}{2}$, the graph is non-amenable and can be embedded into the hyperbolic plane (\cite{DR}). Note that when $m=n=4$, the graph is the square grid embedded into the 2D Euclidean plane. See Figure \ref{3464} for an illustration of the [3,4,6,4] lattice, and \Cref{3474} for the [3,4,7,4] lattice.
\begin{figure}
\includegraphics[width=.5\textwidth]{3464}
\caption{The [3,4,6,4] lattice, the auxiliary hexagonal lattice and triangular lattice. Black lines represent the [3,4,6,4] lattice; dashed red lines represent the triangular lattice; dashed blue lines represent the hexagonal lattice.}
\label{3464}
\end{figure}
\begin{figure}
\includegraphics[width=.6\textwidth]{1280px-H2_tiling_237-5}
\caption{The [3,4,7,4] lattice (picture from the wikipedia)}
\label{3474}
\end{figure}
Let $G=(V,E)$ be an $[m,4,n,4]$ lattice. We color all the faces with degree $m$ or $n$ by white and all the other faces by black, such that any two faces sharing an edge have different colors. We consider the site percolation on $V$ satisfying the following constraint (see Figure \ref{lcc}):
\begin{itemize}
\item around each black face, there are six allowed configurations $(0000)$, $(1111)$, $(0011)$, $(1100)$, $(0110)$, $(1001)$, where the digits from the left to the right correspond to vertices in clockwise order around the black face, starting from the lower left corner. See Figure \ref{lcc}.
\end{itemize}
\begin{figure}
\subfloat[0000]{\includegraphics[width=.12\textwidth]{lcp0000}}\qquad\qquad\qquad\qquad
\subfloat[0011]{\includegraphics[width = .12\textwidth]{lcp1100}}\qquad\qquad\qquad\qquad
\subfloat[0110]{\includegraphics[width = .12\textwidth]{lcp0110}}\\
\subfloat[1111]{\includegraphics[width = .12\textwidth]{lcp}} \qquad\qquad\qquad\qquad
\subfloat[1100]{\includegraphics[width = .12\textwidth]{lcp0011}}\qquad\qquad\qquad\qquad
\subfloat[1001]{\includegraphics[width = .12\textwidth]{lcp1001}}
\caption{Local configurations of the constrained percolation around a black square. Red and blue lines mark contours separating 0's and 1's (in $\LL_1$ and $\LL_2$ respectively). Yellow (resp.\ green) disks represent 0's (resp.\ 1's).}
\label{lcc}
\end{figure}
Let $\Omega\subset\{0,1\}^V$ be the probability space consisting of all the site configurations on $G$ satisfying the constraint above.
To the $[m,4,n,4]$ lattice $G$, we associate two auxiliary lattices $\LL_1=(V(\LL_1),E(\LL_1))$ and $\LL_2=(V(\LL_2),E(\LL_2))$ as follows. Each vertex of $\LL_1$ (resp.\ $\LL_2$) is located at the center of each degree-$m$ face (resp. degree-$n$ face) of $G$. Two vertices of $\LL_1$ (resp.\ $\LL_2$) are joined by an edge of $\LL_1$ (resp.\ $\LL_2$) if and only if the two corresponding $m$-faces (resp.\ $n$-faces) of $G$ are adjacent to the same square face of $G$ through a pair of opposite edges (edges of a square face that do not share a vertex), respectively.
We say an edge $e\in E(\LL_1)\cup E(\LL_2)$ \textbf{crosses} a square face of the $[m,4,n,4]$ lattice if the edge $e$ crosses a pair of opposite edges of the square face.
Note that
\begin{enumerate}[label=\roman*]
\item $\LL_1$ (resp.\ $\LL_2$) is a planar lattice in which each face has degree $n$ (resp.\ $m$) and each vertex has degree $m$ (resp.\ $n$).
\item $\LL_1$ and $\LL_2$ are planar dual to each other.
\item Each edge in $E(\LL_1)\cup E(\LL_2)$ crosses a unique square face of the $[m,4,n,4]$ lattice. When $m\neq 4$ and $n\neq 4$, each square face of the $[m,4,n,4]$ lattice is crossed by a unique edge $e_1\in E(\LL_1)$ and a unique edge $e_2\in E(\LL_2)$; and moreover, $e_1$ and $e_2$ are dual to each other.
\end{enumerate}
When $m=n$, both $\LL_1$ and $\LL_2$ are lattices in which each face has degree $n$ and each vertex has degree $n$.
When $m=3$ and $n=6$, $\LL_1$ is the hexagonal lattice and $\LL_2$ is the triangular lattice; see \Cref{3464}.
Let $\Phi\subset \{0,1\}^{E(\LL_1)\cup E(\LL_2)}$ be the set of contour configurations satisfying the condition that each vertex of $V(\LL_1)$ and $V(\LL_2)$ is incident to an even number of present edges, and present edges in $E(\LL_1)$ and $E(\LL_2)$ never cross. Any constrained percolation configuration $\omega\in \Omega$ is mapped to a contour configuration $\phi(\omega)\in \Phi$, where an edge $e$ in $E(\LL_1)$ or $E(\LL_2)$ is present (i.e., have state 1) if and only if the following condition holds
\begin{itemize}
\item Let $S$ be the square face of $G$ crossed by $e$. Let $v_1,v_2,v_3,v_4$ be the four vertices of $S$, such that $v_1$ and $v_2$ are on one side of $e$ and $v_3$, $v_4$ are on the other side of $e$. Then $v_1$ and $v_2$ have the same state, $v_3$ and $v_4$ have the same state, and $v_1$ and $v_3$ have different states.
\end{itemize}
See \Cref{3464contour} for a contour configuration obtained from a constrained percolation configuration on the $[3,4,6,4]$ lattice. Note that the mapping $\phi:\Omega\rightarrow \Phi$ is 2-to-1 since $\phi(\omega)=\phi(1-\omega)$.
A \df{contour} is a connected component of present edges in a contour configuration in $\Phi$. A contour may be finite or infinite depending on the number of edges in the contour. Since present edges of a contour configuration in $E(\LL_1)$ and in $E(\LL_2)$ never cross, either all the edges in a contour are edges of $\LL_1$, or all the edges in a contour are edges in $\LL_2$. We call a contour \df{primal contour} (resp.\ \df{dual contour}) if all the edges in the contour are edges of $\LL_1$ (resp.\ $\LL_2$).
\begin{figure}
\includegraphics[width=.5\textwidth]{3464contour}
\caption{A constrained percolation configuration on the [3,4,6,4] lattice. Red lines represent contours on the triangular lattice. Blue lines represent contours on the hexagonal lattice.}
\label{3464contour}
\end{figure}
Let $\Gamma$ be the automorphism group $\mathrm{Aut}(G)$ of the graph $G$. For $i\in\{1,2\}$, let $\Gamma_i$ be the automorphism group $\mathrm{Aut}(\LL_i)$ of the graph $\LL_i$. Let $\mu$ be a probability measure on $\Omega$. We may assume that $\mu$ satisfies the following conditions:
\begin{enumerate}[label=({A}{\arabic*})]
\item $\mu$ is $\Gamma$-invariant;
\item $\mu$ is $\Gamma_i$-ergodic for $i=1,2$; i.e.\ any $\Gamma_i$-invariant event has $\mu$-probability 0 or 1;
\item $\mu$ is symmetric: let $\theta:\Omega\rightarrow\Omega$ be the map defined by $\theta(\omega)(v)=1-\omega(v)$, for each $v\in \ZZ^2$, then $\mu$ is invariant under $\theta$, that is, for any event $A$, $\mu(A)=\mu(\theta(A))$.
\end{enumerate}
Let $\Phi_1$ (resp.\ $\Phi_2$) be the set of all contour configurations on $\LL_1$ (resp.\ $\LL_2$) satisfying the condition that each vertex of $\LL_1$ (resp.\ $\LL_2$) has an even number of incident present edges. For each contour configuration $\psi\in\Phi$, we have $\psi=\psi_1\cup\psi_2$, where $\psi_1\in \Phi_1$ and $\psi_2\in\Phi_2$; moreover, $\psi_1\cap\psi_2=\emptyset$.
Let $\nu_1$ (resp.\ $\nu_2$) be the marginal distribution of $\mu$ on $\Phi_1$ (resp.\ $\Phi_2$). When $\frac{1}{m}+\frac{1}{n}=\frac{1}{2}$, the $[m,4,n,4]$ lattice is amenable.
It is not hard to see that if $\frac{1}{m}+\frac{1}{n}=\frac{1}{2}$, then $(m,n)\in \{(4,4),(3,6),(6,3)\}$. When $m=n=4$, the $[m,4,n,4]$ lattice is the 2D square grid, on which the constrained percolation models was discussed in \cite{HL16}.
Now we consider the case when $(m,n)=(3,6)$. As discussed before, in this case $\LL_1$ is the hexagonal lattice $\HH$, and $\LL_2$ is the triangular lattice $\TT$.
We may assume that $\nu_1$ or $\nu_2$ has \df{finite energy} as follows.
\begin{enumerate}[label=({A}{\arabic*})]
\setcounter{enumi}{3}
\item $\nu_1$ has \df{finite energy} in the following sense: let $S$ be a hexagon face of $\HH=(V(\HH),E(\HH))$, and $\partial S\subset E(\HH)$ be the set of six sides of the hexagon $S$. Let $\phi\in \Phi_{1}$. Define $\phi_S$ to be the configuration obtained by switching the states of each element of $\partial S$, i.e.\ $\phi_S(e)=1-\phi(e)$ if $e\in\partial S$, and $\phi_S(e)=\phi(e)$ otherwise. Let $E$ be an event, and define
\begin{eqnarray}
E_S=\{\phi_S:\phi\in E\}.\label{ef}
\end{eqnarray}
Then $\nu_1(E_S)>0$ whenever $\nu_{1}(E)>0$.
\item $\nu_2$ has \df{finite energy} in the following sense: let $S$ be a triangle face of $\TT=(V(\TT),E(\TT))$, and $\partial S\subset E(\TT)$ be the set of three sides of the triangle $S$. Let $\phi\in \Phi_{2}$. Define $\phi_S$ to be the configuration obtained by switching the states of each element of $\partial S$, i.e.\ $\phi_S(e)=1-\phi(e)$ if $e\in\partial S$, and $\phi_S(e)=\phi(e)$ otherwise. Let $E$ be an event, and define
\begin{eqnarray}
E_S=\{\phi_S:\phi\in E\}.\label{ef}
\end{eqnarray}
Then $\nu_2(E_S)>0$ whenever $\nu_2(E)>0$.
\end{enumerate}
See \Cref{feh,fet} for illustrations of the configuration-changing process on the hexagonal lattice $\HH$ and the triangular lattice $\TT$, respectively.
\begin{figure}
\centering
\includegraphics[width=0.2\textwidth]{feh1}\qquad\qquad\qquad\qquad \includegraphics[width=0.2\textwidth]{feh2}
\caption{Change of contour configurations in $\LL_1=\HH$}\label{feh}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.15\textwidth]{fet1}\qquad\qquad\qquad\qquad\qquad \includegraphics[width=0.15\textwidth]{fet2}
\caption{Change of contour configurations in $\LL_2=\TT$}\label{fet}
\end{figure}
For a random contour configuration $\psi\in\Phi_1$ (resp.\ $\psi\in \Phi_2$) whose distribution is the marginal distribution $\nu_1$ (resp.\ $\nu_2$) of $\mu$ on $\Phi_1$ (resp.\ $\Phi_2$), $\psi$ induces a random constrained configuration $\omega\in \phi^{-1}(\psi)$ as follows. Let $v_0$ be a fixed vertex of $G$. Assume that $\omega(v_0)=1$ with probability $\frac{1}{2}$, and $\omega(v_0)=0$ with probability $\frac{1}{2}$, and is independent of $\psi$. For two vertices $v_1,v_2$ of $G$ joined by an edge $e$, $v_1$ and $v_2$ have different states if and only if $e$ crosses a present edge in $\psi$. Let $\lambda_1$ (resp.\ $\lambda_2$) be the distribution of $\omega$. We may further make the following assumptions
\begin{enumerate}[label=({A}{\arabic*})]
\setcounter{enumi}{5}
\item $\lambda_1$ is $\Gamma_1$-ergodic;
\item $\lambda_2$ is $\Gamma_2$-ergodic.
\end{enumerate}
Also we may sometimes assume that
\begin{enumerate}[label=({A}{\arabic*})]
\setcounter{enumi}{7}
\item $\mu$ is $\Gamma_1$-invariant.
\end{enumerate}
The main theorems of this section are stated as follows.
\begin{theorem}\label{m23}Let $G$ be the $[3,4,n,4]$ lattice with $n\geq 7$. Let $s_0$ (resp.\ $s_1$) be the number of infinite 0-clusters (resp.\ 1-clusters). Let $t_1$ (resp.\ $t_2$) be the number of infinite $\LL_1$-contours (resp.\ $\LL_2$-contours).
\begin{enumerate}
\item Let $\mu$ be a probability measure on $\Omega$ satisfying (A2),(A3),(A7),(A8). Then $\mu$-a.s. $(s_0,s_1,t_1)=(\infty,\infty,\infty)$.
\item Let $\mu$ be a probability measure on $\Omega$ satisfying (A2),(A3),(A6),(A7),(A8). Then $\mu$-a.s. $(s_0,s_1,t_1,t_2)\in\{(\infty,\infty,\infty,1),(\infty,\infty,\infty,\infty)\}$.
\end{enumerate}
\end{theorem}
\begin{theorem}\label{m21}Let $G$ be the $[m,4,n,4]$ lattice such that
\begin{eqnarray}
&&m\geq 3,\qquad n\geq 3;\label{mn1}\\
&&\frac{1}{m}+\frac{1}{n}=\frac{1}{2}\notag
\end{eqnarray}
Let $\mu$ be a probability measure on $\Omega$. Then
\begin{enumerate}
\item if $\mu$ satisfies (A1)-(A6), then almost surely there are no infinite contours in $\LL_2$;
\item if $\mu$ satisfies (A1)-(A5) and (A7), then almost surely there are no infinite contours in $\LL_1$;
\item if $\mu$ satisfies (A1)-(A7), almost surely there are neither infinite contours nor infinite clusters.
\end{enumerate}
\end{theorem}
\begin{theorem}\label{m22}Let $G$ be the $[m,4,n,4]$ lattice satisfying
\begin{eqnarray*}
m=n\geq 5.
\end{eqnarray*}
Let $\mu$ be a probability measure on $\Omega$. Let $s_0$ (resp.\ $s_1$) be the number of infinite 0-clusters (resp.\ 1-clusters), and let $t_1$ (resp.\ $t_2$) be the number of infinite $\LL_1$-contours (resp.\ $\LL_2$-contours). If $\mu$ satisfies (A1)-(A3); then $\mu$-a.s. $(s_0,s_1,t_1,t_2)=(\infty,\infty,\infty,\infty)$.
\end{theorem}
\Cref{m23} is proved in \Cref{p23}. \Cref{m21} is proved in \Cref{p212}, and \Cref{m22} is proved in \Cref{p211}.
\section{Random-cluster model on a non-amenable graph with $q\geq 1$}\label{rcq}
In this section, we summarize basic facts about the Fortuin-Kasteleyn random cluster model, which is a probability measure on bond configurations of a graph, and the related Potts model. See \cite{GrGrc} for more information.
The \textbf{random cluster measure} $RC:=RC_{p,q}^{G_0}$ on a finite graph $G_0=(V_0,E_0)$ with parameters $p\in[0,1]$ and $q\geq 1$ is the probability measure on $\{0,1\}^{E_0}$ which to each $\xi\in \{0,1\}^{E_0}$ assigns probability
\begin{eqnarray}
RC(\xi):\propto q^{k(\xi)}\prod_{e\in E_0}p^{\xi(e)}(1-p)^{1-\xi(e)}.\label{drc}
\end{eqnarray}
where $k(\xi)$ is the number of connected components in $\xi$.
Let $G=(V,E)$ be an infinite, locally finite, connected graph. For each $q\in[1,\infty)$ and each $p\in (0,1)$, let $WRC_{p,q}^G$ be the random cluster measure with the wired boundary condition, and let $FRC_{p,q}^G$ be the random cluster measure with the free boundary condition. More precisely, $WRC_{p,q}^G$ (resp.\ $FRC_{p,q}^G$) is the weak limit of $RC$'s defined by (\ref{drc}) on larger and larger finite subgraphs approximating $G$, where we assume that all the edges outside each finite subgraph are present (resp.\ absent).
Let $e\in E$ be an edge of $G$ with endvertices $x,y\in V$.
It is known that $FRC_{p,q}^G$ and $WRC_{p,q}^G$ admit conditional probabilities satisfying
\begin{eqnarray}
&&FRC_{p,q}^G(X(e)=1|X(E\setminus \{e\})=\xi)=\left\{\begin{array}{cc}p,&\mathrm{if}\ x\leftrightarrow y\\ \frac{p}{p+(1-p)q},&\mathrm{otherwise} \end{array}\right.\label{fs}\\
&&WRC_{p,q}^G(X(e)=1|X(E\setminus \{e\})=\xi)=\left\{\begin{array}{cc}p,&\mathrm{if}\ x\leftrightarrow y,\ \mathrm{or}\ x\leftrightarrow\infty\leftrightarrow y\\ \frac{p}{p+(1-p)q},&\mathrm{otherwise} \end{array};\right.\label{ws}
\end{eqnarray}
where $x\leftrightarrow y$ means that there exists an open path (each edge along the path has state 1) joining $x$ and $y$ in $E\setminus \{e\}$, and $x\leftrightarrow \infty\leftrightarrow y$ means there exists an infinite open path passing through $x$, and an infinite open path passing through $y$. See expressions (6) and (7) in \cite{HJL02}.
When $q=1$, $WRC_{p,q}^G=FRC_{p,q}^G$, which is exactly the i.i.d. Bernoulli bond percolation measure on $G$.
The Gibbs measure $\mu^{+}$ (resp.\ $\mu^{-}$) for the Ising model on $G$ with coupling constant $J\geq 0$ on each edge and ``$+$''-boundary conditions (resp.\ ``$-$''-boundary conditions) can be obtained by considering a random configuration of present and absent edges according to the law $WRC_{p,2}^G$, $p=1-e^{-2J}$, and assigning to all the vertices in each infinite cluster the state ``$+$'' (resp.\ ``$-$''), and to all the vertices in each finite cluster a state from $\{+,-\}$, chosen uniformly at random for each cluster and independently for different clusters.
The Gibbs measure $\mu^{f}$ for the Ising model on $G$ with coupling constant $J\geq 0$ on each edge and free boundary conditions can be obtained by considering a random configuration of present and absent edges according to the law $FRC_{p,2}^G$, $p=1-e^{-2J}$, and assigning to all the vertices in each cluster a state from $\{+,-\}$, chosen uniformly at random for each cluster and independently for different clusters.
When there is no confusion, we may write $FRC_{p,q}^G$ and $WRC_{p,q}^G$ as $FRC_{p,q}$ and $WRC_{p,q}$ for simplicity.
Assume that $G$ is transitive. Then measures $FRC_{p,q}$ and $WRC_{p,q}$ are $\mathrm{Aut}(G)$-invariant, and $\mathrm{Aut}(G)$-ergodic; see the explanations on Page 295 of \cite{RS01}.
If we further assume that $G$ is unimodular, nonamenable and planar,
it is known that there exists $p_{c,q}^{w},\ p_{c,q}^{f},\ p_{u,q}^{w},\ p_{c,q}^{f}\in[0,1]$, such that $FRC_{p,q}$-a.s. the number of infinite clusters equals
\begin{eqnarray}
\left\{\begin{array}{cc}0&\mathrm{for}\ p\leq p_{c,q}^f\\ \infty&\mathrm{for}\ p\in (p_{c,q}^f,p_{u,q}^f)\\1&\mathrm{for}\ p> p_{u,q}^f; \end{array}\right.\label{frc}
\end{eqnarray}
and $WRC_{p,q}$-a.s. the number of infinite clusters equals
\begin{eqnarray}
\left\{\begin{array}{cc}0&\mathrm{for}\ p< p_{c,q}^w\\ \infty&\mathrm{for}\ p\in (p_{c,q}^w,p_{u,q}^w)\\1&\mathrm{for}\ p\geq p_{u,q}^w.\end{array}\right.;\label{wrc}
\end{eqnarray}
see expressions (17),(18), Theorem 3.1 and Corollary 3.7 of \cite{HJL02}.
\section{Ising model on transitive, triangular tilings of the hyperbolic plane}\label{Is}
In this section, we state the main result concerning the percolation properties of the Ising model on transitive, triangular tilings of the hyperbolic plane. These results, as given in \Cref{ipl}, will be proved in \Cref{pipl}.
The random cluster representation of an Ising model on a transitive, triangular tiling of the hyperbolic plane can be defined as in \cite{HJL02}; see also \Cref{rcq}.
It is well known that for the i.i.d Bernoulli percolation on a infinite, connected, locally finite transitive graph $G$, there exist $p_c,p_u$ such that
\begin{enumerate}[label=(\roman*)]
\item $0<p_c\leq p_u\leq 1$;
\item for $p\in[0,p_c)$ there is no infinite cluster a.s.
\item for $p\in(p_c,p_u)$ there are infinitely many infinite clusters, a.s.
\item for $p\in(p_u,1]$, there is exactly one infinite cluster, a.s.
\end{enumerate}
The monotonicity in $p$ of the uniqueness of the infinite cluster was proved in \cite{HP,Sch99}. It is proved that $p_c=p_u$ for amenable transitive graphs (see \cite{BS96}); and conjectured that $p_c<p_u$ for transitive non-amenable graphs. The conjecture $p_c<p_u$ was proved for transitive planar graphs (see \cite{BS20}) and non-amenable Cayley graphs with small spectral radii (see \cite{PSN,RS01,Th15}) or large girths (see \cite{NP12}).
\begin{theorem}\label{ipl}Let $\LL_2$ be the triangulation of the hyperbolic plane such that each vertex has degree $n\geq 7$. Consider the Ising model with spins located on vertices of $\LL_2$ and coupling constant $J\in \RR$ on each edge. Let $p_c$, $p_u$ be critical i.i.d Bernoulli site percolation probabilities on $\LL_2$ as defined by (i)-(iv) above.
\begin{enumerate}
\item Let $h>0$ satisfy
\begin{eqnarray}
\frac{e^{-h}}{e^{h}+e^{-h}}=p_c\label{pch}
\end{eqnarray}
Let $\mu^+$ (resp.\ $\mu^-$. $\mu^f$) be the infinite-volume Ising Gibbs measure with ``$+$''-boundary conditions (resp.\ ``$-$'' boundary conditions, free boundary conditions). If
\begin{eqnarray}
n|J|<h,\label{jh}
\end{eqnarray}
then $\mu$-a.s. there are infinitely many infinite ``$+$''-clusters, infinitely many infinite ``$-$''-clusters and infinitely many infinite contours, where $\mu$ is an arbitrary $\mathrm{Aut}(\LL_2)$-invariant Gibbs measure for the Ising model on $\LL_2$ with coupling constant $J$.
\item Assume $J\geq 0$. If one of the following conditions
\begin{enumerate}[label=(\alph*)]
\item $\mu^f$ is $\mathrm{Aut}(\LL_2)$-ergodic;
\item $\mathrm{inf}_{u,v\in V(\LL_2)}\langle \sigma_{u}\sigma_{v}\rangle_{\mu^f}=0$, where $\sigma_{u}$ and $\sigma_{v}$ are two spins associated to vertices $u,v\in V(\LL_2)$ in the Ising model;
\item $0\leq J<\frac{1}{2}\ln\left(\frac{1}{1-p_{u,2}^{f}}\right)$, where $p_{u,2}^{f}$ is the critical probability for the existence of a unique infinite open cluster of the corresponding random cluster representation of the Ising model on $\LL_2$, with free boundary conditions as given in (\ref{frc});
\item $0\leq J<\frac{1}{2}\ln\left(\frac{1}{1-p_{u,1}}\right)$, where $p_{u,1}$ is the critical probability for the existence of a unique infinite open cluster for the i.i.d Bernoulli bond percolation on $\LL_2$;
\end{enumerate}
holds, then $\mu^f$-a.s. there are infinitely many infinite ``$+$''-clusters and infinitely many infinite ``$-$''-clusters.
\item Assume $J\geq 0$. If
\begin{eqnarray}
J< \frac{1}{2}\ln\left(\frac{1}{1-p_{c,2}^{w}}\right),\label{jj3}
\end{eqnarray}
then for any Gibbs measure $\mu$ for the Ising model on $\LL_2$ with coupling constant $J$, $\mu$-a.s. there are infinitely many infinite ``$+$''-clusters, infinitely many infinite ``$-$''-clusters and infinitely many infinite contours. Here $p_{c,2}^{w}$ is defined as in (\ref{wrc}).
\item If
\begin{eqnarray}
J\geq\frac{1}{2}\ln\left(\frac{1}{1-p_{u,2}^{w}}\right)\label{jj1}
\end{eqnarray}
Let $\sA_+$ be the event that there is a unique infinite ``$+$''-cluster, no infinite ``$-$''-clusters and no infinite contours; and let $\sA_-$ be the event that there is a unique infinite ``$-$''-cluster, no infinite ``$+$''-clusters and no infinite contours.
then
\begin{eqnarray}
\mu^+(\sA_+)&=&1.\label{+1}\\
\mu^-(\sA_-)&=&1. \label{-1}
\end{eqnarray}
\item If
\begin{eqnarray}
J>\frac{1}{2}\ln\left(\frac{1}{1-p_{u,2}^{f}}\right)\label{jj2}
\end{eqnarray}
then (\ref{+1}) and (\ref{-1}) hold, and moreover,
\begin{eqnarray}
\mu^{f}(\sA_+)=\mu^{f}(\sA_-)=\frac{1}{2}.\label{hmf}
\end{eqnarray}
\end{enumerate}
\end{theorem}
\section{XOR Ising model on transitive, triangular tilings of the hyperbolic plane}\label{xIs}
In this section, we state the main result concerning the percolation properties of the XOR Ising model on transitive, triangular tilings of the hyperbolic plane. These results, as given in \Cref{coii,xorc}, will be proved in \Cref{pxorc} as applications of \Cref{m23}.
Throughout this section, we let $\LL_1$ be the $[n,n,n]$ regular tiling of the hyperbolic plane, such that each face has degree $n\geq 7$, and each vertex has degree 3. Let $\LL_2$ be the planar dual graph of $\LL_1$. More precisely, $\LL_2$ is the vertex-transitive triangular tiling of the hyperbolic plane such that each vertex has degree $n\geq 7$. An \textbf{XOR Ising model} on $\LL_2$ is a probability measure on $\sigma_{XOR}\in \{\pm 1\}^{V(\LL_2)}$, such that
\begin{eqnarray*}
\sigma_{XOR}(v)=\sigma_1(v)\sigma_2(v),\qquad\forall v\in V(\LL_2),
\end{eqnarray*}
where $\sigma_1$, $\sigma_2$ are two i.i.d.\ Ising models with spins located on $V(\LL_2)$. A \textbf{contour configuration} of the XOR Ising configuration on $\LL_2$ is a subset of edges of $\LL_1$ in which each edge has a dual edge in $E(\LL_2)$ joining two vertices $u,v\in V(\LL_2)$ satisfying $\sigma_{XOR}(u)=-\sigma_{XOR}(v)$. A connected component in a contour configuration is called a \textbf{contour}. Obviously each vertex of $\LL_1$ has 0 or 2 incident present edges in a contour configuration of an XOR Ising configuration, since $\LL_1$ has vertex degree 3. Each contour of an XOR Ising configuration on $\LL_2$ is either a self-avoiding cycle or a doubly-infinite self-avoiding path.
We can similarly define an XOR Ising model with spins located on vertices of $\LL_1$, and its contours to be even-degree subgraphs of $\LL_2$.
\begin{theorem}\label{coii} Let $\sigma_1$, $\sigma_2$ be two i.i.d.\ Ising models with spins located on vertices of $\LL_2$, coupling constant $J\in [0,\infty)$ and free boundary conditions. Let $\mu_1^f$ (resp.\ $\mu_2^f$) be the distribution of $\sigma_1$ (resp.\ $\sigma_2$). Assume that one of the following cases occurs
\begin{enumerate}
\item If $\mu_1^f\times \mu_2^f$ is $\mathrm{Aut}(\LL_2)$-ergodic; or
\item $\liminf_{|i-j|\rightarrow\infty}\langle \sigma_{1,i}\sigma_{1,j}\rangle_{\mu_{1,f}}=0$, where $\sigma_{1,i}$ and $\sigma_{1,j}$ are two spins in the Ising model $\sigma_1$ with distance $|i-j|$; or
\item $J$ satisfies Condition (c) of \Cref{ipl} II.
\item $J$ sasifies Condition (d) of \Cref{ipl} II.
\end{enumerate}
then $\mu_1^f\times \mu_2^f$-a.s. there are infinitely many infinite ``$+$''-clusters and infinitely many infinite ``$-$''-clusters.
\end{theorem}
\begin{theorem}\label{xorc} Let $\sigma_1$, $\sigma_2$ be two i.i.d.\ Ising models with spins located on vertices of $\LL_1$, and coupling constant $K\geq 0$. For $i=1,2$, let $\mu_{i,+}$ (resp.\ $\mu_{i,-}$) be the distribution of $\sigma_i$ with ``$+$''-boundary conditions (resp.\ ``$-$''-boundary conditions). Let $J\geq 0$ be given by
\begin{eqnarray}
e^{-2J}=\frac{1-e^{-2K}}{1+e^{-2K}},\label{jkr}
\end{eqnarray}
and let $t$ be the number of infinite contours. Let $\mu_{++}$ (resp.\ $\mu_{--}$, $\mu_{+-}$) be the product measure of $\mu_{1,+}$ and $\mu_{2,+}$ (resp.\ $\mu_{1,-}$ and $\mu_{2,-}$, $\mu_{1,+}$ and $\mu_{2,-}$). Assume $J$ satisfies the assumption of \Cref{coii}, then we have
\begin{eqnarray*}
\mu_{++}(t\in\{0,\infty\})=\mu_{--}(t\in\{0,\infty\})=\mu_{+-}(t\in\{0,\infty\})=1.
\end{eqnarray*}
\end{theorem}
\section{Dimer model on the [4,6,12] lattice}\label{db}
In this section, we introduce the $[4,6,12]$ lattice, explain the relation between perfect matchings on the [4,6,12] lattice and constrained percolation configurations in the [3,4,6,4] lattice as discussed in \Cref{xorh}, and then state the main result proved in this paper concerning the dimer model on the [4,6,12] lattice.
A [4,6,12] lattice is a graph that can be embedded into the Euclidean plane $\RR^2$ such that each vertex is incident to 3 faces with degrees 4, 6, and 12, respectively. See \Cref{4612}.
\begin{figure}
\centering
\includegraphics[width=.5\textwidth]{hh4612}
\caption{[3,4,6,4] lattice (represented by dashed lines) and the [4,6,12] lattice $\mathbb{A}$ (represented by black lines).}
\label{4612}
\end{figure}
A \textbf{perfect matching}, or a \textbf{dimer configuration} on a [4,6,12] lattice is a subset of edges such that each vertex of the [4,6,12] lattice is incident to exactly one edge in the subset.
A \df{Type-I} edge of a [4,6,12] lattice is an edge of a square face. Any other edge of the [4,6,12] lattice is a \df{Type-II} edge. We say two Type-II edges $e_1,e_2$ are \df{adjacent} if there exists an Type-I edge $e_3$, such that both $e_1$ and $e_2$ share a vertex with $e_3$ in the [4,6,12] lattice. A subset of Type-II edges is \df{connected} if for any two edges $e$ and $f$ in the subset, there exist a sequence of Type-II edges $e_0 (=e),\ e_1,\ \ldots,\ e_n(=f)$ in the subset, such that $e_i$ and $e_{i-1}$ are adjacent, for $1\leq i\leq n$. A \df{Type-II cluster} is a maximal connected set of present Type-II edges in a perfect matching.
The restriction of any dimer configuration on the [4,6,12] lattice to Type-II edges naturally correspond to a constrained percolation configuration on the [3,4,6,4] lattice in $\Omega$. The [3,4,6,4] lattice is constructed as follows. A vertex of the [3,4,6,4] lattice is placed at the midpoint of each Type-II edge of the [4,6,12] lattice. Two vertices of the [3,4,6,4] lattices are joined by an edge if and only if they are midpoints of two adjacent Type-II edges. See \Cref{4612}. A Type-II edge is present in a dimer configuration if and only if its midpoint has state ``1'' in the corresponding constrained percolation configuration. It is straightforward to check that this way we obtain a bijection between restrictions to Type-II edges of dimer configurations on the [4,6,12] lattice and constrained percolation configurations on the [3,4,6,4] lattice in $\Omega$. Recall that constrained percolation configurations on the [3,4,6,4] lattice induces contour configurations on the hexagonal lattice $\HH$ and the triangular lattice $\TT$ by a 2-to-1 mapping $\phi:\Omega\rightarrow\Phi$. See \Cref{3464contour}.
From the connection of the [4,6,12] lattice and the [3,4,6,4] lattice, as well as the connection of the [3,4,6,4] lattice with the hexagonal lattice $\HH$ and the dual triangular lattice $\TT$ as described in \Cref{xorh}, we can see that each square face of the $[4,6,12]$ lattice is crossed by a unique edge of $\HH$ and a unique edge of $\TT$. Each vertex of $\HH$ is located at the center of a hexagon face of the [4,6,12] lattice, and each vertex of $\TT$ is located in the center of a degree-12 face of the [4,6,12] lattice.
Note that the $\HH$ is a bipartite graph; i.e. all the vertices can be colored black and white such that the vertices of the same color are not adjacent. Let $\Gamma$ be the translation group of the hexagonal lattice generated by translations along two different directions, such that the set of black vertices and the set of white vertices form two distinct orbits under the action of $\Gamma$. Note that $\Gamma$ acts transitively on the dual triangular lattice $\TT$.
In order to define a probability measure for perfect matchings on the [4,6,12] lattice, we introduce edge weights. We assign weight 1 to each Type-II edge, and weight $w_e$ to the Type-I edge $e$. Assume that the edge weights of the [4,6,12] lattice satisfy the following conditions.
\begin{enumerate}[label=({B}{\arabic*})]
\item The edge weights are $\Gamma$-invariant.
\item If $e_1$, $e_2$ are two parallel Type-I edges around the same square face, then $w_{e_1}=w_{e_2}$.
\item If $e_1$, $e_2$ are two perpendicular Type-I edges around the same square face, then $w_{e_1}^2+w_{e_2}^2=1$.
\end{enumerate}
The reason we assume (B1) is to define a $\Gamma$-translation-invariant probability measure. The reason we assume (B2) and (B3) is to define a probability measure for dimer configurations of the [4,6,12] lattice, which, under the connection described above to constrained percolation configurations in $\Omega$, will induce a probability measure on $\Omega$ satisfying the symmetry assumption (A3).
Under (B1)--(B3), the edge weights are described by three independent parameters. We may sometimes assume that the parameters satisfy the identity below, which reduces the number of independent parameters to two.
\begin{enumerate}[label=({B}{\arabic*})]
\setcounter{enumi}{3}
\item Let
\begin{eqnarray}
h(x,y,z)=x+y+z+xy+xz+yz-xyz-1.\label{hxyz}
\end{eqnarray}
For each edge $e$ of the hexagonal lattice $\HH$, let $e_1$ (resp.\ $e_2$) be a Type-I edge of the [4,6,12] lattice parallel (resp.\ perpendicular) to $e$. Let
\begin{eqnarray*}
t_e=\frac{1-w_{e_1}}{w_{e_2}},
\end{eqnarray*}
where $w_{e_1}$ (resp.\ $w_{e_2}$) is the edge weight of $e_1$ (resp.\ $e_2$) for dimer configurations on the [4,6,12] lattice. Under the assumption (B1), $t_e$ is uniquely defined independent of the $e_1$, $e_2$ chosen - as long as $e_1$ is parallel to $e$ and $e_2$ is perpendicular to $e$. Let $e_a, e_b, e_c$ be three edges of $\HH$ with distinct orientations in the embedding of $\HH$ into $\RR^2$. Then $h(t_{e_a},t_{e_b},t_{e_c})=0$.
\end{enumerate}
In \cite{KOS06}, the authors define a probability measure for any bi-periodic, bipartite, 2-dimensional lattice. Specializing to our case, let $\mu_{n,D}$ be the probability measure of dimer configurations on a toroidal $n\times n$ [4,6,12] lattice $\mathbb{A}_n$ (see \cite{KOS06} for details). Let $\mathcal{M}_n$ be the set of all perfect matchings on $\mathbb{A}_n$, and let $M\in \mathcal{M}_n$ be dimer configuration, then
\begin{eqnarray}
\mu_{n,D}(M)=\frac{\prod_{e\in M}w_{e}}{\sum_{M\in \mathcal{M}_n}\prod_{e\in M}w_e},\label{mnd}
\end{eqnarray}
where $w_e$ is the weight of the edge $e$. As $n\rightarrow\infty$, $\mu_{n,D}$ converges weakly to a translation-invariant measure $\mu_{D}$ (see \cite{KOS06}).
\begin{theorem}\label{m31}For the dimer model on the [4,6,12] lattice.
\begin{enumerate}
\item If the edge weights satisfy (B1)-(B4), $\mu_D$ almost surely there are neither infinite Type-II clusters nor infinite contours.
\item If the edge weights satisfy (B1)-(B3), $\mu_D$ almost surely there exists at most one infinite contour.
\end{enumerate}
\end{theorem}
\Cref{m31} is proved in \Cref{pm31}.
\section{XOR Ising models on the hexagonal and triangular lattices}\label{xori}
In this section, we define the XOR Ising models on the hexagonal and triangular lattices, and state the main results proved in this paper concerning the percolation properties of these models.
Let $\sigma_1$, $\sigma_2$ be two i.i.d.\ ferromagnetic Ising models with spins located on vertices of the hexagonal lattice $\HH=(V_{\HH}, E_{\HH})$. The hexagonal lattice has edges in three different directions. Assume that both $\sigma_1$ and $\sigma_2$ have nonnegative coupling constants $J_a$, $J_b$, $J_c$ on edges of $\HH$ with the three different directions, respectively. Assume also that the distributions of both $\sigma_1$ and $\sigma_2$ are weak limits of Gibbs measures under periodic boundary conditions. Define the XOR Ising model $\sigma_{XOR}(v)=\sigma_1(v)\sigma_2(v)$, for $v\in V_{\HH}$.
A \df{contour configuration} for an XOR Ising configuration, $\sigma_{XOR}$, defined on $\HH$ (resp.\ $\TT$), is a subset of $\{0,1\}^{E(\TT)}$ (resp.\ $\{0,1\}^{E(\HH)}$), whose state-1-edges (present edges) are edges of $\TT$ (resp.\ $\HH$) separating neighboring vertices of $\HH$ (resp.\ $\TT$) with different states in $\sigma_{XOR}$. (Note that $\HH$ and $\TT$ are planar duals of each other.) Contour configurations of the XOR Ising model were first studied in \cite{DBW11}, in which the scaling limits of contours of the critical XOR Ising model are conjectured to be level lines of Gaussian free field. It is proved in \cite{bd14} that the contours of the XOR Ising model on a plane graph correspond to level lines of height functions of the dimer model on a decorated graph, inspired by the correspondence between Ising model and bipartite dimer model in \cite{Dub}. We will study the percolation properties of the XOR Ising model on $\HH$ and $\TT$, as an application of the main theorems proved in this paper for the general constrained percolation process.
Let
\begin{eqnarray}
&&f(x,y,z)=e^{-2x}+e^{-2y}+e^{-2z}+e^{-2(x+y)}+e^{-2(x+z)}+e^{-2(y+z)}-e^{-2(x+y+z)}-1.\label{fxyz}\\
&&g(x,y,z)=e^{2x}+e^{2y}+e^{2z}-e^{2(x+y+z)}.\label{gxyz}
\end{eqnarray}
We say the XOR Ising model on $\HH$ with coupling constants $(J_a,J_b,J_c)$ is in the \df{high-temperature state} (resp.\ \df{low-temperature state}, \df{critical state}) if $f(J_a,J_b,J_c)>0$ (resp.\ $f(J_a,J_b,J_c)<0$, $f(J_a,J_b,J_c)=0$).
Let $\TT=(V_{\TT},E_{\TT})$ be the dual triangular lattice of $\HH$. We also consider the XOR Ising model with spins located on $V_{\TT}$. Assume that the coupling constants on edges with 3 different directions are $K_a$, $K_b$ and $K_c$, respectively, such that $K_a, K_b, K_c\geq 0$. Also for $i\in\{a,b,c\}$, assume that $K_i$ is the coupling constant on an edge of $\TT$ dual to an edge of $\HH$ with coupling constant $J_i$. We say the XOR Ising model on the triangular lattice is in the \df{low-temperature state} (resp.\ \df{high-temperature state}, \df{critical state}) if $g(K_a,K_b,K_c)<0$ (resp. $g(K_a,K_b,K_c)>0$, $g(K_a,K_b,K_c)=0$).
Similar to the square grid case, in the high temperature state, the Ising model on the hexagonal lattice or the triangular lattice has a unique Gibbs measure, and the spontaneous magnetization vanishes; while in the low temperature state, the Gibbs measures are not unique and the spontaneous magnetization is strictly positive under the ``$+$''-boundary condition. See \cite{JL72, Ai80,ZL12,DC13}.
If
\begin{eqnarray}
e^{-2K_{\tau}}=\frac{1-e^{-2J_{\tau}}}{1+e^{-2J_{\tau}}},\qquad\mathrm{for}\ \tau=a,b,c,\label{dua}
\end{eqnarray}
then the XOR Ising model on $\HH$ with coupling constants $(J_a,J_b,J_c)$ is in the low-temperature state (resp.\ high-temperature state, critical state) if and only if the XOR Ising model on the triangular lattice with coupling constants $(K_a,K_b,K_c)$ is in the high-temperature state (resp.\ low-temperature state, critical state).
A \df{cluster} of an XOR Ising configuration on $\HH$ (resp.\ $\TT$) is a maximal connected set of vertices on $\HH$ (resp.\ $\TT$), such that every vertex has the same state. A \df{contour} of an XOR Ising configuration on $\HH$ (resp.\ $\TT$) is a maximal connected set in the contour configuration on $\TT$ (resp.\ $\HH$) associated to the XOR Ising configuration on $\HH$ (resp.\ $\TT$).
We have the following theorems.
\begin{theorem}\label{chi}Consider the critical XOR Ising model on $\HH$ or $\TT$. Then
\begin{enumerate}
\item almost surely there are no infinite clusters;
\item almost surely there are no infinite contours.
\end{enumerate}
\end{theorem}
\begin{theorem}\label{lth}In the low-temperature XOR Ising model on $\HH$ or on $\TT$, almost surely there are no infinite contours.
\end{theorem}
\Cref{chi,lth} are proved in \Cref{p412}.
\section{Percolation on non-amenable graphs}\label{pnc}
In this section, we discuss known results about percolation on non-amenable graphs that will be used to prove \Cref{m23,m21,m22}.
The following lemma is proved in \cite{BS20}.
\begin{lemma}\label{lbs}Let G be a quasi-transitive, non-amenable, planar graph with one end,
and let $\omega$ be an invariant percolation on G. Then a.s. the number of infinite 1-clusters
of $\omega$ is 0, 1, or $\infty$.
\end{lemma}
\begin{lemma}(Threshold for bond percolation on non-amenable graphs)\label{tbng} Let $G=(V,E)$ be a non-amenable graph. Let $\Gamma\subseteq \mathrm{Aut}(G)$ be a closed unimodular quasi-transitive subgroup, and let $o_1,\ldots,o_L$ be a complete set of representatives in $V$ of the orbits of $G$. For $1\leq i\leq L$, let
\begin{eqnarray*}
\mathrm{Stab}(o_i):&=&\{\gamma\in \Gamma: \gamma o_i=o_i\};\\
\eta_i:&=&|\mathrm{Stab}(o_i)|.
\end{eqnarray*}
Let $\mathbb{P}$ be a bond percolation on $G$ whose distribution is $\Gamma$-invariant. Let $D_i$ be the random degree of $o_i$ in the percolation subgraph, and let $d_i$ be the degree of $o_i$ in $G$. Write $p_{\infty,v}$ for the probability that $v\in V$ is in an infinite component. Let $p_{\infty,i}$ be the probability that $o_i$ is in an infinite cluster. Then
\begin{eqnarray}
\sum_{i=1}^{L}\frac{(d_i-\alpha(G))p_{\infty,i}}{\eta_i}\geq \sum_{j=1}^{L}\frac{\mathbb{E}D_j-\alpha(G)}{\eta_j}\label{cic}
\end{eqnarray}
where $i_E(G)$ is a constant depending on the structure of the graph $G$ defined by
\begin{eqnarray*}
\alpha_K:&=&\frac{1}{|K|}\sum_{x\in K}\deg_K(x)\\
\alpha(G):&=&\sup\{\alpha_K: K\subset G\ \mathrm{is\ finite}\}
\end{eqnarray*}
In particular, if the right-hand side of (\ref{cic}) is positive, then there is an infinite component in the percolation subgraph with positive probability.
\end{lemma}
\begin{proof}See Theorem 4.1 of \cite{blps}.
\end{proof}
\section{Square tilings of the hyperbolic plane}\label{sthp}
In this section, we introduce the square tilings of the hyperbolic plane, and then state and prove properties of the constrained percolation models on such graphs.
Let $G=(V,E)$ be a graph corresponding to a square tiling of the hyperbolic plane. Assume that
\begin{enumerate}
\item each face of $G$ has 4 edges; and
\item each vertex of $G$ is incident to $2n$ faces, where $n\geq 3$.
\end{enumerate}
See Figure \ref{fig:46} for an example of such a graph when $n=3$.
\begin{figure}
\includegraphics[width=.6\textwidth]{46}
\caption{The [4,4,4,4,4,4] lattice: each face has degree 4, and each vertex have degree 6. Figure from the Wikipedia.}
\label{fig:46}
\end{figure}
We can color all the faces of $G$ by black and white such that black faces can share edges only with white faces and vice versa. Let $G=(V,E)$ denote the graph embedded into the hyperbolic plane as described above.
We consider the site configurations in $\{0,1\}^V$. We impose the following constraint on site configurations
\begin{itemize}
\item Around each black face, there are six allowed configurations $(0000)$, $(1111)$, $(0011)$, $(1100)$, $(0110)$, $(1001)$, where the digits from the left to the right correspond to vertices in clockwise order around the black face, starting from the lower left corner. See Figure \ref{lcc}.
\end{itemize}
Let $\Omega\subset\{0,1\}^V$ be the set of all configurations satisfying the constraint above.
Note that $G$ is a vertex-transitive graph, in the sense that the automorphism group of $G$, $\mathrm{Aut}(G)$, acts transitively on $G$. Since each face of $G$ has an even number of edges, $G$ itself is a bipartite graph - we can color the vertices of $G$ by red and green such that red vertices are adjacent only to green vertices and vice versa. We assign an integer in $1,2,\ldots, n$ to each white face of $G$ according to the following rules
\begin{enumerate}
\item around each red vertex of $G$, white faces are assigned integers $1,2,\ldots,n$ clockwise; and
\item around each green vertex of $G$, white faces are assigned integers $1,2,\ldots,n$ counterclockwise; and
\item any two white faces adjacent to the same black face along two opposite edges have the same assigned integer.
\end{enumerate}
See Figure \ref{fig:label} for an example of assignments of integers $1,2,3$ to the white faces of the $[4,4,4,4,4,4]$ lattice.
\begin{figure}
\includegraphics[width=.4\textwidth]{label}
\caption{Labels of white faces of the [4,4,4,4,4,4] lattice}
\label{fig:label}
\end{figure}
For $1\leq i\leq n$, we construct a graph $\LL_i$ as follows. The vertex set of $\LL_i$ consists of all the white faces of $G$ whose assigned integers are $i$. Two vertices of $\LL_i$ are joined by an edge of $\LL_i$ if and only if they correspond to two white faces of $G$ adjacent to the same black face along two opposite edges. Note that when $n\geq 3$, each component of $\LL_i$ is a regular tree of degree 4. For any integer $i$ satisfying $1\leq i\leq n$, the edges of $\LL_{i-1}$ (if $i=1$, $\LL_{i-1}:=\LL_n$) and $\LL_i$ cross; the edges of $\LL_{i+1}$ (if $i=n$, $\LL_{i+1}:=\LL_1$) and $\LL_i$ cross.
Any constrained percolation configuration in $\Omega$ gives rise to a contour configuration on $\cup_{i=1}^n\LL_{i}$. An edge $e$ in $\cup_{i=1}^n\LL_i$ is present in the contour configuration if and only if it crosses a black face $b$ in $G$, such that the states of the vertices of $b$ on the two sides separated by $e$ in the configuration are different, and any two vertices of $b$ on the same side of $e$ have the same state. This is a contour configuration satisfying the condition that each vertex in $\cup_{i=1}^n\LL_i$ has an even number of incident present edges. For any $1\leq i<j\leq n$, present edges in $\LL_i$ and $\LL_j$ can never cross.
A \df{cluster} is a maximal connected set of vertices in $G$ in which every vertex has the same state in a constrained percolation configuration. A \df{contour} is a maximal connected set of edges in $\cup_{i=1}^n\LL_i$ in which every edge is present in the contour configuration. Note that each contour must be a connected subgraph of $\LL_i$, for some $1\leq i\leq n$. Hence each contour must be a tree. Since each vertex in a contour has an even number of incident present edges in the contour, each contour must be an infinite tree.
Let $\mu$ be a probability measure on $\Omega$. We may assume that $\mu$ satisfies the following conditions
\begin{enumerate}[label=(D\arabic*)]
\item $\mu$ is $\mathrm{Aut}(G)$-invariant;
\item $\mu$ is $\mathrm{Aut}(\LL_i)$-ergodic, for $1\leq i\leq n$;
\item $\mu$ is symmetric, i.e.\ let $\theta:\Omega\rightarrow\Omega$ be the map defined by $\theta(\omega)(v)=1-\omega(v)$, for each $v\in V$, then $\mu$ is invariant under $\theta$, that is, for any event $A$, $\mu(A)=\mu(\theta(A))$.
\end{enumerate}
Note that when $n\geq 3$, the graph $G$ is a non-amenable group. Recall that the number of \df{ends} of a connected graph is the supremum over its finite subgraphs of the number of infinite components that remain after removing the subgraph.
Here is the main theorem concerning the properties of constrained percolations on the square tilings of the hyperbolic plane.
\begin{theorem}\label{m51}
\begin{enumerate}[label=(\alph*)]
\item Let $\mu$ be a probability measure on $\Omega$ satisfying (D1). Let $n_0$ (resp.\ $n_1$) be the number of infinite 0-clusters (resp. 1-clusters). Then $\mu$-a.s. $(n_0,n_1)\in\{(0,1),(1,0),(1,\infty),(\infty,1),(\infty,\infty)\}$.
\item Let $\nu$ be a probability measure on $\Omega$ satisfying (D1) - (D3). The $\nu$-a.s. there are infinitely many infinite 0-clusters and infinitely many infinite 1-clusters.
\end{enumerate}
\end{theorem}
In order to prove \ref{m51}, we first prove a few lemmas.
\begin{lemma}\label{cit}In a contour configuration in $\cup_{i=1}^n\LL_i$ as described above, any contour must be an infinite tree (a tree consisting of infinite many edges of $\cup_{i=1}^n \LL_i$) in which each vertex has degree 2 or 4.
\end{lemma}
\begin{proof}This lemma is straightforward from the facts that each contour is a connected subgraph of $\LL_i$ for some $i\in\{1,\ldots,n\}$; each component of $\LL_i$ $1\leq i\leq n$ is a regular tree of degree 4, and each vertex in a contour has an even number of incident present edges.
\end{proof}
\Cref{l2} below is proved in \cite{blps} and \cite{blps99} using the mass transport principle.
\begin{lemma}\label{l2}Let G be a nonamenable graph whose automorphism group has a closed subgroup acting transitively and unimodularly on $G$, and
let $\omega$ be an invariant percolation on G which has a single component a.s. Then
$p_c(\omega) < 1$ a.s., where $p_c(\cdot)$ is the critical i.i.d.\ Bernoulli percolation probability on a graph.
\end{lemma}
\begin{proof}See Theorem 1.5 of \cite{blps}.
\end{proof}
\begin{lemma}\label{01c}Let $G=(V,E)$ be a square tiling of the hyperbolic plane satisfying I and II. Let $\omega\in \Omega$ be a constrained percolation configuration on $G$. If there exists a contour in the corresponding contour configuration of $\omega$, then there exist at least one infinite 1-cluster and at least on infinite 0-cluster in $\omega$.
\end{lemma}
\begin{proof}Let $C$ be a contour in the corresponding contour configurations. By \Cref{cit}, $C$ is an infinite tree in which each vertex is incident to 2 or 4 edges. Since $C$ has no cycles, the complement $\HH^2\setminus C$ of $C$ in the hyperbolic plane $\HH^2$ has no bounded components.
We claim that each unbounded component of $\HH^2\setminus C$ contains at least one infinite cluster. Let $\Lambda$ be an unbounded component of $\HH^2\setminus C$. Let $V_{\Lambda,C}\subset V$ consist of all the vertices in $\Lambda$ that are also in a face of $G$ intersecting $C$. Then all the vertices in $V_{\Lambda,C}$ are in the same cluster of $\omega$ and $|V_{\Lambda,C}|=\infty$. Hence there exists an infinite cluster in $\omega$ containing every vertex in $V_{\Lambda,C}$.
Let $e\in C$ be an edge crossing a pair of opposite edges of a black face $b$ of $G$. Let $v_1,v_2,v_3,v_4$ be the 4 vertices of $b$. Assume that $v_1$ and $v_2$ are on one side of $e$ while $v_3$ and $v_4$ are on the other side of $e$. By the arguments above, $v_1$ and $v_2$ are in an infinite cluster $\xi_1$ of $\omega$; similarly, $v_3$ and $v_4$ are in an infinite cluster $\xi_2$ of $\omega$. Moreover, $e\in C$ implies that $v_1$ and $v_3$ have different state; and therefore exactly one of $\xi_1$ and $\xi_2$ is an infinite 0-cluster, and the other is an infinite 1-cluster.
\end{proof}
\begin{lemma}\label{ifc}Let $G=(V,E)$ be a square tiling of the hyperbolic plane satisfying I and II. Let $\omega\in \Omega$ be a constrained percolation configuration on $G$, and let $\phi$ be the corresponding contour configuration of $\omega$. Then each component of $\HH^2\setminus \phi$ contains an infinite cluster in $\omega$.
\end{lemma}
\begin{proof}By \Cref{cit}, $\phi$ is the disjoint union of infinite trees, in which each vertex has degree 2 or 4. Since $\phi$ contains no cycles, each component of $\HH^2\setminus \phi$ is unbounded.
Let $\Lambda$ be an unbounded component of $\HH^2\setminus \phi$. Let $V_{\Lambda,\phi}\subset V$ consist of all the vertices in $\Lambda$ that are also in a face of $G$ intersecting $\phi$. Then all the vertices in $V_{\Lambda,\phi}$ are in the same cluster of $\omega$ and $|V_{\Lambda,\phi}|=\infty$. Hence there exists an infinite cluster in $\omega$ containing every vertex in $V_{\Lambda,\phi}$. Since every vertex in $V_{\Lambda,\phi}$ is in $\Lambda$, and any cluster intersecting $\Lambda$ is completely in $\Lambda$, we conclude that $\Lambda$ contains an infinite cluster of $\omega$.
\end{proof}
\bigskip
\noindent\textbf{Proof of Theorem \ref{m51}} First we show that Part (a) of the theorem together with Assumptions (D2), (D3) implies Part (b). Let $\nu$ be a probability measure on $\Omega$ satisfying (D1) - (D3). By Assumption (D2) and (D3), there exists a positive integer $k$ (possibly infinite), such that $\nu(n_0=n_1=k)=1$. Then Part (b) follows from Part (a).
Now we prove Part (a). Obviously $(n_0,n_1)\in\{(0,1),(1,0)\}$ if there are no contours. Now assume that contours do exist. By \Cref{lbs}, $n_0,n_1\in\{0,1,\infty\}$. By \Cref{01c}, $n_0,n_1\in\{1,\infty\}$. Let $\phi$ be the contour configuration. If there are infinitely many contours in $\phi$, or there exists a contour of $\phi$ in which infinitely many vertices have degree 4, then $\HH^2\setminus \phi$ has infinitely many unbounded components. By \Cref{ifc}, $n_0+n_1=\infty$. Therefore $\{n_0,n_1\}\in\{(1,\infty),(\infty,1),(\infty,\infty)\}$ in this case.
Now consider the case that the number of contours is finite and nonzero, and on each contour only finitely many vertices have degree 4. Fix an $i$ satisfying $1\leq i\leq n$, and conditional on the event that the number of contours on $\LL_i$ is finite and nonzero. Choose a contour $\tau$ on $\LL_i$ uniformly at random; then $\tau$ forms an invariant bond percolation on $\LL_i$ which has a single component. By \Cref{l2}, almost surely $\tau$ has infinitely many vertices with degree 4 - since otherwise $p_c(\tau)=1$. Therefore this case does not occur a.s.
$\hfill\Box$
\section{Contours and Clusters}\label{ctcl}
In this section, we prove results concerning contours and clusters for the constrained percolation model on the $[m,4,n,4]$ lattice in preparation to prove \Cref{m21}.
Let $P$ be the underlying plane into which the $[m,4,n,4]$ lattice is embedded. When $\frac{1}{m}+\frac{1}{n}=\frac{1}{2}$, $P$ is the Euclidean plane $\RR^2$ and the graph $G$ is amenable. When $\frac{1}{m}+\frac{1}{n}<\frac{1}{2}$, $P$ is the hyperbolic plane $\HH^2$ and the graph $G$ is non-amenable.
We consider an embedding of the $[m,4,n,4]$ lattice into $P$ in such a way that each edge has length 1.
Let $\phi\in\Phi$ be a contour configuration, and let $C$ be a contour in $\phi$. To each component of $P\setminus \phi$, we associate an \df{interface}, which is a closed set consisting of all the points in the component whose distance to $C$ is $\frac{1}{4}$. Here by distance, we mean either Euclidean distance or hyperbolic distance depending on whether $P$ is $\RR^2$ or $\HH^2$. Obviously each interface is either a self-avoiding cycle or a doubly infinite self-avoiding walk. See \Cref{3464interface} for an example of interfaces on the $[3,4,6,4]$ lattice.
\begin{figure}
\includegraphics[width=.5\textwidth]{3464interface}
\caption{Contour configuration and interfaces. Blue lines represent contours of $\HH$. Red lines represent contours of $\TT$. Green lines represent interfaces.}
\label{3464interface}
\end{figure}
Note that when $\frac{1}{m}+\frac{1}{n}<\frac{1}{2}$ and $\min\{m,n\}\geq 3$, both $\LL_1$ and $\LL_2$ are vertex-transitive, non-amenable, planar graphs with one end.
\begin{lemma}\label{rl}Whenever we have an interface $I$, let $F_{I}$ be the set consisting of all the vertices of $G$ whose (Euclidean or hyperbolic) distance to the interface is $\frac{1}{4}$. Then all the vertices of $F_{I}$ are in the same cluster. If $I$ is a doubly-infinite self-avoiding path, then $F_{I}$ is part of an infinite cluster.
\end{lemma}
\begin{proof}Recall that the interface $I$ is either a self-avoiding cycle or a doubly-infinite self-avoiding walk. Give $I$ a fixed direction. Moving along $I$ following the fixed direction, let $\{S_j\}_{j\in J}$ be the set of faces crossed by $I$ in order, where $J\subseteq \ZZ$ is a set of integers, such that for any $j_1<j_2$, $j_1,j_2\in J$, $I$ crosses $S_{j_1}$ first, and then crosses $S_{j_2}$. Note that it is possible to have $S_{j_1}=S_{j_2}$ for $j_1\neq j_2$.
For any two vertices $u,v\in F_{I}$, we can find a sequence of indices $j_1<j_2<\ldots<j_k\in J$, and a sequence of vertices of $G$,
\begin{eqnarray}
u=v_{j_1,1},v_{j_1,2}(=v_{j_2,1}),v_{j_2,2}(=v_{j_3,1})=\ldots=v_{j_k,2}=v,\label{vss}
\end{eqnarray}
such that for any $1\leq i\leq k$, $v_{j_i,1}$ and $v_{j_i,2}$ are two vertices (which may not be distinct) in $F_I\cap \partial S_{j_i}$, and there exists a path $\ell_{v_{j_i,1}v_{j_i,2}}$ connecting $v_{j_i,1}$ and $v_{j_i,2}\subset \partial S_{j_i}$, and $\ell_{v_{j_i,1}v_{j_i,2}}\cap I=\emptyset$. Note that any two consecutive vertices in (\ref{vss}) are in the same cluster, therefore $u$ and $v$ are in the same cluster.
If $I$ is a doubly-infinite self-avoiding path, then $I$ crosses infinitely many faces. Each face crossed by $I$ has at least one boundary vertex in $F_I$. Each vertex of $F_I$ is a boundary vertex of at most 4 faces. Therefore $|F_I|=\infty$. Since all the vertices in $F_I$ are in the same cluster, $F_I$ is part of an infinite cluster.
\end{proof}
In the following lemma, contours may be primal or dual as usual.
\begin{lemma}\label{cc}If there exist at least two infinite contours, then there exists an infinite 0-cluster or an infinite 1-cluster. Moreover, if $C_1$ and $C_2$ are two infinite contours, then there exists an infinite cluster incident to $C_1$.
\end{lemma}
\begin{proof}If there exist at least two infinite contours, then we can find two distinct infinite contours $C_1$ and $C_2$, two points $x\in C_1$ and $y\in C_2$ and a self-avoiding path $p_{xy}$, consisting of edges of $G$ and two half-edges, one starting at $x$ and the other ending at $y$, and connecting $x$ and $y$, such that $p_{xy}$ does not intersect any infinite contours except at $x$ and at $y$. Indeed, we may take any path intersecting two distinct contours, and then take a minimal subpath with this property.
Let $v\in V$ be the first vertex of $G$ along $p_{xy}$ starting from $x$. Let $u$ be the point along the line segment $[v,x]$ lying on an interface of $C_1$. Let $\ell_u$ be the interface of $C_1$ containing $u$. Then $\ell_u$ is either a doubly-infinite self-avoiding path or a self-avoiding cycle.
We consider these two cases separately. Firstly, if $\ell_u$ is a doubly-infinite self-avoiding path, then we claim that $v$ is in an infinite (0 or 1-)cluster of the constrained site configuration on $G$. Indeed, this follows from \Cref{rl}.
Secondly, if $\ell_u$ is a self-avoiding cycle, then $P\setminus \ell_u$ has two components, $Q_v$ and $Q_v'$, where $Q_v$ is the component including $v$. Since $\ell_u$ is a cycle, exactly one of $Q_v$ and $Q_v'$ is bounded, the other is unbounded. Since $C_1\subseteq Q_v'$, and $C_1$ is an infinite contour, we deduce that $Q_v'$ is unbounded, and $Q_v$ is bounded. Since $y\notin \ell_u$, either $y\in Q_v$, or $y\in Q_v'$. If $y\in Q_v'$, then any path, consisting of edges of $G$ and one half-edge incident to $y$, connecting $v$ and $y$ must cross $C_1$. In particular, $p_{xy}$ crosses $C_1$ not only at $x$, but also at some point other than $x$. This contradicts the definition of $p_{xy}$. Hence $y\in Q_v$. Since $C_1\cap C_2=\emptyset$, this implies $C_2\subseteq Q_v$; because if $C_2\cap Q_v'\neq \emptyset$, then $C_2\cap C_1\neq \emptyset$. But $C_2\subseteq Q_v$ is impossible since $C_2$ is infinite and $Q_v$ is bounded. Hence this second case is impossible.
Therefore we conclude that if there exist at least two infinite contours, then there exists an infinite (0 or 1)-cluster.
\end{proof}
\begin{lemma}\label{io} Let $x\in V$ be in the infinite 0-cluster, let $y\in V$ be in the infinite 1-cluster, and let $\ell_{xy}$ be a path, consisting of edges of $G$ and connecting $x$ and $y$. Then $\ell_{xy}$ has an odd number of crossings with infinite contours in total.
In particular, if there exist both an infinite 0-cluster and an infinite 1-cluster in a constrained percolation configuration $\omega\in\Omega$, then there exists an infinite contour in $\phi(\omega)\in\Phi$.
\end{lemma}
\begin{proof}Same as Lemma 2.8 of \cite{HL16}.
\end{proof}
\begin{lemma}\label{icic}Let $C_{\infty}$ be an infinite contour. Then each infinite component of $G\setminus C_{\infty}$ contains an infinite cluster that is incident to $C_{\infty}$.
\end{lemma}
\begin{proof} The lemma can be proved using similar technique as in Lemma 2.7 of \cite{HL16}.
\end{proof}
\begin{lemma}\label{icazo}Let $\omega\in \Omega$. Assume that there is exactly one infinite 0-cluster and exactly one infinite 1-cluster in $\omega$. Assume that there exist a vertex $x$ in the infinite 0-cluster, a vertex $y$ in the infinite 1-cluster, and a path $\ell_{xy}$, consisting of edges of $G$ and joining $x$ and $y$, such that $\ell_{xy}$ crosses exactly one infinite contour, $C_{\infty}$. Then $C_{\infty}$ is incident to both the infinite 0-cluster and the infinite 1-cluster.
\end{lemma}
\begin{proof}By \Cref{icic}, there is an infinite cluster in each infinite component of $G\setminus C_{\infty}$. Since there are exactly two infinite clusters, $G\setminus C_{\infty}$ has at most 2 infinite components. Since each infinite cluster lies in some infinite component of $G\setminus C_{\infty}$, the number of infinite components of $G\setminus C_{\infty}$ is at least one.
If $G\setminus C_{\infty}$ has exactly two infinite components, then we can construct two infinite connected set of vertices in the two infinite components of $G\setminus C_{\infty}$, as a consequence of \Cref{icic}, denoted by $W_1$ and $W_2$, such that $C_{\infty}$ is incident to both $W_1$ and $W_2$. Moreover, $W_1$ and $W_2$ are exactly part of the infinite 0-cluster and part of the infinite 1-cluster. Therefore $C_{\infty}$ is incident to both the infinite 0-cluster and the infinite 1-cluster.
If $G\setminus C_{\infty}$ has exactly one infinite component, denoted by $R$, then both the infinite 0-cluster and the infinite 1-cluster lie in $R$, and in particular $x,y\in R$. We can find a path $\ell'_{xy}$, connecting $x$ and $y$, using edges of $G$, such that $\ell'_{xy}$ does not cross $C_{\infty}$ at all. We can change path from $\ell'_{xy}$ to $\ell_{xy}$ by choosing finitely many faces $S_1$,$S_2$,\ldots, $S_k$ of $G$; along the boundary of each face, make every present edge in the path absent and every absent edge in the path present; and we perform this procedure for $S_1,\ldots,S_k$ one by one. Such a the path modification procedure does not change the parity of the number of crossings of the path with $C_{\infty}$. Hence we infer that $\ell_{xy}$ intersects $C_{\infty}$ an even number of times. But this is a contradiction to Lemma \ref{io} which says $\ell_{xy}$ crosses $C_{\infty}$ an odd number of times, since we assume that $\ell_{xy}$ crosses exactly one infinite contour $C_{\infty}$.
\end{proof}
\begin{lemma}\label{isc}Assume that $\xi$ is an infinite cluster, and $C$ is an infinite contour. Assume that $x$ is a vertex of $G$ in $\xi$, and let $y\in C$ be the midpoint of an edge of $G$. Assume that there exists a path $p_{xy}$ connecting $x$ and $y$, consisting of edges of $G$ and a half-edge incident to $y$, such that $p_{xy}$ crosses no infinite contours except at $y$. Let $z$ be the first vertex of $G$ along $p_{xy}$ starting from $y$. Then $z\in\xi$.
\end{lemma}
\begin{proof}Since $p_{xy}$ crosses no infinite contours except at $y$, let $C_{1},\ldots,C_{m}$ be all the finite contours crossing $p_{xy}$. We claim that $P\setminus \cup_{i=1}^{m}C_{i}$ has a unique unbounded component, which contains both $x$ and $y$. Indeed, since $x\in \xi$ and $y\in C$; neither the infinite cluster $\xi$ nor the infinite contour $C$ can lie in a bounded component of $P\setminus \cup_{i=1}^{m}C_{i}$.
Let $I$ be the intersection of the union of the interfaces of $C_1,\ldots, C_m$ with the unique unbounded component of $P\setminus \cup_{i=1}^{m}C_{i}$. Since each $C_{i}$, $1\leq i\leq m$, is a finite contour, each interface of $C_{i}$ is finite. In particular, $I$ consists of finitely many disjoint self-avoiding cycles, denoted by $D_1,\ldots,D_t$. For $1\leq i\leq t$, $P\setminus D_i$ has exactly one unbounded component, and one bounded component. Moreover, for $i\neq j$, $D_i$ and $D_j$ come from interfaces of distinct contours.
Let $B_i$ be the bounded component of $P\setminus D_i$. We claim that each $B_i$ is simply-connected, and $B_i\cap B_j=\emptyset$, for $i\neq j$. Indeed, $B_i$ is simply connected, since the boundary of $B_i$, $D_i$ is a self-avoiding cycle, whose embedding on the plane is a simple closed curve, for $1\leq i\leq t$. Let $1\leq i<j\leq t$. Since $D_i$ and $D_j$ are disjoint, either $B_i\cap B_j=\emptyset$, or one of $B_i$ and $B_j$ is a proper subset of the other. Without loss of generality, assume $B_i$ is a proper subset of $B_j$. Then $D_i$ is a proper subset of $B_j$. Hence $D_i$ is in a bounded component of $P\setminus\cup_{k=1}^{m}C_k$, which contradicts the definition of $D_i$.
Let $R_i$ be the set of faces $F$ of $G$, for which $B_i\cap F\neq \emptyset$. Let $\widetilde{B}_i=\cup_{F\in R_i}F$. Note that for $1\leq i\leq t$, each $\widetilde{B}_i$ is a simply-connected, closed set. Let $B_i'$ be the interior of $\widetilde{B}_i$. Then each $B_i'$ is a simply-connected, open set; moreover, $B_i'\cap B_j'=\emptyset$, if $i\neq j$. This follows from the fact that for $i\neq j$, $D_i$ and $D_j$ come from interfaces of distinct contours, and the fact that $B_i\cap B_j=\emptyset$, for $i\neq j$.
Let $B'=\cup_{i=1}^{t}B'_i$. Then $B'$ is open, and $x,y,z\in P\setminus B'$, although $x$ and $z$ may be on the boundary of $B'$.
There is a path $p_{xy}'\subseteq [p_{xy}\cap (P\setminus B')]\cup \partial B'$, connecting $x$ and $y$, where $\partial B'$ is the boundary of $B'$. More precisely, $p_{xy}$ is divided by $\partial B'$ into segments; on each segment of $p_{xy}$ in $P\setminus B'$, $p'_{xy}$ follows the path of $p_{xy}$; for each segment of $p_{xy}$ in $B'$, $p'_{xy}$ follows the boundary of $B'$ to connect the two endpoints of the segment. This is possible since $B'$ consists of bounded, disjoint, simply-connected, open sets $B_i'$, for $1\leq i\leq t$, and both $x$ and $v$ are in the complement of $B'$ in $P$.
All the vertices along $p'_{xy}$ are in the same cluster. In particular, this implies that $x$ and $z$ are in the same infinite cluster $\xi$.
\end{proof}
\begin{lemma}\label{ct}If there exist exactly two infinite contours, then there exists an infinite cluster incident to both infinite contours.
\end{lemma}
\begin{proof}Let $C_1$, $C_2$ be the two infinite contours. Since there exist only two infinite contours, we can find two points $x\in C_1$, $y\in C_2$, and a self-avoiding path $p_{xy}$, consisting of edges of $G$ and two half-edges, one starting at $x$ and the other ending at $y$, and connecting $x$ and $y$, such that $p_{xy}$ does not intersect infinite contours except at $x$ and at $y$.
Let $v\in V$ be the first vertex of $G$ along $p_{xy}$ starting from $x$. By the proof of \Cref{cc}, $v$ is in an infinite cluster $\xi$ incident to $C_1$. Let $z$ be the first vertex of $G$ along $p_{xy}$. By \Cref{isc}, $z\in \xi$, and therefore $\xi$ is an infinite cluster incident to both $C_1$ and $C_2$.
\end{proof}
\begin{lemma}\label{cac}Let $\omega\in \Omega$. If there is exactly one infinite 0-cluster and exactly one infinite 1-cluster in $\omega$, then there exists an infinite contour that is incident to both the infinite 0-cluster and the infinite 1-cluster in $\omega$.
\end{lemma}
\begin{proof} Let $x$ be a vertex in the infinite 0-cluster, and let $y$ be a vertex in the infinite 1-cluster. Let $\ell_{xy}$ be a path joining $x$ and $y$ and consisting of edges of $G$.
By \Cref{io}, $\ell_{xy}$ must cross infinite contours an odd number of times. By \Cref{icazo}, if $\ell_{xy}$ crosses exactly one infinite contour, $C_{\infty}$, then $C_{\infty}$ is incident to both the infinite 0-cluster and the infinite 1-cluster, and so the lemma is proved in this case.
Suppose that there exist more than one infinite contour crossing $\ell_{xy}$. Let $C_1$ and $C_2$ be two distinct infinite contours crossing $\ell_{xy}$.
Let $u\in C_1\cap \ell_{xy}$ and $v\in C_2\cap \ell_{xy}$ (Here we interpret the contours and the paths as their embeddings to $P$, so that $u,v$ are points in $P$), such that the portion of $\ell_{xy}$ between $u$ and $v$, $p_{uv}$, does not cross any infinite contours except at $u$ and at $v$. As in the proof of Lemma \ref{cc}, let $u_1$ be the first vertex of $G$ along $p_{uv}$, starting from $u$; and let $v_1$ be the first vertex of $G$ along $p_{uv}$ starting from $v$. Let $u_2$ (resp.\ $v_2$) be the point along the line segment $[u,u_1]$ (resp.\ $[v,v_1]$) lying on an interface. Following the procedure in the proof of Lemma \ref{cc}, we can find an infinite cluster $\xi_1$, such that $u_1\in \xi_1$. The following cases might happen:
\begin{enumerate}[label=\Roman*]
\item $x\notin \xi_1$ and $y\notin \xi_1$;
\item $x\notin \xi_1$ and $y\in \xi_1$;
\item $x\in \xi_1$ and $y\notin \xi_1$;
\item $x\in \xi_1$ and $y\in\xi_1$.
\end{enumerate}
First of all, Case IV is impossible because we assume $x$ and $y$ are in two distinct infinite clusters. Secondly, if Case I is true, then there exist at least three infinite clusters, which is a contradiction to our assumption.
Case II and Case III can be handled using similar arguments, and we write down the proof of Case II here.
If Case II is true, first note that $y\in\xi_1$ implies that $C_1$ is incident to the infinite 1-cluster. Let $z$ be the first point in $C_1\cap \ell_{xy}$ (again interpret edges as line segments), when traveling along $\ell_{xy}$ starting from $x$. Let $p_{xz}$ be the portion of $\ell_{xy}$ between $x$ and $z$.
Next, we will prove the following claim by induction on the number of complete edges of $G$ along $p_{xz}$ (in contrast to the half edge along $p_{xz}$ with an endpoint $z$).
\begin{claim}\label{cl}
Under Case II, there is an infinite contour
incident to both the infinite 0-cluster and the infinite 1-cluster.
\end{claim}
Assume that the number of complete edges of $G$ along $p_{xz}$ is $n$, where $n=0,1,2,\ldots$.
First of all, consider the case when $n=0$. This implies that $C_1$ is incident to the infinite 0-cluster at $x$. Recall that $C_1$ is also incident to the infinite 1-cluster at $y$, and so \Cref{cl} is proved.
We make the following induction hypothesis:
\begin{itemize}
\item Claim \ref{cl} holds for $n\leq k$, where $k\geq 0$.
\end{itemize}
Now we consider the case when $n=k+1$.
The \df{interior points} of $p_{xz}$ are all points along $p_{xz}$ except $x$ and $z$. We consider two cases:
\begin{enumerate}[label=(\alph*)]
\item at interior points, $p_{xz}$ crosses only finite contours but not infinite contours;
\item at interior points, $p_{xz}$ crosses infinite contours.
\end{enumerate}
We claim that if Case (a) occurs, then $C_1$ is incident to both the infinite 0-cluster and the infinite 1-cluster. It suffices to show that $C_1$ is incident to the infinite 0-cluster.
Let $z_1$ be the first vertex in $V$ along $p_{xz}$ starting from $z$. According to \Cref{isc}, both $x$ and $z_1$ are in the infinite 0-cluster. We infer that $C_1$ is incident to the infinite 0-cluster at $x$, if $p_{xz}$ intersects only finite contours at interior points.
Now we consider Case (b). Let $C_3$ be an infinite contour crossing $p_{xz}$ at interior points. Obviously, $C_3$ and $C_1$ are distinct, because $C_1$ crosses $p_{xz}$ only at $z$. Let $w$ be the last point in $C_3\cap p_{xz}$, when traveling along $p_{xz}$, starting from $x$, and let $p_{wz}$ be the portion of $p_{xz}$ between $w$ and $z$. Assume $p_{wz}$ does not cross infinite contours at interior points.
Let $w_1$ be the first vertex of $G$ along $p_{wz}$, starting from $w$, and let $w_2$ be the midpoint of $w$ and $w_1$. According to the proof of Lemma \ref{cc}, we can find an infinite cluster $\xi_3$ including $w_1$. The following cases might happen:
\begin{enumerate}[label=\roman*]
\item $x\notin \xi_3$, and $y\notin\xi_3$;
\item $x\in \xi_3$, and $y\notin\xi_3$;
\item $x\notin \xi_3$, and $y\in\xi_3$;
\item $x\in \xi_3$, and $y\in\xi_3$.
\end{enumerate}
First of all, Case iv is impossible because we assume $x$ and $y$ are in two distinct infinite clusters. Secondly, if Case i is true, then there exist at least three infinite clusters, which is a contradiction to the assumption that there exists exactly one infinite 0-cluster and one infinite 1-cluster.
If Case ii is true, then $C_3$ is incident to the infinite 0-cluster including $x$. Since $w_1\in \xi_3$, and $p_{wz}$ does not cross infinite contours except at $w$ and $z$, by \Cref{isc}, we infer that $z\in \xi_3$, and $\xi_3$ is exactly the infinite 0-cluster including $x$. We conclude that $C_1$ is incident to the infinite 0-cluster including $x$ as well, and Claim \ref{cl} is proved.
If Case iii is true, then $C_3$ is incident to the infinite 1-cluster including $y$. Let $t$ be the first vertex in $p_{xz}\cap C_3$, when traveling from $p_{xz}$, starting at $x$, and let $p_{xt}$ be the portion of $p_{xz}$ between $x$ and $t$. We explore the path $p_{xt}$ as we have done for $p_{xz}$. Since the length of $p_{xz}$ is finite, and the number of full edges of $G$ along $p_{xt}$ is less than that of $p_{xz}$ by at least 1, we apply the induction hypothesis with $C_1$ replaced by $C_3$, $C_2$ replaced by $C_1$, $\xi_1$ replaced by $\xi_3$, $p_{xz}$ replaced by $p_{xt}$, and we conclude that there exists an infinite contour adjacent to both the infinite 0-cluster and infinite 1-cluster.
\end{proof}
\begin{lemma}\label{2c2c}Let $C_1$ and $C_2$ be two infinite contours, and let $\xi_0$ and $\xi_1$ be two infinite clusters. The following two cases cannot occur simultaneously.
\begin{itemize}
\item $\xi_0$ is incident to both $C_1$ and $C_2$;
\item $\xi_1$ is incident to both $C_1$ and $C_2$.
\end{itemize}
\end{lemma}
\begin{proof}Same arguments as Lemma 6.3 of \cite{HL16}.
\end{proof}
\section{Proof of \Cref{m23}}\label{p23}
In this section, we prove \Cref{m23}. We start with the following lemma.
\begin{lemma}\label{nooc}Let $G$ be the $[m,4,n,4]$ lattice satisfying (\ref{cmn1}) and
\begin{eqnarray}
\frac{1}{m}+\frac{1}{n}<\frac{1}{2}.\label{cmn3}
\end{eqnarray}
Let $\mu$ be a probability measure on $\Omega$ satisfying (A1). Let $t_1$ (resp.\ $t_2$) be the number of infinite $\LL_1$-contours (resp.\ $\LL_2$-contours). Then
\begin{eqnarray*}
\mu((t_1,t_2))=(1,1)=0.
\end{eqnarray*}
\end{lemma}
\begin{proof}
The proof is inspired by the proof of Corollary 3.6 of \cite{BS20}.
We embed $\LL_1$ and $\LL_2$ in the hyperbolic plane in such a way that every edge $e$ intersects its dual edge $e^*$ at one point $v_e$, and there are no other intersections of $\LL_1$ and $\LL_2$. We define a new graph $\hat{G}=(\hat{V},\hat{E})$, where $\hat{V}=V(\LL_1)\cup V(\LL_2)\cup\{v_e,e\in E(\LL_1)\}$, and an edge in $\hat{E}$ is either a half-edge of $E(\LL_1)$ joining a vertex in $V(\LL_1)$ and a vertex in $\cup\{v_e,e\in E(\LL_1)\}$, or a half-edge of $E(\LL_2)$ joining a vertex in $V(\LL_2)$ and a vertex in $\cup\{v_e,e\in E(\LL_1)\}$.
Let
\begin{eqnarray*}
\hat{\phi}:=\{[v,v_e]\in \hat{E}: v\in V(\LL_1),e\in \phi_1\}\cup\{[v_*,v_e]\in \hat{E}: v_*\in V(\LL_2),e_*\in \phi_2\}
\end{eqnarray*}
We say $\hat{\phi}$ is a \textbf{contour configuration} on $\hat{G}$, and each connected component of $\hat{\phi}$ is called a \textbf{contour}.
Then $\hat{\phi}$ is an invariant bond percolation on the quasi-transitive, non-amenable, planar, one-ended graph $\hat{G}$. Note that the number of infinite components of $\hat{\phi}$ is the number of infinite contours of $\phi_1$ plus the number of infinite contours of $\phi_2$. If there is a positive probability that $(t_0,t_1)=(1,1)$, then the number of infinite component in $\hat{\phi}$ is 2; which contradicts \Cref{lbs}.
\end{proof}
\begin{proposition}\label{pp}Let $G$ be the $[m,4,n,4]$ lattice with $m,n$ satisfying (\ref{cmn1}) and (\ref{cmn3}). Let $\omega\in \Omega$ be a $\Gamma$-invariant, $\Gamma_1$-ergodic constrained percolation on $G$. Let $s_0$ (resp.\ $s_1$) be the number of infinite 0-clusters (resp.\ 1-clusters) in $\omega$, and let $t_1$ (resp.\ $t_2$) be the number of infinite $\LL_1$-contours (resp.\ infinite $\LL_2$-contours) in $\omega$. Then almost surely $(s_0,s_1,t_1,t_2)\neq (0,0,0,0)$.
\end{proposition}
\begin{proof}The proof is inspired by Lemma 3.3 of \cite{BS20}. Let $\hat{G}=(\hat{V},\hat{E})$, $\hat{\phi}$ be defined as in the proof of \Cref{nooc}. Note that when $m,n$ satisfy (\ref{cmn1}) and (\ref{cmn3}), $\hat{G}$ is a quasi-transitive, non-amenable, planar and one-ended graph; and that the $[m,4,n,4]$ lattice is exactly the dual graph of $\hat{G}$. It is also known that quasi-transitive planar graphs with one end are unimodular; see \cite{LP}.
Define a \textbf{generalized contour} in a contour configuration $\hat{\phi}$ of $\hat{G}$ to be either a single vertex in $\hat{V}$ which has no incident present edges in $\hat{\phi}$, or a contour in $\hat{\phi}$. This way each vertex $v\in \hat{V}$ has a unique generalized contour in $\hat{\phi}$ passing through the vertex $v$.
Suppose that $(s_0,s_1,t_1,t_2)=(0,0,0,0)$ a.s. Then a.s.\ given a generalized contour $C$ of $\hat{\phi}$, there is a unique cluster $C'$ of $\omega$ surrounding it. Similarly, for every cluster $C$ in $\omega$, there is a unique contour $C'$ in $\hat{\phi}$ that surrounds it. Let $\mathcal{C}_0$ denote the set of all generalized contours of $\hat{\phi}$. We set
\begin{eqnarray*}
\mathcal{C}_{j+1}:=\{C'':C\in\mathcal{C}_j\}.
\end{eqnarray*}
For $C\in \mathcal{C}_0$ and $v\in \hat{V}$, let $r(C):=\sup\{j:C\in \mathcal{C}_j\}$, and define $r(v):=r(C)$ if $C$ is the generalized contour of $v$ in $\hat{\phi}$. For each $r$ let $\omega^r$ be the set of edges in $\hat{E}$ whose both endpoints $u,v\in \hat{V}$ satisfy $r(v)\leq r$ and $r(u)\leq r$. Then $\omega^r$ is an invariant bond percolation and for any $v\in \hat{V}$,
\begin{eqnarray*}
\lim_{r\rightarrow\infty}\mathbf{E}[\deg_{\omega^r}v]=\deg_Gv.
\end{eqnarray*}
By Lemma \ref{tbng}, we deduce that $\omega^r$ has infinite components with positive probability for all sufficiently large $r$.
However, since $(s_0,s_1,t_1,t_2)=(0,0,0,0)$, by the arguments above each vertex in $\hat{V}$ is surrounded by infinitely many contours. This implies that for any $r\in \mathbb{N}$, for any vertex $v\in \hat{V}$, there exists a contour $C$ surrounding $v$, such that $r(C)>r$, and therefore $r(C)\cap \omega^r=\emptyset$. As a result, the components in $\omega^r$ including $v$ is finite. Then the proposition follows from the contradiction.
\end{proof}
\begin{lemma}\label{lm23}Let $G$ be the $[m,4,n,4]$ lattice with $m,n$ satisfying (\ref{cmn1}), (\ref{cmn3}). Let $(s_0,s_1,t_1,t_2)$ be given as in \Cref{m23}.
\begin{enumerate}
\item Let $\mu$ be a probability measure on $\Omega$ satisfying (A2)(A7). Then
\begin{eqnarray*}
\mu((t_1,t_2)=(0,k))=0.
\end{eqnarray*}
for any integer $1\leq k\leq \infty$.
\item Let $\mu$ be a probability measure on $\Omega$ satisfying (A2)(A6). Then
\begin{eqnarray*}
\mu((t_1,t_2)=(k,0))=0.
\end{eqnarray*}
for any integer $1\leq k\leq \infty$.
\end{enumerate}
\end{lemma}
\begin{proof}We prove Part I here; Part II can be proved using exactly the same technique.
By (A2), either $\mu((t_1,t_2)=(0,k))=0$ or $\mu((t_1,t_2)=(0,k))=1$. Assume that $\mu((t_1,t_2)=(0,k))=1$; we shall obtain a contradiction. Since there exists an infinite $\LL_2$-contour; hence there exists an infinite cluster in $\lambda_2$ containing the infinite $\LL_2$-contour. By (A7) and the symmetry of $\lambda_2$, there exist an infinite 0-cluster and an infinite 1-cluster in $\lambda_2$. Note that the configuration in $\lambda_2\in\{0,1\}^{V(\LL_2)}$ naturally induces a configuration $\omega\in \Omega$ by the condition that the contour configurations corresponding to $\lambda_2$ and $\omega$ are the same. We can see that if in $\lambda_2$ there exist both an infinite 0-cluster and an infinite 1-cluster, then in the induced constrained configuration $\omega\in\Omega$, there is both an infinite 0-cluster and an infinite 1-cluster. By \Cref{io}, there exist an infinite $\LL_1$-contour. But this is a contradiction to the fact that $t_1=0$.
\end{proof}
\begin{lemma}\label{n0111}Let $G=(V,E)$ be the $[m,4,n,4]$ lattice with $m,n$ satisfying (\ref{cmn1}), (\ref{cmn3}). Let $\mu$ be a probability measure on $\Omega$ satisfying (A2), (A8). Let $(s_0,s_1,t_1,t_2)$ be given as in \Cref{m23}. Then
\begin{eqnarray*}
\mu((s_0,s_1)=(1,1))=0.
\end{eqnarray*}
\end{lemma}
\begin{proof}By (A2), either $\mu((s_0,s_1)=(1,1))=0$ or $\mu((s_0,s_1)=(1,1))=1$. Assume that $\mu((s_0,s_1)=(1,1))=1$; we shall obtain a contradiction.
Let $\omega\in\Omega$. We first construct a bond configuration $\omega_b\in \{0,1\}^E$ by letting an edge $e\in E$ to be present if and only if it joins two edges in $\omega$ with the same state; i.e.\ either both its endpoints has state 0; or both its end points has state 1. It is easy to check that the (0 or 1) clusters in $\omega$ are exactly the components in $\omega_b$. Then $\omega_b$ forms a $\Gamma_1$-invariant percolation on $G$. If $(s_0,s_1)=(1,1)$, then $\omega_b$ has exactly two infinite components. But this is a contradiction to \Cref{lbs}.
\end{proof}
\noindent\textbf{Proof of \Cref{m23} I.} Assume that $\mu$ is a probability measure on $\Omega$ satisfying (A2),(A3),(A7),(A8).
Let $(s_0,s_1,t_1,t_2)$ be given as in the theorem. By Lemma \ref{lbs}, we have $\mu$-a.s.\ $s_0\in\{0,1,\infty\}$, $s_1\in\{0,1,\infty\}$ and $t_1\in\{0,1,\infty\}$.
By (A2) (A3), we have $\mu$-a.s.\ $(s_0,s_1)\in\{(0,0),(1,1),(\infty,\infty)\}$. Hence we need to rule out the case that $(s_0,s_1)=(1,1)$ and the case that $(s_0,s_1)=(0,0)$. Almost surely we have $(s_0,s_1)\neq (1,1)$ by \Cref{n0111}. Now we show that almost surely $(s_0,s_1)\neq (0,0)$.
We claim that $\mu$-a.s.\ $t_1\in \{0,\infty\}$. Assume that $\mu$-a.s.\ $t_1=1$, we shall obtain a contradiction. Let $\tau$ be the unique infinite $\LL_1$-contour. Then $\tau$ forms an invariant bond percolation on $\LL_1$ which has a single component a.s.. By Lemma \ref{l2}, $p_c(\tau)<1$ a.s. However, $\tau$ is an even-degree subgraph of $\LL_1$ and $\LL_1$ has vertex-degree 3; as a result, $\tau$ must be a doubly-infinite self-avoiding path. This is a contradiction to the fact that $p_c(\tau)<1$. Therefore we have either $\mu$-a.s.\ $t_1=0$ or $\mu$-a.s.\ $t_1=\infty$.
If $\mu$-a.s.\ $t_1=\infty$, let $\phi$ be the contour configuration on $\LL_1\cup\LL_2$ corresponding to the constrained percolation configuration. Since each infinite contour in $\phi$ is a doubly-infinite self-avoiding path, if there are infinitely many infinite contours, then $\HH^2\setminus \phi$ has infinitely many unbounded components. Note also that there exists an infinite cluster in each infinite component of $\HH^2\setminus \phi$; hence $\mu$-a.s. $(s_0,s_1,t_1)=(\infty,\infty,\infty)$ in this case.
Now consider the case that $\mu$-a.s. $t_1=0$.
We assume that $\mu$-a.s.\ $(s_0,s_1,t_1)=(0,0,0)$ and shall again obtain a contradiction. By \Cref{pp}, a.s.\ $(s_0,s_1,t_1,t_2)\neq(0,0,0,0)$. Moreover, it is impossible to have $(s_0,s_1,t_1,t_2)=(0,0,0,\infty)$ since if $t_2=\infty$, then there are infinitely many infinite clusters. By \Cref{lm23}, a.s.\ $(s_0,s_1,t_1,t_2)\neq(0,0,0,1)$. Therefore $\mu((s_0,s_1,t_1)=(0,0,0))=0$.
We next assume that $\mu$-a.s.\ $(s_0,s_1,t_1)=(\infty,\infty,0)$. By \Cref{lm23}, $\mu$-a.s.\ $(s_0,s_1,t_1,t_2)=(\infty,\infty,0,0)$. Since there exists an infinite 0-cluster and an infinite 1-cluster, by \Cref{io}, there exists an infinite contour, and $s_0+s_1>0$. The contradiction implies that $\mu((s_0,s_1,t_1)=(\infty,\infty,0))=0$. This completes the proof of Part I of \Cref{m23}. $\hfill\Box$
\bigskip
\noindent\textbf{Proof of \Cref{m23} I.} Assume that $\mu$ is a probability measure on $\Omega$ satisfying (A2),(A3),(A6),(A7),(A8). By \Cref{m23} I, $\mu$-a.s.\ $(s_0,s_1,t_1)=(\infty,\infty,\infty)$. Part II of \Cref{m23} then follows from \Cref{lm23}.
$\hfill\Box$
\bigskip
\section{Proof of \Cref{m21} }\label{p212}
In this section, we prove \Cref{m21}.
We first prove that Parts (a) and (b) implies Part (c). If $\mu$ satisfies (A1)-(A7), then by (a) and (b), $\mu$-a.s. there are neither infinite primal contours nor infinite dual contours. Therefore $\mu$-a.s. there are no infinite contours.
Let $\mathcal{E}_0$ (resp.\ $\mathcal{E}_1$) be the event that there exists an infinite 0-cluster (resp.\ infinite 1-cluster). Assume that $\mu(\mathcal{E}_0\cup\mathcal{E}_1)>0$. Then by (A2),
\begin{eqnarray}
\mu(\mathcal{E}_0\cup\mathcal{E}_1)=1.\label{e01}
\end{eqnarray}
By (A3), $\mu(\mathcal{E}_0)=\mu(\mathcal{E}_1)$. By (A2), either $\mu(\mathcal{E}_0)=\mu(\mathcal{E}_1)=1$ or $\mu(\mathcal{E}_0)=\mu(\mathcal{E}_1)=0$. By (\ref{e01}), we have $\mu(\mathcal{E}_0)=\mu(\mathcal{E}_1)=1$. By \Cref{io}, $\mu$-a.s. there exists an infinite contour. But this is a contradiction to the fact that $\mu$-a.s. there are no infinite contours. Therefore $\mu$-a.s. there are no infinite clusters.
Next we prove (a) and (b). Note that the $[m,4,n,4]$ lattice $G$ is amenable if and only if
\begin{eqnarray}
\frac{1}{m}+\frac{1}{n}=\frac{1}{2}.\label{am2}
\end{eqnarray}
When $m,n$ are positive integers greater than or equal to 3, the only pairs of $(m,n)$ satisfying (\ref{am2}) are $(m,n)=(4,4)$, $(m,n)=(3,6)$ and $(m,n)=(6,3)$. When $(m,n)=(4,4)$, $G$ is the square grid embedded into $\RR^2$. In this case (a) and (b) were proved in \cite{HL16}. Then cases $(m,n)=(3,6)$ and $(m,n)=(6,3)$ can be proved in the same way. We write down the proof of the case when $(m,n)=(3,6)$ here.
When $(m,n)=(3,6)$, $\LL_1$ is the hexagonal lattice $\HH=(V(\HH),E(\HH))$ and $\LL_2$ is the triangular lattice $\TT=(V(\TT),E(\TT))$.
Recall that the number of \df{ends} of a connected graph is the supreme over its finite subgraphs of the number of infinite components that remain after removing the subgraph.
\begin{lemma}\label{l81}Assume that $(m,n)=(3,6)$. When $\mu$ satisfies (A1), (A2) and (A5), almost surely there exists at most one infinite contour in $\TT$.
\end{lemma}
\begin{proof}Let $\mathbb{N}$ be the set of all nonnegative integers. Let $\mathcal{N}$ be the number of infinite contours in $\TT$. By (A2), there exists $k_0\in\mathbb{N}\cup\{\infty\}$, s.t. $\mu(\mathcal{N}=k_0)=1$.
By \cite{blps} (see also Exercise 7.24 of \cite{LP}), (A1) and the fact that the triangular lattice $\TT$ is transitive and amenable, $\mu$-a.s. no infinite contours has more than 2 ends.
The triangular lattice $\TT$ can be obtained from a square grid $\mathbb{S}$ by adding a diagonal in each square face of $\mathbb{S}$.
Let $B_n$ be an $n\times n$ box of $\mathbb{S}$. Let $\widetilde{B}_n$ be the corresponding box in $\TT$, i.e.\ $\widetilde{B}_n$ can be obtained from $B_n$ by adding a diagonal edge on each square face of $B_n$.
Let $\phi$ (resp.\ $\widetilde{\phi}$) be a contour configuration on $\mathbb{S}$ (resp.\ $\TT$), such that $\phi$ and $\widetilde{\phi}$ satisfy the following conditions (note that the vertices in $\partial B_n$ and $\partial\widetilde{B}_n$ are in 1-1 correspondence)
\begin{itemize}
\item for each vertex $v\in\partial B_n$, no edges incident to $v$ outside $\widetilde{B}_n$ are present in $\widetilde{\phi}$ if and only if no edges incident to $v$ outside $B_n$ are present in $\phi$;
\item for each vertex $v\in\partial B_n$, if there are incident present edges of $v$ in $\widetilde{\phi}$ outside $\widetilde{B}_n$, then the parity of the number of incident present edges of $v$ outside $\widetilde{B}_n$ in $\widetilde{\phi}$ is the same as the parity of the number of incident present edges of $v$ outside $B_n$ in $\phi$; i.e.\ either both numbers are even or both are odd.
\end{itemize}
Let $n\geq 2$. Given $\phi$, we can find a configuration $\xi$ in $B_n$, such that $[\phi\setminus B_n]\cup\xi$ is a contour configuration on $\mathbb{S}$ (i.e.\ each vertex of $\mathbb{S}$ has an even number of incident present edges in $[\phi\setminus B_n]\cup\xi$), and all the incident present edges of $\partial B_n$ outside $B_n$ are in the same contour; see Lemma 4.2 of \cite{HL16}. If $\widetilde{\phi}$ and $\phi$ satisfy the conditions described above, then $[\widetilde{\phi}\setminus \widetilde{B}_n]\cup\xi$ is a contour configuration on $\TT$, and all the incident present edges of $\partial\widetilde{B}_n$ outside $\widetilde{B}_n$ are in the same contour.
Note that $\xi$ can be obtained from $\widetilde{\phi}\cap \widetilde{B}_n$ by changing configurations on finitely many triangles in $\widetilde{B}$ as described in (A6). That is because any contour configuration on $\TT$ naturally induces two site configurations $\omega$, $1-\omega$, in $\{0,1\}^{V(\HH)}$, such that two adjacent vertices in $\HH$ have different states if and only if the edge in $\TT$ separating the two vertices are present in the contour configuration. Any two site configurations in $\{0,1\}^{V(\HH)}$ differ only in $\widetilde{B_n}$ can be obtained from each other by changing states on finitely many vertices in $V(\HH)\cap\widetilde{B}_n$. Changing the state at a vertex in $V(\HH)$ corresponds to changing the states on all the edges of the dual triangle face including the vertex in the contour configuration of $\TT$.
We claim that $k_0\in\{0,1\}$. Indeed, if $2\leq k_0<\infty$, we can find a box $\widetilde{B}_n$ in $\TT$, such that $\widetilde{B}_n$ intersects all the $k_0$ infinite contours. Then we can change configurations on finitely many triangles in $\widetilde{B}_n$, such that after the configuration change, there is exactly one infinite contour. By the finite energy assumption (A5), with positive probability, there exists exactly one infinite contour, but this is a contradiction to $\mu(\mathcal{N}=k_0)=1$, where $2\leq k_0<\infty$.
If $k_0=\infty$, we can find a box $\widetilde{B}_m$ in $\TT$, such that $\widetilde{B}_m$ intersects at least 3 infinite contours. Then we can change configurations on finitely many triangles in $\tilde{B}_n$, such that after the configuration change, all the infinite contours intersecting $\tilde{B}_m$ merge into one infinite contour, which has at least 3 ends. By (A5), with positive probability there exists an infinite contour with more than 2 ends. But this is a contradiction to the fact that almost surely no infinite contours have more than two ends.
\end{proof}
\begin{lemma}\label{l82}Assume that $(m,n)=(3,6)$. When $\mu$ satisfies (A1) and (A4), almost surely there exists at most one infinite contour in $\HH$.
\end{lemma}
\begin{proof}Recall that a contour is a connected set of edges in which each vertex has an even number of incident present edges. Since the hexagonal lattice $\HH$ is a cubic graph, i.e.\ each vertex has 3 incident edges; each vertex in a contour of $\HH$ has 2 incident present edges. As a result, each contour in $\HH$ is either a self-avoiding cycle or a doubly-infinite self-avoiding path. In particular, each infinite contour in $\HH$ is a doubly-infinite self-avoiding path.
Each contour configuration in $\HH$ naturally induces two site configurations $\omega,1-\omega$ in $\{0,1\}^{V(\TT)}$, in which two adjacent vertices of $\TT$ have the same state if and only if the dual edge in $\HH$ separating the two vertices is absent in the contour configuration. The finite energy assumption (A4) implies the finite energy in the induced site configuration in $\{0,1\}^{V(\TT)}$; see \cite{bk89} for a definition. When $\mu$ satisfies (A1) (A4), by the result in \cite{bk89}, almost surely there exists at most one infinite 1-cluster and at most one infinite 0-cluster. In particular, there exist at most two infinite clusters. However, if in $\HH$ there are more than one infinite contour, then there are at least two doubly-infinite self-avoiding paths in $\HH$. As a result, the number of infinite clusters in the induced site configuration on $\TT$ is at least 3. The contradiction implies the lemma.
\end{proof}
\begin{lemma}\label{l83}Let $\omega\in\Omega$ be a constrained percolation configuration on the $[3,4,6,4]$ lattice $G$. Let $\psi=\phi(\omega)\in\Phi$ be the corresponding contour configuration in $E(\HH)\cup E(\TT)$. Assume that $\psi=\psi_1\cup\psi_2$, where $\psi_1$ (resp.\ $\psi_2$) is the contour configuration in $\HH$ (resp. $\TT$). If there is an infinite contour in $\psi_1$ (resp.\ $\psi_2$), then there is an infinite cluster in $\phi^{-1}(\psi_2)$ (resp.\ $\phi^{-1}(\psi_1)$).
\end{lemma}
\begin{proof}Assume that there is an infinite contour $C$ in $\HH$ (resp.\ $\TT$). Let $V_{C}$ be the set consisting of all the vertices of $G$ such that
\begin{itemize}
\item $v\in V_C$ if and only if $v$ is a vertex of a face of $G$ crossed by an edge present in the contour $C$.
\end{itemize}
Let $F$ be a square face of $G$ crossed by $C$; then all the vertices in $F$ are in the same cluster of $\phi^{-1}(\psi_2)$ (resp.\ $\phi^{-1}(\psi_1)$). That is because $\psi_1\cap\psi_2=\emptyset$, if $F$ is crossed by $C\subseteq\psi_1$ (resp.\ $C\subseteq\psi_2$), then $F\cap \psi_2=\emptyset$ (resp. $F\cap\psi_1=\emptyset$).
Let $F'$ be a triangle (resp.\ hexagon) face of $G$ crossed by $C$; then all the vertices in $F$ are also in the same cluster of $\phi^{-1}(\psi_2)$ (resp.\ $\phi^{-1}(\psi_1)$). That is because the boundary edges of $F'$ cannot be crossed by edges of $\TT$ (resp.\ $\HH$) at all.
We claim that all the vertices in $V_C$ are in the same cluster in $\phi^{-1}(\psi_2)$ (resp.\ $\phi^{-1}(\psi_1)$). Indeed, for any two vertices $u,v\in V_C$, we can find a sequence of faces $F_0,F_1,\ldots,F_k$, such that
\begin{itemize}
\item $u\in F_0$ and $v\in F_k$; and
\item for $0\leq i\leq k$, $F_i$ is crossed by $C$;
\item for $1\leq j\leq k$, $F_j$ and $F_{j-1}$ share a vertex.
\end{itemize}
Then $u$ and $v$ are in the same cluster in $\phi^{-1}(\psi_2)$ (resp.\ $\phi^{-1}(\psi_1)$) since all the vertices in $\cup_{i=0}^{n}F_i$ are in the same cluster in $\phi^{-1}(\psi_2)$ (resp.\ $\phi^{-1}(\psi_1)$).
Moreover, $|V_C|=\infty$ since $C$ is an infinite contour. Therefore, $\phi^{-1}(\psi_2)$ (resp.\ $\phi^{-1}(\psi_1)$) has an infinite cluster.
\end{proof}
Parts (a) and (b) can be proved using similar techniques; we write down the proof of (a) here.
Let $\mu$ be a probability measure on $\Omega$ satisfying (A1)-(A6). Assume that with strictly positive probability, there exist infinite contours in $\TT$. Then by (A2), $\mu$-a.s. there exist infinite contours in $\TT$. By \Cref{l81}, $\mu$-a.s.\ there exists exactly one infinite contour $C_1$ in $\TT$. By \Cref{l83}, $\mu$-a.s. there exist infinite clusters in $\phi^{-1}(\psi_1)$. Let $\mathcal{F}_0$ (resp.\ $\mathcal{F}_1$) be the event that there exists an infinite 0-cluster (resp.\ infinite 1-cluster) in $\phi^{-1}(\psi_1)$, then
\begin{eqnarray}
\lambda_1(\mathcal{F}_0\cup\mathcal{F}_1)=1\label{f01},
\end{eqnarray}
where the probability measure $\lambda_1$ is defined before (A6). By (A6), and the symmetry of $\lambda_1$, either $\lambda_1(\mathcal{F}_0)=\lambda_1(\mathcal{F}_1)=0$, or $\lambda_1(\mathcal{F}_0)=\lambda_1(\mathcal{F}_1)=1$. By (\ref{f01}), we have $\lambda_1(\mathcal{F}_0\cap\mathcal{F}_1)=1$. By \Cref{io}, $\mu$-a.s. there are infinite contours in $\HH$. By \Cref{l82}, $\mu$-a.s. there is exactly one infinite contour $C_2$ in $\HH$.
Hence there is exactly one infinite contour $C_2$ in $\HH$ and exactly one infinite contour $C_1$ in $\TT$. By \Cref{ct}, there exists
an infinite cluster incident to both $C_1$ and $C_2$.
Let $\mathcal{H}_0$ (resp.\ $\mathcal{H}_1$) be the event that there exists an infinite 0-cluster (resp.\ infinite 1-cluster) in $\omega$ incident to both the infinite contour in $\HH$ and the infinite contour in $\TT$. Then
\begin{eqnarray}
\mu(\mathcal{H}_0\cup\mathcal{H}_1)=1.\label{eq:h01}
\end{eqnarray}
By (A3), $\mu(\mathcal{H}_0)=\mu(\mathcal{H}_1)$. By (A2), either $\mu(\mathcal{H}_0)=\mu(\mathcal{H}_1)=0$, or $\mu(\mathcal{H}_0)=\mu(\mathcal{H}_1)=1$. By (\ref{eq:h01}), $\mu(\mathcal{H}_0)=\mu(\mathcal{H}_1)=1$, therefore $\mu(\mathcal{H}_0\cap\mathcal{H}_1)=1$, i.e.\ there exist an infinite 0-cluster $\xi_0$ and an infinite 1-cluster $\xi_1$, such that $\xi_0$ is incident to both $C_1$ and $C_2$, and $\xi_1$ is incident to both $C_1$ and $C_2$. But this is a contradiction to \Cref{2c2c}. Therefore we conclude that $\mu$-a.s. there are no infinite contours in $\TT$; this completes the proof of Part (a).
\section{Proof of \Cref{m22}}\label{p211}
In this section, we prove \Cref{m22}.
\begin{lemma}\label{l71}Let $G$ be an $[m,4,m,4]$ lattice with $m\geq 5$. Let $\mu$ be a probability measure on $\Omega$ satisfying (A1)-(A3). Then the distribution of infinite clusters can only be one of the following 2 cases.
\begin{enumerate}
\item There are no infinite clusters $\mu$-a.s.
\item There are infinitely many infinite 1-clusters and infinitely many infinite 0-clusters $\mu$-a.s.
\end{enumerate}
\end{lemma}
\Cref{l71} can be obtained from \Cref{n0111}; it may also be proved using different arguments below.
\begin{proof}Let $s_0$ (resp.\ $s_1$) be the number of infinite 0-clusters (resp.\ 1-clusters). By \Cref{lbs} and (A2) (A3), there exist $k\in\{0,1,\infty\}$, such that $\mu((s_0,s_1)=(k,k))=1$. It suffices to show that $k\neq 1$.
Let $\sA$ be the event that $(s_0,s_1)=(1,1)$. Assume that $\mu(\sA)=1$, we will obtain a contradiction.
As explained before the constrained site configurations on $G$ correspond to contour configurations in $\LL_1\cup\LL_2$.
Since $\mu$-a.s.\ there exists exactly one infinite 0-cluster and exactly one infinite 1-cluster simultaneously, then by \Cref{cac}, $\mu$-a.s.\ there exists an infinite primal or dual contour incident to both the infinite 0-cluster and the infinite 1-cluster. Let $\sD_1$ (resp.\ $\sD_2$) be the event that there exists an infinite primal (resp.\ dual) contour in $\LL_1$ (resp.\ $\LL_2$), incident to both the infinite 0-cluster and the infinite 1-cluster. So we have
\begin{eqnarray}
\mu(\sD_1\cup\sD_2)=1.\label{c1}
\end{eqnarray}
By (A1), we have
\begin{eqnarray}
\mu(\sD_1)=\mu(\sD_2).\label{c2}
\end{eqnarray}
Moreover, by (A2), we have either
\begin{eqnarray}
\mu(\sD_1)=0\ \mathrm{or}\ \mu(\sD_1)=1.\label{c3}
\end{eqnarray}
Combining (\ref{c1}), (\ref{c2}) and (\ref{c3}), we have
\begin{eqnarray}
\mu(\sD_1\cap\sD_2)=1. \label{dcd}
\end{eqnarray}
Thus, by (\ref{dcd}), we have exactly one infinite 1-cluster on $G$, denoted by $\xi_1$ and exactly one infinite 0-cluster on $G$, denoted by $\xi_0$. There is an infinite primal contour incident to both $\xi_0$ and $\xi_1$, denoted by $C_1$; as well as an infinite dual contour incident to both $\xi_0$ and $\xi_1$, denoted by $C_2$. But this is impossible by \Cref{2c2c}. The contradiction implies the lemma.
\end{proof}
\begin{lemma}\label{l72}Let $\mu$ be a probability measure on $\Omega$ satisfying (A1)-(A3). Then the distribution of infinite contours can only be one of the following 2 cases.
\begin{enumerate}
\item There are no infinite primal contours or infinite dual contours $\mu$-a.s..
\item There are infinitely many infinite primal contours and infinitely many infinite dual contours $\mu$-a.s..
\end{enumerate}
\end{lemma}
\begin{proof}The primal (resp.\ dual) contours form an invariant bond percolation on $\LL_1$ (resp.\ $\LL_2$) under $\mu$. Let $t_1$ (resp.\ $t_2$) be the number of infinite primal (resp.\ dual) contours. By \Cref{lbs} and (A1)-(A3), only 3 cases may occur:
\begin{enumerate}[label=\roman*.]
\item $\mu$-a.s. $(t_1,t_2)=(0,0)$;
\item $\mu$-a.s. $(t_1,t_2)=(\infty,\infty)$;
\item $\mu$-a.s. $(t_1,t_2)=(1,1)$.
\end{enumerate}
It remains to exclude in Case iii.. Assume that Case iii.\ occurs. Let $C_1$ (resp. $C_2$) be the unique infinite primal (resp.\ dual) contour. By \Cref{ct}, there exists an infinite cluster incident to both $C_1$ and $C_2$. Moreover, by (A2)-(A3), $\mu$-a.s.\ there exists an infinite 0-cluster incident to both $C_1$ and $C_2$, as well as an infinite 1-cluster incident to both $C_1$ and $C_2$. But this is impossible by \Cref{2c2c}.
\end{proof}
\bigskip
\noindent\textbf{Proof of \Cref{m22}.} By \Cref{l71,l72,pp}, it suffices to show that there exists an infinite cluster $\mu$-a.s.\ if and only if there exists an infinite contour $\mu$-a.s.\ if $\mu$ satisfies (A1)-(A3).
First assume that there exists an infinite cluster $\mu$-a.s. By (A2)-(A3), $\mu$-a.s.\ there exist both an infinite 0-cluster and an infinite 1-cluster. By \Cref{io}, $\mu$-a.s.\ there exists an infinite contour.
Now assume that there exists an infinite contour $\mu$-a.s. By (A1)-(A2), $\mu$-a.s.\ there exist both an infinite primal contour and an infinite dual contour. By \Cref{cc}, $\mu$-a.s. there exists an infinite cluster.
$\hfill\Box$
\section{Percolation on transitive, triangular tilings of hyperbolic plane}\label{peh}
In this section, we discuss the applications of the techniques developed in the proof of \Cref{m23}
to prove results concerning unconstrained site percolation on vertex-transitive, triangular tilings of the hyperbolic plane. We also discuss results about Bernoulli percolation on such graphs. These results will be used to prove \Cref{ipl,coii,xorc}.
\begin{lemma}\label{nzz}\label{lnz}Let $\LL_2$ a vertex-transitive, regular tiling of the hyperbolic plane with triangles, such that each vertex has degree $n\geq 7$. Consider an $\mathrm{Aut}(\LL_2)$-invariant site percolation $\omega$ on $\LL_2$ with distribution $\mu$. Assume that $\mu$ is $\mathrm{Aut}(\LL_2)$-ergodic. Let $s_0$ (resp.\ $s_1$) be the number of infinite 0-clusters (resp.\ infinite 1-clusters) in the percolation. Then
\begin{eqnarray*}
\mu((s_0,s_1)=(0,0))=0.
\end{eqnarray*}
\end{lemma}
\begin{proof}Since the event $\{(s_0,s_1)=(0,0)\}$ is $\mathrm{Aut}(\LL_2)$-invariant, and $\mu$ is $\mathrm{Aut}(\LL_2)$-ergodic, either $\mu((s_0,s_1)=(0,0))=0$ or $\mu((s_0,s_1)=(0,0))=1$. Assume that $\mu((s_0,s_1)=(0,0))=1$; we shall obtain a contradiction.
Note that the dual graph $\LL_1$ of $\LL_2$ is a vertex-transitive, non-amenable, planar graph in which each vertex has degree 3. A contour configuration $\phi(\omega)\subset E(\LL_1)$ is a subset of edges of $\LL_1$ in which each present edge has a dual edge in $E(\LL_2)$ joining exactly one vertex with state 0 and one vertex with state 1 in $\omega$. As usual, a contour is a maximal connected component of present edges in a contour configuration. Each contour configuration, by definition, must be an even-degree subgraph of $\LL_1$. Given that $\LL_1$ has vertex-degree 3, each vertex in $\LL_1$ is incident to zero or two present edges in a contour configuration. Let $t$ be the number of infinite contours. Each infinite contour on $\LL_1$ must be a doubly infinite self-avoiding path.
If $t\geq 1$, let $C_{\infty}$ be an infinite contour. Then $\HH^2\setminus C_{\infty}$ has exactly two unbounded components, since $C_{\infty}$ is a doubly-infinite self-avoiding path. Let $V_{\infty}$ be the set of all vertices of $\LL_2$ lying on a face crossed by $C_{\infty}$. Given $C_{\infty}$ a fixed orientation. Let $V_{\infty}^{+}$ (resp.\ $V_{\infty}^{-}$) be the subset of $V_{\infty}$ consisting of all the vertices in $V_{\infty}$ on the left hand side (resp.\ right hand side) of $C_{\infty}$ when traversing $C_{\infty}$ along the given orientation. Then exactly one of $V_{\infty}^{+}$ and $V_{\infty}^{-}$ is part of an infinite 1-cluster, and the other is part of an infinite 0-cluster. Therefore we have $s_0\geq1$ and $s_1\geq 1$ if $t\geq 1$.
Define a \textbf{generalized contour} in a contour configuration $\phi\subset E(\LL_1)$ to be either a single vertex in $V(\LL_1)$ which has no incident present edges in $\phi$, or a contour in $\phi$. This way for each vertex $v\in V(\LL_1)$, and each contour configuration $\phi\subset E(\LL_1)$, there is a unique generalized contour passing through $v$.
By the arguments in the last paragraph, if $(s_0,s_1)=(0,0)$, then $t=0$.
Now consider the case when $(s_0, s_1, t)=(0,0,0)$. Given a cluster $C$ in $\omega$, there is a unique contour $C'$ of $\phi(\omega)$ surrounding $\xi$. Similarly, for every generalized contour $C'$, there is a cluster $C''$ that surrounds $C'$. Let $\mathcal{C}_0$ denote the set of all generalized contours of $\omega$. We set
\begin{eqnarray*}
\mathcal{C}_{j+1}:=\{C'':K\in\mathcal{C}_j\}.
\end{eqnarray*}
For $C\in \mathcal{C}_0$ and $v\in V(\LL_1)$, let $r(C):=\sup\{j:C\in \mathcal{C}_j\}$, and define $r(v):=r(C)$ if $C$ is the generalized contour of $v$ in $\phi(\omega)$. For each $r$ let $\omega^r$ be the set of edges in $E(\LL_1)$ whose both end-vertices $u,v\in V(\LL_1)$ satisfy $r(v)\leq r$ and $r(u)\leq r$. Then $\omega^r$ is an invariant bond percolation and
\begin{eqnarray*}
\lim_{r\rightarrow\infty}\mathbf{E}[\deg_{\omega^r}v]=3.
\end{eqnarray*}
By Lemma \ref{tbng}, we deduce that $\omega^r$ has infinite components with positive probability for all sufficiently large $r$.
However, since $(s_0,s_1,t)=(0,0,0)$, by the arguments above each vertex in $v\in V(\LL_1)$ is surrounded by infinitely many contours. This implies that for any $r\in \mathbb{N}$, for any vertex $v\in V(\LL_1)$, there exists a contour $C$ surrounding $v$, such that $r(C)>r$, and therefore $r(C)\cap \omega^r=\emptyset$. As a result, the components in $\omega^r$ including $v$ is finite. The contradiction implies the lemma.
\end{proof}
\begin{lemma}\label{t01}Let $\LL_2$ be the regular tiling of the hyperbolic plane with triangles, such that each vertex has degree $n\geq 7$. Consider an $\mathrm{Aut}(\LL_2)$-invariant site percolation $\omega$ on $\LL_2$ with distribution $\mu$. Let $t$ be the number of infinite contours. Then $\mu$-a.s.\ $t\in\{0,\infty\}$
\end{lemma}
\begin{proof}By \Cref{lbs}, $\mu$-a.s.\ $t\in \{0,1,\infty\}$. Let $\sA$ be the event that $t=1$. Assume $\mu(\sA)>0$, we shall obtain a contradiction. Conditional on the event $\sA$, let $\tau$ be the unique infinite contour. Since $\tau$ is a infinite, connected, even-degree subgraph of $\LL_1$, and each vertex of $\LL_1$ has degree 3, $\tau$ must be a doubly infinite self-avoiding path. Then $p_c(\tau)=1$. But this is a contradiction to \Cref{l2}. Then the lemma follows.
\end{proof}
\begin{lemma}\label{ozz}Let $\LL_2$ be the regular tiling of the hyperbolic plane with triangles, such that each vertex has degree $n\geq 7$. Consider an $\mathrm{Aut}(\LL_2)$-invariant site percolation $\omega$ on $\LL_2$ with distribution $\mu$. Assume that $\mu$ is $\mathrm{Aut}(\LL_2)$-ergodic. Let $s_0$ (resp.\ $s_1$) be the number of infinite 0-clusters (resp.\ infinite 1-clusters) in the percolation. Then
\begin{eqnarray*}
\mu((s_0,s_1)=(1,1))=0.
\end{eqnarray*}
\end{lemma}
\begin{proof}We may construct a $[3,4,n,4]$ lattice embedded into the hyperbolic plane $\HH^2$, such that the $[3,4,n,4]$ lattice, $\LL_1$ and $\LL_2$ satisfy the conditions as described in \Cref{xorh}. Then each percolation configuration $\omega$ in $\{0,1\}^{V(\LL_2)}$ induces a constrained configuration $\widetilde{\omega}\in \Omega$ by the condition that $\omega$ and $\widetilde{\omega}$ has the same contour configuration; and that $\omega(v)=1$ for $v\in V(\LL_2)$ if and only if all the vertices of the $[3,4,n,4]$ lattice in the degree-$n$ face of $\LL_1$ containing $v$ have state 1 in $\widetilde{\omega}$.
Then the lemma follows from \Cref{n0111}.
\end{proof}
\begin{lemma}\label{nzi}\label{lnz}Let $\LL_2$ be the regular tiling of the hyperbolic plane with triangles, such that each vertex has degree $n\geq 7$. Consider a site percolation $\omega$ on $\LL_2$ with distribution $\mu$. Let $s_0$ (resp.\ $s_1$) be the number of infinite 0-clusters (resp.\ infinite 1-clusters) in the percolation. Then
it is not possible that
\begin{eqnarray*}
(s_0,s_1)=(0,k);\ \mathrm{for}\ 2\leq k\leq \infty.
\end{eqnarray*}
\end{lemma}
\begin{proof}Assume that $s_0=0$, and that there exist at least two distinct infinite 1-clusters $C_1$ and $C_2$. Let $\ell$ be a path consisting of edges of $\LL_2$ joining a vertex $x\in C_1$ and a vertex $y\in C_2$. If $\ell$ does not cross infinite contours, then we can find another path $\ell'$ joining $x$ and $y$ such that $\ell'$ does not cross contours at all. Then $C_1=C_2$. The contradiction implies that there exists at least one infinite contour in $\LL_1$. Since each infinite contour in $\LL_1$ is a doubly infinite self-avoiding path, if there exists an infinite contour, then there exist at least one infinite 0-cluster and at least one infinite 1-cluster. But this is a contradiction to the fact that $s_0=0$.
\end{proof}
\begin{definition}Let $G=(V,E)$ be a graph. Given a set $A\in 2^V$, and a vertex $v\in V$, denote $\Pi_{v}A=A\cup\{v\}$. For $\sA\subset 2^V$, we write $\Pi_v\sA=\{\Pi_v A: A\in \mathcal{A}\}$. A site percolation process $(\mathbf{P},\omega)$ on $G$ is \textbf{insertion-tolerant} in $\mathbf{P}(\Pi_v\sA)>0$ for every $v\in V$ and every measure $\sA\subset 2^V$ satisfying $\mathbf{P}(\sA)>0$.
\end{definition}
We can similarly define an insertion tolerant bond percolation by replacing a vertex with an edge in the above definition. A bond percolation is \textbf{deletion tolerant} if $\mathbf{P}[\Pi_{\neg e}\sA]>0$ whenever $e\in E(G)$ and $\mathbf{P}(\sA)>0$, where $\Pi_{\neg e}\omega=\omega\setminus\{e\}$.
\begin{lemma}\label{icie}Let $G$ be a graph with a transitive, unimodular, closed automorphism group $\Gamma\subset \mathrm{Aut}(G)$. If $(\mathbf{P},\omega)$ is a $\Gamma$-invariant, insertion-tolerant percolation process on $G$ with infinitely many infinite clusters a.s., then a.s. every infinite cluster has infinitely many ends.
\end{lemma}
\begin{proof}See Proposition 3.10 of \cite{LS99}; see also \cite{HP} and \cite{HPS}.
\end{proof}
\begin{lemma}\label{l147}Let $\LL_2$ be the regular tiling of the hyperbolic plane with triangles, such that each vertex has degree $n\geq 7$. Let $\omega\in\{0,1\}^{V(\LL_2)}$ be a site percolation configuration on $\LL_2$. Assume that there exists an infinite 1-cluster $\eta\subseteq\omega$ with infinitely many ends. Then there exist at least two infinite 0-clusters in $\omega$.
\end{lemma}
\begin{proof}Since $\eta\subseteq \omega$ has infinitely many ends, there exists a finite box $B$ of $\LL_1$, such that $\eta\setminus B$ has at least two distinct infinite components. Let $X$, $Y$ be two distinct infinite components of $\eta\setminus B$. Define the boundary of $X$ (resp.\ Y), $\partial X$ (resp.\ $\partial Y$) to be the set of all edges in $\LL_1$, such that each dual edge has exactly one endpoint in $X$ (resp.\ Y), and one endpoint in $\LL_2\setminus B$ and not in $X$ (resp.\ not in $Y$). Then $\partial X$ and $\partial Y$ are part of contours - each vertex of $\LL_1$ in $\partial X$ and $\partial Y$ has degree 1 or 2, whose degree-1 vertices are along $\partial B$ (here $\partial B$ consists of all the edges of $\LL_1$ on the boundary of the finite box $B$).
Each component of $\partial X$ or $\partial Y$ (here we assume that points on $\partial B$ are not included in $\partial X$ or $\partial Y$) must be one of the following three cases:
\begin{enumerate}
\item a finite component; or
\item a doubly infinite self-avoiding path which does not intersect $\partial B$;
\item a singly-infinite self-avoiding path starting from a vertex along $\partial B$.
\end{enumerate}
Let $B_1\supset B$ be a box of $\LL_2$ containing $B$. Then the embedding of $\partial B_1$ into $\HH^2$ is a simple closed curve consisting of edges in $\LL_2$. Since $X$ and $Y$ are two infinite components of $\eta\setminus B$, we deduce that $\partial B_1\cap X\neq \emptyset$ and $\partial B_1\cap Y\neq \emptyset$. Since $\partial B_1$ is a closed curve, there exist $x_1,x_2\in \partial B_1\cap X$ and $y_1,y_2\in \partial B_1\cap Y$, such that there are segments $p_{x_1y_1}\subset \partial B_1$ joining $x_1$ and $y_1$, and $p_{x_2y_2}\subset \partial B_1$ joining $x_2$ and $y_2$ such that $p_{x_1y_1}$ and $p_{x_2y_2}$ does not intersect each other except possibly at $x_1,x_2,y_1,y_2$; $p_{x_1y_1}$ does not intersect $X\cup Y$ except at $x_1,y_1$; and that $p_{x_2y_2}$ does not intersect $X\cup Y$ except at $x_2,y_2$.
Then we claim that both $p_{x_1y_1}$ and $p_{x_2y_2}$ cross a component of $\partial X$ of Type III, and a component of $\partial Y$ of Type III. To see why that is true, assume that $p_{x_1y_1}$ crosses only components of $\partial X$ of Type I or II, then we can find a path $q_{x_1y_1}$ consisting of edges in $\LL_2\setminus B$ and joining $x_1$ and $y_1$ such that $q_{x_1y_1}$ does not cross $\partial X$ at all. Then $Y$ and $X$ are the same component of $\eta\setminus B$. The contradiction implies that $p_{x_1y_1}$ must cross a component of $\partial X$ of Type III. Similarly, $p_{x_1y_1}$ must cross a component of $\partial Y$ of Type III; $p_{x_2y_2}$ must cross a component of $\partial X$ of Type III, and a component of $\partial Y$ of Type III.
Let $\ell_1$ (resp.\ $\ell_2$) be a component of $\partial X$ of Type III crossed by $p_{x_1y_1}$ (resp.\ $p_{x_2y_2}$). Let $V_1$ (resp.\ $V_2$) consist of all the vertices of $\LL_2$ on a triangle face crossed by $\ell_1$ (resp.\ $\ell_2$) but not in $X$. Then $V_1$ is part of an infinite 0-cluster $\xi$, and $V_2$ is part of an infinite 0-cluster $\zeta$; and moreover, $\xi$ and $\zeta$ are distinct since both $p_{x_1y_1}$ and $p_{x_2y_2}$ also cross components of $\partial Y$ of Type III. This completes the proof.
\end{proof}
\begin{lemma}\label{oiz}\label{lnz}Let $\LL_2$ be the regular tiling of the hyperbolic plane with triangles, such that each vertex has degree $n\geq 7$. Consider an $\mathrm{Aut}(\LL_2)$-invariant, insertion-tolerant site percolation $\omega$ on $\LL_2$ with distribution $\mu$. Assume that $\mu$ is $\mathrm{Aut}(\LL_2)$-ergodic. Let $s_0$ (resp.\ $s_1$) be the number of infinite 0-clusters (resp.\ infinite 1-clusters) in the percolation. Then
\begin{eqnarray*}
\mu((s_0,s_1)=(1,\infty))=0.
\end{eqnarray*}
\end{lemma}
\begin{proof}Assume that $\mu(s_0,s_1)=(1,\infty)=1$; we shall obtain a contradiction. By \Cref{icie}, a.s. every infinite 1-cluster has infinitely many ends. Then the lemma follows from \Cref{l147}.
\end{proof}
\begin{proposition}\label{p118}Let $\LL_2$ be the regular tiling of the hyperbolic plane with triangles, such that each vertex has degree $n\geq 7$. Consider an $\mathrm{Aut}(\LL_2)$-invariant, $\mathrm{Aut}(\LL_2)$-ergodic, insertion-tolerant site percolation $\omega$ on $\LL_2$ with distribution $\mu$. Let $s_0,s_1,t$ be given as in \Cref{nzz,t01}, then
\begin{eqnarray*}
(s_0,s_1,t)\in\{(0,1,0),(1,0,0),(\infty,\infty,\infty)\}\ a.s.
\end{eqnarray*}
\end{proposition}
\begin{proof} By \Cref{lbs}, we have $\mu$-a.s.\ $s_0,s_1,t\in\{0,1,\infty\}$. By \Cref{nzz,ozz,nzi,oiz}, we have $\mu$-a.s.\ $(s_0,s_1)\in\{(0,1),(1,0),(\infty,\infty)\}$. By \Cref{t01}, $\mu$-a.s.\ $t\in\{0,\infty\}$. Moreover, since each infinite contour must be a doubly infinite self-avoiding path, if $t=\infty$, then $s_0+s_1=\infty$. This implies that if $s_0+s_1=1$, then $t=0$, a.s. Moreover, if $s_0+s_1\geq 2$, then $t\geq1$. Then the proposition follows.
\end{proof}
\begin{example}\label{exl2}Consider the i.i.d.\ Bernoulli site percolation on the regular tiling $\LL_2$ of the hyperbolic plane with triangles, such that each vertex has degree $n\geq 7$. Assume that each vertex of $\LL_2$ takes value 1 with probability $\frac{1}{2}$. The corresponding contour configuration on the dual graph $\LL_1$ to the site percolation on $\LL_2$ induces a constrained configuration in the $[3,4,n,4]$ lattice satisfying (A8),(A2),(A3),(A7). Then by \Cref{m23} $\mu$-a.s. $(s_0,s_1)=(\infty,\infty)$. This implies that
\begin{eqnarray*}
\frac{1}{n-1}\leq p_c<\frac{1}{2}<p_u=1-p_c\leq \frac{n}{n-1}
\end{eqnarray*}
by Theorems 1.1, 1.2. and 1.3 of \cite{BS20}.
By \Cref{p118}, we have
\begin{itemize}
\item if $p\in [0,p_c]$, a.s. $(s_0,s_1,t)=(1,0,0)$;
\item if $p\in (p_c,p_u)$, a.s. $(s_0,s_1,t)=(\infty,\infty,\infty)$;
\item if $p\in[p_u,1]$, a.s. $(s_0,s_1,t)=(0,1,0)$.
\end{itemize}
\end{example}
\section{Proof of \Cref{ipl}}\label{pipl}
In this section, we prove \Cref{ipl}. We shall first review the stochastic domination we use to prove these results.
\begin{definition}(Stochastic Domination) Let $G=(V,E)$ be a graph. Let $\Omega=\{0,1\}^E$ (resp.\ $\Omega=\{0,1\}^V$). Then the configuration space $\Omega$ is a partially ordered set with partial order given by $\omega_1\leq \omega_2$ if $\omega_1(e)\leq \omega_2(e)$ for all $e\in E$ (resp.\ $\omega_1(v)\leq \omega_2(v)$ for all $v\in V$). A random variable $X:\Omega\rightarrow\RR$ is called increasing if $X(\omega_1)\leq X(\omega_2)$ whenever $\omega_1\leq \omega_2$. An event $A\subset\Omega$ is called increasing (respectively, decreasing) if its indicator function $1_A$ is increasing (respectively, decreasing). Given two probability measures $\mu_1$, $\mu_2$ on $\Omega$, we write $\mu_1\prec \mu_2$, and we say that $\mu_2$ \textbf{stochastically dominates} $\mu_1$, if $\mu_1(A)\leq \mu_2(A)$ for all increasing events $A\subset \Omega$.
\end{definition}
\begin{lemma}(Holley inequality)\label{hln} Let $G=(V,E)$ be a finite graph. Let $\Omega=\{0,1\}^E$ (resp.\ $\Omega=\{0,1\}^V$). Let $\mu_1$ and $\mu_2$ be strictly positive probability measures on $\Omega$ such that
\begin{eqnarray}
\mu_2(\max\{\omega_1,\omega_2\})\mu_1(\min(\omega_1,\omega_2))\geq \mu_1(\omega_1)\mu_2(\omega_2),\qquad \omega_1,\omega_2\in\Omega,\label{hli}
\end{eqnarray}
Then
\begin{eqnarray*}
\mu_1\prec \mu_2.
\end{eqnarray*}
\end{lemma}
\begin{proof}See Theorem (2.1) of \cite{GrGrc}; see also \cite{HR74}.
\end{proof}
\begin{lemma}\label{shln}Let $G=(V,E)$ be a finite graph. Let $\Omega=\{0,1\}^E$. Let $\mu_1$ and $\mu_2$ be strictly positive probability measures on $\Omega$ such that
\begin{eqnarray}
\mu_2(\omega\cup\{e\})\mu_1(\omega\setminus\{e\})\geq \mu_1(\omega\cup\{e\})\mu_2(\omega\setminus\{e\}),\qquad\omega\in \Omega, e\in E.\label{sc1}
\end{eqnarray}
Here $\omega$ is interpreted as the subset of $E$ consisting of all the edges with state 1. If either $\mu_1$ or $\mu_2$ satisfies
\begin{eqnarray}
\mu(\omega\cup\{e,f\})\mu(\omega\setminus\{e,f\})\geq \mu(\omega\cup{e}\setminus \{f\})\mu(\omega\cup\{f\}\setminus\{e\}),\label{sc2}
\end{eqnarray}
then (\ref{hli}) holds.
\end{lemma}
\begin{proof}See Theorem 2.6 of \cite{GrGrc}.
\end{proof}
For a planar graph $G$, let $G_*$ be its planar dual graph. The following lemmas concerning planar duality, are proved in \cite{BS20,HJL02}.
\begin{lemma}\label{pd1}Let $G$ be a planar nonamenable quasi-transitive graph, and let $p,p_*\in (0,1)$ satisfy
\begin{eqnarray}
p_*=\frac{(1-p)q}{p+(1-p)q}\label{pdp}
\end{eqnarray}
In the natural coupling of $FRC_{p,q}^G$ and $WRC_{p_*,q}^{G_*}$ as dual measures (i.e.\ a dual edge is present if and only if the corresponding primal edge is absent), the number of infinite clusters with respect to each is a.s.\ one of the following: $(0,1)$, $(1,0)$, $(\infty,\infty)$.
\end{lemma}
\begin{proof}See Proposition 3.5 of \cite{HJL02}, which is proved using the same technique as the proof of Theorem 3.7 of \cite{BS20}.
\end{proof}
\begin{lemma}\label{pd2}Let
\begin{eqnarray*}
h(x):=\frac{x}{1-x}.
\end{eqnarray*}
For any planar non-amenable quasi-transitive graph $G$,
\begin{eqnarray*}
&&h(p_{c,q}^w(G))h(p_{u,q}^f(G_*))=h(p_{u,q}^w(G_*))h(p_{c,q}^f(G))=1,\\
&&0<p_{c,q}^w(G)\leq p_{c,q}^f(G)<1,\ \mathrm{and}\ 0<p_{u,q}^w(G)\leq p_{u,q}^f(G)<1
\end{eqnarray*}
\end{lemma}
\begin{proof}See Corollary 3.6 of \cite{HJL02}.
\end{proof}
We start the proof of \Cref{ipl} with the following lemma.
\begin{lemma}\label{sz}Let $\LL_2$ be a vertex-transitive, triangular tiling of the hyperbolic plane such that each vertex has degree $n\geq 7$. Let $q\geq 1$.
Assume that $WRC^{\LL_2}_{p,q}$-a.s. there is a unique infinite open cluster for the random cluster model on $\LL_2$. Let $\tau$ be the unique infinite open cluster in the random cluster configuration $\omega$. We define a site percolation configuration $\xi$ on $V(\LL_2)$, by letting all the vertices in $\tau$ have state 1, and all the other vertices have state 0. Then a.s. $\xi$ has no infinite 0-cluster.
\end{lemma}
\begin{proof}Let $\sA_0$ be the event that $\xi$ has an infinite 0-cluster. By ergodicity of $WRC_{p,q}^{\LL_2}$, either $WRC_{p,q}^{\LL_2}(\sA_0)=0$ or $WRC_{p,q}^{\LL_2}(\sA_0)=1$. Assume that $WRC_{p,q}^{\LL_2}(\sA_0)=1$; we shall obtain a contradiction.
The dual configuration to the random cluster configuration on $\LL_2$ is a bond configuration on $\LL_1$ such that an edge in $\LL_1$ is present if and only if its dual edge in $\LL_2$ is absent in the random cluster configuration of $\LL_2$. Note also that the free boundary condition is dual to the wired boundary condition by the relation between dual configurations described above. Moreover, if the random cluster configuration on $\LL_2$ has distribution $WRC_{p,q}^{\LL_2}$, then its dual configuration on $\LL_1$ has distribution $FRC_{p_*,q}^{\LL_1}$; where $p,p_*$ satisfy (\ref{pdp}).
Let $k$ be the number of infinite clusters in the bond configuration in $\LL_2$, and let $k^{*}$ be the number of infinite clusters in the corresponding dual configuration in $\LL_1$.
By \Cref{pd1}, a.s.\ $(k,k^*)\in \{(0,1),(1,0),(\infty,\infty)\}$. (Indeed this is true for any $\mathrm{Aut}(\LL_2)$-invariant, insertion tolerant and deletion tolerant bond configuration.) Hence if $WRC_{p,q}^{\LL_2}$-a.s. there is a unique infinite open cluster, then $FRC_{p_*,q}^{\LL_1}$-a.s. there is no infinite cluster in the corresponding dual configuration.
Since $\tau$ is an infinite cluster, there exists an infinite 1-cluster in $\xi$ by construction. If there exists an infinite 0-cluster in $\xi$ as well, by \Cref{io}, there exists an infinite contour $C$ consisting of edges of $\LL_1$ in which each edge has a dual edge joining a vertex of $V(\LL_2)$ with state 1 in $\xi$ and a vertex of $V(\LL_2)$ with state 0 in $\xi$. Moreover, all the edges in $C$ must be present in the dual configuration of $\omega$, since every edge in $C$ is dual to an edge of $V(\LL_2)$ not open in $\omega$.
Then we have $k=1$ and $k^*\geq 1$. But this is a contradiction to \Cref{pd1}. Hence a.s.\ there are no infinite 0-clusters in $\xi$.
\end{proof}
\begin{lemma}\label{1l2}Let $\LL_2$ be a vertex-transitive, triangular tiling of the hyperbolic plane such that each vertex has degree $n\geq 7$.
Then for the graph $\LL_2$,
\begin{eqnarray*}
p_{u,1} \leq p_{u,2}^{w}\leq p_{u,2}^{f}
\end{eqnarray*}
\end{lemma}
\begin{proof}The fact that $p_{u,2}^w\leq p_{u,2}^f$ follows from \Cref{pd2}.
Now we prove that $p_{u,1}\leq p_{u,2}^w$. Note that the following stochastic monotonicity result holds:
\begin{eqnarray}
WRC_{p,2}\prec WRC_{p,1}=FRC_{p,1},\label{st21}
\end{eqnarray}
by (4.1) of \cite{RS01}.
Let $\sA_1$ be the event that there exists a unique infinite cluster in the random cluster configuration in $\LL_2$.
By ergodicity of $WRC_{p,2}$ and $WRC_{p,1}$ and the monotonicity of $WRC_{p,2}(\sA_1)$ and $WRC_{p,1}(\sA_1)$ with respect to $p$, to show that $p_{u,1}\leq p_{u,2}^{w}$, it suffices to show that whenever $WRC_{p,2}(\sA_1)=1$, then $WRC_{p,1}(\sA_1)=1$.
Let $\sA_{1,0}\subset \sA_1$ be the event that there exists a unique infinite cluster $\tau$ in the random cluster configuration on $\LL_2$; moreover, let $\xi\in \{0,1\}^{V(\LL_2)}$ be the site configuration obtained by assigning the state 1 to all the vertices in $\tau$, and the state 0 to all the vertices not in $\tau$; then there exists no infinite 0-cluster in $\xi$.
By \Cref{sz}, if $WRC_{p,2}(\sA_1)=1$, then $WRC_{p,2}(\sA_{1,0})=1$. Since $\sA_{1,0}$ is an increasing event, by (\ref{st21}) we have $WRC_{p,2}(\sA_{1,0})=1$, then $WRC_{p,1}(\sA_{1,0})=1$. Since $\sA_{1,0}\subset \sA_1$, we have $WRC_{p,1}(\sA_1)=1$. This completes the proof.
\end{proof}
\begin{lemma}\label{labc} Let $\LL_2$ be a vertex-transitive, triangular tiling of the hyperbolic plane such that each vertex has degree $n\geq 7$. Consider Conditions (a) (b) (c) (d) in \Cref{ipl} II. We have
\begin{eqnarray*}
(d)\Rightarrow(c)\Rightarrow(b)\Rightarrow (a);
\end{eqnarray*}
where $A\Rightarrow B$ means that if $A$ holds, then $B$ holds.
\end{lemma}
\begin{proof}The statement $(b)\Rightarrow (a)$ follows from Theorem 4.1 of \cite{RS01}.
The fact that $(c)\Rightarrow (a)$ follows from Theorem 3.2 (v) of \cite{HJL02}; while the fact that $(c)\Rightarrow (b)$ follows from Theorem 4.1 and Lemma 6.4 of \cite{LS99}.
The fact that $(d)\Rightarrow (c)$ follows from \Cref{1l2}.
\end{proof}
\subsection{Proof of Part I of \Cref{ipl}.}
First note that if (\ref{pch}) hold, then
\begin{eqnarray}
\frac{e^{h}}{e^{h}+e^{-h}}=p_u.\label{puh}
\end{eqnarray}
by the fact that $p_c+p_u=1$.
Let $\nu_1$ (resp.\ $\nu_2$) be the probability measure for the i.i.d.\ Bernoulli site percolation on $\LL_2$ in which each vertex takes the value ``$+$'' with probability $p_1$ (resp.\ $p_2$) satisfying
\begin{eqnarray*}
&&\frac{e^{nJ}}{e^{nJ}+e^{-nJ}}<p_1<p_u\\
&& p_c<p_2<\frac{e^{-nJ}}{e^{nJ}+e^{-nJ}}
\end{eqnarray*}
and the value ``$-$'' with probability $1-p_1$ (resp.\ $1-p_2$). Such $p_1$ and $p_2$ exist by (\ref{jh}).
Fix a triangle face $F_0$ of $\LL_2$. Let $B_R=(V(B_R),E(B_R))$ be the finite subgraph of $\LL_2$ consisting of all the faces of $\LL_2$ whose graph distance to $F_0$ is at most $R$. Let $\nu_{1,R}$ (resp.\ $\nu_{2,R}$) be the restriction of $\nu_1$ (resp.\ $\nu_2$) on $B_R$. Let $\mu_{R}^{+}$ (resp. $\mu_{R}^{-})$ be the probability measure for the Ising model on $B_R$ with respect to the coupling constant $J$ and the ``+'' boundary condition (resp.\ the ``$-$'' boundary condition). Let $\omega_1$, $\omega_2$ be two configurations in $\{-1,1\}^{V(B_R)}$. Then by \Cref{hln,shln}, we can check the F.K.G. lattice conditions below
\begin{eqnarray*}
\nu_{1,R}(\max\{\omega_1,\omega_2\})\mu_{R}^+(\min\{\omega_1,\omega_2\})\geq \nu_{1,R}(\omega_1)\mu_R^+(\omega_2)\\
\mu_{R}^-(\max\{\omega_1,\omega_2\})\nu_{2,R}(\min\{\omega_1,\omega_2\})\geq \mu_R^-(\omega_1)\nu_{2,R}(\omega_1).
\end{eqnarray*}
Then we obtain the following stochastic domination result:
\begin{eqnarray*}
\nu_{2,R}\prec\mu_R^{-}\prec \mu_R^+\prec\nu_{1,R}.
\end{eqnarray*}
Letting $R\rightarrow\infty$, for any $\mathrm{Aut}(\LL_2)$-invariant Gibbs measure $\mu$ for the Ising model on $\LL_2$ with coupling constant $J$, we have
\begin{eqnarray*}
\nu_2\prec \mu^{-}\prec \mu\prec \mu^+\prec\nu_{1}.
\end{eqnarray*}
Since $\nu_2$-a.s. there are infinite ``$+$''-clusters, $\mu$-a.s. there are infinite ``$+$''-clusters. Similarly, $\mu$-a.s. there are infinite ``$-$''-clusters, since $\nu_1$-a.s. there are infinite ``$-$''-clusters. By \Cref{p118}, we conclude that when (\ref{jh}) hold, $\mu$-a.s. there are infinitely many infinite ``$+$''-clusters, infinitely many infinite ``$-$''-clusters and infinitely many infinite contours. This completes the proof of Part I.
\subsection{Proof of Part II of \Cref{ipl}.}
We first assume that $\mu^f$ is $\mathrm{Aut}(\LL_2)$-ergodic.
Since $\mu^f$ is also symmetric in switching ``$+$'' and ``$-$'' states, $\mu^f$-a.s.\ the number of infinite ``$+$''-clusters and the number of infinite ``$-$''-clusters are equal. Then the conclusion under Condition (a) follows from \Cref{p118}. The conclusion holds under Conditions (b), (c) and (d) by \Cref{labc}.
\subsection{Proof of Part III of \Cref{ipl}.}
By Proposition 3.2 (i) of \cite{HJL02}, if (\ref{jj3}) holds, then there is a unique infinite-volume Gibbs measure. Since
\begin{eqnarray*}
p_{c,2}^{w}\leq p_{c,2}^{f}\leq p_{u,2}^{f},
\end{eqnarray*}
Part III now follows from Part II.
\subsection{Proof of Part IV of \Cref{ipl}.}
Since $p=1-e^{-2J}$, when (\ref{jj1}) holds, we have $p\geq p_{u,2}^w$. By Corollary 3.7 of \cite{HJL02} (see also (\ref{wrc})), there exists a unique infinite open cluster $\tau$ in the random cluster representation of the Ising model with wired boundary conditions $WRC_{p,2}$-a.s.
By the correspondence of random-cluster configurations and Ising configurations as described in \Cref{rcq}, each infinite cluster in the random cluster representation must be a subset of an infinite cluster in the Ising model. Hence if $WRC_{p,2}$-a.s. there is a unique infinite open cluster, $\mu^+$ a.s.\ there exists an infinite ``$+$''-cluster in the Ising model, and $\mu^{-}$-a.s.\ there exists an infinite ``$-$''-cluster in the Ising model.
Let $\tau$ be the unique infinite open cluster in the random cluster configuration $\omega$. We define a site percolation configuration $\xi$ on $V(\LL_2)$, by letting all the vertices in $\tau$ have state 1, and all the other vertices have state 0. By \Cref{sz}, a.s.\ $\xi$ has no-infinite 0-clusters. Again by the correspondence of the random cluster configuration and the Ising configuration and \cref{p118}, we obtain $\mu^{+}(\sA_+)=1$, and $\mu^{-}(\sA_-)=1$.
\subsection{Proof of Part V of \Cref{ipl}.}
Now we prove Part V. The identities (\ref{+1}) and (\ref{-1}) follows Part I and the fact that
\begin{eqnarray*}
p_{u,2}^{w}\leq p_{u,2}^f;
\end{eqnarray*}
and (\ref{hmf}) follows from the fact that when (\ref{jj2}) hold,
\begin{eqnarray*}
\mu^f=\frac{\mu^++\mu^-}{2};
\end{eqnarray*}
See Theorem 4.2 of \cite{RS01}, expressions (17) (18) of \cite{HJL02}.
\section{Proof of \Cref{coii,xorc}}\label{pxorc}
In this section, we prove \Cref{coii,xorc}.
\bigskip
\noindent\textbf{Proof of \cref{coii}.} We first prove Part I of the theorem. Let $s_{+}$ (resp.\ $s_{-}$) be the number of infinite ``$+$''-clusters (resp.\ infinite ``$-$''-clusters). By ergodicity, $\mathrm{Aut}(\LL_2)$-invariance and symmetry in ``$+$'' and ``$-$'' of $\mu_{1,f}\times \mu_{2,f}$, as well as \Cref{lbs}, one of the following cases occurs:
\begin{enumerate}[label=(\roman*)]
\item $\mu_1^f\times \mu_2^f((s_+,s_-)=(0,0))=1$; or
\item $\mu_1^f\times \mu_2^f((s_+,s_-)=(1,1))=1$; or
\item $\mu_1^f\times \mu_2^f((s_+,s_-)=(\infty,\infty))=1$
\end{enumerate}
Case (i) is impossible to occur by \Cref{nzz}. Case (ii) is impossible to occur by \Cref{ozz}. This completes the proof of Part I.
Now we show that Assumption II implies Assumption I. This follows from Theorem 4.1 of \cite{Sch99}, and the fact that the XOR Ising measure is the product measure of two i.i.d Ising models.
The fact that Assumption III implies Assumption II follows from \cref{labc}. This completes the proof of the theorem.
$\hfill\Box$
\bigskip
Before proving \Cref{xorc}, we shall first introduce the following definition and proposition in \cite{LS99} (see Theorem 3.3, Remark 3.4).
\begin{definition}\label{idd}Let $G=(V,E)$ be a graph and $\Gamma$ a transitive group acting on $G$. Suppose that $X$ is either $V$, $E$ or $V\cup E$. Let $Q$ be a measurable space and $\Omega:=2^V\times Q^X$. A probability measure $\mathbf{P}$ on $\Omega$ will be called a \textbf{site percolation with scenery} on $G$. The projection onto $2^V$ is the underlying percolation and the projection onto $Q^X$ is the scenery. If $(\omega,q)\in \Omega$, we set $\Pi_v(\omega,q)=(\Pi_v \omega,q)$. We say the percolation with scenery $\mathbf{P}$ is \textbf{insertion-tolerant} if $\mathbf{P}(\Pi_v\mathcal{B})>0$ for every measurable $\mathcal{B}\subset \Omega$ with positive measure. We say that $\mathbf{P}$ has \textbf{indistinguishable infinite clusters} if for every $\mathcal{A}\subset 2^V\times 2^V\times Q^X$ that is invariant under diagonal actions of $\Gamma$, for $\mathbf{P}$-a.e. $(\omega,q)$, either all infinite clusters $C$ of $\omega$ satisfy $(C,\omega,q)\in \mathcal{A}$, or they all satisfy $(C,\omega,q)\notin \mathcal{A}$.
\end{definition}
\begin{proposition}\label{idc}Let $\mathbf{P}$ be a site percolation with scenery on a graph $G=(V,E)$ with state space $\Omega:=2^V\times Q^X$, where $Q$ is a measurable space and $X$ is either $V$, $E$ or $V\cup E$. If $\mathbf{P}$ is $\Gamma$-invariant and insertion tolerant, then $\mathbf{P}$ has indistinguishable infinite clusters.
\end{proposition}
\noindent\textbf{Proof of \Cref{xorc}} Let $\Lambda=(V_{\Lambda},E_{\Lambda})$ be a subgraph of $\LL_2$ consisting of faces of $\LL_2$. Let $\Lambda_{*}=(V_{\Lambda_*},E_{\Lambda_*})$ be the dual graph of $\Lambda$, such that there is a vertex in $V_{\Lambda}$ corresponding to each triangle face in $\Lambda$, as well as the unbounded face; the edges in $E_{\Lambda}$ and $E_{\Lambda_*}$ are in 1-1 correspondence by duality.
Consider an XOR Ising model on $\Lambda$ with respect to two i.i.d. Ising models $\sigma_3$, $\sigma_4$ with free boundary conditions and coupling constants $J\geq 0$ satisfying the assumption of \Cref{coii}. The partition function of the XOR Ising model can be computed by
\begin{eqnarray*}
Z_{\Lambda,f}=\sum_{\sigma_3,\sigma_4\in \{\pm 1\}^{V_{\Lambda}}}\prod_{(u,v)\in E_{\Lambda}}e^{J(\sigma_{3,u}\sigma_{3,v}+\sigma_{4,u}\sigma_{4,v})}.
\end{eqnarray*}
Following the same computations as in \cite{bd14}, we obtain
\begin{eqnarray}
Z_{\Lambda,f}= C_1\sum_{P_*\in \mathcal{P}_{*},P\in \mathcal{P},P\cap P_*=\emptyset}\left(\frac{2 e^{-2J}}{1+e^{-4J}}\right)^{|P_*|}\left(\frac{1-e^{-4J}}{1+e^{-4J}}\right)^{|P|}.\label{zlf}
\end{eqnarray}
where $\mathcal{P}_*$ (resp.\ $\mathcal{P}$) consists of all the contour configurations on $E_{\Lambda_*}$ (resp.\ $E_{\Lambda}$) such that each vertex of $V_{\Lambda_*}$ (resp.\ $V_{\Lambda}$) has an even number of incident present edges, and $C_1=2^{|V_{\Lambda}|-|E_{\Lambda}|+2}(e^{2j}-e^{-2J})^{|E_{\Lambda}|}$ is a constant.
When $J,K$ satisfies (\ref{jkr}), we have
\begin{eqnarray*}
\frac{2e^{-2J}}{1+e^{-4J}}&=&\frac{1-e^{-4K}}{1+e^{-4K}};\\
\frac{2e^{-2K}}{1+e^{-4K}}&=&\frac{1-e^{-4J}}{1+e^{-4J}}.
\end{eqnarray*}
Thus the partition function $Z_{\Lambda,f}$, up to a multiplicative constant, is the same as the partition function of the XOR Ising model on $\Lambda_*$ with coupling constant $K$.
Recall that there is exactly one vertex $v_{\infty}\in V_{\Lambda_{*}}$ corresponding to the unbounded face in $\Lambda$. The XOR Ising model $\sigma_{XOR}=\sigma_1\sigma_2$ on $\Lambda_*$, corresponds to an XOR Ising model on $\Lambda_{*}\setminus \{v_{\infty}\}$ (which is a subgraph of $\LL_1$) with the boundary condition that all the boundary vertices have the same state in $\sigma_1$ and all the boundary vertices have the same state in $\sigma_2$. Hence the boundary condition must be a mixture of $++$, $+-$, $-+$ and $--$. However, each one of the 4 possible boundary conditions gives the same distribution of contours in the XOR Ising model. From the expression (\ref{zlf}), we can see that there is a natural probability measure on the set of contours $\Phi=\{(P,P_*):P\in\mathcal{P},P_*\in \mathcal{P}_*,P\cap P_*=\emptyset\}$, such that the probability of each pair of contours $(P,P_*)\in \Phi$ is proportional to $\left(\frac{2 e^{-2J}}{1+e^{-4J}}\right)^{|P_*|}\left(\frac{1-e^{-4J}}{1+e^{-4J}}\right)^{|P|}$, and the marginal distribution on $\mathcal{P}$ is the distribution of contours for the XOR Ising model on $\LL_2$ with coupling constant $J$ and free boundary conditions, while the marginal distribution on $\mathcal{P}_*$ is the distribution of contours for the XOR Ising model on $\LL_1$ with coupling constant $K$ and $++$ boundary conditions.
We let $\Lambda$ and $\Lambda_*\setminus \{v_{\infty}\}$ increase and approximate the graph $\LL_2$ and $\LL_1$, respectively. If with a positive $\mu_{++}$ probability, there exists exactly one infinite contour $C$ in $\LL_2$, then $\mu_f$-a.s. there exists an infinite cluster in $\LL_2$ containing $C$, since contours in $\LL_1$ and $\LL_2$ are disjoint. Consider the XOR Ising spin configuration as a site percolation on $\LL_2$, with scenery given by contour configurations of $\LL_2$ within the ``$+$'' clusters of $\LL_2$. In the notation of \Cref{idd}, $Q=\{0,1\}$, and $X=E(\LL_2)$. An edge in $E(\LL_2)$ is present (has state ``1'') if and only if it is in a ``$+$''-cluster of the XOR Ising configuration on $\LL_2$ and present in the contour configuration of $\LL_2$. This way we obtain an automorphism-invariant and insertion-tolerant percolation with scenery. Let $\mathcal{A}\subset 2^V(\LL_2)\times 2^V(\LL_2)\times 2^E(\LL_2)$ be the triple $(C,\omega, q)$ such that
\begin{itemize}
\item $\omega$ is an XOR Ising spin configuration on $\LL_2$; and
\item $C$ is an infinite ``$+$''-cluster; and
\item $q$ is the $\LL_2$-contour configuration within ``$+$''-clusters of $\omega$; and
\item $C$ contains an infinite contour in $q$.
\end{itemize}
Obviously $\mathcal{A}$ is invariant under diagonal actions of automorphisms.
By \Cref{coii}, $\mu_f$-a.s. there exists infinitely many infinite ``$+$''-clusters in $\LL_2$. By \Cref{idc}, either all the infinite clusters are in $\mathcal{A}$, or no infinite clusters are in $\mathcal{A}$. Similar arguments applies for ``$-$''-clusters of $\LL_2$. Hence almost surely the number of infinite contours in $\LL_2$ is 0 or $\infty$. Since the distribution of infinite contours in $\LL_2$ is exactly that of contours for the XOR Ising model on $\LL_1$ with coupling constant $K$ and $++$ (or $+-$, $-+$, $--$) boundary condition, the theorem follows.
$\hfill\Box$
\section{Proof of \Cref{m31}}\label{pm31}
In this section, we prove \Cref{m31}.
Let $\widetilde{\mu}_D$ be the marginal distribution of $\mu_D$ restricted on Type-II edges. Recall that the restriction of dimer configurations to Type-II edges on the [4,6,12] lattice, are in 1-1 correspondence with constrained percolation configurations on the $[3,4,6,4]$ lattice in $\Omega$, as described in \Cref{db}. See also \Cref{4612}.
Also recall that $\widetilde{\mu}_D$ is the weak limit of measures on larger and larger tori; since the edge weights are translation-invariant, the measures on tori are translation-invariant. Hence $\widetilde{\mu}_D$ satisfies (A1) if edge weights satisfies (B1).
The measure $\widetilde{\mu}_D$ is both translation-invariant and mixing (see \cite{KOS06}), hence $\widetilde{\mu}_D$ is totally ergodic and satisfies (A2).
If the edge weights satisfy (B2) and (B3), the measures on tori are symmetric under $\theta$. Hence $\widetilde{\mu}_D$ satisfies (A3).
By the results in \cite{bd14}, the marginal distribution of contours in $\HH$ (resp.\ $\TT$) under $\widetilde{\mu}_D$ is the same as the distribution of contours of an XOR Ising model $\sigma_{XOR,\TT}$ (resp. $\sigma_{XOR,\HH}$) with spins located on vertices of $\TT$ (resp.\ $\HH$), if the dimer edge weights and Ising coupling constants satisfy the following conditions:
\begin{itemize}
\item each Type-II edge has weight 1;
\item each Type-I edge parallel to an edge of $e$ with coupling constants $J_e$ has weight $w_e$ such that $w_e=\frac{1-e^{-4J_e}}{1+e^{-4J_e}}$;
\item each Type-I edge perpendicular to an edge $e$ with coupling constants $J_e$ has weight $w_e$ such that $w_e=\frac{2e^{-2J_e}}{1+e^{-4J_e}}$.
\end{itemize}
Moreover, when the edge weights satisfy (B1)-(B3), $\sigma_{XOR,\TT}$ and $\sigma_{XOR,\HH}$ are dual to each other, i.e.\ the coupling constants $J_{\tau}$ and $K_{\tau}$ on a pair of dual edges $e\in \HH,e^*\in \TT$ satisfy (\ref{dua}). The finite energy assumptions (A4) and (A5) follows from the finite energy of $\sigma_{XOR,\HH}$ and $\sigma_{XOR,\TT}$.
Since $\sigma_{XOR,\HH}$ and $\sigma_{XOR,\TT}$ are dual to each other, one of the following cases might occur
\begin{enumerate}
\item $\sigma_{XOR,\HH}$ is in the low-temperature state, and $\sigma_{XOR,\TT}$ is in the high-temperature state;
\item $\sigma_{XOR,\HH}$ is in the high-temperature state, and $\sigma_{XOR,\TT}$ is in the low-temperature state;
\item both $\sigma_{XOR,\HH}$ and $\sigma_{XOR,\TT}$ are in the critical state.
\end{enumerate}
See \Cref{xori} for definitions of the low-temperature state, high-temperature state and critical state for XOR Ising models.
In Case III, both (A6) and (A7) are satisfied because of the ergodicity of measures for the critical XOR Ising model on $\HH$ and $\TT$; see \Cref{ec}. Moreover, Case III occurs if and only if the edge weights satisfy (B4). Hence when the edge weights satisfy (B1)-(B4), $\widetilde{\mu}_D$ satisfy (A1)-(A7). \Cref{m31} I follows from \Cref{m21}II(c).
Note that the measures for the high-temperature XOR Ising models on $\HH$ and $\TT$ are also ergodic; see \Cref{he}. Therefore, in each case of I, II and III, at least one of (A6) and (A7) is satisfied. Then \Cref{m31} II follows from \Cref{m21} II(a)(b) and \Cref{l81,l82}.
\section{Proof of \Cref{chi,lth}}\label{p412}
In this section, we prove \Cref{chi,lth}.
Consider the XOR Ising model with spins located on vertices of the hexagonal lattice $\HH$ with coupling constants $J_a$, $J_b$, $J_c$ on horizontal, NW/SE, NE/SW edges, see \Cref{htst} for an embedding of $\HH$ into the plane such that all the edges are either horizontal, NW/SE or NE/SW.
We construct a (4,6,12) lattice $\mathbb{A}$ as in \Cref{4612}. Each Type-II edge is assigned weight 1. Each Type-I edge parallel to an edge $e$ of $\HH$ is assigned weight $\frac{1-e^{-4J_e}}{1+e^{-4J_e}}$. Each Type-I edge perpendicular to an edge $e$ of $\HH$ is assigned weight $\frac{2e^{-2J_e}}{1+e^{-4J_e}}$. Here $J_e$ is the coupling constant on the edge $e$.
In order to prove the \Cref{chi,lth}, we prove the following lemmas.
\begin{lemma}\label{ec}The measure for the critical XOR Ising model on $\HH$ (resp.\ $\TT$), obtained as the weak limit of measures on tori, is ergodic.
\end{lemma}
\begin{proof}Let $\rho=\frac{\sigma+1}{2}$, where $\sigma:V_{\HH}\rightarrow\{\pm 1\}$ is the spin configuration for an Ising model on $\HH$.
Following the same arguments as in the proof of Lemma 10.2 in \cite{HL16}, it suffices to show that for the critical Ising model on $\HH$ (resp.\ $\TT$), we have
\begin{eqnarray*}
\lim_{|u-v|\rightarrow\infty} |\langle\rho(u)\rho(v) \rangle-\langle\rho(u)\rangle\langle\rho(v)\rangle|=0,
\end{eqnarray*}
which is equivalent to show that
\begin{eqnarray}
\lim_{|u-v|\rightarrow\infty}\langle \sigma(u)\sigma(v) \rangle=0.\label{vs}
\end{eqnarray}
Note that $\langle\sigma(u)\sigma(v)\rangle$ is an even spin correlation function (i.e.\ the expectation of the product of spins on an even number of vertices), and hence for all the infinite-volume, translation-invariant Gibbs measures of the Ising model on $\HH$ (resp.\ $\TT$) corresponding to the given coupling constant, $\langle\sigma_u\sigma_v\rangle$ has a unique value; see \cite{Leb77}.
Consider the FK random cluster representation of the Ising model with $q=2$, the two-point spin correlation $\langle\sigma(u)\sigma(v)\rangle$ is exactly the connectivity probability of $u$ and $v$ in the random cluster model, up to a multiplicative constant; see Chapter 1.4 of \cite{GrGrc}. Therefore in order to show (\ref{vs}), it suffices to show that the connectivity probabilities of two vertices in the corresponding random cluster model, as the distances of the two vertices go to infinity, converge to zero.
By Theorem 4 of \cite{BR07}, we infer that the connectivity probabilities of two vertices in the $q=2$ random cluster model corresponding to the critical Ising model on the triangular lattices with coupling constants $K_a,K_b,K_c$ satisfying $g(K_a,K_b,K_c)=0$ converge to zero as the distances of the two vertices go to infinity.
Note that the hexagonal lattice is a bipartite graph, i.e., all the vertices can be colored black and white such that vertices of the same color can never be adjacent.
Recall that the star-triangle transformation is a replacement of each black vertex of $\HH$, as well as its incident edges, into a triangle. The resulting graph is a triangular lattice $\TT'$; see the right graph of \Cref{htst}. The parameters of the random cluster model on $\HH$ and the random cluster model $\TT'$ satisfy certain identities, such that the probabilities of connections of any two adjacent vertices in $\TT'$ (which are also vertices in $\HH$) internal to each triangle face in $\TT'$ which has a black vertex of $\HH$ in the center are the same for the random cluster model on $\HH$ and the random cluster model on $\TT'$; see Page 160-161 of \cite{GrGrc}.
Using a star-triangle transformation (see the right graph of \Cref{htst}), and (6.69) of \cite{GrGrc}, we deduce that the connectivity probabilities of two vertices in the $q=2$ random cluster model on the hexagonal lattice corresponding to the critical Ising model on $\HH$ with coupling constants $J_a,J_b,J_c$ satisfying $f(J_a,J_b,J_c)=0$ converge to zero as the distances of the two vertices go to infinity. Note that the weak limit of of measures with free boundary conditions is known to exist and translation-invariant, see Theorem (4.19) of \cite{GrGrc}. By the uniqueness of $\langle \sigma(u)\sigma(v) \rangle$ under all the translation-invariant measures, we obtain that (\ref{vs}) holds under the measure obtained as the weak limit of measures with periodic boundary conditions.
\end{proof}
\begin{figure}
\centering
\subfloat[$\HH$ and $\TT$]{\includegraphics[width=.45\textwidth]{HHTT}}\qquad
\subfloat[$\TT$ and $\HH'$]{\includegraphics[width = .45\textwidth]{TTst}}
\caption{Hexagonal lattice $\HH$ (represented by black lines), dual triangular lattice $\TT$ (represented by red lines) and hexagonal lattice $\HH'$ obtained from the star-triangle transformation (represented by blue lines).}
\label{htst}
\end{figure}
\begin{lemma}\label{he}The measure for the high-temperature XOR Ising model on $\HH$ (resp.\ $\TT$), obtained as the weak limit of measures on tori, is ergodic.
\end{lemma}
\begin{proof}The identity (\ref{vs}) holds under the measure for the high-temperature Ising model on $\HH$ (resp.\ $\TT$); see \cite{ZL12,DC13}.
\end{proof}
\noindent\textbf{Proof of \Cref{chi}.} We first show that in the critical XOR Ising model on $\HH$ or $\TT$, almost surely there are no infinite contours. It is proved in \cite{bd14} that the contours of XOR Ising model with spins located on $\HH$ (resp.\ $\TT$) have the same distribution as contours in $\TT$ (resp.\ $\HH$) for the Type-II clusters of dimer configurations on the $[4,6,12]$ lattice, if the coupling constants of the XOR Ising model and the edge weights of the $[4,6,12]$ lattice satisfy the condition as described in \Cref{pm31}. It is not hard to check that when the coupling constants of the XOR Ising model on $\TT$ (resp.\ $\HH$) are critical, then the edge weights of the corresponding dimer model on the [4,6,12] lattice satisfy (B1)-(B4). By \Cref{m31} I, almost surely there are no infinite contours in the critical XOR Ising model on $\TT$ or $\HH$.
Now we prove that almost surely there are no infinite clusters for the critical XOR Ising model on $\HH$ or $\TT$. We write down the proof for the critical XOR Ising model on $\HH$ here, the case for the XOR Ising model on $\TT$ can be proved in a similar way.
Let $\sA$ be the event that there exists an infinite cluster for the XOR Ising model on $\HH$. Assume that $\mu(\sA)>0$; we will obtain a contradiction. By translation-invariance of $\sA$ and \Cref{ec}, if $\mu(\sA)>0$ then $\mu(\sA)=1$. Let $\sA_1$ (resp. $\sA_2$) be the event that there exists an infinite ``$+$''-cluster (resp.\ ``$-$''-cluster)
for the XOR Ising model on $\HH$, then
\begin{eqnarray}
\mu(\sA_1\cup\sA_2)=1.\label{u12}
\end{eqnarray}
By symmetry $\mu(\sA_1)=\mu(\sA_2)$. By translation-invariance of $\sA_1$, $\sA_2$ and \Cref{ec}, either $\mu(\sA_1)=\mu(\sA_2)=1$, or $\mu(\sA_1)=\mu(\sA_2)=0$. By (\ref{u12}), we have $\mu(\sA_1)=\mu(\sA_2)=1$, hence $\mu(\sA_1\cap\sA_2)=1$, i.e. $\mu$-a.s.\ there exist both an infinite ``$+$''-cluster and an infinite ``$-$''-cluster in the critical XOR Ising configuration on $\HH$.
Let $\phi_{\TT}$ be the contour configuration associated to the critical XOR Ising configuration on $\HH$. Let $\omega\in \phi^{-1}(\phi_{\TT})$ be a constrained percolation configuration on the $[3,4,6,4]$ lattice whose contour configuration is $\phi_{\TT}$. It is not hard to check that in $\omega$ there exist both an infinite 1-cluster and an infinite 0-cluster if in the original XOR Ising model on $\HH$, there exist both an infinite ``$+$''-cluster and an infinite ``$-$''-cluster. By \Cref{io}, $\mu$-a.s.\ there exists an infinite contour in $\phi_{\TT}$. The contradiction implies that $\mu$-a.s. there are no infinite clusters in the critical XOR Ising model on $\HH$.
$\hfill\Box$
\bigskip
\noindent\textbf{Proof of \Cref{lth}.} By the correspondence betwween contours in an XOR Ising model with spins located on vertices of $\HH$ (resp.\ $\TT$) and contours on $\TT$ (resp.\ $\HH$) for the Type-II clusters of dimer configurations on the $[4,6,12]$ lattice, as proved in \cite{bd14}, as well as correspondence between Type-II clusters of dimer configurations on the [4,6,12] lattice and clusters of constrained configurations on the $[3,4,6,4]$ lattice, it suffices to show that the measure and \Cref{m21}, it suffices to show that the probability measure for the low-temperature XOR Ising model on $\HH$ (resp.\ $\TT$) satisfies (A1)-(A5) and (A7) (resp.\ (A1)-(A6)), for $\LL_1=\HH$.
It is straightforward to verify (A1)-(A5). The assumption (A6) (resp.\ (A7)) follows from \Cref{he}.
$\hfill\Box$
\bigskip
\noindent\textbf{Acknowledgements.} ZL thanks Alexander Holroyd for stimulating discussions in the preparation of the paper, and Geoffrey Grimmett and Russ Lyons for comments. ZL's research is supported by Simons Foundation grant 351813 and National Science Foundation grant 1608896.
|
\section{Introduction}
At low mass accretion rates $\dot{m} \equiv \dot{M}/\dot{M}_{\mathrm{Edd}}\footnote{$\dot{M}$ is the accretion rate. $\dot{M}_{\mathrm{Edd}}\equiv 4 \pi G M m_p / \eta \sigma_T c$, where $M$ is the black hole mass and $\eta$ is the nominal efficiency. We adopt $\eta=0.1$; $\dot{M}_{\mathrm{Edd}} = 1.4\times10^{18}(M/M_{\odot})~\mathrm{g/s} = 2.2\times10^{-8}(M/M_{\odot})~M_{\odot}/\mathrm{yr}$.}\lesssim 10^{-3}$, thermally stable accretion onto black holes {is generally believed to form} a geometrically thick, optically thin, radiatively inefficient accretion flow (RIAF, or ADAF\footnote{Advection-Dominated Accretion Flow}; \citealt{Ichimaru1977}, \citealt{NarayanYi1994}, \citealt{YuanNarayan2014}).
Due in part to the two-temperature nature (e.g.\ \citealt{ShapiroLightmanEardley1976}, \citealt{MahadevanQuataert1997}, \citealt{Ressler2015}) of such flows, RIAFs are nearly virial and the liberated gravitational energy is {either} advected across the event horizon or lost through mechanical outflows. Such accretion flows are probably well-represented across the range of astrophysical black hole masses (\citealt{McClintockRemillard2006}, \citealt{Ho2009}).
Analytic and semi-analytic RIAF models have been profitably applied in the study of low-luminosity accretion flows (e.g.\ \citealt{NarayanBarretMcClintock1997}, \citealt{NarayanMahadevanEtAl1998}). However, a limitation of such studies is the reliance on an $\alpha$ viscosity (\citealt{ShakuraSunyaev1973}) to represent angular momentum transport, probably due to magnetohydrodynamic (MHD) turbulence generated by the magnetorotational instability (MRI; \citealt{BalbusHawley1991}).
Additionally, analytic models typically neglect or approximate general relativity, with potential consequences for interpreting observations as much of a RIAF's outgoing radiation may originate near the black hole ({e.g.\ }\citealt{MoscibrodzkaEtAl2009}).
Global general relativistic numerical simulations have been widely used to study RIAFs driven self-consistently by magnetorotational turbulence (e.g.\ \citealt{KoideShibataKudoh1999}, \citealt{DeVilliersHawleyKrolik2003}, \citealt{McKinneyGammie2004}, \citealt{NarayanEtAl2012}). In the absence of significant mean fields and cooling, such calculations generically recover the hot, nearly Keplerian, nearly axisymmetric (but see \citealt{FragileEtAl2007}) accretion disk {anticipated by analytic models}. {Nonetheless, electron thermodynamics in such calculations has remained a challenge. These flows are collisionless and likely two-temperature (\citealt{Quataert1998}). Historically, constant proton to electron temperature ratios, or other local prescriptions mapping the fluid state to electron temperature (e.g.\ \citealt{MoscibrodzkaEtAl2009}, \citealt{Shcherbakov+2012}, \citealt{Moscibrodzka2014}, \citealt{Chan+2015})} have been employed. Recently, however, \cite{Ressler2015} introduced a method to track numerical dissipation in conservative relativistic MHD schemes, interpret it according to local kinetic plasma studies, and thus separately evolve the electron temperature {(see also \citealt{Sadowski2017} for a similar method). While this provides for physically motivated electron heating, it still assumes a thermal distribution of electrons, whereas these collisionless flows may have a significant population of nonthermal electrons (e.g. \citealt{Kunz2016}, \citealt{Chael2017})}.
Radiative losses are negligible at sufficiently low accretion rates. Towards the Eddington rate, however, radiative processes become important to the dynamics of the flow. Significantly below Eddington, the flow is still optically thin and the electrons are relativistic near the black hole. The dominant energy loss mechanisms are synchrotron emission and Compton upscattering. \cite{Ohsuga2009} first demonstrated that radiation leads to thick/thin disk transitions in numerical models. \cite{FragileMeier2009} found a cooling state inconsistent with either a pure RIAF or a thin disk, and compared {it} to a magnetically-dominated accretion flow in the inner disk. \cite{Moscibrodzka2011} studied accretion rates targeting the supermassive black hole at the center of M87. \cite{Dibi2012} identified $\dot{m} \approx 10^{-7}$ as a critical accretion rate above which radiative losses matter in GRMHD simulations. \cite{Wu2016} targeted the near-Eddington state transition in X-ray binaries in Newtonian MHD with local cooling. Recently, \cite{Sadowski2017} addressed cooling in RIAFs with self-consistent electron heating and a gray M1 radiation closure, while \cite{SadowskiGaspari2017} use a similar model except with constant proton-to-electron temperature ratios to study the {transition to radiatively efficient flows}.
These studies integrate over frequency and adopt a local cooling function or approximate the radiation as a fluid.
In this paper we do not use either of these approximations. Instead, we introduce a scheme that couples a global, albeit axisymmetric, model with electron heating (\citealt{Ressler2015}) for the flow to a Monte Carlo radiation MHD scheme (\citealt{Ryan2015}), yielding a frequency-dependent, full transport solution to the equations of two-temperature relativistic radiation MHD.
We apply this new scheme, {\tt ebhlight}{}, to RIAFs across a range of accretion rates. Section \ref{sec:eqns} presents the governing equations. Section \ref{sec:numerics} describes our numerical implementation. Section \ref{sec:results} {contains} our {results}. Section \ref{sec:conclusion} concludes.
\section{Governing Equations}
\label{sec:eqns}
We solve the equations of general relativistic radiation ideal magnetohydrodynamics with full radiation transport. We include a separate electron energy equation (\citealt{Ressler2015}) and electron-photon interactions. {Hereafter, we adopt units such that $c = k_B = 1$ and absorb a factor $\sqrt{4 \pi}$ into definitions of magnetic field strength}.
The radiation and fluid are coupled through exchange of four-momentum. The electron energy density is sourced by numerical dissipation, and electrons and protons exchange energy through Coulomb interactions, as in \cite{Sadowski2017}, allowing transfer of energy between protons and electrons according to the transrelativistic rate of \cite{StepneyGuilbert1983}. Although we track electron and proton temperatures separately, we assume a single four-velocity for the fluid dynamics (\citealt{Ressler2015}).
The dynamical variables in our model are the fluid rest-mass density $\rho_0$, the fluid four-velocity $u^{\mu}$, the fluid internal energy $u$ (equivalently, the fluid pressure $P = (\gamma - 1)u$), the magnetic field three-vector $B^i$,
$\kappa_e \equiv \exp((\gamma_e-1)s_e) = P_e/\rho_0^{\gamma_e}$ ($s_e \equiv$ electron entropy),
and the radiation specific intensity $I_{\nu}$. We adopt three adiabatic indices: $\gamma_e = 4/3$ for the (relativistic) electrons, $\gamma_p = 5/3$ for the (non-relativistic) protons, and $\gamma = 13/9$ for the total fluid. Although our approximation of three constant $\gamma$ is likely not valid everywhere, previous studies (\citealt{Shiokawa2012}, \citealt{Sadowski2017}) suggest that variable $\gamma$ do not significantly alter conclusions drawn from numerical (GRMHD, GRRMHD) calculations.
Our full set of governing equations is (written in a coordinate basis):
\begin{align}
\partial_t \left( \sqrt{-g} \rho_0 u^t \right) &= -\partial_i \left( \sqrt{-g} \rho_0 u^i \right), \label{eqn:massConservation}\\
\begin{split}\label{eqn:stressEnergyConservation}
\partial_t \left( \sqrt{-g} T^t_{~\nu} \right) &={} \partial_i \left( \sqrt{-g} T^i_{~\nu} \right) + \sqrt{-g}T^{\kappa}_{~\lambda} \Gamma^{\lambda}_{~\nu\kappa}\\
& \quad - \sqrt{-g}R^{\mu}_{~\nu;\mu},
\end{split}\\
\partial_t \left( \sqrt{-g} B^i \right) &= \partial_j \left[ \sqrt{-g} \left( b^j u^i - b^i u^j \right) \right], \label{eqn:fluxConservation} \\
\partial_i \left( \sqrt{-g} B^i \right) &= 0, \label{eqn:monopoleConstraint} \\
\frac{dx^{\mu}}{d\lambda} &= k^{\mu}, \label{eqn:photonMotion} \\
\frac{dk^{\mu}}{d\lambda} &= -\Gamma^{\lambda}_{~\mu\nu}k^{\mu}k^{\nu}, \label{eqn:geodesic} \\
\frac{D}{d\lambda}\left(\frac{I_{\nu}}{\nu^3}\right) &= \frac{\eta_{\nu}(T_e)}{\nu^2} -\frac{I_{\nu}\chi_{\nu}(T_e)}{\nu^2}, \label{eqn:radiativeTransfer} \\
\frac{\rho^{\gamma_e}}{\gamma_e - 1} u^{\mu} \partial_{\mu} \kappa_e &= f_e Q_H + Q_C(T_e, T_p) - u^{\nu}R^{\mu}_{~\nu;\mu}, \label{eqn:electronEntropy}
\end{align}
where $D/d\lambda$ is the convective derivative in phase space, the GRMHD stress-energy tensor
\begin{align}
\begin{split}
T^{\mu}_{~\nu} &= \left( \rho_0 + u + P + b^{\lambda}b_{\lambda}\right)u^{\mu}u_{\nu} \\
&\quad + \left(P + \frac{b^{\lambda}b_{\lambda}}{2} \right)g^{\mu}_{~\nu} - b^{\mu}b_{\nu}
\end{split}
\end{align}
with $b^{\mu}$ the magnetic field four-vector (see \citealt{Gammie2003}), and the radiation stress-energy tensor
\begin{align}
R^{\mu}_{~\nu} = \int \frac{d^3 p}{\sqrt{-g}p^t}p^{\mu}p_{\nu} \left(\frac{I_{\nu}}{h^4 \nu^3}\right).
\end{align}
$Q_H$ and $Q_C$ are, respectively, dissipative and Coulomb volumetric heating rates. Temperature dependencies of interaction terms are shown for clarity. Note that $T_e$ is calculated not from $P$ and $\rho_0$ as in \cite{Ryan2015}, but rather from $\rho_0$ and $\kappa_e$ as $T_e = \rho_0^{\gamma_e-1} \kappa_e$. {$T_p = (\gamma_p - 1)(u - u_e)/\rho$ is the proton temperature, only needed for Coulomb coupling.}
For $T_e = (\gamma_e-1)u_e/\rho_0$, $\Theta_e = m_p T_e / m_e$ where $\Theta_e \equiv$ electron temperature in units of $m_e c^2$.
Note that the radiation four-force $R^{\mu}_{~\nu;\mu}$ is applied to both the electron and total energy equations; $T^{\mu}_{~\nu}$ incorporates both electrons and protons.
We consider synchrotron emission and absorption. We also include Compton scattering, which for $\Theta_e \gg 1$ and $h \nu \ll k_b T_e$ has a mean amplification factor $\delta E_{\gamma} / E_{\gamma} \approx 16 \Theta_e^2$.
\section{Numerical Method}
\label{sec:numerics}
Our calculations are performed with {\tt ebhlight}{}, an extension of {\tt bhlight}{} (\citealt{Ryan2015}) that includes the electron heating model of \cite{Ressler2015}. {\tt ebhlight}{} solves the equations of GRMHD (Equations \ref{eqn:massConservation} - \ref{eqn:monopoleConstraint}) with the flux-conservative second-order-accurate {\tt harm}{} scheme (\citealt{Gammie2003}). The radiative transfer and photon-electron interactions (Equations \ref{eqn:photonMotion} - \ref{eqn:radiativeTransfer}) are evaluated with the Monte Carlo scheme {\tt grmonty}{} (\citealt{Dolence2009}; we term radiation samples ``superphotons''). The electron heating (Equation \ref{eqn:electronEntropy}) is evaluated as in \cite{Ressler2015}, with Coulomb heating introduced in a separate explicit second-order step.
We neglect electron and ion conduction, as RIAF simulations have found both to be suppressed by misaligned magnetic fields and temperature gradients (\citealt{Ressler2015}, \citealt{Foucart2017}).
The radiation four-force is evaluated with time-centered fluid quantities, and applied to the total fluid and the electron energy in a first-order operator-split fashion. Emission, absorption, and scattering are treated probabilistically as in \cite{Ryan2015}.
\subsection{Coordinates}
We perform our calculation in horizon-penetrating Modified Kerr-Schild (MKS) coordinates (\citealt{McKinneyGammie2004}). The inner {boundary} is placed inside the event horizon, the outer {boundary at} $r = 200 GM/c^3$. The MKS $h$ parameter is $0.3$. To avoid wasting computational resources advancing many superphotons in the outer region where radiative interactions are relatively unimportant ($\Theta_e \lesssim 1$), we evaluate the radiation sector only inside a smaller outer radius, either $40$ or $100 GM/c^2$, as required to capture at least 95\% of the bolometric luminosity. We employ a spatial resolution $388\times256$ zones.
\subsection{Initial Conditions}
{\tt ebhlight}{} is {currently} axisymmetric; the useful time integration window is thus of the order $t\sim1000 GM/c^3$, after which MRI turbulence decays (see \citealt{GuanGammie2007} for details of MRI-driven turbulence in axisymmetry).
The timescale for viscous electron heating to equilibrate is longer than this beyond $10-15 GM/c^2$.
To address this issue, we initialize our simulation with axisymmetrized data from a 3D {\it nonradiative} GRMHD run with electron heating using the method described in \citealt{Ressler2016}. We consider a low net magnetic flux configuration (i.e.\ SANE rather than MAD; see e.g.\ \citealt{NarayanEtAl2012}). For $\rho_0$, $u$, $\kappa_e$, and $u^i$, axisymmetrization is a straightforward average in $\phi$.
For $B^i$, we construct a vector potential from the 3D data, average that, and then evaluate the axisymmetric field.
The accretion rate is controlled by varying the mass unit conversion from the scale-free GRMHD data. No radiation is present initially; the radiation field equilibrates on the light crossing time. We set the black hole mass to $10^8 M_{\odot}$ (near the turnover of the supermassive black hole mass function; e.g.\ \citealt{Kelly2012}) and dimensionless black hole spin $a_{\star} = 0.5$.
\subsection{Pathologies}
We employ the drift-frame floors described in \cite{Ressler2016} to repair unphysical {total fluid} densities and energies. Capturing numerical dissipation for electron heating is especially challenging{:} {\tt harm}{}-like schemes {can} violate the second law of thermodynamics locally {at the truncation error level}, and in our scheme the electrons may also be {\it cooled} anomalously near large fluid entropy gradients, such as at the funnel wall. See \cite{Ressler2016} for more details. We enforce $\Theta_{e,\mathrm{max}} < 1000$ in the radiation sector, and $T_p/T_e > 0.01$. Additionally, our explicit radiation-fluid coupling may yield negative electron energies. This is difficult to prevent except by increasing superphoton resolution. We monitor such ``supercooling'' events to ensure they are never a significant fraction of the total radiation energy budget. This diagnostic is used to set superphoton resolution, which is related to the cooling time of the flow.
\section{Results}
\label{sec:results}
\begin{deluxetable}{ccccc}
{
\tablehead{
\colhead{$\overline{\dot{m}}$} &
\colhead{$\overline{L}/L_{\mathrm{Edd}}$} &
\colhead{$\overline{\epsilon}$} &
\colhead{$\langle \overline{\Theta_e} \rangle_J$} &
\colhead{$\overline{L_{\mathrm{em}}}/\overline{L_{\mathrm{sc}}}$}
}
\tablecaption{Time-averaged Results\label{tab:results}}
\startdata
$1.25\times10^{-9}$ & $3.01\times10^{-14}$ & $2.45\times10^{-6}$ & $13.1$ & $1.51\times10^{4}$ \\
$1.08\times10^{-8}$ & $4.27\times10^{-12}$ & $4.45\times10^{-5}$ & $14.9$ & $1.07\times10^{3}$ \\
$1.18\times10^{-7}$ & $2.86\times10^{-10}$ & $2.60\times10^{-4}$ & $12.4$ & $1.42\times10^{2}$ \\
$9.33\times10^{-7}$ & $1.39\times10^{-8}$ & $1.61\times10^{-3}$ & $12.2$ & $1.79\times10^{1}$ \\
$1.01\times10^{-5}$ & $4.89\times10^{-7}$ & $5.07\times10^{-3}$ & $7.64$ & $1.74$ \\
\enddata
\tablecomments{{Accretion rate, luminosity, radiative efficiency, emission-weighted electron temperature, and ratio of emission to scattering process{es} for all simulations. Throughout, models are identified by $\overline{\dot{m}}$ rounded to the nearest power of $10$.}}
}
\end{deluxetable}
We consider the same initial conditions except at five accretion rates: $\dot{m} \sim (10^{-9}, 10^{-8}, 10^{-7},\\ 10^{-6}, 10^{-5})$. Each calculation extends for $1000 GM/c^3$. To {gauge} the importance of cooling, we run these models both with and without radiative cooling. Luminosities from models without cooling are post-processed using {\tt grmonty}{} with a $5 GM/c^3$ cadence.
Each superphoton records the $i,j$ indices of the zone of its last interaction; over a time interval $\Delta t$, superphotons captured at the outer radial boundary are used to compute volumetric radiative energy exchange rates in each zone $J_{i,j} \equiv \sum_n -w_n k_{0,n}/(\sqrt{-g}\Delta t \Delta x^1 \Delta x^2 \Delta x^3)$ with the sum taken over the $n$ recorded photons tagged with $i,j$. $J_{\mathrm{em}}$ is that due to emission (with self-absorption subtracted) and $J_{\mathrm{sc}}$ is that due to scattering. As above, heating rates are $Q \equiv du/d\tau$ {due to each process}. Luminosities $L$ are $\int R^1_{~0} \sqrt{-g} dx^2 dx^3$ evaluated at the outer radial radiation boundary. The mass accretion rate $\dot{M} = \int \rho_0 u^1 \sqrt{-g} dx^2 dx^3$ {is} evaluated at the inner radial boundary. Radiative efficiency $\epsilon \equiv L/\dot{M}$.
We begin time averages at the time at which global quantities ($\dot{m}${}, $L$, $\epsilon$) appear relatively steady; time averages (denoted as $\overline{f}$ for a quantity $f$) are always for $600 \leq tc^3/GM \leq 1000$. We also consider weighted spatial averages,
\begin{align}
\langle f \rangle_{\phi} \equiv \frac{\int dx^1dx^2dx^3\sqrt{-g} f \, \phi}{\int dx^1dx^2dx^3\sqrt{-g} \,\phi}.
\end{align}
For simple volume averages inside a radius $r$ denoted $\langle f \rangle_r$, $\phi = 1$ and $r$ sets the upper radial bound of the integrals. $r=10 GM/c^2$ is a natural choice , as it is approximately the radius inside of which viscous equilibrium is achieved.
\begin{figure}
\includegraphics[width=\columnwidth]{f1.pdf}
\caption{Globally averaged quantities. The top left panel shows radiative efficiency $\epsilon$ versus $\overline{\dot{m}}$ for models with and without radiative cooling, along with the thin disk efficiency ($\epsilon = 8.2\%$ for $a_{\star} = 0.5$; \citealt{NovikovThorne}). The top right panel shows the ratio between viscous and Coulomb heating. The bottom right panel shows the emissivity-weighted electron temperature, and the bottom left panel shows the ratio of outgoing radiation due to synchrotron and Compton processes.}
\label{fig:tavgs}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{f2.pdf}
\caption{Accretion rate, luminosity, and radiative efficiency as a function of time. {Time-averaging window is shown as the shaded region.} {Thin disk efficiency is shown as dashed line in the bottom panel.}}
\label{fig:mdot_lum_eff}
\end{figure}
Figure 1 compares the radiative efficiency $\epsilon$ for radiative and nonradiative models versus $\dot{m}$. Up to $\dot{m} \approx 10^{-6}$, the models are equivalent. At higher $\dot{m}$, however, radiative cooling significantly affects the bolometric luminosity. Therefore, for $\dot{m} \gtrsim 10^{-6}$, self-consistency requires the inclusion of radiative cooling. Note that this value is somewhat higher than the condition $\dot{m} \gtrsim 10^{-7}$ identified by \cite{Dibi2012}, possibly due to differing prescriptions for $T_e$. Additionally, our $\epsilon$ are a factor $\sim 5$ larger at comparable $\dot{m}$ than the $T_p/T_e = 10$, $a_{\star}=0$, 3D models of \cite{SadowskiGaspari2017}. Compton scattering becomes commensurate with synchrotron emission at $\dot{m} \approx 10^{-5}$, and Coulomb heating becomes energetically significant at the $\sim 10\%$ level. The emission-weighted electron temperature $\langle \overline{\Theta_e} \rangle_{J}$ decreases significantly for $\dot{m} \approx 10^{-5}$.
Electrons inside $r \sim 15 GM/c^2$ achieve thermal equilibrium in our models. These are the radiating electrons for all but the $\dot{m} \approx 10^{-5}$ model, where electrons out to $\sim 30 GM/c^2$ contribute to the luminosity. At $t=1000 GM/c^3$, these electrons are still heating slightly due to Coulomb coupling.
Hereafter we ignore the $\dot{m} \approx 10^{-9}$ model, as flow properties are nearly independent of $\dot{m}${} at such low rates in our model since radiation is negligible.
Figure \ref{fig:mdot_lum_eff} shows $\dot{m}$, luminosity $L$, and radiative efficiency $\epsilon$ as a function of time. $L$ scales superlinearly with $\dot{m}${} for all simulations reported here {($L \sim \dot{m}^2$, and therefore $\epsilon \sim \dot{m}$, for low $\dot{m}${}, as expected for synchrotron-dominated weak cooling)}{, consistent with the increase in $\epsilon$ with $\dot{m}${} seen in Figure \ref{fig:tavgs}}. Across this range of $\dot{m}$ the flow {transitions} from very radiatively inefficient to a nearly radiatively efficient luminous state.
\begin{figure}
\includegraphics[width=\columnwidth]{f3.pdf}
\caption{Time-averaged electron temperature for all models, also averaged about the midplane. Coulomb collisions heat up the disk at higher $\dot{m}${}
. The dashed line shows the funnel wall, defined as $b^2/\rho = 1$.}
\label{fig:Thetae}
\end{figure}
Figure \ref{fig:Thetae} shows the global structure of {the} electron temperature $\overline{\Theta_e}$ in the accretion flows near the black hole. The electron heating model used here leads to hotter electrons in the more magnetized corona and cooler electrons in the less-magnetized disk midplane (see \citealt{Ressler2015, Ressler2016} for more details). At the highest accretion rates, however, the midplane electrons are significantly hotter (at $r=20 GM/c^2$, in the midplane, $\Theta_e(\dot{m}=10^{-5})/\Theta_e(\dot{m}=10^{-8}) \approx 8$) due to Coulomb heating, and cooling lowers $\Theta_e$ in the inner regions of the flow.
\begin{figure}
\includegraphics[width=\columnwidth]{f4.pdf}
\caption{Spectral energy distributions for all models. {Both total $\nu L_{\nu}$ and that due to individual interactions (emission, one scattering event, etc.) are shown. The logarithmic interval in $\nu L_{\nu}$ is common to all panels.} At high $\dot{m}${}, multiple Compton scattering events form a high-energy, nearly power-law spectral component}
\label{fig:spectra}
\end{figure}
Figure \ref{fig:spectra} shows the spectra of emergent radiation for an observer nearly in the midplane of the disk. These are evaluated from the same superphotons present in the simulations. At low accretion rates the spectrum is very soft, with distinct Compton bumps, consistent with previous models {where radiation was calculated in post-processing without solving self-consistently for the electron temperature} (e.g.~\citealt{MoscibrodzkaEtAl2009}). As the accretion rate increases, the slope of the high energy tail shifts upwards.
These trends are consistent with spectral models of 1D RIAFs (e.g.\ \citealt{Esin+1997}, \citealt{Yuan+2004}).
\section{Conclusion}
\label{sec:conclusion}
We have presented general relativistic radiation magnetohydrodynamic simulations of radiatively inefficient black hole accretion flows. We have considered a black hole of mass $10^8 M_{\odot}$ and spin $a_{\star} = 0.5$, and
accretion rates up to and including those for which radiative cooling is important. In particular, our inclusion of frequency-dependent full radiation transport addresses an important uncertainty in simulations of RIAFs.
We have found that RIAF models depart from self-consistency at an accretion rate $\dot{m} \approx 10^{-6}${, in the sense that self-consistent calculations with cooling are needed to predict the radiative efficiency and spectrum}.
By $\dot{m} \approx 10^{-5}$, the cooling of these flows becomes dominated by Compton scattering, rather than emission, and the flow achieves nearly 1\%-level radiative efficiency.
Our results suggest that Coulomb collisions will become as important as viscous heating at an accretion rate of $\dot{m} \approx 10^{-4}$ (extrapolating Figure \ref{fig:tavgs} to somewhat higher $\dot{m}${}). This is well below what is traditionally assumed in semi-analytic models (for example, \cite{Esin+1997} assume that Coulomb collisions dominate for $\dot{m}${} $\gtrsim$ 0.1).
This is probably due to the different density and temperature profiles for analytic models and numerical simulations (e.g.\ \citealt{NarayanEtAl2012}).
Future work should study this in 3D simulations and assess the implications of this behavior for observations, including the phenomenology of state transitions in X-ray binaries.
Our study is limited to axisymmetry. To minimize this weakness{,}
we have used as initial conditions long-duration 3D nonradiative two-temperature GRMHD simulations.
Nonetheless, we achieve viscous and inflow equilibria only within $r \sim 15 GM/c^2$. This has potential consequences mostly for the $\dot{m} \sim 10^{-5}$ model, for which $\sim 20\%$ of the luminosity is generated beyond $15 GM/c^2$. In this model the electrons at large radius are still heating up; thermal equilibrium would imply a slightly higher radiative efficiency. Should the flow change after viscous equilibration (probably towards reduced proton pressure), the luminosity could be suppressed by $\sim 20\%$, mostly in the high-energy tail of the spectrum.
Our survey is not comprehensive. Black hole mass, spin, accretion disk tilt, and net magnetic flux may all significantly affect these results. We will study these dependencies in future work.
We have directly demonstrated that radiative cooling plays an important role
in RIAFs. The whole range of accretion rates considered in this work is probably populated by astrophysical sources, and the technique presented here will be valuable in interpreting observations of both stellar mass and supermassive black holes from the mm to the $\gamma$-ray.
\acknowledgments
It is a pleasure to thank M. Chandra, A. S{\c a}dowski, and J. Stone for useful discussions, as well as the anonymous referee for a very useful report. Work at Los Alamos National Laboratory was done under the auspices of the National Nuclear Security Administration of the US Department of Energy. SMR is supported in part by the NASA Earth and Space Science Fellowship. JD acknowledges support from the Laboratory Directed Research and Development program at Los Alamos National Laboratory. Support for AT was provided by NASA through Einstein Postdoctoral Fellowship grant number PF3-140131 awarded by the Chandra X-ray Center, which is operated by the Smithsonian Astrophysical Observatory for NASA under contract NAS8-03060, and the TAC fellowship, and by NSF through an XSEDE computational time allocation TG-AST100040 on TACC Stampede. This work was made possible by computing time granted by UCB on the Savio cluster. CFG's work was also supported in part by a Romano Professorial Scholar appointment, a Simons Fellowship in Theoretical Physics, and a Visiting Fellowship at All Souls College, Oxford. EQ is supported in part by a Simons Investigator Award from the Simons Foundation and the David and Lucile Packard Foundation. This work was supported in part by NSF grant AST 13-33612. This research used resources provided by the Los Alamos National Laboratory Institutional Computing Program, which is supported by the U.S. Department of Energy National Nuclear Security Administration under Contract No.\ DE-AC52-06NA25396.
\newpage
|
\section{Introduction}
In the past two decades, gradings on Lie algebras by arbitrary abelian groups have been extensively studied. For finite-dimensional simple Lie algebras over an algebraically closed field $\FF$, the classification of fine gradings up to equivalence has recently been completed (assuming $\Char \FF = 0$) by efforts of many authors --- see the monograph \cite[Chapters 3--6]{livromicha} and the references therein, and also \cite{YuExc} and \cite{E14}. For a fixed abelian group $G$, the classification of $G$-gradings up to isomorphism is also known (assuming $\Char \FF \neq 2$), except for types $E_6$, $E_7$ and $E_8$ --- see \cite{livromicha} and \cite{EK_d4}.
This paper is devoted to gradings on finite-dimensional simple Lie superalgebras. Over an algebraically closed field of characteristic $0$, such superalgebras were classified by V.~G.~Kac in \cite{kacZ,artigokac} (see also \cite{livrosuperalgebra}). In \cite{kacZ}, there is also a classification of $\ZZ$-gradings on these superalgebras. More recently, gradings by arbitrary abelian groups have been considered. Fine gradings on the exceptional simple Lie superalgebras, namely, $D(2,1;\alpha)$, $G(3)$ and $F(4)$, were classified in \cite{artigoelduque} and all gradings on the series $Q(n)$, $n\geq 2$, were classified in \cite{paper-Qn}. A description of gradings on matrix superalgebras, here denoted by $M(m,n)$ (see Section \ref{sec:Mmn}), was given in \cite{BS}, but the isomorphism problem was left open and fine gradings were not considered.
The initial goal of this work was to classify abelian group gradings on the series $P(n)$, $n\geq 2$, and thereby complete the classification of gradings on the so-called ``strange Lie superalgebras''. Our approach led us to the study of gradings on the associative superalgebras $M(m,n)$ and the closely related Lie superalgebras $A(m,n)$.
Throughout this work, the canonical $\ZZ_2$-grading of a superalgebra will be denoted by superscripts, reserving subscripts for the components of other gradings. Thus, a $G$-grading on a superalgebra $A = A\even \oplus A\odd$ is a vector space decomposition $\Gamma:\,A = \bigoplus_{g \in G} A_g$ such that $A_g A_h\subseteq A_{gh}$, for all $g,h\in G$, and each $A_g$ is compatible with the superalgebra structure, i.e., $A_g=A_g^\bz \oplus A_g^\bo$. Note that $G$-gradings on a superalgebra can be seen as $G\times \ZZ_2$-gradings on the underlying algebra. For the superalgebras under consideration, namely, $M(m,n)$, $A(m,n)$ and $P(n)$, the canonical $\ZZ_2$-grading can be refined to a canonical $\ZZ$-grading, whose components will be denoted by superscripts $-1, 0, 1$.
Only gradings by abelian groups are discussed in this work, which is no loss of generality in the case of simple Lie superalgebras, because the support always generates an abelian group.
All the (super)algebras and vector (super)spaces are assumed to be finite-dimensional over a fixed algebraically closed field $\FF$.
When dealing with the Lie superalgebras $A(m,n)$ and $P(n)$, we will also assume $\Char \FF = 0$.
The paper is structured as follows. Sections \ref{sec:generalities} and \ref{sec:gradings-on-matrix-algebras} have no original results. In the former, we introduce all basic definitions and a few general results for future reference, and the latter is a review of the classification of gradings on matrix algebras closely following \cite[Chapter 2]{livromicha}, with a slight change in notation.
Section \ref{sec:Mmn} is devoted to the associative superalgebras $M(m,n)$, which
have two kinds of gradings: the \emph{even gradings} are compatible with the canonical $\ZZ$-grading and the \emph{odd gradings} are not.
(The latter can occur only if $m=n$.)
The classification results for even gradings are Theorems \ref{thm:even-assc-iso} ($G$-gradings up to isomorphism)
and \ref{thm:class-fine-even} (fine gradings up to equivalence).
We present two descriptions of odd gradings: one as $G\times \ZZ_2$-gradings on the underlying matrix algebra (see Subsection \ref{ssec:grds-on-superalgebras})
and the other purely in terms of the group $G$ (see Subsection \ref{ssec:second-odd}). We classify odd gradings in Theorems \ref{thm:first-odd-iso} and \ref{thm:2nd-odd-iso} ($G$-gradings up to isomorphism) and in Theorem \ref{thm:class-fine-odd} (fine gradings up to equivalence).
In Section \ref{sec:Amn}, we consider gradings on the Lie superalgebras $A(m,n)$, but only those that are induced from $M(m+1, n+1)$ (see Definition \ref{def:Type-I}). We classify them up to isomorphism in Theorem \ref{thm:even-Lie-iso} (even gradings) and in Theorem \ref{thm:first-odd-Lie-iso} and Corollary \ref{cor:2nd-odd-Lie-iso} (odd gradings).
In Section \ref{sec:Pn}, we classify gradings on the Lie superalgebras $P(n)$: see Theorem \ref{thm:Pn-iso} for $G$-gradings up to isomorphism
and Theorem \ref{thm:class-fine-Pn} for fine gradings up to equivalence.
\section{Generalities on gradings}\label{sec:generalities}
The purpose of this section is to fix notation and recall definitions concerning graded algebras and graded modules.
\subsection{Gradings on vector spaces and (bi)modules}\label{subsec:graded-bimodules}
Let $G$ be a group. By a \emph{$G$-grading} on a vector space $V$ we mean simply a vector space decomposition $\Gamma:\,V = \bigoplus_{g \in G} V_g$ where the summands are labeled by elements of $G$. If $\Gamma$ is fixed, $V$ is referred to as a {\em $G$-graded vector space}. A subspace $W \subseteq V$ is said to be \emph{graded} if $W = \bigoplus_{g \in G} (W \cap V_g)$. We will refer to $\ZZ_2$-graded vector spaces as \emph{superspaces} and their graded subspaces as \emph{subsuperspaces}.
An element $v$ in a graded vector space $V = \bigoplus_{g \in G} V_g$ is said to be \emph{homogeneous} if $v\in V_g$ for some $g\in G$.
If $0\ne v\in V_g$, we will say that $g$ is the \emph{degree} of $v$ and write $\deg v = g$.
In reference to the canonical $\ZZ_2$-grading of a superspace, we will instead speak of the \emph{parity} of $v$ and write $|v| = g$.
Every time we write $\deg v$ or $|v|$, it should be understood that $v$ is a nonzero homogeneous element.
\begin{defi}
Given two $G$-graded vector spaces, $V=\bigoplus_{g\in G} V_g$ and $W=\bigoplus_{g\in G} W_g$, we define their tensor product to be the vector space $V\otimes W$
together with the $G$-grading given by $(V \otimes W)_g = \bigoplus_{ab=g} V_{a} \otimes W_{b}$.
\end{defi}
The concept of grading on a vector space is connected to gradings on algebras by means of the following:
\begin{defi}
If $V=\bigoplus_{g\in G} V_{g}$ and $W=\bigoplus_{g\in G} W_{g}$ are two graded vector spaces and $T: V\rightarrow W$ is a linear map, we say that $T$ is \emph{homogeneous of degree $t$}, for some $t\in G$, if $T(V_g)\subseteq W_{tg}$ for all $g\in G$.
\end{defi}
If $S: U\rightarrow V$ and $T: V\rightarrow W$ are homogeneous linear maps of degrees $s$ and $t$, respectively,
then the composition $T\circ S$ is homogeneous of degree $ts$.
We define the {\em space of graded linear transformations} from $V$ to $W$ to be:
\[ \Hom^{\text{gr}} (V,W) = \bigoplus_{g\in G} \Hom (V,W)_{g}\]
where $\Hom (V,W)_{g}$ denotes the set of all linear maps from $V$ to $W$ that are homogeneous of degree $g$.
If we assume $V$ to be finite-dimensional then we have $\Hom(V,W)=\Hom^{\gr}(V,W)$ and, in particular, $\End (V) = \bigoplus_{g\in G} \End (V)_g$ is a graded algebra.
We also note that $V$ becomes a graded module over $\End(V)$ in the following sense:
\begin{defi}
Let $A$ be a $G$-graded algebra (associative or Lie) and let $V$ be a (left) module over $A$ that is also a $G$-graded vector space. We say that $V$ is a \emph{graded $A$-module} if $A_g \cdot V_h \subseteq V_{gh}$, for all $g$,$h\in G$. The concept of $G$-\emph{graded bimodule} is defined similarly.
\end{defi}
If we have a $G$-grading on a Lie superalgebra $L=L\even \oplus L\odd$ then, in particular, we have a grading on the Lie algebra $L\even$ and a grading on the space $L\odd$ that makes it a graded $L\even$-module. If we have a $G$-grading on an associative superalgebra $C=C\even \oplus C\odd$, then $C\odd$ becomes a graded bimodule over $C\even$.
If $ \Gamma$ is a $G$-grading on a vector space $V$ and $g\in G$, we denote by $\Gamma^{[g]} $ the grading given by relabeling the component
$V_h$ as $V_{hg}$, for all $h \in G$. This is called the \emph{(right) shift of the grading $\Gamma$ by $g$}.
We denote the graded space $(V, \, \Gamma^{[g]})$ by $V^{[g]}$.
From now on, we assume that $G$ is abelian.
If $V$ is a graded module over a graded algebra (or a graded bimodule over a pair of graded algebras), then $V^{[g]}$ is also a graded (bi)module.
We will make use of the following partial converse (see e.g. \cite[Proposition 3.5]{paper-Qn}):
\begin{lemma}\label{lemma:simplebimodule}
Let $A$ and $B$ be $G$-graded algebras and let $V$ be a finite-dimensional (ungraded) simple $A$-module or $(A,B)$-bimodule. If $\Gamma$ and $\Gamma'$ are two $G$-gradings that make $V$ a graded (bi)module, then $\Gamma'$ is a shift of $\Gamma$.\qed
\end{lemma}
Certain shifts of grading may be applied to graded $\ZZ$- or $\ZZ_2$-superalgebras. In the case of a $\ZZ$-superalgebra $L=L^{-1}\oplus L^{0}\oplus L^{1}$, we have the following:
\begin{lemma}\label{lemma:opposite-directions}
Let $L=L^{-1}\oplus L^0\oplus L^1$ be a $\ZZ$-superalgebra such that $L^1\, L^{-1}\neq 0$. If we shift the grading on $L^1$ by $g\in G$ and the grading on $L^{-1}$ by $g' \in G$, then we have a grading on $L$ if and only if $g' = g^{-1}$. \qed
\end{lemma}
We will describe this situation as \emph{shift in opposite directions}.
\subsection{Universal grading group, equivalence and isomorphism of gradings}
There is a concept of grading not involving groups. A \emph{set grading} on a (super)algebra $A$ is a decomposition $\Gamma:\,A=\bigoplus_{s\in S}A_s$ as a direct sum of sub\-(su\-per)\-spa\-ces indexed by a set $S$ and having the property that, for any $s_1,s_2\in S$ with $A_{s_1}A_{s_2}\ne 0$, there exists $s_3\in S$ such that $A_{s_1}A_{s_2}\subseteq A_{s_3}$. The \emph{support} of $\Gamma$ (or of $A$) is defined to be the set $\supp(\Gamma) := \{s\in S \mid A_s \neq 0\}$.
Similarly, $\supp_\bz(\Gamma) := \{s\in S \mid A_s^\bz \neq 0\}$ and $\supp_\bo(\Gamma) := \{s\in S \mid A_s^\bo \neq 0\}$.
For a set grading $\Gamma:\,A=\bigoplus_{s\in S}A_s$, there may or may not exist a group $G$ containing $\supp(\Gamma)$ that makes $\Gamma$ a $G$-grading.
If such a group exists, $\Gamma$ is said to be a {\em group grading}. (As already mentioned, we only consider abelian group gradings in this paper.)
However, $G$ is usually not unique even if we require that it should be generated by $\supp(\Gamma)$.
The {\em universal (abelian) grading group} of $\Gamma$ is generated by $\supp(\Gamma)$ and has the defining relations
$s_1s_2=s_3$ for all $s_1,s_2,s_3\in S$ such that $0\neq A_{s_1}A_{s_2}\subseteq A_{s_3}$.
This group is universal among all (abelian) groups that realize the grading $\Gamma$ (see e.g. \cite[Chapter 1]{livromicha} for details).
Let $\Gamma:\,A=\bigoplus_{g\in G} A_g$ and $\Delta:\,B=\bigoplus_{h\in H} B_h$ be two group gradings on the (super)algebras $A$ and $B$, with supports $S$ and $T$, respectively.
We say that $\Gamma$ and $\Delta$ are {\em equivalent} if there exists an isomorphism of (super)algebras $\vphi: A\to B$ and a bijection $\alpha: S\to T$ such that $\vphi(A_s)=B_{\alpha(s)}$ for all $s\in S$. If $G$ and $H$ are universal grading groups then $\alpha$ extends to an isomorphism $G\to H$. In the case $G=H$, the $G$-gradings $\Gamma$ and $\Delta$ are {\em isomorphic} if $A$ and $B$ are isomorphic as $G$-graded (super)algebras, i.e., if there exists an isomorphism of (super)algebras $\vphi: A\to B$ such that $\vphi(A_g)=B_g$ for all $g\in G$.
If $\Gamma:\,A=\bigoplus_{g\in G} A_g$ and $\Gamma':\,A=\bigoplus_{h\in H} A'_h$ are two gradings on the same (super)algebra $A$, with supports $S$ and $T$, respectively, then we will say that $\Gamma'$ is a {\em refinement} of $\Gamma$ (or $\Gamma$ is a {\em coarsening} of $\Gamma'$) if, for any $t\in T$, there exists (unique) $s\in S$ such that $A'_t\subseteq A_s$. If, moreover, $A'_t\ne A_s$ for at least one $t\in T$, then the refinement is said to be {\em proper}. A grading $\Gamma$ is said to be {\em fine} if it does not admit any proper refinements.
Note that if $A$ is a superalgebra then $A=\bigoplus_{(g,i)\in G\times\mathbb{Z}_2}A_g^i$ is a refinement of $\Gamma$.
It follows that if $\Gamma$ is fine then the sets $\supp_\bz(\Gamma)$ and $\supp_\bo(\Gamma)$ are disjoint.
If, moreover, $G$ is the universal group of $\Gamma$, then the superalgebra structure on $A$ is given by the unique homomorphism $p: G \to \ZZ_2$
that sends $\supp_\bz(\Gamma)$ to $\bar 0$ and $\supp_\bo(\Gamma)$ to $\bar 1$.
\begin{defi}
Let $G$ and $H$ be groups, $\alpha:G\to H$ be a group homomorphism and $\Gamma:\,A=\bigoplus_{g\in G} A_g$ be a $G$-grading. The \emph{coarsening of $\Gamma$ induced by $\alpha$} is the $H$-grading ${}^\alpha \Gamma: A= \bigoplus_{h\in H} B_h$ where
$ B_h = \bigoplus_{g\in \alpha\inv (h)} A_g$. (This coarsening is not necessarily proper.)
\end{defi}
The following result appears to be ``folklore''. We include a proof for completeness.
\begin{lemma}\label{lemma:universal-grp}
Let $\mathcal{F}=\{\Gamma_i\}_{i\in I}$, be a family of pairwise nonequivalent fine (abelian) group gradings on a (super)algebra $A$, where $\Gamma_i$ is a $G_i$-grading and $G_i$ is generated by $\supp(\Gamma_i)$. Suppose that $\mathcal{F}$ has the following property:
for any grading $\Gamma$ on $A$ by an (abelian) group $H$, there exists $i\in I$ and a homomorphism $\alpha:G_i\to H$ such that $\Gamma$
is isomorphic to ${}^\alpha\Gamma_i$. Then
%
\begin{enumerate}[(i)]
\item every fine (abelian) group grading on $A$ is equivalent to a unique $\Gamma_i$;
\item for all $i$, $G_i$ is the universal (abelian) group of $\Gamma_i$.
\end{enumerate}
\end{lemma}
\begin{proof}
Let $\Gamma$ be a fine grading on $A$, realized over its universal group $H$. Then there is $i\in I$ and $\alpha: G_i \to H$ such that ${}^\alpha \Gamma_i \iso \Gamma$. Writing $\Gamma_i: A = \bigoplus_{g\in G_i} A_g$ and $\Gamma: A = \bigoplus_{h\in H} B_h$, we then have $\vphi \in \Aut(A)$ such that
\[
\vphi\,\big( \bigoplus_{g\in \alpha\inv (h)} A_g \big) = B_h
\]
for all $h\in H$. Since $\Gamma$ is fine, we must have $B_h \neq 0$ if, and only if, there is a unique $g\in G_i$ such that $\alpha(g) = h$, $A_g\neq 0$ and $\vphi(A_g) = B_h$. Equivalently, $\alpha$ restricts to a bijection $\supp(\Gamma_i) \to \supp(\Gamma)$ and $\vphi(A_g) = B_{\alpha(g)}$ for all $g \in S_i:= \supp (\Gamma_i)$. This proves assertion $(i)$.
Let $G$ be the universal group of $\Gamma_i$. It follows that, for all $s_1, s_2, s_3 \in S_i$,
%
\begin{equation*} \label{eq:relations-unvrsl-grp}
\begin{split}
& s_1s_2 = s_3 \text{ is a defining relation of } G \\
\iff & 0 \neq A_{s_1} A_{s_2} \subseteq A_{s_3}\\
\iff & 0 \neq B_{\alpha(s_1)} B_{\alpha(s_2)} \subseteq B_{\alpha (s_3)}\\
\iff & \alpha(s_1)\alpha(s_2) = \alpha(s_3) \text{ is a defining relation of } H.
\end{split}
\end{equation*}
%
Therefore, the bijection $\alpha\restriction_{S_i}$ extends uniquely to an isomorphism $\beta: G\rightarrow H$.
By the universal property of $G$, there is a unique homomorphism $\gamma: G\to G_i$ that restricts to the identity on $S_i$. Hence, the following diagram commutes:
%
\begin{center}
\begin{tikzcd}
G \arrow[to=Gi, "\gamma"] \arrow[to = H, "\beta"]&&\\
&& |[alias=H]|H\\
|[alias=Gi]|G_i \arrow[to=H, "\alpha"]&&
\end{tikzcd}
\end{center}
%
Since $\beta$ is an isomorphism, $\gamma$ must be injective. But $\gamma$ is also surjective since $S_i$ generates $G_i$. Hence $G_i$ is isomorphic to $G$. Since $\Gamma$ was an arbitrary fine grading, for each given $j\in I$, we can take $\Gamma = \Gamma_j$ (hence, $i=j$ and $H=G$). This concludes the proof of $(ii)$.
\end{proof}
\begin{defi}
Let $\Gamma$ be a grading on an algebra $A$. We define $\Aut(\Gamma)$ as the group of all self-equivalences of $\Gamma$, i.e., automorphisms of $A$ that
permute the components of $\Gamma$. Let $\operatorname{Stab}(\Gamma)$ be the subgroup of $\Aut(\Gamma)$ consisting of the automorphisms that fix
each component of $\Gamma$. Clearly, $\operatorname{Stab}(\Gamma)$ is a normal subgroup of $\Aut(\Gamma)$, so we can define the \emph{Weil group} of
$\Gamma$ by $\operatorname W (\Gamma) := \frac{\Aut(\Gamma)}{\operatorname{Stab}(\Gamma)}$. The group $\operatorname W (\Gamma)$ can be seen as a subgroup
of the permutation group of the support and also as a subgroup of the automorphism group of the universal group of $\Gamma$.
\end{defi}
\subsection{Correspondence between $G$-gradings and $\widehat G$-actions}\label{ssec:G-hat-action}
One of the most important tools for dealing with gradings by abelian groups on (super)algebras is to translate a $G$-grading into a $\widehat G$-action, where $\widehat G$ is the algebraic group of characters of $G$, \ie, group homomorphisms $G \rightarrow \FF^{\times}$. The group $\widehat{G}$ acts on any $G$-graded (super)algebra $A = \bigoplus_{g\in G} A_g$ by $\chi \cdot a = \chi(g) a$ for all $a\in A_g$ (extended to arbitrary $a\in A$ by linearity). The map given by the action of a character $\chi \in \widehat{G}$ is an automorphism of $A$. If $\FF$ is algebraically closed and $\Char \FF = 0$, then $A_g = \{ a\in A \mid \chi \cdot a = \chi (g) a\}$, so the grading can be recovered from the action.
For example, if $A=A\even \oplus A\odd$ is a superalgebra, the action of the nontrivial character of $\ZZ_2$ yields the \emph{parity automorphism} $\upsilon$, which acts as the identity on $A\even$ and as the negative identity on $A\odd$. If $A$ is a $\ZZ$-graded algebra, we get a representation $\widehat \ZZ = \FF^\times \rightarrow \Aut (A)$ given by $\lambda \mapsto \upsilon_\lambda$ where $\upsilon_{\lambda}$ acts as $\lambda^i \id$ on $A^i$.
A grading on a (super)algebra over an algebraically closed field of characteristic $0$ is said to be \emph{inner} if it corresponds to an action by inner automorphisms. For example, the inner gradings on $\Sl(n)$ (also known as Type I gradings) are precisely the restrictions of gradings on the associative algebra $M_n(\FF)$.
\section{Gradings on matrix algebras}
\label{sec:gradings-on-matrix-algebras}
In this section we will recall the classification of gradings on matrix algebras. We will follow \cite[Chapter 2]{livromicha} but use slightly different notation, which we will extend to superalgebras in Section \ref{sec:Mmn}.
The following is the graded version of a classical result (see e.g. \cite[Theorem 2.6]{livromicha}).
We recall that a \emph{graded division algebra} is a graded unital associative algebra such that every nonzero homogeneous element is invertible.
\begin{thm}\label{thm:End-over-D}
Let $G$ be a group and let $R$ be a $G$-graded associative algebra that has no nontrivial graded ideals and satisfies the descending chain condition on
graded left ideals. Then there is a $G$-graded division algebra $\D$ and a graded (right) $\D$-module $\mc{V}$ such that $R \simeq \End_{\D} (\mc{V})$ as graded algebras.\qed
\end{thm}
We apply this result to the algebra $R=M_n(\FF)$ equipped with a grading by an abelian group $G$. We will now introduce the parameters that determine
$\mc D$ and $\mc V$, and give an explicit isomorphism $\End_{\D} (\mc{V})\simeq M_n(\FF)$ (see Definition \ref{def:explicit-grd-assoc}).
Let $\D$ be a finite-dimensional $G$-graded division algebra. It is easy to see that $T= \supp \D$ is a finite subgroup of $G$.
Also, since we are over an algebraically closed field, each homogeneous component $\D_t$, for $t\in T$, is one-dimensional. We can choose a generator $X_t$ for each $\D_t$. It follows that, for every $u,v\in T$, there is a unique nonzero scalar $\beta (u,v)$ such that $X_u X_v = \beta (u,v) X_v X_u$. Clearly, $\beta (u,v)$ does not depend on the choice of $X_u$ and $X_v$.
The map $\beta: T\times T \rightarrow \FF^{\times}$ is a \emph{bicharacter}, \ie, both maps $\beta(t,\cdot)$ and $\beta(\cdot,t)$ are characters for every $t \in T$. It is also \emph{alternating} in the sense that $\beta (t,t) = 1$ for all $t\in T$. We define the \emph{radical} of $\beta$ as the set $\rad \beta = \{ t\in T \mid \beta(t, T) = 1 \}$. In the case we are interested in, where $\D$ is simple as an algebra, the bicharacter $\beta$ is \emph{nondegenerate}, \ie, $\rad \beta = \{e\} $. The isomorphism classes of $G$-graded division algebras that are finite-dimensional and simple as algebras are in one-to-one correspondence with the pairs $(T,\beta)$ where $T$ is a finite subgroup of $G$ and $\beta$ is an alternating nondegenerate bicharacter on $T$ (see e.g. \cite[Section 2.2]{livromicha} for a proof).
Using that the bicharacter $\beta$ is nondegenerate, we can decompose the group $T$ as $A\times B$, where the restrictions of $\beta$ to each of the subgroups $A$ and $B$ is trivial, and hence $A$ and $B$ are in duality by $\beta$. We can choose the elements $X_t\in \D_t$ in a convenient way (see \cite[Remark 2.16]{livromicha} and
\cite[Remark 18]{EK15}) such that $X_{ab}=X_aX_b$ for all $a\in A$ and $b\in B$. Using this choice, we can define an action of $\D$ on the vector space underlying the group algebra $\FF B$, by declaring $X_a\cdot e_{b'} = \beta(a, b') e_{b'}$ and $X_b\cdot e_{b'} = e_{bb'}$.
This action allows us to identify $\D$ with $\End{(\FF B)}$. Using the basis $\{e_{b}\mid b\in B\}$ in $\FF B$, we can see it as a matrix algebra, where
\[X_{ab}= \sum_{b'\in B} \beta(a, bb') E_{bb', b'}\]
and $E_{b'', b'}$ with $b'$, $b'' \in B$, is a matrix unit, namely, the matrix of the operator that sends $e_{b'}$ to $e_{b''}$ and sends all other basis elements to zero.
\begin{defi}
We will refer to these matrix models of $\mc D$ as its \emph{standard realizations}.
\end{defi}
\begin{remark}\label{rmk:2-grp-transp}
The matrix transposition is always an involution of the algebra structure. As to the grading, we have
%
\[
X_{ab}\transp = \sum_{b'\in B} \beta(a, bb') E_{b',bb'}
= \beta(a,b) \sum_{b''\in B} \beta(a, b^{-1}b'') E_{b^{-1}b'', b''} = \beta(a,b) X_{ab^{-1}}.
\]
%
It follows that if $T$ is an elementary 2-group, then the transposition preserves the degree.
In this case, we will use it to fix an identification between the graded algebras $\D$ and $\D\op$.
\end{remark}
Graded modules over a graded division algebra $\mc D$ behave similarly to vector spaces. The usual proof that every vector space has a basis, with obvious modifications, shows that every graded $\mc D$-module has a \emph{homogeneous basis}, \ie, a basis formed by homogeneous elements.
Let $\mc V$ be such a module of finite rank $k$, fix a homogeneous basis $\mc B = \{v_1, \ldots, v_k\}$ and let $g_i := \operatorname{deg} v_i$. We then have $\mc{V}\iso \ \D^{[g_1]}\oplus\cdots\oplus\D^{[g_k]}$, so, the graded $\mc D$-module $\mc V$ is determined by the $k$-tuple $\gamma = (g_1,\ldots, g_k)$. The tuple $\gamma$ is not unique. To capture the precise information that determines the isomorphism class of $\mc V$, we use the concept of \emph{multiset}, \ie, a set together with a map from it to the set of positive integers. If $\gamma = (g_1,\ldots, g_k)$ and $T=\supp \D$, we denote by $\Xi(\gamma)$ the multiset whose underlying set is $\{g_1 T,\ldots, g_k T\} \subseteq G/T$ and the multiplicity of $g_i T$, for $1\leq i\leq k$, is the number of entries of $\gamma$ that are congruent to $g_i$ modulo $T$.
Using $\mc B$ to represent the linear maps by matrices in $M_k(\D) = M_k(\FF)\tensor \D$, we now construct an explicit matrix model for $\End_{\D}(\mc V)$.
\begin{defi}\label{def:explicit-grd-assoc}
Let $T \subseteq G$ be a finite subgroup, $\beta$ a nondegenerate alternating bicharacter on $T$, and $\gamma = (g_1, \ldots, g_k)$ a $k$-tuple of elements of $G$. Let $\D$ be a standard realization of a graded division algebra associated to $(T, \beta)$. Identify $M_k(\FF)\tensor \D \iso M_n(\FF)$ by means of the Kronecker product, where $n=k\sqrt{|T|}$. We will denote by $\Gamma(T, \beta, \gamma)$ the grading on $M_n(\FF)$ given by $\deg (E_{ij} \tensor d) := g_i (\deg d) g_j\inv$ for $i,j\in \{1, \ldots , k\}$ and homogeneous $d\in \D$, where $E_{ij}$ is the $(i,j)$-th matrix unit.
\end{defi}
If $\End(V)$, equipped with a grading, is isomorphic to $M_n(\FF)$ with $\Gamma(T, \beta, \gamma)$, we may abuse notation and also denote the grading on $\End(V)$ by $\Gamma(T,\beta,\gamma)$.
We restate \cite[Theorem 2.27]{livromicha} using our notation:
\begin{thm}\label{thm:classification-matrix}
Two gradings, $\Gamma(T,\beta,\gamma)$ and $\Gamma(T',\beta',\gamma')$, on the algebra $M_n(\FF)$ are isomorphic if, and only if, $T=T'$, $\beta=\beta'$ and there is an element $g\in G$ such that $g \Xi(\gamma)=\Xi(\gamma')$.\qed
\end{thm}
The proof of this theorem is based on the following result (see Theorem 2.10 and Proposition 2.18 from \cite{livromicha}), which will also be needed:
\begin{prop}\label{prop:inner-automorphism}
If $\phi: \End_\D (\mc V) \rightarrow \End_\D (\mc V')$ is an isomorphism of graded algebras, then there is a homogeneous invertible
$\D$-linear map $\psi: \mc V\rightarrow \mc V'$ such that $\phi(r)=\psi \circ r \circ \psi\inv$, for all $r\in \End_\D (\mc V)$.\qed
\end{prop}
\section{Gradings on $M(m,n)$}\label{sec:Mmn}
\subsection{The associative superalgebra $M(m,n)$}\label{M(m,n)}
Let $U = U\even \oplus U\odd$ be a superspace.
The algebra of endomorphisms of $U$ has an induced $\Zmod2$-grading, so it can be regarded as a superalgebra. It is convenient to write it in matrix form:
\begin{equation}\label{eq:End_U}
\End(U) = \left(\begin{matrix}
\End(U\even) & \Hom(U\odd,U\even)\\
\Hom(U\even,U\odd) & \End(U\odd)\\
\end{matrix}
\right).
\end{equation}
Choosing bases, we may assume that $U\even=\FF^m$ and $U\odd=\FF^n$, so the superalgebra $\End(U)$ can be seen as a matrix superalgebra, which is denoted by $M(m,n)$.
We may also regard $U$ as a $\ZZ$-graded vector space, putting $U^0=U\even$ and $U^1=U\odd$. By doing so, we obtain an induced $\ZZ$-grading on $M(m,n) = \End (U)$ such that
\[(\End\, U)\even =(\End\, U)^0 =
\left(\begin{matrix}
\End(U\even) & 0\\
0 & \End(U\odd)\\
\end{matrix}
\right)
\]
and $(\End\, U)\odd = (\End\, U)^{-1}\oplus (\End\, U)^1$ where
\[(\End U)^{1}=
\left(\begin{matrix}
0 & 0\\
\Hom(U\even,U\odd) & 0\\
\end{matrix}
\right) \,\text{ and }\, (\End U)^{-1}=
\left(\begin{matrix}
0 & \Hom(U\odd,U\even)\\
0 & 0 \\
\end{matrix}
\right).
\]
This grading will be called the \emph{canonical $\ZZ$-grading} on $M(m,n)$
\subsection{Automorphisms of $M(m,n)$}
It is known that the automorphisms of the superalgebra $\End(U)$ are conjugations by invertible homogeneous operators.
(This follows, for example, from Proposition \ref{prop:inner-automorphism}.) The invertible even operators are of the form $\left( \begin{matrix}
a&0\\
0&d\\
\end{matrix}\right)$ where $a\in \GL(m)$ and $d\in \GL(n)$. The corresponding inner automorphisms of $M(m,n)$ will be called \emph{even automorphisms}.
They form a normal subgroup of $\Aut(M(m,n))$, which we denote by $\mc E$.
The inner automorphisms given by odd operators will be called \emph{odd automorphisms}.
Note that an invertible odd operator must be of the form $\left( \begin{matrix}
0&b\\
c&0\\
\end{matrix}\right)$ where both $b$ and $c$ are invertible, and this forces $m=n$.
In this case, the set of odd automorphisms is a coset of $\mc E$, namely, $\pi \mc E$,
where $\pi$ is the conjugation by the matrix $\left( \begin{matrix}
0_n & I_n\\
I_n & 0_n\\
\end{matrix}\right)$. This automorphism is called the \emph{parity transpose} and is usually denoted by superscript:
\begin{equation*}
\left( \begin{matrix}
a&b\\
c&d\\
\end{matrix}\right)^\pi = \left( \begin{matrix}
d&c\\
b&a\\
\end{matrix}\right).
\end{equation*}
Thus, $\Aut (M(m,n)) = \mc E$ if $m\neq n$, and $\Aut (M(n,n)) = \mc E \rtimes \langle \pi \rangle$.
\begin{remark}\label{rmk:Aut-ZZ-superalgebra}
It is worth noting that $\mc E$ is the automorphism group of the $\ZZ$-\-su\-per\-al\-gebra structure of $M(m,n)$, regardless of the values $m$ and $n$. Indeed, the elements of this group are conjugations by homogeneous matrices with respect to the canonical $\ZZ$-grading, but all the matrices of degree $-1$ or $1$ are degenerate.
\end{remark}
\subsection{Gradings on matrix superalgebras}\label{ssec:grds-on-superalgebras}
We are now going to generalize the results of Section \ref{sec:gradings-on-matrix-algebras} to the superalgebra $\M(m,n)$. It is clear that a $G$-graded associative superalgebra is equivalent to a $(G \times \ZZ_2)$-graded associative algebra, hence one could think that there is no new problem. But the description of gradings on matrix algebras presented in Section \ref{sec:gradings-on-matrix-algebras} does not allow us to readily see the gradings on the even and odd components of the superalgebra, so we are going to refine that description. We will denote the group $G\times \ZZ_2$ by $G^\#$ and the projection on the second factor by $p\colon G^\# \rightarrow \ZZ_2$. Also, we will abuse notation and identify $G$ with $G\times \{\barr 0\} \subseteq G^\#$.
\begin{remark}
If the canonical $\ZZ_2$-grading is a coarsening of the $G$-grading by means of a homomorphism $p\colon G\rightarrow \ZZ_2$ (referred to as the \emph{parity homomorphism}), then we have another isomorphic copy of $G$ in $G^\#$, namely, the image of the embedding $g\mapsto (g, p (g))$, which contains the support of the $G^\#$-grading. In this case, we do not need $G^\#$ and can work with the original $G$-grading.
\end{remark}
A $G$-graded superalgebra $\mathcal D$ is called a \emph{graded division superalgebra} if every nonzero homogeneous element in $\D\even \cup \D\odd$ is invertible --- in other words, $\D$ is a $G^\#$-graded division algebra.
We separate the gradings on $\M(m,n)$ in two classes depending on the superalgebra structure on $\D$: if $\D\odd = 0$, we say that we have an \emph{even grading} and, if $\D\odd \ne 0$, we have an \emph{odd grading}.
To see the difference between even and odd gradings, consider the $G^\#$-graded algebra $E=\End_\D (\mc U)$, where $\D$ is a $G^{\#}$-graded division algebra and $\mc U$ is a graded module over $\D$. Define
\[
\mc U\even = \bigoplus_{g\in G^\#} \{u\in \mc U_g \mid p(g)=\barr 0\}\,\, \text{and} \,\,\mc U\odd = \bigoplus_{g\in G^\#} \{u\in \mc U_g \mid p(g)=\barr 1\}.
\]
Then $\mc U\even$ and $\mc U\odd$ are $\D\even$-modules, but they are $\D$-modules if and only if $\D\odd=0$. So, in the case of an even grading, $\mc U$ is as a direct sum of $\D$-modules, and all the information related to the canonical $\ZZ_2$-grading on $\End_\D (\mc U)$ comes from the decomposition $\mc U=\mc U\even \oplus \mc U\odd$.
\begin{defi}\label{def:even-grd-on-Mmn}
Similarly to Definition \ref{def:explicit-grd-assoc}, we will parametrize the even gradings on $M(m,n)$ as $\Gamma(T,\beta, \gamma_0, \gamma_1)$, where the pair $(T,\beta)$ characterizes $\D$ and $\gamma_0$ and $\gamma_1$ are tuples of elements of $G$ corresponding to the degrees of homogeneous bases for $\mc U\even$ and $\mc U\odd$, respectively. Here $\gamma_0$ is a $k_0$-tuple and $\gamma_1$ is a $k_1$-tuple, with $k_0\sqrt{|T|}=m$ and $k_1\sqrt{|T|}=n$.
\end{defi}
On the other hand, in the case of an odd grading, the information about the canonical $\ZZ_2$-grading is encoded in $\D$. To see that, take a homogeneous $\D$-basis of $\mc U$ and multiply all the odd elements by some nonzero homogeneous element in $\D\odd$. This way we get a homogeneous $\D$-basis of $\mc U$ such that the degrees are all in the subgroup $G$ of $G^\#$. If we denote the $\FF$-span of this new basis by $\widetilde U$, then $E\iso \End (\widetilde U)\tensor \D$ where the first factor has the trivial $\ZZ_2$-grading.
\begin{defi}\label{def:odd-grd-on-Mmn-1}
We parametrize the odd gradings by $\Gamma(T, \beta, \gamma)$ where $T\subseteq G^\#$ but $T\subsetneq G$, the pair $(T,\beta)$ characterizes $\D$, and $\gamma$ is a tuple of elements of $G = G\times \{\bar 0\}$ corresponding to the degrees of a homogeneous basis of $\mc U$ with only even elements.
\end{defi}
Clearly, it is impossible for an even grading to be isomorphic to an odd grading. The classification of even gradings is the following:
\begin{thm}\label{thm:even-assc-iso}
Every even $G$-grading on the superalgebra $M(m,n)$ is isomorphic to some $\Gamma(T,\beta, \gamma_0, \gamma_1)$ as in Definition \ref{def:even-grd-on-Mmn}.
Two even gradings, $\Gamma = \Gamma(T,\beta, \gamma_0, \gamma_1)$ and $\Gamma' = \Gamma(T',\beta', \gamma_0', \gamma_1')$,
are isomorphic if, and only if, $T=T'$, $\beta=\beta'$, and there is $g\in G$ such that
\begin{enumerate}[(i)]
\item for $m\neq n$: $g \Xi(\gamma_0)=\Xi(\gamma_0')$ and $g \Xi(\gamma_1)=\Xi(\gamma_1')$;
\item for $m = n$: either $g \Xi(\gamma_0)=\Xi(\gamma_0')$ and $g \Xi(\gamma_1)=\Xi(\gamma_1')$ or $g\Xi(\gamma_0)=\Xi(\gamma_1')$ and $g \Xi(\gamma_1)=\Xi(\gamma_0')$.
\end{enumerate}
\end{thm}
\begin{proof}
We have already proved the first assertion. For the second assertion, we
consider $\Gamma$ and $\Gamma'$ as $G^\#$-gradings on the algebra $M(m+n)$ and use Theorem \ref{thm:classification-matrix} to conclude that they are isomorphic if, and only if, $T=T'$, $\beta=\beta'$ and there is $(g,s)\in G^\#$ such that $(g,s)\Xi(\gamma)=\Xi(\gamma')$, where $\gamma$ is the concatenation of $\gamma_0$ and $\gamma_1$, where we regard the entries as elements of $G^\# = G\times \ZZ_2$ appending $\barr{0}$ in the second coordinate of the entries of $\gamma_0$ and $\bar 1$ in the second coordinates of the entries of $\gamma_1$.
If $m\neq n$, the condition $(g,s)\Xi(\gamma)=\Xi(\gamma)$ must have $s=\barr0$, since the size of $\gamma_0$ is different from the size of $\gamma_1$.
If $m=n$, the condition $(g,s)\Xi(\gamma)=\Xi(\gamma')$ becomes $g \Xi(\gamma_1)=\Xi(\gamma_1')$ if $s=\barr0$ and $g \Xi(\gamma_1)=\Xi(\gamma_0')$ if $s=\barr1$.
\end{proof}
We now turn to the classification of odd gradings. Recall that here we choose the tuple $\gamma$ to consist of elements of $G$. The corresponding multiset $\Xi(\gamma)$ is contained in $\frac{G^\#}{T} \iso \frac{G}{T \cap G}$.
\begin{thm}\label{thm:first-odd-iso}
Every odd $G$-grading on the superalgebra $M(m,n)$ is isomorphic to some $\Gamma(T,\beta, \gamma)$ as in Definition \ref{def:odd-grd-on-Mmn-1}.
Two odd gradings, $\Gamma = \Gamma(T,\beta, \gamma)$ and $\Gamma' = \Gamma(T',\beta', \gamma')$,
are isomorphic if, and only if, $T=T'$, $\beta=\beta'$, and there is $g\in G$ such that $g \Xi(\gamma)=\Xi(\gamma')$.
\end{thm}
\begin{proof}
We have already proved the first assertion. For the second assertion,
we again consider $\Gamma$ and $\Gamma'$ as $G^\#$-gradings and use Theorem \ref{thm:classification-matrix}: they are isomorphic if, and only if, $T=T'$, $\beta=\beta'$ and there is $(g,s)\in G^\#$ such that $(g,s)\Xi(\gamma)=\Xi(\gamma')$.
Since $T$ contains an element $t_1$ with $p(t_1) = \barr 1$, we may assume $s=\barr 0$.
\end{proof}
In Subsection \ref{subsec:odd-gradings}, we will show that odd gradings can exist only if $m=n$. It may be desirable to express the classification in terms of $G$ rather than $G^\#$ (as we did for even gradings). We will return to this in Subsection \ref{ssec:second-odd}.
\subsection{Even gradings and Morita context.}\label{subsec:even-gradings}
First we observe that every grading on $M(m,n)$ compatible with the $\ZZ$-superalgebra structure is an even grading. This follows from the fact that $T=\supp \D$ is a finite group, and if a finite group is contained in $G\times \ZZ$, then it must be contained in $G\times \{0\}$. Hence, when we look at the corresponding $(G\times\ZZ_2)$-grading, we have that $T\subseteq G$, so no element of $\D$ has an odd degree.
The converse is also true. Actually, we can prove a stronger assertion: if we write $\M(m,n)$ as in Equation \eqref{eq:End_U}, the subspaces given by each of the four blocks are graded. To capture this information, it is convenient to use the concepts of Morita context and Morita algebra.
Recall that a \emph{Morita context} is a sextuple $\mathcal{C} = (R, S, M, N, \vphi, \psi )$ where $R$ and $S$ are unital associative algebras, $M$ is an $(R,S)$-bimodule, $N$ is a $(S,R)$-bimodule and $\vphi: M\tensor_{S} N\rightarrow R$ and $\psi: N\tensor_{R} M\rightarrow S$ are bilinear maps satisfying the necessary and sufficient conditions for
\begin{equation*}
C = \left(\begin{matrix}\label{eq:morita-algebra}
R & M\\
N & S\\
\end{matrix}
\right)
\end{equation*}
to be an associative algebra, \ie,
\[\vphi(m_1\tensor n_1)\cdot m_2 = m_1\cdot \psi(n_1\tensor m_2) \text{ and }\psi(n_1\tensor m_1)\cdot n_2 = n_1\cdot \vphi(m_1\tensor n_2)\]
\noindent for all $m_1,m_2\in M$ and $n_1,n_2\in N$.
We can associate a Morita context to a superspace $U = U\even \oplus U\odd$ by taking $R = \End(U\even)$, $S = \End(U\odd)$, $M = \Hom(U\odd, U\even)$,
$N = \Hom (U\even, U\odd)$, with $\vphi$ and $\psi$ given by composition of operators.
Given an algebra $C$ as above and the idempotent
$
\epsilon = \left(\begin{matrix}
1 & 0\\
0 & 0\\
\end{matrix}
\right)
$,
we can recover all the data of the Morita context (up to isomorphism): $R \iso \epsilon C \epsilon$, $S \iso (1 - \epsilon) C (1 - \epsilon)$, $M \iso \epsilon C (1-\epsilon)$, $N \iso (1-\epsilon) C \epsilon$ and $\phi$ and $\psi$ are given by multiplication in $C$. In other words, the concept of Morita context is equivalent to the concept of \emph{Morita algebra}, which is a pair $(C,\epsilon)$ where $C$ is a unital associative algebra and $\epsilon\in C$ is an idempotent.
For example, we may consider $\M(m,n)$ as a Morita algebra by fixing the idempotent
$
\epsilon = \left(\begin{matrix}
I_m & 0_{m\times n}\\
0_{n\times m} & 0_n\\
\end{matrix}
\right)
$,
\ie, $M(m,n)$ is the Morita algebra corresponding to the Morita context associated to the superspace $U = \FF^m \oplus \FF^n$.
\begin{defi}
A Morita context $(R, S, M, N, \vphi, \psi )$ is said to be $G$-\emph{graded} if the algebras $R$ and $S$ are graded, the bimodules $M$ and $N$ are graded, and the maps $\vphi$ and $\psi$ are homogeneous of degree $e$. A Morita algebra $(C,\epsilon)$ is said to be $G$-\emph{graded} if $C$ is $G$-graded and $\epsilon$ is a homogeneous element (necessarily of degree $e$).
\end{defi}
Clearly, a Morita context is graded if, and only if, the corresponding Morita algebra is graded.
\begin{remark}\label{remarkk}
For every graded Morita algebra $(C,\epsilon)$, we can define a $\ZZ$-grading by taking $C^{-1} = \epsilon C (1-\epsilon )$, $C^0 = \epsilon C \epsilon \oplus (1-\epsilon)C(1-\epsilon)$ and $C^1=(1-\epsilon)C\epsilon$. In the case of $M(m,n)$, this is precisely the canonical $\ZZ$-grading.
\end{remark}
\begin{prop}\label{prop:3-equiv-even-morita-action}
Let $\Gamma$ be a grading on the superalgebra $M(m,n)$. The following are equivalent:
%
\begin{enumerate}[(i)]
\item $\Gamma$ is compatible with the canonical $\ZZ$-grading;
\item $\Gamma$ is even;
\item $M(m,n)$ equipped with $\Gamma$ is a graded Morita algebra.
\vspace{1mm}
\par\vbox{\parbox[t]{\linewidth}{Further, if we assume $\Char\FF=0$, the above statements are also equivalent to:}}
\vspace{1mm}
\item $\Gamma$ corresponds to a $\widehat G$-action by even automorphisms.
\end{enumerate}
\end{prop}
\begin{proof}
~\\ \vspace{-2.5mm}
\textit{(i) $\Rightarrow$ (ii):}
See the beginning of this subsection.
\vspace{2mm}
\textit{(ii) $\Rightarrow$ (iii):}
Regard $\Gamma$ as a $G^\#$-grading. By Theorem \ref{thm:End-over-D}, there is a graded division algebra $\D$ and a graded right $\D$-module $\mc U$ such that $\End_{\mc D} (\mc U) \simeq M(m,n)$. Take an isomorphism of graded algebras $\phi: \End_{\mc D} (\mc U) \rightarrow M(m,n)$. Since $\Gamma$ is even, $\mc U\even$ and $\mc U\odd$ are graded $\mc D$-submodules. Take $\epsilon' \in \End_{\mc D} (\mc U)$ to be the projection onto $\mc U\even$ associated to the decomposition $\mc U= \mc U\even \oplus \mc U\odd$. Clearly, $\epsilon'$ is a central idempotent of $\End_{\mc D} (\mc U)\even$, hence $\phi(\epsilon')$ is a central idempotent of $M(m,n)\even$, so either $\phi(\epsilon')=\epsilon$ or $\phi(\epsilon')=1-\epsilon$. Either way, $\phi^{-1}(\epsilon)$ is homogeneous, hence so is $\epsilon$.
\vspace{2mm}
\textit{(iii) $\Rightarrow$ (i):}
Follows from Remark \ref{remarkk}.
\vspace{2mm}
\textit{(i) $\Leftrightarrow$ (iv):}
This follows from the fact that the group of even automorphisms is precisely the group of automorphisms of the $\ZZ$-superalgebra structure on $M(m,n)$ (see Remark \ref{rmk:Aut-ZZ-superalgebra}).
\end{proof}
\begin{remark}
It follows from Proposition \ref{prop:3-equiv-even-morita-action} that, if $\Char \FF = 0$, odd gradings exist only if $m=n$. In Subsection \ref{subsec:odd-gradings} we will give a characteristic-independent proof of this fact.
\end{remark}
We now know that the gradings on the $\ZZ$-superalgebra $M(m,n)$ are precisely the even gradings, but since the automorphism group is different
from the $\ZZ_2$-superalgebra case, the classification of gradings up to isomorphism is also different.
The proof of the next result is similar to the proof of Theorem \ref{thm:even-assc-iso}.
\begin{thm}
Let $\Gamma(T,\beta,\gamma_0,\gamma_1)$ and $\Gamma'(T',\beta',\gamma_0',\gamma_1')$ be $G$-gradings on the $\ZZ$-superalgebra $M(m,n)$. Then $\Gamma$ and $\Gamma'$ are isomorphic if, and only if, $T=T'$, $\beta=\beta'$, and there is $g\in G$ such that $g\Xi (\gamma_i) = \Xi (\gamma_i')$ for $i=0,1$. \qed
\end{thm}
As we mentioned in Subsection \ref{subsec:graded-bimodules}, we can always shift the grading on a graded (bi)module and still have a graded (bi)module.
In a graded Morita context, as in the case of a graded superalgebra (see Lemma \ref{lemma:opposite-directions}), we have more structure to preserve: if we shift one of the bimodules by an element $g\in G$ and at least one of the bilinear maps is nonzero, then we are forced to shift the other bimodule by $g^{-1}$. As in the superalgebra case, we will refer to this situation as \emph{shift in opposite directions}.
\begin{thm}\label{thm:graded-morita}
Let $\mathcal{C}=(R, S, M, N, \vphi, \psi )$ be the Morita context associated with a superspace $U$ and fix gradings on $R$ and $S$ making them graded algebras. The bimodules $M$ and $N$ admit $G$-gradings so that $\mathcal{C}$ becomes a graded Morita context if, and only if, there exists a graded division algebra $\D$ and graded right $\D$-modules $\mc V$ and $\mc W$ such that $R\iso \End_{\D}(\mc V)$ and $S\iso \End_{\D} (\mc W)$ as graded algebras.
Moreover, all such gradings on $M$ and $N$ have the form $M\iso \Hom_{\D}(\mc W, \mc V)^{[g]}$ and $N\iso \Hom_{\D}(\mc V, \mc W)^{[g^{-1}]}$
as graded bimodules, where $g\in G$ is arbitrary.
\end{thm}
\begin{proof}
Suppose $M$ and $N$ admit $G$-gradings so that the Morita algebra $(C, \epsilon)$ associated to $\mc C$ becomes $G$-graded. By Theorem \ref{thm:End-over-D} there exists a graded division algebra $\D$ and a graded $\D$-module $\mc U$ such that $C \iso \End_{\D} (\mc U)$. Denote the image of $\epsilon$ under this isomorphism by $\epsilon'$ and let $\mc V = \epsilon'(\mc U)$ and $\mc W = (1 -\epsilon')(\mc U)$. Since $\epsilon$ is homogeneous, so is $\epsilon'$, hence $\mc V$ and $\mc W$ are graded $\D$-modules. It follows that $R \iso \epsilon M(m,n) \epsilon \iso \epsilon' \End_{\D} (\mc U) \epsilon' \iso \End_{\D} (\mc V)$ and, analogously, $S \iso \End_{\D} (\mc W)$.
For the converse, write $C$ in matrix form by fixing a basis in $U$ and identify $\End_{\D} (\mc V)$ and $\End_{\D} (\mc W)$ with matrix algebras as in Definition \ref{def:explicit-grd-assoc}. Suppose there exist isomorphisms of graded algebras $\theta_1 \colon R\rightarrow \End_{\D} (\mc V) $ and $\theta_2 \colon S\rightarrow \End_{\D} (\mc W)$. Then there are $x\in \GL (m)$ and $y\in \GL (n)$ such that $\theta_1$ is the conjugation by $x$ and $\theta_2$ is the conjugation by $y$. It follows that the conjugation by
$\begin{pmatrix}
x & 0\\
0 & y
\end{pmatrix}$
\noindent is an isomorphism of algebras between $C\iso M(m,n)$ and
\[\End_{\D} (\mc V \oplus \mc W) =
\begin{pmatrix}
\End_\D(\mc V) & \Hom_\D(\mc W, \mc V)\\
\Hom_\D(\mc V, \mc W) & \End_\D(\mc W)
\end{pmatrix},\]
\noindent hence we transport the gradings on $\Hom_\D(\mc W, \mc V)$ and $\Hom_\D(\mc V, \mc W)$ to $M$ and $N$, respectively.
It remains to prove that the gradings on $M$ and $N$ are determined up to shift in opposite directions. Since in our case the Morita algebra $C$ is simple, $M$ and $N$ are simple bimodules.
By Lemma \ref{lemma:simplebimodule}, the gradings on $M$ and $N$ are determined up to shifts, and the shifts have to be in opposite directions in order for $\vphi$ and $\psi$ to be degree-preserving.
\end{proof}
\subsection{Odd gradings}\label{subsec:odd-gradings}
Let $\Gamma$ be an odd grading on $M(m,n)$. We saw in Subsection \ref{ssec:grds-on-superalgebras} that, as a $G^\#$-graded algebra, $M(m,n)$ is isomorphic to $E\iso \End(\tilde U)\tensor \D$ where the first factor has the trivial $\ZZ_2$-grading and $\D=\D\even\oplus \D\odd$, with $\D\odd\neq 0$, is a $G^\#$-graded division algebra that is simple as an algebra. Let $T\subseteq G^\#$ be the support of $\D$ and $\beta: T\times T \rightarrow \FF^\times$ be the associated bicharacter.
We write $T^+ = \{t\in T \mid p(t)=\barr 0\} = T \cap G$ and $T^- = \{t\in T \mid p(t)=\barr 1\}$,
and denote the restriction of $\beta$ to $T^+\times T^+$ by $\beta^+$.
Note that there are no odd gradings if $\Char \FF =2$. Indeed, in this case, there is no nondegenerate bicharacter on $T$ because the characteristic of the field divides $|T|=2|T^+|$. From now on, we suppose $\Char \FF \neq 2$.
For a subgroup $A\subseteq T$, we denote by $A'$ its orthogonal complement in $T$ with respect to $\beta$, i.e., $A' = \{t\in T\mid \beta(t, A) =1\}$.
This is the inverse image of the subgroup $A^\perp\subseteq \widehat T$ under the isomorphism $T\rightarrow \widehat T$ given by $t\mapsto \beta(t,\cdot)$.
In particular, $|A'| = [T:A]$.
From these considerations, we have $(T^+)' = \langle t_0 \rangle$ where $t_0$ is an element of order 2. It follows that $\beta(t_0, t) = 1$ if $t\in T^+$ and $\beta(t_0, t) = -1$ if $t\in T^-$. For this reason, we call $t_0$ the \emph{parity element} of the odd grading $\Gamma$. Note that $\rad \beta^+ = T^+\cap (T^+)' = \langle t_0 \rangle$.
Fix an element $0\neq d_0\in \D$ of degree $t_0$. By the definition of $\beta$, $d_0$ commutes with all elements of $\D\even$ and anticommutes with all elements of $\D\odd$. Since $d_0^2\in \D_e = \FF$, we may rescale $d_0$ so that $d_0^2=1$. Then $\epsilon := \frac{1}{2}(1+d_0)$ is a central idempotent of $\D\even$. Take a homogeneous element $0\neq d_1\in\D\odd$. Then $d_1\epsilon d_1\inv = \frac{1}{2}(1-d_0)=1-\epsilon$, which is another central idempotent of $\D\even$ and must have the same rank as $\epsilon$.
Hence, $\D\even\iso \epsilon\D\even\oplus (1-\epsilon)\D\even$ (direct sum of ideals) and, consequently, $E\even \iso \End(\tilde U)\tensor \D\even = \End(\tilde U)\tensor \epsilon\D\even \oplus \End(\tilde U)\tensor (1-\epsilon)\D\even$, where the two summands have the same dimension. Therefore, odd gradings exist only if $m=n$.
Also note that we have
\begin{equation}\label{eq:D1eps}
\D\odd \epsilon = (1-\epsilon) \D\odd.
\end{equation}
We are now going construct an even grading by coarsening a given odd grading. The reverse of this construction will be used in Subsection \ref{ssec:second-odd}.
Let $H$ be a group and suppose we have an even grading $\Gamma'$ on $M(n,n)$ that is the coarsening of $\Gamma$ induced by a group homomorphism $\alpha: G\rightarrow H$. Since $\Gamma'$ is even, then the idempotent $\id_{\tilde U}\tensor\epsilon$ must be homogeneous with respect to $\Gamma'$. This means that $\alpha(t_0)=e$, so $\alpha$ factors through $\barr G := G/\langle t_0 \rangle$. This motivates the following definition:
\begin{defi}
Let $\Gamma$ be an odd $G$-grading on $M(n,n)$ with parity element $t_0$. The \emph{finest even coarsening of $\Gamma$} is the $\barr G$-grading ${}^\theta \Gamma$, where $\barr G := G/\langle t_0 \rangle$ and $\theta: G \to \barr G$ is the natural homomorphism.
\end{defi}
\begin{thm}
Let $\Gamma = \Gamma(T, \beta, \gamma)$ be an odd grading on $M(n,n)$ with parity element $t_0$. Then its finest even coarsening is isomorphic to $\barr \Gamma = \Gamma(\barr T, \barr \beta, \barr \gamma, \barr u\barr \gamma)$, where $\barr T= \frac{T^+}{\langle t_0 \rangle}$, $\barr\beta$ is the nondegenerate bicharacter on $\barr T$ induced by $\beta^+$, $\barr\gamma$ is the tuple whose entries are the images of the entries of $\gamma$ under $\theta$, and $u \in G$ is
any element such that $(u, \barr 1) \in T^-$.
\end{thm}
\begin{proof}
Let us focus our attention on the $G$-graded division algebra $\D$. We now consider it as a $\barr G$-graded algebra, which has a decomposition $\D=\D\epsilon \oplus \D(1-\epsilon)$ as a graded left module over itself.
\setcounter{claim}{0}
\begin{claim}
The $\D$-module $\D\epsilon$ is simple as a graded module.
\end{claim}
To see this, consider a nontrivial graded submodule $V\subseteq \D\epsilon$ and take a homogeneous element $0\neq v\in V$. Then we can write $v=d\epsilon$ where $d$ is a $\barr G$-homogeneous element of $\D$, so $d = d' + \lambda d' d_0$ where $d'$ is a $G$-homogeneous element and $\lambda\in \FF$. Hence, $v = d'\epsilon + \lambda d'd_0\epsilon = (1+\lambda)d'\epsilon$. Clearly, $(1+\lambda)d'\neq 0$, so it has an inverse in $\D$. We conclude that $\epsilon\in V$, hence $V=\D\epsilon$.\qedclaim
Let $\barr \D := \epsilon \D \epsilon \iso \End_{\D}(\D\epsilon)$, where we are using the convention of writing endomorphisms of a left module on the right. By Claim 1 and the graded analog of Schur's Lemma (see \eg \cite[Lemma 2.4]{livromicha}), $\barr \D$ is a $\barr G$-graded division algebra.
\begin{claim}
The support of $\barr \D$ is $\barr T= \frac{T^+}{\langle t_0 \rangle}$ and the bicharacter $\barr \beta: \barr T\times \barr T\rightarrow \FF^\times$ is induced by $\beta^+: T^+\times T^+ \rightarrow \FF^\times$.
\end{claim}
We have $\barr \D = \epsilon \D\even \epsilon + \epsilon \D\odd \epsilon$ and $\epsilon \D\odd \epsilon = 0$ by Equation \eqref{eq:D1eps}, so $\supp \barr \D \subseteq \barr T$. On the other hand, for every $0\neq d\in \D\even$ with $G$-degree $t\in T^+$, we have that $\epsilon d\epsilon = d\epsilon = \frac{1}{2}(d+dd_0)\neq 0$, since the component of degree $t$ is different from zero. Hence $\supp \barr \D = \barr T$. Since $\epsilon$ is central in $\D\even$, we obtain $\barr\beta (\barr t,\barr s) = \beta (s, t) = \beta^+ (s, t)$ for all $t, s\in T^+$.\qedclaim
We now consider $\D\epsilon$ as a graded right $\barr \D$-module. Then we have the decomposition $\D\epsilon = \epsilon \D\epsilon \oplus (1-\epsilon) \D\epsilon$. The set $\{\epsilon\}$ is clearly a basis of $\epsilon \D\epsilon$. To find a basis for $(1-\epsilon)\D\epsilon$, fix any $G$-homogeneous $0\neq d_1\in \D\odd$ with $\deg d_1 = t_1\in T^-$. Then we have $(1-\epsilon)\D\epsilon = (1-\epsilon)\D\even \epsilon + (1-\epsilon)\D\odd \epsilon = (1-\epsilon)\D\odd \epsilon = \D\odd \epsilon$ by Equation \eqref{eq:D1eps}. Since $d_1$ is invertible, $\{d_1\epsilon\}$ is a basis for $(1-\epsilon) \D\epsilon$. We conclude that $\{\epsilon, d_1\epsilon\}$ is a basis for $\D\epsilon$.
Using the graded analog of the Density Theorem (see e.g. \cite[Theorem 2.5]{livromicha}), we have $\D\iso \End_{\barr \D}(\D\epsilon)\iso \End(\FF\epsilon\oplus \FF d_1\epsilon)\tensor \barr\D$. Hence,
%
\[
\begin{split}
\End_\D(\mc U)&\iso\End (\tilde U) \tensor \D \iso \End (\tilde U) \tensor \End(\FF\epsilon \oplus \FF d_1\epsilon) \tensor \barr\D \\
&\iso \End(\tilde U\tensor \epsilon \oplus \tilde U\tensor d_1\epsilon) \tensor \barr\D
\end{split}
\]
%
as $\barr G$-graded algebras. The result follows.
\end{proof}
In the next section, we will show how to recover $\Gamma$ from $\barr\Gamma$ and some extra data. The following definition and result will be used there.
\begin{defi}
For every abelian group $A$ we put $A^{[2]} = \{a^2 \mid a\in A\}$ and $A_{[2]} = \{a\in A \mid a^2 = e \}$.
\end{defi}
Note that $T^{[2]}\subseteq T^+$, but $T^{[2]}$ can be larger than $(T^+)^{[2]}$ since it also includes the squares of elements of $T^-$. Also, the subgroup $\barr S = \{\barr t \in \barr T \mid t \in T^{[2]}\}$ of $\barr T$ can be larger than $\barr T^{[2]}$, but we will show that, surprisingly, it does not depend on $T^-$.
\begin{lemma}\label{lemma:square-subgroup}
Let $\theta: T^+\rightarrow \barr T=\frac{T^+}{\langle t_0 \rangle}$ be the natural homomorphism.
Consider the subgroups $\barr S = \theta(T^{[2]})$ and $\barr R=\theta(T^+_{[2]})$ of $\barr T$.
Then $\barr S$ is the orthogonal complement of $\barr R$ with respect to the nondegenerate bicharacter $\barr\beta$.
\end{lemma}
\begin{proof}
We claim that $\barr S' = \barr R$. Indeed,
%
\[
\begin{split}
\barr S' & = \{ \theta(t) \mid t\in T^+ \AND \barr\beta(\theta (t), \theta (s^2)) =1 \text{ for all }s\in T\}\\ & = \{\theta(t) \mid t\in T^+ \AND \beta (t, s^2) =1 \text{ for all }s\in T\}\\ & = \{ \theta(t) \mid t\in T^+ \AND \beta (t^2, s) =1 \text{ for all }s\in T\}\\ & = \{ \theta(t) \mid t\in T^+ \AND t^2=e \}\\ & = \barr R\,.
\end{split}
\]
It follows that $\barr S = \barr R'$, as desired.
\end{proof}
\subsection{A description of odd gradings in terms of $G$}\label{ssec:second-odd}
Our second description of an odd grading consists of its finest even coarsening and the data necessary to recover the odd grading from this coarsening. All parameters will be obtained in terms of $G$ rather than its extension $G^\#=G\times \ZZ_2$.
Let $t_0\in G$ be an arbitrarily fixed element of order 2 and set $\barr G = \frac{G}{\langle t_0 \rangle}$. Let $\barr T \subseteq \barr G$ be a finite subgroup and let $\barr \beta: \barr T \times \barr T \rightarrow \FF^\times$ be a nondegenerate alternating bicharacter. We define $T^+\subseteq G$ to be the inverse image of $\barr T$ under the natural homomorphism $\theta: G\rightarrow \barr G$. Note that $\barr \beta$ gives rise to a bicharacter $\beta^+$ on $T^+$ whose radical is generated by the element $t_0$. We wish to define $T^-\subseteq G\times \{\barr 1\}$ so that $T=T^+\cup T^-$ is a subgroup of $G^\#$ and $\beta^+$ extends to a nondegenerate alternating bicharacter on $T$.
From Lemma \ref{lemma:square-subgroup}, we have a necessary condition for the existence of such $T^-$, namely, for $\barr R=\frac{T^+_{[2]}}{\langle t_0 \rangle}$, we need $\barr R' \subseteq \barr G^{[2]}$ (indeed, $\barr S$ is a subgroup of $\overline {G^{[2]}} = \barr G^{[2]}$). We will now prove that this condition is also sufficient.
\begin{prop}\label{prop:square-subgroup-converse}
If $\left( \frac{T^+_{[2]} }{\langle t_0 \rangle}\right)'\subseteq \barr G^{[2]}$, then there exists an element $t_1\in G\times \{\barr 1\} \subseteq G^\#$ such that $T= T^+ \cup t_1\, T^+$ is a subgroup of $G^\#$ and $\beta^+$ extends to a nondegenerate alternating bicharacter $\beta:T\times T\rightarrow \FF^\times$.
\end{prop}
\begin{proof}
Let $\chi\in \widehat {T^+}$ be such that $\chi(t_0) = -1$. Since $\chi^2(t_0)=1$, we can consider $\chi^2$ as a character of the group $\barr T = \frac{T^+}{\langle t_0 \rangle}$, hence there is $a\in T^+$ such that $\chi^2(\barr t) = \barr\beta(\barr a, \barr t)$ for all $\barr t\in \barr T$. Note that $\chi (a) = \pm 1$ and hence, changing $a$ to $a t_0$ if necessary, we may assume $\chi (a) = 1$.
\bigskip
\textit{(i) Existence of $t_1$}:
\medskip
As before, let $\barr R = \frac{T^+_{[2]}}{\langle t_0 \rangle}$. Then $\barr a \in \barr R'$. Indeed, if $b\in T^+_{[2]}$, then $\barr\beta(\barr a,\barr b) = \chi^2 (\barr b) = \chi (b^2) = \chi (e) =1$. By our assumption, we conclude that $\barr a\in \barr G^{[2]}$. We are going to prove that, actually, $a\in G^{[2]}$. Pick $u\in G$ such that $\barr u^2 = \barr a$. Then, either $a=u^2$ or $a=u^2t_0$. If $t_0 = c^2$ for some $c\in G$, then replacing $u$ by $uc$ if necessary, we can make $u^2 = a$. Otherwise, $t_0$ has no square root in $T^+$, which implies that $\barr R=\barr T_{[2]}$. Hence $\barr R' = (\barr T_{[2]})' = \barr T^{[2]} = \theta ((T^+)^{[2]})$. Thus, in this case, we can assume $u\in T^+$. Then $\chi(u^2) = \chi^2(u) = \barr \beta (\barr a, \barr u) = \barr \beta (\barr u^2, \barr u) =1$, hence $u^2 = a$. Finally, we set $t_1=(u,\barr 1) \in G^\#$.
\bigskip
\textit{(ii) Existence of $\beta$}:
\medskip
We wish to extend $\beta^+$ to $T=T^+ \cup t_1\, T^+$ by setting $\beta(t_1, t) = \chi (t)$ for all $t\in T^+$. It is clear that there is at most one alternating bicharacter on $T$ with this property that extends $\beta^+$. To show that it exists and is nondegenerate, we will first introduce an auxiliary group $\widetilde T$ and a bicharacter $\tilde\beta$.
Let $\widetilde T$ be the direct product of $T^+$ and the infinite cyclic group generated by a new symbol $\tau$. We define $\tilde\beta:\widetilde T\times \widetilde T \rightarrow \FF^\times$ by $ \tilde\beta(s\tau^i,t\tau^j) = \beta^+(s,t)\, \chi (s)^{-j}\, \chi (t)^i$, where $s,t\in T^+$. It is clear that $\tilde\beta$ is an alternating bicharacter.
\begin{claim*}
$\langle a\tau^{-2} \rangle = \rad \tilde \beta\,$.
\end{claim*}
Let $t\in T^+$ and $\ell\in \ZZ$. Then
\[
\tilde \beta (a\tau^{-2},t\tau^\ell) =
\beta^+(a, t)\,\, \chi(t)^{-2} \, \chi(a)^{-\ell} =
\barr\beta(\barr a, \barr t)\,\, \chi(t)^{-2} = \chi(t)^2 \, \chi(t)^{-2} = 1,
\]
hence, $\langle a\tau^{-2} \rangle \subseteq \rad \tilde \beta$.
Conversely, if $s\tau^k \in \rad \tilde\beta$, then, $1 = \tilde \beta (s\tau^k, t_0) = \beta^+(s,t_0)\, \chi(t_0)^k = (-1)^k$, hence $k$ is even. From the previous paragraph, we know that $a\tau^{-2} \in \rad \tilde\beta$, hence $a^\frac{k}{2} \tau^{-k} \in \rad \tilde\beta$ and $s a^\frac{k}{2} = (s \tau^k) (a^\frac{k}{2} \tau^{-k}) \in \rad \tilde\beta$. Since $s a^\frac{k}{2} \in T^+$, we get $s a^\frac{k}{2} \in \rad \beta^+ = \{ e, t_0 \}$. But, if $sa^\frac{k}{2} = t_0$, we have $1 = \tilde\beta (sa^\frac{k}{2}, \tau) = \tilde\beta (t_0, \tau) = \chi(t_0)\inv = -1$, a contradiction. It follows that $sa^\frac{k}{2} = e$ and, hence, $s\tau^k = a^{-\frac{k}{2}}\tau^k = (a\tau^{-2})^{\frac{k}{2}}$, concluding the proof of the claim.
\qedclaim
We have a homomorphism $\vphi:\widetilde T\rightarrow T$ that is the identity on $T^+$ and sends $\tau$ to $t_1$. Clearly, $\ker \vphi = \langle a\tau^{-2} \rangle$. By the above claim, $\tilde\beta$ induces a nondegenerate alternating bicharacter on $\frac{\widetilde T}{\langle a\tau^{-2} \rangle}$, which can be transferred via $\vphi$ to a nondegenerate alternating bicharacter on $T$ that extends $\beta^+$.
\end{proof}
Now fix $\chi\in \widehat {T^+}$ with $\chi(t_0)=-1$ and let $a$ be the unique element of $T^+$ such that $\chi(a)=1$ and $\chi^2(\barr t) = \barr\beta (\barr a, \barr t)$ for all $t\in T^+$. Suppose that the condition of Proposition \ref{prop:square-subgroup-converse} is satisfied.
Then part (i) of the proof shows that there exists $u\in G$ such that $u^2=a$. Moreover, part (ii) shows that there exists an extension of $\beta^+$ to
a nondegenerate alternating bicharacter $\beta$ on $T=T^+\cup t_1T^+$, where $t_1=(u,\bar 1)$, such that $\beta(t_1,t)=\chi(t)$ for all $t\in T^+$.
Clearly, such an extension is unique. We will denote it by $\beta_u$ and its domain by $T_u$.
\begin{prop}\label{prop:roots-of-a}
For every $T\subseteq G^\#$ such that $T\subsetneq G$ and $T\cap G=T^+$ and for every extension of $\beta^+$ to a nondegenerate alternating bicharacter $\beta$
on $T$, there exists $u\in G$ such that $u^2=a$, $T=T_u$ and $\beta=\beta_u$.
Moreover, $\beta_u=\beta_{\tilde{u}}$ if, and only if, $u \equiv \tilde{u} \pmod{\langle t_0 \rangle}$.
\end{prop}
\begin{proof}
We have $T=T^+ \cup T^-$ where $T^-\subseteq G\times \{\barr 1\}$ is a coset of $T^+$.
We can extend $\chi$ to a character of $T$, which we still denote by $\chi$, and, since $\beta$ is nondegenerate,
there is $t_1\in T$ such that $\beta(t_1, t) = \chi(t)$ for all $t\in T$. We have $t_1\in T^-$ since $\beta(t_1,t_0)=\chi(t_0)=-1$, so $t_1=(u,\bar 1)$,
for some $u\in G$, and hence $T=T_u$. We claim that $t_1^2=a$. Indeed, $\chi(t_1^2) = \beta(t_1,t_1^2)=1$ and, for every $t\in T^+$,
\[
\chi^2(\barr t) = \chi(t)^2 = \beta (t_1, t)^2 = \beta (t_1^2, t) = \barr\beta (\,\overline {(t_1^2)},\, \barr t)\,,
\]
so $t_1^2$ satisfies the definition of the element $a$. This completes the proof of the first assertion.
Now suppose $\beta_u=\beta_{\tilde{u}}$, so in particular $t_1\,T^+=\tilde{t}_1\,T^+$ where $t_1 = (u, \barr 1)$ and $\tilde{t}_1 = (\tilde u, \barr 1)$.
Then there is $r\in T^+$ such that $\tilde{t}_1 = t_1\,r$. Also, for every $t\in T^+$,
\[
\chi(t) = \beta_{\tilde{u}}(\tilde{t}_1,t) = \beta_u (t_1\,r, t)
= \beta_u(t_1, t)\,\beta_u(r,t) = \chi(t) \beta^+(r, t)
\]
and, hence, $\beta^+(r, t)=1$ for all $t\in T^+$. This means that $r = u\inv \tilde{u} \in \langle t_0 \rangle$.
Conversely, if $\tilde u = u r$ for some $r\in \langle t_0 \rangle$, then $t_1\, T^+ = \tilde t_1\, T^+$. Also, for all $t\in T^+$,
\[
\beta_u(t_1, t) = \chi(t) = \beta_{\tilde{u}}(\tilde{t}_1, t) = \beta_{\tilde{u}}(t_1r, t) =
\beta_{\tilde{u}}(t_1, t)\, \beta^+(r, t) = \beta_{\tilde{u}}(t_1, t).
\]
It follows that $\beta_u=\beta_{\tilde{u}}$.
\end{proof}
Note that, keeping the character $\chi \in \widehat {T^+}$ with $\chi(t_0) = -1$ fixed, we have a surjective map from the square roots of $a$ to all possible pairs $(T,\beta)$. If we had started with a different character above, we would have obtained a different surjective map. Hence, for parametrization purposes, $\chi$ (and, hence, $a$) will be fixed.
We are now in a position to give a classification of odd gradings in terms of $G$ only. We already have the following parameters: an element $t_0\in G$ of order $2$ and a pair $(\barr T, \barr\beta)$. For each $t_0$ and $\barr T$, we fix a character $\chi\in \widehat {T^+}$ satisfying $\chi(t_0) = -1$. The next parameter is an element $u\in G$ such that $u^2 = a$, where $a$ is the unique element of $T^+$ such that $\chi(a)=1$ and $\chi^2(\barr t) = \barr\beta (\barr a, \barr t)$ for all $t\in T^+$. Finally, let $\gamma = (g_1, \ldots, g_k)$ be a $k$-tuple of elements of $G$. With these data, we construct the grading $\Gamma (t_0, \barr T, \barr \beta, u, \gamma)$ as follows:
\begin{defi}\label{def:odd-grd-on-Mmn-2}
Let $\D$ be a standard realization of the $G^\#$-graded division algebra with parameters $(T_u,\beta_u)$. Take the graded $\D$-module $\mathcal U = \D^{[g_1]}\oplus \cdots \oplus \D^{[g_k]}$. Then $\End_\D (\mathcal U)$ is a $G^\#$-graded algebra, hence a superalgebra by means of $p:G^\# \rightarrow \ZZ_2$. As a superalgebra, it is isomorphic to $M(n,n)$ where $n=k\sqrt{|\barr T|}$. We define $\Gamma (t_0, \barr T, \barr \beta, u, \gamma)$ as the corresponding $G$-grading on $M(n,n)$.
\end{defi}
Theorem \ref{thm:first-odd-iso} together with Proposition \ref{prop:roots-of-a} give the following result:
\begin{thm}\label{thm:2nd-odd-iso}
Every odd $G$-grading on the superalgebra $M(n,n)$ is isomorphic to some $\Gamma (t_0, \barr T, \barr \beta, u, \gamma)$ as in Definition \ref{def:odd-grd-on-Mmn-2}.
Two odd gradings, $\Gamma (t_0, \barr T, \barr \beta, u, \gamma)$ and $\Gamma (t_0', \barr T', \barr \beta', u', \gamma')$,
are isomorphic if, and only if, $t_0=t_0'$, $\barr T = \barr T'$, $\barr\beta = \barr\beta'$, $u \equiv u' \pmod{\langle t_0 \rangle}$,
and there is $g\in G$ such that $g\, \Xi(\gamma) = \Xi(\gamma')$.\qed
\end{thm}
\subsection{Fine gradings up to equivalence}
We start by investigating the gradings on the superalgebra $M(m,n)$ that are fine among even gradings.
By Proposition \ref{prop:3-equiv-even-morita-action}, this is the same as fine gradings on $M(m,n)$ as a $\ZZ$-superalgebra, and,
by the discussion in Subsection \ref{subsec:odd-gradings}, the same as fine gradings if $m\neq n$ or $\Char \FF = 2$.
We will use the following notation. Let $H$ be a finite abelian group whose order is not divisible by $\Char \FF$. Set $T_H = H\times \widehat H$
and define $\beta_H: T_H\times T_H \to \FF^\times$ by
\[
\beta_H((h_1, \chi_1), (h_2, \chi_2)) = \chi_1(h_2)\, \chi_2 (h_1)\inv.
\]
Then $\beta_H$ is a nondegenerate alternating bicharacter on $T_H$.
\begin{defi}\label{def:even-fine-grd-on-Mmn}
Let $\ell \mid \operatorname{gcd}(m,n)$ be a natural number such that $\Char \FF\nmid\ell$ and put $k_0 := \frac{m}{\ell}$ and $k_1 := \frac{n}{\ell}$.
Let $\Theta_\ell$ be a set of representatives of the isomorphism classes of abelian groups of order $\ell$.
For every $H$ in $\Theta_\ell$, we define $\Gamma(H, k_0, k_1)$ to be the even $T_H\times \ZZ^{k_0 + k_1}$-grading
$\Gamma(T_H, \beta_H, (e_1, \ldots, e_{k_0}), (e_{k_0 + 1}, \dots, e_{k_0 + k_1}))$ on $M(m,n)$,
where $\{e_1, \ldots, e_{k_0+k_1}\}$ is the standard basis of $\ZZ^{k_0 + k_1}$.
If $m$ and $n$ are clear from the context, we will simply write $\Gamma(H)$.
\end{defi}
Let $G_H$ be the subgroup of $T_H\times \ZZ^{k_0 + k_1}$ generated by the support of $\Gamma(H,k_0,k_1)$, i.e.,
$G_H = T_H\times \ZZ^{k_0 + k_1}_0$, where $\ZZ^k_0 := \{ (x_1, \ldots, x_k) \in \ZZ^k \mid x_1 + \cdots + x_k = 0\}$.
\begin{thm}\label{thm:class-fine-even}
The fine gradings on $M(m,n)$ as a $\ZZ$-superalgebra are precisely the even fine gradings.
Every such grading is equivalent to a unique $\Gamma(H)$ as in Definition \ref{def:even-fine-grd-on-Mmn}.
Moreover, every grading $\Gamma(H)$ is fine, and $G_H$ is its universal group.
\end{thm}
\begin{proof}
By \cite[Proposition 2.35]{livromicha}, if we consider $\Gamma(H)$ as a grading on the algebra $M_{n+m}(\FF)$, it is a fine grading and
$G_H$ is its universal group. It follows that the same is true of $\Gamma(H)$ as a grading on the superalgebra $M(m,n)$.
Let $\Gamma = \Gamma(T, \beta, \gamma_0, \gamma_1)$ be any even $G$-grading on $M(m,n)$. We can write $T = A\times B$ where the restrictions of $\beta$ to
the subgroups $A$ and $B$ are trivial and, hence, there is an isomorphism $\alpha: T_A \to T$ such that $\beta_A=\beta\circ(\alpha\times\alpha)$. We can extend $\alpha$ to a homomorphism $G_A \to G$ (also denoted by $\alpha$) by sending the elements $e_1, \ldots, e_{k_0}$ to the entries of $\gamma_0$ and the elements $e_{k_0+1}, \ldots, e_{k_0+k_1}$ to the entries of $\gamma_1$. It follows that ${}^\alpha \Gamma(A) \iso \Gamma$. Since all $\Gamma(H)$ are fine and pairwise nonequivalent (because their universal groups are pairwise nonisomorphic), we can apply Lemma \ref{lemma:universal-grp}, concluding that every fine grading
on $M(m,n)$ as a $\ZZ$-superalgebra is equivalent to a unique $\Gamma(H)$.
\end{proof}
We now consider odd fine gradings on $M(n,n)$, so $\Char\FF\ne 2$. We first define some gradings on the algebra $M_{2n}(\FF)$ and then impose a superalgebra structure.
\begin{defi}\label{def:param-fine-odd}
Let $\ell\mid n$ be a natural number such that $\Char \FF\nmid\ell$ and put $k:= \frac{n}{\ell}$.
Let $\Theta_{2\ell}$ be a set of representatives of the isomorphism classes of abelian groups of order $2\ell$.
For every $H$ in $\Theta_{2\ell}$, we consider the $T_H\times \ZZ^k$-grading $\Gamma = \Gamma(T_H, \beta_H, (e_1, \ldots, e_k))$ on $M_{2n}(\FF)$,
where $\{e_1, \ldots, e_k\}$ is the standard basis of $\ZZ^k$.
Then we choose an element $t_0 \in T$ of order $2$ and define a group homomorphism $p: T_H\times\ZZ^k \to \ZZ_2$ by
\[
p(t, x_1, \ldots, x_k) =
\begin{cases*}
\bar 0 & if $\beta(t_0, t) = 1$,\\
\bar 1 & if $\beta(t_0, t) = -1$.
\end{cases*}
\]
This defines a superalgebra structure on $M_{2n}(\FF)$. By construction, $\Gamma$ is odd as a grading on this superalgebra $(M_{2n}(\FF),p)$,
and this forces the superalgebra to be isomorphic to $M(n,n)$. We denote by $\Gamma(H, t_0, k)$ the grading $\Gamma$ considered as a grading on $M(n,n)$.
If $n$ is clear from the context, we will simply write $\Gamma(H, t_0)$.
\end{defi}
Note that the parameter $t_0$ of $\Gamma(H, t_0, k)$ does not affect the grading on the algebra $M_{2n}(\FF)$, but, as we will see in Proposition \ref{prop:equiv-with-same-H}, different choices of $t_0$ can yield nonequivalent gradings on the superalgebra $M(n,n)$.
\begin{prop}\label{prop:all-fine-odd}
Each grading $\Gamma(H, t_0)$ on $M(n,n)$ is fine, and its universal group is $G_H = T_H\times \ZZ^k_0$.
Every odd fine grading on $M(n,n)$ is equivalent to at least one $\Gamma(H, t_0)$.
\end{prop}
\begin{proof}
As in the proof of Theorem \ref{thm:class-fine-even}, the first assertion follows from \cite[Proposition 2.35]{livromicha}.
Let $\Gamma(T,\beta, \gamma)$ be an odd $G$-grading on $M(n,n)$ and let $t_0$ be its parity element.
Then we can find subgroups $A$ and $B$ such that $T=A\times B$ and there exists an isomorphism $\alpha: T_A \to T$ such that $\beta_A=\beta\circ(\alpha\times\alpha)$. We define $t_0' := \alpha\inv(t_0)$ and extend $\alpha$ to a homomorphism $G_A \to G$ (also denoted by $\alpha$)
by sending the elements $e_1, \ldots, e_k$ to the entries of $\gamma$. Then ${}^\alpha \Gamma(A, t_0') \iso \Gamma$.
Selecting a representative from each equivalence class of gradings of the form $\Gamma(H, t_0)$, we can apply Lemma \ref{lemma:universal-grp},
which proves the second assertion.
\end{proof}
It remains to determine which of the gradings $\Gamma(H, t_0)$ are equivalent to each other.
\begin{prop}\label{prop:equiv-with-same-H}
The gradings $\Gamma = \Gamma(H, t_0)$ and $\Gamma' = \Gamma(H, t_0')$ on $M(n,n)$ are equivalent if, and only if, there is $\alpha \in \Aut(T_H, \beta_H)$ such that $\alpha(t_0) = t_0'$.
\end{prop}
\begin{proof}
We will denote by $p: G_H \to \ZZ_2$ the parity homomorphism associated to the grading $\Gamma$ and by
$p': G_H\to \ZZ_2$ the one associated to $\Gamma'$.
If $\Gamma$ is equivalent to $\Gamma'$, there is an isomorphism $\vphi: (M_{2n}(\FF),p) \to (M_{2n}(\FF),p')$ of superalgebras that is a
self-equivalence of the grading on $M_{2n}(\FF)$. Hence, we have the corresponding group automorphism $\alpha: G_H \to G_H$ in the Weyl group of the grading,
and the following diagram commutes:
%
\begin{equation}\label{diag:parity}
\begin{tikzcd}
G_H \arrow[to = 2G, "\alpha"] \arrow[to = Z2, "p"'] && |[alias = 2G]| G_H \arrow[to = Z2, "p'"]\\
& |[alias = Z2]|\ZZ_2 &
\end{tikzcd}
\end{equation}
%
By the definition of $p$ and $p'$, this is equivalent to $\alpha(t_0) = \alpha(t_0')$.
The automorphism $\alpha$ must send the torsion subgroup of $G_H$ to itself, so we can consider the restriction $\alpha\restriction_{T_H}$.
By \cite[Corrolary 2.45]{livromicha}, this restriction is in $\Aut(T_H, \beta_H)$.
For the converse, we use the same \cite[Corrolary 2.45]{livromicha} to extend $\alpha$ to an automorphism $G_H\to G_H$ in the Weyl group.
Hence, there is an automorphism $\vphi$ of the algebra $M_{2n}(\FF)$ that permutes the components of the grading according to $\alpha$.
The condition $\alpha(t_0) = \alpha(t_0')$ is equivalent to Diagram \eqref{diag:parity} being commutative, which shows that $\vphi: (M_{2n}(\FF),p) \to (M_{2n}(\FF),p')$ is an isomorphism of superalgebras.
\end{proof}
Combining Propositions \ref{prop:all-fine-odd} and \ref{prop:equiv-with-same-H}, we obtain:
\begin{thm}\label{thm:class-fine-odd}
Every odd fine grading on $M(n,n)$ is equivalent to some $\Gamma(H,t_0)$ as in Definition \ref{def:param-fine-odd}.
Every grading $\Gamma(H,t_0)$ is fine, and $G_H$ is its universal group.
Two gradings, $\Gamma(H, t_0)$ and $\Gamma(H', t_0')$, are equivalent if, and only if, $H=H'$ and $t_0'$ lies in the orbit of $t_0$
under the natural action of $\Aut(T_H, \beta_H)$.\qed
\end{thm}
For a matrix description of the group $\Aut(T_H, \beta_H)$, we refer the reader to \cite[Remark 2.46]{livromicha}.
\section{Gradings on $A(m,n)$}\label{sec:Amn}
Throughout this section it will be assumed that $\Char \FF = 0$.
\subsection{The Lie superalgebra $A(m,n)$}\label{ssec:def-A}
Let $U = U\even \oplus U\odd$ be a finite dimentional superspace. Recall that the \emph{general linear Lie superalgebra}, denoted by $\gl\, (U)$, is the superspace $\End(U)$ with product given by the \emph{supercommutator}:
\[
[a,b] = ab - (-1)^{\abs{a}\abs{b}}ba.
\]
If $U\even=\FF^m$ and $U\odd=\FF^n$, then $\gl (U)$ is also denoted by $\gl(m|n)$.
The \emph{special linear Lie superalgebra}, denoted by $\Sl (U)$, is the derived algebra of $\gl(U)$. As in the Lie algebra case, we describe it as an algebra of ``traceless'' operators. The analog of trace in the ``super'' setting is the so called \emph{supertrace}:%
\[ \str \left(\begin{matrix}
a & b\\
c & d\\
\end{matrix}\right) = \tr a - \tr d,
\]
and we have $\Sl (U) = \{ T\in \gl (U) \mid \Str T = 0\}$.
Again, if $U\even=\FF^m$ and $U\odd=\FF^n$ then $\Sl (U)$ is also denoted by $\Sl(m|n)$.
If one of the parameters $m$ or $n$ is zero, we get a Lie algebra, so we assume this is not the case. If $m\neq n$ then $\Sl(m|n)$ is a simple Lie superalgebra. If $m=n$, the identity map $I_{2n}\in \Sl(n|n)$ is a central element and hence $\Sl(n|n)$ is not simple, but the quotient $\mathfrak{psl}(n|n) := \Sl(n|n)/ \FF I_{2n}$ is simple if $n>1$.
For $m$,$n\geq 0$ (not both zero), the simple Lie superalgebra $A(m,n)$ is $\Sl(m+1|n+1)$ if $m\neq n$, and $\mathfrak{psl}(n+1|n+1)$ if $m=n$.
\begin{defi}\label{def:Type-I}
If $\Gamma$ is a $G$-grading on $M(m,n)$, then, since $G$ is abelian, it is also a grading on $\gl(m|n)$ and, hence, restricts to its derived superalgebra $\Sl(m|n)$. If $m=n$, then the grading on $\Sl(m|n)$ induces a grading on $\mathfrak{psl}(n|n)$.
If a grading on $\Sl(m|n)$ or $\mathfrak{psl}(n|n)$ is obtained in this way, we will call it a \emph{Type I} grading and, otherwise, a \emph{Type II} grading.
\end{defi}
\subsection{Automorphisms of $A(m,n)$}\label{ssec:auto-Amn}
As in the Lie algebra case, the group of automorphisms of the Lie superalgebra $A(m,n)$ is bigger than the group of automorphisms of the associative superalgebra $\End(U)$.
We define the \emph{supertranspose} of a matrix in $\End(U)$ by
\begin{equation*}
\left( \begin{matrix}
a&b\\
c&d\\
\end{matrix}\right)^{s\top} = \left( \begin{matrix}
a\transp & -c\transp\\
b\transp & d\transp\\
\end{matrix}\right).
\end{equation*}
The supertranspose map $\End(U) \to \End(U)$ is an example of a \emph{super-anti-automorphism}, \ie, it is $\FF$-linear and
\[
(XY)^{s\top} = (-1)^{|X||Y|} Y^{s\top} X^{s\top}.
\]
Hence, the map $\tau:\Sl(m+1,n+1)\rightarrow \Sl(m+1,n+1)$ given by $\tau(X) = - X^{s\top}$ is an automorphism.
By \cite[Theorem 1]{serganova}, the group of automorphisms of $A(m,n)$ is generated by $\tau$ and the automorphisms of $\End(U)$, which are restricted to traceless operators and, if necessary, taken modulo the center. In other words, if $m\neq n$, $\Aut(A(m,n))$ is generated by $\mc E \cup \{\tau\}$ and, if $m=n$, by $\mc E \cup \{\pi\,,\,\tau\}$. In both cases, $\mc E$ is a normal subgroup of $\Aut(A(m,n))$. Note that $\pi^2 = \id$, $\tau^2=\upsilon$ (the parity automorphism) and $\pi \tau = \upsilon \tau \pi$. Hence $\frac{\Aut(A)}{\mc E}$ is isomorphic to $\ZZ_{2}$ if $m\neq n$ and $\ZZ_2\times\ZZ_2$ if $m=n$.
Note that a $G$-grading on $A(m,n)$ is of Type I if, and only if, it corresponds to a $\widehat{G}$-action on $A(m,n)$ by automorphisms that belong to
the subgroup $\mc E$ if $m\ne n$ and to $\mc E\rtimes\langle\pi\rangle$ if $m=n$. If $\widehat{G}$ acts by automorphisms that belong to $\mc E$ then the
Type I grading is said to be \emph{even} and, otherwise, \emph{odd}.
\subsection{Superdual of a graded module}\label{ssec:superdual}
We will need the following concepts. Let $\D$ be an associative superalgebra with a grading by an abelian group $G$, so we may consider $\D$ graded by the group $G^\# = G\times \ZZ_2$. Let $\U$ be a $G^\#$-graded \emph{right} $\D$-module. The parity $|x|$ of a homogeneous element $x \in \D$ or $x\in \U$ is determined by $\deg x \in G^\#$. The \emph{superdual module} of $\U$ is $\U\Star = \Hom_\D (\U,\D)$, with its natural $G^\#$-grading and the $\D$-action defined on the \emph{left}: if $d \in \D$ and $f \in \U \Star$, then $(df)(u) = d\, f(u)$ for all $u\in \mc U$.
We define the \emph{opposite superalgebra} of $\D$, denoted by $\D\sop$, to be the same graded superspace $\D$, but with a new product $a*b = (-1)^{|a||b|} ba$ for every pair of $\ZZ_2$-homogeneous elements $a,b \in \D$. The left $\D$-module $\U\Star$ can be considered as a right $\D\sop$-module by means of the action defined by $f\cdot d := (-1)^{|d||f|} df$, for every $\ZZ_2$-homogeneous $d\in \D$ and $f\in \U\Star$.
\begin{lemma}\label{lemma:Dsop}
If $\D$ is a graded division superalgebra associated to the pair $(T,\beta)$, then $\D\sop$ is associated to the pair $(T,\beta\inv)$.\qed
\end{lemma}
If $\U$ has a homogeneous $\D$-basis $\mc B = \{e_1, \ldots, e_k\}$, we can consider its \emph{superdual basis} $\mc B\Star = \{e_1\Star, \ldots, e_k\Star\}$ in $\U\Star$, where $e_i\Star : \U \rightarrow \D$ is defined by $e_i\Star (e_j) = (-1)^{|e_i||e_j|} \delta_{ij}$.
\begin{remark}\label{rmk:gamma-inv}
The superdual basis is a homogeneous basis of $\U\Star$, with $\deg e_i\Star = (\deg e_i)\inv$. So, if $\gamma = (g_1, \ldots, g_k)$ is the $k$-tuple of degrees of $\mc B$, then $\gamma\inv = (g_1\inv, \ldots, g_k\inv)$ is the $k$-tuple of degrees of $\mc B\Star$.
\end{remark}
For graded right $\D$-modules $\U$ and $\V$, we consider $\U\Star$ and $\V\Star$ as right $\D\sop$-modules as defined above. If $L:\U \rightarrow \V$ is a $\ZZ_2$-homogeneous $\D$-linear map, then the \emph{superadjoint} of $L$ is the $\D\sop$-linear map $L\Star: \V\Star \rightarrow \U\Star$ defined by $L\Star (f) = (-1)^{|L||f|} f \circ L$. We extend the definition of superadjoint to any map in $\Hom_\D (\U, \V)$ by linearity.
\begin{remark}
In the case $\D=\FF$, if we denote by $[L]$ the matrix of $L$ with respect to the homogeneous bases $\mc B$ of $\U$ and $\mc C$ of $\V$, then the supertranspose $[L]\sT$ is the matrix corresponding to $L\Star$ with respect to the superdual bases $\mc C\Star$ and $\mc B\Star$.
\end{remark}
We denote by $\vphi: \End_\D (\U) \rightarrow \End_{\D\sop} (\U\Star)$ the map $L \mapsto L\Star$. It is clearly a degree-preserving super-anti-isomorphism. It follows that, if we consider the Lie superalgebras $\End_\D (U)^{(-)}$ and $\End_{\D\sop} (U\Star)^{(-)}$, the map $-\vphi$ is an isomorphism.
We summarize these considerations in the following result:
\begin{lemma}\label{lemma:iso-inv}
If $\Gamma = \Gamma(T,\beta,\gamma)$ and $\Gamma' = \Gamma(T,\beta\inv,\gamma\inv)$ are $G$-gradings (considered as $G^\#$-gradings) on the associative superalgebra $M(m,n)$, then, as gradings on the Lie superalgebra $M(m,n)^{(-)}$, $\Gamma$ and $\Gamma'$ are isomorphic via an automorphism of $M(m,n)^{(-)}$ that is the negative of a super-anti-automorphism of $M(m,n)$.
\end{lemma}
\begin{proof}
Let $\D$ be a graded division superalgebra associated to $(T,\beta)$ and let $\U$ be the graded right $\D$-module associated to $\gamma$. The grading $\Gamma$ is obtained by an identification $\psi: M(m, n) \xrightarrow{\sim} \End_\D (\U)$. By Lemma \ref{lemma:Dsop} and Remark \ref{rmk:gamma-inv}, $\Gamma'$ is obtained by an identification $\psi': M(m, n) \xrightarrow{\sim} \End_{\D\sop} (\U\Star)$. Hence we have the diagram:
\begin{center}
\begin{tikzcd}
& \End_\D (\U) \arrow[to=3-2, "-\vphi"]\\
M(m, n) \arrow[ur, "\psi"] \arrow[dr, "\psi'"]\\
& \End_{\D\sop} (\U\Star)
\end{tikzcd}
\end{center}
Thus, the composition $(\psi')\inv \, (-\vphi) \, \psi$ is an automorphism of the Lie superalgebra $M(m,n)^{(-)}$ sending $\Gamma$ to $\Gamma'$.
\end{proof}
\subsection{Type I gradings on $A(m,n)$}
In this work, we only classify the gradings on $A(m,n)$ that are induced from the associative algebra $M(m+1, n+1)$.
\begin{defi}\label{def:grd-on-Amn-I}
If $\Gamma(T, \beta, \gamma_0,\gamma_1)$ is an even grading on $M(m+1,n+1)$ (see Definition \ref{def:even-grd-on-Mmn}), we denote by $\Gamma_A (T, \beta, \gamma_0,\gamma_1)$ the induced grading on $A(m,n)$. Analogously, if $\Gamma(T, \beta, \gamma)$, or alternatively $\Gamma(t_0, \barr T, \barr\beta, u, \gamma)$, is an odd grading on $M (n+1,n+1)$ (see Definitions \ref{def:odd-grd-on-Mmn-1} and \ref{def:odd-grd-on-Mmn-2}), we denote by $\Gamma_A (T, \beta, \gamma)$, respectively $\Gamma_A (t_0, \barr T, \barr\beta, u, \gamma)$, the induced grading on $A(n,n)$. (Recall that odd gradings can occur only if $m=n$.)
\end{defi}
\begin{thm}\label{thm:even-Lie-iso}
If a $G$-grading of Type I on the Lie superalgebra $A(m,n)$ is even, then it is isomorphic to some $\Gamma_A(T,\beta, \gamma_0, \gamma_1)$ as in Definition \ref{def:grd-on-Amn-I}.
Two such gradings, $\Gamma=\Gamma_A(T,\beta, \gamma_0, \gamma_1)$ and $\Gamma'=\Gamma_A (T',\beta', \gamma_0', \gamma_1')$, are isomorphic if, and only if, $T=T'$ and there are $\delta\in \{\pm 1\}$ and $g\in G$ such that $\beta^\delta=\beta'$ and
\begin{enumerate}[(i)]
\item for $m\neq n$: $g \Xi (\gamma_0^\delta) =\Xi(\gamma_0')$ and $g \Xi (\gamma_1^\delta) =\Xi(\gamma_1')$;
\item for $m = n$: either $g \Xi(\gamma_0^\delta)=\Xi(\gamma_0')$ and $g \Xi(\gamma_1^\delta)=\Xi(\gamma_1')$ or $g\Xi(\gamma_0^\delta)=\Xi(\gamma_1')$ and $g \Xi(\gamma_1^\delta)=\Xi(\gamma_0')$.
\end{enumerate}
\end{thm}
\begin{proof}
Let $M = M(m+1, n+1)$. Since any automorphism of $M$ induces an automorphism of $A(m,n)$, the first assertion follows from Theorem \ref{thm:even-assc-iso} and the definition of Type I grading.
We know from Subsection \ref{ssec:auto-Amn} that every automorphism of $A(m, n)$ arises from an automorphism of $M$ or the negative of a super-anti-automorphism of $M$. Moreover, this automorphism or super-anti-automorphism is uniquely determined and, hence, any Type I grading on $A(m,n)$ is induced by a unique grading on $M$. It follows that $\Gamma$ and $\Gamma'$ are isomorphic if, and only if, there exists either $(a)$ an automorphism or $(b)$ a super-anti-automorphism of $M$ sending $\Gamma (T,\beta, \gamma_0, \gamma_1)$ to $\Gamma (T',\beta', \gamma_0', \gamma_1')$.
From Theorem \ref{thm:even-assc-iso}, we know that case $(a)$ holds if, and only if, the above conditions are satisfied with $\delta = 1$.
From Lemma \ref{lemma:iso-inv}, there is an automorphism of $A(m,n)$ coming from a super-anti-automorphism of $M$ that sends $\Gamma (T,\beta, \gamma_0, \gamma_1)$ to $\Gamma (T,\beta\inv, \gamma_0\inv, \gamma_1\inv)$. It follows that case $(b)$ holds if, and only if, the above conditions are satisfied with $\delta = -1$.
\end{proof}
\begin{thm}\label{thm:first-odd-Lie-iso}
If a $G$-grading of Type I on the Lie superalgebra $A(n,n)$ is odd, then it is isomorphic to some $\Gamma_A(T,\beta,\gamma)$ as in Definition \ref{def:grd-on-Amn-I}.
Two such gradings, $\Gamma_A (T,\beta, \gamma)$ and $\Gamma_A (T',\beta', \gamma')$, are isomorphic if, and only if, $T=T'$, and there are $\delta \in \{\pm 1\}$ and $g\in G$ such that $\beta^\delta=\beta'$ and $g \Xi(\gamma^\delta)=\Xi(\gamma')$.
\end{thm}
\begin{proof}
The same as for Theorem \ref{thm:even-Lie-iso}, but referring to Theorem \ref{thm:first-odd-iso} instead of Theorem \ref{thm:even-assc-iso}.
\end{proof}
The parameters $T$, $\beta$ and $\gamma$ in Theorem \ref{thm:first-odd-Lie-iso} are associated to the group $G^\#$, not $G$. Below we use parameters associated to $G$, as we did in Subsection \ref{ssec:second-odd}.
\begin{cor}\label{cor:2nd-odd-Lie-iso}
If a $G$-grading of Type I on the Lie superalgebra $A(n,n)$ is odd, then it is isomorphic to some $\Gamma_A (t_0, \barr T, \barr\beta, u, \gamma)$.
Two such gradings, $\Gamma_A (t_0, \barr T, \barr \beta, u, \gamma)$ and $\Gamma_A (t_0', \barr T', \barr \beta', u', \gamma')$, are isomorphic if, and only if, $t_0=t_0'$, $\barr T = \barr T'$, and there are $\delta \in \{\pm 1\}$ and $g\in G$ such that $\barr\beta^\delta = \barr\beta'$, $u^\delta \equiv u' \,\, (\operatorname{mod}\,\, \langle t_0 \rangle)$ and $g\, \Xi(\gamma^\delta) = \Xi(\gamma')$.
\end{cor}
\begin{proof}
Follows from Theorems \ref{thm:first-odd-Lie-iso} and \ref{thm:2nd-odd-iso}.
\end{proof}
\section{Gradings on $P(n)$}\label{sec:Pn}
Throughout this section it will be assumed that $\Char \FF = 0$.
\subsection{The Lie superalgebra $P(n)$}\label{subseq:Pn}
Let $U = U\even \oplus U\odd$ be a superspace and let $\langle\, , \rangle: U\times U\rightarrow \FF$ be a bilinear form that is homogeneous with respect to the $\ZZ_2$-grading, i.e., has parity as a linear map $U\tensor U \rightarrow \FF$.
We say that $\langle\, , \rangle$ is \emph{supersymmetric} if $\langle x,y\rangle = (-1)^{\abs{x}\abs{y}} \langle y,x \rangle$ for all homogeneous elements $x,y\in U$.
From now on, we suppose that $\langle\, , \rangle$ is supersymmetric, nondegenerate, and odd.
The \emph{periplectic Lie superalgebra} $\mathfrak{p}(U)$ is defined as $\mathfrak{p}(U)\even \oplus \mathfrak{p}(U)\odd$ where
\[
\mathfrak{p}(U)^{i} = \{L\in \gl(U)^i\mid \langle L(x),y\rangle = - (-1)^{i\abs{x}} \langle x,L(y)\rangle\}
\]
for all $i\in\Zmod2$.
The superalgebra $\mathfrak{p}(U)$ is not simple, but its derived superalgebra $P(U) = [\mathfrak{p}(U),\mathfrak{p}(U)]$ is simple if $\Dim U \geq 6$.
Since $\langle\, , \rangle$ is nondegenerate and odd, it is clear that $U\odd$ is isomorphic to $(U\even)^*$ by $u \mapsto \langle u, \cdot\rangle $.
Writing $U\even = V$, we can identify $U$ with $V\oplus V^*$.
Since $\langle \, , \rangle$ is supersymmetric, with this identification we have
\[
\langle v_1+v^*_1,v_2 + v_2^* \rangle = v_1^* (v_2) + v_2^*(v_1)
\]
for all $v_1, v_2\in V$ and $v_1^*, v_2^*\in V^*$.
Hence, $P(U)$ is a subsuperspace of
\[
\End(U) = \End(V \oplus V^*) =
\begin{pmatrix}
\End (V) & \Hom (V^*, V)\\
\Hom (V, V^*) & \End(V^*)
\end{pmatrix}
\]
given by
\[
P(U) = \left\{\left(\begin{matrix}
a & b \\
c & -a^*\\
\end{matrix}
\right)\Big| \,\tr a = 0,\, b=b^* \AND c=-c^*\right\}.
\]
In the case $V=\FF^{n+1}$, we write $\mathfrak{p}(n)$ for $\mathfrak{p}(U)$ and define $P(n) = [\mathfrak{p}(n),\mathfrak{p}(n)]$, where $n\geq 2$.
Using the standard basis of $V$, we can identify $P(n)$ with the following subsuperalgebra of $M(n+1,n+1)^{(-)}$:
\begin{equation}\label{eq:Pn-abstract}
P(n) = \left\{\left(\begin{matrix}
a & b \\
c & -a\transp\\
\end{matrix}
\right)\Big| \,\tr a = 0,\, b=b\transp \AND c=-c\transp\right\}.
\end{equation}
One can readily check that $P(U)$ is a graded subspace of $\End (U)$ equipped with its canonical $\ZZ$-grading, so we have $P(U) = P(U)\inv \oplus P(U)^0 \oplus P(U)^1$.
Also, the map $\iota: \Sl(n+1) \rightarrow P(n)^0$ given by
$
\iota(a) = \left(\begin{matrix}
a & 0 \\
0 & -a\transp\\
\end{matrix}
\right)
$
is an isomorphism of Lie algebras.
If we identify $\Sl(n+1)$ and $P(n)^0$ via this map, then $P(n)^{-1} \iso \mathrm{S}^2 (U\even) \iso V_{2\pi_1}$ and $P(n)^1 \iso \Exterior^2 (U\odd) \iso V_{\pi_{n-1}}$ as modules over $P(n)^0$, where $\pi_i$ denotes the $i$-th fundamental weight of $\Sl(n+1)$.
\subsection{Automorphisms of $P(n)$}
The automorphisms of $P(n)$ were originally described by V.
Serganova (see \cite[Theorem 1]{serganova}).
We give a more explicit description of the automorphism group that we will use for our purposes.
\begin{lemma}\label{lemma:Pn-generates-Mmn}
Let $U$ be a finite-dimensional superspace equipped with a supersymmetric nondegenerate odd bilinear form.
The subset $P(U)$ generates $\End (U)$ as an associative superalgebra.
\end{lemma}
\begin{proof}
Denote by $R$ the associative superalgebra generated by $P(U)$.
We claim that $U$ is a simple $R$-module.
Indeed, since $P(U)\even\iso \Sl(n+1)$, we have that $U\even \iso V_{\pi_1}$ and $U\odd \iso V_{\pi_n}$ are simple non-isomorphic modules over the Lie algebra $P(U)\even$.
Also, the action of $P(U)\odd$ moves elements from $U\even$ to $U\odd$ and vice-versa, so $U$ does not have nonzero proper subspaces invariant under $P(U)$.
By Density Theorem, since we are over an algebraically closed field, we conclude that $R = \End (U)$.
\end{proof}
\begin{prop}\label{prop:Aut-Pn}
The group of automorphisms of $P(n)$ is $\frac{\GL (n+1)}{\{-1,+1\}}$ where $a\in \GL(n+1)$ acts as the conjugation by
$\left( \begin{matrix}
a&0\\
0&(a\transp)^{-1}\\
\end{matrix}\right)$.
\end{prop}
\begin{proof}
Let $P=P(n)$ and $\vphi: P \rightarrow P$ be a Lie superalgebra automorphism.
Since it preserves the canonical $\ZZ_2$-grading, taking its restrictions, we obtain a Lie algebra automorphism $\vphi\subeven : P\even \rightarrow P\even$ and an invertible linear map $\vphi\subodd: P\odd \rightarrow P\odd$.
\setcounter{claim}{0}
\begin{claim}
The components $P^{-1}$ and $P^1$ of the canonical $\ZZ$-grading are invariant under $\vphi$.
\end{claim}
We denote by $(P\odd)^{\vphi\subeven}$ the $P\even$-module $P\odd$ twisted by $\vphi\subeven$, \ie, the space $P\odd$ with a new action given by $\ell \cdot x = \vphi\subeven (\ell)x$ for all $\ell \in P\even$ and $x \in P\odd$.
Clearly, the map $\vphi\subodd: P\odd \rightarrow (P\odd)^{\vphi\subeven}$ is a $P\even$-module isomorphism.
In particular, $(P\odd)^{\vphi\subeven} = \vphi\subodd (P^{-1}) \oplus \vphi\subodd (P^{1})$, where $\vphi\subodd (P^{-1})$ and $\vphi\subodd (P^{1})$ are simple and non-isomorphic.
It follows that either $(P^{-1})^{\vphi\subeven} = \vphi\subodd (P^{-1})$ or $(P^{-1})^{\vphi\subeven} = \vphi\subodd (P^{1})$.
By dimension count, we have $(P^{-1})^{\vphi\subeven} = \vphi\subodd (P^{-1})$ and, similarly, $(P^{1})^{\vphi\subeven} = \vphi\subodd (P^{1})$
\begin{claim}
The automorphism $\vphi\subeven$ is inner.
\end{claim}
If we identify $\Sl(n+1)$ with $P^0$ via the map $\iota$ defined in Subsection \ref{subseq:Pn}, we have $P^{-1}\iso V_{2\pi_1}$ as an $\Sl(n+1)$-module.
By Claim 1, we know that $\vphi\subodd \restriction_{P\inv} : P\inv \rightarrow (P\inv)^{\vphi\subeven}$ is an isomorphism of modules, but if $\vphi\subeven$ were an outer automorphism, we would have $(V_{2\pi_1})^{\vphi\subeven} \simeq V_{2\pi_n}$, which would force $n=1$, a contradiction.
\begin{claim}
If $\varphi\subeven = \id$, then $\varphi = \upsilon_\lambda$ for some $\lambda\in \FF^\times$.
\end{claim}
Recall from Subsection \ref{ssec:G-hat-action} that $\upsilon_{\lambda}$ acts as $\lambda^i \id$ on $P^i$.
Since $\varphi\subeven = \id$, $\vphi_{\bar1}: P\odd \rightarrow P\odd$ is a $P^0$-module automorphism.
By Claim 1 and Schur's Lemma, $\vphi_{\bar1}\restriction_{P^{-1}}$ and $\vphi_{\bar1}\restriction_{P^{1}}$ are scalar operators.
Due to the superalgebra structure, these two scalars must be inverses of each other, concluding the proof of the claim.\qedclaim
By Claim 2, we know that there is an invertible $a\in \End(U\even)$ such that $\vphi\subeven$ is the conjugation by $A=\left( \begin{matrix}
a&0\\
0&(a\transp)^{-1}\\
\end{matrix}\right)$.
By Claim 3, $\vphi$ must be this conjugation composed with $\upsilon_\lambda$ for some $\lambda\in \FF^\times$.
But $\upsilon_\lambda$ is the conjugation by $\left( \begin{matrix}
\mu\inv\id &0\\
0&\mu \id\\
\end{matrix}
\right)$ where $\mu^2=\lambda$, so we can adjust $a$ and assume that $\vphi$ is the conjugation by $A$.
Since $P$ generates $M(n+1,n+1)$ as an associative superalgebra (Lemma \ref{lemma:Pn-generates-Mmn}), $A$ is determined up to scalar and, clearly, the only possible scalars are $-1$ and $1$.
\end{proof}
\begin{remark}
The images of $\upsilon_\lambda$, $\lambda \in \FF^\times$, cover the group of outer automorphisms of $P(n)$ (see \cite[Theorem 1]{serganova}).
\end{remark}
\subsection{Restriction of gradings from $M(n+1,n+1)$ to $P(n)$}
We start with a consequence of Proposition \ref{prop:Aut-Pn}.
\begin{cor}\label{cor:automorphisms-Pn
Every automorphism of $P(n)$ is the restriction of a unique even automorphism of $M(n+1, n+1)$ and every grading on $P(n)$ is the restriction of a unique even grading on $M(n+1, n+1)$.
\end{cor}
\begin{proof}
Consider the embedding $\Aut(P(n))\rightarrow \Aut(M(n+1,n+1))$ that follows from Proposition \ref{prop:Aut-Pn}.
The image consists of even automorphisms, so Proposition \ref{prop:3-equiv-even-morita-action}(iv) implies that every $G$-grading on $P(n)$ extends to an even grading on $M(n+1, n+1)$. The uniqueness follows from Lemma \ref{lemma:Pn-generates-Mmn}.
\end{proof}
Of course, not every even grading on $M(n+1,n+1)$ restricts to $P(n)$.
We are going to obtain necessary and sufficient conditions for such restriction to be possible.
Let $\D$ be a finite-dimensional graded division algebra.
The concept of \emph{dual of a graded $\D$-module} is a special case of the concept of \emph{superdual} discussed in Subsection \ref{ssec:superdual},
which arises when the gradings on $\D$ and its graded modules are even.
Furthermore, in our situation $T$ must be an elementary $2$-group (see Theorem \ref{thm:Pn-elem-2-grp}).
Let us recall the definitions and specialize them to the case at hand.
Let $\mc V$ be a \emph{right} graded $\D$-module. Then $\mc V^{\star}=\Hom_{\D} (\mc V, \D)$ is a \emph{left} $\D$-module with the action given by $(d\cdot f) (v) = d f(v)$ for all $d\in \D$, $f\in\mc V^\star$ and $v\in\mc V$.
If $\mc B = \{v_1, \ldots, v_k\}$ is a homogeneous basis for $\mc V$, the dual basis $\mc B\Star \subseteq \mc V \Star$ consists of the elements $v_i\Star$, $1\leq i \leq k$, defined by $v_i\Star (v_j) = \delta_{ij}$.
Note that $\operatorname {deg} v_i\Star = (\operatorname {deg} v_i)\inv$.
Given two right $\D$-modules, $\mc V$ and $\mc W$, and a $\D$-linear map $L:\mc V\rightarrow \mc W$, we have the adjoint $L^\star: \mc W^\star\rightarrow \mc V^\star$ defined by $L^\star(f) = f\circ L$, for every $f\in\mc W^\star$.
We now assume that $\D$ is a standard realization associated to a pair $(T, \beta)$ such that $T$ is an elementary $2$-group.
With this we can identify $\D$ with $\D\op$ via transposition (see Remark \ref{rmk:2-grp-transp}) and, hence, we can regard left $\D$-modules as right $\D$-modules.
In particular, if $\mc V$ is a graded right $\D$-module, then $\mc V\Star$ is a graded right $\D$-module via $(f \cdot d)(v) = d\transp f(v)$ for all $f\in \mc V\Star$, $d\in\D$ and $v\in \mc V$.
Consider the space $\Hom_\D(\mc V, \mc W)$.
Fixing homogeneous $\D$-bases $\mc B = \{v_1, \ldots, v_k\}$ and $\mc C = \{w_1, \ldots, w_\ell\}$ for $\mc V$ and $\mc W$, respectively, we obtain an isomorphism between $\Hom_\D(\mc V, \mc W)$ and $\M_{\ell \times k} (\D)$.
The latter is naturally isomorphic to $\M_{\ell \times k} (\FF) \tensor \D$, so we will identify them.
\begin{lemma}\label{lemma:D-transp}
Let $L: \mc V\rightarrow \mc W$ be a $\D$-linear map.
We fix homogeneous $\D$-bases $\mc B$ and $\mc C$ on $\mc V$ and $\mc W$, respectively, and their dual bases in $\mc V\Star$ and $\mc W\Star$.
If $A\tensor d \in \M_{\ell \times k} (\FF) \tensor \D$ represents $L$, then $A\transp \tensor d\transp$ represents $L\Star$.\qed
\end{lemma}
We can regard the elements of $\M_{\ell \times k} (\FF) \tensor \D$ as matrices over $\FF$ via Kronecker product (as in Definition \ref{def:explicit-grd-assoc}).
Then we have $A\transp \tensor d\transp = (A\tensor d)\transp$.
\begin{thm}\label{thm:Pn-elem-2-grp}
Let $U$ be a finite-dimensional superspace and let $\Gamma = \Gamma (T, \beta, \gamma_0, \gamma_1)$ be an even $G$-grading on $\End(U)$.
The superspace $U$ admits a supersymmetric nondegenerate odd bilinear form such that $P(U)$ is a $G$-graded subsuperalgebra of $\End(U)^{(-)}$ if, and only if, $T$ is an elementary $2$-group and there is $g_0\in G$ such that $\Xi(\gamma_1) = g_0 \, \Xi(\gamma_0\inv)$.
Moreover, if there are two supersymmetric nondegenerate odd bilinear forms on $U$ such that the corresponding $P_1(U)$ and $P_2(U)$ are $G$-graded subsuperalgebras, then $P_1(U)$ and $P_2(U)$ are ismorphic up to shift in opposite directions.
\end{thm}
\begin{proof}
Assume that, for some form, $P(U)$ is a $G$-graded subsuperalgebra.
Let $V=U\even$ and consider the identification of $U\odd$ with $V^*$ presented in Subsection \ref{subseq:Pn}.
This way $\Gamma = \Gamma (T, \beta, \gamma_0, \gamma_1)$ is an even grading on
\[
\End(U) = \End(V \oplus V^*) =
\begin{pmatrix}
\End (V) & \Hom (V^*, V)\\
\Hom (V, V^*) & \End(V^*)
\end{pmatrix}.
\]
In particular, $\End(V)$ and $\End(V^*)$ are graded subspaces of $\End(U)\even$, with gradings $\Gamma (T, \beta, \gamma_0)$ and
$\Gamma (T, \beta, \gamma_1)$, respectively. If
\[
x = \left(\begin{matrix}
a & 0 \\
0 & -a^*\\
\end{matrix}
\right)
\]
is a homogeneous element in $P(U)\even$, then both $u(x) := a \in \Sl(V) \subseteq \End(V)$ and $v(x) := -a^* \in \Sl(V^*) \subseteq \End(V^*)$ are homogeneous elements of the same degree.
In other words, the maps $u: P(n)\even \rightarrow \Sl(V)$ and $v: P(n)\even \rightarrow \Sl(V^*)$ are homogeneous of degree $e$.
Consider the algebra isomorphism $\vphi: \End(V)\op \rightarrow \End(V^*)$ associating to each operator its adjoint.
Clearly, $\vphi(a) = - (v \circ u\inv) (a)$ for all $a\in \Sl(V)$.
Since $\End(V) = \Sl(V)\, \oplus\, \FF\id_V$ and $\vphi(\id_V) = \id_{V^*}$, we see that $\vphi$ is homogeneous of degree $e$.
From Lemma \ref{lemma:Dsop} and Remark \ref{rmk:gamma-inv}, we conclude that $\Gamma (T, \beta\inv, \gamma_0\inv) \iso \Gamma (T, \beta, \gamma_0)$, and hence, by Theorem \ref{thm:classification-matrix}, $\beta\inv = \beta$ and there is $g_0\in G$ such that $g_0\,\Xi(\gamma_0\inv) = \Xi(\gamma_1)$.
Since $\beta$ is nondegenerate, $\beta\inv = \beta$ if, and only if, $T$ is an elementary $2$-group.
Note that the $G$-graded algebra $P(U)\even$ is isomorphic (via the map $u$) to the $G$-graded subalgebra $\Sl(V)$ of $\End(V)^{(-)}$, where the grading on $\End(V)$ is $\Gamma(T,\beta, \gamma_0)$.
Therefore, if we have two forms such that the corresponding $P_1(U)$ and $P_2(U)$ are $G$-graded subsuperalgebras, then their even parts are isomorphic as $G$-graded algebras.
Using Lemmas \ref{lemma:simplebimodule} and \ref{lemma:opposite-directions}, we conclude the ``moreover'' part.
Now assume, conversely, that $T$ is an elementary $2$-group and $\Xi(\gamma_1) = g_0 \, \Xi(\gamma_0\inv)$.
We can adjust $\gamma_1$, if necessary, so that $\gamma_1 = g_0\, \gamma_0\inv$ and the isomorphism class of $\Gamma$ does not change.
Let $\D$ be a standard realization of a graded division algebra associated to $(T,\beta)$ and let $\mc V$ be a graded right $\D$-module
with a homogeneous basis $\mc B$ whose degrees are given by $\gamma_0$.
Define $\mc U = \mc U\even \oplus \mc U\odd$ with $\mc U\even = \mc V$ and $\mc U\odd = (\mc V\Star)^{[g_0]}$.
The $G$-grading $\Gamma$ on $\End(U)$ is defined by means of an isomorphism:
\[
\begin{split}
\End(U) \iso \End_\D (\mc U) &= \begin{pmatrix}
\End_\D (\mc V) & \Hom_\D ((\mc V\Star)^{[g_0]}, \mc V)\\
\Hom_\D (\mc V, (\mc V\Star)^{[g_0]}) & \End_\D ((\mc V\Star)^{[g_0]})
\end{pmatrix}\\
&= \begin{pmatrix}
\End_\D (\mc V) & \Hom_\D (\mc V\Star, \mc V)^{[g_0\inv]}\\
\Hom_\D (\mc V, \mc V\Star)^{[g_0]} & \End_\D (\mc V\Star)
\end{pmatrix}.
\end{split}
\]
Using the homogeneous $\D$-bases $\mc B$ for $\mc V$ and $\mc B\Star$ for $\mc V\Star$ to represent $\D$-linear maps by matrices in $M_k(\D) = M_k(\FF) \tensor \D$ and using the Kronecker product to identify the latter with $M_{n+1}(\FF)$, we obtain an isomorphism $\End(U) \xrightarrow{\sim} M(n+1, n+1)$, and $M(n+1, n+1)$ contains $\mathfrak{p}(n)$ and $P(n) = [\mathfrak{p}(n), \mathfrak{p}(n)]$ as in Equation \eqref{eq:Pn-abstract}.
The above isomorphism $\End(U)\xrightarrow{\sim} M(n+1,n+1)$ of superagebras is given by an isomorphism of superspaces $U \xrightarrow{\sim} \FF^{n+1} \oplus \FF^{n+1}$.
Hence, there exists a supersymmetric nondegenerate odd bilinear form on $U$ such that $P(U)$ corresponds to $P(n)$ under the above isomorphism.
Finally, we have to show that $P(U)$ is a $G$-graded subsuperspace of $\End(U)$.
Clearly, it is sufficient to prove the same for $\mathfrak{p}(U)$.
But $\mathfrak{p}(U)$ corresponds to
\begin{equation*}
\mathfrak{p}(n) = \left\{\left(\begin{matrix}
a & b \\
c & -a\transp\\
\end{matrix}
\right)\Big| \,a,b,c\in M_{n+1}(\FF),\, b=b\transp \AND c=-c\transp\right\} \subseteq M(n+1,n+1),
\end{equation*}
which, in view of Lemma \ref{lemma:D-transp}, corresponds to the subsuperspace
\[\begin{split}
\bigg \{\left(\begin{matrix}
a & b \\
c & -a\Star\\
\end{matrix}
\right)\Big| \,a\in \End_\D(\mc U),\, b=b\Star\in \Hom_\D(\mc V\Star, \mc V),\, c=-c\Star\in \Hom_\D(\mc V, \mc V\Star) \bigg\}
\end{split}
\]
of $\End_\D(\mc U)$, which is clearly a $G$-graded subsuperspace.
\end{proof}
\subsection{$G$-gradings up to isomorphism}
\begin{defi}\label{def:grd-Pn}
Let $T\subseteq G$ be a finite elementary $2$-subgroup, $\beta$ be a nondegenerate alternating bicharacter on $T$, $\gamma$ be a $k$-tuple of elements of $G$, and $g_0\in G$. We will denote by $\Gamma_P (T, \beta, \gamma, g_0)$ the grading on the superalgebra $P(n)$ obtained by restricting the grading $\Gamma(T,\beta,\gamma,g_0\gamma\inv)$ on $M(n+1,n+1)$ as in the proof of Theorem \ref{thm:Pn-elem-2-grp}.
Explicitly, write $\gamma = (g_1, \ldots, g_k)$ and take a standard realization of a graded division algebra $\D$ associated to $(T, \beta)$.
Then $M_{n+1}(\FF)\iso M_k(\FF) \tensor \D$ by means of Kronecker product, and
\[
M(n+1, n+1) \iso
%
\begin{pmatrix}
M_k (\FF)\tensor \D & M_k (\FF)\tensor \D\\
M_k (\FF)\tensor \D & M_k (\FF)\tensor \D
\end{pmatrix}
\]
Denote by $E_{ij}$ the $(i,j)$-th matrix unit in $M_k (\FF)$. The grading $\Gamma(T,\beta,\gamma,g_0\gamma\inv)$ is given by:
\begin{center}
\begin{tabular}{@{$\bullet$ }ll}
$\deg (E_{ij}\tensor d) = g_i (\deg d) g_j\inv$ & in the upper left corner;\\
$\deg (E_{ij}\tensor d) = g_i (\deg d) g_j \, g_0\inv$ & in the upper right corner;\\
$\deg (E_{ij}\tensor d) = g_i\inv (\deg d) g_j\inv g_0$ & in the lower left corner;\\
$\deg (E_{ij}\tensor d) = g_i\inv (\deg d) g_j$ & in the lower right corner.
\end{tabular}
\end{center}
\end{defi}
Note that the restriction of $\Gamma_P(T, \beta, \gamma, g_0)$ to the even part is the inner grading on $\Sl(n+1)$ with parameters $(T, \beta, \gamma)$.
\begin{thm}\label{thm:Pn-iso}
Every $G$-grading on the Lie superalgebra $P(n)$ is isomorphic to some $\Gamma_P (T, \beta, \gamma, g_0)$ as in Definition \ref{def:grd-Pn}.
Two gradings, $\Gamma = \Gamma_P (T,\beta,\gamma,g_0)$ and $\Gamma' = \Gamma_P (T',\beta',\gamma',g_0')$, are isomorphic if and only if $T=T'$, $\beta = \beta'$, and there is $g\in G$ such that $g^2 g_0 = g_0'$ and $g\,\Xi(\gamma)=\Xi(\gamma')$.
\end{thm}
\begin{proof}
The first assertion follows from Corollary \ref{cor:automorphisms-Pn} and Theorem \ref{thm:Pn-elem-2-grp}.
For the second assertion, recall that $\Gamma$ and $\Gamma'$ are, respectively, the restrictions of the gradings
$\widetilde \Gamma = \Gamma (T, \beta, \gamma, g_0 \gamma\inv)$ and $\widetilde \Gamma' = \Gamma (T', \beta', \gamma', g_0' (\gamma')\inv)$
on $M(n+1, n+1)$.
$(\Rightarrow)$:
Suppose $\Gamma \iso \Gamma'$.
Since every automorphism of $P(n)$ extends to an automorphism of $M(n+1, n+1)$ (Corollary \ref{cor:automorphisms-Pn}), we have
$\widetilde \Gamma \iso \widetilde \Gamma'$, which implies $T=T'$ and $\beta = \beta'$ by Theorem \ref{thm:even-assc-iso}.
Let $\D$ be a standard realization associated to $(T, \beta)$ and let $\mc V$ be a right $\D$-module with basis $\mc B = \{v_1, \ldots, v_k\}$,
which is graded by assigning $\deg v_i = g_i$.
The same module, but with $\deg v_i = g_i'$, will be denoted by $\mc W$.
Then $E = \End_\D (\mc V \oplus (\mc V\Star)^{[g_0]})$ and $E' = \End_\D (\mc W \oplus (\mc W\Star)^{[g_0']})$ are graded superalgebras.
Using the bases $\mc B$ and $\mc B\Star$ and the Kronecker product, we can identify them with $M(n+1, n+1)$.
The first identification gives the grading $\widetilde \Gamma$ on $M(n+1, n+1)$ and the second gives $\widetilde \Gamma'$.
Let $\Phi$ be an automorphism of $M(n+1, n+1)$ that sends $\widetilde\Gamma$ to $\widetilde\Gamma'$.
By Proposition \ref{prop:Aut-Pn}, $\Phi$ is the conjugation by
\[
A =
\begin{pmatrix}
a & 0\\
0 & (a\transp)\inv
\end{pmatrix}
\]
for some $a\in \GL(n+1)$.
By Lemma \ref{lemma:D-transp}, $\Phi$ corresponds to the isomorphism $E\rightarrow E'$ that is the conjugation by
\[
\phi =
\begin{pmatrix}
\alpha & 0\\
0 & (\alpha\Star)\inv
\end{pmatrix}
\]
where $\alpha: \mc V\rightarrow \mc W $ and $(\alpha\Star)\inv: (\mc V\Star)^{[g_0]} \rightarrow (\mc W\Star)^{[g_0']} $ are $\D$-linear maps.
On the other hand, by Proposition \ref{prop:inner-automorphism}, this isomorphism $E\to E'$ is the conjugation by a homogeneous bijective $\D$-linear map
\[
\psi=
\left( \begin{matrix}
\psi_{11}&\psi_{12}\\
\psi_{21}&\psi_{22}\\
\end{matrix}\right).
\]
It follows that there is $\lambda\in \FF$ such that $\phi = \lambda\psi$, and, hence, $\phi$ is homogeneous.
Let us denote its degree by $g$.
Then both $\alpha$ and $(\alpha\Star)\inv$ must be homogeneous of degree $g$.
Hence, $\alpha: \mc V^{[g]}\to\mc W $ is an isomorphism of graded $\D$-modules, so we conclude that $g \Xi(\gamma) = \Xi(\gamma')$.
Considered as a map $\mc V\Star \rightarrow \mc W\Star$, $(\alpha\Star)\inv$ would have degree $g\inv$, so taking into account the shifts, it has degree $g\inv g_0\inv g_0'$, which must be equal to $g$, so $g_0' = g^2 g_0$.
$(\Leftarrow)$:
We may suppose $\D=\D'$. Since $g\Xi(\gamma) = \Xi(\gamma')$, we have an isomorphism of graded $\D$-modules $\alpha: \mc V^{[g]} \rightarrow \mc W$.
As a map from $\mc V$ to $\mc W$, $\alpha$ is homogeneous of degree $g$, hence $(\alpha\Star)\inv: \mc V\Star \rightarrow \mc W\Star$ has degree $g\inv$.
It follows that, as a map from $(\mc V\Star)^{[g_0]}$ to $(\mc W\Star)^{[g_0']}$, $(\mc \alpha\Star)\inv$ has degree $g\inv g_0\inv g_0' = g$.
The desired automorphism of $P(n)$ that sends $\Gamma$ to $\Gamma'$ is the conjugation by the matrix
$\psi=\left( \begin{matrix}
\alpha&0\\
0&(\alpha\Star)\inv\\
\end{matrix}\right).$
\end{proof}
\subsection{Fine gradings up to equivalence}
For every integer $\ell\ge 0$, we set $T_{(\ell)}=\ZZ_2^{2\ell}$ and fix a nondegenerate alternating bicharacter $\beta_{(\ell)}$, say,
\[
\beta_{(\ell)} (x,y)=(-1)^{\sum_{i=1}^{2\ell} x_i y_{2\ell-i+1}}.
\]
\begin{defi}\label{def:fine-grd-Pn}
For every $\ell$ such that $2^\ell$ is a divisor of $n+1$, put $k:=\frac{n+1}{2^\ell}$ and $\tilde{G}_{(\ell)}=T_{(\ell)}\times \ZZ^k$.
Let $\{e_1,\ldots,e_k\}$ be the standard basis of $\ZZ^k$ and let $\langle e_0\rangle$ be the infinite cyclic group generated by a new symbol $e_0$.
We define $\Gamma_P(\ell,k)$ to be the $\tilde{G}_{(\ell)}\times\langle e_0 \rangle$-grading $\Gamma_P (T_{(\ell)},\beta_{(\ell)},(e_1, \ldots, e_k), e_0)$ on $P(n)$.
If $n$ is clear from the context, we will simply write $\Gamma_P(\ell)$.
\end{defi}
The subgroup of $\tilde{G}_{(\ell)} \times \langle e_0 \rangle$ generated by the support of $\Gamma_P(\ell,k)$ is
\[
G_{(\ell)} := (T_{(\ell)}\times \ZZ^k_0)\oplus \langle 2e_1 - e_0 \rangle \iso \ZZ_2^{2\ell}\times \ZZ^k.
\]
\begin{prop}\label{prop:P-fine}
The gradings $\Gamma_P(\ell)$ on $P(n)$ are fine and pairwise nonequivalent.
\end{prop}
\begin{proof}
We can write $\Gamma_P (\ell) = \Gamma^{-1} \oplus \Gamma^0 \oplus \Gamma^1$ where $\Gamma^i$ is the restriction of $\Gamma_P(\ell)$ to
the $i$-th component of the canonical $\ZZ$-grading of $P(n)$.
We identify $ P(n)^0 = P(n)\even$ with $ \Sl(n+1)$ via de map $\iota$ defined in Subsection \ref{subseq:Pn}.
Then the grading $\Gamma^0$ on $P(n)^0$ is the restriction to $\Sl(n+1)$ of a fine grading on $M_{n+1}(\FF)$ with universal group $T_{(\ell)}\times\ZZ^k_0$ (\cite[Proposition 2.35]{livromicha}), so it has no proper refinements among the inner gradings on $\Sl(n+1)$.
Also, $\Gamma_P(\ell)$ and $\Gamma_P(\ell')$ are nonequivalent if $\ell\ne\ell'$, because their restrictions to $P(n)^0$ are nonequivalent.
Note that the supports of $\Gamma^{-1}$, $\Gamma^0$ and $\Gamma^1$ are pairwise disjoint since they project to, respectively, $-e_0$, $0$, and $e_0$
in the direct summand $\langle e_0 \rangle $ of $\tilde{G}_{(\ell)}\times\langle e_0\rangle$.
Suppose that the grading $\Gamma_P (\ell)$ admits a refinement $\Delta = \Delta\inv \oplus \Delta^0 \oplus \Delta^1$.
Then $\Delta^0$ is an inner grading that is a refinement of $\Gamma^0$, hence they are the same grading (up to relabeling).
Using Lemma \ref{lemma:simplebimodule}, we conclude that $\Gamma$ and $\Delta$ are the same grading, proving that $\Gamma$ is fine.
\end{proof}
\begin{thm}\label{thm:class-fine-Pn}
Every fine grading on $P(n)$ is equivalent to a unique $\Gamma_P(\ell)$ as in Definition \ref{def:fine-grd-Pn}.
Moreover, every grading $\Gamma_P(\ell)$ is fine, and $G_{(\ell)}$ is its universal group.
\end{thm}
\begin{proof}
Let $\Gamma=\Gamma_P(G,T,\beta,\gamma,g_0)$ be any $G$-grading on $P(n)$.
Since $T$ is an elementary $2$-group of even rank, we have an isomorphism $\alpha:T_{(\ell)}\to T$, for some $\ell$, such that $\beta_{(\ell)}=\beta\circ(\alpha\times\alpha)$. We can extend $\alpha$ to a homomorphism $G_{(\ell)}\rightarrow G$ (also denoted by $\alpha$) by sending the elements $e_1,\ldots,e_k$ to the entries of $\gamma$, and $e_0$ to $g_0$.
By construction, ${}^\alpha\Gamma_P {(\ell)}\iso\Gamma$.
It remains to apply Proposition \ref{prop:P-fine} and Lemma \ref{lemma:universal-grp}.
\end{proof}
\section*{Acknowledgments}
The first two authors were Ph.D. students at Memorial University of Newfoundland while working on this paper. Helen Samara Dos Santos would like to thank her co-supervisor, Yuri Bahturin, for help and guidance during her Ph.D. program. All authors are grateful to Yuri Bahturin and Alberto Elduque for useful discussions.
\newcommand{\noop}[1]{} \def$'${$'$}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Introduction}
Beginning in the 1970's, physicists have calculated that seemingly small changes to the fundamental constants of nature and the initial conditions of the cosmos would have dramatic effects on the universe. In particular, the complexity and stability required by any known or thus-far conceived form of life can be rather easily erased. For example, the masses of the fundamental constituents of ordinary matter --- up quarks, down quarks and electrons --- must be constrained to lie in a very small section of parameter space for nuclei, atoms, molecules, and chemistry to be possible at all. Similarly, the vacuum energy of the universe must be extraordinarily small compared to its ``natural" Planck-scale value for the universe to have any structure. This \emph{fine-tuning of the universe for life} was first investigated by \citet{Carter74}, \citet{Silk1977}, \citet{Carr79}, and \citet{BT86}, and has been reviewed recently by \citet{Hogan2000}, \citet{Barnes2012}, \citet{Schellekens2013} and \citet{Lewis2016}.
The fine-tuning of the universe for life has been pressed into the service of a variety of conclusions --- physical, cosmological and philosophical. For example, fine-tuning has been offered as evidence for a multiverse: the universe as a whole consists of an enormous number of variegated ``pocket" universes, each with different constants and cosmic conditions, and within which observers must see local conditions that are conducive to life forms \citep[see, for example, the brief history given by][]{Linde2015}. It has also been argued that fine-tuning is evidence for a cosmic designer, whose purposes for this universe include the existence of embodied moral agents \citep{Swinburne2004,Collins2009}.
A crucial motivator for these arguments is the intuition that fine-tuning demonstrates that a life-permitting universe is extraordinarily \emph{improbable}. A universe drawn blindly from a big barrel of possible universes is unlikely to have the right forces, particles and cosmic initial conditions for life to develop, it seems. However, a number of philosophers have cast doubt on whether this intuition can be made rigorous. \citet[][hereafter MMV]{MMV} and \citet[][hereafter CGP]{CGP} have argued that the relevant probability measure, because it is spread evenly over an infinitely large range, cannot be normalized and hence the relevant probabilities cannot be calculated. These papers are mostly concerned with the fine-tuning argument for the existence of God, but apply equally to the inference from fine-tuning to the existence of a multiverse. In a similar vein, \citet{Halvorson2014} has argued that a correct understanding of the probabilities shows that a life-permitting universe is unlikely on any assumptions about its origin, and thus fine-tuning cannot be used to argue for anything deeper than the laws of nature.
Our goal here is to use Bayesian probability theory, as it is employed in theory testing (or \emph{model selection}) in the physical sciences, to show that we can make rigorous the claim that the fine-tuning of the universe for intelligent life renders our universe exceedingly improbable. There really is something to be explained. Precisely what (or Who) that explanation is, is left as an exercise for the reader.
\section{Probabilities for Model Selection in Physics} \label{S:modelselect}
We begin with an overview of model selection in the physical sciences; the reader should be mindful of differences between the approaches to probability in philosophy and physics.
In recent decades, Bayesian approaches to probability theory have significantly changed both the principles and the practice of how physicists analyze data and draw scientific conclusions. Bayesian probabilities $p(B|A)$ in physics are not taken to quantify some aspect of the psychological state of someone who believes $A$ and is considering $B$. Rather, they are presented \citep[for example, by][]{Jaynes2003} as an extension to classical logic, quantifing the \emph{degree of plausibility} of the proposition $B$ given the truth of the proposition $A$. Just as symbolic logic's material condition $A \rightarrow B$ says nothing about whether $A$ is known by anyone, but instead denotes a connection between the truth values of the propositions $A$ and $B$, so $p(B|A)$ quantifies a relationship between these propositions. Note that these are not degrees of truth; $A$ and $B$ are in fact either true or false.
Why think that degrees of plausibility can be modelled by probabilities? There are a number of approaches that lead Bayesians to the probability axioms of \citet{Kolmogorov1933} or similar, such as Dutch book arguments and representational theorems which trace back to \citet{Ramsey1926}. More common among physicists is the theorem of \citet{Cox1946} \citep[see also][]{Jaynes2003,Caticha2009,Knuth2012}. We propose that degrees of plausibility obey the following desiderata :
\begin{itemize} \setlength{\itemsep}{-2pt}
\item[D1.] Degrees of plausibility are represented by real numbers. This ensures that they can be compared on a single scale.
\item[D2.] Degrees of plausibility change in common sense ways. For example, if learning $C$ makes $B$ more likely, but doesn't change how likely $A$ is, then learning $C$ should make ($A$ and $B$) more likely.
\item[D3.] If a conclusion can be reasoned out in more than one way, then every possible way must lead to the same result.
\item[D4.] Information must not be arbitrarily ignored. All given evidence must be taken into account.
\item[D5.] Identical states of knowledge (except perhaps for the labeling of the propositions) should result in identical assigned degrees of plausibility.
\end{itemize}
Cox's theorem shows that degrees of plausibility are probabilities, or more precisely, they obey the usual rules of probability. (Hereafter, we will use the term probabilty unless further distinction is required.) We have a rule for each of the Boolean operations `and' ($AB$), `or' ($A+B$) and 'not' ($\bar{A}$)\footnote{As our presentation of Bayesian probability is somewhat different to the usual philosophical presentation, we've used the notation most familiar to physicists.},
\begin{align}
p(AB|C) &\equiv p(A|BC) ~p(B|C) \equiv p(B|AC) ~p(A|C) \label{eq:prod}\\
p(A+B|C) &\equiv p(A|C) + p(B|C) - p(AB|C) \label{eq:sum} \\
p(\bar{A}|C) &\equiv 1 - p(A|C) ~. \label{eq:not}
\end{align}
From Equation \eqref{eq:prod} we can derive Bayes' theorem (assuming $p(B | C) \neq 0$),
\begin{equation} \label{eq:bayes}
p(A|BC) = \frac{p(B|AC) ~ p(A | C)}{p(B | C)} ~.
\end{equation}
These are \emph{identities}, holding for any propositions $A$, $B$ and $C$ for which the relevant quantities are defined. While Bayes' theorem often comes attached to a narrative about prior beliefs, updating and conditioning, none of this is essential. In fact, an insistence that Bayes' theorem must be applied in chronological order (``updating") is contrary to D3, which is so crucial to the Bayesian (and Coxian) approach that \citet{Skilling2014} goes so far as to claim that ``probability calculus is forced upon us as the only method which lets us learn from data irrespective of their order". Even if one does not agree with Skilling, assigning known propositions to $B$ and $C$ in Equation \eqref{eq:bayes} is purely for convenience.
When Bayes' theorem is used to calculate the probability of some hypothesis or theory $T$, given evidence $E$ and background information $B$, the corresponding terms in Equation \eqref{eq:bayes} are commonly given the following names: $p(T|EB)$ is the \emph{posterior probability}, $p(T|B)$ is the \emph{prior probability}, $p(E|TB)$ is the \emph{likelihood}, and $p(E|B)$ is the \emph{marginal likelihood}. As noted above, the conjunction $EB$ represents everything that the posterior treats as ``known'', and the separation into $E$ and $B$ (into evidence and background) is purely for convenience.
While not our primary focus here, the assignment of prior probabilities is (at best) an active research problem for the Bayesian. Here, we note \emph{one} important aspect of prior probabilities: they are crucial for penalizing ad-hoc theories. Consider a simple case: suppose that a physical theory attempts to ``cheat the likelihood" by simply adding the data to the theory, $T_\textrm{new} = TU$. This gives the new theory a perfect likelihood: $p(U|T_\textrm{new}B) = p(U|TUB) = 1$. While the likelihood is fooled, the posterior is not because it depends on the prior: $p(T_\textrm{new}|B) = p(TU|B) = p(U|TB) p(T|B)$, and thus $p(T_\textrm{new}|UB) = p(T|UB)$. The lesson is this: don't smuggle data into your theory.\footnote{Consistent application of D3 avoids the ``problem of old evidence" for the objective Bayesian. \citet{Glymour1980} argues that, if we already know evidence $E$, then $p(E|T) = 1$ and $p(E) = 1$, and thus $p(T|E) = p(T)$. This is not how to use Bayes theorem. Even if $E$ is known, we should not take $E$ as given in every probability we calculate. The posterior $p(T|EB)$ takes $E$ as given, but the likelihood $p(E|TB)$ and marginal likelihood $p(E|B)$ do not. Calculating the likelihood uses the same probability function that comes from Cox's theorem; it does not require a new ``ur-probability'' function, generated by supposing ``that one does not fully believe that E'' \citep{Monton2006}. Objective Bayesian probabilities aren't about what any individual knows or believes; they are about what one proposition implies about the plausibility of another. Similarly, there is no need to argue in chronological order, taking life as background information and fine-tuning as new information \citep{Roberts2011}.}
We turn now to testing physical theories. Let,
\begin{itemize} \setlength{\itemsep}{-2pt} \renewcommand{\labelitemi}{$\bullet$}
\item $T$= the theory to be tested. For our purposes, the important thing about a physical theory is that it implies certain expectations about physical scenarios. As a specific example, $T$ may represent a set of symmetry principles, from which we can derive the mathematical form of a Lagrangian (or, equivalently, the dynamical equations), but not the values of its free parameters.
\item $U$ = our observations of this Universe.
\item $B$ = everything else we know. For example, the findings of mathematics and theoretical physics are included in $B$. As I have defined it for our purposes here, the information in $B$ does not give us any information about which possible world is actual. The theoretical physicist can explore models of the universe mathematically, without concern for whether they describe reality.
\end{itemize}
To help in calculating the posterior, we turn to Bayes Theorem,
\begin{equation} \label{eq:bayesTUB}
p(T|UB) = \frac{p(U|TB) ~ p(T | B)} {p(U | B)} ~.
\end{equation}
A physical theory describes the physical world, and so should make claims about which physical scenarios (including empirical data) are more or less to be expected. In the Bayesian framework, this implies that calculating \emph{likelihoods} $p(U|TB)$ is part of the job description of any physical theory. Note that there are several, conceptually distinct sources of the uncertainty quantified by the likelihood. Firstly, there may be stochasticity within the theory itself. This could be because the theory is intrinsically indeterminate, such as (some interpretations of) quantum mechanics, or because $T$ is an \emph{effective} theory that describes physical systems by averaging over --- and thus blurring out --- microphysical details. Secondly, our observations are imprecise, and thus consistent with a range of physically possible universes.
We require two further ingredients to calculate $p(U|TB)$. Firstly, physical theories typically contain free parameters, that is, mathematical constants that appear in the Lagrangian. We will call them, collectively, \ensuremath{\boldsymbol{\alpha}_T}\xspace. If the Lagrangian in question is a fundamental theory of physics, that is, a theory that is not an approximation to another known theory, then these parameters are called \emph{fundamental constants of nature}. Secondly, a description of the physical universe is given by a \emph{solution} to the equations of the theory. The set of solutions to $T$ describes the set of universes that are physically possible according to $T$. To specify a particular solution, we usually require a few more numbers (denoted \ensuremath{\boldsymbol{\beta}_T}\xspace) in the form of initial conditions. (It is a useful simplification to consider \ensuremath{\boldsymbol{\beta}_T}\xspace to be simply a set of numbers.)
To calculate $p(U|TB)$, we \emph{marginalize} over the constants \ensuremath{\boldsymbol{\alpha}_T}\xspace and \ensuremath{\boldsymbol{\beta}_T}\xspace, treating them as nuisance parameters. By the law of total probability,
\begin{equation} \label{eq:totalprob}
p(U|TB) = \int p(U | \ensuremath{\boldsymbol{\alpha}_T}\xspace \ensuremath{\boldsymbol{\beta}_T}\xspace T B) ~ p( \ensuremath{\boldsymbol{\alpha}_T}\xspace \ensuremath{\boldsymbol{\beta}_T}\xspace | T B) \ensuremath{\ \textrm{d}}\xspace \ensuremath{\boldsymbol{\alpha}_T}\xspace \ensuremath{\ \textrm{d}}\xspace \ensuremath{\boldsymbol{\beta}_T}\xspace ~.
\end{equation}
We can call $p( \ensuremath{\boldsymbol{\alpha}_T}\xspace \ensuremath{\boldsymbol{\beta}_T}\xspace | T B)$ the \emph{prior probability distribution of the free parameters} \ensuremath{\boldsymbol{\alpha}_T}\xspace and \ensuremath{\boldsymbol{\beta}_T}\xspace, given $T$. Note that we also need this distribution to infer the values of physical parameters from experiment by calculating their posterior probability.\footnote{Thus, we cannot use experimental constraints to say that we have empirical constraints on the prior right from the start. As the formula shows, we cannot turn empirical evidence $U$ into information about the value of the constant in our universe $p(\alpha|UTB)$ without a prior derived purely from the theory and theoretical background information $p( \alpha | T B)$.} For example, in one dimension,
\begin{equation} \label{eq:paramposterior}
p(\alpha|UTB) = \frac{p(U | \alpha T B) p( \alpha | T B)} { \int p(U | \alpha T B) p( \alpha | T B) \ensuremath{\ \textrm{d}}\xspace \alpha} ~.
\end{equation}
\section{Fine-Tuning as Physics Jargon} \label{S:FTjargon}
Philosophical discussions of the fine-tuning of the universe for life have often failed to recognize the context in which physicists have made their claims. `Fine-tuning', a metaphor that brings to mind a precisely-set analogue radio dial, is used as a technical term in physics. \citet{Donoghue2007} discusses the case of a theory in which a positive measurable quantity $x$ is calculated to be the sum of an unknown bare value ($x_0$) and an estimatable quantum correction ($x_q$). These quantities $x_0$ and $x_q$ are, according to the theory, unrelated. If, however, we discover that the measured value is much smaller than the quantum correction, then the theory is fine-tuned or \emph{unnatural} in a technical sense. To explain the data ($x$), we require that the following coincidence holds: $x_0 \approx -x_q$. However, this cancellation is unexpected and unexplained by the theory, suggesting that we should search for a theory that implies a deeper relationship between these quantities.
We can formalize this intuition using probability theory. Consider a simplified scenario in which a theory $T$ has a free parameter $\alpha$ that ranges from 0 to $R_\alpha$. The theory itself and our background mathematical knowledge $B$ give no reason to prefer any particular value of $\alpha$ in this range, so we represent our state of knowledge with a uniform probability distribution: $p(\alpha|T B) \ensuremath{\ \textrm{d}}\xspace \alpha = \ensuremath{\ \textrm{d}}\xspace \alpha / R_\alpha$. Suppose that the likelihood of the observed data is equal to 1 in a range of width $\Delta \alpha$, and zero outside.\footnote{The reader is invited to generalize this lesson to a narrow Gaussian likelihood, and to the case of multiple, coincidentally-related parameters, as was the case with $x$ above. Remember that the likelihood is normalized over \emph{data} $D$, not over free parameters. That is, it need not be the case that $\int p(D|\alpha T B) \ensuremath{\ \textrm{d}}\xspace \alpha = 1$.} Then,
\begin{equation} \label{eq:totalprobD}
p(D|TB) = \int_0^{R_\alpha} p(D | \alpha T B) p( \alpha | T B) \ensuremath{\ \textrm{d}}\xspace \alpha \approx \frac{\Delta \alpha}{R_\alpha} ~.
\end{equation}
If $\Delta \alpha \ll R$, then $p(D|TB) \ll 1$. That is what a physicist means by \emph{fine-tuned}. It is a special case of Bayesian theory testing: fine-tuned or unnatural theories have very small likelihoods, and the more fine-tuned, the smaller the likelihood. Now, a small likelihood is not a sufficient reason to discard a theory. Rather, it presents a explainable but as-yet-unexplained fact, a good reason to examine our background assumptions and/or search for an alternative theory, perhaps one whose free parameters are not as fine-tuned or in which the quantity $\alpha$ is not a free parameter at all.
In particular, the likelihood of the data given a theory with a free parameter becomes smaller when experiments measure the free parameter more precisely. This doesn't necessarily mean that the posterior probability of the theory becomes smaller. For example, if the relevant parameter is free in all known, viable physical theories, then measuring it more precisely won't make much of a difference. But if an alternative theory doesn't have the free parameter, then better measurement can increase its posterior probability. Suppose that in theory $T_1$, $\alpha$ is a free parameter which ranges between 0 and 1, while in theory $T_2$, $\alpha$ is constrained by a symmetry principle to be precisely 0.5. Suppose data $D$ implies $\alpha = 0.45 \pm 0.2$. (Continuing the simplification above, consider a step function that is one in the quoted range and zero outside.) This data is hardly decisive. But suppose the data $D_\textrm{new}$ from an improved experiment implies $\alpha = 0.50002 \pm 0.00005$. Supporters of $T_2$ are suitably thrilled, because while $p(D_\textrm{new}|T_2 B) = p(D|T_2 B)
= 1$, $p(D_\textrm{new}|T_1 B) \ll p(D | T_1 B)$.
How are we to guard against theories that ``pre-cook" their parameters to match the data? As with the ad hoc theories discussed above, the prior is the key. Consider again theory $T_1$, and create a new theory $T'_1 = TA$ by adding the statement $A = $``$\alpha = 0.5 \pm 0.00005$". Then, the theory's likelihood is ideal $p(D_\textrm{new} | T'_1 B)$. However, the prior probability of the `theory' is now $p(T'_1 | B) = p(A|TB) p(T | B)$, and $p(A|T_1B) \approx p(D | T_1 B)$. The pre-cooked theory fares no better.
\section{Free Parameters and their Limits}
In the examples given above, we simply postulated the ranges of the free parameters $R_\alpha$. However, what if the range is infinite? Assigning a constant, normalized probability measure over an infinite range is impossible: there is no probability distribution $p$ such that a) $p(x)$ is constant, and b) $\int_0^\infty p(x) \ensuremath{\ \textrm{d}}\xspace x = 1$. So, how should we test a theory in such circumstances?
Under certain circumstances, a non-normalizable probability distribution can be tolerated. For example, if we are interested in using Equation \eqref{eq:paramposterior} to calculate the posterior probability distribution $p(\alpha | UTB)$ of the free parameter $\alpha$, then we may only need to assume that the prior probability distribution is constant near the peak of the likelihood $p(D | \alpha T B)$. In fact, many non-normalizable prior probability distributions (not just constant ones) can give sensible answers in such circumstances. However, this will not do when it comes to calculating the likelihood of the theory $p(D | T B)$ via Equation \eqref{eq:totalprobD}, since the constant value of the prior does not cancel out.
Infinities are well-known to produce apparent paradoxes in probability theory. In Chapter 15 of \citet{Jaynes2003}, a number of these paradoxes are carefully reviewed and discussed. Jaynes notes their common cause: passing from a well-defined probability problem (finite or convergent) to a limit –-- infinite magnitude, infinite set, zero measure, improper probability distribution function (PDF), or some other kind --– without specifying how the limit is approached. If one then forgets the limiting process, and asks a question whose answer depends on said process, contradictory results are inevitable but hardly surprising.\footnote{``It is not surprising that those who persist in trying to evaluate probabilities directly on infinite sets [and] trying to calculate probabilities conditional on propositions of probability zero, have before them an unlimited field of opportunities for scholarly looking research and publication --- without hope of any meaningful or useful results." (pg. 485)}
Jaynes's conclusion: ``based on some 40 years of mathematical efforts and experience with real problems –-- is that, at least in probability theory, an infinite set should be thought of only as the limit of a specific (i.e. unambiguously specified) sequence of finite sets." In particular, measure theory is to be used with caution, as it can easily disguise infinities. Further, ``in practice we will always have some kind of prior knowledge \ldots [that implies that the location and scale] parameters $(a,b)$ cannot vary over a truly infinite range.'' We can confine our attention to ``the range which \ldots expresses our prior ignorance" (pg. 396).
We intend here to take Jaynes's advice. But what could this extra knowledge be that defines the limiting process or constrains the free parameters to a finite range? Keep in mind that fundamental constants and initial conditions are defined by their theories; they have no ``theory-independent" existence.\footnote{The \emph{measurement} of a particular constant, stripped to its barest elements, is theory independent. For example, we can still measure the deflection of a torsion balance in the presence of a given mass. But only Newton's theory of gravity will tell us how to combine our measurements into the fundamental constant $G$.} Thus, in the context of testing a physical theory by calculating the likelihood of the data $p(D|TB)$, the only ``extra'' information that could constrain the parameters is \emph{the theory itself}. A physical theory, to be testable, must be sufficiently well-defined as to allow probabilities of data (likelihoods) to be calculated, at least in principle. Otherwise, the theory cannot tell us what data we should expect to observe, and so cannot connect with the physical universe. If the theory contains free parameters, then since the prior probability distribution of the free parameter is a necessary ingredient in calculating the likelihood of the data, the theory must justify a prior. In summary, a theory whose likelihoods are rendered undefined by untamed infinities simply fails to be testable. In essence, it fails to be a physical theory at all.
Note that we have not yet raised the fine-tuning of the universe for life. These are the conditions for a viable, testable physical theory.
How, then, do current theories of physics avoid problematic infinities? Consider Newton's theory of gravity ($N$), and its free parameter $G$. At first glance, it may appear that $G$ could be any positive real number, and no specific number is preferred, and hence Newtonian gravity fails to provide a normalizable prior probability distribution for its parameter $p(G|NB)$. However, $G$ has physical units: in SI units, m$^3$ kg$^{-1}$ s$^{-2}$. By changing our system of units, we can give $G$ any value we please. Thus, $p(G|NB)$ is ill-posed because $G$ is arbitrary. We need more physics to specify a system of units (such as the Planck units), or a specific experimental setup.
Suppose we are presented with the Cavendish experiment $C$: a rod of length $L$ with masses $m$ at each end is suspended from a thin wire, whose torsion creates a natural oscillation period $T$. Two masses $M$ are brought to a distance $r$ from the masses $m$, and the angular deviation of the rod $\theta$ is measured. The inferred value of $G$ is $2\pi L r^2 \theta / (M T^2)$. Knowing only the details of the setup $C$, but not the data $\theta$, is the possible range of $G$ infinite? No, because if $G$ were sufficiently large, the masses $m$ would crush the rod. Since we are taking $C$ as given [$p(G|CNB)$], the fact that $C$ is a stable experimental apparatus \emph{at all} places an upper limit on $G$, removing the problem of an infinite possible range for the free parameter.\footnote{It also places a lower limit on negative values for $G$, since strongly repulsive gravity would also destroy the rod. Note that this is not the usual Bayesian way of testing theories, as we would be required to calculate the prior $p(N|CB)$. This is almost ``cheating" --- we are smuggling data (the existence of the apparatus) into the calculation of the prior. At best, given that the apparatus, considered in isolation from the rest of the universe, relies only on non-gravitational physics, we might argue that $C$ is irrelevant and hence $p(N|CB) = p(N|B)$. The point of this example is that $p(G|NB)$ is not well-posed; this point is unaffected.}
The moral of the story is that, whether measuring constants and testing the theory itself, many physical theories cannot be considered in isolation. Our ``theory" $T$ must be able to describe not just the system of interest but also the measuring apparatus. For example, in calculating the likelihood of galaxy redshifts in cosmology, we not only need Einstein's General Relativity to describe expanding space, but also quantum theory to permit the interpretation of atomic emission spectra, from which redshifts are derived.
In modern physical theories, we can define our system of units using three fundamental parameters. A particularly useful system involves setting $G = c = \hbar = 1$, where $c$ is the speed of light and $\hbar$ is the reduced Planck constant. This defines \emph{Planck units}, in which the standard units of mass, length and time are $m_\textrm{Pl} = 2.17651 \times 10^{-8}$ kg, $1.616199 \times 10^{-35}$ m and $5.39106 \times 10^{-44}$ s, respectively. It then makes no sense to vary $(G,c,\hbar)$, to ponder what would happen if they were different, or to ask for their prior probabilities. The advantage is that, once units have been fixed, varying the other fundamental parameters (electron mass, quark masses, strengths of the fundamental forces etc.) makes a real difference to the universe, rather than simply making an essentially identical, scale-model of our universe.
\citet{Tegmark2006} list 26 parameters of the standard model of particle physics, and five parameters of the standard model of cosmology.\footnote{Cosmologists usually consider six parameters, adding the Hubble parameter $H_0$ to Tegmark et al's list. This parameter is not a constant, and essentially measures the age of the universe today. Its value is thus linked to our existence as observers, and so is not predicted by fundamental theory. If it is predicted at all, it is by anthropic arguments \citep{Lineweaver2007}.} Of these, one particle physics parameter and four cosmological parameters have units: the Quadratic Higgs coefficient (or, equivalently, the Higgs vacuum expectation value $v$), the cosmic dark energy density $\rho_\Lambda$, baryon mass per photon $\xi_b$, cold dark matter mass per photon $\xi_c$, and neutrino mass per photon $\xi_\nu$. In fact, the cosmological ``per photon'' quantities are the result of asymmetry in physics. For example, $\xi_b$ is derived from matter-antimatter symmetry: $\xi_b \sim m_\textrm{proton} ~ \eta$, where $\eta = (n_b - n_{\bar{b}}) / n_{\gamma}$ is the dimensionless Baryon asymmetry parameter. \citet{Tegmark2006} consider a model for dark matter in which $\xi_c$ is linked to the Peccei-Quinn (PQ) symmetry breaking scale ($f_a$). We thus will consider these parameters to be derived from more fundamental, dimensionless asymmetry parameters.
This leaves the dimensional parameters $v$ and $\rho_\Lambda$. Is there anything in the relevant physical theories that limits their range of possible values? Yes --- the Planck scale. Famously, we do not have a quantum theory of gravity. That is, we do not know how to describe gravity within a quantum framework. Naively combining quantum mechanics and general relativity, we can calculate that if a single particle were to have a mass equal to the Planck mass, then its black hole (Schwarzschild) radius would be larger than its quantum size (Compton length), and thus it would become its own black hole. The point is not that we think that this would actually happen, but rather that can't possibly trust our theories in this case. The Planck mass represents an upper boundary to any single-particle mass scale in our current theories. A lower boundary is provided by zero, since quantum field theory breaks down for negative masses; even if it didn't, a lower bound would be given by $-m_\textrm{Planck}$. Thus, the theory itself constricts the value of $v$ to the range $[0,m_\textrm{Planck})$, and $\rho_\Lambda$ to the range $[0,m_\textrm{Planck}^4)$ \citep{Wilson1979,Weinberg1989,Dine2015}. Outside of these ranges, our current theories cannot be trusted. Indeed, the very concepts that underlie our constants --- space, time, mass, energy --- may cease to be meaningful in the quantum gravity regime. This constraint comes from $T$ itself
What about dimensionless parameters? Some are phase angles, which have a finite range $[0, 2\pi)$. Others are coupling constants, which might vary over an infinite range. Two considerations can result in a normalizable prior distribution. Firstly, the masses of certain particles depend on the coupling constants, and increasing the constant beyond a certain value will send the mass beyond the Planck scale. For example, the mass of the electron is $m_e = v \Gamma_e /\sqrt{2}$, where $\Gamma_e$ is the dimensionless electron Yukawa coupling. If $\Gamma_e > \sqrt{2} m_\textrm{Planck} / v$, then the electron mass is greater than the Planck mass, and we have reached the border of our theory. Secondly, dimensionless parameters are expected to be of order unity. This is the idea behind the definition of \emph{naturalness} due to \citet{tHooft1980}:
\begin{quote}
a physical parameter or set of physical parameters is allowed to be very small [compared to unity] only if the replacement [setting it to zero] would increase the symmetry of the theory.
\end{quote}
\citet{BT86} offer the following justification for a preference for order-unity parameters in physics. When physicists use dimensional analysis to predict the form of an equation, we expect that the associated dimensionless constant is a combination of a few geometric factors (2, $\pi$, etc), with each contribution equally likely to be multiplied or divided. Thus, we consider dimensionless constants that are many orders of magnitude away from one to be \emph{unnatural} \citep[see, for example, the discussion in][]{Dine2015}. Perhaps simplicity also plays a role: a constant that is equal to unity is effectively no constant at all! Even if the possible range for a dimensionless constant is infinite, the principle of indifference does not force us to put a constant prior probability distribution over this range. We can debate the most appropriate form of the distribution, but there are plenty of functions that are non-zero over an infinite range and yet normalizable.
In short, there are normalization problems for a physical theory with free parameters that \emph{both} vary over an infinite range and are uniformly distributed. The standard models of particle physics and cosmology avoid these problems as follows. Dimensional parameters do not vary over an infinite range; they are bounded by the Planck scale. Dimensionless parameters might not vary over an infinite range, and common practice in the physical sciences assumes that parameters of order unity are more probable, so a uniform probability distribution is not forced upon us by the principle of indifference.
What about reparameterizations? Given any finite range of a free parameter, say $\alpha \in [a,b]$, we can write an equivalent theory in terms of a parameter $\gamma$ which varies over an infinite range. Here, we again rely on the theory to tell us which of these parameterizations is simpler. Often there are symmetries and invariances to guide our choice, for example the famous prior of \citet{Jeffreys1946}. Given that it cannot be the case that both $p(\alpha|TB) \ensuremath{\ \textrm{d}}\xspace \alpha$ and $p(\gamma|TB) \ensuremath{\ \textrm{d}}\xspace \gamma$ are constant, both cannot honestly represent our state of knowledge. We must decide, or more exactly, the theory itself must aid our determination of the prior. More than this cannot be said without considering the details of individual theories. But as \citet{Collins2009} has pointed out, by reparameterizing a theory we could make any prediction, even one as successful as QED's one in a billion prediction of the gyromagnetic moment of the electron, seem trivial.
\section{Fine-Tuning for Life}
We turn to the fine-tuning of the universe for life. We can follow the formalism of Section \ref{S:FTjargon}, applying it to this fact about our universe: it has developed and supports physical life forms ($L$). For our purposes, a typical dictionary definition will do: a living entity has the capacity to grow, metabolize, actively resist outside disturbance, and reproduce. A precise definition of life is not required, for the following reason. We would like to be able to place firm boundaries in parameter space between possible universes that would develop and support life and those that would not. However, this is not practically possible, as we do not know the sufficient conditions for abiogenesis. What we can do is consider a conservative outer boundary associated with sufficient conditions for lifelessness. For example, if the cosmological constant were negative, and its absolute value $10^{90}$ times smaller than the Planck scale (rather than $10^{120}$ in our universe), space would recollapse into a big crunch in one minute. This, it seems, is a sufficient condition for a lifeless, physical-observer-less universe.
We consider the likelihood that a particular universe is life-permitting (or, in practice, not life-prohibiting), given the laws of nature as we know them. As above, we marginalize over the free parameters,
\begin{equation} \label{eq:totalprobL}
p(L|TB) = \int p(L | \ensuremath{\boldsymbol{\alpha}_T}\xspace \ensuremath{\boldsymbol{\beta}_T}\xspace T B) ~ p( \ensuremath{\boldsymbol{\alpha}_T}\xspace \ensuremath{\boldsymbol{\beta}_T}\xspace | T B) \ensuremath{\ \textrm{d}}\xspace \ensuremath{\boldsymbol{\alpha}_T}\xspace \ensuremath{\ \textrm{d}}\xspace \ensuremath{\boldsymbol{\beta}_T}\xspace ~.
\end{equation}
The scientific literature on fine-tuning has identified life-permitting regions of the constants, outside of which life is seemingly physically impossible, that is, $p(L | \ensuremath{\boldsymbol{\alpha}_T}\xspace \ensuremath{\boldsymbol{\beta}_T}\xspace T B)$ is extremely small. The most significant constants on which such constraints can be made are as follows. From particle physics: the Higgs vacuum expectation value (vev), the masses (or, equivalently, Yukawa parameters) of the electron, neutrinos (the sum of the three species), and up-, down- and strange-quark, the strong force coupling constant, and the fine-structure constant. From cosmology: the cosmological constant, the scalar fluctuation amplitude (“lumpiness”, $Q$), the number of spacetime dimensions, the baryonic and dark matter mass-to-photon ratios, and the initial entropy of the universe. These cases are discussed in \citet{Hogan2000}, \citet{Tegmark2006}, \citet{Barnes2012} and \citet{Schellekens2013}.
In particular, for the Higgs vev ($v$) and the cosmological constant ($\rho_\Lambda$), the Planck scale is not merely the maximum possible value of the parameter. Quantum corrections contribute terms of order $m_\ro{Planck}^2$ and $m_\ro{Planck}^4$ to $v^2$ and $\rho_\Lambda$ respectively, meaning that the Planck scale is the \emph{natural} scale for these parameters \citep{Dine2015}. The smallness of these parameters with respect to the Planck scale is known as the hierarchy problem and the cosmological constant problem respectively. Their life-permitting limits are as follows. If $v^2/m_\ro{Planck}^2 \lesssim 6 \ten{-35}$, then hydrogen is unstable to electron capture; if $v^2/m_\ro{Planck}^2 \gtrsim 10^{-33}$ then no nuclei are bound and the periodic table is erased. If $\rho_\Lambda / \rho_\ro{Planck} \lesssim - 10^{-90}$, the universe would recollapse after 1 second; if $\rho_\Lambda / \rho_\ro{Planck} \gtrsim 10^{-110}$, then no structure whatsoever would form in the universe. These limits come from \citet{Hogan2000}, \citet{Tegmark2006}, \citet{Barnes2012}, \citet{Schellekens2013}, and references therein.
In light of the naturalness of the Planck scale, our ignorance of these parameters can be honestly represented by prior probability distributions that are uniform in $v^2/m_\ro{Planck}^2$ and $\rho_\Lambda / \rho_\ro{Planck}$, resulting in probabilities $\sim 10^{-33}$ and $10^{-90}$ respectively. Indeed, a uniform prior seems to be conservative. We could argue for a prior that peaks at the Planck scale, which would decrease these probabilities even further.
These probabilities are surprisingly small in the following sense. $L$ could have been a generic fact about universes which obey the laws of nature as we know them. That is, it could have been the case that some form of life would exist for a wide range of values of the free parameters. The smallness of the likelihood of $L$ is not the result of considering a highly specified, precise outcome. For example, the likelihood of experimental data can depend on the precision of our measurements. The probability of a radioactive nucleus decaying in a given 1 second interval is smaller than the probability in a given 1 minute inverval. By contrast, the likelihood of $L$ does not depend on the precision of our measurements.\footnote{Note that the degree of fine-tuning is not the degree of accuracy of various order-of-magnitude models used in the early fine-tuning literature, contra \citet{Klee2002}.}
The smallness of these likelihoods, as noted above, proves nothing by itself. But very small probabilities should make us suspicious --- perhaps something unlikely happened, or perhaps one of the assumptions that went into the probability calculation needs to be challenged.
\section{Responses to Critics}
\subsection{McGrew, McGrew \& Vestrup, and Colyvan, Garfield \& Priest}
MMV consider the fine-tuning argument, contending that we cannot conclude anything from the narrow life-permitting ranges of the fundamental parameters because ``the Euclidean measure function \ldots is not normalizable. If we assume every value of every variable to be as likely as every other --- more precisely, if we assume that for each variable, every small interval of radius $e$ on $R$ has the same measure as every other --- there is no way to `add up' the regions of $R^K_+$ so as to make them sum to one. If they have any sum, it is infinite."
Similarly, CGP argue that ``the probability of finding the constant in question in \emph{any} finite interval is zero. This makes a mockery of the claim that \emph{the class of life-permitting universes, in particular}, is improbable."
MMV and CGP's concerns are perfectly legitimate, but are not specifically a problem with fine-tuning. They have identified the conditions under which a physical theory, scuppered by infinities, fails to produce likelihoods of data --- any data. Consider Equations \eqref{eq:totalprob} and \eqref{eq:totalprobL}: if $p( \ensuremath{\boldsymbol{\alpha}_T}\xspace \ensuremath{\boldsymbol{\beta}_T}\xspace | T B)$ is undefined, we can't calculate $p(L|TB)$ because we can't calculate $p(D | TB)$ for any data $D$. If the fundamental theories of modern physics cannot \emph{in principle} justify a normalizable prior distribution over their free parameters, then we have bigger problems than fine-tuning. We don't have a testable theory at all. We can't derive predictions, can't model data, and must go back to square one.
We have seen above how modern theories avoid the problems raised by MMV and CGP. The set of possible values of a constant is dictated by the theory. If the model (or combination of models, such as gravity plus quantum mechanics) fails to be mathematically consistent beyond a certain value of the constant, then this limits the constant's possible range. The constant only lives within the theory, so where the theory ceases to be coherent, the constant ceases to be possible. In particular, mass-energy (including energy density) parameters are bounded between zero and the Planck scale. With dimensionless parameters, we should not assume that every value is as likely as every other, because `order-unity' ($\sim 1$) values are more likely.
CGP consider the idea that ``the laws of physics themselves set limits on the values certain constants can take'' as a possible solution to the problem of non-normalizability. They reply that ``without some independent argument for the shape of the distribution in question, this version of the argument simply begs the question."
On the contrary, \emph{some} resolution to this problem is necessary for any law of physics to be tested using Bayesian probabilities. Without a prior distribution for the free parameters, no likelihood can be calculated and no predictions made. Moreover, the shape of the distribution need only honestly reflect our ignorance of the parameter in the absence of experimental data. We can debate the shape that fulfills this requirement; the problem of prior probabilities in Bayesian data analysis is an open research question, with ``catalogues" of distributions available to the discerning scientist \citep{Yang1997}. We have argued above that, for ``natural'' parameters (in the physicist's sense of the word), a uniform distribution over the finite range dictated by the laws themselves is reasonable and indeed conservative. But for such a distribution not to exist at all would hamstring the entire Bayesian approach to testing physical theories.
\subsection{Halvorson}
Many of the criticisms of \citet{Halvorson2014} are relevant only to the fine-tuning argument for God, which is beyond our purview here. We comment only on the important distinction between credences and chances in testing physical theories in general and in the problem of fine-tuning in particular. Here, we use the following definition of David Lewis's Principal Principle: ``a person Ms degree of belief (at t) that A conditional on the proposition that the chance of A (at t) is x should be equal to x; with the qualification that at t she has no information about A that is inadmissible. Information about A is inadmissible if it is information about A over and above information about A's chance.'' \citep{Loewer2004}.
Halvorson considers an urn containing with 1 yellow ball and 99 purple balls. Because the balls are otherwise identical, ``any two balls are equally likely to be drawn''. Thus, ``according to the principal principle, the rational credence for drawing a yellow ball, given [background information], is 0.01." This probability is unaffected by any additional hypothesis regarding how the balls were placed in the urn. Similarly, says Halvorson, the probability of the universe and its constants is set by the laws of nature, and is unaffected by any deeper story about how the universe came to be.
Note that Halvorson's view allows for a positive scientific case to be made for the multiverse. If a cosmologist proposes a theory on which a multiverse is almost inevitable, then the theory could assign an almost unity likelihood to our observation of a life-permitting universe. Thus, that theory would be confirmed over a theory that implied a single, fine-tuned universe. But note carefully what the cosmologist cannot do: they cannot take a certain physical theory, find a set of initial conditions that will produce a multiverse, and then propose that such initial conditions should be privileged in some way. This, for Halvorson, is inadmissible: the theory sets the probability of initial conditions, and no further assumption (theism, multiverse or whatever) can change this assignment. It should be noted that advocates of the \emph{past hypothesis} do exactly what Halvorson prohibits: ``the distribution of probabilities over all of the possible exact initial microconditions of the world is uniform \ldots over those possible microconditions of the universe which are compatible with the initial [very low-entropy] macrocondition \ldots, and zero elsewhere.'' \citep{Albert2015}
My central concern here is the claim that chances of initial conditions are set by the theory. This is mistaken, and for the same reason that the claim about the urn is mistaken. It does not follow from the fact that the ball are otherwise identical that any two balls are equally likely to be drawn. That assumption is \emph{additional} to the setup specified by Halvorson's background information (1 yellow, 99 purple, otherwise identical). Putting balls in an urn, even near-identical ones, does not specify the mechanism by which a ball is selected, and so does not specify any chances at all.
If we remove the assumption that ``any two balls are equally likely to be drawn'', we do not have any chances, but we can still reason about credences as follows. Either the ball will be drawn by a process that is indifferent $(I)$ to the colour of the ball, or it will not $(\bar{I})$. On $I$, the chances of all balls are equal, and so by the principle of indifference, $p(Y|I ~ B) = 0.01$, where $B$ is the relevant background information about the urn and its contents. On $\bar{I}$, however, we have a process that is trying to select one colour of ball. We do not know which colour is preferred, so we say that the probability of a yellow preference ($P_Y$) is equal to the probability of a blue preference ($P_B$), so that $\bar{I} = P_Y + P_B$. Then, by the law of total probability:
\begin{equation} \label{eq:probyellow}
p(Y | B) = p(Y | IB)~ p(I|B) + p(Y|P_Y B) ~ p(P_Y | B) + p(Y | P_B B) ~ p(P_B | B) ~,
\end{equation}
where, by assumption, $p(I|B) = 1/2$ and $p(P_Y | B) = p(P_B | B) = 1/4$. The remaining unknowns are $p(Y|P_Y B)$ and $p(Y|P_B B)$, the probability of someone biased towards yellow successfully drawing yellow, and the probability of someone biased towards blue accidentally drawing yellow. These depend on the urn --- if locating and selecting the yellow ball is made near-impossible by (say) a deep, dark interior and a narrow neck, then even a biased selector will do no better than random: $p(Y|P_Y B) = p(Y|P_B B) \approx 0.01$, in which case $p(Y | B) = 0.01$. However, if the urn is transparent and the balls easily manoeuvred, then $p(Y|P_Y B) \approx 1$ and $p(Y|P_B B) \approx 0$, in which case $p(Y | B) \approx 0.25$.
This is the Bayesian approach: we model our state of knowledge, marginalizing over our assumptions. What we do not do is invoke the principal principle, set $p(Y | B) = 0.01$, and use this as a constraint on the probability of the various scenarios for how the ball was drawn. Further, the existence of a measure $\mu$ over the set of possible draws from the urn which respects a certain symmetry (in this case, permutation independently of colour) implies neither chances nor credences. The assumption of equal chances is \emph{additional} to the facts about the urn, including the measure.
We apply this to the laws of nature. A measure over the set of initial conditions of a theory, even one that follows naturally from the theory itself in some sense (respecting symmetry, for example), does not imply the assignment of chances to those initial conditions. Indeed, such a claim would be highly problematic. Are we to think that all theories imply an actually existing ensemble of universes, with properties distributed in accordance with the measure over initial conditions? That would add significant ontological baggage to our laws. But surely this is mistaken. Suppose I set up a computer simulation of the universe. I choose the laws that my simulation will follow, and those laws may motivate a measure over their space of initial conditions. It does not follow that I am obliged to select initial conditions according to a chancy process that respects that measure. Furthermore, by definition, no physical mechanism sets initial conditions, especially of the whole universe. So it is difficult to understand how this could be understood as a ``single-case'' probability or chance or propensity.
The connection between the measure and credences is even less tight. It is sometimes supposed that fine-tuning arguments assume that for our universe to be improbable, one must postulate that the universe is selected at random by some stochastic physical mechanism. Our ignorance of that mechanism, then, would seem to make fine-tuning a purely speculative exercise. But Bayesian model selection does not rely solely on physical chances. Indeed, testing physical theories cannot use chances alone if it wants to calculate the probability \emph{of} a theory. Outcomes, or physical events, can be chancy, but theories are not.
\section{Conclusion}
When faced with the question ``why is the universe the way it is?", we might want to consider the related question ``what other ways could the universe have been?". To approach this question systematically, we can take the deepest known laws that describe how the physical universe works and look for ways to vary them. A particularly tractable way is to vary their free parameters. The set of solutions to those laws, usually represented by the set of possible initial conditions, is precisely the set of scenarios that the law deems ``physically possible''. And the constants that appear in the laws themselves have long attracted the attention of theoretical physicists as being in need of deeper explanation.
It is an interesting fact, then, that this search for other ways that the universe could have been has overwhelmingly found lifelessness. This lifelessness is surprising in the way that any fine-tuned parameter in physics is surprising: it is improbable. By the standards of Bayesian model selection, an explanation of this fact should be sought.
\section*{Acknowledgments}
Supported by a grant from the John Templeton Foundation. This publication was made possible through the support of a grant from the John Templeton Foundation. The opinions expressed in this publication are those of the author and do not necessarily reflect the views of the John Templeton Foundation.
|
\section{Introduction}
Since the Hubbard model, a simple tight-binding model with
on-site repulsion,
was formulated in the early 1960s,
numerous attempts have been made on it to understand
mechanisms for itinerant-electron ferromagnetism \cite{Lieb95,Tasaki98a,Tasaki98b}.
To date, some rigorous examples of ferrimagnetism \cite{Lieb89},
ferromagnetism \cite{Nagaoka66,Mielke91a,Mielke91b,Mielke92,Tasaki92,Mielke93,MielkeTasaki93,Tasaki95,Tasaki2003,TanakaUeda2003,Lu2009}
and metallic ferromagnetism \cite{TanakaTasaki2007,TanakaTasaki2016}
in the Hubbard model have been proposed,
and it is well recognized that an interplay between
the quantum mechanical motion of electrons and repulsive
interaction between electrons does generate ferromagnetism.
Among the others, flat-band models
proposed by Mielke and Tasaki
provided a significant breakthrough in understanding of itinerant-electron
ferromagnetism \cite{Mielke91a,Mielke91b,Mielke92,Tasaki92,Mielke93,MielkeTasaki93}.
The flat-band models in common have multi single-electron bands
including a flat (highly degenerate) lowest band.
In these models, one finds that
electrons occupy the flat band in order to minimize
the kinetic energy
and then
the spins of electrons occupying the flat band
align parallel to each other, i.e., align ferromagnetically,
in order to avoid an increase of energy
due to on-site repulsion.
The next important step was also taken by Tasaki, who proved
that the ferromagnetism in flat-band models on lattices constructed
by the cell-construction method
is stable against perturbations which turn flat bands into
dispersive ones
provided that the on-site repulsion is sufficiently large \cite{Tasaki95,Tasaki2003}.
Note that
an electron has a tendency to occupy
a lower kinetic energy state when the repulsion is small.
Tasaki's models which have dispersive bands thus exhibit
the Pauli paramagnetism when the on-site repulsion is vanishing,
and remain nonferromagnetic when the on-site repulsion is small.
Tasaki's models could describe the true competition between
the kinetic energy and the on-site repulsion,
and established that the spin-independent repulsion
can cause the itinerant-electron ferromagnetism.
Although the stability of ferromagnetism in Tasaki's models
is shown in rather general settings,
as for Mielke's flat-band models on line graphs,
less is known about stability or instability of ferromagnetism
in perturbed nearly-flat-band models.
Here we note that
there are no band gaps above flat lowest bands
in Mielke's flat-band models,
whereas
there are finite band gaps in
Tasaki's models and corresponding (unperturbed) flat-band models.
This is an essential difference.
The occurrence of ferromagnetism in Tasaki's models,
at least at a heuristic level,
can be understood as a consequence
of the band gap as follows.
The band gap enforces the electrons to occupy
the lowest nearly-flat band while the on-site repulsion
forbids double occupancy
of sites.
Then, despite that the lowest band is dispersive,
the situation is almost as in the flat-band models,
and the system exhibits ferromagnetism.
In fact,
a certain parameter in hopping amplitudes which controls the energy gap
in the band structure plays
an important role
in the rigorous proof of ferromagnetism in Tasaki's models.
On the other hand,
it is more subtle and difficult
to show the stability of ferromagnetism against a perturbation
in Mielke's flat-band models
since there might be various low energy excitation modes
reflecting a gapless nature of the band structure.
Some examples of nearly-flat-band models related to Mielke's flat-band models
in two dimensions, such as models on the kagome lattice and
on the regular lattice of corner sharing tetrahedra \cite{TanakaUeda2003,Lu2009},
have been proposed and proved to have ferromagnetic ground states;
as far as we know, however, there are no examples
in three or more dimensions.
In this paper, we propose a version of the Hubbard model which
has a gapless nearly-flat lowest band
and exhibits ferromagnetism in the ground state in two or more dimensions.
Our model is defined on a lattice obtained by placing
a site on each edge of the hypercubic lattice.
In a certain case, it is reduced
to a gapless flat-band model.
In two dimensions the corresponding flat-band model
is Mielke's flat-band model on the regular lattice of
corner sharing tetrahedra.
In three or more dimensions, the corresponding flat-band
model is slightly different from Mielke's one,
but they are closely related.
Although the perturbative hopping term considered in this paper
is rather special,
our model is simple such that all the lattice sites are identical
and electron hopping is assumed to be only between nearest
and next nearest neighbor sites.
We believe that our model is helpful for a better understanding
of ferromagnetism in the Hubbard model.
\section{Definition of the model and the main result}
\label{s:Def}
\subsection{Definition of the lattice}
\label{subs:Def lat}
We start by describing the lattice~$\La$ on which our
Hubbard Hamiltonian is defined.
Let $G=(\calV,\calE)$
be a graph,
where $\calV$ is a set of vertices and $\calE$ is a set of edges.
The vertex set $\calV$ is assumed to be
a subset of $\Bbb{Z}^\nu$ with $\nu\ge2$,
\begin{equation}
\calV=\left\{ \alpha=(\alpha_1,\dots,\alpha_\nu)~|~
\alpha_l\in\Bbb{Z},
-\frac{L-1}{2} \le \alpha_l \le \frac{L-1}{2}
\mbox{ for $l=1,\dots,\nu$}
\right\} \, ,
\end{equation}
where $L$ is an odd intege
\footnote
The reason why we choose $L$ to be an odd integer is
only technical.
We can treat the case of even $L$ with a slight modification.}.
We impose periodic boundary conditions in all direction
\footnote
This condition is also imposed to simplify the argument.
It is easy to apply the present method to the model
with open boundary conditions.}.
Let $e(\alpha,\beta)$ be a line segment between
nearest neighbor vertices~$\alpha$ and $\beta$ in $\calV$.
The edge set $\calE$ is a collection of these line segment
\footnote
For $\alpha$ in $\calV$ or $x$ in $\La$ defined by~\eqref{eq:Lambda},
$|\alpha|$ or $|x|$ denotes the usual Euclidean distance.
For a set $X$,
we use the same symbol $|X|$ to denote the number of
elements in $X$.},
\begin{equation}
\calE=\left\{ \bond{\alpha,\beta}~|~\alpha,\beta\in {\calV},
|\alpha-\beta|=1
\right\}\, .
\end{equation}
For each edge~$\bond{\alpha,\beta}$ in $\calE$,
we denote by $m(\alpha,\beta)$ the point taken
in the middle of $\alpha$ and $\beta$.
Then we define the lattice~$\La$ as a collection of these mid-points
\begin{equation}
\label{eq:Lambda}
\La = \{ x = m(\alpha,\beta)~|~\bond{\alpha,\beta}\in {\calE}\}\, .
\end{equation}
Before proceeding to the definition of our Hubbard Hamiltonian,
let us introduce some more notation.
Let $\calP$ be the set of all regular squares with side-length 1
whose corners are located at vertices in $\calV$.
For four edges of a regular square $p$ in $\calP$,
we find the corresponding mid-points,
sites in $\La$, which we denote by $x_i(p)$ with $i=1,2,3,4$.
For convenience, we assume that $x_3(p)$ denotes
the site opposite to that denoted by $x_1(p)$.
Each edge whose mid-point is $x$ is shared by $2(\nu-1)$
regular squares in $\calP$;
we denote by $\calP(x)$ the collection of these regular squares.
Each vertex $\alpha$ in $\calV$ is shared by $2\nu(\nu-1)$ regular
squares in $\calP$;
we denote by $\calP_\alpha$ the collection of
these regular squares.
Each vertex $\alpha$ in $\calV$ is shared by $2\nu$ edges,
$\bond{\alpha,\alpha+\delta_l}$ and
$\bond{\alpha,\alpha-\delta_l}$ with $l=1,\dots,\nu$,
where $\delta_l$ denotes the unit vector along the $l$-axis.
We define $\calC_\alpha$ as the collection of the mid-points of
these edges
\begin{equation}
\calC_\alpha = \{ x = m(\alpha,y)~|~y=\alpha\pm\delta_l, l=1,\dots,\nu\}\, .
\end{equation}
The lattice structure for $\nu=2$ is shown in Fig.~\ref{fig:set of squares}.
\begin{figure}
\begin{center}
\includegraphics[scale=1]{square1.eps}
\end{center}
\caption{The lattice structure for $\nu=2$.
Thin lines represent edges in $\calE$
and filled circles represent lattice sites.
For example, the subsets
$\calP(x)$ and
$\calP_\alpha$ of $\calP$
are given by
$\calP(x)=\{p_1,p_2\}$ and
$\calP_\alpha = \{p_1,p_2,p_3,p_4\}$, respectively.
The subset $\calC_\alpha$ of the lattice sites is
$\{w,x,y,z\}$.}
\label{fig:set of squares}
\end{figure}
\subsection{Definition of the Hamiltonian
and the main result}
\label{subs:Def Ham}
Let $\cxs$ and $\cxsd$ be annihilation and creation
operators, respectively, of an electron with spin~$\sigma$
at a site~$x$ in $\La$.
These operators satisfy the anticommutation relations
\begin{equation}
\acom{c_{x,\sigma},c_{y,\tau}} = \acom{c_{x,\sigma}^\dagger,c_{y,\tau}^\dagger}=0
\end{equation}
and
\begin{equation}
\acom{c_{x,\sigma}^\dagger,c_{y,\tau}} = \delta_{x,y}\delta_{\sigma,\tau}
\end{equation}
for $x,y\in\La$ and $\sigma,\tau=\up,\dn$.
The number operator is defined by $\nxs = \cxsd\cxs$.
The total spin operators
$\vecS_\mathrm{tot}=(\Stot^{(1)},\Stot^{(2)},\Stot^{(3)})$
of the system are defined by
\begin{equation}
\Stot^{(i)} = \frac{1}{2}
\sum_{x\in\La}
\sum_{\sigma,\tau = \up,\dn}
c_{x,\sigma}^\dagger
\mathsf{p}_{\sigma\tau}^{(i)}
c_{x,\tau}
\end{equation}
for $i=1,2,3$, where $\mathsf{p}^{(i)}=[\mathsf{p}_{\sigma\tau}^{(i)}]_{\sigma,\tau=\up,\dn}$
are the Pauli matrices
\begin{equation}
\mathsf{p}^{(1)}=\left(\begin{array}{cc}
0 & 1 \\
1 & 0
\end{array}
\right)\,,\hspace*{1em}
\mathsf{p}^{(2)}=\left(\begin{array}{cc}
0 & -\rmi \\
\rmi & 0
\end{array}
\right)\,,\hspace*{1em}
\mathsf{p}^{(3)}=\left(\begin{array}{cc}
1 & 0 \\
0 & -1
\end{array}
\right)\,.
\end{equation}
We also define the raising and lowering operators on
the eigenvalues of $\Stot^{(3)}$ by
${\Stot^{+}=\Stot^{(1)}+\rmi\Stot^{(2)}}$
and ${\Stot^{-}=\Stot^{(1)}-\rmi\Stot^{(2)}}$,
respectively.
We denote by $\Stot(\Stot+1)$ the eigenvalue of $(\vecS_\mathrm{tot})^2$.
We consider the Hubbard Hamiltonian on the lattice $\La$ given by
\begin{equation}
\label{eq:Hamiltonian}
H = \Hhop + \Hint\, ,
\end{equation}
where the first term in the right-hand side,
\begin{equation}
\Hhop = \sumsigma
\sum_{x,y\in\La} t_{x,y}c_{x,\sigma}^\dagger c_{y,\sigma}
\end{equation}
with real parameters $t_{x,y}$, which we will specify below,
represents energy associated with
the quantum mechanical motion of electrons,
and the second term,
\begin{equation}
\Hint = U \sum_{x\in\La} \nxup\nxdn
\end{equation}
with $U>0$, represents a repulsive interaction between
electrons with up- and down-spin at
the same site.
The parameter $t_{x,y}$, which is called the hopping amplitude, is defined
as follows.
First we introduce new fermion operators.
For each vertex $\alpha$ in $\calV$, let
\begin{equation}
\aas = \sum_{x\in \calC_\alpha} c_{x,\sigma}\,,
\end{equation}
and for each regular square $p$ in $\calP$, let
\begin{equation}
\bps = \sum_{i=1}^4(-1)^i c_{x_i(p),\sigma}
\end{equation}
(see Fig.\ref{fig:ab states}).
\begin{figure}
(a)
\begin{center}
\includegraphics[scale=1]{ab2d.eps}
\end{center}
(b)
\begin{center}
\includegraphics[scale=1]{ab3d.eps}
\end{center}
\caption{The states corresponding to $a$- and $b$-operators.
Thin lines represent edges in $\calE$.
(a) $\nu=2$. (b) $\nu=3$. }
\label{fig:ab states}
\end{figure}
We note that the following anticommutation relations are satisfied:
\begin{equation}
\label{eq:anticom aa}
\acom{a_{\alpha,\sigma}^\dagger,a_{\beta,\sigma}}
=\left\{ \begin{array}{ll}
2\nu & \mbox{if $\alpha=\beta$};\\
1 & \mbox{if $|\alpha-\beta|=1$};\\
0 & \mbox{otherwise}\, ,
\end{array}
\right.
\end{equation}
and
\begin{equation}
\acom{b_{p,\sigma}^\dagger,b_{q,\sigma}}
=\left\{ \begin{array}{ll}
4 & \mbox{if $p=q$};\\
\mu[p,q] & \mbox{if $p$ and $q$ share a common edge};\\
0 & \mbox{otherwise}\, ,
\end{array}
\right.
\end{equation}
where $\mu[p,q]=(-1)^{i+j}$ when
the mid-point of the common edge
corresponds to $x_i(p)$ as well as $x_j(q)$.
We also note that the anticommutation relation
\begin{equation}
\label{eq:anticom ab}
\acom{a_{\alpha,\sigma}^\dagger,b_{p,\tau}} = 0
\end{equation}
holds for any $\alpha\in \calV$, any $p\in \calP$ and $\sigma,\tau=\up,\dn$.
Then we define the hopping amplitudes $t_{x,y}$ so
that the hopping term $\Hhop$ can be
written as
\begin{equation}
\Hhop = -s \sumsigma \sum_{\alpha\in \calV} \aasd\aas
+ t \sumsigma\sum_{p\in \calP} \bpsd\bps
\end{equation}
with parameters $s\ge0$ and $t>0$.
Explicitly the hopping amplitudes are given by
\begin{equation}
\label{eq:hopping matrix}
t_{x,y} = \left\{\begin{array}{ll}
-2s+2(\nu-1)\,t & \mbox{ if $x=y$}; \\
-s & \mbox{ if $|x-y|=1$ and
$m(x,y)\in \calV$};\\
t & \mbox{ if $|x-y|=1$ and
$m(x,y)\notin \calV$}; \\
-(s+t)& \mbox{ if $|x-y|=\frac{\sqrt{2}}{2}$};\\
0 & \mbox{ otherwise},
\end{array}
\right.
\end{equation}
where $m(x,y)$ is defined to
be the point taken in the middle of sites $x$ and $y$ in $\La$,
as in the case of $m(\alpha,\beta)$ for
$\alpha,\beta\in \calV$ (see Fig.~\ref{fig:lattice2d}).
\begin{figure}[t]
\begin{center}
\includegraphics[scale=1]{la2db.eps}
\end{center}
\caption{The hopping amplitudes for $\nu=2$.
Filled circles represent lattice sites.}
\label{fig:lattice2d}
\end{figure}
The main result of this paper is the occurrence of
ferromagnetism in the model defined above
at zero temperature:
\begin{theorem}
\label{th:main}
Consider the Hubbard Hamiltonian $H$ with the
hopping matrix given by~\eqref{eq:hopping matrix},
and
suppose that the number $\Ne$ of electrons is $|\calV|$.
Then, for each value of $s$
there exist $t_\mathrm{c}$ and $U_\mathrm{c}$
which are independent of the lattice size $|\calV|$
such that for $t>t_\mathrm{c}$ and $U>U_\mathrm{c}$,
the following are valid:
\begin{enumerate}
\item the ground state energy is $-2\nu s|\calV|$,
\item the ground states have the maximal total spin $\Stot = \Ne/2$,
\item the ground state is unique apart {from} $(2\Stot+1)$-fold
degeneracy
due to the spin rotation symmetry.
\end{enumerate}
\end{theorem}
In the case $s=0$,
the single-electron ground state energy is zero and
$|\calV|$-fold degenerate, as we will see in the next section.
In this case,
the model corresponds to
the flat-band Hubbard model
and the ground state exhibits saturated ferromagnetism
for all positive values of $t$ and $U$
($t_\mathrm{c}=0,U_\mathrm{c}=0$).
In particular,
in the case $\nu=2$, our model with $s=0$
is equivalent to
the flat-band Hubbard model on the line graph
of the square lattic
\footnote
One finds that our model with $s=0$ is unitary equivalent to
Mielke's flat-band model
by performing a gauge transformation $c_{x,\sigma}\to -c_{x,\sigma}$
at all sites in one of the two sublattices, say,
$\{x=m(\alpha,\alpha+\delta_1)|\alpha\in\calV\}$.
},
and the occurrence
of saturated ferromagnetism in the model was proved
by Mielke~\cite{Mielke91a,Mielke91b}.
Our model in three or more dimensions, however,
does not belong to the flat-band models on line graphs.
It is also noted that our model does not belong to
the class of flat-band models discussed in detail
by Mielke and Tasaki~\cite{MielkeTasaki93}.
We will give the proof of Theorem~\ref{th:main} for $s=0$
in Appendix~\ref{s:flat-band ferro}.
On the other hand, the lowest band of
the model is dispersive for $s>0$, and the proof for such cases
is quite different {from} that for the flat-band case.
Here we adopt the same idea which is used for proving the occurrence of ferromagnetism
in Tasaki's models~\cite{Tasaki95,Tasaki2003}.
We first decompose the Hamiltonian $H$ into local Hamiltonians, and investigate
their properties, in particular, conditions for states to attain
the local minimum energy in detail.
Then the ferromagnetic state is shown to attain the minimum energy
simultaneously for all the local Hamiltonians
for $t>t_\mathrm{c}$ and $U>U_\mathrm{c}$.
That is, the model is ``frustration free''
when the lower band supporting the ferromagnetism is almost flat
and the on-site repulsion is sufficiently large.
It is noted that, compared with Tasaki's nearly-flat band case,
we need more lengthy discussion about
local properties because of the gapless nature of our model.
We will devote ourselves to such discussion
in section~\ref{s:local properties}
and complete the proof for the case $s>0$
in section~\ref{s:proof of theorem}.
\section{Single-electron dispersion relations
and basis states of the $\Ne$-electron space}
\label{s:single electron}
Let us consider the single-electron
Schr\"odinger equation
for our Hubbard Hamiltonian.
Let $\Phi_{1,\sigma}$ be a single-electron state
\begin{equation}
\Phi_{1,\sigma}=\sum_{x\in\La} \phi_x \cxsd \Phi_0\,,
\end{equation}
where $\phi_x$ is a complex coefficient
and $\Phi_0$ is the state with no electron in $\La$.
Applying $H$ to $\Phi_{1,\sigma}$, we obtain
the eigenvalue equation
\begin{equation}
\sum_{y\in\La}t_{x,y} \phi_y = \varepsilon \phi_x\,,
\end{equation}
where $\varepsilon$ denotes a single-electron energy eigenvalue.
Because of the translation invariance of the Hamiltonian,
we can write an eigenstate $(\phi_x)_{x\in\La}$
in the form of the Bloch state as
$\phi_x = \rme^{\rmi k\cdot x}v_x(k)$,
where $v_x(k)$ satisfies $v_x(k)=v_{x+z}(k)$ for
any $z\in\Bbb{Z}^d$ and $k$ is an element in $\calK$ defined by
\begin{equation}
\calK=
\left\{k=(k_1,\dots,k_\nu)~|~
k_l = \frac{2\pi}{L}n_l,~n_l=0,\pm1,\dots, \pm\frac{L-1}{2}
\mbox{ for $l=1,\dots,\nu$}
\right\}\,.
\end{equation}
Then the eigenvalue equation is reduced to
\begin{equation}
\label{eq:eigenvalue equation}
\left(
-4s\cos^2\frac{k_l}{2}
+4t\sumtwo{m=1}{m\ne l}^\nu \cos^2\frac{k_m}{2}
\right) v_l
-4(s+t)\cos\frac{k_l}{2}
\sumtwo{m=1}{m\ne l}^\nu
\left(\cos\frac{k_m}{2}
\right) v_m = \varepsilon v_l
\end{equation}
with $l=1,\dots,\nu$, where we write $v_l$
for $v_{{\delta_l}/{2}}(k)$ for notational simplicity.
Setting $\mathsf{A}_{l,m} = \cos(k_l/2)\cos(k_m/2)$
and
\begin{equation}
\varepsilon^\prime =
-\frac{1}{s+t}\left(
\frac{\varepsilon}{4}-t\sum_{l=1}^\nu \cos^2 \frac{k_l}{2}
\right)\,,
\end{equation}
we can further rewrite \eqref{eq:eigenvalue equation} as
\begin{equation}
\sum_{m=1}^\nu \mathsf{A}_{l,m} v_m = \varepsilon^\prime v_l\,.
\end{equation}
It is easy to see that the rank of matrix
$\mathsf{A}=[\mathsf{A}_{l,m}]_{1\le l,m\le \nu}$
is $1$, and therefore $\varepsilon^\prime=0$ is the $(\nu-1)$-fold degenerate
eigenvalue of $\mathsf{A}$.
It is also easy to see that the vector $(v_m)_{m=1}^{\nu}=(\cos(k_1/2),\cos(k_2/2),\dots,\cos(k_\nu/2))$
is an eigenstate of $\mathsf{A}$ and
its eigenvalue is given by
\begin{equation}
\varepsilon^\prime=\sum_{l=1}^\nu \cos^2 \frac{k_l}{2}\,.
\end{equation}
As a result we obtain the single-electron energy eigenvalues
which are characterized by the dispersion relations
\begin{equation}
\varepsilon_0(k) = -4s \sum_{l=1}^\nu \cos^2\frac{k_l}{2}
= -2s \sum_{l=1}^\nu (1+\cos k_l)
\end{equation}
and
\begin{equation}
\varepsilon_1(k) = 4t \sum_{l=1}^\nu \cos^2\frac{k_l}{2}
= 2t \sum_{l=1}^\nu (1+\cos k_l)\,.
\end{equation}
The eigenvalue $\varepsilon_1(k)$
is $(\nu-1)$-fold degenerate for each $k$.
See Fig.~\ref{fig:dispersion} for the dispersion relations
for $\nu=2$.
\begin{figure}
\begin{center}
\includegraphics[scale=1]{disper.eps}
\end{center}
\caption{The dispersion relations for $\nu=2$.}
\label{fig:dispersion}
\end{figure}
In the case of $s=0$, $\varepsilon_0(k)=0$ for all $k\in \calK$,
i.e., the single-electron ground state energy is zero and
$|\calV|$-fold degenerate.
In the rest of this section we comment
on a construction of $\Ne$-electron states.
Let us denote by $f_{(0,k),\sigma}$
a fermion operator corresponding to the eigenstate
with the eigenvalue $\varepsilon_0(k)$.
We also denote by $f_{(l,k),\sigma}$ with $l=1,\dots,\nu-1$
fermion operators corresponding to linearly independent
$\nu-1$ eigenstates with the eigenvalue $\varepsilon_1(k)$.
Then, owing to linear independence of the energy eigenstates,
any $\Ne$-electron state can be
represented as a linear combination of the states
\begin{equation}
\left(\prod_{i\in A_\up} f_{i,\up}^\dagger\right)
\left(\prod_{i\in A_\dn} f_{i,\dn}^\dagger\right)\Phi_0
\end{equation}
with $|A_\up|+|A_\dn|=\Ne$, where $A_\up$ and $A_\dn$
are arbitrary subsets of
\begin{equation}
\{i=(l,k)~|~l=0,\dots,\nu-1,k\in \calK\}\,.
\end{equation}
It is crucial to find that the fermion operators $f_{(0,k),\sigma}$
can be expanded as a linear combination of $\aas$.
This is observed as follows.
First we note that
$\sum_{\alpha\in \calV} u_\alpha \aasd\Phi_0$
is zero if and only if $u_\alpha = 0$
for all $\alpha\in \calV$, i.e.,
the states $\aasd\Phi_0$ are linearly independent.
Next consider the single-electron
Schr\"odinger equation
\begin{equation}
H \sum_{\beta\in \calV}\phi_\beta a_{\beta,\sigma}^\dagger \Phi_0
= \varepsilon \sum_{\beta\in \calV}
\phi_\beta a_{\beta,\sigma}^\dagger\Phi_0\,,
\end{equation}
where $\phi_\beta$ are complex coefficients
and $\varepsilon$ is real.
By using the anticommutation relations \eqref{eq:anticom aa} and
\eqref{eq:anticom ab}, we reduce the left-hand side of the above
equation to a linear combination of $a_{\beta,\sigma}^\dagger\Phi_0$.
Then, comparing the coefficients of $a_{\beta,\sigma}^\dagger\Phi_0$
in the left- and the right-hand sides, we obtain
\begin{equation}
-2\nu s \phi_{\beta}
-s\sumtwo{\alpha\in \calV}{|\alpha-\beta|=1}
\phi_\alpha
= \varepsilon \phi_\beta \,.
\end{equation}
By solving the above equation we obtain
the eigenvalue $\varepsilon_0(k)$ with $k\in \calK$.
Therefore the sets
$\{f_{(0,k),\sigma}^\dagger\Phi_0\}_{k\in \calK}$
and $\{a_{\alpha,\sigma}^\dagger\Phi_0\}_{\alpha\in \calV}$
span the same space.
This also indicates the fact that
any $\Ne$-electron state can be
represented as a linear combination of the states
\begin{equation}
\label{eq:basis}
\Phi(V_\up,B_\up;V_\dn,B_\dn)
=
\left(\prod_{\alpha\in V_\up} a_{\alpha,\up}^\dagger\right)
\left(\prod_{i\in B_\up} f_{i,\up}^\dagger\right)
\left(\prod_{\alpha\in V_\dn} a_{\alpha,\dn}^\dagger\right)
\left(\prod_{i\in B_\dn} f_{i,\dn}^\dagger\right)\Phi_0
\end{equation}
with $|V_\up|+|V_\dn|+|B_\up|+|B_\dn|=\Ne$,
where $V_\up$ and $V_\dn$ are arbitrary subsets of $\calV$,
and
$B_\up$ and $B_\dn$
are arbitrary subsets of
\begin{equation}
\{i=(l,k)~|~l=1,\dots,\nu-1,k\in \calK\}\,.
\end{equation}
\section{Local properties of the model}
\label{s:local properties}
In this section we represent
the Hamiltonian \eqref{eq:Hamiltonian}
as a sum of local Hamiltonians
and investigate properties of the local Hamiltonians.
As a consequence we will obtain a lemma which will
play a central role in the proof of
Theorem~\ref{th:main} for $s>0$.
\subsection{Decomposition of the Hamiltonian}
For each vertex $\alpha\in \calV$,
we define the sublattice $\Laa$ as
\begin{equation}
\Laa = \{ x=x_i(p)~|~i=1,\dots,4,~p\in \calP_\alpha\}\,,
\end{equation}
which is the collection of the sites on the edges of the
regular squares which share vertex $\alpha$.
(In the case $\nu=2$, $\Laa$ consists of 12 filled circles
depicted in Fig.\ref{fig:lattice2d}.)
Then we define the local Hamiltonian $\halph$ on the lattice $\Laa$ by
\begin{equation}
\label{eq:local Hamiltonian}
\halph =
-s\sumsigma\aasd\aas
+ \frac{t}{4}\sumsigma\sum_{p\in \calP_\alpha}
b_{p,\sigma}^\dagger b_{p,\sigma}
+\frac{U}{2(2\nu-1)}\sum_{x\in\Laa} \nxup\nxdn\, .
\end{equation}
By using $\halph$, the Hamiltonian $H$ is decomposed as
\begin{equation}
H = \sum_{\alpha\in \calV} \halph\, .
\end{equation}
It is noted that the local Hamiltonians do not commute with one another,
and it is thus not trivial to find out something about
the whole system, such as the energy and magnetic
properties of the ground state, by using these local Hamiltonians.
The properties of the local Hamiltonian $\halph$
for sufficiently large values of $U/s$ and $t/s$ are
summarized in the following lemma.
\begin{lemma}
\label{lemma:main}
Consider the local Hamiltonian $h_\alpha$
on the Hilbert space where
the electron number on the local lattice $\La_\alpha$
is not fixed.
Suppose that $t/s$ and $U/s$ are sufficiently large.
Then the minimum eigenvalue of $\halph$ is $-2\nu s$ and
any eigenstate $\Phi$ with this eigenvalue can be expressed as
\begin{equation}
\label{eq:condition1 in Lemma}
\Phi = a_{\alpha,\up}^\dagger\Phi_\up + a_{\alpha,\dn}^\dagger\Phi_\dn,
\end{equation}
by using some states $\Phi_\up$ with $a_{\alpha,\dn}\Phi_\up=0$
and $\Phi_\dn$ with $a_{\alpha,\up}\Phi_\dn=0$.
Furthermore, $\Phi$ satisfies
\begin{equation}
\label{eq:condition2 in Lemma}
c_{x,\dn}c_{x,\up}\Phi = 0
\end{equation}
for all $x\in\La_\alpha$.
\end{lemma}
The above lemma claims that electrons do not
doubly occupy the localized states corresponding
to $\aas$
when the strength of the repulsive interaction is
sufficiently large and
the high-energy part of local single-electron energy eigenvalues
is well separated from the lower levels.
It is remarked that a similar lemma is proved
for Tasaki's models and
the proof of our lemma proceeds in almost the same way as
in Refs.~\cite{Tasaki95,Tasaki2003}.
However, the proof for our case is much more involved.
The difference is due to the following reason.
The localized states corresponding to $a$-operators in Tasaki's models
contain a parameter which controls a band gap or
hopping amplitudes.
By taking a certain limit of this parameter,
the local Hamiltonians of Tasaki's models are
reduced to ``atomic Hamiltonians'', which have no hoppings of electrons.
This fact makes it somewhat easier to treat the local Hamiltonians.
On the other hand, our model does not contain such a parameter, and
we have to prove the lemma by directly using eigenstates of~$h_\alpha$.
It is also remarked that the local Hamiltonian $h_\alpha$ is supported on
the sublattice consisting of $2\nu(2\nu-1)$ sites,
which becomes a 12-site lattice even in two dimensions, and
the decomposition with smaller sublattices probably does not work.
It is a non-trivial task to analytically solve a problem of interacting electrons
on this lattice siz
\footnote
In Ref.~\cite{Lu2009}, a similar lemma is proved relying on numerical calculations.
}.
\subsection{Single-electron problem for the local Hamiltonian}
Because of the translation invariance, it suffices to
investigate $h_0$, the local Hamiltonian associated with the origin $0=(0,\dots,0)$
of $\calV$, in order to prove Lemma~\ref{lemma:main}.
Since $h_0$ is defined only on $\Lambda_0$,
we consider the space constructed by using
fermion operators
$c_{x,\sigma}$ with $x\in\La_0$, $\sigma=\up,\dn$.
In this subsection
we will solve the single-electron problem for $h_0$.
As usual we can solve it
by using basis states $\{\cxsd\Phi_0\}_{x\in\La_0}$
of the single-electron Hilbert space $\calH$ on $\La_0$,
but we adopt a slightly different method here.
To prove Lemma~\ref{lemma:main},
we will consider finite energy states
in the limit $t\to\infty$.
So let us firstly characterize single-electron states
which have infinitely large energy in the limit $t\to\infty$.
Let $\calH_b$ be the single-electron Hilbert space spanned by the states
$\bpsd\Phi_0$ with $p\in \calP_0$, which are linearly independent.
Since all the creation operators $\bpsd$
anticommute with $a_{0,\sigma}$,
we find that
$h_0 \Phi_{1,\sigma}\in \calH_b$
for any $\Phi_{1,\sigma}\in \calH_b$.
This implies that it is possible to obtain eigenstates of $h_0$
within $\calH_b$.
Since the operator $t\sum_{p\in \calP_0}\bpsd\bps$
restricted on $\calH_b$
is positive definite,
corresponding energy eigenvalues are proportional to $t$,
and thus $\calH_b$ consists of high energy states.
Let $\calH_b^{\perp}$ be the orthogonal complement
of $\calH_b$ within $\calH$.
In the following, we solve the single-electron problem for $h_0$
within $\calH_b^{\perp}$, by constructing states orthogonal to
those in $\calH_b$
(or equivalently
by constructing local fermion operators anticommuting
with $\bpsd$).
Let $\dLao$ be the collection of the ``boundary'' sites
in $\Lambda_0$
\begin{equation}
\dLao = \Lambda_0 \backslash \calC_0\,.
\label{eq:boundary}
\end{equation}
Without loss of generality, we can assume
that ${x}_3(p),{x}_4(p)\in\dLao$ for all $p\in \calP_0$.
Then, we define
\begin{equation}
\label{eq:dps}
\dps = \frac{1}{\sqrt{2}}(c_{{x}_3(p)} + c_{{x}_4(p)})
\end{equation}
for each $p\in \calP_0$ and
\begin{equation}
\label{eq:bdxs}
\bar{d}_{x,\sigma} = 2 \cxs + \sum_{p\in \calP(x)}
\mu[x,p](c_{x_3(p),\sigma} - c_{x_4(p),\sigma})\, ,
\end{equation}
for each $x\in \calC_0$ ,
where $\mu[x,p]$ takes $-1$ if $x$ corresponds to $x_1(p)$
and 1 if $x$ corresponds to $x_2(p)$.
By a straightforward calculation, we can check that
\begin{equation}
\acom{b_{p,\sigma}^\dagger, d_{q,\tau}} =
\acom{b_{p,\sigma}^\dagger, \bar{d}_{x,\tau}} = 0
\end{equation}
for any $p,q\in \calP_0$ and $x\in \calC_0$.
It is also not difficult to check that
$|\calP_0|+|\calC_0|$ states, ${d}_{p,\sigma}^\dagger\Phi_0$ with $p\in \calP_0$
and $\bar{d}_{x,\sigma}^\dagger\Phi_0$ with $x\in \calC_0$, are
linearly independent.
Noting that the dimension of $\calH_b$ is $|\calP_0|$,
and $|\calP_0|+(|\calP_0|+|\calC_0|) = |\La_0|$,
we find that
these states
span the orthogonal complement $\calH_b^\perp$ of $\calH_b$.
Since the operator~$s a_{0,\sigma}^\dagger a_{0,\sigma}$
restricted to $\calH$
is a projection onto the space spanned by
the single state $a_{0,\sigma}^\dagger\Phi_0$,
$h_0$ restricted to $\calH_b^\perp$
has exactly two eigenvalue; one is 0 and
the other is $-2\nu s$ to which $a_{0,\sigma}^\dagger\Phi_0$
belongs.
One easily finds that all $\dpsd$ with $p\in\calP_0$,
which are supported only on the sites in $\partial\La_0$,
anticommute
with not only $\bps$ but also $a_{0,\sigma}$,
so that the states $\dpsd\Phi$
are the eigenstates of $h_0$
with eigenvalue 0.
Furthermore, it follows {from} $\acom{\dpsd,d_{q,\sigma}}=\delta_{p,q}$
that the set of states $\left\{\dpsd\Phi_0\right\}_{p\in\calP_0}$
is orthonormal.
Let us determine the rest of eigenstates with eigenvalue 0.
Since the anticommutation relation
\begin{equation}
\acom{d_{p,\sigma}^\dagger, \bar{d}_{x,\sigma}} = 0
\end{equation}
holds for $p\in \calP_0$ and $x\in \calC_0$,
our task is to construct fermion operators
which anticommute with $a_{0,\sigma}^\dagger$,
by using $\bar{d}_{x,\sigma}$
with $x\in \calC_0$.
In the following, we write $x_l$ for $m(0,\delta_l)$
and $x_{-l}$ for $m(0,-\delta_l)$ with $l=1,\dots \nu$.
We note that
\begin{equation}
\calC_0 = \{ x_l~|~l=\pm1,\dots,\pm \nu\}\,.
\end{equation}
By using $\bar{d}_{x,\sigma}$, we form fermion operators which anticommute
with one another.
We write again $d_{i,\sigma}$ for these fermion operators.
To label new $d$-operators which we will define below,
we introduce a set $\calD=\left\{1,\dots,\nu\right\}$, and also
\begin{equation}
\calJ=\left\{0,\pm\frac{2\pi}{\nu},\dots,\pm\frac{2\pi}{\nu}\frac{(\nu-1)}{2} \right\}
\end{equation}
for odd $\nu$,
and
\begin{equation}
\calJ=\left\{0,\pm\frac{2\pi}{\nu},\dots,\pm\frac{2\pi}{\nu}\frac{\nu-2}{2},\pi \right\}
\end{equation}
for even $\nu$.
Now, for $l\in \calD$, let us
define
\begin{equation}
\label{eq:dls}
d_{l,\sigma} = \frac{1}{2\sqrt{2\nu}}
(\bar{d}_{x_l,\sigma}-\bar{d}_{x_{-l},\sigma})\,.
\end{equation}
We also define
\begin{equation}
\label{eq:dks}
d_{k,\sigma} = \frac{1}{2\sqrt{2\nu(\nu+1-\nu\delta_{k,0})}}
\sum_{l=1}^\nu
\rme^{\rmi kl} (\bar{d}_{x_l,\sigma}
+\bar{d}_{x_{-l},\sigma})
\end{equation}
for $k\in\calJ$.
We note that $d_{k,\sigma}$ with $k=0$ is
equal to $(1/\sqrt{2\nu})a_{0,\sigma}$.
Furthermore, as we will see below, we have the anticommutation relations
\begin{equation}
\acom{d_{i,\sigma}^\dagger, d_{j,\sigma}} = \delta_{i,j}
\label{eq:anticom}
\end{equation}
for any $i,j\in \calD\cup \calJ$.
Therefore, the single-electron states
$d_{i,\sigma}^\dagger\Phi_0$ with $i\in \calD\cup\calJ\backslash\{0\}$
in $\calH_b^\perp$
are orthonormal eigenstates of $h_0$ with eigenvalue 0.
To prove~\eqref{eq:anticom}, let us first calculate the anticommutation
relations for $\bar{d}_{x,\sigma}^\dagger$
and $\bar{d}_{y,\sigma}$ with $x,y\in \calC_0$.
Noting that $|\calP(x_l)|=2(\nu-1)$ and
$|\calP(x_l)\cap \calP(x_{-l})|=\emptyset$
for $l=\pm1,\dots,\pm \nu$,
we find
\begin{equation}
\label{eq:anticom td1}
\acom{\bar{d}_{x_l,\sigma}^\dagger,
\bar{d}_{x_l,\sigma}}=4+2|\calP(x_l)|=4\nu\,,
\end{equation}
and
\begin{equation}
\label{eq:anticom td2}
\acom{\bar{d}_{x_l,\sigma}^\dagger, \bar{d}_{x_{-l},\sigma}} = 0\, .
\end{equation}
Next, noting that, for
$l,l^\prime=\pm1,\dots,\pm \nu$ and $l\ne \pm l^\prime$,
$\calP(x_l)\cap \calP(x_{l^\prime})$ always has
only one element, say $p$, such that
either $x_l=x_1(p),x_{l^\prime}=x_2(p)$ or
$x_l=x_2(p),x_{l^\prime}=x_1(p)$
hold
\footnote{
The regular square $p$ is in the plane including the $l$-axis and
the $l^\prime$-axis.
},
we have
\begin{equation}
\label{eq:anticom td3}
\acom{\bar{d}_{x_l,\sigma}^\dagger,
\bar{d}_{x_{l^\prime},\sigma}} = 2\mu[x_l,p]\mu[x_{l^\prime},p]
=-2.
\end{equation}
Then, by using anticommutation
relations~\eqref{eq:anticom td1},
\eqref{eq:anticom td2} and
\eqref{eq:anticom td3},
we obtain
\begin{equation}
\label{eq:anticom td4}
\{(\bar{d}_{x_l,\sigma}^\dagger-\bar{d}_{x_{-l},\sigma}^\dagger),
(\bar{d}_{x_{l^\prime},\sigma}-\bar{d}_{x_{-l^\prime},\sigma})\}
={8\nu}\delta_{l,l^\prime},
\end{equation}
\begin{equation}
\label{eq:anticom td5}
\{(\bar{d}_{x_l,\sigma}^\dagger+\bar{d}_{x_{-l},\sigma}^\dagger),
(\bar{d}_{x_{l^\prime},\sigma}+\bar{d}_{x_{-l^\prime},\sigma})\}
={8\nu}\delta_{l,l^\prime}-8(1-\delta_{l,l^\prime}),
\end{equation}
and
\begin{equation}
\label{eq:anticom td6}
\{(\bar{d}_{x_l,\sigma}^\dagger+\bar{d}_{x_{-l},\sigma}^\dagger),
(\bar{d}_{x_{l^\prime},\sigma}-\bar{d}_{x_{-l^\prime},\sigma})\}
=0
\end{equation}
for $l,l^\prime\in\calD$.
From, \eqref{eq:anticom td4},
\eqref{eq:anticom td5} and
\eqref{eq:anticom td6},
we obtain the desired relations~\eqref{eq:anticom}.
The results in this subsection are summarized
as follows.
Let $\calI=\calP_0\cup \calD \cup \calJ$.
The fermion operators defined by \eqref{eq:dps}, \eqref{eq:dls} and \eqref{eq:dks}
satisfy
the anticommutation relation
\begin{equation}
\acom{d_{i,\sigma}^\dagger, d_{j,\sigma}} = \delta_{i,j}
\end{equation}
for any $i,j\in \calI$.
The single-electron states $d_{i,\sigma}^\dagger\Phi_0$
with $i\in \calI$
satisfy
\begin{equation}
h_0 d_{i,\sigma}^\dagger\Phi_0
= \varepsilon_i d_{i,\sigma}^\dagger\Phi_0\,,
\end{equation}
where $\varepsilon_i$ are given by
\begin{equation}
\varepsilon_i = \left\{\begin{array}{ll}
-2\nu s & \mbox{if $i=0$};\\
0 & \mbox{otherwise},
\end{array}
\right.
\end{equation}
and they form an orthonormal basi
\footnote
Note that $|\calI|=|\calP_0|+|\calD|+|\calJ|=|\calP_0|+|\calC_0|$ which equals the dimension of $\calH_b^\perp$.
}
for subspace $\calH_b^\perp$.
\subsection{Proof of Lemma~\ref{lemma:main}}
By using the local fermion operators introduced in the previous section,
we solve a many-electron problem for $h_0$.
First, we will consider the problem in the limit $t,U\to\infty$,
where it will be proved that the minimum expectation value of $h_0$
is equal to or greater than $-2\nu s$
and that any state which attains the minimum expectation value
is an eigenstate of $h_0$.
Let $\Phi$ be a state on $\La_0$ with a finite energy
in the limit $t,U\to\infty$.
Consider representing $\Phi$ by
using the fermion operators
$\{d_{i,\sigma}^\dagger\}_{i\in \calI}$
and
$\{b_{p,\sigma}^\dagger\}_{p\in \calP_0}$.
In the limit $t\to\infty$, since the terms
$t b_{p,\sigma}^\dagger b_{p,\sigma}$ in $h_0$ are positive semidefinite,
$\Phi$ with a finite energy must satisfy
$b_{p,\sigma}\Phi=0$ for all $p\in \calP_0$ and
$\sigma=\up,\dn$.
This means that a finite energy state $\Phi$ is written as
\begin{equation}
\label{eq:finite energy state}
\Phi = \sum_{I_\up,I_\dn\subset \calI}
g(I_\up;I_\dn) \Phi(I_\up,I_\dn)
\end{equation}
with complex coefficients $g(I_\up;I_\dn)$,
where
\begin{equation}
\label{eq:d basis}
\Phi(I_\up;I_\dn) = \left(\prod_{i\in I_\up}d_{i,\up}^\dagger\right)
\left(\prod_{j\in I_\dn}d_{j,\dn}^\dagger\right)\Phi_0 \,.
\end{equation}
Here, to avoid an ambiguity which may arise due to
the exchange of fermion operators,
we adopt the following rule for the product
of fermion operators.
Let $\theta$ be a one-to-one mapping {from} $\calI$ to $\Bbb{Z}$.
Then, we assume that the products of
the fermion operators in $\Phi(I_\up;I_\dn)$ are ordered
in such a way that
$d_{i,\up}^\dagger$ (respectively, $d_{i,\dn}^\dagger$)
is always on the left of
$d_{j,\up}^\dagger$ (respectively, $d_{j,\dn}^\dagger$)
if $\theta(i)<\theta(j)$.
For later use, we also define
\begin{equation}
\label{eq:sign factor S}
\mathsf{S}_{I_\sigma}^{i}
=\prod_{j\in I_{\sigma};\theta(j)<\theta(i)}(-1).
\end{equation}
For example, we have
\begin{equation}
\left(\prod_{j\in I_{\sigma}} d_{j,\sigma}^\dagger\right)
=
\mathsf{S}_{I_\sigma}^{i}
d_{i,\sigma}^\dagger
\left(\prod_{j\in I_{\sigma}\backslash \{i\}} d_{j,\sigma}^\dagger\right).
\end{equation}
Let us consider furthermore the limit $U\to\infty$.
Since the on-site interaction
$U\nxup\nxdn=Uc_{x,\up}^\dagger c_{x,\dn}^\dagger c_{x,\dn}c_{x,\up}$
is positive semidefinite,
a state $\Phi$ with finite energy satisfies
$c_{x,\dn}c_{x,\up}\Phi=0$ for any $x\in\La_0$.
Substituting \eqref{eq:finite energy state}
into this equation, we find that the coefficients
$g(I_\up;I_\dn)$ must be chosen so that
\begin{equation}
\sum_{I_\up,I_\dn\subset \calI}
g(I_\up;I_\dn)
c_{x,\dn}c_{x,\up}\Phi(I_\up;I_\dn)= 0\,
\label{eq:finite energy condition on g}
\end{equation}
will always hold for any $x\in\La_0$.
By using the coefficients $g(I_\up;I_\dn)$
which satisfy the condition~\eqref{eq:finite energy condition on g},
let us now express
an expectation value $E[\Phi]$ of $h_0$ for $\Phi$
in the limit of $t,U\to\infty$.
Since the eigenvalues of
$-s \sum_{\sigma} a_{0,\sigma}^\dagger a_{0,\sigma}$
for $\Phi(I_\up;I_\dn)$
are simply given by
\begin{equation}
\left\{ \begin{array}{ll}
-4\nu s & \mbox{if $0\in I_\up\cap I_\dn$}; \\
-2\nu s & \mbox{if $0\in I_\up\cup I_\dn$ and $0\notin I_\up\cap I_\dn$}; \\
0 & \mbox{otherwise},
\end{array}
\right.
\end{equation}
and the set of all states $\Phi(I_\up;I_\dn)$ is orthonormal,
we have
\begin{eqnarray}
E[\Phi] & = & \frac{\langle \Phi, h_0 \Phi \rangle}
{\langle \Phi, \Phi \rangle} \\
& = & \left(-4\nu s \sumtwo{I_\up,I_\dn\subset \calI}
{0\in I_\up\cap I_\dn} |g(I_\up;I_\dn)|^2
-2\nu s \sumtwo {I_\up,I_\dn\subset \calI}
{0\in I_\up\cup I_\dn, 0\notin I_\up\cap I_\dn}
|g(I_\up;I_\dn)|^2
\right)||\Phi||^{-2}
\end{eqnarray}
with
\begin{equation}
||\Phi||=\sqrt{\langle \Phi, \Phi \rangle}=\left(\sumIud|g(I_\up;I_\dn)|^2 \right)^\frac{1}{2}.
\end{equation}
By noting that
\begin{equation}
\sumtwo{I_\up,I_\dn\subset \calI}{0\in I_\up\cap I_\dn}
+\sumtwo{I_\up,I_\dn\subset \calI}{0\in I_\up\cup I_\dn, 0\notin I_\up\cap I_\dn}
+\sumtwo{I_\up,I_\dn\subset \calI}{0\notin I_\up\cup I_\dn}
=
\sum_{I_\up,I_\dn\subset \calI},
\end{equation}
we rewrite $E[\Phi]$ as
\begin{equation}
E[\Phi]=-2\nu s + 2\nu s F[\Phi]||\Phi||^{-2}
\end{equation}
where
\begin{eqnarray}
F[\Phi]
&=& \sumtwo{I_\up,I_\dn\subset \calI}{0\notin I_\up\cup I_\dn}|g(I_\up;I_\dn)|^2
-\sumtwo{I_\up,I_\dn\subset \calI}{0\in I_\up\cap I_\dn}|g(I_\up;I_\dn)|^2
\nonumber\\
&=& \sumIudo |g(I_\up;I_\dn)|^2
-\sumIudo |\g{0}{0}|^2 \,.
\end{eqnarray}
In the following, we show $F[\Phi]\ge0$. This implies $E[\Phi]\ge -2\nu s$ since $s>0$.
In order to prove $F[\Phi]\ge 0$ we investigate in detail the conditions on
$g(I_\up;I_\dn)$ imposed by~\eqref{eq:finite energy condition on g}.
The left-hand side of~\eqref{eq:finite energy condition on g} is calculated as
\begin{equation}
\label{eq:ccPhi}
\sumtwo{I_\up,I_\dn\subset \calI}{|I_\up|\ge1,|I_\dn|\ge1}
g(I_\up;I_\dn)
\sum_{n\in I_\up}\sum_{m\in I_\dn}
(-1)^{|I_\up|-1}{\mathsf{S}_{I_\up}^n}{\mathsf{S}_{I_\dn}^m}
(\varphi_x^{(n)})^\ast(\varphi_x^{(m)})^\ast
\Phi(I_\up\backslash\{n\};I_\dn\backslash\{m\})\, ,
\end{equation}
where
$(\varphi_x^{(i)})^\ast=\{d_{i,\sigma}^\dagger,c_{x,\sigma}\}$.
The expression \eqref{eq:ccPhi} is further calculated as
\begin{eqnarray}
\label{eq:ccPhi2}
\lefteqn{\sum_{n,m\in \calI}
(\varphi_x^{(n)})^\ast(\varphi_x^{(m)})^\ast
\sumtwo{I_\up,I_\dn\subset \calI}{|I_\up|\ge1,|I_\dn|\ge1}
(-1)^{|I_\up|-1}
{\mathsf{S}_{I_\up}^n}{\mathsf{S}_{I_\dn}^m}
\chi[n\in I_\up]\chi[m\in I_\dn]}\nonumber\\
&& \hspace*{7cm}\times
g(I_\up;I_\dn)
\Phi(I_\up\backslash\{n\};I_\dn\backslash\{m\})\nonumber\\ \nonumber\\
&=&
\sum_{n,m\in \calI}
(\varphi_x^{(n)})^\ast(\varphi_x^{(m)})^\ast
\sum_{I_\up\subset {\calI}\backslash\{n\}}
\sum_{I_\dn\subset {\calI}\backslash\{m\}}
(-1)^{|I_\up|}
{\mathsf{S}_{I_\up}^n}{\mathsf{S}_{I_\dn}^m}
g(\{n\}\cup I_\up;\{m\}\cup I_\dn)
\Phi(I_\up;I_\dn) \nonumber\\
&=&
\sumIud
\sum_{n,m\in \calI}
(\varphi_x^{(n)})^\ast(\varphi_x^{(m)})^\ast
\tg{n}{m}\Phi(I_\up;I_\dn)\, ,
\end{eqnarray}
where $\chi[\textrm{event}]$ is the indicator function
which takes 1 if event is true and 0 otherwise, and
we have introduced the subsidiary coefficients defined as
\begin{equation}
\label{eq:tg1}
\tg{n}{m} = \left\{
\begin{array}{ll}
0 & \mbox{if $n\in I_\up$ or $m\in I_\dn$;}\\
(-1)^{|I_\up|}
{\mathsf{S}_{I_\up}^n}{\mathsf{S}_{I_\dn}^m}
\g{n}{m}
&\mbox{otherwise.}
\end{array}
\right.
\end{equation}
In the first equality of \eqref{eq:ccPhi2} we have used
${\mathsf{S}_{I_\sigma\cup\{i\}}^i}={\mathsf{S}_{I_\sigma}^i}$.
Since all the states $\Phi(I_\up;I_\dn)$ are linearly independent,
we find that the condition~\eqref{eq:finite energy condition on g}
becomes
\begin{equation}
\label{eq:finite energy condition}
\sum_{n,m\in \calI}
(\varphi_x^{(n)})^\ast(\varphi_x^{(m)})^\ast
\tg{n}{m} = 0
\end{equation}
for any $I_\up,I_\dn\subset {\calI}$ and for any $x\in\La_0$.
By choosing the sites in $\calC_0$ as $x$,
we rewrite condition~\eqref{eq:finite energy condition}
more concretely.
Take $x_l=m(0,\delta_l)\in \calC_0$ as $x$ in condition~\eqref{eq:finite energy condition}.
Noting that $\varphi_{x_l}^{(i)}$ is vanishing if
$i\in \calP_0$ or $i\in \calD\backslash\{l\}
\footnote
We calculate the coefficients $\varphi_{x_l}^{(i)}=\{d_{i,\sigma}^\dagger, c_{x_l}\}$
by using
\eqref{eq:dps}, \eqref{eq:bdxs}, \eqref{eq:dls} and \eqref{eq:dks}.
},
we obtain
\if0
\begin{eqnarray}
\lefteqn{\frac{1}{2\nu}\tilde{g}(l;l)
+ \frac{1}{2\nu}\tilde{g}(0;0)
+ \frac{1}{2\nu(\nu+1)}\sum_{k,k^\prime\in\calJ\backslash\{0\}}
\rme^{-\rmi kl}\rme^{-\rmi k^\prime l}
\tilde{g}(k;k^\prime)} \nonumber \\
&&
+ \frac{1}{2\nu}(\tilde{g}(l;0)+\tilde{g}(0;l))
+ \frac{1}{2\nu\sqrt{\nu+1}}
\sumJo
\rme^{-\rmi kl}(\tilde{g}(l;k)+\tilde{g}(k;l)) \nonumber\\
&&
+ \frac{1}{2\nu\sqrt{\nu+1}}
\sumJo
\rme^{-\rmi kl}(\tilde{g}(0;k)+\tilde{g}(k;0)) = 0\, ,
\end{eqnarray}
\fi
\begin{eqnarray}
\lefteqn{\tilde{g}(l;l)
+ \tilde{g}(0;0)
+ \frac{1}{\nu+1}\sum_{k,k^\prime\in\calJ\backslash\{0\}}
\rme^{-\rmi kl}\rme^{-\rmi k^\prime l}
\tilde{g}(k;k^\prime)} \nonumber \\
&&
+ (\tilde{g}(l;0)+\tilde{g}(0;l))
+ \frac{1}{\sqrt{\nu+1}}
\sumJo
\rme^{-\rmi kl}(\tilde{g}(l;k)+\tilde{g}(k;l)) \nonumber\\
&&
+ \frac{1}{\sqrt{\nu+1}}
\sumJo
\rme^{-\rmi kl}(\tilde{g}(0;k)+\tilde{g}(k;0)) = 0\, ,
\end{eqnarray}
where we write $\tilde{g}(n;m)$ for
$\tg{n}{m}$.
Similarly, for $x_{-l}=m(0,-\delta_l)\in \calC_0$, we obtain
\if0
\begin{eqnarray}
\lefteqn{\frac{1}{2\nu}\tilde{g}(l;l)
+ \frac{1}{2\nu}\tilde{g}(0;0)
+ \frac{1}{2\nu(\nu+1)}\sum_{k,k^\prime\in\calJ\backslash\{0\}}
\rme^{-\rmi kl}\rme^{-\rmi k^\prime l}
\tilde{g}(k;k^\prime)} \nonumber \\
&&
- \frac{1}{2\nu}(\tilde{g}(l;0)+\tilde{g}(0;l))
- \frac{1}{2\nu\sqrt{\nu+1}}
\sumJo
\rme^{-\rmi kl}(\tilde{g}(l;k)+\tilde{g}(k;l)) \nonumber\\
&&
+ \frac{1}{2\nu\sqrt{\nu+1}}
\sumJo
\rme^{-\rmi kl}(\tilde{g}(0;k)+\tilde{g}(k;0)) = 0\, .
\end{eqnarray}
\fi
\begin{eqnarray}
\lefteqn{\tilde{g}(l;l)
+ \tilde{g}(0;0)
+ \frac{1}{\nu+1}\sum_{k,k^\prime\in\calJ\backslash\{0\}}
\rme^{-\rmi kl}\rme^{-\rmi k^\prime l}
\tilde{g}(k;k^\prime)} \nonumber \\
&&
- (\tilde{g}(l;0)+\tilde{g}(0;l))
- \frac{1}{\sqrt{\nu+1}}
\sumJo
\rme^{-\rmi kl}(\tilde{g}(l;k)+\tilde{g}(k;l)) \nonumber\\
&&
+ \frac{1}{\sqrt{\nu+1}}
\sumJo
\rme^{-\rmi kl}(\tilde{g}(0;k)+\tilde{g}(k;0)) = 0\, .
\end{eqnarray}
Then, summing the above two equations over $l\in \calD$
we obtain the conditio
\footnote
We think of $-\pi$ as $\pi$ when $\nu$ is even ($\pi$ is in $\calJ$).}
\begin{equation}
\label{eq:final condition}
\tg{0}{0} =
- \frac{1}{\nu}\sum_{l\in \calD}\tg{l}{l}
- \frac{1}{(\nu+1)}\sumJo \tg{k}{-k}\,.
\end{equation}
Our analysis below highly relies on the above relation.
Let $\bar{I}_\dn$ be the set defined by
\begin{equation}
\bar{I}_\dn = \{ i ~|~i\in I_\dn\cap (\calP_0\cup D)\}
\cup
\{ -i ~|~i\in I_\dn\cap \calJ\}\,,
\end{equation}
and let $N(I_\up;I_\dn)$ be the number of elements in
$I_\up\cap\bar{I}_\dn \cap (\calD\cup\calJ\backslash\{0\})$.
Condition~\eqref{eq:final condition} relates
$\g{0}{0}$ with
$N(I_\up;I_\dn)=r$ and
${g}(I_\up^\prime;I_\dn^\prime)$
with $N(I_\up^\prime;I_\dn^\prime)=r+1$.
This motivates us to decompose $F[\Phi]$ as
\begin{equation}
\label{eq:decomposition of F}
F[\Phi] = F^\prime[\Phi]+\sum_{r=0}^{2\nu-1} F_r[\Phi]\,,
\end{equation}
where
\begin{equation}
\label{eq:Fr}
F_r[\Phi]=\sumtwo{I_\up,I_\dn\subset \calI\backslash\{0\}}
{N(I_\up;I_\dn)=r+1}
|g(I_\up;I_\dn)|^2
-\sumtwo{I_\up,I_\dn\subset \calI\backslash\{0\}}
{N(I_\up;I_\dn)=r}
|\g{0}{0}|^2
\,,
\end{equation}
and
\begin{equation}
\label{eq:Fprime}
F^\prime[\Phi] = \sumtwo{I_\up,I_\dn\subset \calI\backslash\{0\}}
{N(I_\up;I_\dn)=0}|g(I_\up;I_\dn)|^2\,.
\end{equation}
The term $F^\prime[\Phi]$ is
apparently non-negative,
and therefore $F[\Phi]\ge0$ is implied by $F_r[\Phi]\ge0$
for $r=0,\dots,2\nu-1$.
We will prove $F_r[\Phi]\ge0$ by using~\eqref{eq:final condition}.
First we note that, for a pair of $I_\up$ and $I_\dn$ such that
$N(I_\up;I_\dn)=r$, the number of non-zero $\tilde{g}$
in the right-hand side of~\eqref{eq:final condition} is at most
$2\nu-r-1$ (see the definition \eqref{eq:tg1} of $\tilde{g}$).
For such a pair of $I_\up$ and $I_\dn$, using the Schwarz inequality, we have
\begin{eqnarray}
\label{eq:inequality for tilde g 1}
|\tg{0}{0}|^2
& \le & \frac{2\nu-r-1}{\nu^2}\sum_{l\in \calD}
|\tg{l}{l}|^2 \nonumber\\
&& +\frac{2\nu-r-1}{(\nu+1)^2}
\sumJo
|\tg{k}{-k}|^2.
\end{eqnarray}
Then,
noting that
$0 \le \frac{2\nu-r-1}{(\nu+1)^2}\le\frac{2\nu-r-1}{\nu^2}$,
we find that
\begin{eqnarray}
\label{eq:Fr lower bound}
\sumtwo{I_\up,I_\dn\subset \calI\backslash\{0\}}
{N(I_\up;I_\dn)=r}
|\g{0}{0}|^2
& = &
\sumtwo{I_\up,I_\dn\subset \calI\backslash\{0\}}
{N(I_\up;I_\dn)=r}
|\tg{0}{0}|^2 \nonumber\\
& \le & \frac{2\nu-r-1}{\nu^2}
\sumtwo{I_\up,I_\dn\subset \calI\backslash\{0\}}
{N(I_\up;I_\dn)=r}
\sum_{l\in \calD}
|\tg{l}{l}|^2 \nonumber\\
&& +\frac{2\nu-r-1}{(\nu+1)^2}
\sumtwo{I_\up,I_\dn\subset \calI\backslash\{0\}}
{N(I_\up;I_\dn)=r}
\sumJo
|\tg{k}{-k}|^2 \nonumber \\
& \le & \frac{(2\nu-r-1)(r+1)}{\nu^2}
\sumtwo{I_\up,I_\dn\subset \calI\backslash\{0\}}
{N(I_\up;I_\dn)=r+1}
|g(I_\up;I_\dn)|^2\,.
\end{eqnarray}
To get the final line, we have used the fact that,
for a pair of $I_\up$ and $I_\dn$ with $N(I_\up;I_\dn)=r+1$,
there are $(r+1)$ elements $n$ in $\calD\cup(\calJ\backslash\{0\})$
for which we can find a suitable pair $I^\prime_\up$ and $I^\prime_\dn$ with
$N(I^\prime_\up;I^\prime_\dn)=r$ such that either
$\{n\}\cup I^\prime_\up = I_\up$ and $\{n\}\cup I^\prime_\dn = I_\dn$
or
$\{n\}\cup I^\prime_\up = I_\up$ and $\{-n\}\cup I^\prime_\dn = I_\dn$
holds.
Therefore we obtain
\begin{equation}
\label{eq:Fr ge 0}
F_r[\Phi] \ge \left( 1-\frac{(2\nu-r-1)(r+1)}{\nu^2} \right)
\sumtwo{I_\up,I_\dn\subset \calI\backslash\{0\}}
{N(I_\up;I_\dn)=r+1}
|{g}(I_\up;I_\dn)|^2\,.
\end{equation}
Since $\frac{(2\nu-r-1)(r+1)}{\nu^2}\le1$,
we conclude that $F_r[\Phi]\ge0$.
Finally, we examine the condition for equality $F[\Phi] = 0
\footnote
We could easily find the condition for this equality
in the case of the kagome lattice~\cite{TanakaUeda2003},
but we have to consider this carefully in the present model.
The condition for the equality depends on
the form of the finite energy condition~\eqref{eq:finite energy condition}.
This point should be kept in mind in applying the present method
to other models.
}.
Since $F^\prime[\Phi]\ge0$ by the definition
and we have shown $F_r[\Phi]\ge 0$,
$F[\Phi]=0$ holds only when $F^\prime[\Phi]$ and $F_r[\Phi]$ are vanishing.
Here it is easy to see that $F^\prime[\Phi]=0$ holds only if all the coefficients
appearing in the sum in the right-hand side of~\eqref{eq:Fprime},
i.e., coefficients $g(I_\up;I_\dn)$ for $I_\up$ and $I_\dn$ such that
$0 \notin I_\up\cup I_\dn$ and $N(I_\up;I_\dn)=0$,
are identically zero.
To see when $F_r[\Phi]=0$ holds,
let us look into the inequality~\eqref{eq:Fr ge 0}.
The factor $\frac{(2\nu-r-1)(r+1)}{\nu^2}$ in~\eqref{eq:Fr ge 0} is equal to 1
if $r=\nu-1$ and is less than 1 otherwise.
Therefore we find that $F_r[\Phi]$ with $r\ne \nu-1$
is vanishing only if all the coefficients $g(I_\up;I_\dn)$
in the sum in the right-hand side
of~\eqref{eq:Fr} are zero.
On the other hand,
we need further consideration for $r=\nu-1$.
By using the inequality ~\eqref{eq:inequality for tilde g 1} with $r=\nu-1$,
we obtain
\begin{eqnarray}
\label{eq:r=nu-1}
\sumtwo{I_\up,I_\dn\subset \calI\backslash\{0\}}
{N(I_\up;I_\dn)=\nu-1}
|\g{0}{0}|^2
& \le & \frac{1}{\nu}
\sumtwo{I_\up,I_\dn\subset \calI\backslash\{0\}}
{N(I_\up;I_\dn)=\nu-1}
\sum_{l\in \calD}
|\tg{l}{l}|^2 \nonumber\\
&& +\frac{\nu}{(\nu+1)^2}
\sumtwo{I_\up,I_\dn\subset \calI\backslash\{0\}}
{N(I_\up;I_\dn)=\nu-1}
\sumJo
|\tg{k}{-k}|^2.
\nonumber\\
\end{eqnarray}
The right-hand side is rewritten a
\footnote{
Note that $|I_\up\cap I_\dn\cap\calD|=n$ implies $|I_\up\cap {\bar{I}_\dn}\cap(\calJ\backslash\{0\})|=\nu-n$
when $N(I_\up;I_\dn)=\nu$.
Then, considering in a similar way below~\eqref{eq:Fr lower bound}, we obtain $R_n$.
}
\begin{equation}
\sum_{n=0}^{\nu}
\sumtwo{I_\up,I_\dn\subset \calI\backslash\{0\}}
{N(I_\up;I_\dn)=\nu,~|I_\up\cap I_\dn\cap\calD|=n}
R_n |{g}(I_\up;I_\dn)|^2
\end{equation}
with $R_n=n/\nu+(\nu-n)\nu/(\nu+1)^2$.
Noting that $R_\nu=1$ we obtain
\begin{equation}
F_{\nu-1}[\Phi]\ge\sum_{n=0}^{\nu-1} \left(1-R_n\right)
\sumtwo{I_\up,I_\dn\subset \calI\backslash\{0\}}
{N(I_\up;I_\dn)=\nu,~|I_\up\cap I_\dn\cap\calD|=n}
|{g}(I_\up;I_\dn)|^2.
\end{equation}
Now suppose that $F_{\nu-1}[\Phi]=0$.
Since $R_n<1$ for $n=0,\dots,\nu-1$,
we find that
all the coefficients ${g}(I_\up;I_\dn)$ in the sum in the right-hand side
of the above inequality should be zero.
We thus have $\tilde{g}(k,I_\up;-k,I_\dn)=0$
with $k\in\calJ\backslash\{0\}$
for any $I_\up,I_\dn\subset \calI\backslash\{0\}$ such that $N(I_\up;I_\dn)=\nu-1
\footnote
\label{fn:memo}
Put $I_\up^\prime=\{k\}\cup I_\up$ and $I_\dn^\prime=\{-k\}\cup I_\dn$.
If $k\in I_\up$ or $-k\in I_\dn$, we have $\tilde{g}(k,I_\up;-k,I_\dn)=0$ by the definition;
otherwise, since $N(I_\up^\prime;I_\dn^\prime)=\nu$ and
$|I_\up^\prime \cap I_\dn^\prime \cap \calD|\ne\nu$, we have $g(I_\up^\prime;I_\dn^\prime)=0$, which
implies $\tilde{g}(k,I_\up;-k,I_\dn)=0$.
}.
Then, \eqref{eq:final condition} becomes
\begin{equation}
\label{eq:new condition on tg}
\tg{0}{0}
= -\frac{1}{\nu}
\sum_{l\in \calD}
\tg{l}{l}
\end{equation}
for $N(I_\up;I_\dn)=\nu-1$. Note that the number of non-zero $\tilde{g}$
in the right-hand side of \eqref{eq:new condition on tg} is at most on
\footnote
Put $I_\up^\prime=\{l\}\cup I_\up$ and $I_\dn^\prime=\{l\}\cup I_\dn$.
If $I_\up^\prime \cap I_\dn^\prime \cap \calD\ne\calD$,
we have $\tilde{g}(l,I_\up;l,I_\dn)=0$ (see also footnote \ref{fn:memo}).
If $I_\up^\prime \cap I_\dn^\prime \cap \calD=\calD$, $|I_\up \cap I_\dn \cap \calD|=\nu-1$
and there is only one element $l$ in $\calD$ for
which $\tg{l}{l}$ can be non-zero.
}.
Repeating the same argument as above with the relation \eqref{eq:new condition on tg},
we find that
$F_{\nu-1}[\Phi]=0$
only when all the coefficients in the sum in the right-hand side
of \eqref{eq:Fr}
are zero.
Therefore, we conclude that $F[\Phi]=0$ holds
only if $g(I_\up;I_\dn)=0$ for any pair of
$I_\up$ and $I_\dn$ such that $0\in I_\up\cap I_\dn$
and
$0\notin I_\up\cup I_\dn$.
So far we have proved that $E[\Phi]\ge -2\nu s$ and that the minimum
expectation value $-2\nu s$ is attained by states $\Phi$
which satisfy finite energy state
condition~\eqref{eq:finite energy condition} and can be expanded as
\begin{equation}
\Phi = \sumtwo{I_\up,I_\dn\subset \calI}
{0\in I_\up\cup I_\dn,0\notin I_\up\cap I_\dn}
g(I_\up;I_\dn)\Phi(I_\up;I_\dn).
\end{equation}
It follows {from} the above expression that
any state which attains
the minimum expectation value $-2\nu s$ is
an eigenstate of $h_0$ as well as
${-s\sum_{\sigma}a_{0,\sigma}^\dagger a_{0,\sigma}}$.
It is noted that
there really exist such eigenstates
under condition~\eqref{eq:finite energy condition};
$d_{0,\up}^\dagger{\Phi}_0$ and
$\left(\prod_{i\in \calI}d_{i,\up}^\dagger\right){\Phi}_0$
are examples.
By the continuity of eigenvalues
we conclude that $-2\nu s$ is
the minimum eigenvalue of $h_0$ for sufficiently large
values of $t/s$ and
$U/s$, and that
the corresponding eigenstates are the same as those
which give the minimum expectation value of $h_0$
in the limit $t/s,U/s\to\infty$.
It is easy to check that such eigenstates have
the properties stated in Lemma~\ref{lemma:main}.
Because of the translation invariance,
the same holds for all $\halph$ and
this completes
the proof of Lemma~\ref{lemma:main}.
\section{Proof of {Theorem~\ref{th:main}} for $s>0$; nearly-flat-band ferromagnetism}
\label{s:proof of theorem}
Suppose that the number $\Ne$ of electrons is $|\calV|$ and
that the values of $t/s$ and $U/s$
are so large that Lemma~\ref{lemma:main} holds.
We note that Lemma~\ref{lemma:main} is associated only
with the local Hamiltonians
and that how large $t/s$ and $U/s$
should be is independent of
the size of the whole lattice.
Since the minimum eigenvalue of
the local Hamiltonians $\halph$
is $-2\nu s$, as claimed in Lemma~\ref{lemma:main},
the eigenvalue of
the Hamiltonian $H=\sum_{\alpha\in \calV}\halph$
is bounded {from} below by $-2\nu s|\calV|$.
On the other hand, taking the fully saturated ferromagnetic state
\begin{equation}
\label{eq:ferro state}
\Phi_\mathrm{ferro}
= \left(\prod_{\alpha\in \calV}a_{\alpha,\up}^\dagger\right)\Phi_0
\end{equation}
as a variational state for $H$ and noting that
\begin{equation}
-s a_{\beta,\dn}^\dagger a_{\beta,\dn}\Phi_\mathrm{ferro} = 0
\end{equation}
and
\begin{equation}
-s a_{\beta,\up}^\dagger a_{\beta,\up}\Phi_\mathrm{ferro}
= -s(2\nu-a_{\beta,\up} a_{\beta,\up}^\dagger)\Phi_\mathrm{ferro}
=-2\nu s \Phi_\mathrm{ferro}
\end{equation}
for all $\beta\in \calV$,
we find that
the upper bound on the ground state energy
is also given by $-2\nu s|\calV|$.
Therefore,
the ground state energy
is exactly $-2\nu s|\calV|$.
It is easy to see that the state $\Phi_\mathrm{ferro}$ has
the maximal total spin $\Stot=\Ne/2$.
Our remaining task is to prove
the uniqueness of the ground state up to $(\Ne+1)$-fold degeneracy
due to the spin rotation symmetry.
Let $\PhiG$ be a ground state,
and assume that
$\PhiG$ is expanded in terms of the basis states
$\Phi(V_\up,B_\up;V_\dn,B_\dn)$ in~\eqref{eq:basis}
with $\Ne=|\calV|$.
Since $H\PhiG=-2\nu s|\calV|\PhiG$, $\PhiG$ must satisfy
$\halph\PhiG = -2\nu s \PhiG$ for all $\alpha\in \calV$.
Thus $\PhiG$ must satisfy the properties stated in Lemma~\ref{lemma:main}.
The condition~\eqref{eq:condition1 in Lemma} implies
that
\begin{equation}
a_{\alpha,\up}^\dagger a_{\alpha,\dn}^\dagger\PhiG=0
\end{equation}
for any $\alpha\in\calV$.
Since $a_{\alpha,\up}^\dagger a_{\alpha,\dn}^\dagger\Phi(V_\up,B_\up;V_\dn,B_\dn)\ne0$
if $\alpha\notin V_\up\cup V_\dn$ and the states of the form
$a_{\alpha,\up}^\dagger a_{\alpha,\dn}^\dagger\Phi(V_\up,B_\up;V_\dn,B_\dn)$ are
linearly independent with each other,
$\Phi$ is expanded only in terms of the basis states $\Phi(V_\up,B_\up;V_\dn,B_\dn)$
with $V_\up \cup V_\dn=\calV$.
Then, taking into account $\Ne=|\calV|$,
we find that $\PhiG$ is written as
\begin{equation}
\label{eq:PhiG spin representation}
\PhiG = \sum_{\{\sigma\}} \varphi(\{\sigma\})
\left(\prod_{\alpha\in \calV}
a_{\alpha,\sigma_\alpha}^\dagger
\right)\Phi_0 \,,
\end{equation}
where $\{\sigma\}$ is a shorthand for a spin configuration
$\{\sigma_\alpha\}_{\alpha\in \calV}$ of electrons each of
which is singly occupying a state corresponding to
${a}_{\alpha,\sigma_\alpha}$, the summation
is taken over all spin configurations, and $\varphi(\{\sigma\})$
is a new complex coefficient.
As for the product of fermion operators, we adopt the same rule as
in~\eqref{eq:d basis} with another one-to-one mapping $\theta^\prime$ from $\calV$ to $\Bbb{Z}$.
Let us impose
the condition~\eqref{eq:condition2 in Lemma} on $\PhiG$
in the form of~\eqref{eq:PhiG spin representation}.
Note that \eqref{eq:condition2 in Lemma} must hold for sites in all $\La_\alpha$,
i.e., for all sites in $\La$.
Choose site $x=m(\beta,\gamma)$.
By the definition of $a_{\alpha,\sigma}$, one finds that
$\{c_{m(\beta,\gamma)},a_{\alpha,\sigma}^\dagger\}$ is 1 if $\alpha$ is either $\beta$ or $\gamma$,
and zero otherwise.
Then, we have
\begin{eqnarray}
\lefteqn{c_{m(\beta,\gamma),\dn}c_{m(\beta,\gamma),\up}\PhiG} \nonumber\\
&=&
\sum_{\{\sigma\}}
\mathsf{S}_{\calV}^\beta \mathsf{S}_{\calV \backslash\{\beta\} }^\gamma
\varphi(\{\sigma\})
c_{m(\beta,\gamma),\dn}c_{m(\beta,\gamma),\up}
a_{\beta,\sigma_\beta}^\dagger a_{\gamma,\sigma_\gamma}^\dagger
\left(
\prod_{\alpha\in \calV\backslash\{\beta,\gamma\}}
{a}_{\alpha,\sigma_\alpha}^\dagger
\right)\Phi_0
\nonumber\\
&=&
\sum_{\{\sigma\}}
\mathsf{S}_{\calV}^\beta \mathsf{S}_{\calV \backslash\{\beta\} }^\gamma
\varphi(\{\sigma\})
\left\{
\chi[\sigma_\beta=\up,\sigma_\gamma=\dn]
-
\chi[\sigma_\beta=\dn,\sigma_\gamma=\up]
\right\}
\left(
\prod_{\alpha\in \calV\backslash\{\beta,\gamma\}}
{a}_{\alpha,\sigma_\alpha}^\dagger
\right)\Phi_0
\nonumber\\
\end{eqnarray}
where $\mathsf{S}_{\calV}^\beta$ and $\mathsf{S}_{\calV \backslash\{\beta\}}^\gamma$
are defined similarly as in~\eqref{eq:sign factor S}
with $\theta$ replaced by $\theta^\prime$.
Since all the states
$\left(
\prod_{\alpha\in \calV\backslash\{\beta,\gamma\}}
{a}_{\alpha,\sigma_\alpha}^\dagger
\right)\Phi_0$
in the last line are linearly independent,
it follows from
the condition $c_{m(\beta,\gamma),\dn}c_{m(\beta,\gamma),\up}\PhiG=0$
that
$
\varphi(\{\sigma\})=\varphi(\{\tau\})
$
for any pair of spin configurations
$\{\sigma\}$ and $\{\tau\}$ such that
$\sigma_\beta=\tau_\gamma, \sigma_\gamma=\tau_\beta$,
and $\sigma_\alpha=\tau_\alpha$ for $\alpha\ne\beta,\gamma$.
Repeating the same argument for all sites in $\La$,
we find that
\begin{equation}
\label{eq:condition on coefficients}
\varphi(\{\sigma\}) = \varphi(\{\tau\})
\end{equation}
if
$\sum_{\alpha\in \calV}\sigma_\alpha=\sum_{\alpha\in \calV}\tau_\alpha$
(we regard $\up$ as +1 and $\dn$ as $-1$ in the sum).
Therefore $\PhiG$ is always expanded as
\begin{equation}
\PhiG = \sum_{M=-\Ne}^{\Ne}\varphi_M (S_\mathrm{tot}^-)^{\Ne-M}
\Phi_\mathrm{ferro}
\end{equation}
with complex coefficients $\varphi_M$.
This completes the proof of Theorem~\ref{th:main} for $s>0$.
\jour{
\bigskip\\
{\small
\textbf{Acknowledgements}~~I would like to thank Hal Tasaki for valuable discussions.The present work was supported by JSPS Grants-in-Aid for Scientific Research no.~25400407.}
}
{
\begin{acknowledgements}
I would like to thank Hal Tasaki for valuable discussions.The present work was supported by JSPS Grants-in-Aid for Scientific Research no.~25400407.
\end{acknowledgements}
\begin{appendices}
\renewcommand{\theequation}{\thesection.\arabic{equation}}
\setcounter{equation}{0}
\section{Proof of {Theorem~\ref{th:main}} for $s=0$; flat-band ferromagnetism}
\label{s:flat-band ferro}
Throughout this section, we assume that $s=0$
and the other parameters $t$ and $U$ take
arbitrary positive values.
As we have seen in section~\ref{s:single electron},
the single-electron ground state energy is zero
and $|\calV|$-fold degenerate for $s=0$, and the single-electron
ground states
are given by $\aasd\Phi_0$ with $\alpha\in \calV$.
Since both $\Hhop$ with $s=0$ and $\Hint$ are positive semidefinite,
we find that the energy eigenvalue of $H$ is bounded from below by zero.
Taking a state $\Phi_\mathrm{ferro}$ in \eqref{eq:ferro state}
as a variational state, we conclude that the ground state energy of $H$ for $\Ne=|\calV|$
is exactly zero.
To show the uniqueness of the ground state,
we will use the following property which was pointed out by Mielke~\cite{Mielke93}.
Consider Hubbard models with $M$-fold degenerate
single-electron ground state energy.
Then, we can construct
a set of $M$ single-electron ground states
which satisfy the following condition.
Let $\Phi_{1,\sigma}^{(i)}$ with
$i=1,2,\dots,M$ be linearly independent
single-electron ground states with spin $\sigma$.
For each $i$, there exists site $x_i$ such
that $c_{x_i,\sigma}\Phi_{1,\sigma}^{(i)} \ne 0 $ and
$c_{x_i,\sigma}\Phi_{1,\sigma}^{(j)}=0$ for $j\ne i$.
In our model, we can construct single-electron ground states
which possess the above property in the following manner.
For each $\alpha\in {\calV}$, let us
define a new fermion operator by
\begin{equation}
\bar{a}_{\alpha,\sigma}
= a_{\alpha,\sigma}
-\sum_{n=1}^{L-1}(-1)^{n}a_{\alpha+n\delta_\nu,\sigma}
\end{equation}
(recall that $\calV$ is periodic). Since $\bar{a}_{\alpha,\sigma}$ is a linear combination
of $a_{\beta,\sigma}$ with $\beta\in \calV$,
the single-electron states $\bar{a}_{\alpha,\sigma}^\dagger\Phi_0$
apparently have energy 0.
Furthermore, we have
\begin{equation}
\label{eq:c bar a}
\acom{\bar{a}_{\beta,\sigma}^\dagger,c_{m(\alpha,\alpha+\delta_\nu),\sigma}}=
\left\{
\begin{array}{@{\,}ll}
2 & \mbox{if $\beta=\alpha$;}\\
0 & \mbox{otherwise.}
\end{array}
\right.
\end{equation}
These anticommutation relations imply that
the single-electron states $\bar{a}_{\alpha,\sigma}^\dagger\Phi_0$
with $\alpha\in \calV$ have the desired property.
We note that \eqref{eq:c bar a} also implies
the linear independence of these single-electron states.
Let us prove the uniqueness of the ground state
by using $\{\bar{a}_{\alpha,\sigma}^\dagger\}_{\alpha\in{\calV}}$
introduced as above.
Let $\PhiG$ be a ground state, which must be
a zero energy state for both $\Hhop$ and $\Hint$.
Since $\PhiG$ is a zero energy state for $\Hhop$, we expand it as
\begin{equation}
\PhiG = \sumtwo{V_\up,V_\dn\subset \calV}
{|V_\up|+|V_\dn|=|\calV|}
\varphi^\prime(V_\up;V_\dn)
\left(
\prod_{\alpha\in V_\up}
\bar{a}_{\alpha,\up}^\dagger
\right)
\left(
\prod_{\alpha\in V_\dn}
\bar{a}_{\alpha,\dn}^\dagger
\right)
\Phi_0
\label{eq:flat-band PhiG}
\end{equation}
with complex coefficients $\varphi^\prime(V_\up;V_\dn)$.
To be a zero energy state of the on-site interaction, $\PhiG$
in the form of~\eqref{eq:flat-band PhiG}
must further satisfy the condition $c_{x,\dn}c_{x,\up}\PhiG=0$ for all $x$ in $\La$.
We examine this condition for sites $x=m(\alpha,\alpha+\delta_\nu)$ with $\alpha\in\calV$.
Then, taking account of \eqref{eq:c bar a},
we find that $\varphi^\prime(V_\up;V_\dn)$ is vanishing if $V_\up\cap V_\dn\ne\emptyset$.
A ground state $\PhiG$ is thus rewritten as
\begin{equation}
\PhiG = \sum_{\{\sigma\}} \varphi^\prime(\{\sigma\})
\left(
\prod_{\alpha \in \calV}
\bar{a}_{\alpha,\sigma_\alpha}^\dagger
\right)
\Phi_0
\end{equation}
with new coefficients $\varphi^\prime(\{\sigma\})$.
Now we have expressed $\PhiG$
in the same fashion as in~\eqref{eq:PhiG spin representation}.
Examining repeatedly the condition $c_{x,\dn}c_{x,\up}\PhiG=0$ for sites $x=m(\beta,\beta+\delta_l)$
with $l\ne\nu$,
we obtain the same relation for $\varphi^\prime(\{\sigma\})$
as that in~\eqref{eq:condition on coefficients}.
Therefore, we conclude that the ground state is unique apart from the degeneracy due to
the spin rotation symmetry.
This completes the proof of Theorem~\ref{th:main} for $s=0$.
\end{appendices}
|
\section*{Introduction}
The understanding of molecular structures provides life sciences with tools to explain complex cell-biology phenomena. Biological complexity of mesoscopic objects along with the quantum behavior of their basic elements leads to interesting unsolved questions awaiting comprehensive answers \cite{Whaley2010}. The interdisciplinary field of ``quantum biology'' is the natural area for combining quantum physical methods and tools to investigate, model and simulate biological systems on a mesoscopic level \cite{Arndt2009,Plenio2013,Li2016}.
Many biological processes are powered by the energy released from the hydrolysis of adenosine triphosphate (ATP). On a physical level, it can be viewed as a vibrational bound state of the ATP molecule to a protein with energy equal to $0.49$ eV. In the 1970's Davydov proposed a mechanism for the localization and transport of the associated vibrational energy in the $\alpha$-helix region of a protein by means of a so-called Davydov soliton \cite{Davydov1969,Davydov1973a,Davydov1973b,Davydov1976}.
Although this model has been used for a theoretical description of experimentally observed unconventional absorption bands in proteins \cite{Scott1983,Scott1984}, direct experimental evidence for existence of this soliton is still missing.
The Davydov soliton is a subclass of richer {\it polaron} phenomena, \textit{i.e}., excitations mediated by phonons originally introduced by Landau \cite{Landau}. Polarons have been broadly studied theoretically as well as experimentally in a condensed matter context \cite{Alexandrov} and recently in different areas of ultracold physics: ultracold ions \cite{Cirac2004,Zoller2008,Monroe2009,Hague2012,Solano2012,Cirac2012,Lamata2014}, polar molecules \cite{Herrera2010a,Herrera2010b,Herrera2011,Lesanovsky2012,Herrera2013}, ultracold Rydberg gases \cite{MacCormick2012,Hague2014,Wuster2015,Graetzle2015,Pohl2015,Molmer2017}, and strongly-interacting ultracold Bose and Fermi gases \cite{Zwerlein2009,Kohl2012,Deborah2016,Jorgensen2016,Levinsen2016,Nakano2017,Demler2017,Roati2017}.
In this paper we show that the Davydov soliton can be created and observed in a suitably prepared system of ultracold atoms, confined in an optical lattice and off-resonantly coupled to a Rydberg state \cite{Henkel2010,Johnson2010,Honer2010,Pupillo2010}. Such a system can be regarded as a dedicated quantum simulator within the broader class of the Holstein-Su-Schrieffer-Heeger (HSSH) Hamiltonian \cite{Holstein,SSH_model}.
First, we investigate dynamical properties of the system within the semi-classical Davydov approach assuming an infinite number of phonons. Second, we use an exact evolution approach to study the dynamics of the system in the non-classical few-phonons regime. In a particular, experimentally accessible parameter regime, both approaches confirm the existence of soliton solutions. This also indicates that with this Rydberg quantum simulator it is possible to study the regime in between the so-called small and large polaron regimes of the HSSH model, which cannot be described by perturbative theoretical methods.
\section{The simulator}
We mimic the behavior of the above-mentioned bio-molecules with a system of ultracold atoms confined in a very deep one-dimensional optical lattice potential $V(x)=V_0\sin^2(2\pi x/R_0)$ where each lattice site is occupied exactly by one atom. We assume that the spatial dynamics of atoms is not completely frozen, \textit{i.e.}, atoms may oscillate in vicinities of local minima with frequency $\omega_0 = \sqrt{2V_0\pi^2/mR_0^2}$. This motion is however quantized and therefore it is driven by a simple harmonic oscillator-like Hamiltonian:
\begin{equation}
\hat{\cal H}_{\text{vib}} = \sum_i\left(\frac{\hat p_i^2}{2m}+\frac{m\omega_0^2}{2}\hat{u}_i^2\right) = \sum_i \hbar\omega_0 \hat{b}_i^\dagger\hat{b}_i,
\end{equation}
where $\hat{u}_i = l_0(\hat{b}^\dagger_i + \hat{b}_i)/\sqrt{2}$ and $\hat{p}_i = i\hbar(\hat{b}^\dagger_i - \hat{b}_i)/(l_0\sqrt{2})$ are the position and momenta operators related to $i$-th atom, while operator $\hat{b}_i$ annihilates vibrational excitation of the $i$-th atom. Local motion defines a natural scale of length, $l_0 = \sqrt{\hbar/m\omega_0}$.
Besides spatial motion, each atom may exhibit changes of its internal state due to the long-range interactions between neighboring atoms via Rydberg dressing mechanism \cite{Pohl2014,Buchler2010,Rost2014,Gross2016,Ates2008}. Following the work by W\"{u}ster {\it et al.}\cite{Wuster2011} coupling of the internal state of atoms with their spatial motion can be realized by the off-resonant coupling of two different but degenerated internal Zeeman levels in the different hyperfine states of the ground-state manifold of the atoms. The large hyperfine splitting will permit selective addressing of the levels $|g\rangle$ and $|g'\rangle$ to two precisely selected, highly excited Rydberg states $|nS\rangle$ or $|nP\rangle$ (via two- and single-photon transition, respectively) with principal quantum number $n$ and angular momentum equal to $0$ or $\hbar$, respectively. A perturbation analysis shows that this coupling results in a quite small admixture of a Rydberg state to the atomic ground states and, as a consequence, atom can be found in one of the two dressed states\cite{Wuster2011}:
\begin{align}
|\mathtt{0}\rangle \approx |g\rangle + \alpha_s |nS\rangle, \qquad
|\mathtt{1}\rangle \approx |g'\rangle + \alpha_p |nP\rangle,
\end{align}
where amplitudes $\alpha_{l} = \varTheta_{l}/2\varDelta_{l}$ ($l\in\{s,p\}$) are determined by a total Rabi frequency of a driving field $\varTheta_{l}$ and a total laser detuning $\varDelta_{l}$.
In this basis of dressed states the dipole-dipole interaction between neighboring atoms is $C_3^{sp}/R^3$ ($R$ is the spatial distance between the atoms), besides additional contribution to the energy gap between local states $|\mathtt{0}_i\rangle$ and $|\mathtt{1}_i\rangle$, may induce transitions (excitation hoppings) between internal states of neighboring atoms $|\mathtt{0}_i\rangle|\mathtt{1}_{i+1}\rangle\leftrightarrow |\mathtt{1}_i\rangle|\mathtt{0}_{i+1}\rangle$. In consequence, the excitation $|\mathtt{1}\rangle$ can be effectively transported across the lattice. This effect is driven by the following Hamiltonian of internal motion of all atoms:
\begin{equation}\label{Hamiltonian_1}
\hat{\cal H}_\mathrm{exc} = \sum_i W_i\hat{a}^\dagger_i\hat{a}_i + \sum_iJ_{i+1,i}(\hat{a}^\dagger_{i+1}\hat{a}_i+\hat{a}^\dagger_{i}\hat{a}_{i+1}),
\end{equation}
where an annihilation operator of an excitation $\hat a_i$ can be viewed as a local transition operator $|\mathtt{1}_i\rangle\langle\mathtt{0}_i|$ between dressed Rydberg states. The spatial dependent parameters $W_i$ and $J_{i,i+1}$ are related to the dipole-dipole forces induced by Rydberg dressing and are given by\cite{Wuster2011}:
\begin{align}\label{JiW}
W_i &=\frac{\alpha^4\hbar\varDelta }{2}\left(\frac{1}{1-\kappa(R_0+u_{i+1}-u_i)^2} + \frac{1}{1-\kappa(R_0+u_{i}-u_{i-1})^2}\right), \nonumber \\
J_{i+1,i} &=\frac{\alpha^4 C_3^{sp}}{|R_0+u_{i+1}-u_i|^3} \frac{1}{1-\kappa(R_0+u_{i+1}-u_i)^2},
\end{align}
where $\kappa(R) = (C_3^{sp}/\hbar\varDelta)/ R^3$ and $\varDelta = \varDelta_s + \varDelta_p$.
In a static situation, when all atom positions are frozen, the energies $W_i$ and $J_{i+1,i}$ are site-independent with values controlled by dipole-dipole interactions between neighboring atoms at fixed lattice spacing $R_0$. However, due to the vibrational motion of atoms, these parameters are position dependent and they couple internal states of atoms with their motional degrees of freedom. In the lowest order of approximation they can be written as $W_i = W_0 + g_W(u_{i+1}+ u_{i-1})$, $J_{i+1,i} = -J_0 + g_J (u_{i+1}- u_i)$, where $g_W$ and $g_J$ are the appropriate Taylor expansion coefficients of \eqref{JiW} around $R_0$.
Moreover, since the vibrational motion is quantized, the parameters have an operator character when acting in the subspace of spatial motion of atoms.
By inserting expanded $W_i$ and $J_{i+1,i}$ to the Hamiltonian \eqref{Hamiltonian_1} one obtains HSSH Hamiltonian
\begin{equation}\label{Hamiltonian_final}
\begin{split}
\cal{\hat{H}} = & \sum_i W_0 \hat{a}^\dagger_i\hat{a}_i -J_0(\hat{a}^\dagger_{i+1}\hat{a}_i + h.c.) + \sum_i g_W(u_{i+1} - u_{i-1})\hat{a}^\dagger_i\hat{a}_i + \sum_i g_J(u_{i+1} - u_{i})(\hat{a}^\dagger_{i+1}\hat{a}_i + h.c.)
\end{split}
\end{equation}
where the first term describes the excitation dynamics on the lattice, the second term describes the excitation-vibration coupling via the on-site energy, and the last term represents the off-site coupling via the excitation-vibration coupling through the hopping energy. Our implementation can be regarded as a dedicated quantum simulator to study excitation stabilization by vibrations related to the $\alpha$-helix protein.
For the Rydberg parameters that we consider, the second order terms in the Taylor expansion are significantly smaller than the first-order corrections, which justifies a linear approximation.
For the moment we comment on a special case of Hamiltonian (\ref{Hamiltonian_final}) with zero off-site coupling $g_J = 0$, \textit{i.e.} Holstein model. In this model the excitation is dressed by phonons forming a polaron quasiparticle. Two limiting cases have exact solutions: $(i)$ For zero excitation-vibration coupling $g_W = 0$ eigenstates are well described by product states in Fourier space \textit{i.e.} $\hat{a}^{\dagger}_k|0\rangle, \hat{a}^\dagger_{k-q}\hat{b}^\dagger_q|0\rangle, \hat{a}^\dagger_{k-q-q'}\hat{b}^\dagger_q \hat{b}^\dagger_{q'}|0\rangle$, etc. These states construct a good basis for perturbation expansions in the $g_\text{W}/J_0$ parameter for the weak coupling limit; $(ii)$ a second limiting case, called the small polaron, corresponds to a zero hopping energy term $J_0 = 0$ in which the Hamiltonian can be diagonalized in the dressed-polaron picture, \textit{i.e.}
\begin{equation}
\begin{split}
\cal{\hat{H}} & = \sum_{i}\hbar\omega_0(\hat{b}_{i}^\dagger \hat{b}_{i}) + g_w \sum_{i} \hat{a}_i^\dagger a_i (\hat{b}_{i}^\dagger + \hat{b}_{i}) = \hbar\omega_0\hat{b}_0^\dagger\hat{b}_0 + g_W(\hat{b}_0^\dagger+\hat{b}_0) + \hbar\omega_0\sum_{i\neq 0}(\hat{b}_i^\dagger\hat{b}_i+1/2) \\
& = \hbar\omega_0\hat{B}_0^\dagger\hat{B}_0 - \frac{g_W^2}{\hbar\omega_0} + \hbar\omega_0\sum_{i\neq 0}(\hat{b}_i^\dagger\hat{b}_i),
\end{split}
\end{equation}
where $\hat{B}_0 = \hat{b}_0 + g_W/\hbar\omega_0$. The dressed-polaron eigenstates $\hat{a}^\dagger_0 |0\rangle_{ex}\hat{B}^\dagger_0 \prod_{i\neq 0}\hat{b}^{\dagger^{m_i}}|\tilde{0}\rangle_{ph}$ with eigenenergies
$E_n=-g_W^2/\hbar\omega_0 + n\hbar\omega_0$ form a good basis for a perturbative description in the small $g_W/\hbar\omega_0$ parameter .
A perturbative calculation shows that the excitation-vibration coupling is given by a dressed hopping amplitude $J_0e^{-(g_W/\hbar\omega_0)^2}$\cite{Alexandrov}, where for $\omega_0$ large enough, the hopping amplitude vanishes. For the HSSH model in the regime of parameters giving rise to a Davydov soliton, \textit{i.e.} where all parameters are of the same order, a good small parameter for a perturbative description is lacking, and therefore a variational approach is preferred.
The Rydberg dressing is responsible for the coupling between vibrational degrees of freedom of neighboring atoms. The resulting dressed soft-core interaction is proportional to $(R^6+R_b^6)^{-1}$, with $R_b$ the Rydberg blockade radius \cite{Honer2010}, gives also rise to an additional energy shift. However, this shift is negligible compared to $\hbar\omega_0$, and therefore we omit it. In the following, all energies are expressed in units of $J_0$, and time is measured in units of $\hbar/J_0$, {\it i.e.} we set $J_0=\hbar=m=1$.
\section{Dynamical properties of the system}
An important question related to the dynamics of the HSSH Hamiltonian is whether the lattice vibrations are able to stabilize the excitation that is initially localized on a specific site $K$
$
|\boldsymbol{\Psi}_0\rangle = \hat{a}_K^\dagger|\mathtt{vac}\rangle,
$
or slightly delocalized on two neighboring sites
$
|\tilde{\boldsymbol{\Psi}}_0\rangle = \frac{1}{\sqrt{2}}(\hat{a}_K^\dagger + \hat{a}_{K+1}^\dagger)|\mathtt{vac}\rangle,
$
where $|\mathtt{vac}\rangle$ is the vacuum state of the system fulfilling the condition $\hat{a}_i|\mathtt{vac}\rangle=\hat{b}_i|\mathtt{vac}\rangle=0$ for any $i$.
For certain parameters, a system prepared in these initial states evolves in such a way that the excitation does not spread across the protein. This is attributed to a specific ratio of the interactions of excitation and vibrational degrees of freedom, giving rise to a soliton. This spreading or non-spreading behavior can be extracted from information encoded in the time-dependent density profile $\rho_i(t) = \langle\boldsymbol{\Psi}(t)|\hat{a}^\dagger_i\hat{a}_i|\boldsymbol{\Psi}(t)\rangle,
$
where the state of the system at given time $t$ can be formally written as
$|\boldsymbol{\Psi}(t)\rangle = \mathrm{exp}\left(-i\hat{\cal H}t\right)|\boldsymbol{\Psi}_\mathtt{ini}\rangle$,
where $|\boldsymbol{\Psi}_\mathtt{ini}\rangle$ is one of the considered initial states.
Temporal spreading of the excitation is then given by an effective width of the spatial density profile
$
\sigma(t) = (N\left[\sum_i \rho_i^2(t)\right])^{-1}.
$
This quantity takes the value $1/N$ for an excitation localized at exactly one lattice site and $1$ when fully delocalized. In principle, by analyzing the time-dependence of $\sigma(t)$ one can easily determine whether the excitation remains localized or whether it spreads across the system. Numerically exact solutions of the evolution problem
are very challenging due to the strong non-linear quantum-mechanical coupling between excitation and vibrational degrees of freedom. Therefore generally the evolution of the system cannot be found exactly and some approximation methods have to be adopted.
\subsection*{The Davydov approach}
We discuss here the two-step Davydov approach \cite{Davydov1969}, which results in a semiclassical description of the system. In the first step one assumes that the state of the system $|\boldsymbol{\Psi}(t)\rangle$ can be well approximated by the product of two independent states $|\psi(t)\rangle$ and $|\phi(t)\rangle$ for excitation and vibrational degrees of freedom, respectively, $
|\boldsymbol{\Psi}(t)\rangle = |\psi(t)\rangle|\phi(t)\rangle$. Since the system is initially prepared in the state with precisely one excitation and the number of excitations is conserved, the state $|\psi(t)\rangle$ can be decomposed in the single-particle subspace, $|\psi(t)\rangle=\sum_i\psi_i(t)\hat{a}_i^\dagger|\mathtt{vac}\rangle$, where time-dependent functions $\psi_i(t)$ play the role of probability amplitudes for finding an excitation at site $i$. Consequently $\rho_i(t)=|\psi_i(t)|^2$. The second step relays on a semi-classical treatment of the vibrational degrees of the system. In analogy to other quantum field theories, we assume that the state $|\phi(t)\rangle$ has classical features, {\it i.e.}, it can be well approximated by the product of independent coherent states:
$
|\phi(t)\rangle =\mathrm{exp}\left[-i\sum_i( u_i(t)\hat{p}_i-p_i(t)\hat{u}_i)\right]|\mathtt{vac}\rangle,
$
where amplitudes $u_i(t)$ and $p_i(t)$ are expectation values of appropriate operators in the state $|\phi(t)\rangle$. Within these approximations, we calculate the expectation value of the many-body Hamiltonian on the system state and approximate the resulting equation of motion by the classical Hamilton equations \cite{Zhang1988}, we obtain
set of coupled differential equations of the form:
\begin{equation} \label{DavidovEq}
\begin{aligned}
i\frac{\mathrm{d}\psi_i(t)}{\mathrm{d}t} & = -(\psi_{i+1} + \psi_{i-1}) + g_W(u_{i+1} - u_{i-1})\psi_i + g_J[\psi_{i+1}(u_{i+1} - u_i) + \psi_{i-1}(u_i - u_{i-1})], \\
\frac{\mathrm{d}u^2_i(t)}{\mathrm{d}t^2} &= -\omega_0^2 u_i(t) + g_W \omega_0(|\psi_{i+1}|^2 - |\psi_{i-1}|^2) + g_J \omega_0[\psi_i^*(\psi_{i+1} - \psi_{i-1}) + \psi_i(\psi_{i+1}^* - \psi_{i-1}^*)].
\end{aligned}
\end{equation}
These are Davydov type equations \cite{Davydov1969,Davydov1973a,Davydov1973b,Davydov1976}, which
describe the dynamics of an excitation $\psi_i$ coupled to a gradient of a classical phonon field $u_i$ forming an effective self-trapping potential. An alternative derivation provided by Kerr \cite{Kerr1987} is based on the Heisenberg equations of phonon position and momentum operators. A complementary approach to the above Davydov equations, which are based on the Lagrangian variational principle, is one based on the Dirac-Frenkel-McLachan (DFM) variational principle.
This approach is commonly used in quantum molecular dynamics \cite{McLachlan1964,Lubich2004,Raab2000,Zhao2004,Perroni2004,Stojanovic2008,Zhou2012}, in which equations of motion for the variational parameters are obtained from the minimization of $\langle\delta\phi|\hat{\cal H}-i\partial_t|\phi\rangle$, where $\delta\phi$ denotes possible
variations of $\phi$ with respect to the variational parameters.
\subsection*{Phase diagram}
We perform a semi-classical evolution of the system governed by Eqns.~\eqref{DavidovEq}, which allows us to observe spreading or non-spreading evolution of an effective width of the spatial density profile as a function of the parameters $\{ \omega_0,g_W,g_J \}$. The results can be visualized by the phase diagrams presented in Figs.~\ref{fig1} and~\ref{fig2}. These diagrams are obtained by plotting the maximal value of the $\sigma(t)$ reached during the evolution up to maximal time $T_{\mathrm{max}}=10$. In order to avoid interference effects during the evolution caused by the boundaries, all calculations are performed with a sufficiently large lattice of $N=50$ lattice sites and with periodic boundary conditions.
We checked that the numerical results are insensitive to enlarging $N$ on the time-scales of study, and therefore the obtained results are also valid for infinitely large systems. Moreover, the chosen lattice size is similar to current experimental efforts in this direction.
\begin{figure}
\includegraphics[scale=0.3]{Figure1.eps}
\caption{(Color online) Left panel: Maximal value of an effective width of the spatial density profile $\mathrm{max}\left[\sigma(t)\right]$ as a function of $\omega_0$ and $g_W$ for vanishing coupling $g_J = 0$. A sharp crossover between non-spreading excitations (blue) and spreading excitation (dark red) is clearly visible. Different regions of the phase-diagram (bordered with white lines) correspond to a distinct nature of the exciton-vibration dynamics.
Right panel: Evolution of the excitation width $\sigma(t)$ for different points on the phase diagram (marked as white squares on the left panel).}
\label{fig1}
\end{figure}
First, we focus on the case of a completely localized initial state $|\boldsymbol{\Psi}_0\rangle$ for $g_J=0$ (Fig.~\ref{fig1}).
We qualitatively indicate five different regions on the phase diagram (left panel): {\it (I)} and {\it (II)} where the excitation is dressed by a cloud of vibrations and the excitation does spread; {\it (III)} where due to an exponential reduction of the hopping amplitude the excitation is localized in its initial position \cite{Fesser1982}; {\it (IV)} where the sum of vibration energy and exciton-vibration coupling is larger than the hopping energy, giving rise to Davydov-like soliton behavior; {\it (V)} where $g_W\gtrsim\omega_0$ corresponding to the Discrete Breathers-like behavior \cite{Flach1998,Juanico2007}. Distinct behavior of the system is also visible in these selected areas in the time evolution of $\sigma(t)$ (right panel of Fig.~\ref{fig1}).
This picture can be generalized to non-vanishing coupling $g_J$, which we investigate for the the second initial state $|\tilde{\boldsymbol{\Psi}}_0\rangle$ (Fig.~\ref{fig2}).
As can be seen, a slight delocalization of the initial state together with non-local coupling $g_J$ dramatically enhance the non-spreading behavior of the wave packet. It is a direct consequence of the non-local terms in \eqref{DavidovEq}.
\begin{figure}
\includegraphics[scale=0.45]{Figure2.eps}
\caption{(Color scale) Maximal value of the wave packet width $\mathrm{max}\left[\sigma(t)\right]$ for different initial states $|\boldsymbol{\Psi}_0\rangle$ and $|\tilde{\boldsymbol{\Psi}}_0\rangle$ (top and bottom row, respectively) and different non-local interactions $g_J=\{0,3,5\}$ (appropriate columns from left to right). Note that strong enhancement of the non-spreading behavior takes place for stronger $g_J$ and for a smeared out initial state.}
\label{fig2}
\end{figure}
\subsection*{Numerically exact approach}
The results obtained in the framework of the semi-classical Davydov approach can be supported by numerically exact dynamics governed by the many-body Hamiltonian \eqref{Hamiltonian_final}. In this approach we represent Hamiltonian $\hat{\cal H}$ as a matrix in the Fock basis spanned by many-body states
$|i\rangle|m_1,\ldots,m_N\rangle = \hat{a}^\dagger_i(\hat{b}^\dagger_1)^{m_1}\cdots(\hat{b}^\dagger_N)^{m_N} |\mathtt{vac}\rangle$,
{\it i.e.}, states with an excitation located exactly at site $i$ and with selected vibrational states $m_i$ for all sites. An arbitrary state of the system can be expressed as an appropriate superposition of the basis states. Since the operator of a total number of vibrations in the system $
\hat{\cal N}_\mathrm{vib}=\sum_i \hat{b}_i^\dagger\hat{b}_i
$
does not commute with the Hamiltonian \eqref{Hamiltonian_final}, therefore an exact evolution is obtained only in the limit where all Fock states are taken into account. In practice, for numerical purposes, we assume that the total number of excitations cannot be larger than some well defined cut-off $M$. Then the results are treated as exact if increasing $M$ does not change the outcome noticeably \cite{Sowinski2,Dobrzyniecki}. Therefore, for a given $M$, one can perform calculations only for a small range of parameters for which creations of vibrations is limited. It is worth noticing that numerical complexity grows exponentially with the cut-off $M$. For our parameters, $N=50$ and $M=3$ the size of the corresponding Hilbert space exceeds 1.1 million. Other approaches based on quazi-exact dynamics are presented in \cite{Trugman2013,Trugman2015}.
As already expected from Fig.(\ref{fig2}), the numerically exact approach confirms that the on-site coupling ($g_W\ne 0)$ plays a dominant role in the excitation stabilization process. Therefore, without loss of generality, we consider now the minimal-coupling scenario with $g_J=0$
In Fig.~\ref{fig3} we show the time evolution of an initially localized excitation for $\omega_0=3$ and for three different values of the local coupling parameter $g_W=\{0.1, 0.75, 1.5\}$. It is clearly visible that for larger $g_W$ the wave packet of the excitation becomes more stable and spreads less. This effect is directly reflected in the number of vibrational modes created, which can be seen in the bottom row of Fig.~\ref{fig3}.
One can observe that increasing fluctuations of the total vibrations in the system stabilize excitation.
Since we reached the limits of our computational method with this size of the Hilbert space, we cannot increase the coupling parameter further. From Fig.~\ref{fig3} it can be seen that the total number of vibrations for $g_W=2$ is close to the limiting cut-off. At the same time, however, this is a strong argument for employing a quantum simulator, such as proposed in this letter, to validate the predictions of the semi-classical approach.
\begin{figure}
\includegraphics[scale=2]{Figure3.eps}
\caption{Exact evolution of the effective width of the spatial density profile governed by the Hamiltonian \eqref{Hamiltonian_final} for the initial state $|\boldsymbol{\Psi}_0\rangle$. The bottom row shows the total number of vibrations $\hat{\cal N}_\mathtt{vib}$ created in the system during the evolution. Consecutive columns correspond to different local couplings $g_W=\{0.1,0.75,1.5\}$. All calculations performed for $g_J=0$ and $\omega_0=3$. Note that for stronger interactions evident stabilization of the excitation density profile, along with increasing number of created vibrations, is observed. \label{fig3}}
\end{figure}
\section{Experimental parameters} The numerical predictions for the model described by the Hamiltonian \eqref{Hamiltonian_final} are quite general. For a quantum simulator we consider $^{87}$Rb atoms confined in an optical lattice determined by lattice spacing $R_0=1\,\mathrm{\mu m}$ and $V_0=100 E_R$ (recoil energy $E_R=2\pi^2\hbar^2/mR_0^2$) \cite{Pohl2014}, \textit{i.e.}, the local trap frequency is equal to $6.2\,\mathrm{kHz}$. We assume Rydberg states with principal quantum number $n=50$ for which $C_3^{sp} = 3.224\,\mathrm{GHz\,\mu m^3}$ \cite{Buchler2016}. We choose the dressing parameters as $\alpha = 0.015$ and $\varDelta/2=\varDelta_s = \varDelta_p = 2.5\,\mathrm{GHz}$.
With these values, the system mimics the Hamiltonian \eqref{Hamiltonian_final} with dimensionless parameters $\omega_0 = 4.7$, $g_W = 5.6$, and $g_J = 5.6$. These parameters can be easily tuned since they strongly depend on the lattice spacing $R_0$ and on the set of laser detunings. In this way, a large and interesting area of the phase diagrams presented in Fig.~\ref{fig2} can be covered.
The estimated lifetime of Rydberg atoms excited to states with $n=50$ is $\tau_S = 65\,\mathrm{\mu s}$ and $\tau_P = 86\,\mathrm{\mu s}$ \cite{Beterov2009}. The effective lifetime of a Rydberg dressed state is scaled by a factor $\alpha^{-2}$ and is sufficiently long to observe the non-spreading excitation behavior. It is worth noting that also other experimental realizations, based for example on Rydberg microtrap arrays \cite{Leung2011}, can be considered as proper candidates for simulating this system.
\section*{Summary} We show that a system of ultracold Rydberg atoms confined in a one-dimensional optical lattice may serve as a dedicated quantum simulator for excitation-vibration dynamics, which is a subclass of polaron dynamics. Since effective parameters of the resulting model can be easily tuned, the system can be used to mimic transport of excitation in biologically active proteins and to perform full quantum mechanical tests of the semi-classical predictions.
The proposed scheme may serve as a platform to investigate the HSSH bi- and many-polaron system \cite{Berciu2017a,Berciu2017b}. In particular, the character of the bi-polaron interactions can be tuned from repulsive to attractive by the experimental control parameters $g_W$ and $g_J$. Finally, we note that also disorder effects in the HSSH Hamiltonian \cite{Berciu2009} can be studied by introducing incommensurate optical lattices.
\bibliographystyle{abbrv}
|
\section{Introduction}\label{sec:intro}
For a graph $G$, let $|G| = |V(G)|$, $\| G \| = |E(G)|$, and $\delta(G)$ be the minimum degree of a vertex in $G$.
For a positive integer $k$, define $H_{k}(G)$ to be the subset of vertices with degree at least $2k$ and $L_{k}(G)$ to be the subset of vertices of degree at most $2k - 2$.
Two graphs are \emph{disjoint} if they have no common vertices.
Every graph with minimum degree at least $2$ contains a cycle. The following seminal result of Corr\'{a}di and Hajnal \cite{C-H} generalizes this fact.
\begin{thm}\cite{C-H}\label{thm:C-H}
Let $G$ be a graph and $k$ a positive integer. If $|G| \geq 3k$ and $\delta(G) \geq 2k$, then $G$ contains $k$ disjoint cycles.
\end{thm}
Both conditions in Theorem~\ref{thm:C-H} are sharp. The condition $|G| \geq 3k$ is necessary as every cycle contains at least $3$ vertices.
Further, there are infinitely many graphs that satisfy $|G| \geq 3k$ and $\delta(G) = 2k - 1$, but contain at most $k-1$ disjoint cycles.
For example, for any $n\ge 3k$, let $G_n=K_{n}-E(K_{n-2k+1})$ where $K_{n-2k+1}\subseteq K_n$.
The Corr\'{a}di-Hajnal Theorem inspired several results related to the existence of disjoint cycles in a graph
(e.g.~\cite{D-E, Di, HSz, Enomoto, Wang, KK-Ore, CFKS, KK-refCH, K-K-Y, KKMY}).
This paper focuses on the following theorem of Dirac and Erd\H{o}s~\cite{D-E}, one of the first attempts to generalize Theorem~\ref{thm:C-H}.
\begin{thm}\cite{D-E}\label{thm:D-E}
Let $k \geq 3$ be an integer and $G$ be a graph with $|H_{k}(G)| - |L_{k}(G)| \geq k^{2} + 2k - 4$. Then $G$ contains $k$ disjoint cycles.
\end{thm}
Dirac and Erd\H{o}s suggested that the bound $k^2+2k-4$ is not best possible and also
constructed an infinite sequence of graphs $G_k(n)$ with $h_{k}(G_k(n)) - \ell_{k}(G_k(n)) = 2k-1$
such that $G_k(n)$ does not have $k$ disjoint cycles.
They did not explicitly pose problems, and
it seems that Erd\H{o}s regretted not doing so, as later in \cite{Erdos} he remarked (about \cite{D-E}):
``This paper was perhaps undeservedly neglected; one reason was that we have few easily quotable theorems there, and do not state any unsolved problems.''
Here we consider questions that are implicit in \cite{D-E}.
For small graphs, the bound of $|H_k(G)|-|L_k(G)|\ge2k$
is not sufficient to guarantee the existence of $k$ disjoint cycles. Indeed, $K_{3k-1}$ contains at most $k - 1$ disjoint cycles,
so for small graphs, a bound of at least $3k$ is necessary. The authors~\cite{KKM} recently proved that $3k$ is also sufficient.
\begin{thm}\cite{KKM}\label{thm:3k}
Let $k \geq 2$ be an integer and $G$ be a graph with $|H_k (G)| - |L_{k}(G)| \geq 3k$. Then $G$ contains $k$ disjoint cycles.
\end{thm}
There exist graphs $G$ with at least $3k$ vertices and $|H_{k}(G)| - |L_{k}(G)| \geq 2k$ that do not contain $k$ disjoint cycles.
For example, consider the graph $G_0(k)$ obtained from $K_{3k-1}$ by selecting a subset $S \subseteq V(K_{3k-1})$ with $|S| = k$, removing all edges in $G[ S ]$,
adding an extra vertex $x$ and the edges from $x$ to each vertex in $S$. Then $|H_{k}(G_0(k))| - |L_{k}(G_0(k))| = 3k - 2$ and $|G_0(k)| = 3k$,
but $x$ is not in a triangle, so $G_0(k)$ contains at most $k-1$ disjoint cycles.
In~\cite{KKM}, the authors describe another graph $G_1(k)$, obtained from $G_0(k)$ by adding $k$ vertices of degree $1$, each adjacent to $x$.
The graph $G_1(k)$ still contains only $k-1$ disjoint cycles, but has $4k$ vertices and $|H_{k}(G_1(k))| - |L_{k}(G_1(k))| = 2k$.
However, in the special case that $G$ is planar, it is shown in~\cite{KKM} that the bound of $2k$ is sufficient.
\begin{thm}\cite{KKM}\label{thm:planar}
Let $k \geq 2$ be an integer and $G$ be a planar graph. If
\[ |H_{k}(G)| - |L_{k}(G)| \geq 2k,\]
then $G$ contains $k$ disjoint cycles.
\end{thm}
Further, when $k \geq 3$, a bound of $2k$ is also sufficient for graphs with no two disjoint triangles.
\begin{thm}\cite{KKM}\label{thm:1tri}
Let $k \geq 3$ be an integer and $G$ be a graph such that $G$ does not contain two disjoint triangles. If
\[ |H_{k}(G)| - |L_{k}(G)| \geq 2k,\]
then $G$ contains $k$ disjoint cycles.
\end{thm}
In general, the bound of $2k$ is the best we may hope for, as witnessed by $K_{n - 2k + 1, 2k-1}$ for $n \geq 4k$.
Further, the graph $G_1(k)$ described above shows that a difference of $2k$ is not sufficient when $|G|$ is small.
In~\cite{KKM}, we were not able
to determine whether for each $k$ there are only finitely many such examples. In order to attract attention to this
problem and based on known examples, we raised the following question.
\begin{question}\cite{KKM}\label{q:main}
Is it true that every graph $G$ with $|G|\geq 4k+1$ and $|H_k(G)| - |L_k(G)| \geq 2k$ has $k$ disjoint cycles?
\end{question}
The goal of this paper is to confirm that indeed for every $k \geq 2$,
there are only finitely many graphs $G$ with $h_{k}(G) - \ell_{k}(G) \geq 2k$ but no $k$ disjoint cycles.
We do this by answering Question~\ref{q:main} for graphs with at least $19k$ vertices.
\begin{thm}\label{thm:2k}
Let $k \geq 2$ be an integer and $G$ be a graph with $|G| \geq 19k$ and
\[ |H_{k} (G)| - |L_{k}(G)| \geq 2k.\]
Then $G$ contains $k$ disjoint cycles.
\end{thm}
The remainder of this paper is organized as follows. The next two sections outline notation and previous results that will be used in the proof of Theorem~\ref{thm:2k}.
We also introduce Theorem~\ref{thm:2k induct}, which is a more technical version of Theorem~\ref{thm:2k}.
Theorem~\ref{thm:2k induct} is proved in Section~\ref{sec:proof}.
The proof builds on the techniques of Dirac and Erd\H{o}s \cite{D-E} and uses Theorem~\ref{thm:3k} as the base case for our induction.
\section{Notation}\label{sec:notation}
We mostly use the standard notation. For a graph $G$ and $x\in V(G)$, $N_{G}(x)$ is the set of all vertices adjacent to $x$ in $G$, and the \emph{degree} of $x$, denoted $d_{G}(x)$, is $|N_{G}(x)|$.
When the choice of $G$ is clear, we simplify the notation to $N(x)$ and $d(x)$, respectively.
The complement of a graph $G$ is denoted by $\overline{G}$.
For an edge $xy\in E(G)$, $G\diagup xy$ denotes the graph obtained from $G$ by contracting $xy$; the new vertex is denoted by
$v_{xy}$.
For disjoint sets $U, U' \subseteq V(G)$, we write $\| U, U' \|_{G}$ for the number of edges from $U$ to $U'$. When the choice of $G$ is clear, we will write $\| U, U' \|$ instead.
If $U = \{ u \}$, then we will write $\| u, U' \|$ instead of $\| \{ u \}, U' \|$.
The \emph{join} $G \vee G'$ of two graphs is $G \cup G' \cup \{xx' : x \in V(G)~\mbox{and}~x' \in V(G')\}$.
Let $SK_{m}$ denote the graph obtained by subdividing one edge of the complete $m$-vertex graph $K_{m}$.
Given an integer $k$, we say a vertex in $H_{k}(G)$ is \emph{high}, and set $h_{k}(G)=|H_{k}(G)|$. A vertex in $L_{k}(G)$ is \emph{low}. Set $\ell_{k}(G)=|L_{k}(G)|$. A vertex $v$ is in $V^{i}(G)$ if $d_{G}(v) = i$. Similarly, $v \in V^{\leq i}(G)$ if $d_{G}(v) \leq i$ and $v \in V^{\geq i}(G)$ if $d_{G}(v) \geq i$.
In these terms, $H_k(G)=V^{\ge2k}(G)$ and $L_k(G)=V^{\le2k-2}(G)$.
We say that $x, y, z \in V(G)$ \emph{form a triangle} $T = xyzx$ in $G$ if $G[\{x,y,z\}]$ is a triangle.
If $v \in \{x, y, z \}$, then we say $v \in T$.
A \emph{set $\mathcal{T}$ of disjoint triangles} is a set of subgraphs of $G$ such that each subgraph is a triangle and all the triangles are disjoint.
For a set $\mathcal S$ of graphs, let $ \bigcup \mathcal S = \bigcup \{V(S):S \in \mathcal S\}$.
For a graph $G$, let $c(G)$ be the maximum number of disjoint cycles in $G$ and $t(G)$ be the maximum number of disjoint triangles in $G$.
When the graph $G$ and integer $k$ are clear from the context, we use $H$ and $L$ for $H_{k}(G)$ and $L_{k}(G)$, respectively.
The sizes of $H$ and $L$ will be denoted by $h$ and $\ell$, respectively.
\section{Preliminaries}\label{sec:pre}
As shown in \cite{K-K-Y}, if a graph $G$ with $|G| \geq 3k$ and $\delta(G) \geq 2k - 1$ does not contain a large independent set, then with two exceptions,
$G$ contains $k$ disjoint cycles:
\begin{thm}\cite{K-K-Y}\label{thm:K-K-Y}
Let $k \geq 2$. Let $G$ be a graph with $|G| \geq 3k$ and $\delta(G) \geq 2k -1$ such that $G$ does not contain $k$ disjoint cycles. Then
\begin{enumerate}
\item $G$ contains an independent set of size at least $|G| - 2k + 1$, or
\item $k$ is odd and $G = 2K_{k} \vee \overline{K_{k}}$, or
\item $k =2$ and $G$ is a wheel.
\end{enumerate}
\end{thm}
The theorem gives the following corollary.
\begin{cor}\label{cor:K-K-Y}
Let $k \geq 2$ be an integer and $G$ be a graph with $|G| \geq 3k$. If $h \geq 2k$ and $\delta(G) \geq 2k - 1$ (i.e. $L = \emptyset$),
then $G$ contains $k$ disjoint cycles.
\end{cor}
This corollary, along with the following theorem from \cite{KKM} will be used in the proof.
\begin{thm}\cite{KKM}
\label{thm:2k+t} Let $k\geq2$ be an integer and $G$ be a graph such that $|G| \geq 3k$. If
\[
h - \ell \geq 2k + t(G),
\]
then $G$ contains $k$ disjoint cycles.
\end{thm}
We prove the following technical statement that implies Theorem~\ref{thm:2k}, but is more amenable to induction.
\begin{thm}
\label{thm:2k induct} Suppose $i,k\in\mathbb Z$, $k\geq i$ and $k\geq 2$. Let $\alpha = 16$ be a constant. If $G$ is a graph with $|G| \geq \alpha k + 3i$ and $h \geq \ell + 3k - i$,
then $c(G)\geq k$.
\end{thm}
Theorem~\ref{thm:2k} is the special case of Theorem~\ref{thm:2k induct} for $i = k$.
The heart of this paper will be a proof of Theorem~\ref{thm:2k induct}. In the remainder of this section we organize the induction and
establish some preliminary results.
We argue by induction on $i$. The base case $i \leq 0$ follows from Theorem~\ref{thm:3k}. Now suppose $i\ge 1$. The equations $|G| \geq h + \ell$ and $h - \ell \geq 2k$ give
\begin{equation}
\label{eq:less than half}
\ell \leq \frac{|G|}{2} - k.
\end{equation}
The \emph{2-core} of a graph $G$ is the largest subgraph $G'\subseteq G$ with $\delta(G') \geq 2$. It can be obtained from $G$ by iterative deletion of vertices of degree at most $1$. The following Lemma was proved in \cite{KKM}.
\begin{lem}\cite{KKM}\label{lemma:2-core}
Suppose the $2$-core of $G$ contains at least $6$ vertices
and is not isomorphic to $SK_{5}$. If $h_{2}(G) - \ell_{2}(G) \geq 4$ then $c(G)\geq2$.
\end{lem}
Now, we prove a result regarding minimal counterexamples to Theorem~\ref{thm:2k induct}. Call a triangle $T$ \emph{good} if $T \cap L_{k}(G) \neq \emptyset$.
\begin{lem}
\label{lem:minimal}Suppose $i,k\in\mathbb Z$, $k\geq i$ and $k\geq 2$. Let $\alpha = 16$. If a graph $G$ satisfies all of:
\begin{enumerate}[ (a)]
\item $|G| \geq \alpha k + 3i$,
\item $h \geq \ell + 3k - i$,
\item $c(G) < k$, and
\item subject to (a\textendash c), $\sigma := ( k, i, |G|+\left\Vert G\right\Vert )$ is lexicographically minimum,
\end{enumerate}
then all of the following hold:
\begin{enumerate}[(i)]
\item $G$ has no isolated vertices; \label{minimal 1}
\item $k\geq3$; \label{minimal 2}
\item $L(G)\cup V^{\geq 2k+1}(G)$ is independent; \label{minimal 3}
\item if $x\in L(G)$, $d(x)\geq2$, and $xy\in E$, then $xy$ is in a triangle; and \label{minimal 4}
\item if $\mathcal{T}$ is a set of disjoint good triangles in $G$ and $X:=\bigcup\mathcal{T}$, then $\left\Vert v ,X\right\Vert \geq2|\mathcal{T}|+1$
for at least two vertices $v\in V\smallsetminus X$. \label{minimal 5}
\end{enumerate}
\end{lem}
\begin{proof}
Assume (a--d) hold. Using Theorem~\ref{thm:3k}, (a--c) imply $i \geq 1$; so the minimum in (d) is well defined.
If (\ref{minimal 1}) fails, then let $v$ be an isolated vertex in $G$. Now $G': = G-v$ and $i': = i -1$ satisfy conditions (a\textendash c), contradicting (d).
Hence,~(\ref{minimal 1}) holds.
For (\ref{minimal 2}), suppose $k=2$. Then $t(G)\le c(G) \leq 1$. If $i =1$ then $h - \ell \geq 3k - i \geq 2k + t(G)$, so $c(G) \geq 2$ by Theorem~\ref{thm:2k+t}.
Thus $i = 2$ and $h - \ell = 4$. Using~\eqref{eq:less than half} and~(\ref{minimal 1}),
{\allowdisplaybreaks
\begin{align}
\nonumber \| G \| &\geq \frac{1}{2} ( \ell + 3( |G| - \ell ) + h ) = \frac{1}{2} ( 3|G| + h - 2\ell ) \\
\nonumber &= \frac{1}{2} ( 3|G| - \ell + 4 ) \geq \frac{1}{2} \left( 3|G| - \left( \frac{|G|}{2} - 2 \right) + 4 \right) \\
\label{j20} &= |G| + \frac{ |G| }{4} + 3
\geq |G| + \frac{\alpha}{2} + \frac{3i}{4} + 3 = |G| + \frac{\alpha}{2} + \frac{9}{2}.
\end{align}
}If $G'$ is the $2$-core of $G$, then $\| G' \| - |G'| \ge \| G \| - |G|$. Since $\alpha > 1$,~\eqref{j20} yields $\|G' \| > |G'| + 5$; so $|G'| > 5$ and $G' \not\cong SK_{5}$.
By Lemma~\ref{lemma:2-core}, $c(G)\ge 2$, contradicting (c).
For (\ref{minimal 3}), suppose $e \in E(G[L\cup V^{\geq 2k+1}(G)])$, and set $G':=G - e$. Since $G'$ is a spanning subgraph of $G$, it satisfies (a) and (c).
Moreover, by the definition of $G'$, $h_k(G')=h$ and $ \ell_k(G')=\ell$, so (b) holds for $G'$, which means
(d) fails for $G$.
If (\ref{minimal 4}) fails, then let $G' = G \diagup xy$ and $i' = i - 1$. Since $d_{G'} (v_{xy}) \geq d(y)$ and the degrees of all other vertices in $G'$ are unchanged,
$G'$ and $i'$ satisfy (a\textendash c), contradicting (d).
Finally, suppose (\ref{minimal 5}) fails, and let $u\in V\smallsetminus X$ with
$\left\Vert u,X\right\Vert $ maximum. Then $\left\Vert v,X\right\Vert \leq2|\mathcal{T}|$
for all $v\in V\smallsetminus(X+u)$. Set $G'=G-X$,
$k'=k-|\mathcal{T}|$, and $i' = i - |\mathcal{T}| \leq k'$.
Then $H\cap V(G')-u\subseteq H_{k'}(G')$ and $L_{k'}(G')-u\subseteq L\cap V(G')$.
Since $\alpha \geq 3$, we have $|G'| \geq \alpha k' + 3i$; so $G'$ satisfies (a). Let $\beta_{1}=1$ if $u\in H\smallsetminus H_{k'}(G')$;
else $\beta_{1}=0$. Let $\beta_{2}=1$ if $u\in L_{k'}(G)\smallsetminus L$;
else $\beta_{2}=0$. Then $\beta_{1}+\beta_{2}\leq|\mathcal{T}|$ and so
{\allowdisplaybreaks
\begin{align*}
h_{k'}(G') &\geq h - 2|\mathcal{T}| - \beta_{1}
\geq \ell + 3k - i - 2| \mathcal{T} | - \beta_{1}\geq ( \ell_{k'}(G')+ |\mathcal{T} |-\beta_2)+3k-i
-2| \mathcal{T} | - \beta_{1}\\
&=\ell_{k'}(G') - |\mathcal{T} | + 3(k'+|\mathcal{T} |) - (i'+ |\mathcal{T}|) - \beta_{1}-\beta_{2}
\geq \ell_{k'}(G') + 3k' - i'.
\end{align*}}
\noindent
This means $G'$ satisfies (b).
As $c(G')+|\mathcal{T}|\leq c(G)<k$, $c(G')<k'$. Thus $G'$ satisfies (c). If $k' \geq 2$, then this contradicts the choice of $k$ in (d), so (\ref{minimal 5}) holds.
Otherwise, $| \mathcal{T} | = k - 1$ and so $|X | = 3k -3$. Since each triangle in $\mathcal{T}$ has a low vertex,
$|L\cap X|\geq |\mathcal{T}|$, and by (iii), $d_G(x)\leq 2k$ for each $x\in X$.
Thus
\begin{equation}\label{10
\| X,V(G')\|<2k|X|<6k^2.
\end{equation}
By (b), $|H\cap V(G')|- |L\cap V(G')|\ge 3k -i-|H\cap X|+ |L\cap X|\ge 2k-i$.
So,
$$\sum_{v\in V(G')\cap (H\cup L)}d_G(v)\geq 2k|H\cap V(G')|\geq 2k\frac{|V(G')\cap(H\cup L)|+(2k-i)}{2}.
$$
By this and~\eqref{10}, we get
\begin{equation}\label{11}
2\|G'\|=\sum_{v\in V(G')}d_G(v)-\|X,V(G')\|\ge k(|G'|+2k-i)-\|X,V(G')\|\ge k(|G'|-4k-i).
\end{equation}
By (c), $c(G)\le k-1$, so $G'$ has no cycle. Thus by~\eqref{11},
\[2|G'|>2\|G'\|\ge k(|G'|-4k-i).\]
By (a), $|G'|\ge|G|-3k\ge(\alpha-3)k+3i=13k+3i$. Solving yields
\begin{align}\notag
k(4k+i)&>(k-2)|G'|\ge (k-2)(13k+3i)\\
26k&>9k^2+i(2k-6).\notag
\end{align}
As $i\ge 0$, and $k\ge3$ by \eqref{minimal 2}, this is a contradiction.
\end{proof}
\section{Proof of Theorem~\ref{thm:2k induct}}\label{sec:proof}
Fix $k$, $i$, and $G=(V,E)$ satisfying the hypotheses of Lemma~\ref{lem:minimal}. First choose
a set $\mathcal{S}$ of disjoint good triangles with $s:=|\mathcal{S}|$ maximum,
and put $S=\bigcup\mathcal{S}$.
Next choose a set $\mathcal{S}'$ of disjoint triangles,
each contained in $V^{\leq2k}(G)\smallsetminus S$, with $s':=|\mathcal{S}'|$
maximum, and put $S'=\bigcup\mathcal{S}'$.
Say $\mathcal{S}=\{T_{1},\dots,T_{s}\}$ and $\mathcal{S}'=\{T_{s+1},\dots,T_{s+s'}\}$.
Let $\mathcal{H}$ be the directed graph defined on vertex set $\mathcal{S}$ by
$CD\in E(\mathcal{H})$ if and only if there is $v\in C$ with $\left\Vert v,D\right\Vert =3$.
Here we allow graphs with no vertices.
A vertex $C'$ is \emph{reachable} from a vertex $C$ if
$\mathcal{H}$ contains a directed $CC'$-path.
\begin{fact}
\label{2}If $x\in L\smallsetminus S$ and $d(x)\geq2$ then $N(x)\subseteq S$.
\end{fact}
\begin{proof}
Suppose $y\in N(x)\smallsetminus S$. As $x$ is low, $x \notin S'$. By
Lemma~\ref{lem:minimal}(\ref{minimal 4}), $xy$ is in a triangle $xyzx$. As\emph{ $\mathcal{S}$ }is maximal, $z\in S$,
so $z\in C$ for some $C\in\mathcal{S}$. Let
\[ \mathcal{S}_{0}=\{C'\in\mathcal{S}:C \text{~is reachable from \ensuremath{C'} in}\,\mathcal{H}\}.\]
By Lemma~\ref{lem:minimal}(\ref{minimal 5}), there is $w\in(V\smallsetminus\bigcup\mathcal{S}_{0})-y$ with $\left\Vert w,\bigcup\mathcal{S}_{0}\right\Vert \geq2|\mathcal{S}_{0}|+1$.
Then $\left\Vert w,D\right\Vert =3$ for some $D\in\mathcal{S}_{0}$. By Lemma~\ref{lem:minimal}(\ref{minimal 3}), $w\ne x$.
Further, $w\notin S$ as otherwise the triangle in $\mathcal{S}$ containing $w$ is in $\mathcal{S}_{0}$, contradicting that $w \notin \bigcup \mathcal{S}_{0}$.
Let $D = C_{1}, \ldots, C_{j} = C$ be a $D,C$-path in $\mathcal{H}$, and for $i \in [j - 1]$ let $x_{i} \in C_{i}$ with $\| x_{i}, C_{i+1} \| = 3$. If $C_{1}' = C_{1} - x_{1} + w$, $C_{j}' = C_{j} - z + x_{j-1}$ and $C_{i}' = C_{i} - x_{i} + x_{i-1}$ for $i \in \{ 2, \ldots, j-1\}$, then $\left( \mathcal{S} \smallsetminus \bigcup_{i = 1}^{j} C_{i} \right) \cup \bigcup_{i =1}^{j} C_{i}' \cup \{xyzx\}$ is a set of $s+1$ disjoint good triangles.
This contradicts the maximality of $\mathcal{S}$.
\end{proof}
\begin{fact}\label{fact:2 leaves}
Each $v \in V$ is adjacent to at most $2$ leaves. Moreover, if $v$ is adjacent to $2$ leaves, then $v \in V^{2k}$.
\end{fact}
\begin{proof}
Let $v$ be adjacent to a leaf. By Lemma~\ref{lem:minimal}(\ref{minimal 3}), $v \in V^{2k-1} \cup V^{2k}$. Let $X$ be the set of leaves adjacent to $v$, and put $G'=G-X$.
Let $i' = i - (|X| - 1 - |\{ v \} \cap V^{2k}|)$. Observe
{\allowdisplaybreaks
\begin{align*}
h_{k} ( G' ) - \ell_{k} (G' ) &\geq (h - |\{ v \} \cap V^{2k}| ) - (\ell + 1 - |X| ) \\
&= h - \ell - |\{ v \} \cap V^{2k}| + |X| - 1 \\
&\geq 3k - i - |\{ v \} \cap V^{2k}| + |X| - 1 \\
&\geq 3k - i',
\end{align*}}
so (b) holds for $G', k$ and $i'$. Now, $|G'| \geq \alpha k + 3i - |X| = \alpha k + 3i' + 2|X| - 3( 1 + |\{v \} \cap V^{2k}|)$.
If $|X| \geq 3$, then $2|X| - 3( 1 + |\{ v \} \cap V^{2k}|) \geq 0$, so $|G'| \geq \alpha k + 3i'$ and (a) holds. As $i'$ is at most $i$ and $G' \subset G$, (d) does not hold for $G,k$, and $i$, a contradiction.
Similarly, if $v \in V^{2k-1}$ and $|X| = 2$, then $|G'| \geq \alpha k + 3i'$, so (a) still holds and $G', k$ and $i'$.
Thus this also contradicts (d) for $G$.
\end{proof}
Let $G_{1} = G - V^{1}$.
Let $H^1 = V^{\geq2k}(G_{1})$,
$R^1=V^{2k-1}(G_{1})$, $L^1 = L_{k}(G_{1}) \cap L$, and $M = L_{k}(G_{1})\smallsetminus L^1$.
Then $G_{1} = G[H^1\cup R^1 \cup M \cup L^1]$ and $V^{\geq2k-1}(G) = H^1 \cup R^1 \cup M$.
Since deleting a leaf does not decrease the difference $h- \ell$,
\begin{equation}
\label{eq:G1}
h_{k}(G_{1}) - \ell_{k} (G_{1}) \geq 3k - i.
\end{equation}
\begin{fact}\label{fact:new low in triangle}
If $x \in M$, then $x$ is in a triangle $xyzx$ in $G$ with $d(x), d(y), d(z) \leq 2k$.
\end{fact}
\begin{proof}
Suppose $x \in M$. By Fact~\ref{fact:2 leaves}, either (i) $x \in V^{2k-1}$ and is adjacent to one leaf or (ii) $x \in V^{2k}$ and is adjacent to two leaves. Thus $d(x)\le 2k$. We first claim:
\begin{equation}
\label{CF3.3}
\mbox{\em $x$ has a neighbor $y$ such that $2\leq d(y) \leq 2k$.}
\end{equation}
Suppose not. Let $X$ be the set consisting of $x$ and the leaves adjacent to $x$. For each vertex $v \not\in X$, $d_{G-X}(v) \geq d(v) - 1$, with equality if $v \in N(x)$. Moreover, if $v \in N(x)$, then $d_{G-X} (v) \geq 2k$. Therefore, $h_{k}( G - X ) = h - | \{ x \} \cap V^{2k}|$ and $\ell_{k}( G - X ) = \ell - ( |X| - 1 )$. So
\[ h_{k}(G - X ) - \ell_{k} (G - X) = h - \ell + 1 \geq 3k - ( i - 1 ) \]
and
$|G - X| \geq |G| - 3 \geq \alpha k + 3( i - 1 )$,
contradicting the minimality of $i$. So \eqref{CF3.3} holds.
Now, suppose $xy$ is not in a triangle. Let $G'$ be formed from $G$ by removing the leaves adjacent to $x$ and contracting $xy$. By Fact~\ref{fact:2 leaves}, $|G'| \geq |G| - 3$. Since $d(x) \geq 2k - 1$ and $x$ does not share neighbors with $y$, $d_{G'}(v_{xy}) \geq d(y)$. Similarly, $d_{G'}(v) = d(v)$ for all $v \in V(G') - v_{xy}$. Now, $h_{k}(G') - \ell_{k}(G') = h - \ell + 1 \geq 3k - (i-1)$, contradicting the choice of $i$.
Let $xyzx$ be a triangle containing $xy$. If $d(z) \leq 2k$, we are done.
Otherwise, let $G''$ be the graph obtained from $G$ by removing the leaves adjacent to $x$ and deleting the vertices $x, y$, and $z$. Observe $|G''| \geq |G| - 5 \geq \alpha (k-1) + 3(i - 1)$. If there exists a vertex $u \in H \setminus H_{k-1}(G'')$, then $N(u) \supseteq \{ x, y, z \}$, and $d(u) \leq 2k$, since $d(z) \geq 2k+ 1$. In this case $xyux$ is the desired triangle. Similarly, if $v \in L_{k-1}(G'') \setminus L$, then $xyvx$ is the desired triangle. Thus $h - h_{k-1}(G'') \leq 2 + |\{ x \} \cap V^{2k}|$ and
$\ell - \ell_{k-1}(G'') \geq 1 + |\{ x \} \cap V^{2k}|$. Now,
\[ h_{k-1}( G'') - \ell_{k-1}(G'') \geq h - \ell - 1 \geq 3k - i - 1 = 3(k-1) - (i - 2).\]
By the minimality of $G$, $c(G'') \geq k-1$. Hence $c(G) \geq k$, a contradiction. We conclude that $xyzx$ is a triangle with $d(x), d(y), d(z) \leq 2k$.
\end{proof}
\begin{fact}\label{fact:s plus s prime}
$s + s' \geq 1$.
\end{fact}
\begin{proof}
Suppose $s + s' = 0$. In this case, Fact~\ref{fact:new low in triangle} implies $M = \emptyset$: indeed, if $v \in M$, there exists a triangle $vuwv$ with $d(v), d(u), d(w) \leq 2k$, contradicting the choice of $\mathcal{S}'$. By Fact~\ref{2} and since $\mathcal{S} = \emptyset$, all vertices in $L$ have degree at most $1$. By Lemma~\ref{lem:minimal}(\ref{minimal 1}), all vertices in $L$ are leaves in $G$ and $L^1 = \emptyset$.
Now, for every $x \in H - H_{k}(G_{1})$, there is a leaf $y \in L - L_{k}(G_{1})$ such that $xy \in E(G)$. Hence,
\[ h_{k}(G_{1}) \geq h_{k}(G_{1}) - \ell_{k}(G_{1}) \geq h - \ell \geq 2k. \]
By~\eqref{eq:less than half} and since $\alpha \geq 4$, $|G_{1}| \geq |G| - \ell \geq |G|/2 + k \geq \alpha k / 2 + k\geq 3k$. Finally, $L_{k}(G_{1}) = L^1 \cup M = \emptyset$, so Corollary~\ref{cor:K-K-Y} implies $G_{1}$ (and also $G$) contains $k$ disjoint cycles.
\end{proof}
Let $G_{2} = G\smallsetminus(L\smallsetminus S)$.
By~\eqref{eq:less than half},
\begin{equation}
|G_{2}| \geq \frac{\alpha +2}{2}k +\frac{3i}{2}. \label{eq:B}
\end{equation}
\begin{proof}[Proof of Theorem~\ref{thm:2k induct}]
Put $s^{*}=\max\{1,s\}$. Let $\mathcal{S}^{*}=\{T_{1},\dots,T_{s^{*}}\}$; by
Fact~\ref{fact:s plus s prime}, $T_{s^{*}}$ exists. Put $S^{*}=\bigcup\mathcal{S}^{*}$.
Let $W=V(G_{2})\smallsetminus S^{*}$, $F=G[W]$ and $k'=k-s^{*}$. It suffices to
prove $c(F)\geq k'$. \\
\noindent \emph{Case 1:} $s^{*} = k-1$. Since $k \geq 3$, $s^{*} \geq 2$. Thus, $s = s^{*} = k-1$. By Fact~\ref{fact:2 leaves}, all vertices in $M$ have degree $2k-2$. Let $M' = M \cap W$ and $H' = H(G_{2}) \cap W$. Fact~\ref{2} implies that if $v \in W$, then $d_{G_{1}}(v) = d_{G_{2}}(v)$. Thus
\[ H' = H^1 \cap W \text{ and } L(G_{1})\cap W = L(G_{2}) \cap W.\]
Hence, by~\eqref{eq:G1},
{\allowdisplaybreaks
\begin{align}
2k \leq h(G_{1}) - \ell( G_{1} ) &\leq ( |H(G_{1}) \cap S| + |H'| ) - ( |L(G_{1}) \cap S| + |M \cap W| + |L^1 \smallsetminus S| ) \nonumber \\
&= (|H(G_{1}) \cap S| - |L(G_{1}) \cap S| ) + |H'| - |M'| - |L^1 \smallsetminus S| \label{eq:H minus M} \\
&\leq (k-1) + |H'| - |M'|. \nonumber
\end{align}
}
Here, the last inequality holds because $S$ contains $s = k-1$
low vertices and at most $2s = 2k - 2$ high vertices. Equation~\eqref{eq:H minus M} implies $|H'| - |M'| \geq k + 1$. Further, if $W$ does not contain a cycle, then
\begin{align}
\| W, S \|_{G_{2}} &\geq \sum_{v \in W} d_{G_{2}} (v) - 2( |W| - 1 ) \nonumber \\
&\geq ( (2k-1)|W| + |H'| - |M'| ) - 2(|W| - 1) \nonumber \\
&\geq ( (2k-1)|W| + k + 1 ) - 2(|W| - 1) \label{eq:edges upper} \\
&\geq (2k - 3 )|W| + k + 3. \nonumber
\end{align}
On the other hand, using Lemma~\ref{lem:minimal}\eqref{minimal 3},
\begin{equation}
\label{eq:edges lower}
\| W, S \|_{G_{2}} \leq \sum_{w \in S} (d_{G_{2}} (w) - 2) \leq ( k-1 )( 6k - 8).
\end{equation}
Therefore, combining~\eqref{eq:edges upper} and~\eqref{eq:edges lower}, $|W| \leq 3( k - 1) - \frac{4}{2k- 3}$. Since $|S| = 3(k - 1)$ and $|G_{2}| = |S| + |W|$, this contradicts \eqref{eq:B} when $\alpha \geq 10$. \\
\noindent \emph{Case 2: } $s^{*} \leq k - 2$. Consider a vertex $v$ in $V^{\leq 2k'-2}(F)$. Since every vertex in $F$ has degree at least $2k - 2$ in $G_{2}$, $v$ must be adjacent to at least $2s^{*}$ vertices in $S^{*}$. Further, every vertex in $S^{*}$ is adjacent to at most $2k - 2$ vertices outside of $S^{*}$. Therefore,
\begin{equation}
\label{eq:V2k-2edge}
2s^{*} |V^{\leq 2k'-2}(F) | \leq \| V^{\leq 2k'-2} (F), S^{*} \| \leq 3s^{*} (2k - 2),
\end{equation}
and so
\begin{equation}
\label{eq:V2k-2}
|V^{ \leq 2k'-2}(F)| \leq 3k - 3.
\end{equation}
Similarly, if $u \in V^{2k'-1}(F)$, then $u$ is adjacent to at least $2s^{*} - 1$ vertices in $S^{*}$. Moreover,
there are at most $3s^{*}(2k - 2) - \| V^{\leq 2k'-2}(F), S^{*} \|$
edges from $V^{2k-1}(F)$ to $S^{*}$. So,
\[(2s^{*} - 1) | V^{2k'-1}(F) | \leq \| V^{2k'-1} (F), S^{*} \| \leq 3s^{*} (2k - 2) - \| V^{\leq 2k'-2} (F), S^{*} \| ,\]
and, combining with~\eqref{eq:V2k-2edge} gives,
\begin{align}
|V^{2k'-1}| &\leq \frac{2s^{*}(3k - 3)}{2s^{*} - 1} - \frac{2s^{*} |V^{\leq 2k'-2}(F)|}{2s^{*} - 1} \nonumber \\
&= 3k - 3 + \frac{3k-3}{2s^{*} - 1} - \frac{2s^{*} |V^{\leq 2k'-2}(F)|}{2s^{*} - 1}. \label{eq:V2k-1}
\end{align}
Using~\eqref{eq:V2k-2} and~\eqref{eq:V2k-1}, we see that
\begin{align*}
h_{k'}(F) - \ell_{k'}(F) &= |W| - 2|V^{\leq 2k'-2}(F)| - |V^{2k'-1}(F)| \\
&\geq |W| - 2|V^{\leq 2k'-2}(F)| - \left(3k - 3 + \frac{3k-3}{2s^{*} - 1} - \frac{2s^{*} |V^{\leq 2k'-2}(F)|}{2s^{*} - 1}\right) \\
&= |W| - \frac{ (2s^{*} - 2 )|V^{\leq 2k'-2}(F)|}{2s^{*} - 1} - 3k + 3 - \frac{3k-3}{2s^{*} - 1} \\
&\geq |W| - \frac{ (2s^{*} - 2 )(3k - 3)}{2s^{*} - 1} - 3k + 3 - \frac{3k-3}{2s^{*} - 1} \\
&= |W| + \left( -( 3k - 3) + \frac{ 3k - 3}{2s^{*} - 1} \right) - 3k + 3 - \frac{3k-3}{2s^{*} - 1} \\
&= |W| - 6k + 6 \\
&\geq \left( \frac{\alpha +2}{2}k +\frac{3i}{2} - 3s^{*} \right)- 6k + 6 \\
&\geq \frac{\alpha +2}{2}k +\frac{3i}{2} - 9k + 6 + 3k'. \\
\end{align*}
When $\alpha \geq 16$, this is at least $3k'$ and $F$ contains $k'$ disjoint cycles by Theorem~\ref{thm:3k}.
\end{proof}
\section*{Acknowledgement}
The authors thank Jaehoon Kim for helpful comments.
\bibliographystyle{abbrv}
|
\section{Introduction}
The geometry of normed spaces, also called \emph{Minkowski geometry} (see \cite{thompson}), is a research field which presents a lot of interesting directions to explore. Recently, we are concerned with giving a still missing systematic study of the differential geometry referring to these spaces. Although this topic was already partially studied throughout the past decades by researchers as Busemann (cf. \cite{Bus3}) and Petty (cf. \cite{Pet}), the existing works on it are, in a certain way, dispersed through the literature. The first step towards a systematic treatment is given in \cite{Ba-Ma-Sho}, where various concepts of curvature for regular curves in normed planes are studied. The present work aims to present an approach to the geometry of surfaces in three-dimensional Minkowski spaces, and this effort can be seen from the perspectives of different areas, namely Minkowski geometry, classical differential geometry, and Finsler geometry. In view of Minkowski geometry, we are extending some concepts from Euclidean geometry and investigate their behavior, asking what properties the general case has, and by which properties the inner product subcase can be characterized. Our construction is a particular equiaffine immersion (as it is defined and investigated in \cite{nomizu}), and we may study whether this particular case has a special behavior. From the viewpoint of Finsler geometry, we are studying curvatures of Finsler manifolds whose geometry is induced by the Minkowski geometry of an ambient space. The geometry in the tangent spaces is therefore defined by the parallel central planar sections of a certain convex body (namely, the unit sphere of the considered norm) that they determine. \\
Let us concretely explain the initial idea behind our constructions. The approach to affine differential geometry given in the book \cite{nomizu} regards hypersurfaces immersed in an affine space $\mathbb{R}^{n+1}$ endowed with a \emph{transversal vector field} ``playing the role" of a normal vector field. Our idea in this paper is to endow a surface in a three-dimensional Minkowski space with the transversal vector field obtained via the Birkhoff orthogonality associated to the norm. As far as the authors know, this construction was originally studied by Biberstein \cite{biberstein}; but this approach was different to the one that we present here. As we will see, such transversal fields give birth to immersions which are \emph{equiaffine}, and this allows us to extend some concepts from classical differential geometry to normed spaces. As it is often the case when dealing with non-Euclidean geometries and, in particular, general Minkowski spaces, one is mainly confronted with two situations: a large variety of concepts obtained as very natural extensions yields also really different analogues of classical concepts, or it yields Minkowskian analogues which remain very similar to their Euclidean subcases. In the present paper, both variants of extensions will occur. \\
The paper is organized as follows. Sections \ref{basics} and \ref{affine} are devoted to background theory and notation from Minkowski geometry and affine differential geometry, respectively. The reader not familiar with some of these topics will find standard references. In particular, by endowing a regular curve with the normal vector field determined by Birkhoff orthogonality, we re-obtain the concept of \emph{circular curvature}, which is one of the curvature types studied for the planar situation in \cite{Ba-Ma-Sho}. This is an important concept for us, since it is central in the definition of an analogue to the \emph{normal curvature} of a surface. \\
The possibly most important part of this paper, Section \ref{birkgaussmap}, is devoted to the study of an analogue to the Gauss map of a surface. We define the principal curvatures and principal directions to be the eigenvalues and eigenvectors of the differential of this map, respectively. Also, we extend the notions of curvature lines, asymptotic directions, and umbilic points. In particular it is proven that a surface whose points are all umbilic must be a Minkowski sphere. In Section \ref{normal} an analogue of normal curvature is obtained by considering the plane curvatures of curves obtained as plane sections of the surface. The relations between the normal curvature and the principal curvatures are also described. Up to assuming a further hypothesis on the (Minkowski) curvature lines of the surface, we prove in Section \ref{rigid} that, similarly to the Euclidean subcase, if the Minkowski Gaussian curvature of a surface is constant, then this surface must be a Minkowski sphere. We also prove that the same holds if the mean curvature is constant, subject to the hypothesis that the Minkowski Gaussian curvature is positive.\\
This is the first of three papers devoted to the study of differential geometry of surfaces immersed in three-dimensional spaces endowed with a norm from several viewpoints. Hence this paper aims to build the ``main core" of the theory, and this is mostly done in Sections \ref{birkgaussmap} and \ref{normal}. Also, the reader will notice that the investigations made in this paper are driven by the inspiration from classical differential geometry, since most of our results are analogues to the ones of this area subject. In \cite{diffgeom2} we adopt the viewpoint of affine differential geometry, which is also very natural, since our Gauss map is constructed using a central idea of this field. The third paper \cite{diffgeom3} deals with further topics in the theory, such as minimal surfaces and metric issues. \\
\section{Notation and basic concepts}\label{basics}
A \emph{normed} (=\emph{Minkowski}) \emph{space} is a finite dimensional vector space endowed with a usual norm. (Thus, to avoid confusion, we note that we are not concerned with the so-called Minkowskian space-time geometry.) For the sake of simplicity, we assume here that the vector space is $\mathbb{R}^{n+1}$, and we will denote the norm by $||\cdot||$. The \emph{unit ball} and the \emph{unit sphere} of a normed space are, respectively, the sets $B = \{v \in \mathbb{R}^{n+1}:||v||\leq 1\}$ and $\partial B=\{v \in \mathbb{R}^{n+1}:||v|| = 1\}$, respectively. Thus, $B$ is a \emph{centered convex body}, i.e., a compact, convex set with non-empty interior centered at the origin. Since we are dealing with aspects of differential geometry, throughout the text we will assume that the norm is \emph{strictly convex} (meaning that the triangle inequality is strict for linearly independent vectors or, equivalently, that the unit sphere does not contain straight line segments) and \emph{smooth}, in the sense that the unit sphere is locally the graph of a $C^{\infty}$ map $f:U\subseteq\mathbb{R}^n\rightarrow\mathbb{R}$ (the $C^{\infty}$ hypothesis can, however, often be relaxed to $C^2$ or to the smallest regularity class such that all involved derivatives make sense). \\
The homothets of the unit ball are called \emph{Minkowski balls}, and their boundaries are the \emph{Minkowski spheres}. Given a non-zero vector $v \in \mathbb{R}^{n+1}$ and a hyperplane $H \subseteq \mathbb{R}^{n+1}$, we say that $v$ is \emph{Birkhoff orthogonal} (or simply \emph{orthogonal}) to $H$ if the unit ball is supported by $H$ at $v/||v||$ (see Figure \ref{birkhoff}). We denote this relation by $v \dashv_B H$. From the hypothesis of smoothness and strict convexity it follows that Birkhoff orthogonality is unique, both at left and at right (as a relation between directions and planes). \\
\begin{figure}[h]
\centering
\includegraphics{birkhoff.eps}
\caption{$v \dashv_B H$.}
\label{birkhoff}
\end{figure}
Notice that Birkhoff orthogonality can be naturally extended to a relation between vectors: given two non-zero vectors $v, w \in \mathbb{R}^{n+1}$, we say that $v$ is Birkhoff orthogonal to $w$ (denoted by $v \dashv_B w$) if the hyperplane $H$ to which $v$ is Birkhoff orthogonal contains $w$. If the space is two-dimensional, then Birkhoff orthogonality is essentially a (homogeneous) relation between vectors, since in this case the hyperplanes are lines (see \cite{alonso} for more about orthogonality relations in normed spaces). In spaces of dimensions at least 3, Birkhoff orthogonality is a symmetric relation if and only if the norm is Euclidean. In the two-dimensional case, Birkhoff orthogonality is symmetric if and only if the unit circle is a \emph{Radon curve}, yielding the notion of \emph{Radon plane}. Properties and constructions of Radon curves are presented in \cite{martiniantinorms} and \cite{radonbalestro}. Standard references with respect to Minkowski geometry as a whole are \cite{thompson}, \cite{martini2}, and \cite{martini1}. Regarding differential geometry in Minkowski spaces we refer the reader to \cite{keti} and to the papers \cite{Bus3} and \cite{Pet} already mentioned in the introduction. The paper \cite{keti} deals with global theorems for curves in Minkowski spaces (such as the Fenchel and the Fary-Milnor theorem). \\
Now we will introduce a concept of curvature for plane curves that we will use later. For more information on curvature concepts of curves in two-dimensional Minkowski spaces (i.e., in normed planes) we refer to \cite{Ba-Ma-Sho}. An immersed hypersurface in a normed plane is a regular curve $\gamma:[0,c]\rightarrow (\mathbb{R}^2,||\cdot||)$ which, for our purpose, we may assume to be parametrized by arc-length. Up to choice of orientation, the unit transversal vector field to $\gamma$ given by Birkhoff orthogonality is the vector field $\eta:[0,c]\rightarrow \partial B$ such that $\eta(s) \dashv_B \gamma(s)$ and $[\eta(s),\gamma'(s)] >0$ for each $s \in [0,c]$. Since $\eta(s) \in \partial B$, there exists a function $t(s):[0,c]\rightarrow[0,l(\partial B)]$ such that $\eta(s) = \varphi(t(s))$ for each $s \in [0,c]$. Consequently, we have that
\begin{align*} \frac{d\varphi}{dt}(t(s)) = \gamma'(s), \ \ s \in [0,c].
\end{align*}
In \cite{Ba-Ma-Sho} the number $k(s):=t'(s)$ is called the \emph{circular curvature} of $\gamma$ at $p = \gamma(s)$. It is a natural extension of the usual curvature in the Euclidean plane and, when it does not vanish, it can also be regarded as the inverse of the radius of the osculating (Minkowski) circle of $\gamma$ at $p$.
\section{Affine immersions and transversal vector fields} \label{affine}
As explained in the introducion, our motivation to approach differential geometry in normed spaces is inspired mainly by affine differential geometry, and for this theory we refer to the book \cite{nomizu}. In this short section we briefly explain the basic ideas and concepts; for proofs the reader can consult the mentioned book. \\
Assume that $\mathbb{R}^{n+1}$ is the usual $(n+1)$-dimensional Euclidean space endowed with the standard connection $D:C^{\infty}(T\mathbb{R}^{n+1})\times C^{\infty}(T\mathbb{R}^{n+1})\rightarrow C^{\infty}(T\mathbb{R}^{n+1})$, and let $M$ be a $n$-dimensional manifold together with an immersion $f:M\rightarrow\mathbb{R}^{n+1}$. We call $f$ a \emph{hypersurface immersion} of $\mathbb{R}^{n+1}$. In classical differential geometry, one usually uses, for $x \in M$, the decomposition
\begin{align*} T_{f(x)}\mathbb{R}^{n+1} = f_{*}(T_xM) + \mathrm{span}\{\eta_x\},
\end{align*}
where $f_{*}$ denotes the usual push-forward differential map, and $\eta_x$ is a unit normal vector to $f_{*}(T_xM)$. By doing so, one can obtain the Levi-Civita connection and the second fundamental form associated to the metric induced on $M$ by the Euclidean metric in $\mathbb{R}^{n+1}$. \\
However, the point here is that in order to decompose $T_{f(x)}\mathbb{R}^{n+1}$, one does not need to regard the \emph{normal vector} to $f_{*}(T_xM)$, but any \emph{transversal vector} to it. Formally, given a point $x_0 \in M$, we consider a smooth transversal vector field $x \mapsto \xi_x$ defined in a neighborhood $U$ of $x_0$, meaning that $\xi_x \notin f_{*}(T_xM)$. We write then
\begin{align}\label{decomposition} T_{f(x)}\mathbb{R}^{n+1} = f_{*}(T_xM) + \mathrm{span}\{\xi_x\}.
\end{align}
The choice of such a transversal field induces implicitly a connection $\nabla:C^{\infty}(TU)\times C^{\infty}(TU)\rightarrow C^{\infty}(TU)$ by
\begin{align}\label{gauss} D_Xf_{*}(Y) = f_{*}(\nabla_XY) + h(X,Y)\xi.
\end{align}
It is easy to see that $\nabla$ is a torsion-free connection, meaning that $\nabla_XY-\nabla_YX-[X,Y]$ vanishes everywhere, and $h$ is a symmetric bilinear form in $T_xM$, for each $x \in U$. The connection $\nabla$ is called the \emph{induced connection}. Formula (\ref{gauss}) is known as the \emph{formula of Gauss}. The map $h$ is called the \emph{affine fundamental form}, and its rank is called the \emph{rank of the hypersurface}. This number does not depend on the transversal field considered in the hypersurface (this is proved in \cite{nomizu}). \\
The covariant derivative of $\xi$ can be decomposed in view of (\ref{decomposition}) as well. By doing so, we obtain the \emph{Weingarten formula}
\begin{align}\label{weingarten} D_X\xi = -f_{*}(SX) + \tau(X)\xi,
\end{align}
where $S:T_xM\rightarrow T_xM$ is a linear map called the \emph{shape operator}, and $\tau:T_xM\rightarrow\mathbb{R}$ is a $1$-form. As we will show later, this $1$-form will be of no importance for us, since every considered transversal field will be shown to have tangential derivative. Let $\mathrm{det}:\mathbb{R}^{n+1}\rightarrow\mathbb{R}$ denote the usual determinant in $\mathbb{R}^{n+1}$, and define a volume form on $M$ by
\begin{align*} \omega(X_1,...,X_n) = \mathrm{det}[X_1,...,X_n,\xi],
\end{align*}
where we are omitting the push-forward $f_{*}$ from the notation, for the sake of convenience. We call $\theta$ the \emph{induced volume element}. A volume form $\theta$ is said to be \emph{parallel} with respect to a connection $\nabla$ if $\nabla\theta = 0$. For the induced volume element $\omega$, we have the following
\begin{prop} We have $\nabla_X\omega = \tau(X)\omega$ for each $X \in T_xM$. Consequently, $\omega$ is parallel if and only if $D_X\xi$ is always tangential. In this case, we say that $(\nabla,\omega)$ is an \emph{equiaffine structure} on $M$.
\end{prop}
\begin{proof} See \cite[Proposition 1.4]{nomizu}.
\end{proof}
If $\tau = 0$, we say that $f$ is an \emph{equiaffine immersion}. The affine fundamental form $h$ also induces a volume form $\omega_h$ on $M$, where $\omega_h(X_1,...,X_n)$ is the square root of the absolute value of the determinant of the matrix $[h_{ij}]$, whose entries are the numbers $h(X_i,X_j)$, for $i,j = 1,...,n$. An equiaffine immersion for which $\omega = \omega_h$ is said to be a \emph{Blaschke immersion}, and in this case the transversal field $\xi$ is called the \emph{affine normal field}.
\section{The Birkhoff-Gauss map} \label{birkgaussmap}
Let $(\mathbb{R}^3,||\cdot||)$ be a normed space whose norm is smooth and strictly convex, and let $f:M\rightarrow(\mathbb{R}^3,||\cdot||)$ be a surface immersion. In what follows, we naturally identify the manifold $M$ with its image $f(M)$ and, consequently, the tangent space $T_pM$ with $f_*(T_pM) \subseteq T_{f(p)}\mathbb{R}^{3}$; therefore we will omit the push-forward map $f_*$ from our notation. For each $p \in M$, let $\eta(p) \in \partial B$ be a unit vector such that $\eta(p) \dashv_B T_pM$. The choice of such a transversal vector gives a local smooth unit vector field that, in our context, clearly plays the role of the usual normal vector field. We will call this vector field the \emph{Birkhoff normal vector field} of $M$. Of course, this local field can be regarded as global if and only if $M$ is orientable. As in the Euclidean subcase, the Birkhoff normal vector field can be seen as a map $\eta:U\subseteq M\rightarrow \partial B$, which we will call the \emph{Birkhoff-Gauss map}.
\begin{lemma} A surface immersion $f:M\rightarrow(\mathbb{R}^3,||\cdot||)$ endowed with the Birkhoff normal vector field is an equiaffine immersion.
\end{lemma}
\begin{proof} Let $p \in M$, and fix a vector $X \in T_pM$. Let $\gamma:(-\varepsilon,\varepsilon)\rightarrow M$ be any smooth curve such that $\gamma(0) = p$ and $\gamma'(0) = X$. Therefore, we have
\begin{align*} D_X\eta|_p = \frac{d}{dt}\eta\circ\gamma(t)|_{t=0}.
\end{align*}
On the other hand, $\eta\circ\gamma$ is a curve on $\partial B$, and hence the right hand term must be a vector in $T_{\eta(p)}\partial B$. Since $\eta(p) \dashv_B T_{\eta(p)}\partial B$, it follows that we have a natural identification $T_{\eta(p)}\partial B \simeq T_pM$. Then we have indeed $D_X\eta|_p \in T_pM$, and the covariant derivative of the transversal vector field is always tangential. This concludes the proof.
\end{proof}
In other words, the differential map $d\eta_p:T_pM\rightarrow T_{\eta(p)}\partial B$ can be regarded as a (linear) map of $T_pM$ onto itself. Notice also that $D_X\eta|_p = d\eta_p(X)$. In classical differential geometry, the differential of the Gauss map is used to define the curvatures of a surface at a point (see \cite{manfredo}). We will develop a similar theory for our context. We start by proving the key fact that each tangent space $T_pM$ admits a basis of eigenvectors of $d\eta_p$. We separate our approach into cases depending on the structure of the affine fundamental form. First, we assume that $h$ has rank $2$ and is definite. As we will see, for all the other cases we need an auxiliary Euclidean structure.
\begin{prop}\label{rank2} If the rank of the affine fundamental form $h$ equals $2$ at $p \in M$, and $h$ is definite, then the tangent space $T_pM$ admits a basis of eigenvectors of the differential map $d\eta_p:T_pM\rightarrow T_pM$.
\end{prop}
\begin{proof} Recall that the affine fundamental form $h$ is defined intrinsically by the Gauss formula
\begin{align*} D_XY = \nabla_XY + h(X,Y)\eta,
\end{align*}
where $X,Y \in T_pM$ and $\nabla$ is the induced connection. If $h$ has rank $2$, then we may choose vectors $X,Y \in T_pM$ such that $h(X,X) = h(Y,Y) \neq 0$ and $h(X,Y) = 0$. Taking smooth extensions of $X$ and $Y$ to local vector fields in a neighborhood $U$ of $p$, we may write
\begin{align*} D_X\eta = f_1X + f_2Y \ \ \mathrm{and} \\ D_Y\eta = g_1X + g_2Y,
\end{align*}
for some smooth functions $f_1,f_2,g_1,g_2:U\rightarrow \mathbb{R}$. Differentiating the equalities above yields
\begin{align*} D_YD_X\eta = Y(f_1)X + f_1D_YX + Y(f_2)Y + f_2D_YY \ \ \mathrm{and} \\
D_XD_Y\eta = X(g_1)X + g_1D_XX + X(g_2)Y + g_2D_XY.
\end{align*}
Now notice that since $h(X,Y) = 0$, it follows that, at $p$, $D_XY$ and $D_YX$ are tangential. Therefore, the transversal components of the two vectors above are given by $f_2h(Y,Y)\eta$ and $g_1h(X,X)\eta$, respectively. On the other hand, $D$ is a flat connection, and hence
\begin{align*} D_XD_Y\eta - D_YD_X\eta - D_{[X,Y]}\eta = 0.
\end{align*}
Since $D_{[X,Y]}\eta$ is tangential, it follows that $f_2h(Y,Y)\eta = g_1h(X,X)\eta$, and then $f_2 = g_1$ at $p$. This shows that the matrix of $d\eta_p$ written in the basis $\{X,Y\}$ is symmetric, and we have what we wished to prove.
\end{proof}
The rank of the affine fundamental form does not depend on the transversal vector field, as it becomes clear in \cite[Proposition 2.5]{nomizu}. This opens the possibility of working with an auxiliary Euclidean structure in $\mathbb{R}^3$. Let $\langle \cdot,\cdot \rangle$ denote the usual inner product, and denote by $||\cdot||_e$ the induced Euclidean metric. Let $B_e$ and $\partial B_e$ be the Euclidean unit ball and sphere, respectively. Denote by $u:\partial B_e \rightarrow \partial B$ a (smooth) map which carries each vector $v \in \partial B_e$ to a respective vector $u(v) \in \partial B$ at which the supporting hyperplane of $\partial B$ at $u(v)$ is the same supporting hyperplane of $\partial B_e$ at $v$. Notice that there are exactly two choices of such a smooth map $u$, and that this map can be regarded as the Birkhoff-Gauss map of the Euclidean circle in the geometry given by $||\cdot||$, or as the inverse of the Euclidean Gauss map of $\partial B$ as an immersed surface.\\
Writing $d\xi_pX = du^{-1}_{\eta(p)}\circ d\eta_pX$ and recalling that $du^{-1}_{\eta(p)}$ is self-adjoint (because it is the usual Gauss map of the Minkowski unit sphere), we have the following useful expressions for $h$:
\begin{align} \label{exph} h(X,Y) =-\frac{\langle Y,d\xi_pX\rangle}{\langle\eta,\xi \rangle}= -\frac{\langle du^{-1}_{\eta(p)}Y,d\eta_pX\rangle}{\langle \eta,\xi\rangle}.
\end{align}
These equalities come from evaluating the usual inner product of both sides of the Gauss formula (\ref{gauss}) with $\xi$. Notice that it follows from the self-adjointness of $du^{-1}_{\eta(p)}$ that $d\eta_p$ is self-adjoint with respect to $h$. In other words, we have $h(X,d\eta_pY) = h(d\eta_pX,Y)$ for any $X,Y \in T_pM$. In the ``language'' of affine differential geometry this is merely the Ricci equation for an equiaffine immersion (cf. \cite[Theorem 2.4]{nomizu}). \\
Notice also that $d\eta_p = du_{\xi(p)}\circ d\xi_p$. From the strict convexity of $B$ we have that the usual Gaussian curvature of $\partial B$ is non-negative, but may be zero at some isolated point. This may create some ``artificial" direction in the kernel of $d\eta_p$, in the sense that it is not in the kernel of $d\xi_p$. In other words, the Birkhoff-Gauss map can have a rank different to that of the Euclidean Gauss map. For that reason, we will restrict ourselves to the case where $du_v$ is an isomorphism for each $v \in \partial B_e$, which is exactly the same as saying that $\partial B$ has positive Euclidean Gaussian curvature at every point. Norms that give birth to unit spheres with such a property will be called \emph{admissible norms}. In what follows, \textbf{all norms are assumed to be admissible}.
\begin{prop}\label{rank2ind} If the rank of the affine fundamental form $h$ at $p \in M$ equals $2$, and $h$ is indefinite, then we also have that the tangent space $T_pM$ admits a basis of eigenvectors of the differential map $d\eta_p$.
\end{prop}
\begin{proof} The strategy is to use the map $u:\partial B_e\rightarrow \partial B$ of the unit sphere. If $\xi$ denotes the Euclidean Gauss map of $M$, then we have $\eta = u\circ\xi$. Therefore, $d\eta_p = du_{\xi(p)}\circ d\xi_p$. It follows that
\begin{align*} \mathrm{det}(d\eta_p) = \mathrm{det}\left(du_{\xi(p)}\right)\cdot\mathrm{det}(d\xi_p).
\end{align*}
Notice that $\mathrm{det}\left(du_{\xi(p)}\right) > 0$. Indeed, $u$ is the inverse of the usual Gauss map of the strictly convex surface $\partial B$. On the other hand, if $h$ has rank $2$ and is not definite, then we may choose vectors $X,Y \in T_pM$ such that $h(X,X) = -h(Y,Y) \neq 0$ and $h(X,Y) = 0$. From (\ref{exph}) it follows that $\langle X,d\xi_pX\rangle$ and $\langle Y,d\xi_pY\rangle$ have different signs, and $\langle X,d\xi_pY\rangle = \langle d\xi_pX,Y\rangle = 0$. Therefore, we have that $\mathrm{det}(d\xi_p) < 0$. Hence $\mathrm{det}(d\eta_p) < 0$, and this finishes the proof.
\end{proof}
The cases where the rank of $h$ is $0$ or $1$ are discussed next. These cases are somehow special because, as in the Euclidean subcase (where $h$ is the usual second fundamental form), when they occur we can guarantee the existence of a null eigenvalue for $d\eta_p$.
\begin{prop}\label{rank01} Let $f:M\rightarrow (\mathbb{R}^3,||\cdot||)$ be an immersed surface, with Birkhoff normal vector field $\eta$ and induced affine fundamental form $h$. If at $p \in M$ the rank of $h$ is $0$, then $d\eta_p = 0$. In the case that the rank of $h$ at $p\in M$ equals $1$, we have that $d\eta_p$ has two distinct eigenvectors, and one of them is associated to the eigenvalue $0$.
\end{prop}
\begin{proof} On the surface $M$, let $\eta$ denote the Birkhoff-Gauss map and $\xi$ denote the usual Euclidean Gauss map. Then we have $\eta = u \circ \xi$. Let $X,Y \in T_pM$, and denote
by these letters also respective extensions of these vectors to local vectors fields. The Gauss equation for the Euclidean normal map reads
\begin{align*} D_XY = \overline{\nabla}_XY + \overline{h}(X,Y)\xi,
\end{align*}
and hence we may write
\begin{align*} \overline{h}(X,Y) = \langle D_XY,\xi\rangle = -\langle Y,D_X\xi\rangle.
\end{align*}
If, at a point $p \in M$, the bilinear form $h$ has rank $0$, then so has $\overline{h}$. Therefore, in this case we have $\langle Y,d\xi_p(X)\rangle = 0$ for any $X,Y \in T_pM$. It follows that $d\xi_p = 0$. Since $\eta = u\circ\xi$, we have $d\eta_p = du_{\xi(p)}\circ d\xi_p$, and then $d\eta_p = 0$. \\
If the rank of $h$ equals $1$, then let $X$ be the null direction of $\overline{h}$. We have that $\langle Z,D_X\xi\rangle = 0$ for any $Z \in T_pM$, and hence $d\xi_p(X) = 0$. It follows immediately that $d\eta_p(X) = 0$. Since $d\eta_p$ is not null and $du_{\xi(p)}$ is invertible (recall that the norm is admissible), the existence of another eigenvector comes from standard linear algebra arguments.
\end{proof}
\begin{remark} The auxiliary Euclidean structure could have been used to prove also Proposition \ref{rank2}. However, the different method used there emphasizes that in this case the admissibility hypothesis for the norm is not necessary.
\end{remark}
\begin{definition} Let $f:M\rightarrow (\mathbb{R}^3,||\cdot||)$ be a surface immersion with Birkhoff normal vector field $\eta$. Let $\lambda_1,\lambda_2 \in \mathbb{R}$ be the eigenvectors of $d\eta_p$. Then these numbers are called the \emph{principal curvatures} of $M$ at $p$, and the respective eigenvectors $E_1,E_2 \in T_pM$ are called the \emph{principal directions} of $M$ at $p$. The numbers
\begin{align*} K = \lambda_1\lambda_2 \ \ \ \mathrm{and} \ \ \ H = \frac{\lambda_1+\lambda_2}{2}
\end{align*}
are called the \emph{Minkowski Gaussian curvature} and the \emph{Minkowski mean curvature} of $M$ at the point $p$, respectively.
\end{definition}
These definitions are immediate extensions of the Euclidean versions, and we will discuss some of their properties in order to understand them in a better way. First of all, it is easy to see that if we set $\lambda_1 \geq \lambda_2$, then the maps $\lambda_1,\lambda_2:M\rightarrow\mathbb{R}$ are smooth in all points, possibly except for the umbilic ones (since in these points $\lambda_1$ and $\lambda_2$ may ``change their roles''). Regarding the principal directions, it is known that in the Euclidean subcase they must be orthogonal. This is not the case in our context, but we can provide some information about them in terms of the affine fundamental form.
\begin{lemma}\label{princconj} Let $p \in M$ be a point whose principal curvatures are different, and let $V_1,V_2 \in T_pM$ be the principal directions. Then we have $h(V_1,V_2) = 0$.
\end{lemma}
\begin{proof} Let $\lambda_1(p),\lambda_2(p) \in \mathbb{R}$ be the principal curvatures of $M$ at $p$, and assume that $\lambda_1(p) > \lambda_2(p)$. By continuity, this inequality is true in a small neighborhood $U$ containing $p$, and then we may consider local vector fields $V_1,V_2\in C^{\infty}(TU)$ such that $V_1$ and $V_2$ are principal directions associated to the respective principal curvatures at each point $p \in U$. Therefore, we have $D_{V_1}\eta = \lambda_1V_1$ and $D_{V_2}\eta = \lambda_2V_2$ in $U$. Differentiating these expressions, we get
\begin{align*} D_{V_2}D_{V_1}\eta = V_2(\lambda_1)V_1 + \lambda_1D_{V_2}V_1 = V_2(\lambda_1) + \lambda_1\nabla_{V_2}V_1 + \lambda_1h(V_2,V_1)\eta \ \ \mathrm{and}\\
D_{V_1}D_{V_2}\eta = V_1(\lambda_2)V_2 + \lambda_2D_{V_1}V_2 = V_1(\lambda_2)V_2 + \lambda_2\nabla_{V_1}V_2 + \lambda_2h(V_1,V_2)\eta.
\end{align*}
Now, since $0 = D_{V_1}D_{V_2}\eta - D_{V_2}D_{V_1}\eta - D_{[V_1,V_2]}\eta$, and the vectors $V_1,V_2,\nabla_{V_1}V_2,\nabla_{V_2}V_1$ and $D_{[V_1,V_2]}\eta$ are all tangential, it follows that $(\lambda_1-\lambda_2)h(V_1,V_2)\eta = 0$. Since the principal curvatures at $p$ are distinct, the desired follows.
\end{proof}
In Euclidean differential geometry, we say that two directions given by non-zero vectors $X,Y \in T_pM$ are \emph{conjugate} if $\langle X,d\xi_pY\rangle = 0$ (or, equivalently, $\langle Y,d\xi_pX\rangle = 0$), where we recall that $\xi$ is the usual Euclidean Gauss map. However, this definition does not depend on the inner product. Indeed, it is easy to see that $X$ and $Y$ are conjugate if and only if $D_XY$ is tangential. Therefore, we may extend this definition to the Minkowski context, and the conjugate directions in the Minkowski norm will be precisely the same as the conjugate directions in the Euclidean norm. In particular, from the previous lemma it follows that the principal directions at any point are always conjugate directions. We summarize all this as follows.
\begin{lemma} Let $f:M\rightarrow(\mathbb{R}^3,||\cdot||)$ be an immersed surface with Birkhoff-Gauss map $\eta$ and usual Euclidean Gauss map $\xi$. For any non-zero vectors $X,Y \in T_pM$, the following statements are equivalent:\\
\noindent\emph{\textbf{(a)}} the vectors $X$ and $Y$ are conjugate directions in the Euclidean sense, \\
\noindent\emph{\textbf{(b)}} the derivative $D_XY$ is tangential, and \\
\noindent\emph{\textbf{(c)}} $h(X,Y) = 0$. \\
\end{lemma}
\begin{proof} From the equality $D_XY = \nabla_XY + h(X,Y)\eta$ we have that $D_XY$ is tangential if and only if $h(X,Y) = 0$. Now recall that from (\ref{exph}) we have
\begin{align*} h(X,Y) = \frac{\langle D_XY,\xi\rangle}{\langle \eta,\xi \rangle} = -\frac{\langle Y,D_X\xi\rangle}{\langle \eta,\xi\rangle}.
\end{align*}
Since the derivative $D_X\xi$ at a point $p \in M$ is precisely $d\xi_p(X)$, the proof is complete.
\end{proof}
Still in this direction, recall that positive Gaussian curvature has a sort of geometric consequence that can be regarded independently of the norm. Namely, if $M$ has positive Gaussian curvature at $p \in M$, then the normal vector to any curve on $M$ points at $p$ to the side of the tangent plane $T_pM$. This inspires the next proposition, which is important in Section 3 of \cite{diffgeom3}, for proving analogues to Hadamard theorems.
\begin{prop}\label{gaussposit} Assume, as usual, that $||\cdot||$ is an admissible norm, and let $f:M\rightarrow (\mathbb{R}^3,||\cdot||)$ be an immersed surface. Denote by $K$ and $K_e$ the Minkowski and the Euclidean Gaussian curvatures of $M$, respectively. For a point $p \in M$, the following statements are equivalent:\\
\noindent\emph{\textbf{(a)}} $K_e(p) > 0$,\\
\noindent\emph{\textbf{(b)}} $K(p) > 0$, and\\
\noindent\emph{\textbf{(c)}} $h$ is a (positive or negative) definite bilinear form at $p$.
\end{prop}
\begin{proof} The Euclidean second fundamental form is given by $(X,Y)\mapsto\langle Y,d\xi_pX\rangle$, and it is known that this bilinear form is definite if and only if $K_e(p) > 0$. Hence, from equality (\ref{exph}) it follows that \textbf{(a)}$\Leftrightarrow$\textbf{(c)}. \\
Assume that \textbf{(a)} holds, and let $V_1,V_2 \in T_pM$ be such that $d\eta_pV_1 = \lambda_1V_1$ and $d\eta_pV_2 = \lambda_2V_2$. We have to prove that $\lambda_1$ and $\lambda_2$ have the same sign. To do so, we notice that $d\xi_pV_1 = \lambda_1du_{\eta(p)}^{-1}V_1$, and the analogous equality holds for $V_2$. Therefore,
\begin{align*} \langle d\xi_pV_1,V_1\rangle = \lambda_1\langle du^{-1}_{\eta(p)}V_1,V_1\rangle \ \ \mathrm{and}\\
\langle d\xi_pV_2,V_2 \rangle = \lambda_2\langle du^{-1}_{\eta(p)}V_2,V_2\rangle.
\end{align*}
Since $K_e(p) > 0$, it follows that the left hand terms have the same sign. The same holds for the numbers $\langle du^{-1}_{\eta(p)}V_1,V_1 \rangle$ and $\langle du^{-1}_{\eta(p)}V_2,V_2\rangle$, due to the admissibility of the norm. Then $\lambda_1$ and $\lambda_2$ have the same sign, and this implies $K(p) > 0$. To prove \textbf{(b)}$\Rightarrow$\textbf{(a)}, the same argument works. Notice also that, in this case, the question whether $h$ is positive or negative only depends on the orientation chosen to $\eta$.
\end{proof}
Extending another concept from the Euclidean subcase, we say that a non-zero vector $X\in T_pM$ is an \emph{asymptotic direction} if $X$ is conjugate to itself. In other words, we say that $X \in T_pM\setminus\{0\}$ is an asymptotic direction whenever $D_XX \in T_pM$. It turns out that, in view of this definition, the asymptotic directions of a surface with respect to the Minkowski norm are precisely the same as the ones with respect to the Euclidean norm. \\
A connected regular curve $\gamma:J\rightarrow M$ is said to be an \emph{asymptotic curve} if the tangent line of $\gamma$ at any point is an asymptotic direction. It is clear that if the norm is admissible, then the asymptotic lines with respect to the norm are precisely the same as the ones with respect to Euclidean geometry. In Euclidean differential geometry, the asymptotic directions are characterized as the ones for which $\langle X,d\xi_pX\rangle = 0$. In the general case, when the space is endowed with a Minkowski norm, we clearly can characterize the asymptotic directions as the directions $X \in T_pM$ for which $h(X,X) = 0$. Therefore, as in the Euclidean subcase, if there exists an asymptotic direction at a point $p \in M$ where the eigenvalues of $d\eta_p$ are both non-zero, then we may guarantee the existence of a local asymptotic curve passing through $p$ via standard ordinary differential equations theory. \\
Notice that the principal curvatures of a plane are $0$ at any point. Also, any Minkowski sphere has constant, equal principal curvatures at each of its points. Indeed, the Birkhoff-Gauss map can be regarded as the map $\eta:M\rightarrow \partial B$ given by $\eta(x)=\frac{1}{\rho}(x-p)$, where $p$ is the center $M$ and $\rho$ is the radius. Clearly, $d\eta_p = \frac{1}{\rho}\mathrm{Id}_{T_{p}M}$, and hence the principal curvatures of $M$ at any point $p$ are $\frac{1}{\rho}$. \\
We say that a point $p \in M$ is an \emph{umbilic point} if the principal curvatures of $M$ at $p$ have the same value. Equivalently, a point $p \in M$ is umbilic when the differential of the Birkhoff normal vector field at $p$ is a multiple of the identity map. By the previous observation, all the points of a plane or of a Minkowski sphere are umbilic. The next proposition states that, as in the Euclidean subcase, these are the only possible surfaces with such property.
\begin{prop}\label{allumbilic} An immersed connected surface all whose points are umbilic is contained in a plane or in a Minkowski sphere.
\end{prop}
\begin{proof} For each $p \in M$ we have that $d\eta_p(X) = \lambda(p)X$ for any $X \in T_pM$, where $\lambda:M\rightarrow\mathbb{R}$ is a smooth function. Our first step is to prove that the function $\lambda$ is constant. For this sake, fix linearly independent vectors $X,Y \in T_pM$ and denote also by $X$ and $Y$ the parallel transport of $X$ through a curve tangent to $Y$ at $p$, and the parallel transport of $Y$ through a curve tangent to $X$ at $p$, both with respect to the induced connection $\nabla$. We have then $\nabla_XY|_p = \nabla_YX|_p = 0$. Now, extending smoothly both vector fields to an open neighborhood of $p$, we may calculate at $p$
\begin{align*} D_YD_X\eta = D_Y(\lambda X) = Y(\lambda)X + \lambda h(Y,X)\eta \ \ \mathrm{and} \\
D_XD_Y\eta = X(\lambda)Y + \lambda h(X,Y)\eta.
\end{align*}
Since $D$ is a flat connection, we write
\begin{align*} 0 = D_YD_X\eta - D_XD_Y\eta - D_{[X,Y]}\eta = Y(\lambda)X-X(\lambda)Y - \lambda[X,Y].
\end{align*}
Recalling that $\nabla$ is a torsion-free connection, it follows that $[X,Y] = 0$ at $p$. Hence we have $X(\lambda)Y - Y(\lambda)X = 0$, and this gives $X(\lambda) = Y(\lambda) = 0$ (since $X$ and $Y$ are linearly independent). This argument shows that the derivative of the function $\lambda$ at any point $p \in M$ and with respect to any direction $X \in T_pM$ equals $0$. It follows that $\lambda$ is constant. \\
If $\lambda = 0$, then the Birkhoff-Gauss map is constant, and this means that the Birkhoff normal vector is the same for each point of $M$. In particular, the Euclidean normal vector is also the same for every point, and therefore $M$ is contained in a plane (see \cite{manfredo}). If $\lambda \neq 0$, then the map $x \in M \mapsto x - \frac{1}{\lambda}\eta(x) \in \mathbb{R}^3$ is clearly a constant map. Indeed, for any point $p \in M$ and any direction $X \in T_pM$, we have
\begin{align*} D_X\left(x - \frac{1}{\lambda}\eta(x)\right) = X - \frac{1}{\lambda}(\lambda X) = 0.
\end{align*}
Thus, $M$ is contained in the Minkowski sphere whose center is this constant point, and whose radius equals $\frac{1}{\lambda}$.
\end{proof}
A regular connected curve $\gamma:J\subseteq\mathbb{R}\rightarrow M$ is said to be a \emph{curvature line} if for each $t \in J$ the tangent vector $\gamma'(t)$ gives a principal direction at $\gamma(t)$. We will characterize the curvature lines of a surface in a Minkowski space in a similar manner as it is done for the Euclidean subcase.
\begin{prop} Let $\gamma:J\rightarrow M$ be a regular connected curve. Then $\gamma$ is a curvature line of $M$ if and only if there exists a function $\lambda:J\rightarrow\mathbb{R}$ such that
\begin{align*} (\eta\circ\gamma)'(t) = \lambda(t)\gamma'(t),
\end{align*}
for each $t \in J$.
\end{prop}
\begin{proof} First suppose that $\gamma$ is a curvature line. Then, for each $t \in J$, we have that $\gamma'(t)$ is a principal direction, and therefore $(\eta\circ\gamma)'(t) = d\eta_{\gamma(t)}(\gamma'(t)) = \lambda(t) \gamma'(t)$, where $\lambda(t)$ is an eigenvalue of $d\eta_{\gamma(t)}$. \\
Conversely, assume that $\gamma$ is a connected curve for which $(\eta\circ\gamma)'(t) = \lambda(t)\gamma'(t)$ holds for some function $\lambda:J\rightarrow\mathbb{R}$. For each $t \in J$ we have that $\gamma'(t)$ is an eigenvector of $d\eta_{\gamma(t)}$. Thus, $\gamma$ is a curvature line.
\end{proof}
\section{An analogue of the normal curvature}\label{normal}
Throughout this section \textbf{we still always assume that the norm fixed in the space is admissible}. As usual, we let $u:\partial B_e\rightarrow \partial B$ be the inverse of the Euclidean Gauss map of $\partial B$. Recall also that we are denoting by $\langle\cdot,\cdot\rangle$ the \emph{usual inner product in} $\mathbb{R}^3$. Given an immersed surface $f:M\rightarrow\mathbb{R}^3$, we still denote by $\eta$ and $\xi$ the Birkhoff-Gauss and usual Euclidean Gauss maps of $M$, respectively. \\
In Euclidean differential geometry, the \emph{normal curvature} of a surface $M$ in a given point $p \in M$ and a given direction $X \in T_pM$ can be regarded as the (signed) length of the projection of the normal vector of a curve in $M$, passing through $p$ with tangent vector $X$, onto $\xi(p)$. In particular, the considered curve can be taken as the intersection of the plane spanned by $\xi(p)$ and $X$, and therefore the normal curvature is the usual curvature of this (plane) curve at $p$ (see \cite{manfredo}). This observation allows us to extend this notion to our general case. Let $f:M\rightarrow(\mathbb{R}^3,||\cdot||)$ be an immersed surface, and fix $p \in M$ and $X \in S_p\subseteq T_pM$, where $S_p$ denotes the unit circle of $T_pM$. Denote by $H$ the plane spanned by $\eta(p)$ and $X$. Let $\gamma:(-\varepsilon,\varepsilon)\rightarrow M$ be a local arc-length parametrization of the curve given by the intersection of the plane $p\oplus H$ with $M$, and assume that $\gamma(0) = p$ and $\gamma'(0) = X$.
\begin{definition}The \emph{Minkowski normal curvature} of $M$ at $p \in M$ in the direction $X \in S_p$ is the circular curvature of $\gamma$ at $p$ in the plane geometry endowed in $H$ by the norm $||\cdot||$ (in other words, the geometry in $H$ whose unit circle is the intersection of $\partial B$ with $H$). We will denote this number by $k_{M,p}(X)$.
\end{definition}
We will give a formula for the Minkowski normal curvature in terms of the auxiliary Euclidean structure fixed in the plane. To do so, we first notice that this is essentially a problem in the plane $H$. Following \cite{Ba-Ma-Sho}, the circular curvature of $\gamma$ at $p$ is the ratio between its usual plane Euclidean curvature and the usual plane Euclidean curvature of the circle $\partial B\cap H$ at a point whose tangent lies in the direction $X$.
\begin{teo} For any $p \in M$ and $X \in T_pM$ we have
\begin{align}\label{normalcurv} k_{M,p}(X) = \frac{\langle du^{-1}_{\eta(p)}X,d\eta_pX\rangle}{\langle du^{-1}_{\eta(p)}X,X\rangle},
\end{align}
where we are considering the natural identification $T_pM \simeq T_{\eta(p)}\partial B \simeq T_{\xi(p)}\partial B_e$.
\end{teo}
\begin{proof} Let us first look at $\partial B$ as an immersed surface. The Euclidean normal curvature of $\partial B$ at $\eta(p)$ in the direction $X$ is given by $\langle du^{-1}_{\eta(p)}X,X\rangle$, since $u^{-1}$ is the Euclidean Gauss map of $\partial B$. Following \cite{manfredo}, this normal curvature can be obtained from the curve $\varphi:=\partial B\cap H$ as
\begin{align}\label{eq1} -\langle du^{-1}_{\eta(p)}X,X\rangle = k_{\varphi}(\eta(p))\langle \zeta,\xi(p)\rangle,
\end{align}
where $k_{\varphi}(\eta(p))$ is the (plane) Euclidean curvature of the curve $\varphi$ at $\eta(p)$, and $\zeta$ is the unit Euclidean normal vector to $X$ at the plane $H$, which is also the Euclidean normal vector of the curve $\varphi$ at $\eta(p)$. On the other hand, the Euclidean normal curvature of $M$ at $p$ in the direction $X$ is given by $\langle d\xi_pX,X\rangle$, since $\xi$ is the Euclidean Gauss map of $M$. As in the previous argument, this normal curvature can be obtained from the curve $\gamma$ as
\begin{align}\label{eq2} -\langle d\xi_pX,X\rangle = k_{\gamma}(p)\langle \zeta,\xi(p)\rangle,
\end{align}
where $k_{\gamma}(p)$ is the (plane) Euclidean curvature of $\gamma$ at $p$. Now, from (\ref{eq1}) and (\ref{eq2}) we have
\begin{align*} k_{M,p}(X) = \frac{k_{\gamma}(p)}{k_{\varphi}(\eta(p))} = \frac{\langle d\xi_pX,X\rangle}{\langle du^{-1}_{\eta(p)}X,X\rangle}.
\end{align*}
Since $\xi= u^{-1}\circ\eta$, and since the differential of the Euclidean Gauss map of any immersed surface is self-adjoint at any point, it follows that $\langle d\xi_pX,X\rangle = \langle du^{-1}_{\eta(p)}X,d\eta_pX\rangle$. This gives equality (\ref{normalcurv}).
\end{proof}
We will derive three consequences of this formula (see the next three corollaries). First we prove that, as in the Euclidean subcase, if the normal curvature of a surface is constant, then this surface must be a plane or a Minkowski circle. After that, we will find a relation between the principal directions and the normal curvature. Finally, we show that the asymptotic directions at a point can be characterized in terms of the normal curvature.
\begin{coro}\label{normalumb} The Minkowski normal curvature of an immersed connected surface is constant if and only if this surface is contained in a plane or in a Minkowski sphere. The first case occurs if and only if $k_{M,p} = 0$, and in the second case the radius of the sphere is given by $\left|k_{M,p}\right|^{-1}$.
\end{coro}
\begin{proof} First, it is clear that the Minkowski normal curvature of a plane is always zero, since each plane section yields a straight line segment. Also, if $M$ is a Minkowski sphere of radius $\lambda\in\mathbb{R}$, then we may assume for simplicity that it is centered at the origin, and hence the Birkhoff-Gauss map can be regarded as $\eta(p) = \frac{1}{\lambda}p$. Therefore, $d\eta_p(X) = \frac{1}{\lambda}X$, and we have $k_{M,p}(X) = \frac{1}{\lambda}$. \\
Assume now that the Minkowski normal curvature of a surface $M$ equals $\lambda \in \mathbb{R}$. Then we have that
\begin{align}\label{eq3} \langle du^{-1}_{\eta(p)}X,d\eta_pX\rangle = \lambda\langle du^{-1}_{\eta(p)}X,X\rangle,
\end{align}
for any $p \in M$ and non-zero $X \in T_pM$. Recalling that $du^{-1}_{\eta(p)}$ is a self-adjoint operator, we may take orthogonal unit (in the Euclidean norm) vectors $E_1,E_2 \in T_pM$ such that
\begin{align*} du^{-1}_{\eta(p)}E_1 = \mu_1E_1 \ \ \mathrm{and} \\
du^{-1}_{\eta(p)}E_2 = \mu_2E_2,
\end{align*}
for some $\mu_1,\mu_2 \in \mathbb{R}$. Moreover, since the norm is admissible we may assume that $\mu_1,\mu_2 > 0$. From (\ref{eq3}), it follows that
\begin{align*} d\eta_pE_1 = \lambda E_1 + \alpha E_2 \ \ \mathrm{and} \\
d\eta_p E_2 = \lambda E_2 + \beta E_1,
\end{align*}
for some numbers $\alpha,\beta \in \mathbb{R}$. Applying the equality (\ref{eq3}) to the vector $E_1 + E_2$, we get
\begin{align}\label{eq4} \mu_1\beta + \mu_2\alpha = 0.
\end{align}
In order to obtain another relation between these numbers, we recall that $\xi$ is the Euclidean Gauss map of $M$, and hence $d\xi_p$ is self-adjoint for any $p \in M$. We may write then $\langle d\xi_pE_1,E_2\rangle = \langle E_1,d\xi_pE_2\rangle$. But since $\xi = u\circ \eta$, this equality reads
\begin{align*} \left\langle \frac{\lambda}{\mu_1}E_1+\frac{\alpha}{\mu_2}E_2,E_2\right\rangle = \left\langle E_1,\frac{\lambda}{\mu_2}E_2 + \frac{\beta}{\mu_1}E_1\right\rangle.
\end{align*}
Therefore we have $\alpha\mu_1 - \beta\mu_2 = 0$. This equality, together with (\ref{eq4}) and the fact that $\mu_1,\mu_2 \neq 0$, gives that $\alpha = \beta = 0$. This means that, for each $p \in M$, we have $d\eta_p = \lambda \mathrm{Id}|_{T_pM}$. The result comes now from Proposition \ref{allumbilic}.
\end{proof}
\begin{remark} We do not need an auxiliary Euclidean structure to prove that any Minkowski sphere has constant Minkowski normal curvature. Indeed, any normal section $H$ on a Minkowski sphere will yield a curve which is a homothet of the unit circle of $(H,||\cdot||)$. Hence, its circular curvature will be the inverse of the radius.
\end{remark}
\begin{coro}\label{princnorm} Let $V \in T_pM$ be a principal direction at $p \in M$, with associated principal curvature $\lambda$. Then $k_{M,P}(V) = \lambda$.
\end{coro}
\begin{proof} Since $d\eta_p(V) = \lambda V$, it follows that
\begin{align*} k_{M,p}(V) = \frac{\langle du^{-1}_{\eta(p)}V,\lambda V\rangle}{\langle du^{-1}_{\eta(p)}V,V\rangle} = \lambda.
\end{align*}
\end{proof}
\begin{coro}\label{asympnorm} A non-zero vector $X \in T_pM$ is an asymptotic direction if and only if $k_{M,p}(X) = 0$.
\end{coro}
\begin{proof} From (\ref{exph}) and (\ref{normalcurv}) we have the equality
\begin{align*} k_{M,p}(X) = -\frac{h(X,X)\langle \eta,\xi \rangle}{\langle du^{-1}_{\eta(p)}X,X\rangle}.
\end{align*}
Therefore, $k_{M,p}(X) = 0$ if and only if $h(X,X) = 0$.
\end{proof}
Again, here the general Minkowski case presents a similar behavior as the Euclidean one. Also, the umbilic points can be characterized as the points where the normal curvature is the same for every direction. We will prove this now.
\begin{prop}\label{normumb} A point $p \in M$ is umbilic if and only if $k_{M,p}$ is constant in $T_pM\setminus\{0\}$. In this case, $k_{M,p}$ equals the principal curvature of $M$ at $p$.
\end{prop}
\begin{proof} Assume first that $p \in M$ is umbilic, and let $\lambda$ be the value of the principal curvature of $M$ at $p$. If $X \in T_pM$, then
\begin{align*} k_{M,p}(X) = \frac{\langle du^{-1}_{\eta(p)}X,\lambda X\rangle}{\langle du^{-1}_{\eta(p)}X,X\rangle} = \lambda.
\end{align*}
Now suppose that $k_{M,p} = \lambda$ is constant. Then we have
\begin{align*} \langle du^{-1}_{\eta(p)}X,d\eta_pX\rangle = \lambda\langle du^{-1}_{\eta(p)}X,X\rangle
\end{align*}
for every $X \in T_pM$. From the same argument as used in Corollary \ref{normalumb} it follows that $d\eta_p = \lambda\mathrm{Id}|_{T_{p}M}$. Hence $p$ is an umbilic point with principal curvature $\lambda$.
\end{proof}
In the Euclidean subcase, the principal curvatures of $M$ at a point $p$ are precisely the maximum and the minimum values of the normal curvature in this point. It might be a little surprising that this is also true in the general Minkowski case. Namely, we have
\begin{teo} Let $\lambda_1\geq\lambda_2$ be the principal curvatures of $M$ at $p$. Then we have the inequalities $\lambda_2\leq k_{M,p}(V)\leq \lambda_1$, for every $V \in T_pM\setminus\{0\}$. Moreover, equality holds only in the respective principal directions.
\end{teo}
\begin{proof} First notice that if both principal curvatures are equal, then the result comes straightforwardly (Proposition \ref{normumb}). Let $V_1,V_2 \in T_pM$ be the principal directions associated to $\lambda_1$ and $\lambda_2$, respectively, and asssume that the principal curvatures are distinct. Recall that $V_1$ and $V_2$ are conjugate directions, which means that $h(V_1,V_2) = 0$ (Lemma \ref{princconj}). From (\ref{exph}) it follows that
\begin{align*} \langle du^{-1}_{\eta(p)}V_1,d\eta_pV_2\rangle = \langle du^{-1}_{\eta(p)}V_2,d\eta_pV_1\rangle = 0.
\end{align*}
Hence we have
\begin{align*} \langle du^{-1}_{\eta(p)}V_1,V_2\rangle = \frac{1}{\lambda_2}\langle du^{-1}_{\eta(p)}V_1,d\eta_pV_2\rangle = 0,
\end{align*}
and the same argument holds for $\langle du^{-1}_{\eta(p)}V_2,V_1\rangle$. If $V_2$ is, say, associated to the eigenvalue zero, then
\begin{align*} \langle du^{-1}_{\eta(p)}V_1,V_2\rangle = \langle V_1,du^{-1}_{\eta(p)}V_2\rangle = \frac{1}{\lambda_1}\langle d\eta_pV_1,du^{-1}_{\eta(p)}V_2\rangle = 0.
\end{align*}
\noindent Summarizing this, for $\lambda_1\neq\lambda_2$ we necessarily have
\begin{align*}\langle du^{-1}_{\eta(p)}V_1,V_2\rangle = \langle du^{-1}_{\eta(p)}V_2,V_1\rangle = 0.
\end{align*}
Therefore, decomposing a non-zero vector $V \in T_pM$ by $V = \alpha V_1 + \beta V_2$ yields the equality
\begin{align*} k_{M,p}(V) = \frac{\langle du^{-1}_{\eta(p)}V,\alpha\lambda_1V_1 + \beta\lambda_2V_2\rangle}{\langle du^{-1}_{\eta(p)}V,V\rangle} = \frac{\alpha^2\lambda_1\langle du^{-1}_{\eta(p)}V_1,V_1\rangle + \beta^2\lambda_2\langle du^{-1}_{\eta(p)}V_2,V_2\rangle}{\alpha^2\langle du^{-1}_{\eta(p)}V_1,V_1\rangle + \beta^2\langle du^{-1}_{\eta(p)}V_2,V_2\rangle}.
\end{align*}
Using again the fact that $u^{-1}$ is the (Euclidean) Gauss map of $\partial B$, we may assume that the bilinear form associated to $du^{-1}_{\eta(p)}$ is positive definite. Hence the equality above yields immediately $\lambda_2 \leq k_{M,p}(V) \leq \lambda_1$. The claim on the equality cases is also immediate. \\
\end{proof}
\begin{remark} In the Euclidean case, at any point the sum of the normal curvatures in orthogonal of complementary directions is constant (cf. \cite{manfredo}). We find something similar in \cite{diffgeom2}, by endowing the surface with a certain Riemannian metric whose orthogonality relation ``organizes" the directions in the same way.
\end{remark}
\begin{coro} Let $f:M\rightarrow(\mathbb{R}^3,||\cdot||)$ be an immersed surface, and let $G:(\mathbb{R}^3,||\cdot||)\rightarrow(\mathbb{R}^3,||\cdot||)$ be an isometry of the space. Then, up to the sign, the normal curvature of $M$ at each point is invariant under $G$. In particular, the Minkowski Gaussian curvature of $M$ is invariant under the action of $G$.
\end{coro}
\begin{proof} Let $p \in M$ and $X \in T_pM$. We have to prove that $k_{M,p}(X) = k_{G(M),G(p)}(G(X))$. Let $H$ be the plane spanned by $X$ and $\eta(p)$, and calculate the normal curvature $k_{M,p}(X)$ at $p$ as the circular curvature of the curve $\gamma$ determined by the section of $M$ through $H$.\\
Since Birkhoff orthogonality is invariant under isometry, we have that $\eta(G(p)) = G(\eta(p))$, and hence the plane spanned by $G(X)$ and $\eta(G(p))$ is precisely $G(H)$, which is isometric to $H$ by $G$. Also, every isometry of a normed space is a composition of a linear map with a translation (Mazur-Ulam theorem, cf. \cite{thompson}). Therefore, the normal curvature $k_{G(M),G(p)}(G(X))$ is the circular curvature of the curve $G(\gamma)$ at $G(p)$. The desired follows now from the fact that the circular curvature is, up to the sign, invariant under isometries of the plane (see \cite[Theorem 7.2]{Ba-Ma-Sho}). Whether or not the signs at $p$ change depends on the restriction of $G$ to $H$, being or not being orientation preserving. In both cases the Minkowski Gaussian curvature remains the same.
\end{proof}
\begin{remark} The corollary above states that the Minkowski Gaussian curvature is invariant under an isometry of the ambient space. However, a natural extension of the concept of isometry for surfaces immersed in a Minkowski space would be as follows: let $M$, $N$ be two surfaces immersed in $(\mathbb{R}^3,||\cdot||)$. A smooth mapping $F:M\rightarrow N$ is said to be an \emph{isometry} (in the induced norm) if $||dF_pV|| = ||V||$ for any $p \in M$ and $V \in T_pM$. It seems to be difficult to decide whether the Minkowski Gaussian curvature is invariant under isometries.
\end{remark}
\section{Some rigidity theorems} \label{rigid}
In this section we want to prove that, with a certain additional hypothesis, any compact, connected surface immersed in a space endowed with an admissible norm which has constant positive Minkowski Gaussian curvature, or constant mean curvature, is a Minkowski circle. Our proof follows the steps of the proof given in \cite{manfredo} for the Euclidean subcase. We will start with two auxiliary lemmas.
\begin{lemma}\label{point} Let $f:M\rightarrow (\mathbb{R}^3,||\cdot||)$ be an immersion of a compact, connected surface without boundary. Then there exists a point $p \in M$ such that the product of the principal curvatures of $M$ at $p$ is positive.
\end{lemma}
\begin{proof} Let $B_M$ be the smallest (closed) Minkowski ball centered at the origin and containing $M$. Then it is clear that there exists a point $p \in B_M\cap M$, and that $T_pM = T_pB_M$. Taking any normal section, and regarding the normal curvatures of $B_M$ and $M$ the circular curvatures of the respective intersection curves, it becomes clear that the normal curvature of $B_M$ is greater than or equal to the normal curvature of $M$ (indeed, the opposite would make it possible to construct closed, convex curves contradicting \cite[Theorem 8.3]{Ba-Ma-Sho}). Since the principal curvatures of $M$ at $p$ are the normal curvatures in the associated principal directions, it follows that their product is greater than or equal to $1/r^2$, where $r$ is the radius of $B_M$.
\end{proof}
\begin{remark} An easier (but less geometric) proof of this lemma could be based on the usage of the same result for the Euclidean case (as it is proved in \cite{manfredo}) and Proposition \ref{gaussposit}.\\
\end{remark}
Before coming to the next auxiliary lemma, let us present an observation. Assume that the principal curvatures of $M$ at a point $p$ are distinct, and both non-zero. Then, due to continuity, this condition on the principal curvatures is true in an open neighborhood $U$ of $p$. Therefore we may set functions $\lambda_1:U\rightarrow \mathbb{R}$ and $\lambda_2:U\rightarrow\mathbb{R}$ to be the greatest and the smallest principal curvatures at $q \in U$, respectively, and hence we can choose distinct vector fields $V_1,V_2 \in C^{\infty}(U)$ such that, at each point $q \in U$, $V_1$ and $V_2$ are principal directions associated to $\lambda_1$ and $\lambda_2$, respectively. It follows from Lemma \ref{princconj} and the comments below it that $V_1$ and $V_2$ give (linearly independent) conjugate directions at any point of $U$. From Corollary 1 in Section 3.4 of \cite{manfredo} it follows that we may endow $U$ with a parametrization whose coordinate curves are tangent to the respective directions given by $V_1$ and $V_2$. In our language, this is the same as stating that we may re-scale these vector fields in such a way that we have $[V_1,V_2] = 0$ at each point of $U$. \\
Let $V_1$ and $V_2$ be the principal vector fields of a surface $M$. Then its trajectories are the curvature lines of $M$. We say that $M$ has \emph{coercively convex curvature lines} at $p \in M$ if the following hypothesis holds: let $p \in M$ be a point where $D_{V_1}V_2|_p = 0$, and let $\gamma$ be a local trajectory of $V_2$ through $p$. Then the projection of $D_{V_1}V_2|_{\gamma}$ onto $V_1$ has negative derivative at $p$. Intuitively, this means that the curvature line associated to $V_2$ ``forces $D_{V_1}V_2$ to the other side of the curve when one passes through $p$" (see Figure \ref{coercive}).
\begin{figure}[h]
\centering
\includegraphics{coercive.eps}
\caption{Coercive convexity of the curvature lines of $M$ at $p$.}
\label{coercive}
\end{figure}
\begin{lemma}\label{minmax} Let $f:M\rightarrow(\mathbb{R}^3,||\cdot||)$ be an immersed surface, and let $V_1,V_2$ be the vector fields given by the principal directions associated to the principal directions $\lambda_1 \geq \lambda_2$. Let $p \in M$, where the Minkowski Gaussian curvature of $M$ is positive, and assume that $\lambda_1$ has a local maximum at $p$ and that $\lambda_2$ has a local minimum at $p$. Suppose also that $M$ has coercively convex curvature lines at $p$. Then $p$ is an umbilic point.
\end{lemma}
\begin{proof} Suppose that $p$ is not umbilic, take a neighborhood $U$ of $p$ where $\lambda_1 > \lambda_2$, and let $V_1, V_2 \in C^{\infty}(U)$ be vector fields as in the observation above. Since $[V_1,V_2]$ vanishes identically and $D$ is a flat connection, we have that $D_{V_2}D_{V_1}\eta = D_{V_1}D_{V_2}\eta$. This equality reads
\begin{align*} V_2(\lambda_1)V_1 + \lambda_1D_{V_2}V_1 = V_1(\lambda_2)V_2 + \lambda_2D_{V_1}V_2.
\end{align*}
Differentiating this equality in the directions of $V_1$ and $V_2$, respectively, we get the equalities
\begin{align*} V_1(V_2(\lambda_1))V_1 + V_2(\lambda_1)D_{V_1}V_1 + V_1(\lambda_1)D_{V_2}V_1 + \lambda_1D_{V_1}D_{V_2}V_1 = \\ =V_1(V_1(\lambda_2))V_2 + V_1(\lambda_2)D_{V_1}V_2 + V_1(\lambda_2)D_{V_1}V_2 + \lambda_2D_{V_1}D_{V_1}V_2,
\end{align*}
and
\begin{align*} V_2(V_2(\lambda_1))V_1 + V_2(\lambda_1)D_{V_2}V_1 + V_2(\lambda_1)D_{V_2}V_1 + \lambda_1D_{V_2}D_{V_2}V_1 = \\ = V_2(V_1(\lambda_2))V_2 + V_1(\lambda_2)D_{V_2}V_2 + V_2(\lambda_2)D_{V_1}V_2 + \lambda_2D_{V_2}D_{V_1}V_2.
\end{align*}
At $p$ we have $V_1(\lambda_1) = V_2(\lambda_1) = V_1(\lambda_2) = V_2(\lambda_2) = 0$, since $p$ is a local extremum of both $\lambda_1$ and $\lambda_2$. Since we have vectorial equalities, in each of them the projections of both sides onto $\eta$ must be equal. At $p$, it is clear that the only terms which have non-zero projection onto $\eta$ are the terms of the form $D_{V_i}D_{V_j}V_k$. Therefore, by the Gauss equation we have the following: looking at the restriction of the first equality to the direction $\eta$ at $p$ yields
\begin{align*} \lambda_1h(D_{V_2}V_1,V_1) = \lambda_2h(D_{V_1}V_2,V_1).
\end{align*}
Since $\lambda_1 \neq \lambda_2$ and $D_{V_1}V_2 = D_{V_2}V_1$, it follows that $h(D_{V_2}V_1,V_1) = h(D_{V_1}V_2,V_1) = 0$. Hence, at $p$, the vectors $D_{V_1}D_{V_2}V_1$ and $D_{V_1}D_{V_1}V_2$ are tangential. Repeating the argument for the second equality, we see that the same holds for $D_{V_2}D_{V_2}V_1$ and $D_{V_2}D_{V_1}V_2$. \\
Recall that from Lemma \ref{princconj} we have $h(V_1,V_2) = 0$. Evaluating (at $p$) the affine fundamental form with the vectors involved in the first equality and the vector $V_2$, we obtain
\begin{align}\label{main1} \lambda_1h(D_{V_1}D_{V_2}V_1,V_2) = V_1(V_1(\lambda_2))h(V_2,V_2) + \lambda_2h(D_{V_1}D_{V_1}V_2,V_1).
\end{align}
Doing the same with the second equality and the vector $V_1$, we get
\begin{align} \label{main2} V_2(V_2(\lambda_1))h(V_1,V_1) + \lambda_1h(D_{V_2}D_{V_2}V_1,V_1) = \lambda_2h(D_{V_2}D_{V_1}V_2,V_1).
\end{align}
We claim now that
\begin{align*}h(D_{V_1}D_{V_2}V_1,V_2) = h(D_{V_1}D_{V_1}V_2,V_2) = h(D_{V_2}D_{V_2}V_1,V_1) = h(D_{V_2}D_{V_1}V_2,V_1) = C,
\end{align*}
say. Indeed, using again the fact that $[V_1,V_2] = 0$ and recalling that $D$ is a flat connection, we have that $D_{V_2}(D_{V_1}D_{V_2}V_1) = D_{V_1}(D_{V_1}D_{V_1}V_2)$, and hence their projections onto $\eta$ are equal. Then the Gauss equation gives $h(D_{V_1}D_{V_2}V_1,V_2) = h(D_{V_1}D_{V_1}V_2,V_1)$. We use the same argument to derive the other two equalities. Now, summing up (\ref{main1}) and (\ref{main2}), we get
\begin{align}\label{main} V_2(V_2(\lambda_1))h(V_1,V_1) - V_1(V_1(\lambda_2))h(V_2,V_2) = (\lambda_2-\lambda_1)C.
\end{align}
Our last step is to prove that $C < 0$, up to re-orientation of $V_2$. For this sake, we decompose the vector field $D_{V_1}V_2$ in coordinates as $D_{V_1}V_2 = fV_1+gV_2$, for some functions $f,g:U\rightarrow\mathbb{R}$. From the previous equalities $h(D_{V_1}V_2,V_1) = h(D_{V_1}V_2,V_2) = 0$ at $p$, we have that $f(p) = g(p) = 0$. Derivating, we have
\begin{align*} D_{V_2}D_{V_1}V_2 = V_2(f) + fD_{V_2}V_1 + V_2(g)V_2 + gD_{V_2}V_2,
\end{align*}
and hence we have $C = h(D_{V_2}D_{V_1}V_2,V_1) = V_2(f)h(V_1,V_1)$ at $p$. From the coercive convexity of the curvature lines of $M$ at $p$ it follows that $V_2(f) < 0$. Finally, we look at (\ref{main}). Assume that, without loss of generality, $h$ is positive definite. Then the left hand side is non-positive, since $V_2(V_2(\lambda_1)) \leq 0$ and $V_1(V_1(\lambda_2)) \geq 0$. However, the right hand side is strictly positive, since $(\lambda_2 - \lambda_1) < 0$ and $C < 0$. This contradiction concludes the proof.
\end{proof}
We are ready now to state and prove the first theorem of this section.
\begin{teo} Let $f:M\rightarrow(\mathbb{R}^3,||\cdot||)$ be a compact, connected immersed surface without boundary, and assume that the norm $||\cdot||$ is admissible. Assume also that $M$ has coercively convex curvature lines at any point where the greatest principal curvature function attains a global maximum. If the Minkowski Gaussian curvature of $M$ is constant, then $M$ is a Minkowski sphere.
\end{teo}
\begin{proof} From Lemma \ref{point} it follows that the Minkowski Gaussian curvature of $M$ is positive everywhere. Let $\lambda_1,\lambda_2:M\rightarrow\mathbb{R}$ be the principal curvature functions of $M$, with $\lambda_1 \geq \lambda_2$. By compactness, there exists a point $p \in M$ such that $\lambda_1$ attains its maximum over $M$. Since the product $\lambda_1\lambda_2$ is constant, it follows that $\lambda_2$ attains a minimum at $p$. Hence, from Lemma \ref{minmax} we have that $\lambda_1(p) = \lambda_2(p)$. Therefore, for any $q \in M$ we have
\begin{align*} \lambda_2(p) \leq \lambda_2(q) \leq \lambda_1(q) \leq \lambda_1(p) = \lambda_2(p),
\end{align*}
and this shows that every point of $M$ is an umbilic point. From Proposition \ref{allumbilic} it follows that $M$ is contained in a Minkowski sphere. Since $M$ is connected, compact and without boundary, it follows that $M$ is both open and closed in this sphere, and therefore $M$ is the entire Minkowski sphere.
\end{proof}
Under the same hypothesis, we can also prove that a compact, connected immersed surface with positive Minkowski Gaussian curvature and constant mean curvature is a Minkowski sphere.
\begin{teo} Let $f:M\rightarrow (\mathbb{R}^3,||\cdot||)$ be a compact, connected surface immersed in an admissible Minkowski space. Assume also that $M$ has coercively convex curvature lines at any point where the greatest principal curvature function attains a global maximum. If the Minkowski Gaussian curvature is positive and the mean curvature is constant, then $M$ is a Minkowski sphere.
\end{teo}
\begin{proof} Let $p \in M$ be a point where the greatest principal curvature function $\lambda_1$ attains its maximum. Since the mean curvature is constant, it follows that $M$ attains its minimum at $p$. From Lemma \ref{minmax} it follows that $p$ is an umbilic point. As in the proof of the last theorem, it follows that every point of $M$ is umbilic, and then $M$ is a Minkowski sphere.
\end{proof}
|
\section{ Introduction }
In the field of Many-Body-Localization (see the recent reviews \cite{revue_huse,revue_altman,revue_vasseur,revue_imbrie,revue_rademaker,review_mblergo,review_prelovsek,review_rare,review_toulouse} and references therein), one of the important characterization of Many-Body-Localized phases is the existence
of an extensive number of Local Integrals of Motion called LIOMs
\cite{emergent_swingle,emergent_serbyn,emergent_huse,emergent_ent,imbrie,serbyn_quench,emergent_vidal,emergent_ros,
emergent_rademaker,serbyn_powerlawent,c_emergent,ros_remanent,c_liom}.
Since these LIOMS are the building blocks of the whole set of eigenstates,
it is natural to try to identify them via some real-space renormalization procedure.
The Strong Disorder Real-Space RG approach
developed by Daniel Fisher \cite{fisher_AF,fisher,fisherreview} to construct the ground states of random quantum models (see the review \cite{strong_review}) has been thus generalized into the RSRG-X procedure to construct the whole set of excited eigenstates \cite{rsrgx,rsrgx_moore,vasseur_rsrgx,yang_rsrgx,rsrgx_bifurcation} : the idea is that each local renormalization step produces a LIOM that describes the choice between the local energy levels (instead of projecting always onto the lowest energy-level).
The RSRG-t procedure developed by Vosk and Altman \cite{vosk_dyn1,vosk_dyn2}
in order to construct the effective dynamics via the iterative elimination of the degree of freedom oscillating with the highest local eigenfrequency is equivalent to the RSRG-X procedure but gives an interesting different point of view \cite{c_rsrgt}.
Since the purpose of these Strong Disorder RG procedures is to produce an extensive number of LIOMS,
it is clear that their validity is limited to Many-Body-Localized Phases : they cannot be applied in the delocalized ergodic phase,
and they do not allow to analyze the MBL transition towards this delocalized phase.
In particular, it should be stressed that the current RG descriptions of the MBL delocalization transition
are based on completely different RG rules concerning the entanglement \cite {vosk_rgentanglement},
the resonances \cite{vasseur_resonant,vasseur_resonantbis}, or the decomposition into insulating and thermal blocks \cite{huse_rgpaste}.
However, the RSRG-X is very useful in MB-Localized phases to analyse the long-ranged order of the excited eigenstates made of LIOMs
and the possible phase transitions between different Many-Body-Localized phases, as for instance the
transition between the paramagnetic and the spin-glass Many-Body-Localized phases for the one-dimensional generalized quantum Ising model \cite{rsrgx}.
In the present paper, we wish similarly to analyse the transition between the paramagnetic and the spin-glass Many-Body-Localized phases
for the quantum Ising model with random couplings and random transverse fields on the Cayley tree.
Since the standard RSRG-X procedure destroys the tree structure and could only be followed numerically,
we will instead use an RG procedure that preserves the tree structure in order to obtain some simple analytical approximation :
the Pacheco-Fernandez block-RG introduced for the ground state of the one-dimensional chain without disorder
\cite{pacheco,igloiSD} or with disorder \cite{nishiRandom,us_pacheco,us_renyi} is applied here sequentially
\cite{us_watermelon} around the center of the tree in order to construct the whole set of eigenstates.
The paper is organized as follows.
In section \ref{sec_rsrgx}, the real-space RG procedure to construct the set of eigenstates
of the random quantum Ising model on the Cayley tree is described.
Section \ref{sec_pr} is devoted to the large deviation properties of the basic variables that appear in the RG flows.
The statistics of the renormalized couplings and of the renormalized transverse field of the center
are studied in section \ref{sec_jr} and in section \ref{sec_hr} respectively in order to characterize the critical properties
of the transition between the paramagnetic and spin-glass Many-Body-Localized phases.
Our conclusions are summarized in section \ref{sec_conclusion}.
\section{ Real-Space RG procedure to construct the set of eigenstates }
\label{sec_rsrgx}
\subsection{ Model}
We consider the geometry of a Cayley tree of branching ratio $K$ with $L$ generations around the central spin $\sigma_0$.
It is convenient to decompose the quantum Ising Hamiltonian in terms of the contributions
of the various generations
\begin{eqnarray}
H && = \sum_{r=0}^L H_r
\nonumber \\
H_0 && = h_0 \sigma_0^z
\nonumber \\
H_1 && = \sum_{i_1=1}^{K+1} ( J_{i_1} \sigma_0^x \sigma_{i_1}^x + h_{i_1} \sigma_{i_1}^z )
\nonumber \\
H_2 && = \sum_{i_1=1}^{K+1} \sum_{i_2=1}^{K} ( J_{i_1,i_2} \sigma_{i_1}^x \sigma_{i_1,i_2}^x+ h_{i_1,i_2} \sigma_{i_1,i_2}^z)
\nonumber \\
H_r && = \sum_{i_1=1}^{K+1} \sum_{i_2=1}^{K} .. \sum_{i_r=1}^{K}
( J_{i_1,i_2,..,i_r} \sigma_{i_1,..,i_{r-1}}^x \sigma_{i_1,..,i_r}^x
+ h_{i_1,..,i_r} \sigma_{i_1,..,i_r}^z )
\label{htree}
\end{eqnarray}
We consider
that both the transverse fields $h_{i_1,i_2,..,i_r}$ and the couplings $J_{i_1,i_2,..,i_r}$
are random variables drawn with some continuous distributions.
As example, we will focus on the case where the probability distributions of the couplings $J_{i_1,i_2,..,i_r}$ and of the random fields $h_{i_1,i_2,..,i_r}$ are uniformly drawn on $[-J,+J]$ and $[-h,+h]$ respectively
\begin{eqnarray}
\pi_{J}(J_{i_1,i_2,..,i_r}) =\frac{\theta( -J \leq J_{i_1,i_2,..,i_r} \leq J) }{ 2 J}
\nonumber \\
\pi_{h}(h_{i_1,i_2,..,i_r}) =\frac{\theta( -h \leq h_{i_1,i_2,..,i_r} \leq h) }{ 2 h}
\label{box}
\end{eqnarray}
This model is MB-Localized in the two following limits :
(i) When all couplings vanish $J_{i_1,i_2,..,i_r} \to 0$, the model is in a trivial Paramagnetic MB-Localized phase,
where the LIOMs (commuting with themselves and the Hamiltonian)
are the site operators $\sigma^z_{i_1,i_2,..,i_r}$ that are coupled to the random fileds $h_{i_1,..,i_r} $
\begin{eqnarray}
\tau^z_{i_1,i_2,..,i_r} \mathop{\simeq}_{J_{\{ \}} \to 0} \sigma^z_{i_1,i_2,..,i_r}
\label{liomjzero}
\end{eqnarray}
Since all the random fields $h_{i_1,..,i_r} $ are different, there is no degeneracy between the many-body-energy-levels,
and the LIOMs of Eq. \ref{liomjzero} are perturbatively stable
in the presence of small couplings $J_{\{ \}}$ .
Note the difference with the model studied in Ref. \cite{mossi} where the transverse fields all take the same value $h$,
so that the zero-coupling model is the pure paramagnetic model characterized by huge degeneracies in many-body-energy-levels.
(ii) In the opposite limit where all fields vanish $h_{i_1,i_2,..,i_r} \to 0$, the model is in a trivial Spin-Glass Localized phase,
where the LIOMs are the bond operators $\sigma_{i_1,..,i_{r-1}}^x \sigma_{i_1,..,i_r}^x$ associated to the random couplings $J_{i_1,..,i_r} $
\begin{eqnarray}
\tau^z_{i_1,i_2,..,i_r} \mathop{\simeq}_{h_{\{ \}} \to 0} \sigma_{i_1,..,i_{r-1}}^x \sigma_{i_1,..,i_r}^x
\label{liomhzero}
\end{eqnarray}
Since all the couplings $ J_{i_1,..,i_r}$ are different, the many-body-energy-levels are non-degenerate,
and the LIOMs of Eq. \ref{liomhzero} are perturbatively stable in the presence of small fields $h_{\{ \}} $.
Note again the difference with the model studied in Ref. \cite{mossi} where the couplings only take the two values $(\pm J)$,
leading to huge degeneracies in energy many-body-energy-levels.
In summary, the continuous distributions of both random fields and couplings is necessary to avoid degeneracies between many-body-energy-levels and
to identify simple LIOMs in the two limits of vanishing couplings (Eq \ref{liomjzero}) or vanishing fields (Eq. \ref{liomhzero}).
Since these two type of LIOMs correspond to different Long-Ranged order for the corresponding eigenstates, namely Paramagnetic and Spin-Glass,
one expects that the full model containing both fields and couplings will display
a phase transition between two different Many-Body-Localized phases (Paramagnetic and Spin-Glass).
The goal of the present paper is to analyse this transition
via some real-space procedure that constructs the LIOMs and thus the set of eigenstates.
\subsection { First RG step }
The RSRG-X procedure mentioned in the Introduction can be applied in $d>1$, but the changes of the geometry
prevents the finding of any analytical description. The renormalization procedure has to be implemented numerically,
as was done for the RSRG procedure concerning the ground state in $d=2,3,4$ \cite{fisherreview,motrunich,lin,karevski,lin07,yu,kovacsstrip,kovacs2d,kovacs3d,kovacsentropy,kovacsreview}.
Here we wish instead to obtain some analytically solvable RG procedure in order to get more insight
into the mechanism of the transition.
We have thus chosen to apply sequentially \cite{us_watermelon} around the center of the tree
the idea of the Pacheco-Fernandez elementary step \cite{pacheco,igloiSD,nishiRandom,us_pacheco,us_renyi}
in order to keep a simple geometry along the RG flow.
More precisely, the first RG step consists in the diagonalization
of the Hamiltonian $H_1$ Eq. \ref{htree} concerning the center spin and the $(K+1)$ spins
of the first generation
\begin{eqnarray}
H_{1} =
\sum_{i_1=1}^{K+1} \left( J_{i_1} \sigma_0^x \sigma_{i_1}^x + h_{i_1} \sigma_{i_1}^z \right)
\label{h1c}
\end{eqnarray}
Since $H_1$ commutes with $\sigma_0^x$, one needs to consider the two possible values $ \sigma_0^x=S_0^x=\pm 1$,
and to diagonalize the $(K+1)$ remaining effective Hamiltonians involving the single spin $\sigma_{i_1}$
\begin{eqnarray}
H^{eff}_{i_1} = J_{i_1} S_0^x \sigma_{i_1}^x + h_{i_1} \sigma_{i_1}^z
\label{heffi1}
\end{eqnarray}
The two eigenvalues of Eq. \ref{heffi1} do not depend on the value $S_0^x=\pm$ and read
\begin{eqnarray}
\lambda^{(\tau_{i_1}^z) }_{i_1} = \tau_{i_1}^z \sqrt{ J_{i_1}^2+h_{i_1}^2 }
\label{lambdai1}
\end{eqnarray}
where the variable
\begin{eqnarray}
\tau_{i_1}^z = \pm
\label{tauiz}
\end{eqnarray}
labels the choice between the positive or negative energy in Eq. \ref{lambdai1}.
The corresponding eigenvectors depend on the value $S_0^x$
\begin{eqnarray}
\vert \lambda^{(\tau_{i_1}^z) }_{i_1}(S_0^x) > =
\sqrt{ \frac{1}{2} \left( 1+\frac{\tau_{i_1}^z S_0^x J_{i_1} }{\sqrt{ J_{i_1}^2 + h_{i_1}^2 }} \right) } \vert \sigma_{i_1}^x=+
\rangle
+ \tau_{i_1}^z {\rm sgn}(h_{i_1}) \sqrt{ \frac{1}{2} \left( 1- \frac{\tau_{i_1}^z S_0^x J_{i_1} }{\sqrt{ J_{i_1}^2 + h_{i_1}^2 }} \right) } \vert \sigma_{i_1}^x=- \rangle
\label{lambdatauiz}
\end{eqnarray}
To make the link with the Lioms of Eq. \ref{liomjzero} and \ref{liomhzero}, it is usefule to condider the two corresponding limits :
(i) if the coupling vanishes $J_1=0$, the eigenvalues and eigenvectors reduce to
\begin{eqnarray}
\lambda^{(\tau_{i_1}^z) }_{i_1} && \mathop{\simeq}_{J_1=0} \tau_{i_1}^z \vert h_{i_1} \vert = \tau_{i_1}^z {\rm sgn}(h_{i_1}) h_{i_1}
\nonumber \\
\vert \lambda^{(\tau_{i_1}^z) }_{i_1}(S_0^x) > && \mathop{\simeq}_{J_1=0}
\frac{\vert \sigma_{i_1}^x=+ \rangle
+ \tau_{i_1}^z {\rm sgn}(h_{i_1}) \vert \sigma_{i_1}^x=- \rangle}{ \sqrt 2 } = \vert \sigma_{i_1}^z= \tau_{i_1}^z {\rm sgn}(h_{i_1}) \rangle
\label{lambdatauizjzero}
\end{eqnarray}
that is equivalent to Eq. \ref{liomjzero} up to the factor $ {\rm sgn}(h_{i_1})$ that comes from the choice of Eq. \ref{tauiz} to label the sign of the energy of Eq. \ref{lambdai1}.
(ii) if the field vanishes $h_{i_1}=0$, the eigenvalues and eigenvectors become
\begin{eqnarray}
\lambda^{(\tau_{i_1}^z) }_{i_1} && \mathop{\simeq}_{h_1=0} \tau_{i_1}^z \vert J_{i_1} \vert = \tau_{i_1}^z {\rm sgn}(J_{i_1}) J_{i_1}
\\
\vert \lambda^{(\tau_{i_1}^z) }_{i_1}(S_0^x) > && \mathop{\simeq}_{h_1=0}
\sqrt{ \frac{1+\tau_{i_1}^z S_0^x {\rm sgn} ( J_{i_1})}{2} } \vert \sigma_{i_1}^x=+ \rangle
+ \tau_{i_1}^z {\rm sgn}(h_{i_1})
\sqrt{ \frac{1-\tau_{i_1}^z S_0^x {\rm sgn} ( J_{i_1})}{2} } \vert \sigma_{i_1}^x=- \rangle
\propto \vert \sigma_{i_1}^x=\tau_{i_1}^z S_0^x {\rm sgn} ( J_{i_1}) \rangle
\nonumber
\label{lambdatauizhzero}
\end{eqnarray}
that is equivalent to Eq. \ref{liomhzero} up to the factor $ {\rm sgn}(J_{i_1})$ that comes from the choice of Eq. \ref{tauiz} to label the sign of the energy of Eq. \ref{lambdai1}.
When the coupling $J_{i_1}$ and the field $h_{i_1}$ are both non-vanishing,
the LIOM $\tau_{i_1}^z$ defined by Eqs \ref{lambdai1} and \ref{lambdatauiz} can be thus considered as the appropriate
interpolation between these two simple limits (i) and (ii).
Note that in usual Strong-Disorder RG rules for MB-Localized phases \cite{rsrgx}, each LIOM is declared to be associated either to a site variable as in (i)
(if its renormalized transverse field is the biggest among surviving variables) or to a bond variable as in (ii)
(if its renormalized coupling is the biggest among surviving variables), so that each LIOM could be called accordingly 'paramagnetic' or 'spin-glass'.
On the contrary, within the present procedure, the LIOM $\tau_{i_1}^z$ is some interpolation between (i) and (ii)
as in the block-RG procedures of \cite{c_emergent}, and thus cannot be called 'paramagnetic' or 'spin-glass' in itself.
Let us now return to the whole Hamiltonian $H_1$ of Eq. \ref{h1c} : the $2^{K+1}$ energy-levels labelled by the variables $(\tau_1^z,...\tau_{K+1}^z )$
\begin{eqnarray}
E_{1}^{(\tau_1^z...\tau_{K+1}^z ) } = \sum_{i_1=1}^{K+1} \tau_{i_1}^z \sqrt{ J_{i_1}^2 + h_{i_1}^2 }
\label{e1level}
\end{eqnarray}
are independent of $S_0^x=\pm 1$.
To label this degeneracy, it is thus convenient to introduce the renormalized spin $\sigma_{R0}^x$
\begin{eqnarray}
\vert \tau_1^z...\tau_{K+1}^z ; \sigma_{R0}^x=+1\rangle && =\vert S_{0}^x=+1 \rangle
\otimes_{i_1=1}^{K+1} \vert \lambda^{(\tau_{i_1}^z) }_{i_1}(S_0^x=+1) \rangle
\nonumber \\
\vert \tau_1^z...\tau_{K+1}^z ; \sigma_{R0}^x=-1 \rangle && =\vert S_{0}^x=-1 \rangle
\otimes_{i_1=1}^{K+1} \vert \lambda^{(\tau_{i_1}^z) }_{i_1}(S_0^x=-1) \rangle
\label{tausr}
\end{eqnarray}
The projector onto the energy-level $ E_{1}^{(\tau_1^z...\tau_{K+1}^z ) }$ then reads
\begin{eqnarray}
P_{1}^{(\tau_1^z...\tau_{K+1}^z ) } =
\vert \tau_1^z...\tau_{K+1}^z ; \sigma_{R0}^x=+1\rangle \langle \tau_1^z...\tau_{K+1}^z ; \sigma_{R0}^x=+1 \vert
+ \vert \tau_1^z...\tau_{K+1}^z ; \sigma_{R0}^x=-1 \rangle \langle \tau_1^z...\tau_{K+1}^z ; \sigma_{R0}^x=-1 \vert
\label{projector}
\end{eqnarray}
The projection onto the energy-level $ E_{1}^{(\tau_1^z...\tau_{K+1}^z ) }$
of the Hamiltonian of Eq \ref{htree} concerning the whole tree can be obtained from the various contributions
\begin{eqnarray}
P_{1}^{(\tau_1^z...\tau_{K+1}^z ) } H P_{1}^{(\tau_1^z...\tau_{K+1}^z ) }&& = \sum_{r=0}^L P_{1}^{(\tau_1^z...\tau_{K+1}^z ) } H_r P_{1}^{(\tau_1^z...\tau_{K+1}^z ) }
\label{htreeproj}
\end{eqnarray}
The projection of $H_1$ is simply the energy $ E_{1}^{(\tau_1^z...\tau_{K+1}^z ) }$ by construction
\begin{eqnarray}
P_{1}^{(\tau_1^z...\tau_{K+1}^z ) } H_1 P_{1}^{(\tau_1^z...\tau_{K+1}^z ) } && = E_{1}^{(\tau_1^z...\tau_{K+1}^z ) }
\label{h1treeproj}
\end{eqnarray}
while the projection of $H_{r }$ is unchanged for $r \geq 3$
\begin{eqnarray}
P_{1}^{(\tau_1^z...\tau_{K+1}^z ) } H_{r } P_{1}^{(\tau_1^z...\tau_{K+1}^z ) }&& = H_r
\label{hrtree}
\end{eqnarray}
The projection of $H_0$
\begin{eqnarray}
P_{1}^{(\tau_1^z...\tau_{K+1}^z ) } H_0 P_{1}^{(\tau_1^z...\tau_{K+1}^z ) } && = h_0 P_{1}^{(\tau_1^z...\tau_{K+1}^z ) } \sigma_0^z P_{1}^{(\tau_1^z...\tau_{K+1}^z) }
\nonumber \\
&& = h_0
\left( \prod_{i_1=1}^{K+1} \sqrt{\frac{h_{i_1}^2 }{ J_{i_1}^2 + h_{i_1}^2} } \right)
P_{1}^{(\tau_1^z...\tau_{K+1}^z ) } \sigma_{R0}^z P_{1}^{(\tau_1^z...\tau_{K+1}^z ) }
\label{h0tproj}
\end{eqnarray}
gives the renormalized transverse field $h^R_{0} $ associated to the renormalized spin operator $\sigma_{R0}^z $
\begin{eqnarray}
h^R_{0} = h_0 \prod_{i_1=1}^{K+1} \sqrt{\frac{h_{i_1}^2 }{ J_{i_1}^2 + h_{i_1}^2} }
\label{hr0}
\end{eqnarray}
The projection of $H_2$
\begin{eqnarray}
P_{1}^{(\tau_1^z...\tau_{K+1}^z ) }H_2 P_{1}^{(\tau_1^z...\tau_{K+1}^z ) }
&& = \sum_{i_1=1}^{K+1} \sum_{i_2=1}^{K} \left( J_{i_1,i_2}
(P_{1}^{(\tau_1^z...\tau_{K+1}^z ) } \sigma_{i_1}^x P_{1}^{(\tau_1^z...\tau_{K+1}^z ) }) \sigma_{i_1,i_2}^x+ h_{i_1,i_2} \sigma_{i_1,i_2}^z
\right)
\nonumber \\
&& = \sum_{i_1=1}^{K+1} \sum_{i_2=1}^{K} \left( J_{i_1,i_2} \frac{\tau_{i_1}^z J_{i_1} }{\sqrt{ J_{i_1}^2 + h_{i_1}^2 }}
(P_{1}^{(\tau_1^z...\tau_{K+1}^z ) } \sigma_{R0}^x P_{1}^{(\tau_1^z...\tau_{K+1}^z ) }) \sigma_{i_1,i_2}^x+ h_{i_1,i_2} \sigma_{i_1,i_2}^z
\right)
\label{h2proj}
\end{eqnarray}
gives the renormalized coupling between the operators $\sigma_{R0}^x $ and $ \sigma_{i_1,i_2}^x$
\begin{eqnarray}
J_{i_1,i_2}^R = J_{i_1,i_2} \frac{\tau_{i_1}^z J_{i_1} }{ \sqrt{ J_{i_1}^2 + h_{i_1}^2 } }
\label{jr12}
\end{eqnarray}
\subsection { RG rules }
The iteration of the above procedure yields the following RG rules after $r$ RG steps.
The renormalized transverse field $h^{R^r}_{0} $ associated to the renormalized spin operator $\sigma_{R^r0}^z $
evolves
according to
(Eq \ref{hr0})
\begin{eqnarray}
h_0^{R^r} = h_0^{R^{n-1}} \prod_{i_1=1}^{K+1} \prod_{i_2=1}^{K} .. \prod_{i_r=1}^{K}
\sqrt{ \frac{ h_{i_1,..,i_r}^2 }
{ h_{i_1,..,i_r}^2 + [ J_{i_1,..,i_r}^{R^{r-1}} ]^2 } }
\label{rgh}
\end{eqnarray}
while the renormalized coupling between the operators $\sigma_{R^r0}^x $ and $ \sigma_{i_1,i_2,..,i_{r+1}}^x$ reads (Eq \ref{jr12})
\begin{eqnarray}
J_{i_1,..,i_{r+1}}^{R^r} = J_{i_1,...,i_{r+1}} \frac{\tau_{i_1,..,i_r}^z J_{i_1,..,i_{r}}^{R^{r-1}} }
{ \sqrt{ h_{i_1,..,i_r}^2 + [ J_{i_1,..,i_r}^{R^{r-1}} ]^2 } }
\label{rgj}
\end{eqnarray}
\subsection { Solution of the RG rules }
The RG rule of Eq. \ref{rgj} for the couplings involve only the initial transverse fields
and not the renormalized transversed fields, so that it can be solved independently.
The sign
\begin{eqnarray}
{\rm sgn}( J_{i_1,..,i_{r+1}}^{R^r} ) && = \tau_{i_1,..,i_r}^z {\rm sgn}( J_{i_1,...,i_{r+1}} ) {\rm sgn} ( J_{i_1,..,i_{r}}^{R^{r-1}} )
\nonumber \\
&& = \tau_{i_1,..,i_r}^z \tau_{i_1,..,i_{r-1}}^z ... \tau_{i_1}^z
{\rm sgn}( J_{i_1,...,i_{r+1}} ) {\rm sgn}( J_{i_1,...,i_{r}} ) ... {\rm sgn}( J_{i_1} )
\label{rgjsgn}
\end{eqnarray}
is simply the product of all the couplings $ J$ and of all the variables$\tau^z$ along the path between
the sites $0$ and $(i_1,..,i_r)$.
The absolute value reads (Eq. \ref{rgj})
\begin{eqnarray}
\vert J_{i_1,..,i_{r+1}}^{R^r} \vert && =\vert J_{i_1,...,i_{r+1}} \vert C_{i_1,..,i_r}
\label{rgjabs}
\end{eqnarray}
where
\begin{eqnarray}
C_{i_1,..,i_r} && \equiv \left[ 1+\sum_{m=1}^r \prod_{k=m}^r \frac{h_{i_1,..,i_k}^2 }{J_{i_1,..,i_k}^2} \right]^{-\frac{1}{2}}
\nonumber \\
&& = \left[ 1+\frac{h_{i_1,..,i_r}^2 }{J_{i_1,..,i_r}^2} + \frac{h_{i_1,..,i_r}^2 h_{i_1,..,i_{r-1}}^2 }{J_{i_1,..,i_r}^2 J_{i_1,..,i_{r-1}}^2} + .. + \frac{h_{i_1,..,i_r}^2 h_{i_1,..,i_{r-1}}^2..h_{i_1,i_2}^2 h_{i_1}^2 }{J_{i_1,..,i_r}^2 J_{i_1,..,i_{r-1}}^2.. J_{i_1,i_2}^2 J_{i_1}^2}
\right]^{-\frac{1}{2}}
\label{kesten}
\end{eqnarray}
involves in the denominator a so-called Kesten random variable \cite{Kesten,Der_Pom,Bou,Der_Hil,Cal}
that has been much studied in relation with the surface magnetization in the ground-state of the one-dimensional chain
\cite{strong_review,c_microcano,us_watermelon}.
This solution for the renormalized couplings can be plugged into the RG flow of Eq. \ref{rgh}
for the renormalized transverse field to obtain
\begin{eqnarray}
\ln \left( \frac{h_0^{R^r} }{ h_0^{R^{r-1}} } \right) && = \sum_{i_1=1}^{K+1} \sum_{i_2=1}^{K} .. \sum_{i_r=1}^{K}
\ln \left( \frac{ 1 }
{ \sqrt{ 1 +\frac{ J^2_{i_1,..,i_r} }{h_{i_1,..,i_r}^2} C_{i_1,..,i_{r-1}}^2 } } \right)
\nonumber \\
&& = \frac{1}{2} \sum_{i_1=1}^{K+1} \sum_{i_2=1}^{K} .. \sum_{i_r=1}^{K}
\ln \left( 1- C_{i_1,..,i_{r-1},i_r}^2 \right)
\label{rghbis}
\end{eqnarray}
in terms of the Kesten variables of Eq. \ref{kesten}.
\subsection { Reminder on the one-dimensional chain $K=1$ }
For the one-dimensional chain corresponding to $K=1$, the location of paramagnetic/spin-glass quantum phase transition
for the ground state of the quantum Ising model is know to occur
exactly at
\begin{eqnarray}
{\rm Critical \ Point \ in \ one \ dimension : } \ \ \ \overline{\ln \vert J_i \vert } = \overline{\ln \vert h_i \vert}
\label{criti1d}
\end{eqnarray}
as a consequence of self-duality \cite{pfeuty,fisher,strong_review}.
The corresponding Strong Disorder Fixed Point \cite{fisher} is characterized in particular by the activated exponent
\begin{eqnarray}
\psi^{(d=1)} = \frac{1}{2}
\label{psi1d}
\end{eqnarray}
and by the two correlation length exponents
\begin{eqnarray}
\nu^{(d=1)}_{typ} && =1
\nonumber \\
\nu^{(d=1)}_{av} && = 2
\label{nu1d}
\end{eqnarray}
As discussed in \cite{c_emergent}, the phase transition between the Paramagnetic and Spin-glass
Many-Body-Localized phases for the excited eigenstates is the same as the ground state quantum phase transition just described,
and the above renormalization procedure is able to reproduce the exact transition location of Eq. \ref{criti1d} and the exact critical exponents of
Eqs \ref{psi1d} and \ref{nu1d}, together with the exact surface magnetization in terms of Kesten variables as already mentioned above (Eq \ref{kesten}).
\subsection { Solution at lowest order in the couplings for the Cayley tree with branching ratio $K>1$}
We have just recalled that in one dimension, the transition occurs when the typical coupling and the typical fields are equal (Eq. \ref{criti1d}).
For the Cayley tree with branching ratio $K>1$, the transition is thus expected to occur in the region
\begin{eqnarray}
\overline{\ln \vert J_i \vert } < \overline{\ln \vert h_i \vert}
\label{jtypsmaller}
\end{eqnarray}
where the couplings are typically smaller than the transverse fields.
To analyse the RG rules in this region, it is convenient to introduce the products
\begin{eqnarray}
P_{i_1,..,i_r} && \equiv \left \vert \frac{J_{i_1,..,i_r} J_{i_1,..,i_{r-1}}.. J_{i_1,i_2} J_{i_1}}{h_{i_1,..,i_r} h_{i_1,..,i_{r-1}}..h_{i_1,i_2} h_{i_1} }
\right \vert
\label{product}
\end{eqnarray}
In the region of Eq. \ref{jtypsmaller}, the Kesten variable of the denominator in Eq. \ref{kesten} is dominated by the last term,
while it is convenient to keep the term unity to maintain the important bound $C_{i_1,..,i_r } \leq 1$,
so that we make the following approximation at lowest order in the couplings
\begin{eqnarray}
C_{i_1,..,i_r} && \simeq \left[ 1+ \frac{1}{P^2_{i_1,..,i_r}} \right]^{-\frac{1}{2}}
= \frac{P_{i_1,..,i_r} }{\sqrt{ 1+P^2_{i_1,..,i_r} } }
\label{kestenlow}
\end{eqnarray}
Then the absolute values of the renormalized couplings of Eq \ref{rgjabs} become
\begin{eqnarray}
\vert J_{i_1,..,i_{r+1}}^{R^r} \vert && = \vert J_{i_1,...,i_{r+1}} \vert \frac{P_{i_1,..,i_r} }{\sqrt{ 1+P^2_{i_1,..,i_r} }}
\label{jrlow}
\end{eqnarray}
For the ground state, the result $\vert J_{i_1,...,i_{n+1}} \vert P_{i_1,..,i_r} $ (i.e. without the denominator $\sqrt{1+P^2_{i_1,..,i_r} }$)
that involves the product of all couplings in the numerator and all the transverse fields in the denominator
has been obtained in the paramagnetic phase via various approaches including the Cavity-Mean-Field approach \cite{cavity1,cavity2,cavity3},
the Strong Disorder RG framework when only sites are decimated \cite{us_boundarycayley} or simply
lowest perturbation theory in the couplings \cite{us_transverseDP}.
The approximation of Eq \ref{kestenlow}
yields that the RG flow of Eq. \ref{rghbis}
for the renormalized transverse field becomes
\begin{eqnarray}
\ln \left( \frac{h_0^{R^r} }{ h_0^{R^{r-1}} } \right)
&& \simeq - \frac{1}{2} \sum_{i_1=1}^{K+1} \sum_{i_2=1}^{K} .. \sum_{i_r=1}^{K}
\ln \left( 1+ P_{i_1,..,i_{r-1},i_r}^2 \right)
\label{hrlow}
\end{eqnarray}
To analyse the statistical properties of the RG flows Eq \ref{jrlow} and Eq \ref{hrlow},
one needs first to characterize the large deviation properties of the products of Eq. \ref{product}.
\section{ Large deviation analysis }
\label{sec_pr}
In this section, we describe the statistical properties of the product of Eq. \ref{product}
with the simplified notation
\begin{eqnarray}
P(r) &&= \prod_{k=1}^r \left \vert \frac{J_{i_1,..,i_k} }{h_{i_1,..,i_k} } \right \vert
\label{cradial}
\end{eqnarray}
where $r$ represents the number of random variables $\left \vert \frac{J_{i_1,..,i_k} }{h_{i_1,..,i_k} } \right \vert $ in this product.
\subsection{ Typical behavior }
The logarithm of Eq. \ref{cradial}
reduces to a sum of random variables
\begin{eqnarray}
\ln P(r) && \simeq \sum_{k=1}^r (\ln \vert J_{i_1,..,i_k} \vert - \ln \vert h_{i_1,..,i_k} \vert )
\label{lnpr}
\end{eqnarray}
The Central Limit Theorem thus yields the following typical behavior for large $r$
\begin{eqnarray}
\ln P(r) && \mathop{\simeq}_{r \to +\infty} - r a_{0}+ \sqrt{r} u
\label{clt}
\end{eqnarray}
where
\begin{eqnarray}
a_{0} = \overline{ (\ln \vert h_{i} \vert - \ln \vert J_{i} \vert ) }
\label{a0}
\end{eqnarray}
is positive $a_0>0$ in the region under study (Eq. \ref{jtypsmaller}) and
governs the typical exponential decay of $P(r)$, while $u$ is a Gaussian random variable.
For the one-dimensional chain, only this typical behavior is relevant,
but here on the Cayley tree of branching ratio $K>1$
where the number of sites at distance $r$ grows exponentially as $K^r$ with the distance $r$,
one needs to analyze the large deviations properties.
\subsection{ Large deviations }
In the field of large deviations (see the review \cite{touchette} and references therein),
one is interested into the exponentially small probability
to see an exponential decay with some coefficient $a $ different from the typical value $a_0$ of Eq. \ref{a0}
\begin{eqnarray}
{\rm Prob} ( P(r) \propto e^{-a r} ) && \oppropto_{r \to +\infty} e^{-r I(a)}
\label{largedev}
\end{eqnarray}
where the rate function $I(a)$ vanishes at the typical value $a_0$ (Eq \ref{a0})
\begin{eqnarray}
I(a_0)=0
\label{largedevtyp}
\end{eqnarray}
and is strictly positive otherwise $I(a \ne a_0) >0 $.
The standard way to evaluate the rate function $I(a)$ is to consider
the generalized moments that display the following exponential behavior \cite{touchette}
\begin{eqnarray}
\overline{ P^{2q} (r) } && = \left( \overline{ \frac{\vert J_i \vert ^{2q} }{\vert h_i \vert ^{2q} } } \right)^r = e^{ r \lambda(q)}
\label{c2q}
\end{eqnarray}
where
\begin{eqnarray}
\lambda(q) && = \ln \left( \overline{ \frac{\vert J_i \vert ^{2q} }{\vert h_i \vert ^{2q} } } \right)
\label{lambdaq}
\end{eqnarray}
can be explicitly computed from the probability distribution of the couplings $J_i$ and of the random fields $h_i$
(see the example below).
The evaluation of Eq. \ref{c2q} via the saddle-point approximation
\begin{eqnarray}
\overline{ P^{2q} (r) } && \simeq \int_0^{+\infty} da e^{-r I(a)} e^{-a r 2 q } = e^{r \left( \displaystyle \max_a ( - I(a) - 2q a ) \right) }
\label{saddle}
\end{eqnarray}
yields $\lambda_q$ in terms of the saddle-point $a_q$
\begin{eqnarray}
\lambda(q) && = - I(a_q) - 2 q a_q
\nonumber \\
0 && = I'(a_q) + 2 q
\label{legendre}
\end{eqnarray}
The reciprocal Legendre transform yields
\begin{eqnarray}
I(a) && = - \lambda(q_a) - 2 a q_a
\nonumber \\
0 && = \lambda'(q_a) + 2 a
\label{legendrereciproque}
\end{eqnarray}
\subsection{ Explicit example with the two box distributions of Eq. \ref{box} }
Let us now focus on the example where the probability distributions of the couplings and of the random fields
are the two box distributions of parameters $J$ and $h$ respectively (Eq \ref{box}).
In the region $h>J$, the typical decay of the renormalized couplings is governed by (Eq. \ref{a0})
\begin{eqnarray}
a_{0} =\int_0^h \frac{dh_i}{h} \ln h_i - \int_0^J \frac{dJ_i}{J} \ln J_i = \ln \left( \frac{h}{J} \right) >0
\label{a0box}
\end{eqnarray}
The generalized moments of Eq. \ref{c2q}
converge only in the region $-1<2q<1$ and Eq. \ref{lambdaq} becomes
\begin{eqnarray}
e^{ \lambda(q)} && = \overline{ \frac{\vert J_i \vert ^{2q} }{\vert h_i \vert ^{2q} } } = \int_0^J \frac{dJ_i}{J} J_i^{2q} \int_0^h \frac{dh_i}{h} h_i^{-2q}
= \frac{1}{1- 4 q^2} \left( \frac{J }{h} \right)^{2q} = \frac{1}{1- 4 q^2} e^{-2q a_0}
\label{lqbox}
\end{eqnarray}
so that the function $\lambda(q)$ and its derivative read in terms of the typical value $a_0$
\begin{eqnarray}
\lambda(q) && = - 2q a_0 - \ln ( 1- 4 q^2)
\nonumber \\
\lambda'(q) && = - 2 a_0 + \frac{8q}{ 1- 4 q^2}
\label{lambdabox}
\end{eqnarray}
The second equation of the system \ref{legendrereciproque}
\begin{eqnarray}
0 && = 2a+ \lambda'(q_a) = 2 (a-a_0) + \frac{8q_a}{ 1- 4 q_a^2}
\label{eq2box}
\end{eqnarray}
leads to the following second-order equation for $q_a$
\begin{eqnarray}
0 && = q_a^2 - \frac{q_a}{ a-a_0} - \frac{1}{4}
\label{eqqa}
\end{eqnarray}
The appropriate solution $q_a$ that tends to $ q_a \to 0$ when $a \to a_0$ reads
\begin{eqnarray}
q_a = \frac{a_0-a}{2( 1+\sqrt{1+(a_0-a)^2} )}
\label{soluqa}
\end{eqnarray}
The rate function given by the first equation of the system \ref{legendrereciproque} reads
\begin{eqnarray}
I(a) && = - \lambda(q_a) - 2 a q_a = 2q_a (a_0-a) + \ln ( 1- 4 q_a^2)
= 2q_a (a_0-a) + \ln (\frac{ 4 q_a}{ a_0-a} )
\nonumber \\
&& =\frac{(a_0-a)^2 }{ 1+\sqrt{1+(a_0-a)^2} } - \ln \left( \frac{ 1+\sqrt{1+(a_0-a)^2} }{2} \right)
\label{ratebox}
\end{eqnarray}
\section{ Statistical properties of the renormalized couplings }
\label{sec_jr}
In this section, we focus on the absolute values of the renormalized couplings given by Eq \ref{jrlow}
\begin{eqnarray}
\vert J_{i_1,..,i_{r+1}}^{R^r} \vert && = \vert J_{i_1,...,i_{r+1}} \vert \frac{P_{i_1,..,i_r} }{\sqrt{ 1+P^2_{i_1,..,i_r} }}
\label{jrlows}
\end{eqnarray}
\subsection{ Location of the critical point }
On the Cayley tree where the number of points at distance $r$ grows exponentially as $K^r$,
the number of products $P(r)$ displaying the decay $P(r)\propto e^{-a r}$ reads (Eq. \ref{largedev})
\begin{eqnarray}
{\cal N} ( P(r) \propto e^{-a r} ) && \oppropto_{r \to +\infty} K^r e^{-r I(a)} = e^{r (\ln K - I(a) )} \theta(a_{min} \leq a \leq a_{max})
\label{largedevN}
\end{eqnarray}
where the minimum value $a_{min}$ and the maximal value $a_{max}$ are respectively smaller and bigger than the typical value
$a_{min}<a_0<a_{max}$ and satisfy
\begin{eqnarray}
I(a_{min}) = \ln K = I(a_{max})
\label{amin}
\end{eqnarray}
so that they occur only on a finite number $O(1)$ of branches,
while the typical value $a_0$ where $I(a_0)=0$ occur on an extensive $O(K^n)$ number of branches.
From Eq \ref{jrlows}, it is clear that the renormalized coupling $J(r)$ inherits the exponential decay of $P(r) $ of Eq. \ref{largedevN}
as long as $a>0$, while the region $a \leq 0$ produces finite renormalized couplings $O(1)$
so that the critical point corresponds to the vanishing of the minimal value $a_{min}$
\begin{eqnarray}
a_{min}^{criti}=0
\label{amincriti}
\end{eqnarray}
or equivalently in terms of the large deviation function $I(a)$ ( Eq. \ref{amin})
\begin{eqnarray}
I^{criti}(0) = \ln K
\label{criti}
\end{eqnarray}
For the special case of the box distribution of Eq \ref{box}, Eq \ref{ratebox}
yields the following explicit condition
in terms of the control parameter $a_0= \ln \frac{h}{J}$
\begin{eqnarray}
0 && = \ln \left( K \frac{ 1+\sqrt{1+(a^{criti}_0)^2} }{2} \right)- \frac{(a_0^{criti})^2 }{ 1+\sqrt{1+(a_0^{criti})^2} }
\label{critibox}
\end{eqnarray}
\subsection{ Paramagnetic phase for $a_{min}>0$ }
In the paramagnetic phase $a_{min}>0$, all $K^r$ renormalized couplings decay exponentially with $a \geq a_{min}>0$ (Eq \ref{largedevN})
\begin{eqnarray}
{\cal N} ( J(r) \propto e^{-a r} ) && \oppropto_{r \to +\infty} e^{r (\ln K - I(a) )} \theta(a_{min} \leq a \leq a_{max})
\label{largedevNj}
\end{eqnarray}
\subsection{ Spin-Glass phase for $a_{min}<0$ }
In the spin-glass phase $a_{min}<0$, the $K^r$ renormalized couplings can be split into two groups :
the number of finite couplings grows exponentially in $r$ as
\begin{eqnarray}
{\cal N} ( J(r) \propto O(1) ) && \oppropto_{r \to +\infty} \int_{a_{min}}^0 da e^{r (\ln K - I(a) )} \simeq e^{r (\ln K - I(0) )}
= e^{r (I(a_{min}) - I(0) )}
\label{largedevNjfinite}
\end{eqnarray}
while the other branches are still characterized by exponential decays with exponents $a>0$
\begin{eqnarray}
{\cal N} ( J(r) \propto e^{-a r} ) && \oppropto_{r \to +\infty} e^{r (\ln K - I(a) )} \theta(0< a \leq a_{max})
\label{largedevNjp}
\end{eqnarray}
This is the first indication that the ordered spin-glass cluster remains very sparse near the critical point,
as confirmed by the analysis of the renormalized transverse field in the next section.
\section{ Statistical properties of the renormalized transverse field }
\label{sec_hr}
In this section, we focus on the RG flow of Eq. \ref{hrlow} for the renormalized transverse field
\begin{eqnarray}
\ln \left( \frac{h_0^{R^r} }{ h_0^{R^{r-1}} } \right)
&& \simeq - \frac{1}{2} \sum_{i_1=1}^{K+1} \sum_{i_2=1}^{K} .. \sum_{i_r=1}^{K}
\ln \left( 1+ P_{i_1,..,i_{r-1},i_r}^2 \right)
\label{hrlowsec}
\end{eqnarray}
which can be evaluated in terms of the large deviation analysis of Eq. \ref{largedevN}
concerning the $K^r$ products $P(r)$
\begin{eqnarray}
\ln \left( \frac{h_0^{R^r} }{ h_0^{R^{r-1}} } \right)
&& \mathop{\simeq}_{r \to +\infty} - \frac{1}{2} \int_{a_{min}}^{a_{max} } da e^{r (\ln K - I(a) )}
\ln \left( 1+ e^{-2ar} \right)
\label{hrlowlargedev}
\end{eqnarray}
\subsection{ Paramagnetic phase for $a_{min}>0$ }
In the paramagnetic phase $a_{min}>0$, Eq. \ref{hrlowlargedev} becomes
\begin{eqnarray}
\ln \left( \frac{h_0^{R^r} }{ h_0^{R^{r-1}} } \right)
&& \mathop{\simeq}_{r \to +\infty} - \frac{1}{2} \int_{a_{min}}^{a_{max} } da e^{r (\ln K - I(a) -2a )}
= - \frac{1}{2} \int_{a_{min}}^{a_{max} } da e^{r (I(a_{min}) - I(a) -2a )}
\label{hrlowlargedevpara}
\end{eqnarray}
The integral is dominated by the lower boundary $a_{min}$ of the integral, and one obtains the exponential decay
\begin{eqnarray}
\ln \left( \frac{h_0^{R^r} }{ h_0^{R^{r-1}} } \right)
&& \oppropto_{r \to +\infty} - e^{-2 a_{min} r }
\label{hrlowlargedevparaexp}
\end{eqnarray}
By integration, one obtains that $h_0^{R^r} $ remains finite as $r \to +\infty$
\begin{eqnarray}
\ln \left( \frac{h_0^{R^r} }{ h_0 } \right)
&& \oppropto_{r \to +\infty} - \int_{1}^{r} dr' e^{ -2a_{min} r'} \oppropto_{r \to +\infty} - \frac{1-e^{-2 a_{min} r } }{ a_{min} }
\label{hrlowlargedevparafinite}
\end{eqnarray}
The typical asymptotic value $h_0^{R^{\infty}} $ for the renormalized transverse field
diverges with the following essential singularity near the transition $a_{min} \to a_{min}^{criti}=0$
\begin{eqnarray}
\ln \left( \frac{h_0^{R^{\infty}} }{ h_0 } \right) && \oppropto_{a_{min} \to 0} - \frac{1}{ a_{min} }
\label{essential}
\end{eqnarray}
\subsection{ Spin-Glass phase for $a_{min}<0$ }
In the spin-glass phase $a_{min}<0$,
it is convenient to evaluate separately the contributions of the two regions $a<0$ and $a>0$ in
the integral of Eq. \ref{hrlowlargedev}.
The contribution of the region $a>0$ is dominated by the lower boundary $a=0$ of the integral
\begin{eqnarray}
\int_{0}^{a_{max} } da e^{r (\ln K - I(a) )} \ln \left( 1+ e^{-2ar} \right)
\mathop{\simeq}_{r \to +\infty} \int_{0}^{a_{max} } da e^{r (\ln K - I(a)-2a )} \simeq e^{r (\ln K - I(0) )} = e^{r (I(a_{min}) - I(0) )}
\label{hrlowlargedevp}
\end{eqnarray}
corresponding to an exponentially growing term.
The region $a<0$
\begin{eqnarray}
\int_{a_{min}}^{0 } da e^{r (\ln K - I(a) )} \ln \left( 1+ e^{-2ar} \right)
\mathop{\simeq}_{r \to +\infty} \int_{a_{min}}^{0 } da e^{r (\ln K - I(a) )} (-2ar)
\label{hrlowlargedevm}
\end{eqnarray}
is dominated by the upper boundary $a=0$.
So the RG flow of renormalized transverse field of Eq. \ref{hrlowlargedev} is dominated by the exponentially big term
of coefficient $(I(a_{min}) - I(0))>0$ of Eq. \ref{hrlowlargedevp}
\begin{eqnarray}
\ln \left( \frac{h_0^{R^r} }{ h_0 } \right)
&& \oppropto_{r \to +\infty}
- \int_{1}^{r} dr' e^{r' (I(a_{min}) - I(0) )} \oppropto_{r \to +\infty} - \frac{ e^{r (I(a_{min}) - I(0) )} }{(I(a_{min}) - I(0) )}
\label{hrlowlargedevsg}
\end{eqnarray}
\subsection{ Finite-size scaling in the critical region }
The above results for the renormalized transverse field as a function of the radial distance $r$ can be summarized by
the following finite-size scaling form in the critical region
\begin{eqnarray}
\ln \left( \frac{h_0^{R^r} }{ h_0 } \right)
&& \oppropto_{r \to +\infty}
- r^{\psi} G \left( r^{\frac{1}{\nu}} (J-J_c) \right)
\label{hrlowfss}
\end{eqnarray}
with the exponent
\begin{eqnarray}
\psi=1
\label{ps1}
\end{eqnarray}
and the correlation length exponent
\begin{eqnarray}
\nu=1
\label{nu1}
\end{eqnarray}
as in many other phase transitions on the Cayley tree.
The scaling function $G(x)$ is constant at the origin $G(0)=cst$, behaves as
\begin{eqnarray}
G(x) \oppropto_{x \to - \infty} - \frac{1}{x}
\label{gminfty}
\end{eqnarray}
to reproduce the behavior of Eq. \ref{essential} in the paramagnetic phase $J<J_c$, and as
\begin{eqnarray}
G(x) \oppropto_{x \to + \infty} \frac{e^x-1}{x}
\label{gpinfty}
\end{eqnarray}
to reproduce the behavior of Eq. \ref{hrlowlargedevsg} in the spin-glass phase $J>J_c$.
\subsection{ Number $N_{SG}$ of spins involved in this ordered spin-glass cluster }
The renormalized transverse field $h_0^{R^r} $ directly reflects the number $N_{SG}(r)$ of spins involved in this ordered spin-glass cluster
\begin{eqnarray}
\ln \left( \frac{h_0^{R^r} }{ h_0 } \right)
&& \oppropto_{r \to +\infty} - N_{SG}(r)
\label{hnsg}
\end{eqnarray}
In the paramagnetic phase, both remain finite as $r \to +\infty$.
In the spin-glass phase,
The behavior found in Eq. \ref{hrlowlargedevsg}
for the renormalized transverse field
thus confirms the indication of Eqs \ref{largedevNjfinite} and \ref{largedevNjp}
concerning the renormalized couplings : near the critical point, the
ordered spin-glass cluster remains very sparse.
More precisely, the number $N_{SG}$ of spins involved in this ordered spin-glass cluster grows exponentially with the distance $r$
\begin{eqnarray}
N_{SG} \propto e^{r (I(a_{min}) - I(0) )} = e^{r (\ln K - I(0) )}
\label{nsg}
\end{eqnarray}
but is only sub-extensive with respect to the total number of spins $N=K^r$
\begin{eqnarray}
N_{SG} \propto = e^{r (\ln K - I(0) )} =N^{\theta}
\label{nsgn}
\end{eqnarray}
in the whole region of the phase diagram where the continuously varying exponent
\begin{eqnarray}
\theta = 1- \frac{I(0)}{\ln K } = 1 - \frac{I(0)}{I(a_{min}) }
\label{theta}
\end{eqnarray}
remains in the interval
\begin{eqnarray}
\theta^{criti}=0 <\theta < 1 = \theta^{ext}
\label{thetac}
\end{eqnarray}
At criticality, the vanishing exponent $\theta^{criti}=0 $ corresponds to the logarithmic growth with respect to $N$ (Eq \ref{hrlowfss} and \ref{ps1})
\begin{eqnarray}
N_{SG}^{criti} \propto r = \frac{ \ln N }{ \ln K}
\label{nsgnln}
\end{eqnarray}
meaning that only a finite number of the branches sustain the spin-glass order.
The location where the spin-glass-ordered cluster becomes extensive $\theta^{ext}=1$ corresponds to
the vanishing of the large deviation rate function $I^{ext}(0)=0$, i.e. to the vanishing of the typical value $a_0^{ext}=0$
(Eq. \ref{largedevtyp}), i.e. to the location of the transition for the one-dimensional chain (Eq \ref{criti1d})
\begin{eqnarray}
a_0^{ext}=0= a_0^{criti1d}
\label{a0ext}
\end{eqnarray}
The finite region of the phase diagram corresponding to Eq. \ref{thetac} where the ordered spin-glass cluster remains sub-extensive
is somewhat { \it formally } reminiscent of the delocalized non-ergodic phase
existing in the Anderson Localization model defined on the Cayley tree
\cite{levitov,us_cayley,mirlin_tik}, i.e. in exactly the same geometry as in the present paper, and for the same technical reasons
based on large deviations on the branches of the Cayley tree \cite{levitov}.
But of course the physical meaning of the phases is completely different : in the Anderson Localization Model,
the three phases are Localized/Non-Ergodic-Delocalized/Ergodic-Delocalized,
while in the present study, the three phases are all MB-Localized, namely
Paramagnetic-MBL/SG-MBL with sub-extensive SG-cluster/SG-MBL with extensive SG-cluster.
Let us mention however that the existence of the intermediate delocalized non-ergodic phase remains very controversial
for the Anderson Localization model on Random Regular Graphs
\cite{biroli_nonergo,luca,mirlin_ergo,altshuler_nonergo,lemarie,ioffe}
or for Many-Body-Localization models \cite{grover,harrisMBL,c_entropy,c_mblrgeigen,santos},
where an analogy with the Anderson Localization transition in an Hilbert space of 'infinite dimensionality'
has been put forward \cite{levitov,gornyi_fock,vadim,us_mblaoki,gornyi},
while the properties of the delocalized non-ergodic phase can be explicitly computed in some random matrix models
\cite{kravtsov_rosen,biroli_rosen,ossipov_rosen,c_ww}.
For our present study, these results thus indicate that the intermediate SpinGlass-MBLocalized phase with sub-extensive SG-cluster
found here on the Cayley tree might not exist on other tree-like lattices like Random Regular Graphs.
\subsection { Physical meaning of the results }
The above results can be summarized as follows (see Figure \ref{figure}).
\subsubsection{ The two important control parameters }
The two important control parameters for the quantum Ising model on the Cayley tree of branching ratio $K$ are
(1) the typical value of Eq. \ref{a0} where the large deviation function $I(a)$ of Eq. \ref{largedev} vanishes $I(a_0)=0$
\begin{eqnarray}
a_{0} \equiv \overline{ (\ln \vert h_{i} \vert - \ln \vert J_{i} \vert ) }
\label{a0def}
\end{eqnarray}
(2) the minimum value $a_{min}$ defined as the smaller value $a_{min}<a_0$ where the large deviation function $I(a)$ of Eq. \ref{largedev}
takes the value
\begin{eqnarray}
I(a_{min}) = \ln K
\label{amindef}
\end{eqnarray}
\subsubsection{ The three possible MB-Localized phases }
(a) {\it MB-Localized Phase with extensive Spin-Glass Order }
In the region where the typical value $a_0$ of Eq. \ref{a0} is negative
\begin{eqnarray}
a_{0} <0
\label{a0order}
\end{eqnarray}
a typical one-dimensional chain would be spin-glass ordered (Eq. \ref{criti1d}), and thus the whole Cayley tree is also fully ordered
with an extensive spin-glass cluster with respect to the total number of spins $N=K^r$
\begin{eqnarray}
N_{SG} \propto K^r=N
\label{nsgorder}
\end{eqnarray}
(b) {\it Paramagnetic MB-Localized Phase }
In the region where the minimal value $a_{min}$ is positive
\begin{eqnarray}
0< a_{min}
\label{aminpara}
\end{eqnarray}
the drawing of $K^r$ independent random one-dimensional chains would produce only paramagnetic chains,
i.e. even the exponentially-rare best chain would be paramagnetic. Then the whole Cayley tree is also paramagnetic,
and the spin-glass cluster around the origin remains finite
\begin{eqnarray}
N_{SG} \propto O(1)
\label{nsgpara}
\end{eqnarray}
(c) {\it MB-Localized Phase with sub-extensive Spin-Glass Order }
\begin{figure}
\begin{center}
\includegraphics[width=16cm]{MBLfig.pdf}
\end{center}
\caption{ The two important control parameters for the phase diagram are the minimal value $a_{min}$ and the typical value $a_0 = \overline{ (\ln \vert h_{i} \vert - \ln \vert J_{i} \vert ) } $ where the large deviation function $I(a)$ takes respectively the values $I(a_{min})= \ln K$ and $I(a_0)=0$. The critical point $a_{min}=0$ corresponds to the phase transition between the MBL-Paramagnetic phase $a_{min}>0$ and the MBL-Spin-Glass phase $a_{min}<0$, where the Spin-Glass-Order remains sub-extensive $0<\theta=1- \frac{I(0)}{\ln K } <1$ in the whole region $a_{min}<0<a_0$ before becoming extensive in the region $a_0<0$.}
\label{figure}
\end{figure}
In the intermediate region
\begin{eqnarray}
a_{min} <0 < a_0
\label{amininter}
\end{eqnarray}
corresponding to
\begin{eqnarray}
I(a_{min})= \ln K > I(0) > I( a_0 )=0
\label{iamininter}
\end{eqnarray}
the drawing of $K^r$ independent random one-dimensional chains would produce
$K^r e^{- r I(0) )} $ spin-glass ordered chains, while the other (of order $K^r$) would be paramagnetic.
Then on the Cayley tree, the spin-glass cluster around the origin only contains $K^r e^{- r I(0) )} $ leaves out of the $K^r$.
So the size of the spin-glass cluster grows exponentially in $r$ but not as rapidly as $N=K^r$, so that it is subextensive
\begin{eqnarray}
N_{SG} \propto = e^{r (\ln K - I(0) )} =N^{\theta}
\label{nsgninter}
\end{eqnarray}
where the exponent
\begin{eqnarray}
\theta = 1- \frac{I(0)}{\ln K } = 1 - \frac{I(0)}{I(a_{min}) }
\label{thetabis}
\end{eqnarray}
varies continuously between $\theta^{criti}=0 $ [corresponding to $a_{min}=0$ where the transition towards (b) occurs]
and $\theta^{ext}=1 $ [corresponding to $I(0)=0$ i.e. $a_0=0$ where the transition towards (a) occurs].
\section{Conclusion }
\label{sec_conclusion}
We have introduced a simple Real-Space-Renormalization procedure in order to construct
the whole set of eigenstates for
the quantum Ising model with random couplings and random transverse fields on the Cayley tree of branching ratio $K$.
The analysis of the renormalization rules via large deviations was described to obtain the critical properties
of the phase transition between the paramagnetic and the spin-glass Many-Body-Localized phases.
In particular, we have found that the renormalized transverse field of the center site
involves the activated exponent $\psi=1$ and the correlation length exponent $\nu=1$.
The spin-glass-ordered cluster containing $N_{SG}$ spins was found to be extremely sparse with respect to the total number $N \propto K^r$ of spins : its size grows only logarithmically at the critical point $N_{SG}^{criti} \propto \ln N$, meaning that only a finite number $O(1)$ of the branches are long-ranged-ordered, while the other branches display exponentially decaying correlations.
In addition, the size $N_{SG} $ spin-glass-ordered cluster is sub-extensive $N_{SG} \propto N^{\theta}$ in the finite region of the spin-glass phase where the continuously varying exponent $\theta$ remains in the interval $0<\theta<1$.
As a final remark, let us mention that the mere existence of Many-Body-Localized phases in any dimension $d>1$ has been recently challenged
\cite{chandran-dg1,roeck-dg1,imbrie-dg1}, the same arguments being also used to claim the impossibility of mobility edges for MBL in $d=1$ \cite{bubbles}
(as opposed to the numerical phase-diagrams found in Ref. \cite{kjall,alet,luitz_tail,mondragon})
as well as the impossibility of MBL in the presence of power-law interactions \cite{roeck-dg1}
(as opposed to the works \cite{yao,burin_fss,burin_xy,hauke,gutman,moessner,hi,sondhi}).
It is thus essential to study various MBL models in various dimensions $d>1$
in order to solve the controversial issue about the influence of the dimension $d$.
Many-Body-Localized phases have been reported in dimension $d=2$ both numerically \cite{regnault2d} and experimentally \cite{mbl2d_exp},
as well as on Random Regular graphs \cite{mossi} or in the mean-field quantum random energy model \cite{qrem1,qrem2}.
We thus hope that the present work concerning Many-Body-Localized phases on the Cayley tree of effective infinite dimension $d=\infty$
will motivate future studies on this topic.
|
\section{Introduction}
The square-lattice Hubbard model has been the focus of intense theoretical research due to its simplicity and the fact that it demonstrates many of the phenomena that are associated with strongly-correlated electrons \cite{HubbardBook}. Furthermore, it is believed that in some sense, the physics of the high-temperature superconductors can be captured with the Hubbard model \cite{Hubbard_Cuprates_review}.
The Mott-Hubbard metal-insulator transition (MIT) \cite{MetalInsulatorReview}, the formation of a pseudogap \cite{Hubbard_Cuprates_review1, Pseudogap}, and the formation of Hubbard bands \cite{HubbardBands1,HubbardBands2} are all examples of strongly-correlated behavior that are expected to appear in the Hubbard model. In the case of the unfrustrated Hubbard model on the square-lattice, it is believed that strong, long-range antiferromagnetic fluctuations shift the MIT towards zero on-site coupling as the temperature is decreased towards zero \cite{FateMottHubbard}. Furthermore, these spin fluctuations evolve from being Slater-like at small to intermediate coupling to Heisenberg-like at large coupling.
More recent studies have extended the interactions in the Hubbard model beyond the on-site term and have included non-local correlations. These studies are closely related with the development of extended Dynamical Mean Field Theory (EDMFT) \cite{EDMFT1, EDMFT2} along with the EDMFT+$GW$ approach \cite{EDMFTGW}, where GW refers to the approximation of the self-energy by the first-order graph in which there appears one fermion line ($G$) and one screened interaction line ($W$) \cite{GWReview}. Among other phenomena, the inclusion of a long-range Coulomb interaction allows one to study collective charge excitations of the theory. These excitations, known as plasmons, can be accessed via the charge susceptibility $\chi_{\rho}({\bf q},\omega)$ which measures the system's response to a scalar potential $A_0({\bf q},\omega)$. Using the dual-boson approach, which goes beyond the EDMFT+GW approximation, it was found that at half-filling, the plasmons are characterized by a non-trivial dispersion relation \cite{vanLoon}. In this scenario, for a given strength of the Coulomb tail, at values of the on-site interaction $U$ close to the critical coupling for the metal-insulator transition, the plasmon dispersion separates into two branches as one approaches the edge of the Brillouin zone (BZ). It is argued that this feature can be viewed as a consequence of the formation of Hubbard bands.
Although this scenario seems plausible, it would be beneficial to have an independent, fully non-perturbative calculation. Due to the non-local nature of the Coulomb interaction, one can argue that existing methods may not in fact be accounting for all of the physics present in the system. A certain class of algorithms, going under the name of hybrid Monte Carlo (HMC) \cite{DuaneKogut}, are ideally suited for the calculations in strongly-correlated systems with non-local interactions \cite{ITEPRealistic}. Originally applied to the theory of the strong interactions, quantum chromodynamics (QCD), these methods have recently been applied successfully to certain condensed matter systems \cite{BuividovichUlybyshevReview,BuividovichPolikarpov,SmithVonSmekal,CCS0,CCS1, CCS2}.
Such a fully non-perturbative calculation can be used not only as an independent check of the paper \cite{vanLoon} but also as a benchmark for further improvements of EDMFT methods.
In this paper, we perform calculations for the square lattice extended Hubbard model at half-filling using an interaction which includes an on-site term as well as a long-range ``Coulomb-tail" defined by the value of the nearest-neighbor interaction. Using a lattice Monte Carlo setup, we compute the single particle Green's function as well as the charge density-density correlator in Euclidean time. We then use these observables to obtain the density of states (DOS) and the charge susceptibility by directly performing the numerical analytic continuation (NAC). In doing so, we introduce a completely robust and generalized variant of the Backus-Gilbert (BG) method \cite{BackusGilbertOrig} for performing NAC. This scheme has recently been applied in studies of spectral functions of lattice quantum chromodynamics \cite{BrandtBG} and graphene \cite{UlybyshevConductivity}. Here we introduce an improved BG scheme based on the method of Tikhonov regularization \cite{Tikhonov}.
The remainder of the article is organized in the following way. In section \ref{sec:Setup}, we state our conventions and introduce the lattice setup used to perform the calculations. In section \ref{sec:Observables}, we outline the calculation of the fermion Green's function and charge density-density correlator. In section \ref{sec:AC}, we introduce the Green-Kubo (GK) relations and discuss, in general terms, the problem of obtaining real-frequency information from Euclidean correlators. From there, we describe our method for obtaining spectral functions and make comparisons with other closely related approaches. In section \ref{sec:Results}, we present our results for the charge susceptibility and the DOS and attempt to make contact with previous work \cite{vanLoon}. Finally, in section \ref{sec:Conclusion}, we draw conclusions and propose directions for further investigation.
\section{\label{sec:Setup}Lattice Setup}
\subsection{Extended Hubbard Hamiltonian}
We start by introducing the following tight-binding Hamiltonian on the square lattice
\beq \label{SqHubbard}
\hat{\mathcal{H}}_{\text{tb}} = - \kappa \sum_{\sigma} \sum_{\left \langle x, y \right \rangle} \left( \hat{c}^{\dagger}_{x,\sigma} \hat{c}_{y,\sigma} + \text{h.c.} \right),
\eeq
where $\kappa$ is the nearest-neighbor hopping parameter and the sum $\sum_{\left \langle x,y \right \rangle}$ runs over all pairs of nearest neighbors. The creation and annihilation operators satisfy the following anticommutation relations
\beq \label{CommRelations}
\left\{ \hat{c}_{x,\sigma}, \hat{c}^{\dagger}_{y,\sigma'} \right\} = \delta_{x,y} \delta_{\sigma, \sigma'},
\eeq
where $x,y$ refer to the lattice site and $\sigma, \sigma'$ refer to the electron's spin.
We now make the following canonical transformation on the creation and annihilation operators of the up- and down-spin electrons
\beq \label{ParticleHole}
\hat{c}_{x, \uparrow}, \hat{c}^{\dagger}_{x, \uparrow} &\to & \hat{a}_x, \hat{a}^{\dagger}_x, \\
\hat{c}_{x, \downarrow}, \hat{c}^{\dagger}_{x, \downarrow} &\to & \pm \hat{b}^{\dagger}_x, \pm \hat{b}_x,
\eeq
where the $\pm$ refers to whether the site $x$ is ``even" or ``odd". The lattice site $x$ is ``even'' if $(-1)^{x_1+x_2}=1$ and ``odd'' otherwise.
Thus, we can write (\ref{SqHubbard}), after a constant shift, as
\beq
\hat{\mathcal{H}}_{\text{tb}} = - \kappa \sum_{\left \langle x, y \right \rangle} \left( \hat{a}^{\dagger}_x \hat{a}_y + \hat{b}^{\dagger}_x \hat{b}_y + \text{h.c.} \right).
\eeq
The Hilbert space of this tight-binding Hamiltonian can be constructed by first identifying the state satisfying $\hat{a}_x \ket{0} = \hat{b}_x \ket{0} = 0$ as the reference state. Thus, $\ket{0}$ corresponds to a state where each lattice site is occupied by one spin-down particle. The tight-binding Hamiltonian is written in momentum space as
\beq
\hat{\mathcal{H}}_{\text{tb}} = \sum_{{\bf k}} \epsilon_{{\bf k}} \left( \hat{a}^{\dagger}_{{\bf k}} \hat{a}_{{\bf k}} + \hat{b}^{\dagger}_{{\bf k}} \hat{b}_{{\bf k}} \right),
\eeq
where $\epsilon_{{\bf k}} \equiv -2\kappa \sum_{i=1,2} \cos(k_ia)$, with $a$ being the lattice spacing.
We now add two-body interactions between the electrons which are described by the term
\begin{equation}
\hat{\mathcal{H}}_{\text{int}} = \frac{1}{2}\sum_{x,y} \hat{\rho}_x V_{x,y}\hat{\rho}_y,
\end{equation}
where $V_{x,y}$ is a positive-definite potential matrix and $\hat{\rho}_x$ is the electric charge operator at $x$ site which is defined as follows
\beq
\hat{\rho}_x \to \hat{a}^{\dagger}_x \hat{a}_x - \hat{b}^{\dagger}_x \hat{b}_x.
\eeq
In our set up the matrix $V_{x,y}$ is defined completely by the on-site interaction (the Hubbard term) $U \equiv V_{0,0}$ and the nearest-neighbor interaction $V \equiv V_{(0,0), (0,1)} = V_{(0,0), (1,0)}$.
The latter coefficient characterizes the long-range $1/r$ Coulomb tail at any distance: $V_{x,y} = V/|\vec x-\vec y|,~x \neq y$, where the distance $|\vec x - \vec y|$ is dimensionless and evaluated in units of the lattice spacing. In order to obtain a positive-definite matrix $V_{x,y}$, these couplings must satisfy $U/V \gtrsim 1.5$. One also demands that the potential obeys periodic boundary conditions $V_{x+N_x,y} = V_{x,y+N_y} = V_{x,y}$, where $N_x$ and $N_y$ refer to the number of spatial lattice sites in the $x$ and $y$ directions. This slightly modifies the form of the potential relative to the infinite-volume form. Throughout this article, we will take $N_s \equiv N_x = N_y$.
\subsection{Path Integral Representation}
Following the approach of \cite{BuividovichPolikarpov,ITEPRealistic}, we start our construction of the path integral with the following Suzuki-Trotter decomposition of the partition function
\beq \label{BoltzmanFactor} \nn
Z &\equiv& {\rm Tr} \, e^{-\beta\left( \hat{\mathcal{H}}_{\text{tb}} + \hat{\mathcal{H}}_{\text{int}} \right)} = {\rm Tr} \, \left( e^{-\delta_{\tau} \left( \hat{\mathcal{H}}_{\text{tb}} + \hat{\mathcal{H}}_{\text{int}} \right)} \right)^{N_{\tau}} \\ &=& {\rm Tr} \, \left( e^{-\delta_{\tau} \hat{\mathcal{H}}_{tb}} e^{-\delta_{\tau} \hat{\mathcal{H}}_{\text{int}}} e^{-\delta_{\tau} \hat{\mathcal{H}}_{tb}} \dots \right) + O(\delta^2_{\tau}),
\eeq
where $\beta \equiv 1/T$ and $\delta_{\tau} \equiv \beta/N_{\tau}$ defines the step in Euclidean time. To compute the trace, we insert resolutions of the identity using the Grassmann variable coherent state representation
\beq
{\bm 1} = \int d\psi d\eta d {\bar \psi} d {\bar \eta} e^{-\sum_x ({\bar \psi}_x \psi_x + {\bar \eta}_x \eta_x )} \ket{\psi, \eta} \bra{\psi, \eta},
\eeq
\vspace{-0.75cm}
\beq
\ket{\psi, \eta} = e^{-\sum_x (\psi_x \hat{a}^{\dagger}_x +\eta_x \hat{b}^{\dagger}_x ) } \ket{0}.
\eeq
Matrix elements of the form $\bra{\psi', \eta'} e^{-\delta_{\tau} \hat{\mathcal{H}}_{\text{tb}}} \ket{\psi, \eta}$ can be evaluated using the following identity
\beq
\bra{\psi'} e^{\sum_{x,y} \hat{a}^{\dagger}_x A_{x,y} \hat{a}_y} \ket{\psi} = \exp\left( \sum_{x,y} {\bar \psi}'_x \left(e^A \right)_{x,y} \psi_y \right).
\eeq
In order to apply this identity to the case of $\hat{\mathcal{H}}_{\text{int}}$, one must first perform the Hubbard-Stratonovich \cite{Hubbard,Stratonovich} transformation in order to obtain a bilinear in the exponent
\beq \nn
\exp\left( - \tfrac{\delta_{\tau}}{2} \sum_{x,y} \hat{\rho}_x V_{x,y} \hat{\rho}_y \right) \simeq
\eeq
\vspace{-0.75cm}
\beq
\int \mathcal{D}\phi \exp\left( - \tfrac{\delta_{\tau}}{2} \sum_{x,y}\phi_x V^{-1}_{x,y} \phi_y -i \delta_{\tau} \sum_x \phi_x \hat{\rho}_x \right),
\eeq
where $\phi_x$ is a real scalar field living on each site of the lattice. Putting all of this together, one finally arrives at the path integral representation of the partition function given by
\beq \label{PathIntegral1}
Z = \int \mathcal{D}\phi \mathcal{D}\psi \mathcal{D}\eta \mathcal{D}{\bar \psi} \mathcal{D}{\bar \eta} e^{-\left(S_B[\phi] + {\bar \psi} M[\phi] \psi + {\bar \eta} \bar{M}[\phi] \eta \right)},
\eeq
where $S_B[\phi] = \tfrac{\delta_{\tau}}{2} \sum_{x,y,n}\phi_{x,n} V^{-1}_{x,y} \phi_{y,n}$ is the action of the Hubbard field and $n = 0,1,\dots,2N_{\tau}-1$ labels the factors of the identity that were inserted in (\ref{BoltzmanFactor}). We note that the Grassmann variables satisfy anti-periodic boundary conditions in Euclidean time. The fermionic operator $M$ is defined as follows
\beq \label{FermionAction} \nn
&& \sum_{x,y;\tau,\tau'} \bar{\psi}_{x,\tau} M_{x,y; \tau, \tau'} \psi_{y,\tau'} = \\ && \nn \sum^{N_{\tau}-1}_{k=0} \biggl[ \sum_x \bar{\psi}_{x,2k} \left( \psi_{x,2k} - \psi_{x,2k+1} \right) \\ && \nn - \delta_{\tau} \kappa \sum_{\left \langle x, y \right \rangle} \left( \bar{\psi}_{x,2k} \psi_{y, 2k+1} + \bar{\psi}_{y,2k} \psi_{x,2k+1} \right) \\ && \nn+ \sum_x \bar{\psi}_{x,2k+1} \left( \psi_{x,2k+1} - e^{-i\delta \phi_{x,k}}\psi_{x,2k+2} \right) \biggr],
\eeq
where we have used the approximation $\exp\left( -\delta_{\tau} \hat{\mathcal{H}}_{\text{tb}} \right) \approx {\bm 1} - \delta_{\tau} \hat{\mathcal{H}}_{\text{tb}}$. The second fermionic term in the action is constructed using the relation $\bar{M} \equiv M^*$. It has been shown that the discretization errors present in the action (\ref{FermionAction}) are $O(\delta_{\tau})$ \cite{SmithVonSmekal}.
The integration over the Grassmann variables in (\ref{PathIntegral1}) can be performed to obtain the following form of the partition function
\begin{equation} \label{PathIntegral2}
Z = \int \mathcal{D}\phi |\det M[\phi]|^2 e^{-S_B[\phi]},
\end{equation}
where we have used the identity
\beq
\det M[\phi] \det \bar{M}[\phi] = |\det M[\phi]|^2,
\eeq
which follows from particle-hole symmetry. Immediately, one recognizes that the form of (\ref{PathIntegral2}) defines a positive-definite measure. Thus, one can immediately apply the HMC algorithm to study various equilibrium properties of the system at half-filling.
For our lattice Hamiltonian, the fermionic operator can have zero eigenvalues in the presence of a nonzero Hubbard field. In lattice Quantum Chromodynamics (LQCD), one must avoid these so-called ``exceptional configurations", and this is one of the reasons why a mass term for the fermions is introduced by hand \textit{e.g.} for studies regarding the properties of QCD in the deep chiral limit, such as the order of the chiral phase transition. As a consequence of this, one typically needs to extrapolate results to the chiral limit, $m\to 0$, which can be computationally expensive. In the present case, fermionic zero modes lie on an $(N-2)$-dimensional space where $N \equiv N^2_s \times N_{\tau}$ is the dimension of the space of Hubbard fields. This result can be seen by noting that the fermionic determinant $\det M(\phi)$ is a complex number. Thus, the two conditions for the appearance of a zero mode are $\mbox{Re} \det M(\phi) = \mbox{Im} \det M(\phi) = 0$. The fact that the fermionic determinant is a complex number is important here, since fermionic zero modes form an $(N-1)$ dimensional subspace in the case of a purely real fermionic determinant, where only one condition survives \cite{PhysRevB.90.035134}.
As a result, for the complex fermionic determinant, these configurations are avoided in the molecular dynamics evolution and cannot divide the phase space into isolated regions. This is in direct contrast with LQCD where the phase space is divided into regions with different values for the topological charge \cite{DeTarDeGrand}. Thus, we do not need to introduce a mass term in our lattice action to obtain ergodic sampling of the phase space.
\section{\label{sec:Observables}Observables}
Using the form of the partition function developed in the previous section, one can immediately write down an expression for the thermal expectation value of an operator $\mathcal{O}$
\beq
\vev{ \mathcal{O}}_{\phi} = \frac{1}{Z} \int \mathcal{D}\phi |\det M[\phi]|^2\mathcal{O} e^{-S_B[\phi]}.
\eeq
To access the single-particle DOS we calculate the spatial trace of the fermion Green's function
\beq \label{GreensFunction} \nn
G(\tau) &\equiv& -\sum_x \vev{ \hat{a}_{x}(\tau) \hat{a}^{\dagger}_x(0) }_{\phi} \\ &=& \sum_x \vev{ M^{-1}_{x,\tau;x,0} }_{\phi}, ~\tau=0,2,\dots,2N_{\tau},
\eeq
where $\vev{}_{\phi}$ means the averaging over configurations of Hubbard field generated with the statistical weight (\ref{PathIntegral2}). In practice, one evaluates the trace on the right-hand side of (\ref{GreensFunction}) by the use of complex, Gaussian-distributed, stochastic vectors. It typically suffices to use $O(300)$ of these vectors on each configuration for lattice size $N_s=20$.
Since the single-particle DOS for half-filled system is symmetric with respect to zero, we will use the following symmetric Green-Kubo relation which connects the single-particle, momentum-averaged DOS $A(\omega) = \mbox{Im} G_R (\omega) / \pi$ to the Green's function in Euclidean time\cite{KadanoffMartin,MeyerTransport}
\begin{equation} \label{GreenKuboDOS}
G(\tau) = \int_0^\infty d\omega K(\tau,\omega) A(\omega),
\end{equation}
where the kernel for a correlator of the form $\vev{ \mathcal{O}(\tau)\mathcal{O}^{\dagger}(0)}$ is given by
\beq
K(\tau,\omega) \equiv \frac{\cosh\left[\omega(\tau-\beta/2)\right]}{\cosh(\omega \beta/2)}.
\eeq
In the next section we will discuss in detail our method for inverting the relation in (\ref{GreenKuboDOS}) for $A(\omega)$.
To understand the collective charge excitations of the system, we calculate the response of the equilibrium system, to linear order, to an external potential $A^{(\text{ext})}_0({\bf r},t)$. The scalar potential couples linearly to the charge density $\hat{\rho}({\bf r})$. The deviation of the charge density from its equilibrium value due to this time-dependent perturbation is then expressed in momentum space as:
\begin{equation}
\vev{ \delta \hat{\rho}({\bf q},\omega)} = \chi_{\rho}({\bf q},\omega) A^{(\text{ext})}_0({\bf q},\omega),
\end{equation}
where we introduced the charge susceptibility $\chi_{\rho}({\bf q},\omega)$. Translational invariance should be assumed to derive this relation.
From the charge density susceptibility, one can obtain the dielectric function
\begin{equation}
\frac{1}{\epsilon( {\bf q},\omega )} = 1 + V({\bf q}) \chi_{\rho}({\bf q},\omega ),
\end{equation}
where $V({\bf q})$ is the Fourier transform of the electron-electron interaction potential $V_{x,y}$. The peaks in ${\epsilon( {\bf q},\omega )}^{-1}$ give the dispersion relation for collective charge excitations (plasmons).
The quantities above are all defined in real time (frequency). However, in our approach one computes the following Euclidean correlator
\beq \nn
C({\bf q},\tau) &\equiv& \vev{ \hat{{\rho}}_{{\bf q}}(\tau) \hat{{\rho}}_{{\bf q}}(0) } \\ &\equiv& \frac{1}{Z} {\rm Tr} \, \left( e^{\tau \hat{\mathcal{H}}} \hat{{\rho}}_{{\bf q}} e^{-\tau \hat{\mathcal{H}}} \hat{{\rho}}_{{\bf q}} e^{-\beta \hat{\mathcal{H}}} \right),
\eeq
where $\hat{\rho}_{{\bf q}} = \frac{1}{V} \sum_{{\bf x}} e^{-i {\bf q} \cdot {\bf x}} \hat{\rho}_{{\bf x}}$. By using the definition of $\hat{\rho}_{{\bf x}}$ and the properties of the coherent states, one can obtain an expression for the correlator in the path integral representation. The result is given by
\begin{equation}
C({\bf q},\tau) = 2\sum_{{\bf x}} C({\bf x},\tau) \cos({\bf q} \cdot {\bf x}),
\end{equation}
where
\beq \label{DensityDensityPositionSpace} \nn
&&C({\bf x} - {\bf x}',\tau) \equiv \vev{ \hat{\rho}_{{\bf x}}(\tau) \hat{\rho}_{{\bf x}'}(0) } = \\ \nn && - 2 \mbox{Re} \vev{ M^{-1}_{x,\tau;x',0} M^{-1}_{x',0;x,\tau} }_{\phi} + 2 \mbox{Re} \vev{ M^{-1}_{x,\tau;x,\tau} M^{-1}_{x',0;x',0} }_{\phi} \\ &&- 2 \mbox{Re} \vev{ M^{-1}_{x,\tau;x,\tau} \bar{M}^{-1}_{x',0;x',0} }_{\phi}.
\eeq
In this expression we have performed an additional averaging over equivalent points in momentum space ($\pm \bf{q}$). The first term on the right-hand side of (\ref{DensityDensityPositionSpace}) is the connected piece while the other two terms constitute the disconnected part of the charge density-density correlator. Unlike the case of the current-current correlator \cite{UlybyshevConductivity}, both connected and disconnected parts are equally important, so the whole expression can not be calculated with a simple stochastic estimator.
The disconnected piece in (\ref{DensityDensityPositionSpace}) involves the correlation between two spatial traces evaluated on time slices separated by a distance $\tau$ in Euclidean time. It is thus necessary to perform $O(N), ~N \equiv N^2_s \times N_{\tau}$, inversions on each $\phi$ configuration as these pieces involve a fermion propagating from an arbitrary lattice point back to the same point. In LQCD, a variety of techniques have been employed to deal with the equivalent situation where quark disconnected loops are needed to accurately calculate an observable \cite{LiuStochastic, BaliStochastic, Gambhir:2016uwp}.
In this work, we have used a non-iterative solver based on the idea of Schur domain decomposition \cite{Schur_initial}. The solver is then applied to the point sources instead of the usual Gaussian-distributed stochastic ones. At the heart of the Schur complement method is the LU decomposition of dense matrix blocks contained inside the initial fermion operator matrix. The LU decomposition is made only once for a given $\phi$ configuration and is used repeatedly for all point sources. This allows us to work more efficiently in comparison with the commonly employed iterative solvers such as the conjugate gradient method.
Just as in the case of the fermion Green's function, there exists a Green-Kubo relation which connects the Euclidean charge density-density correlator and its spectral function
\begin{equation} \label{GreenKuboDensity}
C({\bf q},\tau) = \frac{1}{\pi}\int d\omega K_\chi(\tau,\omega) \mbox{Im} \chi_{\rho}({\bf q},\omega),
\end{equation}
where the kernel for a correlator of the form $\vev{ \mathcal{O}(\tau)\mathcal{O}(0)}$ is given by
\begin{equation} \label{DensityKernel}
K_\chi(\tau,\omega) \equiv \frac{\cosh\left[\omega(\tau-\beta/2)\right]}{\sinh(\omega \beta/2)}.
\end{equation}
Thus, in full analogy with the paper \cite{vanLoon}, we can plot $\mbox{Im} {\epsilon( {\bf q},\omega ) }^{-1}$ in order to reveal the dispersion relation for the plasmons.
In practice, it is more convenient to first solve the equation with the same kernel as for DOS
\begin{equation} \label{GreenKuboDensity1}
C({\bf q},\tau) = \int d\omega K(\tau,\omega) \tilde\chi_{\rho}({\bf q},\omega),
\end{equation}
and perform the rescaling for the spectral functions:
\begin{equation} \label{GreenKuboRescaling}
\mbox{Im} \chi_{\rho}({\bf q},\omega) = \pi \tanh(\omega \beta /2) \tilde\chi_{\rho}({\bf q},\omega),
\end{equation}
Along with the case of DOS, (\ref{GreenKuboDensity}) will be the object of study in the coming sections.
\section{\label{sec:AC}Analytic Continuation}
The central problem of this paper is obtaining the spectral functions for the fermion Green's function (DOS) and the charge density-density correlator in order to investigate the relationship between the formation of Hubbard bands and the nontrivial dispersion of the plasmons. However, directly inverting the relations in (\ref{GreenKuboDOS}) and (\ref{GreenKuboDensity}) constitutes an ill-posed problem. This is due to the fact that the kernel $K$ in (\ref{GreenKuboDOS}) has a very large condition number and thus even small changes in the Euclidean correlator $G(\tau)$ can lead to large changes in the spectral function $A(\omega)$ in frequency space. For this situation, a least-squares analysis is untenable.
In the context of lattice QCD, several approaches to the solution of this problem have been used. Two prominent examples are the maximum entropy method (MEM) \cite{Hatsuda1,Hatsuda2,Swansea} and the Backus-Gilbert method \cite{BrandtBG}. MEM uses Bayes' theorem to regularize the inverse problem through the introduction of priors on the spectral function. In the end, one hopes to show that the resulting spectral function has little dependence on the form of the priors. It has been found that MEM can successfully identify sharp structures in frequency space, such as peaks, but can fail to identify other, more smooth features \cite{UlybyshevConductivity}.
\begin{figure}
\centering
\includegraphics[scale=0.3, angle=-90]{figure1.pdf}
\caption{The effect of regularization on the spectral functions. We demonstrate the solution of (\ref{GreenKuboDOS}) for DOS using standard Tikhonov regularization. Here we plot the estimator for the spectral function obtained according to (\ref{ConvolutionDeltaFunction}). The frequency $\omega$ corresponds to the center of the resolution function and the filled area shows the statistical error. The example data is taken for the interaction strength $U/\kappa=1.66$, $V/\kappa = 0.62$ and temperature $T/\kappa = 0.046$. The lattice with spatial size $N_s=20$ and $N_{\tau}=160$ Euclidean time slices is used. Resolution functions for $\lambda=10^{-7}$ can be found in Fig.~\ref{fig:DeltaFunctions}. }
\label{fig:LambdaComparison}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.3, angle=270]{figure2.pdf}
\caption{The resolution functions, $\delta(\omega_0,\omega)$, obtained with standard Tikhonov regularization (see (\ref{SVDStandardTikhonov})) during the solution of the Green-Kubo relation for DOS (\ref{GreenKuboDOS}). The parameter $\omega_0$ labels the center of the resolution function. The value of the regularization parameter is $\lambda=10^{-7}$. This set up will be used in all cases where we compute the DOS from the Monte Carlo data. }
\label{fig:DeltaFunctions}
\end{figure}
On the other hand, BG has been found to work well in characterizing the features of spectral functions in a variety of situations. The main advantage of this method is that one does not need to make assumptions about any particular feature of the spectral function. We will illustrate the use of this method via (\ref{GreenKuboDOS}). However, the discussion is not tied to any particular form of the kernel. The method starts by defining an estimator of the spectral function
\begin{equation} \label{ConvolutionDeltaFunction}
\bar{A}(\omega_0) = \int^{\infty}_0 d\omega \delta(\omega_0,\omega) A(\omega).
\end{equation}
Thus, $\bar{A}(\omega_0)$ is the convolution of the exact spectral function $A(\omega)$ with the resolution function $\delta(\omega_0,\omega)$. One expresses the resolution function in the following basis
\begin{equation}
\delta(\omega_0,\omega) = \sum_{j} q_j(\omega_0) K(\tau_j, \omega),
\end{equation}
where the coefficients $q_j(\omega_0)$ will be determined shortly. This definition of the resolution functions introduces the second important feature of the BG method, linearity. Thus, the error estimation is much simpler and it opens up the possibility for other improvements which will be discussed below in the text. Due to the linearity of the GK relations (see (\ref{GreenKuboDOS}) and (\ref{GreenKuboDensity})), one obtains
\begin{equation} \label{Estimator}
\bar{A}(\omega_0) = \sum_j q_j(\omega_0) G(\tau_j),
\end{equation}
where $G(\tau)$ is a generic correlator in Euclidean time (the momentum dependence was been suppressed). The resolution in frequency space is determined by the width of the resolution function around $\omega_0$
\begin{equation} \label{ResolutionFunctionWidth}
D \equiv \int^{\infty}_0 d\omega (\omega-\omega_0)^2 \delta^2(\omega_0,\omega),
\end{equation}
where $\int^{\infty}_0 d\omega \delta(\omega_0,\omega) = 1$. The coefficients in (\ref{Estimator}) are determined by minimizing the width, $\partial_{q_j} D = 0$, keeping the norm of the resolution function fixed. The result of this minimization yields
\begin{equation}
q_j(\omega_0) =\frac{W^{-1}(\omega_0)_{j,k}R_k}{R_n W^{-1}(\omega_0)_{n,m}R_m},
\end{equation}
where
\beq \label{W}
W(\omega_0)_{j,k} &=& \int^{\infty}_0 d\omega (\omega-\omega_0)^2 K(\tau_j,\omega) K(\tau_k,\omega),\\ R_n &=& \int^{\infty}_0 d\omega K(\tau_n,\omega).
\eeq
The matrix $W$ is extremely ill-conditioned, with condition number $C(W) \equiv \frac{\lambda_{\text{max}}}{\lambda_{\text{min}}} \approx O(10^{20})$. It is thus imperative to regularize the method in order to obtain sensible results for a given set of data $G(\tau_i)$ and its associated error $\Delta G(\tau_i)$.
\begin{figure}
\centering
\includegraphics[scale=0.3, angle=-90]{figure3.pdf}
\caption{The half-peak width of the resolution function centered around $\omega_0=0$ versus the statistical error. The calculation is made with different regularization algorithms for the correlator $G(\tau)$ calculated on the lattice with spatial size $N_s=20$ and $N_{\tau}=160$ for the interaction strength $U/\kappa=1.66$, $V/\kappa = 0.62$ and temperature $T/\kappa = 0.046$. The statistical error is controlled by the regularization parameter $\lambda$. The resolution functions are computed in the process of inverting the Green-Kubo relation for DOS (\ref{GreenKuboDOS}). The width is given in units of temperature and one can see that as $\lambda$ vanishes each curve converges to $\sim 2$. This is as expected as the resolution in frequency space is ultimately limited by the temperature.}
\label{fig:ResWidthComparison}
\end{figure}
Previous studies employing the BG method have used the so-called ``covariance" regularization \cite{BrandtBG,UlybyshevConductivity}. In this approach, the following modification is made in (\ref{W})
\begin{equation}
W(\omega_0)_{j,k} \to (1-\lambda) W(\omega_0)_{j,k} + \lambda C_{j,k},
\end{equation}
where $\lambda$ is a small regularization parameter and $C_{j,k}$ is the covariance matrix of the Euclidean correlator $G$. The hope is that this replacement helps improve the condition number of the matrix $W$ while still maintaining a sufficiently small width of the resolution functions in frequency space.
Although covariance regularization performs well, one might wonder as to the merits of other commonly used regularization methods for ill-posed problems. Furthermore, in numerical studies where a covariance matrix cannot be constructed (i.e.~when the Euclidean correlator data are obtained using a non-stochastic procedure), covariance regularization cannot be applied. The regularization method that we propose, the so-called Tikhonov regularization \cite{Tikhonov}, is a widely used approach to ill-posed problems of the form $Ax=b$. In this method, one seeks a solution to the modified least-squares function
\begin{equation}
\text{min} \left( \Vert A x - b \Vert^2_2 + \Vert \Gamma x \Vert^2_2 \right),
\label{Tikh_def}
\end{equation}
where $\Gamma$ is an appropriately chosen matrix. The effect of various types of Tikhonov regularization on the matrix $W$ can be most easily seen by employing the singular value decomposition (SVD). In this procedure
\begin{equation}
W = U \Sigma V^{\top}, ~UU^{\top} = VV^{\top} = {\bm 1},
\end{equation}
where $\Sigma = \text{diag}(\sigma_1,\sigma_2,\dots,\sigma_N),~\sigma_1 \geq \sigma_2 \geq \cdots \geq \sigma_N$. The inverse is thus easily expressed as
\begin{equation}
W^{-1} = V D U^{\top},~D = \text{diag}( \sigma^{-1}_1,\sigma^{-1}_2,\dots,\sigma^{-1}_N ).
\end{equation}
In the standard Tikhonov regularization, one modifies the matrix D in the following way
\begin{equation} \label{SVDStandardTikhonov}
D_{i,j} \to \tilde{D}_{i,j} = \delta_{ij}\frac{\sigma_i}{\sigma^2_i + \lambda^2},
\end{equation}
where $\lambda$ is again the regularization parameter. One can see that the singular values which satisfy $\lambda \gg \sigma_i$ are smoothly cut off. This procedure corresponds to $\Gamma = \lambda {\bm 1}$ in (\ref{Tikh_def}). One thus pays a price for solutions that are not ``smooth". In general, for small $\lambda$, the solutions fit the data well but are oscillatory, while at large $\lambda$, the solutions are smooth but do not fit the data as well.
We have also tested an alternative method which regulates the small singular values of $W$ in a smoother fashion
\begin{equation} \label{SVDModifiedTikhonov}
\tilde{D}_{i,j} = \delta_{ij}\frac{1}{\sigma_i + \lambda}.
\end{equation}
For this choice, which we refer to as ``modified Tikhonov", we give preference to spectral functions which give smooth reconstructed Green's functions in Euclidean time. This method corresponds to $\Gamma = \lambda A$ ($A \to K$ in our case) in (\ref{Tikh_def}).
The effect of regularization on the reconstructed spectral function is shown in Fig.~\ref{fig:LambdaComparison} for the case of DOS. The corresponding resolution functions obtained with ordinary Tikhonov regularization with $\lambda=10^{-7}$ are plotted in Fig.~\ref{fig:DeltaFunctions}. One can clearly see that varying the regularization parameter $\lambda$ has a significant effect on the resulting spectral function. This set up will be used later for further calculations of DOS.
\begin{figure}
\centering
\includegraphics[scale=0.3, angle=270]{figure4a.pdf}
\includegraphics[scale=0.3, angle=270]{figure4b.pdf}
\caption{Two variants of correlator averaging. The charge density-density correlator $C({\bf q},\tau)$ is calculated for $U/\kappa=3.33$, $V/\kappa=1.26$ and $T/\kappa=0.046$ using a lattice with spatial size $N_s=20$ and $N_\tau=160$ Euclidean time slices. The momentum $q$ belongs to the $X-M$ line in the Brillouin zone, which is the most physically interesting case (see below in section \ref{sec:Results}).}
\label{fig:corr1}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.3, angle=270]{figure5.pdf}
\caption{ Spectral functions $\mbox{Im} \chi_\rho (\omega,q)$ computed using the relations (\ref{GreenKuboDensity1}) and (\ref{GreenKuboRescaling}) for two different variants of the averaged correlator shown in Fig.~\ref{fig:corr1}. In both cases ordinary Tikhonov regularization is used with $\lambda=5.0 \times 10^{-6}$.
We plot the estimator for the spectral function obtained according to (\ref{ConvolutionDeltaFunction}). The frequency $\omega$ corresponds to the center of the resolution function and the filled area shows the statistical error.}
\label{fig:compare_average}
\end{figure}
To gain an accurate estimate of the statistical error in our spectral functions, we perform a data binning as follows. Taking our original ensemble of $N_{\text{conf}}$ Hubbard field configurations generated according to the weight (\ref{PathIntegral2}), we construct $\tilde{N}$ blocks of $N_{\text{conf}}/\tilde{N}$ configurations. We are then left with several subsets of Euclidean time correlators
\beq \nn
&& \{ G^{(i)}(\tau_j), i=1,\dots,N_{\text{conf}} \} \to \\ \nn &&\{ G^{(i)}(\tau_j), i=1,\dots,N_{\text{conf}}/\tilde{N} \}, \dots , \\ && \{ G^{(i)}(\tau_j), i=(\tilde{N}-1)N_{\text{conf}}/\tilde{N}+1,\dots,N_{\text{conf}} \}.
\eeq
The number of blocks, $\tilde{N}$, is chosen by examining the autocorrelation of the Euclidean correlator between different Hubbard field configurations and enforcing the condition $N_{\text{conf}}/\tilde{N} \gg l_{\text{max}}$, where $l_{\text{max}}$ is the maximum autocorrelation length pertaining to $G(\tau_i), i=0,1,\dots,N_{\tau}-1$. This condition ensures that the size of each block is such that it contains numerous statistically-independent configurations and each block is statistically independent of all other blocks. Using these blocks of correlators, we construct $\tilde{N}$ spectral functions $\bar{A}_i$ and calculate an average spectral function
\beq
\bar{A}_{\text{avg}} \equiv \frac{1}{\tilde{N}} \sum^{\tilde{N}}_{i=1} \bar{A}_i,
\eeq
and its associated error $\sigma(\bar{A})$ for each frequency. We have found that this procedure yields a much better estimate of the statistical error than simply propagating the error in the Euclidean correlator to the spectral function through the relation in (\ref{Estimator}).
Fig.~\ref{fig:LambdaComparison} demonstrates typical behaviour of the statistical errors when we switch on the regularization. If the regularization is not sufficient, the spectral function has huge statistical errors. Once we increase $\lambda$, the errors are suppressed and the spectral function converges to some stable average value. As mentioned previously (see Fig.~\ref{fig:ResWidthComparison}), the price for this ``smoothing'' is the enlargement of the width of all resolution functions which may imply the loss of some information.
We now compare all three types of regularization using the width of the resolution functions as a metric. The study is summarized in Fig.~\ref{fig:ResWidthComparison}, where we plot the width at half maximum of the resolution function with center at $\omega_0=0$ versus the statistical error of the reconstructed spectral function (again at $\omega_0=0$) for all regularization methods. The data for DOS were used as a test case.
As one can see, both types of Tikhonov regularization work better in the sense that they provide better resolution in frequency being equally efficient in suppressing the statistical error. Or, in other words, for Tikhonov regularization the statistical error is smaller for the same resolution in frequencies. On can see also, that
in practice the difference between the standard (\ref{SVDStandardTikhonov}) and modified Tikhonov (\ref{SVDModifiedTikhonov}) regularizations is negligible. As a consequence, we will use standard Tikhonov approach in all further calculations.
The algorithm for choosing the optimal value of $\lambda$ is based on the ``global relative error" for the spectral function
\beq \label{GlobalRelativeError}
\mathcal{G} \equiv \frac{1}{N_0} \sum_{\omega_0} \frac{\sigma\left( \bar{A}(\omega_0) \right)}{\bar{A}_{\text{avg}}(\omega_0)},
\eeq
where the sum in the above expression runs over the centers of the resolution functions and $N_0$ is the number of resolution functions with different centers $\omega_0$. Our basic criteria for the choice of $\lambda$ is that the ``global relative error" should be within the interval $5-10\%$. We thus sufficiently suppress the statistical error while still maintaining good resolution in frequency space.
Using the quantity in (\ref{GlobalRelativeError}) as a measure, we start from small $\lambda$ and increase it until we have obtained the desired statistical error. Typically, we have taken $\lambda=10^{-7}-10^{-5}$ in obtaining the results in this paper.
\begin{figure}
\centering
\includegraphics[scale=0.28, angle=-90]{figure6a.pdf}
\includegraphics[scale=0.28, angle=-90]{figure6b.pdf}
\includegraphics[scale=0.28, angle=-90]{figure6c.pdf}
\caption{Plot of the DOS for the lattice ensemble with electron-electron interaction
$U/\kappa=0.83$, $V/\kappa = 0.5$ (a), $U/\kappa=1.66$, $V/\kappa = 0.62$ (b) and $U/\kappa=3.33$, $V/\kappa = 1.26$ (c). In all cases we have used a spatial lattice size of
$N_s=20$ with $N_{\tau}=160$ steps in Euclidean time and a temperature of $T/\kappa= 0.046$. Standard Tikhonov regularization with $\lambda=10^{-7}$ is employed during the analytical continuation. Thus, in all cases, we have the same resolution in frequency. No additional averaging in Euclidean time is applied. Filled areas show the statistical error computed with the data binning procedure and the frequency $\omega$ corresponds to the center of the resolution function. The average value represents the estimator for DOS computed according to (\ref{ConvolutionDeltaFunction}).
One can see the formation of the Hubbard bands, characterized by the peak at $E/\kappa \approx 0.5$, indicating that one is in the strongly-coupled phase for the largest interaction strength. For smaller interaction strengths we obtain a strong peak at zero energy indicating that the system is in the metallic phase.}
\label{fig:DOS1}
\end{figure}
This primary regularization is enough to obtain reasonably good results for DOS, as one can see from Fig.~\ref{fig:LambdaComparison}. Unfortunately, for the case of the charge density-density correlator, when we calculate the charge susceptibility by solving (\ref{GreenKuboDensity}) this regularization is not enough. The source of the this problem is the very large autocorrelation between different points in Euclidean time for the correlator $C({\bf q},\tau)$. As a result, it exhibits long-range fluctuations which can be seen in Fig.~\ref{fig:corr1}. We took the example correlator $C({\bf q},\tau)$ for the largest interaction strength considered ($U/\kappa=3.33$, $V/\kappa=1.26$) with the momentum $q$ directed along the $X-M$ line in the Brillouin zone. This is the most physically interesting case and will be discussed in detail in section \ref{sec:Results}. After the application of the analytic continuation procedure, this oscillating correlator leads to a wildly fluctuating spectral function unless the regularization is so large that we can hardly resolve any structure due to very wide resolution functions $\delta(\omega_0,\omega)$.
The way in which we have modified the BG method to alleviate this problem is through the introduction of ``interval averaging". In this procedure, we take the correlator data $\{ G(\tau_i), \Delta G(\tau_i); ~i=0,1,\dots , N_{\tau}-1 \}$ and map this to a new set $\{ \tilde{G}(\tilde{\tau}_j), \Delta \tilde{G}(\tilde{\tau}_j); ~j=1,\dots , N_{\text{int}} \}$ where
\beq
\tilde{G}(\tilde{\tau}_j) &\equiv& \frac{1}{\tilde{N}_j} \sum^{\tilde{N}_j}_{i=1} G(\tau^{(j)}_i),\\ N_{\tau} &=& \sum^{N_{\text{int}}}_{j=1} \tilde{N}_j,~1 \leq \tilde{N}_j < N_{\tau}.
\eeq
After performing the procedure, we are left with a set of averaged values of the correlator calculated over certain intervals in Euclidean time. Due to the linearity of (\ref{GreenKuboDOS}) and (\ref{GreenKuboDensity}), one can construct $\{ \tilde{K}(\tilde{\tau}_j); ~j=1,\dots , N_{\text{int}} \}$ in an analogous manner and use these in the construction of the spectral function via the Tikhonov regularization. As a result, the spectral function will reproduce the averaged Euclidean correlator and will not follow the fluctuations within these intervals.
\begin{figure}
\centering
\includegraphics[scale=0.3, angle=-90]{figure7.pdf}
\caption{Plot of the DOS for the lattice ensemble with electron-electron interaction $U/\kappa=4.4$, $V/\kappa = 1.26$, which corresponds to the smallest interaction strength studied in \cite{vanLoon} ($U^{*}=1.1$ in their notation). We have used a spatial lattice size of
$N_s=20$ with $N_{\tau}=160$ steps in Euclidean time and two times larger temperature $T/\kappa= 0.092$, which is almost equal to the one used in \cite{vanLoon}. The set up for the analytic continuation is exactly the same as for Fig.~\ref{fig:DOS1}.
One can see that in contrast to the results displayed in \cite{vanLoon}, the Hubbard bands are already formed, which suggests that the metal-insulator phase transition is shifted to a smaller interaction strength even at the same temperature.}
\label{fig:DOS_high}
\end{figure}
Typically, the signal-to-noise ratio of a Euclidean correlator $G(\tau)$ becomes worse as one approaches $\tau = \beta/2$. Furthermore, the correlations between adjacent points in Euclidean time are strong as one approaches this point. In light of this, we leave the points near $\tau=0$ untouched, while points closer to $\beta/2$ are bunched into intervals of longer length.
In order to examine the dependence of the results on the choice of the intervals, we have performed the analytic continuation for the two different choices of intervals displayed in Fig.~\ref{fig:corr1}. The results of the analytic continuation are shown in Fig.~\ref{fig:compare_average}. One can see that the two agree within error bars which indicates that the qualitative features of the spectral function do not drastically change for different variants of averaging. In our calculations for the charge susceptibility we will use the first variant of averaging shown in Fig.~\textcolor{red}{4a}.
\begin{figure}
\centering
\includegraphics[trim=1.5cm 3cm 0 0, scale=0.35, angle=0]{figure8a.png}
\includegraphics[trim=1.5cm 3cm 0 0, scale=0.35, angle=0]{figure8b.png}
\includegraphics[trim=1.5cm 3cm 0 0, scale=0.35, angle=0]{figure8c.png}
\caption{ Charge susceptibility for the square lattice Hubbard-Coulomb model for three different interaction strengths. The color scale corresponds to $V(q) \mbox{Im} \chi_\rho (\omega, q)$. For the spectral function $\mbox{Im} \chi_\rho (\omega, q)$ we use the estimated value after analytic continuation (see (\ref{ConvolutionDeltaFunction})). The same lattice ensembles were used as in Fig.~\ref{fig:DOS1}. Standard Tikhonov regularization with $\lambda=5 \times 10^{-6}$ and additional time averaging according to Fig.~\textcolor{red}{4a} is used in all cases.
}
\label{fig:Susceptibility1}
\end{figure}
\section{\label{sec:Results}Results}
In this section, we present our results for the spectral functions.
We perform all of our calculations with a spatial lattice size of $N_s=20$ and $N_\tau=160$ Euclidean time slices. Temperature is equal to $T=0.046$ in units of the hopping parameter $\kappa$. This temperature is smaller then the one used in \cite{vanLoon}, but we really need it since the resolution of the BG method is limited by temperature. In order to justify our conclusions we've made also one calculation for 2 times higher temperature $T=0.092 \kappa$. For each strength of the electron-electron interaction we have generated $\sim 10^3$ Hubbard field configurations for the calculation of the relevant observables. Since the width of the resolution functions (quantity $D$ in (\ref{ResolutionFunctionWidth})) is bounded from below by temperature, the spacing between neighboring values of $\omega_0$ is equal to the temperature in our calculations. The upper bound for $\omega_0$ is defined by the bandwidth, where all considered spectral functions go to zero.
To make contact with the results of \cite{vanLoon}, we state the values of the parameters characterizing the interaction potential in our notation. The three sets of parameters used in \cite{vanLoon} correspond to $V \approx 1.26$ and $U = 4.4, 8.2, 10.4$ (in units of the hopping parameter $\kappa$). In \cite{vanLoon} the Mott transition was observed somewhere between $U = 8.2$ and $U=10.4$. Our first aim is to identify the real position of the phase transition from our non-perturbative Monte-Carlo calculations. For this purpose , we calculate the DOS for several pairs of $U$ and $V$ which characterize the electron-electron interaction. The results from these calculations are presented in Fig.~\ref{fig:DOS1}.
One can clearly see that even for the case of $U = 3.33$, $V=1.26$ (see Fig. \textcolor{red}{6c}), which is smaller then the smallest interaction strength from \cite{vanLoon}, the system is already in the insulating state. In this regime, the Hubbard bands have already formed and the quasiparticle weight at $\omega=0$ practically vanishes (due to finite temperature and the finite width of resolution functions it can not vanish completely). In Fig. \textcolor{red}{6b}, the DOS for the couplings $U = 1.66$, $V=0.62$ demonstrates that one is now firmly in the regime with well-defined quasiparticles at $\omega=0$. Finally, Fig.~\textcolor{red}{6a} shows the DOS for $U=0.83$, $V=0.5$ which is deep in the metallic phase. These results show that the EDMFT analysis from \cite{vanLoon} overestimates the critical coupling $U_c$ of the Mott transition for the Hubbard-Coulomb model.We emphasize the fact that this is not a temperature effect, as it is known that for the square-lattice Hubbard model, the phase transition is shifted to smaller $U$ as the temperature decreases \cite{FateMottHubbard}. The results displayed in Fig.~\ref{fig:DOS_high} show that, at twice the temperature of the ensembles in Fig.~\ref{fig:DOS1}, we are already in the insulating phase. For these same values of the coupling, $U = 4.4$, $V=1.26$, and a slightly smaller temperature, the authors of \cite{vanLoon} find the system to be in the metallic phase.
Another difference is that one can not confirm the situation reported in \cite{vanLoon} for intermediate coupling: near the phase transition the DOS had equally high peaks at zero energy and at some non-zero values $\pm E_0$. It is possible that something similar develops in the Fig.~\textcolor{red}{6b}, but the second peak is too small and is basically within the error bars.
The results for the charge susceptibility in momentum space are presented in Figs.~\ref{fig:Susceptibility1} and \ref{fig:compare_X} for the same lattice set up and interaction strengths as was presented for DOS in Fig.~\ref{fig:DOS1}. The plots start from the center of the Brillouin zone ($\Gamma$-point), continues out towards the edge of the BZ along the $k_x$-axis, moves along the edge of the BZ parallel to the $k_y$-axis, and then finally returns to the center of the BZ along the diagonal connecting the point $M \equiv (\pi,\pi)$ with $\Gamma$. For the smallest interaction strength we have tried to reproduce the $\sqrt{q}$ dispersion relation in the vicinity of the $\Gamma$-point (see Fig.~\textcolor{red}{8a}). The positions of the peaks at a given value of the momentum are depicted with black dots and we have fitted their position with the function $f(q)=C \sqrt{q}$. Despite the fact that the statistical fluctuations in the Monte Carlo data do not allow us to fully justify this fitting procedure, our data at least do not contradict the $\sqrt{q}$ dispersion at low interaction strength. When we move to larger $U$ and $V$, the dispersion relation sufficiently changes and at the largest interaction strength, already in the insulating state, it finally splits into two branches. The splitting is most prominent along the $X-M$ line in the Brillouin zone. The case of the $X$-point is shown in the Fig.~\ref{fig:compare_X} separately for all three strengths of the electron-electron interaction. One can clearly see two peaks near $\omega=\kappa$ and $\omega=5\kappa$.
\begin{figure}
\centering
\includegraphics[scale=0.3, angle=270]{figure9.pdf}
\caption{Comparison of the spectral function for the charge-charge correlator at the X point for different interaction strengths. The lattice set up and the parameters of the analytic continuation procedure are identical to the ones used for Fig.~\ref{fig:Susceptibility1}. The filled areas show the statistical error computed with the data binning procedure and the frequency $\omega$ corresponds to the center of the resolution function. The spectral functions for smaller interaction strengths are rescaled by factors of 2 and 4 in order to fit to the same scale.
}
\label{fig:compare_X}
\end{figure}
\section{\label{sec:Conclusion}Conclusions}
We have studied the Mott metal-insulator transition in the Hubbard-Coulomb model on the square lattice using unbiased quantum Monte Carlo calculations on finite clusters. The hybrid Monte Carlo algorithm was used to effectively study the system which contained a long-range interaction. In order to obtain the real-frequency spectral functions, we have developed a modified version of the Backus-Hilbert method for analytic continuation from Euclidean to real time. The metal-insulator phase transition was observed directly by the calculation the density of states. The decrease of DOS at zero energy and the formation of the Hubbard bands was observed across the phase transition. It was observed that the position of the phase transition is sufficiently shifted towards smaller interaction strength in comparison to previous EDMFT predictions.
The behavior of the momentum-resolved charge susceptibility across the phase transition was also studied. This data was used to reveal the dispersion relation of the plasmons both in the metallic and in the insulating phase. The main aim was to check the predictions from \cite{vanLoon} concerning the splitting of the plasmonic dispersion relation into two bands in the region of the interaction strength close to the phase transition. The splitting was indeed observed in our Monte Carlo data. However, according to our calculations this phenomena tends to emerge in the situation when the material is already in the insulating state.
The data presented in the paper can be used as a benchmark in the further development of methods for strongly-correlated systems with long-range interactions.
The modified BG method developed in this paper can also be used in cases where a non-biased estimate for the spectral function is important. The code used in the current paper is now accessible online \cite{BG_code1}.
\section*{Acknowledgements}
MU acknowledges inspiring and fruitful discussions
with Prof.~Mikhail Katsnelson. The work of MU
was supported by DFG grant BU 2626/2-1. SZ acknowledges support by the National Science Foundation (USA) under grant PHY-1516509 and by the Jefferson Science Associates,
LLC under U.S. DOE Contract $\#$ DE-AC05-06OR23177.
This work was partially supported by the HPC Center of Champagne-Ardenne ROMEO. CW acknowledges the warm hospitality of the University of Regensburg and would like to thank Pavel Buividovich for his financial support and for discussions in the early stages of the project. CW is supported by the University of Kent, School of Physical Sciences.
|
\section{Introduction}
Traditionally, aerospace structures have been largely manufactured by subtractive processes, such as machining,
which is typically applied to metals. Composite materials, which constitute over 50\% of recent aircraft construction,
are created at the same time as the structure itself; typically by an additive layup process such as
automated fibre placement (AFP). This co-assembly of material and structure
is achieved by sequential deposition of a number of fibrous layers which are either pre-impregnated or post-infused with resin, involving
hundreds of additive operations during the production of a large structural part.
Most of these processes take place before the resin is cured, when the influence
of temperature and pressure on the deposition, forming and curing of parts with
complex geometry is not well understood. For this reason production is vulnerable
to defects arising from small process variations and achieving high-rate automation
of the process is limited.
The ability to manufacture composite material-structural
systems with repeatable properties at a reasonable rate and competitive cost,
with the quality required for certification, is a major challenge in the global
aerospace industry. A number of defect types can arise, and use of non-destructive
evaluation is limited in detecting these, hence conservative `knock-down' factors
are applied to strength limits. These factors are currently obtained by
expensive testing at all scales in the Test Pyramid, with large numbers of small
coupon tests and fewer larger-scale structural tests.
Large-scale tests, are extremely expensive and take
place at a stage when it is difficult to make changes and improve designs, while
small-scale coupons do not represent the performance of the material at the
structural scale where, for example, defects can be introduced as a result of
complex forming processes.
Furthermore,
the boundary conditions and loading in a simple element are very different from the performance
of the material within the structure.
One loading of particular interest is corner unfolding,
in which though thickness tensile loading acts to separate the layers as the corner radius
is increased under bending loads. These stresses act in the weak resin-dominated direction of the
laminate and can therefore lead to a limiting design case. The presence of defects such as mis-aligned
fibres in the form of out-of-arc wrinkles can have a significant importance. Such a reduction
in strength is referred to in manufacturing as a knock-down factor.
Mukhopadhyay et al have modelled the effect of defects on compressive \cite{muk1}
and tensile \cite{muk2} strength in flat laminates using 8-node, solid elements
and zero-thickness, 8-node, cohesive elements between plies. Damage modelling accounted
for nonlinear shear in plies,
transverse matrix cracking,
mixed mode delamination, tensile fibre fracture and fibre kinking. The
minimum size of the FE mesh was one ply thickness (0.25 mm) in the vicinity of wrinkles and towards laminate edges.
The compressive strength predictions were in agreement to within 10\%
of experimental results and were able to pick up a mode switch from
fibre failure to delamination when defect misalignment was above
$\sim 9^\circ$. In the tensile case, it was noted that wrinkles
act as local, through-thickness shear stress concentrators.
For multidirectional laminates, the influence
of the defect was exacerbated by edge effects.
We have previously shown that the application of a $3$ mm tough resin layer to the longitudinal
edges of corner unfolding specimens prevents the zero stress requirement along the edge, thus reducing the numerical
singularity at the edge. The finite element solution becomes more realistic for a given
mesh fidelity and failure prediction
methods become more accurate. Such an edge treatment can increase the strength by up to $20\%$
\cite{fletcher2016resin} and provides a more accurate representation of a long structural part.
There is a wide range of defects that can form in composite laminates and a
detailed taxonomy of these is presented in \cite{potter}.
In this paper we evaluate wrinkle defects, which are more likely to form
in curved laminates as a result of consolidation onto a male tool \cite{dodwell}.
They are also likely to form in curved laminates with tapered sections, which
causes double curvature and makes AFP deposition challenging. These wrinkles are
important to the performance of curved laminates.
In this paper we present a new FE analysis tool \textit{dune-composites} for efficient, high-fidelity
modeling of laminated composite parts. We show, using a simple test case that the results of \textit{dune-composites}
match up well with the commercial software package ABAQUS in all six
stresses. For large scale problems, \textit{dune-composites} crucially
relies on robust iterative solvers for the resulting FE systems.
To this end, we introduce the preconditioner GenEO \cite{geneo1,geneo2,geneoparallel}, which we have implemented within DUNE.
This preconditioner has previously been mathematically proven to be
robust for isotropic FE systems.
We show that these results extend to anisotropic problems and that the solver is suitable
for solving large composites problems. Further, we demonstrate its
parallel efficiency and its ability to scale to
thousands of compute cores, allowing the solution of the large problems with defects mentioned above.
We test this module by modeling the unfolding of a curved laminate
part containing manufacturing defects, for which a micron scale mesh is needed to accurately compute stresses. We show that
\textit{dune-composites} is able to accurately predict damage initiation and that it does so
at a fraction of the computational cost required by ABAQUS.
\section{Modelling approach}
In this section, we introduce the new high performance finite element module \textit{dune-composites},
and demonstrate its capability of efficiently and robustly tackling
large-scale simulations of composite structures. In the example simulations presented, the composite strength of pristine
and defected corner radii are accurately predicted. The analysis assumes standard anisotropic
3D linear elasticity and the failure is assessed using a quadratic damage onset
criterion for the initiation of delamination in \cite{Camanho}.
This allows the results to be benchmarked against existing numerical results and experiments, given
in Fletcher et al. \cite{fletcher2016resin}. For this case, the numerical results
of this linear model show good agreement with experimental data.
The failure of curved laminates subjected to corner unfolding has been shown
to be highly unstable with instantaneous propagation following initiation \cite{fletcher2016resin}.
Therefore we focus on capturing the initiation of failure accurately by using a high-fidelity
3D mesh, with $4$ or more elements through the thickness of each ply and interface layer.
Such an approach has been shown to give accurate results for simple flat coupons \cite{esquej}
and 2D models of a few plies \cite{2dply}. However, this level of fidelity is typically dropped when modelling
larger, more complex parts, such as 3D curved laminates, due to computational limitations.
Here, we do not use any of these more complex models in order to reduce the computational requirements.
More complex modelling approaches have been proposed for modelling of defects in the literature,
these include the use of composite shell elements \cite{muk2, muk1},
interface/cohesive elements \cite{hallettcohesive, hallett2008}
and higher-order continuum models \cite{fleck2004}. In particular, cohesive elements are commonly
used to capture propagation \cite{hallettcohesive},
but propagation is considered less important than initiation for the problem.
The formulation of such
elements, whilst more complex, is possible within \textit{dune-composites}. The solution
strategy would require Newton iterations and path-following methods (available in the DUNE library).
However, even for these non-linear models of failure, computational cost
is still dominated by the speed of solving a linearised system of equations for composite materials.
In this paper our focus is therefore on developing and implementing an efficient solution strategy
for these linearised equations arising from such massive composite simulation.
\subsection{{\em dune-composites}: High performance FE modelling of large-scale composite structures}
DUNE (Distributed and Unified Numerics Environment) is an open source modular toolbox
for solving partial differential equations (PDEs) with grid-based methods,
such as the finite element method (FEM) \cite{dune2a,dune2b,dune1}. Written using modern C++
programming techniques,
the core modules of DUNE have been developed by mathematicians and computer scientists
to allow users to implement and use state-of-the-art mathematical methods across large
high performance parallel computing architectures. It is a generic
package that provides a user with the key ingredients for solving any FEM problem,
e.g. grid generation, different types of finite elements, quadrature rules
and a choice of off-the shelf solvers.
Within this platform, we have developed a new module, \textit{dune-composites}, which
solves the linear elasticity equations with general, anisotropic
stiffness tensor, applicable for modelling
composite structures. This module provides an interfaces to handle composite
applications which includes stacking sequences, complex part geometries and
complex boundary conditions such as multi-point constraints, or periodic boundary conditions.
Further, we have implemented a new $20$-node 3D serendipity element
(with full integration) within \textit{dune-pdelab},
which is not prone to shear locking and
allows comparison with ABAQUS's C3D20R element. This element has degrees of freedom
at the $8$ nodes of the element as well as on each of the $12$ edges.
The main advantage of implementing our simulation tool within DUNE
is that it allows us to exploit developments in state of the art solvers
such as Algebraic Multigrid (AMG), see \cite{dune3}, or to implement
new ones. In particular, as part of our new developments, we
implemented a novel, robust preconditioner, called GenEO
\cite{geneo2,geneo1}, within DUNE.
Finally, we will show in Section \ref{sec-time} that DUNE allows for highly parallelised efficiency
on hundreds of computer cores. This allows for the modelling of meshes fine enough to
resolve defects or the modelling of wide parts with sufficient accuracy.
\subsection{Verification of DUNE}
\label{sec-verify}
In this section we show that DUNE produces stress results that are comparable to
those produced by standard FE libraries such as ABAQUS.
For our comparison we examine a corner unfolding test in DUNE
and ABAQUS and compare the solutions as well as the run times. We
initially use a small computationally inexpensive problem for
comparisons with ABAQUS and move to a larger problem to demonstrate the
parallel efficiency of DUNE.
Figure \ref{fig-mpcsetup} shows the setup of our problem. For this small test we use a width
$W = $ 15mm and $L = $ 3mm long limbs. The inner radius $R$ of the curved section is 6.6mm.
The overall thickness $T$ is 2.98mm, where the ply layers are 0.23 mm thick and the resin
interfaces are 0.02 mm thick.
We then apply a resin treatment of 2mm to the free edges of the
laminate. This type of edge treatment has been shown in \cite{fletcher2016resin} to be
advantageous both for reducing conservatism in the design of aircraft
structures, as well as to make FE analyses more reliable.
The mechanical properties for the ply material and for the resin
interfaces of the curved laminate are given in Table \ref{table-1}.
The fibre angles are given by the following $12$-ply stacking sequence
$$\left[\pm45/90/0/\mp45/\mp45/0/90/\pm45\right].$$
\begin{table}[t]
\centering
\begin{floatrow}
\begin{tabular}{ |c|c|c|c| }
\hline
\multicolumn{2}{|c|}{\bf Orthotropic fibrous layer} & \multicolumn{2}{|c|}{\bf Isotropic interface layer}\\
\hline
$E_{11}$ & 162 GPa & $E$ & 10 GPa \\
\hline
$E_{22},\,E_{33}$ & 10 GPa & $\nu$ & 0.35 \\
\hline
$G_{12},\,G_{13}$ & 5.2 GPa & \multicolumn{2}{|c|}{}\\
\hline
$G_{23}$ & 3.5 GPa & \multicolumn{2}{|c|}{\bf Resin edge material}\\
\hline
$\nu_{12},\,\nu_{13}$ & 0.35 & $E$ & 8.5 GPa\\
\hline
$\nu_{23}$ & 0.5 & $\nu$ & 0.35 \\
\hline
\end{tabular}
\caption{Mechanical properties for the curved $12$-ply laminate.}
\label{table-1}
\end{floatrow}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[scale=0.9]{CornerDiagram.pdf}
\caption{Sketch of the edge-treated corner-unfolding specimen (left) and a
cross section of the configuration (right).}
\label{fig-mpcsetup}
\end{figure}
In ABAQUS, the corner unfolding is modelled using a beam-type multi-point constraint (MPC)
at the end of one limb. In DUNE,
a similar effect is
achieved by adding a thin layer of very stiff material at the end of
the same limb. In both cases, a $96.8$ Nmm/mm moment is applied to
that limb, while fixing all degrees of freedom on the other limb,
as shown in Figure \ref{fig-mpcsetup}.
We denote the stresses using the following directional notation:
arc length $s$ (around the curve); radius $r$ through-thickness and width
$l$ across the laminate. These are shown by the coordinate system in Figure
\ref{fig-mpcsetup}a.
DUNE and ABAQUS both use a model with second-order serendipity elements
(element \verb=C3D20R= in ABAQUS, \cite{serendipity}). There is a small difference in that
ABAQUS uses reduced integration while DUNE uses full integration.
Also, the stresses are not recovered in an identical way from the
displacements in the two codes.
We use a tensor-product grid with $56$ elements in the $l$ and $s$ directions
and $6$ elements in each of the 12 fibrous and the 11 interface layers in
the $r$ direction. In total, this mesh contains $432,768$ elements.
To ensure the effects at the free edge and at the material
discontinuities are sufficiently resolved, the FE mesh is graded
towards both edges in the $l$ direction, as well as towards the
interfaces in each of the resin and fibrous layers in the $r$
direction. The bias ratio
gives the ratio between width of the smallest and largest element of the mesh.
We choose a bias ratio of $400$ between the
centre and the edges of the structure in the $l$ direction and a bias
ratio of $10$ between the centre and the interfaces in each layer in
the $r$ direction.
\begin{figure}[p]
\includegraphics[scale=0.93]{stresscomp.pdf}
\caption{Stresses in MPa as a function of radius $r$ from outer radius
at the apex of the curve,
at $2.156$mm from the edge of the resin-edge-treated laminate
(DUNE, solid blue; ABAQUS, dotted red).
The background colours indicate the stacking sequence: $+45$ =
red, $-45$ = blue,
$90$ = green, $0$ = yellow.}
\label{fig-stresses}
\end{figure}
In Figure \ref{fig-stresses}, all six stresses computed with ABAQUS and DUNE
are compared at the apex of the curve (i.e., at an angle of $45^\circ$) near the resin edge, $2.156$mm
away from the edge, which is in the laminate $0.156$mm ($6$ elements)
away from the resin laminate edge. Since we use identical meshes and identical
elements, the numbers of degrees of freedom are also the same. However, there are some differences
in the models, such as the quadrature order, use of single precision in ABAQUS and the stress
reconstruction, which could account for the different results. Nevertheless, the results
show reasonably good agreement for all six different stresses.
\subsection{Efficiency of \textit{dune-composites}: Robust iterative solvers}\label{sec-solvers}
The computational gains of \textit{dune-composites} demonstrated in this
section are a direct consequence of the use of fast and robust
iterative solvers for the resulting large systems of linear equations
\begin{equation}\label{eq-1}
K\underline{u}=\underline{f},
\end{equation}
where $K$ is the global stiffness matrix, $\underline f$ is the load
vector arising from the applied boundary conditions and $\underline u$
is the solution vector containing
the displacement values at each node within the finite element
mesh. There are also some significant gains in the setup time due
to a more efficient, problem-adapted data management.
Solvers for systems of linear algebraic equations can broadly be
classified into direct and iterative ones. Direct solvers, based on
matrix factorisation, are more universally applicable and more robust
to ill-conditioning \cite{DirectSparse}. Thus, they are typically the default in
commercial FE packages. However, the cost of the factorisation and the
memory requirements can quickly become infeasibly high for large
problems. Moreover, it is difficult
to achieve good parallel efficiencies on large multicore computer
architectures. Since they do not require any factorisations, iterative
methods for \eqref{eq-1} have the potential to scale
optimally, both with respect to problem size and with respect to the
number of processors in a parallel implementation, but that crucially
depends on the 'conditioning' of the problem.
In size, we refer to the total number of degrees of freedom $N$ in the
underlying finite element (FE) solution $\underline u$. Below we will see examples,
for which $N$ can be very large, up to 60 million. It is important to
note that $K$ is {\em sparse}, that is, the number of nonzero entries in $K$
is proportional to $N$, which is significantly less than the possible
$N^2$ entries in a full matrix. Iterative methods only require the
storage of the original, sparse stiffness matrix, while the storage of
the factors in direct methods is closer to the $N^2$ entries in a full
matrix. The computational cost of the best direct methods for FE systems
arising from three-dimensional structural calculations, such as those
used in ABAQUS, still grows at least with order $N^{1.5}$, as we will
see below. Iterative solvers that only require multiplication with the
sparse matrix $K$ have the potential to scale linearly with the problem size $N$.
The conditioning is
a property of the global stiffness matrix $K$. We say that $K$ is
ill-conditioned, if the nodal displacements $\underline{u}$ are
highly sensitive to rounding errors and to small changes in
$\underline f$. Generally, simulations of composite structures are
ill-conditioned because of the strong heterogeneity
and the often complex, non-grid aligned anisotropy of the material.
The condition number of $K$, which is defined as the
ratio of the largest and the smallest eigenvalue of $K$, grows roughly
linearly with the size of the largest jump in the entries of the stiffness
tensor from one finite element to an adjacent one.
But the conditioning also worsens systematically, as the problem size
$N$ increases. This growth is typically of order $N^{2/3}$ in 3D
problems. This growth affects the accuracy of the solution in direct
solvers, but it has no effect on the computational
cost.
By contrast, for iterative methods, such as Gauss-Seidel, conjugate gradients
or GMRES methods, the conditioning of $K$ has a direct impact on the
number of iterations and thus on the cost. The basic idea of iterative
methods is to approximate the solution iteratively, starting from some
initial guess, by
sequentially reducing the residual $$r=K\underline{u}-\underline{f}$$
at each iteration \cite{IterativeSparse}. A 'good' or fast iterative solver reduces this
error quickly, in a few iterations. For well-conditioned problems,
standard black-box iterative solvers (as offered by ABAQUS) work
well. But the number of iterations typically grows with the
square-root of the condition number of $K$. For the poorly
conditioned problems of interest in this paper,
standard iterative methods will either converge very slowly or
completely fail to do so.
A standard approach to improve the conditioning of such problems is to
apply a {\em preconditioner} $P$ to (\ref{eq-1}), typically a cheap approximation of
$K^{-1}$, and then to iteratively solve
\begin{equation}\label{eq-pre}
P K\underline{u}=P \underline{f}.
\end{equation}
The better the approximation $P \approx K^{-1}$, the closer the system
matrix in \eqref{eq-pre} is to the identity and thus to a
well-conditioned problem with condition number close to $1$. However,
as outlined above, computing the inverse of $K$ (e.g. via
factorisation) is too expensive. Ideally the cost of applying $P$ and
the memory requirements should again only grow linearly with~$N$.
A good choice of $P$ is problem
dependent, and must be carefully designed and tuned, to obtain
fast solvers which scale well for large problems over multiple cores.
However, some well-understood common design choices for FE
discretisations of elliptic partial differential equations exist. To
achieve independence (or at most logarithmic dependence) of the number of
iterations on the problem size $N$, it is paramount to apply
multilevel preconditioners, either in the form of multigrid
\cite{BrHeMc, Vass} or multilevel domain decomposition methods
\cite{domaindecomp}.
In addition, the memory demands of such iterative
solvers are small (proportional to $N$) and the work
can be easily distributed over multiple cores.
Care has to be taken though to achieve
robustness with respect to heterogeneities and anisotropies in the
material properties. Since black-box multilevel approaches do not scale
indefinitely and are not robust, more tailor-made approaches are needed. The
flexibility to prescribe or implement such tailor-made
preconditioners is not available in commercial software, and therefore the
development of codes like \textit{dune-composites} is pivotal in solving
large-scale problems in composite structures.
\begin{figure}[t]
\centering
\includegraphics[scale = 0.8]{domaindecomp2.pdf}\\
\caption{Decomposition of a mesh into several subdomains
($\Omega$), showing an overlap of one element in each subdomain.}
\label{fig-domaindecomp}
\end{figure}
In \textit{dune-composites} we use the conjugate gradient (CG) iterative
method for \eqref{eq-pre} and we test two types of preconditioner: the
generic Algebraic Multi Grid (AMG) preconditioner \cite{dune3}, as
well as a new two-level overlapping domain decomposition
preconditioner (GenEO) with a
multiscale coarse space based on local, generalised eigensolves in
each of the subdomains~\cite{geneo2}.
For their parallelisation, both methods partition the physical
domain and the FE mesh into subdomains of roughly equal size and distribute
them across multiple cores (see Figure \ref{fig-domaindecomp}).
The key ingredient in both preconditioners is a coarse approximation
of $\underline{u}$ with significantly fewer degrees of freedom that is
cheaper to obtain then $\underline{u}$. This coarse approximation
is then interpolated back to the original FE mesh.
In Algebraic Multigrid, the problem is coarsened repeatedly by
aggregating degrees of freedom in the stiffness matrix $K$, creating a hierarchy of
coarse approximations of decreasing size. The aggregation process is
purely algebraic and based on the fact that the displacement
$u_i$ and $u_j$ at two neighbouring nodes $x_i$ and $x_j$ in the
FE mesh will be similar if the nodes are ``strongly connected'', i.e.
$u_i \approx u_j$ if $|K_{i,j}|$ is
large (relative to $|K_{ii}|$ and $|K_{jj}|$). The degrees of freedom
at two (or more) nodes that are strongly connected are then aggregated
to a single degree freedom to obtain a cheaper approximation. This
process can be applied recursively until a sufficiently coarse
approximation has been obtained that can be efficiently solved using a
direct method. On the intermediate levels and on the original mesh we
only apply a cheap relaxation method such as Gauss-Seidel. The success of the
AMG preconditioner largely depends on how well this aggregation
process works (for details see \cite{dune3,Vass}).
\begin{figure}[t]
\centering
\includegraphics[scale=0.14]{eigen-ref.png}
\includegraphics[scale=0.14]{eigen-3.png}
\includegraphics[scale=0.14]{eigen-1.png}
\includegraphics[scale=0.14]{eigen-2.png}
\caption{Schematic of low energy eigenmodes in a reference block consisting of a layer of resin (white)
between two layers of plies (green: $90^\circ$, blue: $-45^\circ$, red:
$45^\circ$); from left to right, a reference block,
a zero energy mode (rotation), a low energy mode
between two $90^\circ$ plies and a low-energy twisting mode caused by $45^\circ$ and
$-45^\circ$ plies.}
\label{fig-eig}
\end{figure}
The GenEO preconditioner is of overlapping Schwarz-type
\cite{domaindecomp} and uses only two levels. The domain is
partitioned into overlapping subdomains of roughly equal size (see
Figure \ref{fig-domaindecomp}). In each of these subdomains, the
original FE problem subject to zero-displacement boundary conditions
on the artificial subdomain boundaries is solved either exactly using
a direct method or approximately using, for example the AMG
preconditioner. To capture the global behaviour of the solution, a coarse
problem is introduced. For homogeneous structures, standard finite
elements on a coarser mesh are sufficient \cite{domaindecomp}. For composite
structures, especially in the presence of defects, this approach is
too inaccurate. Instead, the GenEO preconditioner computes a few of
the smallest energy (eigen) modes of the structure on each of the
subdomains. This includes the rigid body modes, but also some more exotic
modes that arise due to the particular properties of the composite.
Some of these low energy modes can be seen in Figure
\ref{fig-eig}. Finally, these local modes are combined to
generate a coarse global approximation using a so-called {\em
partition of unity} approach \cite{domaindecomp}. The coarse problem is again solved
using a direct method or an iterative method. For details see the original paper
\cite{geneo2} and the related works \cite{EfGa, BabLip}.
The GenEO coarse space has been shown to lead to a fully
robust two-level Schwarz preconditioner which scales well over
multiple cores \cite{geneo2,geneoparallel}. For isotropic problems
this has been proved rigorously in \cite{geneo2}. The robustness is
due to its good approximation
properties for problems with highly heterogeneous material
parameters. This is in fact of independent interest
\cite{BabLip,Dodwell17}. Details of our specific implementation of GenEO in DUNE are given in
\cite{Seel}.
\subsection{Performance Tests and Parallel Efficiency}\label{sec-time}
First, we compare the performance of various preconditioners for the iterative solver in
\textit{dune-composites} when simulating the problem specified in Section
\ref{sec-verify}, as well as benchmarking them against the direct solver
in ABAQUS. The iterative solvers in ABAQUS only converge for
homogeneous problems or for very small numbers of degrees of freedom.
The measures for comparing the solvers are their overall computational
cost and their parallel efficiency. All the experiments in Figure
\ref{fig-sequential} were carried out on the computer {\tt cts04}
which consists of four 8-core Intel Xeon E5-4627v2 Ivybridge
processors, each running at 1.2 GHz and giving a total of 32 available cores.
On the left in Figure \ref{fig-sequential}, we compare the iterative solver with
AMG preconditioner in DUNE (with tolerance $10^{-4}$) and the sparse direct solver in
ABAQUS for increasing problem sizes. We see that the
iterative solver in DUNE scales linearly with
the problem size $N$ (the number of degrees of freedom) while the
cost of the direct solver in ABAQUS
grows significantly faster. Asymptotically, the growth is more than
quadratic in $N$. Moreover, the direct solver runs out of memory on one core for the
biggest problem tested in DUNE ($N \approx 7.5$ million).
We specifically use the block version of
AMG in our tests in \textit{dune-composites} and
measure strength of connection
between two blocks for the aggregation procedure in
the Frobenius norm. As the smoother, we use two iterations of symmetric
successive overrelaxation (SSOR). The number of
iterations remains almost constant and moderate in this test.
However, AMG has not originally been
designed for serendipity elements, and in its current form it does not
seem to be robust for this element,
especially in parallel.
\begin{figure}
\centering
\includegraphics[scale=0.9]{time-paralell_new.pdf}
\caption{Left: Timing comparison in terms of degrees of freedom between the
sparse direct solver in ABAQUS and the iterative CG solver (with AMG
preconditioner) in DUNE. Right: Timing comparison in terms of numbers of cores between the
sparse direct solver in ABAQUS and the iterative CG solver (with
one- and two-level domain decomposition preconditioner) in DUNE.}
\label{fig-sequential}
\end{figure}
The second comparison, on the right in Figure \ref{fig-sequential}, shows the
parallel scaling of various solvers. Here, we plot the strong scaling,
that is, the reduction in computational cost due to an increase in
the number of cores for a fixed problem size. We see that up to $32$
cores both the
parallel direct solver in
ABAQUS (green curve), as well as the iterative solver with domain
decomposition preconditioner in DUNE (red and blue curves) show almost
optimal parallel scaling, that is, the CPU time is halved every time
the number of cores doubles. We tested two versions of the domain
decomposition preconditioner. The red line corresponds to a one-level
overlapping Schwarz preconditioner with an overlap of two layers of
finite elements. The subdomain problems are solved with UMFPACK \cite{umfpack}, a
sparse direct solver, similar to the one in ABAQUS. For 1-8 cores, we use 8
subdomains and partition them evenly onto the available cores. For 16
and 32 cores, we use 16 and 32 subdomains, respectively (one per
core). The blue line
corresponds to the new GenEO preconditioner, a two-level overlapping
Schwarz preconditioner where, in addition to the subdomain solves, we
also build and solve a problem-adapted coarse problem (see Section
\ref{sec-solvers}) . This coarse
problem is again solved using UMFPACK.
The addition of the domain decomposition preconditioner was essential
for a good parallel efficiency. For some reasons which we cannot fully
explain, the parallel AMG preconditioner, which is available as a default within
DUNE did not scale well in parallel. The addition of the GenEO coarse space
dramatically reduces the number of iterations from about $400$ for the
one-level overlapping Schwarz preconditioner (diamond points) to about $40$
for GenEO (hollow circle). Both those numbers remain roughly
constant across this range of cores. In absolute terms, the lower
number of iterations for GenEO translates into an almost three-fold
reduction in computational time. Moreover, beyond 32 cores, only
the GenEO preconditioner continues to scale. The numbers of iterations
for the one-level Schwarz preconditioner start to increase with the
number of subdomains.
Finally, we note that even for the relatively small
test we have chosen here ($432,768$ elements) the iterative solver in DUNE
preconditioned with GenEO is around $5-10$ times faster than the direct
solver in ABAQUS. Moreover, the restriction to shared-memory
parallelism and the memory requirements of the direct solver in
ABAQUS, combined with the $O(N^{2})$ growth in computational cost,
make it essentially impossible to carry out a corresponding scaling
test with ABAQUS beyond $32$ cores.
By contrast, in Figure \ref{fig-geneo}, we now look at the parallel
scaling of the iterative solver with GenEO preconditioner in
\textit{dune-composites} on thousands of cores solving problems with
up to $173$ million degrees of freedom. We use again UMFPACK for all
subdomain and coarse solves in GenEO.
For this second scaling test, we move to a larger scale
problem.
The model run in this section has $39$ plies and a variable
width. For details on the setup of this
problem see Section \ref{sec-def}. We also move to a larger computer.
All the experiments in Figure \ref{fig-geneo} were carried out on the
University of Bath HPC cluster {\tt Balena}. This consists of
192 nodes each with two 8-core Intel Xeon E5-2650v2 Ivybridge
processors, each running at 2.6 GHz and giving a total of 3072 available cores.
We carry out a weak scaling
experiment, that is, we increase the problem size proportionally to
the number of cores used. For an iterative solver that scales optimally
both with respect to problem size and with the number of cores,
the computational time should remain constant in this experiment.
To scale the problem size as the number of cores $N_\text{cores}$
grows, we increase the width of the laminate. We present results for
two different setups. For the first, larger problem we have used $56$
elements along the radius, $\frac58 N_\text{cores}$ elements across the
width and $4$ elements through thickness in each of the resin and ply
layers, respectively. This setup ensures that the amount of work
handled by each core remains constant as $N_\text{cores}$ varies.
For the second, smaller problem we have halved the number
of elements through the thickness, using only 2 elements in
each resin and ply layer. In Figure \ref{fig-geneo}, we see that after
a slight initial growth the
scaling of the iterative solver in DUNE with GenEO preconditioner
is indeed almost optimal to at least $2048$ cores, allowing us to
increase the size of the tests at a nearly constant run time and thus, to solve a
problem with $173$ million degrees of freedom in just over 2 minutes.
\begin{figure}[t]
\centering
\includegraphics[scale=1.1]{time-geneo.pdf}
\caption{Scaling of {\em dune-composites} using the GenEO
preconditioner on up to $2048$ cores. Solid lines are used for the problem with
$8.4\cdot10^4$ degrees of freedom per core; dashed lines for
$4.2\cdot10^4$ degrees of freedom per core. For the larger problem,
the problem size varies from about $1.3$ million degrees of freedom on
16 cores to about $173$ million degrees of freedom on 2048 cores.}
\label{fig-geneo}
\end{figure}
\section{Defect analysis and the influence of boundaries}\label{sec-def}
In manufacturing, small localised defects in the form of misaligned fibrous layers can occur,
these defects can have a large effect on the strength of the materials.
In this section we investigate the effect that varying the maximum slope and amplitude of
a localised wrinkle with only one oscillation has on the strength of the curved laminate.
We show that we need a very high-fidelity mesh to be able to compute localised stresses and this
problem provides an appropriate application for \textit{dune-composites}.
In order to determine the shape and size of the wrinkle we used an X-ray CT scan of a typical curved
laminate with a wrinkle defect, shown in Figure \ref{fig-def} and described in more detail
below.
\begin{figure}[t]
\begin{floatrow}
\centering
\includegraphics[scale=0.35]{wrinkle-scan.pdf}
\includegraphics[scale=0.17]{defect_s4.png}
\includegraphics[scale=0.17]{defect_s3.png}
\end{floatrow}
\caption{Image of a defect. Left: A scan of a wrinkle with the $\text{sech}^2\,(\cdot)$ superimposed.
Center: A radial cut through the center of the laminate.
Right: Decay of the wrinkle across the width.}
\label{fig-def}
\end{figure}
For the defect analysis we we examine the influence of width-wise boundary conditions on the results.
The mechanical properties for the ply material and the resin interfaces
are the same as those for the previous smaller test, see Table \ref{table-1}.
The geometry of this $39$ ply problem is chosen to give a similar ratio of
width to radius as the previous test. Here we use a width
$W$ of 52mm and 10mm long limbs $L$. The inner radius $R$ of the curved section is 22mm.
The overall thickness $T$ is 9.93 mm, where the ply layers are 0.24 mm thick and the resin
interfaces are 0.015 mm thick.
The fibre angles of plies are given by the following stacking sequence
$$\left[[\mp45/90/0]_2/[\mp45]_2/90/\mp45/90/0/\mp45/0/\pm45/0/90/\pm45/90/
[\pm45]_2/[0/90/\pm45]_2\right].$$
\begin{figure}
\centering
\includegraphics{geometry-generation.pdf}
\caption{Transformation of the flat geometry into a flat geometry with a defect and finally into
the curved geometry with a defect.}
\label{fig-transformation}
\end{figure}
The boundary conditions and loading in a narrow specimen are very different
from those of the full structure. For this reason we initially model using
periodic boundary conditions
at the free edge. This gives an approximation of the effect of the wrinkle in a very wide part.
For the later tests we remove the periodic boundary conditions and add a $3$mm wide
layer of resin to the free edges. This reduces the strength of the singularity
at this edge and for larger wrinkles, ensures that failure occurs
at the center of the laminate, near the defect.
Figure \ref{fig-transformation} shows the process by which the wrinkle is created.
We start with a mesh on a flat plate, then we add the wrinkle to the flat plate and
finally transform to the curved geometry. Let $(s,l,r)$ be coordinates in the unperturbed,
flat geometry. Note, that in the final curved geometry, $s$ corresponds to an arc length,
$r$ is the radius through the part
and $l$ is the width across the part.
In the flat geometry the wrinkle is given by a perturbation in $r$, which depends
on all three coordinates. The perturbed
coordinates $(\hat s, \hat r, \hat l)$
are given by
\begin{equation}\label{eq-wrinkle}
\hat s = s, \quad \hat l = l \quad \text{and} \quad \hat r = r+f(s,l,r).
\end{equation}
If we denote by $(s_\text{def},l_\text{def}, r_\text{def})$
the location where the defect is largest, the perturbation is given by
\begin{equation}\label{eq-f}
f(s,l,r) = d \,\text{sech}^2\left( \pi \frac{s-s_\text{def}}{\pi b_1}\right)
\,\text{sech}^2\left( \pi \frac{r-r_\text{def}}{R b_2}\right)
\,\text{sech}^2\left( \pi \frac{l-l_\text{def}}{W b_3}\right),
\end{equation}
where $d$ gives the amplitude of the defect and $b_i$, $i=1,2,3,$ are parameters
giving the ``extent'' of the defect in the three directions.
For these tests we have chosen $b_2$ so that the wrinkle decays more quickly towards the inner radius,
more precisely
$$b_2 = \begin{cases}\frac{1}{2} & \text{ if } r-r_\text{def}<0\\
\frac{1}{4} & \text{ if } r-r_\text{def}>0\end{cases}.$$ Further,
$b_1=\frac{1}{5}$ and $b_3=\frac{1}{2}$. As shown in Figure \ref{fig-conana}, the sech$^2(\cdot)$
function never reaches zero, however at $\pm b_i$ it is already negligibly small ($<10^{-2}$). The coordinate $s$ ranges from $0$ to $\pi$, $r$ ranges from $R$ to $R+T$ and $l$ ranges from
$0$ to $W$, which implies that a value of $b_3=\frac{1}{2}$ leads to a wrinkle spanning half of the width of the part.
The wrinkled mesh on the cube is finally mapped to the curved section via the mapping
\begin{equation}
\begin{split}
\hat x &= g_1(\hat s, \hat l,\hat r) = L + \hat r\sin\left(\frac{\pi}{2} \hat s\right)\\
\hat y &= g_2(\hat s,\hat l,\hat r) = \hat l\\
\hat z &= g_3(\hat s,\hat l,\hat r) = \hat r \cos\left(\frac{\pi}{2} \hat s\right),
\end{split}
\end{equation}
where
$L$ is the length of the limb.
From the form of the wrinkle we can easily calculate the steepest slope by deriving
\begin{equation}
\tan^{-1}\left(\frac{d}{ds} f(s,l,r)\right) = \tan^{-1}\left(-\frac{2}{b_1}\,
\text{tanh}\left( \pi \frac{s-s_\text{def}}{\pi b_1}\right) f(s,l,r)\right)
\end{equation}
This allows us to classify wrinkles according to the steepest angle produced by the defect.
In order to ensure that the edge effect does not affect the results we give results with
periodic boundary conditions.
In Figure \ref{fig-def} we give a sample defect from a CT scan with an overlay showing
the fit to our proposed parameterisation of the wrinkle. We also show the resulting model with
the largest amplitude (and thus slope) of defect considered in the following numerical tests.
\subsection{Convergence Analysis}
The FE model used for these tests consists of approximately $7.3\cdot10^6$ degrees of freedom
in total. There are $96$ elements along the radius, $40$ across the width and
$4$ each in the resin and ply layers. To resolve the defect the model is refined towards
the location of the defect, with a bias ratio of $2$, i.e. the elements
at the center of the model are half as large as those at the edges. For the edge
treated case there is also refinement towards the resin edge with a bias ratio of
$400$ to resolve the stress concentrations that occur near the edge.
\begin{figure}[tb]
\begin{minipage}{22em}
\includegraphics[scale=1]{sech-plot.pdf}
\end{minipage}
\begin{minipage}{22em}
\includegraphics[scale=.9]{conv-analysis.pdf}
\end{minipage}
\caption{Left: Decay of the sech$^2()$ function for $3$ different values of parameter $B$.
Right: Convergence plot for a pristine part and a part with a defect with maximum slope $18^\circ$, with
labels giving the number of elements through-thickness per ply and interface layer.}
\label{fig-conana}
\end{figure}
To ensure that the FE modelling is sufficiently accurate we include a convergence analysis.
In Figure \ref{fig-conana} (right) we plot the relative error in the peak failure criterion given by
equation (\ref{eq-Camanho}) below.
As an approximation to the exact solution we used the next refinement level of the FE model
($136$ elements along the radius, $60$ across the width and $8$ each in the resin and ply layers).
For the refinement we have chosen (second to last point) the relative error has been reduced
to around $1\%$ for both the pristine model and for the model containing a defect.
For the case with a wrinkle it is important to note that in the
first three refinement levels have very similar peak failure criterion, even though the model
is not yet converged. However,
in these configurations the wrinkle is not yet sufficiently resolved and the failure criterion
output is still closer to that of the pristine model.
The models used in this analysis increase the number of elements through thickness
in each layer as follows: $1$, $1$, $2$, $4$, $6$, $6$, $8$.
The number of elements across the part and around the radius and limbs
are increased simultaneously. We can see that (especially for a part with
a defect) the number of elements through thickness is very important. With fewer than
$4$ elements we do not get a good approximation of the maximum failure criterion.
Further, when fewer elements are used the stresses drop to zero at the surface
of the laminate.
Around the curve enough elements need to be used to ensure that the wrinkle
is resolved, at least $10$ elements should be used along the length of the wrinkle. Even
with a mesh grading towards the wrinkle this still requires a large total number of elements
across the full length of the part.
\subsection{Defect analysis}
The stresses in the vicinity of a defect are complex, including
interlaminar, shear and direct stresses. Therefore a mixed mode failure criterion
is more suitable than a maximum stress criterion. The
strength of the laminates was assessed using a quadratic damage
onset criterion, defined by Camanho et al. \cite{Camanho} as
\begin{equation}\label{eq-Camanho}
\sqrt{\left(\frac{\sigma^+_{r}}{s_{33}}\right)^2+\left(\frac{\tau_{rs}}{s_{13}}\right)^2
+\left(\frac{\tau_{rl}}{s_{13}}\right)^2}=F
\end{equation}
with negative values of $\sigma_{r}$ treated as zero and the following allowables:
\begin{center}
$s_{33}=$ 61 MPa \qquad $s_{13}=$ 97 MPa.
\end{center}
\begin{figure}[tb]
\centering
\includegraphics[scale=0.4]{defect-zoom.pdf}
\caption{Stresses $\tau_{rl}$ and $\sigma_{l}$ for defect angle $4^\circ$ (left) , followed by
stresses $\tau_{rl}$ and $\sigma_{l}$ for
defect angle $18^\circ$ (right).}
\label{fig-stress}
\end{figure}
Curved laminates subjected to corner unfolding are observed to fail by delamination
of the plies \cite{fletcher2016resin}. Since this indicates failure occurs at the interface between plies, we
apply the failure criterion only in the resin-rich interface zones and failure initiates when $F=1$.
Generally, the failure
criterion is applicable in the local coordinates of the material, however since the interface
zone is isotropic, a transformation is not required.
Having established the wrinkle defect model, we now investigate the effect of varying the wrinkle
severity. Figure \ref{fig-stress} shows the effect on the interlaminar shear and direct stresses
with a wrinkle of maximum angle (slope) of $4^\circ$ and $18^\circ$ for which an opening/bending
moment of 9.58 kN mm/mm is applied. The interlaminar shear
maximises in the region of greatest slope, while the interlaminar direct stress maximises in
the region of peak amplitude. These stresses are highly localised and cause a sharp rise in
failure index through-thickness, as shown in Figure \ref{fig-max}. The pristine part is predicted
to fail near the mid-thickness, where direct interlaminar stress maximises. With a wrinkle
of slope $4^\circ$, there is a rise in failure index near the inner radius, however failure
is still predicted to first occur near the mid-thickness (as per the pristine model).
As the slope of the wrinkle is increased, so failure index near the inner radius becomes
higher than that near the mid-thickness. This is also indicated in Table \ref{table-per},
showing failure at interface $16$ for pristine and $4^\circ$ wrinkle models,
and at interface $4$ for greater slopes. For the most severe wrinkle of slope $18^\circ$,
the failure index is almost double that of the pristine model, and
hence strength is halved.
A summary of all models is shown in Table \ref{table-per}. As seen by the maximum recorded
displacement, the overall stiffness of the curved laminate is not significantly affected by
the inclusion of a wrinkle defect. This is because it is very small in comparison to the size
of the whole laminate. However, it does significantly reduce strength, based on initiation of failure.
When tested in isolation such specimens have been found to fail in a highly unstable manner, suggesting once initiation
is reached there is instantaneous propagation and catastrophic failure \cite{fletcher2016resin}.
Note that the two pristine results in Table \ref{table-per} compare well with predictions
acquired by modelling the curved laminates using ABAQUS software \cite{fletcher2016resin}. Note also that the edge-treated pristine result compares well with experimental results \cite{fletcher2016resin}. Table \ref{table-per} shows that the greatest impact of the wrinkle is to introduce significant
interlaminar shear stresses. Interlaminar direct stress is also increased, by approximately
$50\%$ between the pristine and $18^\circ$ wrinkle models; however,
the interlaminar shear stresses go from being negligible in the pristine model,
to becoming greater than the interlaminar direct stress in the $18^\circ$ model.
Note that these results do not take into account thermal pre-stresses, which occur as a result of the
high-temperature curing process and may be significant. However, the purpose of this paper is to illustrate
the importance of mesh refinement for rapidly varying through-thickness stresses.
\section{Conclusions}
In this paper, a new high-performance FE analysis tool for composite
structures
is presented and shown to accurately model the
behaviour of complex laminate structures with manufacturing
defects in a fraction of the computational time required by commercial
software packages, such as ABAQUS.
\begin{figure}[H]
\centering
\includegraphics{Camanho.pdf}
\caption{Through-thickness quadratic damage onset criterion $F$ at the center of each resin interface
(with linear interpolation between resin interfaces). Values are calculated at mid-width along
the angle of steepest slope ($2.1^\circ$ from the apex of the curve), and at a distance $r$ from the
outer radius.}
\label{fig-max}
\end{figure}
\begin{table}[H]
\resizebox{\textwidth}{!}{%
\begin{tabular}[scale=0.5]{ |p{3em}|p{3em}|p{3.4em}|p{3.4em}|p{3.4em}|c|c|c|c|c|c|p{3.4em}|}
\hline
\multicolumn{12}{|c|}{\textbf{FEA (normalised): Periodic BC }}\\
\hline
Defect slope (deg.) & max. disp. (mm) & location & interface (from outer radius) & max. failure crit. $F$
& \begin{tabular}[c]{@{}c@{}}\\[-.5em]$\sigma_{l}$\\ (MPa)\end{tabular}
& \begin{tabular}[c]{@{}c@{}}\\[-.5em]$\sigma_s$\\ (MPa)\end{tabular}
& \begin{tabular}[c]{@{}c@{}}\\[-.5em]$\sigma_r$ \\ (MPa) \end{tabular}
&\begin{tabular}[c]{@{}c@{}}\\[-.5em] $\tau_{sl}$ \\ (MPa) \end{tabular}
& \begin{tabular}[c]{@{}c@{}}\\[-.5em]$\tau_{rs}$ \\ (MPa) \end{tabular}
& \begin{tabular}[c]{@{}c@{}}\\[-.5em]$\tau_{rl}$ \\ (MPa) \end{tabular}
& M(kN)\\[-.2em]
\hline
0 & 3.99 & Mid- width & 23 & 1.00 & 33.76 & 34.90 & 61.00 & 0.19 & 0.02 & $-0.23$ & \textbf{9.58}$^1$ \\
4 & 3.98 & Mid- width & 23 & 1.07 & 37.26 & 40.72 & 65.59 & 0.17 & 0.95 & 0.47 & 8.90 \\
8 & 3.98 & Mid- width & 35 & 1.30 & 51.67 & 113.2 & 65.69 & $-8.75$ & 6.50 & 71.05 & 7.34 \\
12 & 3.98 & Mid- width & 35 & 1.65 & 45.41 & 117.4 & 79.94 & $-12.80$ & 9.97 & 97.59 & 5.78 \\
18 & 3.98 & Mid- width & 35 & 1.94 & 37.31 & 131.6 & 92.99 & $-17.37$ & 14.91 & 116.4 & 4.92 \\
\hline
\multicolumn{12}{|c|}{\textbf{FEA (normalised): Resin edge treated }}\\
\hline
0 & 5.16 & Edge & 23 & 1.16 & 80.06 & 116.25 & 11.13 & 6.75 & 28.13 & -107.5 & \textbf{8.26}$^2$ \\
18 & 5.16 & Mid- width & 35 & 1.99 & 37.68 & 144.15 & 96.67 & $-16.72$ & 116.80 & 0.57 & 4.79 \\
\hline
\multicolumn{12}{r}{}\\[-.5em]
\multicolumn{12}{r}{\footnotesize{$^1$ ABAQUS result was $9.51$ kN \cite{fletcher2016resin}}}\\
\multicolumn{12}{r}{\footnotesize{$^2$ ABAQUS result was $8.25$ kN, test average $8.65$ kN\cite{fletcher2016resin}} }
\end{tabular} %
}
\caption{Effect of the defect angle on peak stresses, on the failure criterion $F$ in interfaces
and on M the predicted moment of initiation of failure. }
\label{table-per}
\end{table}
The stress field around a wrinkle defect in composite materials
is highly localised, requiring a high fidelity mesh to adequately
capture it. Such defects that arise in composite materials during
manufacture are by nature difficult to predict. There is a case
for performing stochastic analysis: running thousands of
simulations with defects in various locations and sizes,
studying the influence on the integrity of a component.
The high number of simulations, combined with the requirement for
high fidelity in each simulation, results in a very demanding
problem computationally. The use of a bespoke FE solver, such
as the one developed here in DUNE, is essential for meeting these demands.
Current industry standard FE tools, e.g. ABAQUS, are not
able to deal with these problem sizes, largely due to the direct
linear equation solvers that are employed and the restricted
parallel scalability. Iterative solvers within ABAQUS only converge in
very simple cases. For this reason, as part of our project, a new
iterative solver had to be implemented within DUNE. The heart of this
solver is a robust preconditioner (GenEO) that leads to an almost
optimal scaling of the computational time both with respect to the
number of degrees of freedom and the number of compute cores. This has
been demonstrated for problems with up to $173$ million degrees of
freedom, achieving run times of just over $2$ minutes on $2048$ cores.
The chosen example problem evaluates the through-thickness stresses
caused by unfolding a curved laminate.
This serves to illustrate the importance of using high-fidelity meshes
and provides insight into the
influence of small manufacturing induced defects.
\section{Acknowledgements}
This work was supported by an EPSRC Maths for Manufacturing grant (EP/K031368/1).
Richard Butler holds a Royal Academy of Engineering-GKN Aerospace Research Chair in
Composites. This research made use of the Balena High Performance Computing
Service at the University of Bath.
\bibliographystyle{abbrv}
|
\section{Introduction}\label{sec:INT}
Available observational data indicating an accelerated expanding flat
universe suggests an existence of dark energy and dark matter, if the
dynamics of the background is according to general relativity~\cite{PC}~(and references therien to follow how the accelerated expansion of the universe have been detected for the first time). However,
recent research on this problem indicates alternative scenarios as well
including modification of general relativity~\cite{M1}~-~\cite{M8} and, for instance,
gravitationally induced particle creation~\cite{P1}~-~\cite{P7}. The aim of all developed
approaches is to generate an appropriate negative pressure to cancel the
attractive nature of gravity. It is clear, that the source of
anti-gravity cannot be arbitrary and cannot destroy the recent universe
and existing symmetries. Moreover, it should have appropriate properties
not to alter the dynamics started from the birth of the universe.
Therefore, suggested models should pass astrophysical and cosmological
tests and the models of dark energy and dark matter should be
constrained from available observational data. On the other hand,
possible tension existing between different observational datasets from
one hand side, and the technological limitations on the other hand,
allows a scanning of the physics of our universe up to some redshifts.
Incompleteness of this kind makes impossible to finalize our knowledge
giving the final models for dark energy and dark matter. Therefore, a
phenomenological approach to parameterize the darkness of the recent
universe in a form of dark energy and dark matter in recent literature
has been used frequently. In particular, recently, a phenomenological
modification of polytropic dark fluid has been suggested and the study
showed that the model with the following equation of state~\cite{Me1}
\begin{equation}
P = -A H^{-k} \rho^{u},
\end{equation}
where $A$, $k$ and $u$ are constants, while $H$ is the Hubble parameter,
can be used as a source of anti-gravity. $A$, $k$ and $u$ parameters
should be determined from the observational data. The interest towards
to polytropic type fluids is related to their applications in
astrophysics. On the other hand, in recent literature alternative models
of dark fluids like Chaplygin gas and various viscous dark fluids, also
able to solve the problem, systematically have been presented~\cite{Me2}~-~~\cite{Me4}~(and references therein). In
general, dark energy can be parameterized via the equation of state,
which provides a functional dependency of the pressure from the energy
density: the examples are Chaplygin gas and polytropic fluid with their different modifications. The second option
includes a parameterization of the energy density of the source and the
examples are ghost dark energy and generalized holographic dark energy
models with Nojiri-Odintsof cut-offs~(Ricci dark energy and other
holographic dark energy models are particular examples of the
holographic dark energy with Nojiri-Odintsov cut-offs)~\cite{Me5}~-~\cite{Me15}. Finally, the
third option describing dark energy can be a parameterization of the
equation of state parameter of dark energy~(see for instance \cite{S1}). It is well known that the
simplest model of dark energy is the cosmological constant $\Lambda$,
which provides results in well correspondence with available
observational data. However, there are two main problems that
$\Lambda$CDM faced and then the need of introducing of dynamical dark
energy models raised. The first attempt to build a dynamical dark energy
model has been based on the idea of varying cosmological constant
$\Lambda(t)$. Various phenomenological models of $\Lambda(t)$ have been
considered in literature successfully and there is a significant attempt
to use of a false vacuum decay to construct models of $\Lambda(t)$.
Sometimes in literature such models represented as the models of dark
energy based on quantum theory~(see for instance~\cite{U1}~-~\cite{U3} and reference therein for more information). Other models of dynamical dark
energy, besides mentioned dark fluid models, are quintessence, phantom,
quintom and k-essence scalar field dark energy models among the others~\cite{DE1}.
On the other, hand mentioned dark energy models are introduced by hand
in to the dynamics of the background, therefore a direct modification of
general relativity accounts as a straightforward way to explain the
accelerated expansion of the large scale universe. Moreover, a
modification of general relativity has proved to be useful also for the
physics of the early universe including a possibility to explain the
inflation~\cite{In1}~-~\cite{In2}~(and references therein). In this paper, we already had mentioned that general
relativity will describe the background dynamics and it is known that
with such models it is very useful to use the idea of non-gravitational
interaction~(we refer the readers to the cited papers for more information on this idea). Usually, it is accounted that the non-gravitational
interaction is a specific type of interaction which is deduced from the
properties of dark energy and dark matter. Therefore, in cosmological
models we usually consider interaction between dark energy and dark
matter only. Non-gravitational interaction is not only a
phenomenological assumption, it also allows to improve theoretical
results, therefore there is also an increasing amount of interest
towards this option/idea. It is believed, that with the understanding of
the structures of dark energy and dark matter, this question will be
understood as well. There are also some models of non-gravitational
interactions which are already can be considered as classical ones.
Moreover, there are new developments in this direction mainly in recent
literature, again based on various phenomenological modifications. In
this paper we will study new cosmological models involving new forms of
non-linear non-gravitational interactions considered in recent
literature and the main goal is to demonstrate the viability of these
models to the problems of the accelerated expansion of the recent
universe. Moreover, we will use $Om$ analysis and the relative change of
this parameter~\cite{Om}
\begin{equation}\label{eq:DeltaOm}
\Delta Om = 100 \times \left [ \frac{Om_{Model}}{Om_{\Lambda CDM}} -
1\right ]
\end{equation}
to study the possible differences between the new models and
$\Lambda$CDM model, for which $Om = \Omega^{(0)}_{dm} = 0.27$. Moreover,
$\chi^{2}$ analysis will be used in order to constrain the parameters of
the models.\\
The paper is organised as follows: In section~\ref{sec:CCDM} a
description of the background dynamics with the datasets in use are
presented. Moreover, the description of $Om$ analysis is also presented
in section~\ref{sec:CCDM}. In section~\ref{sec:MODELS}, the best fit
values of the parameters obtained during $\chi^{2}$ analysis for
appropriate datasets are presented and appropriate key consequences are
discussed for all models. Finally, discussion on obtained results and
possible future extension of considered cosmological model are
summarized in section~\ref{sec:Discussion}.
\section{Background dynamics and datasets}\label{sec:CCDM}
In case on interacting dark energy models we should take into account
that the dynamics of the energy densities of dark energy and dark matter
should be modified. In particular, the following differential equation
should be considered
\begin{equation}\label{eq:DM}
\dot{\rho}_{dm} + 3 H \rho_{dm} = Q,
\end{equation}
and
\begin{equation}\label{eq:DE}
\dot{\rho}_{de} + 3 H (\rho_{de} + P_{de} ) = -Q,
\end{equation}
where $Q$ indicates non-gravitational interaction and dark matter is
assumed to be cold with $P=0$. Such representation directly depends on
the assumption that the effective fluid in the universe will be
described as follows
\begin{equation}
P_{eff} = P_{de},
\end{equation}
\begin{equation}
\rho_{eff} = \rho_{dm} + \rho_{de}.
\end{equation}
On the other hand, in an isotropic and spatially homogeneous flat FRW c
universe, the Friedmann equations are as follows
\begin{equation}
H^{2} = \frac{8\pi G}{3} \rho_{eff},
\end{equation}
\begin{equation}
\frac{\ddot{a}}{a} = -\frac{4 \pi G}{3} ( \rho_{eff} + 3P_{eff}),
\end{equation}
and describe the background dynamics. To separate a physically
reasonable solution in case of a phenomenological assumption it is
necessary to constrain the parameters of the models. In this paper we
will use the following datasets
\begin{enumerate}
\item The differential age of old galaxies, given by $H(z)$.
\item The peak position of baryonic acoustic oscillations (BAO).
\item The SN Ia data.
\item Strong Gravitation Lensing data.
\end{enumerate}
In the case of the Observed Hubble Data, one defines chi-square given by
\begin{equation}
\chi^{2}_{OHD} = \sum \frac{\left ( H(\textbf{P},z) - H_{obs}(z)\right
)^{2}}{\sigma_{OHD}^{2}},
\end{equation}
where $H_{obs}(z)$ is the observed Hubble parameter at redshift $z$ and
$\sigma_{OHD}$ is the error associated with that particular observation,
while $ H(\textbf{P},z)$ is the Hubble parameter obtained from the model
and $\textbf{P}$ is the set of the parameters to be
determined/constrained from the dataset. On the other hand, $7$
measurements have been jointly used determining the BAO~(Baryon Acoustic
Oscillation) peak parameter to constrain the models by
\begin{equation}
\chi^{2}_{BAO} = \sum \frac{\left ( A(\textbf{P},z) - A_{obs}(z)\right
)^{2}}{\sigma_{BAO}^{2}},
\end{equation}
where the theoretical value for the $\textbf{P}$ set of the parameters
$A(\textbf{P},z)$ is determined as
\begin{equation}
A(\textbf{P},z_{1}) = \frac{\sqrt{\Omega_{m}}}{E(z_{1})^{1/3}} \left(
\frac{ \int_{0}^{z_{1}} \frac{dz}{E(z)} }{z_{1}}\right)^{2/3},
\end{equation}
with $E(z) = H(z)/H_{0}$ and $H_{0}$ is the value of the Hubble
parameter at $z=0$. For the Supernovae Data, $\chi^{2}_{\mu}$ is defined
as
\begin{equation}
\chi^{2}_{\mu} = A - \frac{B^{2}}{C},
\end{equation}
where
\begin{equation}
A = \sum {\frac{(\mu(\textbf{P},z) - \mu_{obs})^{2}}{\sigma_{\mu}^{2}}},
\end{equation}
\begin{equation}
B = \sum {\frac{\mu(\textbf{P},z) - \mu_{obs}}{\sigma_{\mu}^{2}}} ,
\end{equation}
and
\begin{equation}
C = \sum {\frac{1}{\sigma_{\mu}^{2}}}.
\end{equation}
In the last $3$ equations $\sigma_{\mu}$ is the uncertainty in the
distance modulus~\cite{Data1}. We will follow to the receipt of Ref.~\cite{Data2} and use the
data presented there in order to obtain constraints on the parameters of
the models from the strong gravitational lensing. To obtain appropriate
constraints, we will look for the set of the values of the parameters of
the models to minimize $\chi^{2}$ function defined as
\begin{equation}
\chi^{2} = \chi^{2}_{OHD} + \chi^{2}_{BAO}+ \chi^{2}_{\mu} +
\chi^{2}_{SGL},
\end{equation}
if we want to obtain the constraints using all datasets presented above.
\section{Models and data fitting}\label{sec:MODELS}
Three different types of models will be analyzed in this paper involving
different forms of non-gravitational interactions between dark energy
and dark matter. The forms of non-gravitational interactions have been considered for the first time in Ref.~\cite{Me2}. The parameters of the models to be constrained using
$\chi^{2}$ statistical tool are as follows $\textbf{P} = \{ H_{0},
\Omega^{(0)}_{dm}, A, b, u, k \}$. In order to simplify the discussion
on the results of the fit to find the best fit values of the parameters
and the discussion on a relative change of $Om$ parameter, we organized
appropriate subsections.
\subsection{Models of the first type}
In the case of the models consider in this subsection the following
general form describing the non-gravitation interaction between dark
energy and dark matter will be taken into account
\begin{equation}\label{eq:M1}
Q = 3 H b \left( \rho_{de} + \rho_{dm} + \frac{\rho_{de}^{2}}{\rho_{de}
+ \rho_{dm}}\right),
\end{equation}
where $b$ is a constant, $H$ is the Hubble parameter, while $\rho_{de}$
and $\rho_{dm}$ represent the energy density of the varying polytropic
dark fluid under the consideration and the energy density of dark
matter, respectively. However, before the presentation of the results
associated to this model we will study other two models as well, where
non-gravitational interactions between dark energy and dark matter are
particular examples obtained from Eq.~(\ref{eq:M1}).
\subsubsection{Case $1$}\label{ssec:1_1}
The study shows, that when the interaction is defined in the following
way
\begin{equation}\label{eq:M1_1}
Q = 3 H b \left( \rho_{de} + \frac{\rho_{de}^{2}}{\rho_{de} +
\rho_{dm}}\right),
\end{equation}
then using all datasets described above give the results presented in
Table~\ref{tab:Table1}.
\begin{table}
\centering
\begin{tabular}{ | l | l | l | l | l | l | p{1cm} |}
\hline
$\chi^{2}$ & $\Omega^{(0)}_{DM} (f)$ & $H_{0} (f)$ & $A$ & $b$ & $u
(f)$ & $k$ \\
\hline
$782.03$ & $0.27$ & $71.9$ & $0.965$ & $0.024$ & $1.0$ & $-0.014$\\
\hline
$781.60$ & $0.27$ & $71.9$ & $0.741$ & $0.033$ & $1.25$ & $0.463$ \\
\hline
$781.37$ & $0.27$ & $71.9$ & $0.879$ & $0.047$ & $1.5$ & $1.047$ \\
\hline
$781.23$ & $0.28$ & $71.9$ & $0.966$ & $0.033$ & $1.0$ & $-0.014$ \\
\hline
$780.90$ & $0.28$ & $71.9$ & $0.655$ & $0.043$ & $1.25$ & $0.437$ \\
\hline
$780.73$ & $0.28$ & $71.9$ & $0.879$ & $0.057$ & $1.5$ & $1.047$ \\
\hline
$782.56$ & $0.29$ & $71.9$ & $0.965$ & $0.043$ & $1.0$ & $-0.014$ \\
\hline
$782.28$ & $0.29$ & $71.9$ & $0.655$ & $0.052$ & $1.25$ & $0.437$ \\
\hline
$782.19$ & $0.29$ & $71.9$ & $0.503$ & $0.062$ & $1.5$ & $0.914$ \\
\hline
$785.81$ & $0.30$ & $71.9$ & $0.943$ & $0.052$ & $1.0$ & $-0.014$ \\
\hline
$785.62$ & $0.30$ & $71.9$ & $0.914$ & $0.066$ & $1.25$ & $0.516$ \\
\hline
$785.59$ & $0.30$ & $71.9$ & $0.542$ & $0.072$ & $1.5$ & $0.914$ \\
\hline
$790.91$ & $0.31$ & $71.9$ & $0.845$ & $0.062$ & $1.0$ & $-0.041$ \\
\hline
$790.79$ & $0.31$ & $71.9$ & $0.517$ & $0.071$ & $1.25$ & $0.384$ \\
\hline
$790.82$ & $0.31$ & $71.9$ & $0.502$ & $0.081$ & $1.5$ & $0.914$ \\
\hline
\end{tabular}
\caption{The best fit results for the model with $Q = 3 H b \left(
\rho_{de} + \frac{\rho_{de}^{2}}{\rho_{de} + \rho_{dm}}\right)$ with
$\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{SGL} + \chi^{2}_{\mu}$. $f$
means that the parameter has been fixed to the presented value in
advance befor the fit has been started.}
\label{tab:Table1}
\end{table}
The fit has been performed having the following priors on $H_{0} =
71.9$, $A \in [-2,2]$ and $b\in [-1,1]$. The presented results in
Table~\ref{tab:Table1} are for $\Omega^{(0)}_{dm} = 0.27$,
$\Omega^{(0)}_{dm} = 0.28$, $\Omega^{(0)}_{dm} = 0.29$,
$\Omega^{(0)}_{dm} = 0.30$, $\Omega^{(0)}_{dm} = 0.31$ for $u = 1$, $u =
1.25$ and $u = 1.5$, respectively. Initial priors for $u$ was $u
\in(0,3]$, while for $\Omega^{(0)}_{DM}$ was $\Omega^{(0)}_{DM} \in
[0.26,0.32]$. The value of the Hubble parameter has been chosen
according to the report of 2016 provided by Hubble Space Telescope
mission~\cite{HST}. From the results presented in Table~\ref{tab:Table1}, we see
that the minimum for $\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{SGL} +
\chi^{2}_{\mu} = 780.73$, for instance, has been obtained for $\{ H_{0},
\Omega^{(0)}_{dm}, A, b, u, k \} = \{ 71.9, 0.28, 0.879, 0.057, 1.5,
1.047 \}$ providing the best fit of theoretical model with observational
data. We would like to mention that the best fit values of the
parameters of the models has not been affected, when we have considered
$\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{\mu}$ i.e. considered
strong gravitational lensing data does not play an important role on the
best fit values of the parameters.
\begin{figure}[h!]
\begin{center}$
\begin{array}{cccc}
\includegraphics[width=80 mm]{Fig1_a.jpeg} &
\includegraphics[width=80 mm]{Fig1_b.jpeg}
\end{array}$
\end{center}
\caption{The graphical behavior of the deceleration parameter $q$ and
$\Delta Om$, Eq.~(\ref{eq:DeltaOm}), against the redshift $z$. The
considered model is free from the cosmological coincidence problem. The
form of non-gravitational interaction is given by Eq.~(\ref{eq:M1_1}).}
\label{fig:Fig1}
\end{figure}
Fig.~(\ref{fig:Fig1}) presents the graphical behavior of the
deceleration parameter $q$ and $\Delta Om$. The behavior of the
deceleration parameter indicates that considered model can explain the
accelerated expansion of the universe. Moreover, a phase transition
between decelerated expanding and accelerated expanding phases with
different redshifts is on face. On the other hand, the graphical
behavior of $\Delta Om$ presented on the right plot of
Fig.~(\ref{fig:Fig1}) represents differences between the new model and
$\Lambda$CDM standard cosmological model. In particular, the analysis
shows that for $\Omega^{(0)}_{dm} = 0.27$ and $\Omega^{(0)}_{dm} = 0.28$
cases the relative change at $z=0.0$ is about $0.1 \%$. On the other
hand, the study showed that $\{ H_{0}, \Omega^{(0)}_{dm}, A, b, u, k \}
= \{ 71.9, 0.30, 0.542, 0.072, 1.5, 0.914 \}$ case satisfies to the
known constraints from a modified two point $Om$ analysis with the
result from BOSS experiment for the Hubble parameter at $z=2.34$~\cite{BOSS}. In
this case the present value of the equation of state parameter for
considered varying polytropic dark fluid is $\omega_{de} = -1.14$ i.e.
in this case the phantom line crossing is possible and $\Delta Om
\approx -24 \%$. Moreover, if we will take into account constraint om
$\omega_{de}$ obtained from PLANCK 2015 experiment, then the
applicability of this model will be under the doubt. On the other hand,
the model with the following parameters $\{ H_{0}, \Omega^{(0)}_{dm}, A,
b, u, k \} = \{ 71.9, 0.31, 0.517, 0.071, 1.25, 0.384 \}$ with
$\omega_{de} = -1.022$ will satisfy to the mentioned
constraints~($\Delta Om \approx 5.5 \%$). Having obtained results, we
conclude that an additional data is needed in order to be able for any
future conclusion.
\subsubsection{Case $2$}\label{ssec:1_2}
On the other hand, the study shows, that when the interaction $Q$ has
the following form
\begin{equation}\label{eq:M1_2}
Q = 3 H b \left( \rho_{dm} + \frac{\rho_{de}^{2}}{\rho_{de} +
\rho_{dm}}\right),
\end{equation}
the, for instance, the minimum $\chi^{2}_{OHD} + \chi^{2}_{BAO} +
\chi^{2}_{SGL} + \chi^{2}_{\mu} = 780.67$ providing the best fit has
been obtained with $\{ H_{0}, \Omega^{(0)}_{dm}, A, b, u, k \} = \{
71.9, 0.28, 0.948, 0.047, 1.25, 0.516 \}$. In Table~\ref{tab:Table2} a
comprehensive information is provided allowing to understand how
$\Omega^{(0)}_{dm}$ and $u$~(both fixed in advance before the fit has
been started) affect on the fit results with fixed value of the Hubble
parameter $H_{0} = 71.9$. On the other hand, imposing the constraints
from a modified two point $Om$ analysis, the result from BOSS experiment
for the Hubble parameter at $z=2.34$ and the constraints on the
$\omega_{de}$ from PLANCK 2015, only one option has been survived among
presented in Table~\ref{tab:Table2}: $\{ H_{0}, \Omega^{(0)}_{dm}, A, b,
u, k \} = \{ 71.9, 0.28, 0.689, 0.057, 1.0, -0.094 \}$ with $\omega_{de}
= -1.03$ and $\Delta Om \approx 3.5$ at $z=0.0$. The transition redshift
in this case is $z_{tr} \approx 0.9$.
\begin{table}
\centering
\begin{tabular}{ | l | l | l | l | l | l | p{1cm} |}
\hline
$\chi^{2}$ & $\Omega^{(0)}_{DM} (f)$ & $H_{0} (f)$ & $A$ & $b$ & $u
(f)$ & $k$ \\
\hline
$781.78$ & $0.27$ & $71.9$ & $0.983$ & $0.019$ & $1.0$ & $-0.014$\\
\hline
$781.37$ & $0.27$ & $71.9$ & $0.672$ & $0.033$ & $1.25$ & $0.437$ \\
\hline
$781.22$ & $0.27$ & $71.9$ & $0.897$ & $0.052$ & $1.5$ & $1.047$ \\
\hline
$780.96$ & $0.28$ & $71.9$ & $0.776$ & $0.028$ & $1.0$ & $-0.067$ \\
\hline
$780.67$ & $0.28$ & $71.9$ & $0.948$ & $0.047$ & $1.25$ & $0.516$ \\
\hline
$780.68$ & $0.28$ & $71.9$ & $0.812$ & $0.062$ & $1.5$ & $1.021$ \\
\hline
$782.26$ & $0.29$ & $71.9$ & $0.776$ & $0.038$ & $1.0$ & $-0.067$ \\
\hline
$782.12$ & $0.29$ & $71.9$ & $0.948$ & $0.057$ & $1.25$ & $0.516$ \\
\hline
$782.22$ & $0.29$ & $71.9$ & $0.914$ & $0.071$ & $1.5$ & $1.047$ \\
\hline
$785.55$ & $0.30$ & $71.9$ & $0.689$ & $0.047$ & $1.0$ & $-0.094$ \\
\hline
$785.54$ & $0.30$ & $71.9$ & $0.534$ & $0.062$ & $1.25$ & $0.384$ \\
\hline
$785.74$ & $0.30$ & $71.9$ & $0.517$ & $0.076$ & $1.5$ & $0.914$ \\
\hline
$790.72$ & $0.31$ & $71.9$ & $0.689$ & $0.057$ & $1.0$ & $-0.094$ \\
\hline
$790.83$ & $0.31$ & $71.9$ & $0.534$ & $0.071$ & $1.25$ & $0.384$ \\
\hline
$791.13$ & $0.31$ & $71.9$ & $0.517$ & $0.085$ & $1.5$ & $0.914$ \\
\hline
\end{tabular}
\caption{The best fit results for the model with $Q = 3 H b \left(
\rho_{dm} + \frac{\rho_{de}^{2}}{\rho_{de} + \rho_{dm}}\right)$ with
$\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{SGL} + \chi^{2}_{\mu}$. $f$
means that the parameter has been fixed to the presented value in
advance befor the fit has been started.}
\label{tab:Table2}
\end{table}
\subsubsection{Case $3$}
The model considered in this subsection admits the interaction between
dark energy and dark matter given by Eq.~(\ref{eq:M1}). Taking into
account the same priors as it was in two other cases considered in
\ref{ssec:1_1} and \ref{ssec:1_2}, we found the best fit of theoretical
results with $\{ H_{0}, \Omega^{(0)}_{dm}, A, b, u, k \} = \{ 71.9,
0.27, 0.569, 0.024, 1.5, 0.943 \}$ giving $\chi^{2} = \chi^{2}_{OHD} +
\chi^{2}_{SGL} + \chi^{2}_{\mu} = 778.18$. The constraining of the
parameters of the model with $\chi^{2} = \chi^{2}_{OHD} + \chi^{2}_{SGL}
+ \chi^{2}_{\mu} + \chi^{2}_{BAO}$ reveals that the best fit does not
affected. On the other, when we used only $H(z)$ data, then the best fit
has been obtained with $\chi^{2}_{OHD} = 15.81$ and $\{ H_{0},
\Omega^{(0)}_{dm}, A, b, u, k \} = \{ 71.9, 0.3, 0.948, 0.037, 1.5,
1.021 \}$. The result presented here do not satisfy the constraints from
the modified two point $Om$ analysis, the result from BOSS experiment
for the Hubble parameter at $z=2.34$ and the constraints on the
$\omega_{de}$ from PLANCK 2015. Therefore, it is important to study the
model in the light of strong gravitational lensing to improve the best
fit values and see how the new results change the situation with
mentioned constraints.
\subsection{Models of the second type}
The main model to be studied in this subsection admits the following
form of non-gravitational interaction between varying polytropic dark
energy and cold dark matter
\begin{equation}\label{eq:M2_3}
Q = 3 H b \left( \rho_{de} + \rho_{dm} +
\frac{\rho_{de}\rho_{dm}}{\rho_{de} + \rho_{dm}}\right),
\end{equation}
where $b$ is the parameter and should be determined from observational
data under consideration. Before to present the main results obtained
for this model, we will discuss other models of non-gravitational
interactions, which are particular examples of more general form
presented by Eq.~(\ref{eq:M2_3}).
\subsubsection{Case $1$}
The comparison of theoretical results with observational data reveals
that when the non-gravitational interaction is given as follows
\begin{equation}\label{eq:M2_1}
Q = 3 H b \left( \rho_{de} + \frac{\rho_{de}\rho_{dm}}{\rho_{de} +
\rho_{dm}}\right),
\end{equation}
and when $H_{0} = 71.9$, $\Omega^{(0)}_{dm} = 0.27$ and $u = 1.5$ are
fixed, then with minimum $\chi^{2}_{OHD} + \chi^{2}_{BAO} +
\chi^{2}_{SGL} + \chi^{2}_{\mu} = 781.30$ the best fit will be obtained
providing the following values $A = 0.879$, $b = 0.052$ and $k = 1.044$
for the rest of the parameters. On the other hand, the best fit values
of the parameters had been obtained for $H_{0} = 71.9$,
$\Omega^{(0)}_{dm} = 0.28$ and $u = 1.5$ with $A = 0.810$, $b = 0.062$
and $k = 1.025$ giving the following minimum value $\chi^{2}_{OHD} +
\chi^{2}_{BAO} + \chi^{2}_{SGL} + \chi^{2}_{\mu} = 780.71$. Moreover, in
case of $H_{0} = 71.9$, $\Omega^{(0)}_{dm} = 0.29$ and $u = 1.5$ with
$\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{SGL} + \chi^{2}_{\mu} =
782.23$ providing the best fit has been found giving $A$, $b$ and $k$ to
be $0.689$, $0.071$ and $0.987$, respectively. Finally, the best fit
values of the parameters of the model have been found when
$\Omega^{(0)}_{dm} = 0.30$ and $\Omega^{(0)}_{dm} = 0.31$ have been
fixed in advance giving $\{A, b, u, k \} = \{ 0.931, 0.071, 1.25, 0.516
\}$ and $\{A, b, u, k \} = \{ 0.931, 0.081, 1.25, 0.516 \}$,
respectively~(the parameter $u$ and $H_{0}$ have been fixed in advance
as well). On the other hand, a joint constraint from the modified two
point $Om$ analysis, the result from BOSS experiment for the Hubble
parameter at $z=2.34$ and the constraints on the $\omega_{de}$ from
PLANCK 2015 imply that $\{H_{0}, \Omega^{(0)}_{dm}, A, b, u, k \} = \{
71.9, 0.30, 0.931, 0.071, 1.25, 0.516 \}$ with $\omega_{de} = -1.047$ is
the best result among obtained once. In this case the transition
redshift $z_{tr} \approx 0.85$ and $\Delta Om = -1.0 \%$.
\subsubsection{Case $2$}
When $H_{0} = 71.9$, $\Omega^{(0)}_{dm} = 0.27$ and $u = 1.5$ are fixed
in advance, then with minimum $\chi^{2}_{OHD} + \chi^{2}_{BAO} +
\chi^{2}_{SGL} + \chi^{2}_{\mu} = 781.22$ the best fit will be obtained
when $A = 0.707$, $b = 0.052$ and $k = 0.987$ for the model described by
the interaction of the following form
\begin{equation}\label{eq:M2_2}
Q = 3 H b \left( \rho_{dm} + \frac{\rho_{de}\rho_{dm}}{\rho_{de} +
\rho_{dm}}\right)
\end{equation}
On the other hand, for the same model with fixed $\Omega^{(0)}_{dm} =
0.28$, the following result have been obtained $\{A, b, u, k \} = \{
0.811, 0.047, 1.25, 0.478 \}$ giving $\chi^{2}_{OHD} + \chi^{2}_{BAO} +
\chi^{2}_{SGL} + \chi^{2}_{\mu} = 780.65$ minimal value. The results
corresponding to fixed $\Omega^{(0)}_{dm} = 0.29, 0.30, 0.31$ with
appropriate minimal value of $\chi^{2}_{OHD} + \chi^{2}_{BAO} +
\chi^{2}_{SGL} + \chi^{2}_{\mu}$ are presented in
Table~\ref{tab:Table3}. On the left plot of Fig.~(\ref{fig:Fig2}) the
graphical behavior of the deceleration parameter $q$ is presented
indicating the accelerated expansion of the large scale universe. On the
other hand, the right plot represents the graphical behavior of $\Delta
Om$. Similar to previously discussed cases additional constraints
indicates, that in future we should take the following $\{H_{0},
\Omega^{(0)}_{dm}, A, b, u, k \} = \{ 71.9, 0.30, 0.914, 0.052, 1.0,
-0.031 \}$ constraints providing the best fit of the theoretical results
with considered observational data. In this case the transition redshift
is $z_{tr} \approx 0.85$, while $\Delta Om \approx -1.0 \%$ and
$\omega_{de} = -1.044$.
\begin{table}
\centering
\begin{tabular}{ | l | l | l | l | l | l | p{1cm} |}
\hline
$\chi^{2}$ & $\Omega^{(0)}_{DM} (f)$ & $H_{0} (f)$ & $A$ & $b$ & $u
(f)$ & $k$ \\
\hline
$782.16$ & $0.29$ & $71.9$ & $0.502$ & $0.051$ & $1.25$ & $0.365$ \\
\hline
$785.56$ & $0.30$ & $71.9$ & $0.914$ & $0.052$ & $1.0$ & $-0.031$ \\
\hline
$790.79$ & $0.31$ & $71.9$ & $0.517$ & $0.057$ & $1.0$ & $-0.162$ \\
\hline
\end{tabular}
\caption{The best fit results for the model with $Q = 3 H b \left(
\rho_{dm} + \frac{\rho_{de}\rho_{dm}}{\rho_{de} + \rho_{dm}}\right)$
with $\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{SGL} +
\chi^{2}_{\mu}$. $f$ means that the parameter has been fixed to the
presented value in advance before the fit has been started.}
\label{tab:Table3}
\end{table}
\begin{figure}[h!]
\begin{center}$
\begin{array}{cccc}
\includegraphics[width=80 mm]{Fig2_a.jpeg} &
\includegraphics[width=80 mm]{Fig2_b.jpeg}
\end{array}$
\end{center}
\caption{The graphical behavior of the deceleration parameter $q$ and
$\Delta Om$, Eq.~(\ref{eq:DeltaOm}), against the redshift $z$. The
considered model is free from the cosmological coincidence problem. The
form of non-gravitational interaction is given by Eq.~(\ref{eq:M2_2}).}
\label{fig:Fig2}
\end{figure}
\subsubsection{Case $3$}
When the non-gravitational interaction is given by Eq.~(\ref{eq:M2_3}),
then the best fit values of the parameters of the model with $H_{0} =
71.9$, $\Omega^{(0)}_{DM} = 0.27$ and $u = 1.5$ fixed in advance are as
follows: $\{A, b, k \} = \{ 0.534, 0.028, 0.924 \}$ giving
$\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{\mu} = 563.29$. On the
other hand, with $H_{0} = 71.9$, $\Omega^{(0)}_{DM} = 0.29$ and $u =
1.5$ fixed in advance the best fit values of the other parameters of the
model are as follows: $\{A, b, k \} = \{ 0.586, 0.043, 0.946 \}$ giving
$\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{\mu} = 564.37$. Moreover,
with $\Omega^{(0)}_{DM} = 0.30$ and $\Omega^{(0)}_{DM} = 0.31$ fixed in
advance the minimal values of $\chi^{2}_{OHD} + \chi^{2}_{BAO} +
\chi^{2}_{\mu}$ with $567.70$ and $572.89$ provided the best fit of
theoretical results with observational data when $\{A, b, u, k \} = \{
0.931, 0.033, 1.0, -0.024 \}$ and $\{A, b, u, k \} = \{ 0.931, 0.038,
1.0, -0.024 \}$, respectively~($u$ also had been fixed in advance). The
results corresponding to $\chi^{2}_{OHD} + \chi^{2}_{BAO} +
\chi^{2}_{SGL} + \chi^{2}_{\mu}$ are presented in
Table~\ref{tab:Table4}. The top panel of Fig.~(\ref{fig:Fig2})
represents the behavior of the deceleration parameter $q$ and $\Delta
Om$ corresponding to the best fit values obtained for the parameters of
the model using the data from the differential age of old galaxies,
given by $H(z)$, the peak position of baryonic acoustic oscillations
(BAO) and the SN Ia data, when $\Omega^{(0)}_{DM}$, $u$ and $H_{0}$ were
fixed in advance. On the other hand the bottom panel of
Fig.~(\ref{fig:Fig3}) represents the graphical behavior of the same
parameters with the same parameters of the model fixed in advance, when
together with mentioned observational datasets the strong gravitational
lensing data has been used. After imposing the constraints as has been
discussed for the other cases we found the following picture
\begin{enumerate}
\item in case of the analysis with $\chi^{2}_{OHD} + \chi^{2}_{BAO} +
\chi^{2}_{\mu}$ we should take $\{A, b, u, k \} = \{ 0.931, 0.033, 1.0,
-0.024 \}$ for the candidate supported from considered constraints
\item in case of the analysis with $\chi^{2}_{OHD} + \chi^{2}_{BAO} +
\chi^{2}_{SGL} + \chi^{2}_{\mu}$ the results presented in
Table~\ref{tab:Table4} for $\Omega^{(0)}_{DM} = 0.29, 0.30, 0.31$ are
the candidates supported from considered constraints. However, only the
result corresponding to $\Omega^{(0)}_{DM} = 0.30$ will be accounted as
a candidate providing the best fit.
\end{enumerate}
\begin{table}
\centering
\begin{tabular}{ | l | l | l | l | l | l | p{1cm} |}
\hline
$\chi^{2}$ & $\Omega^{(0)}_{DM} (f)$ & $H_{0} (f)$ & $A$ & $b$ & $u
(f)$ & $k$ \\
\hline
$781.79$ & $0.27$ & $71.9$ & $0.983$ & $0.014$ & $1.0$ & $-0.012$ \\
\hline
$782.35$ & $0.29$ & $71.9$ & $0.655$ & $0.024$ & $1.0$ & $-0.106$ \\
\hline
$785.62$ & $0.30$ & $71.9$ & $0.983$ & $0.033$ & $1.0$ & $-0.012$ \\
\hline
$790.78$ & $0.31$ & $71.9$ & $0.983$ & $0.038$ & $1.0$ & $-0.012$ \\
\hline
\end{tabular}
\caption{The best fit results for the model with $Q = 3 H b \left(
\rho_{de} + \rho_{dm} + \frac{\rho_{de}\rho_{dm}}{\rho_{de} +
\rho_{dm}}\right)$ with $\chi^{2}_{OHD} + \chi^{2}_{BAO} +
\chi^{2}_{SGL} + \chi^{2}_{\mu}$. $f$ means that the parameter has been
fixed to the presented value in advance before the fit has been
started.}
\label{tab:Table4}
\end{table}
\begin{figure}[h!]
\begin{center}$
\begin{array}{cccc}
\includegraphics[width=80 mm]{Fig3_a.jpeg} &
\includegraphics[width=80 mm]{Fig3_b.jpeg} \\
\includegraphics[width=80 mm]{Fig3_c.jpeg} &
\includegraphics[width=80 mm]{Fig3_d.jpeg} \\
\end{array}$
\end{center}
\caption{The graphical behavior of the deceleration parameter $q$ and
$\Delta Om$, Eq.~(\ref{eq:DeltaOm}), against the redshift $z$. The
considered model is free from the cosmological coincidence problem. The
form of non-gravitational interaction is given by Eq.~(\ref{eq:M2_3}).
The top panel represents the result corresponding to the analysis with
$\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{\mu}$, while the bottom
panel represents the results for the analysis with $\chi^{2}_{OHD} +
\chi^{2}_{BAO} + \chi^{2}_{SGL} + \chi^{2}_{\mu}$.}
\label{fig:Fig3}
\end{figure}
\subsection{Models of the third type}
The third model of this paper admits the following form of
non-gravitational interaction
\begin{equation}\label{eq:M3}
Q = 3 H b \left( \rho_{de} + \rho_{dm} + \frac{\rho_{dm}^{2}}{\rho_{de}
+ \rho_{dm}}\right),
\end{equation}
which gives the best fit of theoretical results with observational data
when $\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{\mu} = 563.25$ and
$H_{0} = 71.9$, $\Omega^{(0)}_{DM} = 0.27$, $u = 1.5$ are fixed in
advance and the rest of the parameters of the model are defined as
follows: $\{A, b, k \} = \{ 0.534, 0.028, 0.924 \}$. In this case we see
that presented result coincidence with the result obtained for the
model, when the non-gravitational interaction is given by
Eq.~(\ref{eq:M2_3}). The consideration of the cases with $u = 1.0$ and
$u = 1.25$~(with fixed $H_{0} = 71.9$ and $\Omega^{(0)}_{DM} = 0.27$)
provided the best fit when $\{A, b, k \} = \{ 0.534, 0.028, 0.924 \}$
and $\{A, b, k \} = \{ 0.931, 0.014, -0.024 \}$, respectively. On the
other hand, with $u = 1.0$~(with fixed $H_{0} = 71.9$ and
$\Omega^{(0)}_{DM} = 0.29$) the best fit has been found with
$\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{\mu} = 564.25$ when $\{A,
b, k \} = \{ 0.776, 0.024, -0.068 \}$, while with fixed $H_{0} = 71.9$,
$\Omega^{(0)}_{DM} = 0.30$ and $u = 1.0$ the best fit has been found
when $\{A, b, k \} = \{ 0.707, 0.029, -0.09 \}$. The last state is
described by $\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{\mu} =
567.61$. For this model, the study of the question how the strong
gravitational lensing data will affect on the best fit values of the
parameters has been left as a topic of another study. Preliminary study
presented here, showed that the best fit result also satisfying to the
constraints imposed from BOSS and PLANCK 2015 experiments, should be
accounted the result corresponding to the fixed $\Omega^{(0)}_{DM} =
0.30$ case with $z_{tr} \approx 0.85$, $q \approx -0.58$ and $\Delta Om
= 1\%$ at $z=0.0$. Moreover, as can be seen from the right plot of
Fig.~(\ref{fig:Fig4}) $\Delta Om$ is an increasing function from the
redshift.
\begin{figure}[h!]
\begin{center}$
\begin{array}{cccc}
\includegraphics[width=80 mm]{Fig4_a.jpeg} &
\includegraphics[width=80 mm]{Fig4_b.jpeg} \\
\end{array}$
\end{center}
\caption{The graphical behavior of the deceleration parameter $q$ and
$\Delta Om$, Eq.~(\ref{eq:DeltaOm}), against the redshift $z$. The
considered model is free from the cosmological coincidence problem. The
form of non-gravitational interaction is given by Eq.~(\ref{eq:M3}). The
presented result corresponds to the analysis with $\chi^{2}_{OHD} +
\chi^{2}_{BAO} + \chi^{2}_{\mu}$. }
\label{fig:Fig4}
\end{figure}
\subsubsection{Case $1$}
In this section we will present the results of the fit for the
phenomenological model, when the non-gravitational interaction is given
in the following way
\begin{equation}\label{eq:M3_1}
Q = 3 H b \left( \rho_{de} + \frac{\rho_{dm}^{2}}{\rho_{de} +
\rho_{dm}}\right),
\end{equation}
which is a particular case of more general form given by
Eq.~(\ref{eq:M3}). First of all we would like to present the results
according to $\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{\mu}$
constrain and compare them with the results obtained from
$\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{SGL} + \chi^{2}_{\mu}$
constrain. For instance, the study shows that when $H_{0} = 71.9$,
$\Omega^{(0)}_{DM} = 0.27$ and $u = 1.5$ the minimal $\chi^{2}_{OHD} +
\chi^{2}_{BAO} + \chi^{2}_{\mu} = 563.24$ will be obtained providing the
best fit. In this case for the rest parameters we obtained $\{A, b, k \}
= \{ 0.586, 0.047, 0.946 \}$. On the other hand, when $\Omega^{(0)}_{DM}
= 0.28$, then the best fit with $\{A, b, u, k \} = \{ 0.896, 0.047,
1.25, 0.505 \}$~($\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{\mu} =
562.70$) will be obtained. Moreover, the study showed that the model
with $\Omega^{(0)}_{DM} = 0.29$, $\Omega^{(0)}_{DM} = 0.30$ and
$\Omega^{(0)}_{DM} = 0.31$ provides the best fit when $\{A, b, u, k \} =
\{ 0.896, 0.059, 1.25, 0.505 \}$~($\chi^{2}_{OHD} + \chi^{2}_{BAO} +
\chi^{2}_{\mu} = 564.20$), $\{A, b, u, k \} = \{ 0.638, 0.047, 1.0,
-0.112 \}$~($\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{\mu} = 567.63$)
and $\{A, b, u, k \} = \{ 0.638, 0.057, 1.0, -0.112 \}$~($\chi^{2}_{OHD}
+ \chi^{2}_{BAO} + \chi^{2}_{\mu} = 572.83$), respectively. Now,
including strong gravitational lensing data, we obtained the following
results. In particular, when $\Omega^{(0)}_{DM} = 0.27$ the best fit
will be obtained when $\{A, b, u, k \} = \{ 0.586, 0.047, 1.5, 0.946
\}$~($\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{SGL} + \chi^{2}_{\mu}
= 781.21$) i.e. the best fit values of the parameters will not be
affected. On the other hand, when $\Omega^{(0)}_{DM} = 0.28$, then the
consideration of strong gravitational lensing data will significantly
affect on the best fit values of the parameters - $\{A, b, u, k \} = \{
0.586, 0.057, 1.5, 0.946 \}$~($\chi^{2}_{OHD} + \chi^{2}_{BAO} +
\chi^{2}_{SGL} + \chi^{2}_{\mu} = 780.64$). The results for
$\Omega^{(0)}_{DM} = 0.29, 0.30, 0.31$ are presented in
Table~\ref{tab:Table5} and it can be seen, that including of strong
gravitational lensing data under the consideration will not affect the
best fit values of the parameters. Future constraints mentioned earlier
in this paper, support the case with $\Omega^{(0)}_{DM} = 0.30$ to be
the candidate for the best fit. On the other hand, the results
corresponding to $\Omega^{(0)}_{DM} = 0.31$ also can be counted to
satisfy to imposed constraints. The left plot of Fig.~(\ref{fig:Fig4})
indicates how the relative change $\Delta Om$ evolves with the evolution
of the universe.
\begin{table}
\centering
\begin{tabular}{ | l | l | l | l | l | l | p{1cm} |}
\hline
$\chi^{2}$ & $\Omega^{(0)}_{DM} (f)$ & $H_{0} (f)$ & $A$ & $b$ & $u
(f)$ & $k$ \\
\hline
$782.11$ & $0.29$ & $71.9$ & $0.896$ & $0.057$ & $1.25$ & $0.505$ \\
\hline
$785.56$ & $0.30$ & $71.9$ & $0.638$ & $0.047$ & $1.0$ & $-0.112$ \\
\hline
$790.73$ & $0.31$ & $71.9$ & $0.638$ & $0.057$ & $1.0$ & $-0.012$ \\
\hline
\end{tabular}
\caption{The best fit results for the model with $Q = 3 H b \left(
\rho_{de} + \frac{\rho_{dm}^{2}}{\rho_{de} + \rho_{dm}}\right)$ with
$\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{SGL} + \chi^{2}_{\mu}$. $f$
means that the parameter has been fixed to the presented value in
advance before the fit has been started.}
\label{tab:Table5}
\end{table}
\begin{figure}[h!]
\begin{center}$
\begin{array}{cccc}
\includegraphics[width=80 mm]{Fig5_a.jpeg} &
\includegraphics[width=80 mm]{Fig5_b.jpeg} \\
\end{array}$
\end{center}
\caption{The graphical behavior of the deceleration parameter $q$ and
$\Delta Om$, Eq.~(\ref{eq:DeltaOm}), against the redshift $z$. The
considered model is free from the cosmological coincidence problem. The
form of non-gravitational interaction is given by Eq.~(\ref{eq:M3_1}).
The presented result corresponds to the analysis with $\chi^{2}_{OHD} +
\chi^{2}_{BAO} + \chi^{2}_{\mu}$.}
\label{fig:Fig5}
\end{figure}
\subsubsection{Case $2$}
The last model studied in this work admits the following form of
non-gravitational interaction
\begin{equation}\label{eq:M3_2}
Q = 3 H b \left( \rho_{dm} + \frac{\rho_{dm}^{2}}{\rho_{de} +
\rho_{dm}}\right).
\end{equation}
During the study of the model, when in addition to $\chi^{2}$ analysis
with $\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{SGL} +
\chi^{2}_{\mu}$, we applied the constraints from BOSS and PLANCK 2015
experiments, and take into account the constrains from modified
two-point $Om$ analysis gives us the best fit values of the parameters
of the model as follows: $\omega_{de} \approx -1.039$ at $z=0.0$ with
$\{ H_{0}, \Omega^{(0)}_{dm}, A, b, u, k \} = \{ 71.9, 0.30, 0.638,
0.047, 1.0, -0.112 \}$. The graphical behavior of the deceleration
parameter is presented on the left plot of Fig.~(\ref{fig:Fig6}), while
the graphical behavior of $\Delta Om$ is presented on the right plot.
Both clearly indicates how the mentioned parameters evolve during the
evolution of the universe, moreover, it is possible also to estimate the
present day values of them very easily. During the study of the behavior
of the equation of state parameter of considered polytropic fluid for
all models we found two possibilities. In particular we observed that
for some models~(the difference between the models is the form of
non-gravitational interaction) $\omega_{de} > 0$ at high redshifts and
there is a phase transition to a phantom dark fluid state satisfying to
the constraints on $\omega_{de}$ according to PLANCK 2015 experiment.
However, there is also possibility to have a phantom - phantom
transitions also providing the accelerated expansion of the universe. We
would like to mention that mentioned phantom - phantom transitions have
been observed firstly in scope of generalized holographic dark energy
models with Nojiri-Odintsov cut-offs.
\begin{table}
\centering
\begin{tabular}{ | l | l | l | l | l | l | p{1cm} |}
\hline
$\chi^{2}$ & $\Omega^{(0)}_{DM} (f)$ & $H_{0} (f)$ & $A$ & $b$ & $u
(f)$ & $k$ \\
\hline
$781.17$ & $0.27$ & $71.9$ & $1.0$ & $0.038$ & $1.25$ & $0.527$ \\
\hline
$780.59$ & $0.28$ & $71.9$ & $0.759$ & $0.043$ & $1.25$ & $0.461$ \\
\hline
$782.10$ & $0.29$ & $71.9$ & $0.948$ & $0.038$ & $1.0$ & $-0.024$ \\
\hline
$785.51$ & $0.30$ & $71.9$ & $0.586$ & $0.043$ & $1.0$ & $-0.134$ \\
\hline
$790.75$ & $0.31$ & $71.9$ & $0.534$ & $0.047$ & $1.0$ & $-0.156$ \\
\hline
\end{tabular}
\caption{The best fit results for the model with $Q = 3 H b \left(
\rho_{dm} + \frac{\rho_{dm}^{2}}{\rho_{de} + \rho_{dm}}\right)$ with
$\chi^{2}_{OHD} + \chi^{2}_{BAO} + \chi^{2}_{SGL} + \chi^{2}_{\mu}$. $f$
means that the parameter has been fixed to the presented value in
advance before the fit has been started.}
\label{tab:Table7}
\end{table}
\begin{figure}[h!]
\begin{center}$
\begin{array}{cccc}
\includegraphics[width=80 mm]{Fig6_a.jpeg} &
\includegraphics[width=80 mm]{Fig6_b.jpeg} \\
\end{array}$
\end{center}
\caption{The graphical behavior of the deceleration parameter $q$ and
$\Delta Om$, Eq.~(\ref{eq:DeltaOm}), against the redshift $z$. The
considered model is free from the cosmological coincidence problem. The
form of non-gravitational interaction is given by Eq.~(\ref{eq:M3_2}).}
\label{fig:Fig6}
\end{figure}
\section{\large{Discussion}}\label{sec:Discussion}
Available observational data suggests to include dark energy and dark
matter in general relativity to explain the accelerated expansion of the
universe. On the other hand, it is possible to introduce
non-gravitationally interacting dark energy and dark matter to solve the
problems of the large scale universe. The duality is on face with its
problematic consequences. From one hand side, seems that it will be easy
to parameterize the dark side of the universe and solve the problems,
but from the other hand side, it appears that existing tension between
observational datasets makes additional complexity. Therefore, there is
ongoing active research in order to find a solution involving many
phenomenological assumptions. In particular, there are different
phenomenological assumptions concerning to the form of dark energy and
non-gravitational interaction. Motivated by recent developments, in this
paper we considered new cosmological models involving new interacting
varying polytropic gas models. In order to obtain the best fit values of
the parameters of the models we used $\chi^{2}$ analysis involving the
differential age of old galaxies, given by $H(z)$, the peak position of
baryonic acoustic oscillations known as BAO data, the SN Ia data and
strong gravitation lensing data. To simplify the analysis we fixed the
values of some of the parameters before the fit has been start and kept
them frozen in future, we involved constraints on the equation of state
parameter of dark fluid $\omega_{de}$ from PLANCK 2015 experiment, then
we took into account reported value for the Hubble parameter at $z =
2.34$ from BOSS experiment. Moreover, we used constraints obtained from
a modified two-point $Om$ analysis giving $Omh^{2}(z_{1};z_{2}) = 0.124
\pm 0.045$,
$Omh^{2}(z_{1};z_{3}) = 0.122 \pm 0.01$ and $Omh^{2}(z_{2};z_{3}) =
0.122 \pm 0.012$ for $z_{1} = 0$, $z_{2} = 0.57$ and $z_{3} = 2.34$,
respectively~\cite{SahniFin}. Mentioned additional constrains allowed us
to establish the best fit values of the parameters of the models with
fixed $H_{0}$, $u$ and $\Omega^{(0)}_{DM}$. The study shows, that
considered models, which are differ from each other by the form of
non-linear non-gravitational interactions between dark energy and dark
matter can explain the accelerated expansion. It is possible to explain
the phase transition between decelerated expanding and accelerated
expanding phases during the evolution of the universe. Moreover, the
study of the relative change of $Om$ parameter shows clear difference
between new models and $\Lambda$CDM standard model of cosmology. The
main interesting result has been observed during the study of the
equation of state parameter of polytropic fluid. In particular, the
study shows that considered models when non-gravitational interactions
are given by Eq.~(\ref{eq:M1_1}), Eq.~(\ref{eq:M1_2}),
Eq.~(\ref{eq:M1}), Eq.~(\ref{eq:M2_1}), respectively, then quintessence
- phantom transition for the equation of state parameter of dark energy
will be observed. On thi other hand, when we consider non-gravitational
interactions given by Eq.~(\ref{eq:M2_2}), Eq.~(\ref{eq:M2_3}),
Eq.~(\ref{eq:M3}), Eq.~(\ref{eq:M3_2}), then phantom - phantom phases
unification will be observed. However, when the interaction, for
instance, is given by Eq.~(\ref{eq:M1_2}), then the results
corresponding to $\Omega^{(0)}_{DM} = 0.31$ satisfying the the
considered constraints, provides dark energy with phantom - phantom
transition. On the other hand, when the interaction is given by
Eq.~(\ref{eq:M1_2}), then we will observe also quintessence - phantom
transition. In summary - we need more observational data in order to be
able to choose the best model of interacting varying politropic dark
fluid model from the models considered in this paper. It can be done,
for instance, involving constraints provided by the study of the
structure formation. Moreover, this will allow to define which one of
mentioned transitions for dark energy is the realistic scenario, because
from the study of the deceleration parameter and the equation of state
parameter this question cannot be answered. If it will be found that a
model with phantom - phantom transition will be the best model among
considered models, then it is necessary to find Nojiri-Odintsov
holographic dark energy representation of the model and study the
application of the model to the inflationary expansion phase of the
universe. In future research reconstruction of modified theories of
gravity for considered models should be performed and since we saw a
non-unique imprint of the type of non-gravitational interaction on the
behavior of the equation of state parameter of dark fluid, then it is
necessary to determine the type of future singularities which will
provide additional sources to extend applied constraints used in this
paper. Mentioned possibilities and tasks are the subject of additional
research and will be reported in another paper.
\section*{\large{Acknowledgments:}}
Martiros Khurshudyan appreciates Prof. K. Urbanowski from Institute of
Physics, University of Zielona Gora, for valuable comments and
suggestions during the preparation of the paper.
\section*{\large{Author Contributions:}}
Martiros Khurshudyan designed the main subject of this research and
mainly wrote the manuscript.
Asatur Khurshudyan commented on the manuscript at all stages with
performing part of the analysis. All authors equally promoted the
research and discussed the results. All authors have read and approved
the final manuscript.
\section*{\large{Conflicts of Interest:}} The authors declare no
conflict of interest.
|
\section{Introduction}
\label{sec:Introduction}
Ultra-cold atomic sources with high state purity are the basis of many experiments that aim to study fundamental physics through the use of quantum simulators, or via precision measurements using atom interferometers \cite{Borde1989, Kasevich1991, Gillot2014, Rosi2014, Kovachy2015, Dutta2016} or atomic clocks \cite{Kasevich1989, Gibble1993, Santarelli1999, Nicholson2012, Hinkley2013}. These fields have grown more interested in exotic species that exhibit interesting properties that can be exploited to study new physical phenomena. However, lighter atoms such as the isotopes of lithium and potassium are rarely proposed as metrological tools. This is primarily because these atoms have closely-spaced energy structures and relatively large recoil velocities---making them notoriously difficult to cool and manipulate \cite{Cataliotti1998, Fort1998, Modugno1999, Landini2011, Gokhroo2011, Nath2013}. Additionally, their compact hyperfine splitting and lighter mass makes them much more sensitive to parasitic effects, such as AC Stark shifts, Zeeman shifts and forces due to electro-magnetic fields. Nevertheless, in some cases these properties can be advantageous---for instance, a larger sensitivity to atomic recoil could lead to a new determination of the fine structure constant \cite{Bouchendira2011, Cassella2016}. Similarly, precision measurements of different hyperfine splittings can shed new light on collisional interactions \cite{Gibble1995, Kokkelmans1997} and the variation of fundamental constants \cite{Blatt2008, Rosenband2008}.
The starting point for many experiments aimed at studying fundamental physics is to prepare a pure sample in terms of its energy, spin and momentum before injecting into an atom interferometer, spectrometer or quantum simulator. Previously, ultra-cold potassium isotopes have been spin-polarized via optical pumping \cite{Modugno2001, Roati2002, Roati2007, Fattori2008, Campbell2010}, adiabatic radio-frequency sweeps \cite{Wille2008}, and spin relaxation in a two-component mixture \cite{Spiegelhalder2010}. However, the majority of preparation schemes demonstrated so far have been carried out on magnetically-sensitive states which are placed in optical or magnetic traps for evaporation to quantum degeneracy. For applications in metrology, it is usually more desirable to use a magnetically-insensitive state to carry out measurements, yet this is a challenging task for species like potassium.
In this work, we report on an all-optical technique to prepare $^{39}$K atoms in the $\ket{F = 1, m_F = 0}$ state with $\sim 95$\% purity, while maintaining a temperature of 6 $\mu$K from a sample initially cooled to $\sim 4$ $\mu$K in a gray molasses. We use this high-purity, low-temperature atomic source in an atom-interferometric gravimeter \cite{LeGouet2008}, where we achieve a sensitivity to gravitational acceleration of 60 $\mu$Gal (1 $\mu$Gal = $10^{-8}$ m/s$^2$) after 4800 s of integration---representing the state-of-the-art achieved with potassium so far. We also perform Ramsey-type hyperfine spectroscopy \cite{Ramsey1949, Kasevich1989} with this source and demonstrate a sensitivity of $4.1 \times 10^{-11}$ to the ground state splitting of $^{39}$K, which corresponds to more than a factor of 30 improvement over previous measurements \cite{Arimondo1977}.
The rest of this article is organized as follows. In Sec.~ \ref{sec:Sequence}, we overview the laser-cooling and purification process. In Sec.~ \ref{sec:Metrology} we discuss experimental results pertaining to two ongoing precision measurements with $^{39}$K. Finally, we give our conclusions and perspectives in Sec.~ \ref{sec:Conclusion}.
\section{Cooling and purification}
\label{sec:Sequence}
\begin{figure*}[!tb]
\centering
\includegraphics[width=0.98\textwidth]{Fig1-KBeams-StatePrepSequence.pdf}
\caption{(a) Orientation of optical beams for cooling, state preparation and interrogation of the atoms. Light from both D1 and D2 lasers is overlapped with the same polarization along six beams used for the MOT loading, molasses cooling and optical pumping stages. The selection beam (D2 light) is configured as either a push beam to remove atoms in $\ket{F = 2}$, or as a co-propagating Raman beam to make coherent two-photon transitions with circular polarization. The counter-propagating Raman beams (D2 light) have perpendicular linear polarizations in order to enhance velocity-sensitive transitions and suppress velocity-insensitive ones. These beams are also used to detect the internal state of the atoms via resonant fluorescence on the $\ket{F = 2} \to \ket{F' = 3}$ transition. An external pair of Helmholtz coils (not shown) are used to generate a uniform magnetic bias field $\bm{B}$ along the vertical direction during the selection and interrogation phases. (b) State preparation sequence. The function of each optical pulse on the atomic states is shown in the first row. The relative intensities and frequencies of the selection and molasses beams are shown in the remaining rows. Note that the selection beam has two frequencies (labelled as $\omega_1$ and $\omega_2$) to make Raman transitions. The steps circled in the red loop indicate the transfer sequence, which is repeated to increase the fraction of atoms in the target state. At the end of this loop, a near-resonant push beam removes atoms remaining in $\ket{F = 2}$ to improve the purity of the target state.}
\label{fig:KBeams+StatePrepSequence}
\end{figure*}
The first magneto-optical traps (MOTs) with the isotopes of potassium were achieved in the late 1990s \cite{Williamson1995, Wang1996, Cataliotti1998, Modugno1999}. Cooling and manipulation of potassium is challenging primarily as a result of the small energy separation between hyperfine levels in the $4^2 P_{3/2}$ excited state. This effectively opens the ``closed'' transitions on the D2 line that are largely forbidden in other alkali atoms, such as rubidium and cesium, and creates a complex sub-Doppler cooling mechanism that involves all of the hyperfine excited states. Various sub-Doppler cooling schemes have previously been explored using standard red-molasses techniques \cite{Fort1998, Landini2011, Gokhroo2011} that have achieved $^{39}$K temperatures as low as 25 $\mu$K. Recently, gray-molasses cooling \cite{Boiron1995} has been demonstrated as an efficient mechanism for alkali atoms with compact level structures, such as the isotopes of lithium \cite{Grier2013, Burchianti2014} and potassium \cite{RioFernandes2012, Nath2013, Salomon2013, Sievers2015}. This sub-Doppler cooling method, which combines the Sisyphus effect \cite{Dalibard1989} with velocity-selective coherent population trapping \cite{Aspect1989}, requires a light source tuned to the blue of the D1 transition. So far, temperatures as low as 6 $\mu$K have been demonstrated in $^{39}$K samples using this technique \cite{Salomon2013}. More recently, an ultra-low temperature of 1.8 $\mu$K was achieved using degenerate Raman sideband cooling in an optical lattice \cite{Grobner2017}. These conditions are ideal for further cooling large fractions to quantum degeneracy, particularly with all-optical techniques \cite{Salomon2014}.
In this work, we have adopted a cooling scheme similar to that of Ref.~ \cite{Salomon2013}. Both our D1 (770 nm) and D2 (767 nm) light sources are coupled through the same fiber splitter to generate a 6-beam MOT in the 1-1-1 configuration, as shown in Fig.~ \ref{fig:KBeams+StatePrepSequence}(a). Potassium-39 atoms are loaded in a 3D MOT from a background vapor in 0.5 s. We then switch off the magnetic gradient coils and the D2 light, and switch on the D1 light to cool the atoms in a gray molasses for 7 ms. During this time, the D1 cooling beam is blue detuned by 20 MHz from the $\ket{F=2} \to \ket{F'=2}$ transition, while the repump beam is detuned by the same amount from the $\ket{F=1} \to \ket{F'=2}$ transition---satisfying the critical Raman resonance condition for efficient gray-molasses cooling. To reduce additional photon scattering during the molasses, the cooling and repump beam intensities are decreased linearly from $18.5 \, I_{\rm{sat}}$ to $2.4 \, I_{\rm{sat}}$ and from $4.5 \, I_{\rm{sat}}$ to $0.6 \, I_{\rm{sat}}$, respectively ($I_{\rm sat} = 1.75$ mW/cm$^2$ \cite{Tiecke2011}). At the end of the molasses, we apply a short ($\sim 300$ $\mu$s) pulse using the D1 cooling light tuned near the $\ket{F=2} \to \ket{F'=2}$ transition in order to depump the atoms into $\ket{F = 1}$. We obtain a sample of $\sim 6 \times 10^7$ atoms at a temperature of $4.3 \pm 0.2$ $\mu$K, roughly equally distributed in population among the magnetic sub-levels of $\ket{F = 1}$. We verify the temperature of the sample using two methods; (i) by using time-of-flight (TOF) images recorded with a charge-coupled device (CCD) camera, and (ii) using velocity-sensitive two-photon Raman spectroscopy \cite{Moler1992} along the vertical direction \footnote{The Raman spectroscopy method of measuring temperature requires an increase in the bias field to $\sim 1$ G in order to resolve the velocity distribution of each magnetic sub-level.}.
\begin{figure*}[!tb]
\centering
\includegraphics[width=0.98\textwidth]{Fig2-KSampleProperties.pdf}
\caption{(a) Velocity-insensitive Raman spectrum of the sample before (black filled circles) and after (red open circles) state preparation using a single transfer sequence and a purifying push pulse. These data were obtained with a circularly-polarized Raman beam and by blocking the retro-reflected light along the vertical axis [see Fig.~ \ref{fig:KBeams+StatePrepSequence}(a)]. We estimate the relative population in each state from the amplitude obtained by least-squares fits to each resonance using a sinc$^2(\delta)$ lineshape. (b) Population in the target state $\ket{1,0}$ (filled symbols) and the total population in $\ket{1,\pm 1}$ (open symbols) as a function of the number of transfer sequences. Circles correspond to measurements made after $N$ transfer sequences with the push pulse replaced by a depump pulse [see Fig.~ \ref{fig:KBeams+StatePrepSequence}(b)]. Square symbols represent separate measurements including the push pulse, which acts to further purify the sample in the target state. (c) Temperature (open circles) and atom number (filled circles) versus the transfer sequence number. The initial decrease in atom number is due to the push pulse, which is applied only for $N > 0$, but thereafter remains roughly constant. The linear increase in temperature is due to the additional D1 optical pumping pulses made with each transfer sequence.}
\label{fig:KSampleProperties}
\end{figure*}
Although a sub-Doppler-cooled sample is sufficient to proceed with some types of measurements \cite{Barrett2011a}, for a broad variety of experiments a distributed population amongst magnetic sub-levels is undesirable and a purification step is often employed. However, manipulating the state populations of potassium (along with other species exhibiting a light mass and a quasi-open set of electric dipole transitions) is complicated by its high susceptibility to scattering-induced heating and optical pumping. Additionally, the use of microwave transitions can be frustrated by the size and material of the vacuum system. For instance, the $\ket{F = 1, m_F = 0} \to \ket{F = 2, m_F = 0}$ clock transition in $^{39}$K at 461.7 MHz has a relatively large wavelength of $\sim 65$ cm, and small-volume metal chambers cannot support this field. We circumvent these challenges by combining optical Raman pulses to select atoms in the desired $m_F = 0$ state with optical pumping between hyperfine manifolds using near-resonant D1 light. The optical selection pulse plays the role of a microwave, since one can induce coherent state inversion with high efficiency and the optical fields can be kept far from single-photon resonances so as to avoid spontaneous emission. The advantage of optical pumping on the D1 line is that there are no cycling transitions, hence atoms scatter a minimum number of photons before falling into the desired ground state manifold.
Our state preparation sequence is outlined in Fig.~ \ref{fig:KBeams+StatePrepSequence}(b). We first apply a magnetic bias field of 0.25 G along the vertical axes in order to quantize and separate the magnetic sub-levels by a few hundred kHz. This field is servo-locked by measuring the field near the atoms with a flux-gate sensor (Bartington MAG-03MCTPB500) and feeding back on the current driving a large pair of compensation coils around the chamber. The sample is initially depumped to $\ket{F = 1}$ using D1 light along the MOT beams. Atoms in the $\ket{F = 1, m_F = 0}$ state are then transferred coherently to $\ket{F = 2, m_F = 0}$ using a co-propagating Raman $\pi$-pulse \footnote{To make $\Delta m_F = 0$ transitions, the two co-propagating Raman beams used for selection must have the same circular polarization, with a component of the electric field perpendicular to the quantization axis defined by the bias magnetic field.}. The atoms remaining in the $\ket{F=1}$ manifold are then optically pumped to $\ket{F = 2}$ using a repump pulse resonant with the $\ket{F = 1} \to \ket{F' = 2}$ D1 transition. Since the optical pumping light is isotropically polarized, this operation distributes the population remaining in $\ket{F = 1}$ roughly equally among the Zeeman states in $\ket{F = 2}$. We then drive the $\ket{F = 2, m_F = 0} \to \ket{F = 1, m_F = 0}$ transition with a second Raman pulse. This ensures that the accumulation of atoms in $m_F = 0$ from the optical pumping pulses is not reversed by subsequent pulses. Henceforth, we refer to this four-step process as a ``transfer'' sequence, as depicted in Fig.~ \ref{fig:KBeams+StatePrepSequence}(b). This transfer sequence can be repeated a number of times to further increase the fraction of atoms in the target state $\ket{F = 1, m_F = 0}$, but at the expense of additional heating to the sample. Finally, we apply a short pulse of D2 light resonant with the $\ket{F = 2} \to \ket{F' = 3}$ transition to push away any atoms remaining in $\ket{F = 2}$. We emphasize that this technique can easily be adapted to transfer atoms into any other target state with a suitable choice of the selection beam frequency, polarization and orientation relative to the $B$-field axis.
\begin{figure*}[!tb]
\centering
\includegraphics[width=0.98\textwidth]{Fig3-KGravimeter.pdf}
\caption{(a) Interference fringes for an interrogation time of $T = 100$ $\mu$s and a $\pi$-pulse duration $\tau_{\pi} = 6$ $\mu$s. Here, each point represents a single repetition of the interferometer as a function of a controlled phase shift $\phi_{\rm las}$ applied to the Raman laser before the final $\pi/2$-pulse. Atoms cooled to $\sim 16$ $\mu$K using a standard red molasses (red open circles) yield a fringe contrast of 10\% and SNR $\simeq$ 20. Cooling the atoms to $\sim 5$ $\mu$K in a gray molasses (blue filled squares) gives 17\% contrast and SNR $\simeq$ 30. Finally, adding a state preparation sequence after the gray molasses (black filled circles) yields a fringe contrast of 38\% and SNR $\simeq$ 45. The central fringe at $\phi_{\rm las} \simeq 1.7$ rad is phase-shifted due to gravity. (b) Allan deviation of phase measurements from the $^{39}$K gravimeter at $T = 22$ ms. Each measurement is composed of two interferometer fringes, one along each excitation direction, with each fringe containing 40 points (see inset). At a cycle time of $T_{\rm cyc} \simeq 2$ s, each measurement required $\tau = 160$ s---corresponding to the first point on the Allan deviation. Black filled circles correspond to the inertially-sensitive half-difference phase $\Delta\Phi$, while the half-sum phase $\Sigma\Phi$ is shown as red open squares, both of which are scaled by $k_{\rm eff} g T^2$ to obtain relative units of $g$. The bias drift of the mechanical accelerometer is corrected on the half-difference phase using our rubidium reference signal. Inset: typical interference fringe at $T = 22$ ms using upward excitation.}
\label{fig:KGravimeter}
\end{figure*}
We characterize the quality of our state preparation sequence in terms of the purity and the temperature of the atoms in their target state. Figure \ref{fig:KSampleProperties}(a) shows a measurement of the relative population in the $\ket{F = 1}$ manifold obtained using velocity-insensitive Raman spectroscopy. Here, $N_2/N_{\rm total}$ is the fraction atoms in $\ket{F = 2}$ compared to the total atom number, and $\delta$ is the frequency difference between the Raman beams. After an initial depump pulse, the atoms are roughly equally distributed amongst the Zeeman states of $\ket{F = 1}$. The population in $\ket{F = 1, m_F = 0}$ increases from 39\% to 63\% after preparation with a single transfer sequence, and is further increased to 92\% by adding a push pulse, as shown in Fig.~ \ref{fig:KSampleProperties}(b). By cycling the transfer sequence the population in the target state can be increased, but saturates after 2-3 cycles around 75\%, as indicated by the black and green circles. This is consistent with our expectations based on a simple optical pumping model. By adding the push pulse we find further diminishing returns per cycle, and a saturation around 95\% after 2 cycles (red filled squares). We attribute this to additional pumping into $\ket{F = 1}$ during the push pulse. We also find that the temperature of the sample increases roughly linearly with each additional transfer sequence ($\sim 1$ $\mu$K per cycle) as a result of scattering-induced heating during the optical pumping pulses [see Fig.~ \ref{fig:KSampleProperties}(c)]. From these data there is clearly a trade off to be made between the purity of the sample and its temperature. Typically, we use two transfer sequences in our experiments in order to optimize the contrast of our atom interferometer signal.
\section{Applications in metrology}
\label{sec:Metrology}
We now discuss two ongoing precision measurements with cold potassium in the fields of inertial sensing and atomic spectroscopy. The first is an atom-interferometric gravimeter, which is part of a dual-species apparatus \cite{Barrett2015, Barrett2016, Lefevre2017} that aims to test Einstein's equivalence principle by comparing the gravitational acceleration of $^{87}$Rb and $^{39}$K. The second is a hyperfine spectroscopy measurement of the ground state splitting between the $4 S_{1/2}$ $\ket{F = 1}$ and $\ket{F = 2}$ ground states of $^{39}$K.
\subsection{Gravimetry}
We realize our potassium gravimeter using a Mach-Zehnder-like $\pi/2 - \pi - \pi/2$ sequence of Raman pulses \cite{Kasevich1991}, each separated by an interrogation time $T$ during which the atoms are in free fall. We demonstrate improvements made to the interferometer by measuring interference fringes with and without the gray-molasses cooling and state preparation steps, as shown in Fig.~ \ref{fig:KGravimeter}(a). These data were taken at a sufficiently small $T$ such that vibrations of the reference frame (defined by the mirror that retro-reflects the Raman beams) caused negligible phase noise. We observe a factor of $\sim 4$ increase in the contrast of the inertially-sensitive fringes by cooling the atoms in a gray molasses and preparing them in $\ket{F = 1, m_F = 0}$. The signal-to-noise ratio (SNR) also improves by more than a factor of 2, but is currently limited by technical noise from our detection system \cite{Antoni-Micollier2017}. Here, the gains in contrast and SNR are linked to the reduction of the velocity selection along the Raman beam's propagation direction. When increasing the interaction time, the transverse expansion of the atomic cloud contributes to the loss of contrast as the atoms experience a spatial variation of the Raman laser intensity \cite{Dickerson2013}. In this context, our preparation method is even more important in order to maintain a relatively high contrast and SNR at the timescales required for high-performance inertial sensors [see insert in Fig.~ \ref{fig:KGravimeter}(b)].
At larger interrogation times ($T \gtrsim 1$ ms), we apply a frequency chirp $\alpha$ to the Raman lasers in order to cancel the increasing Doppler shift of the two-photon resonance due to the atom's acceleration under gravity. The phase shift of the interferometer's central fringe is given by
\begin{equation}
\label{Phiupdown}
\Phi_{\uparrow\downarrow} = \pm (\bm{k}_{\rm eff} \cdot \bm{g} - \alpha) T^2 \pm \phi_{\rm vib} + \phi_{\rm sys}^{\uparrow\downarrow},
\end{equation}
where $\bm{k}_{\rm eff} \simeq (4\pi/\lambda) \hat{\bm{z}}$ is the effective Raman wavevector along the vertical direction, $\bm{g} = g\,\hat{\bm{z}}$ is the acceleration due to gravity, $\phi_{\rm vib}$ is a phase shift due to vibrations of the reference frame, and $\phi_{\rm sys}$ represents the total systematic shift. Since the counter-propagating Raman beams are generated by retro-reflecting a pair of co-propagating lasers, we can choose the direction of excitation (upward $\uparrow$, or downward $\downarrow$) via the Raman frequency during the first $\pi/2$ pulse. We reject systematic effects that are independent of this direction (\textit{e.g.}~ AC Stark shifts, quadratic Zeeman shifts) by alternating between upward and downward excitations for each repetition of the experiment \cite{Peters2001, Louchet-Chauvet2011a}. In the case where $\phi_{\rm sys}^{\uparrow\downarrow} \simeq \phi_{\rm sys}$ is independent of the excitation direction, the half-difference $\Delta\Phi \equiv \frac{1}{2}(\Phi_{\uparrow} - \Phi_{\downarrow}) = (\bm{k}_{\rm eff} \cdot \bm{g} - \alpha) T^2 + \phi_{\rm vib}$ then yields a local measurement of $g$. Typically, one achieves this by suppressing phase noise due to vibrations, and precisely locating the central fringe for which $\alpha = \bm{k}_{\rm eff} \cdot \bm{g}$. Similarly, the half-sum $\Sigma\Phi \equiv \frac{1}{2}(\Phi_{\uparrow} + \Phi_{\downarrow}) = \phi_{\rm sys}$ gives the total (direction-independent) systematic shift. Figure \ref{fig:KGravimeter}(b) displays the Allan deviation of gravity measurements obtained using a total interrogation time of $2T = 44$ ms \footnote{Our maximum interrogation time is limited by the geometry of our science chamber.}, where the inertially-sensitive half-difference phase ($\Delta\Phi$) is shown in black. These data show a sensitivity of approximately $2 \times 10^{-6}\, g/\sqrt{\rm Hz}$, and a statistical uncertainty of 60 $\mu$Gal after 80 minutes of integration.
A unique feature of our gravimeter is that we do not stabilize the reference mirror against vibrations, which causes a root-mean-squared phase noise of several radians at $T = 22$ ms. Instead, we measure these parasitic vibrations using a sensitive mechanical accelerometer (Nanometrics Titan) and later correct for the corresponding phase shifts on the interferometer \cite{LeGouet2008, Barrett2015, Fang2016}. An example of one such vibration-corrected interference fringe is shown in the inset of Fig.~ \ref{fig:KGravimeter}(b). We emphasize that without this correction, inertial phase measurements for $T$ greater than a few ms would not be possible with our apparatus. However, the mechanical accelerometer acts as a ``floating'' phase reference, since its output bias inevitably drifts due to temperature variations. This drift is then imprinted on our gravity measurements---frustrating long-term integration. This problem can be avoided by strongly filtering the DC component of the accelerometer signal, since only the AC component is used for correcting the vibration-induced phase shifts. For these data, we used the output of our simultaneous $^{87}$Rb gravimeter (sensitivity $\sim 30$ $\mu$Gal after 80 min.) as a phase reference for potassium in order to suppress the DC bias drift of the accelerometer \cite{Lefevre2017}. However, this method is also imperfect since any noise in the rubidium phase is imprinted on the potassium signal. We point out that this approach is equivalent to measuring the differential acceleration between the two species, and constitutes a test of the equivalence principle \cite{Schlippert2014a}.
An analysis of the half-sum phase gives an indication of the sensitivity level we could reach with a perfect phase reference, since this quantity is insensitive to inertial phases like $\phi_{\rm vib}$. We estimate this level to be $\sim 40$ $\mu$Gal after 80 minutes from the data shown in Fig.~ \ref{fig:KGravimeter}(b). Additionally, these data give a good indication of the stability of our experiment, since $\Sigma \Phi$ is a direct measure the relatively large one-photon light shifts and quadratic Zeeman shifts in potassium.
State-of-the-art gravimeters based on cold $^{87}$Rb sources can now achieve sub-$\mu$Gal uncertainty after only a few minutes of integration, with accuracies on the order of a few $\mu$Gal \cite{Gillot2014, Fang2016, Freier2016}. However, for potassium, the most precise gravity measurements demonstrated so far are on the order of 490 $\mu$Gal after one hour of averaging \cite{Schlippert2014a}. With our low-temperature, high-purity $^{39}$K sample, we have improved upon these results by about an order of magnitude. In terms of accuracy, we estimate an overall systematic uncertainty of $\sim 50$ $\mu$Gal, where the dominant contributions arise from the two-photon light shift \cite{Gauguet2009}, the curvature of the Raman beam wavefront \cite{Louchet-Chauvet2011a, Trimeche2017}, and the quadratic Zeeman effect \cite{Barrett2016}. Secondary contributions due to the one-photon light shift, gravity gradient and the Coriolis acceleration sum to less than 10 $\mu$Gal for $T = 22$ ms. Further details concerning systematic effects on our potassium gravimeter will be published elsewhere.
\subsection{Hyperfine spectroscopy}
\begin{figure}[!tb]
\centering
\includegraphics[width=0.48\textwidth]{Fig4-KClock.pdf}
\caption{Optical Ramsey fringes using $T_R = 40$ ms and $\tau_R = 3$ $\mu$s. The corresponding sinusoidal fit (solid line) gives a relative statistical error of $\Delta f/f = 4.1 \times 10^{-11}$. Measurements were carried out near the peak of the Ramsey pattern where the SNR was largest. This peak is shifted by $-126.8(2)$ kHz from the central fringe corresponding to the hyperfine frequency splitting at 461.719720 MHz \cite{Arimondo1977} due to the AC Stark effect from the Raman beams. The corresponding shift of the central fringe is $11.2(1)$ Hz \cite{Zanon-Willette2016b}. Similarly, we estimate a shift due to the quadratic Zeeman effect of $-114.3(4)$ Hz.}
\label{fig:KClock}
\end{figure}
We now describe an ongoing precision measurement of the clock transition in $^{39}$K using a Ramsey-type interferometer. The Ramsey technique \cite{Ramsey1949} involves a $\pi/2 - \pi/2$ pulse sequence, with pulse durations $\tau_R$ and free-evolution time $T_R$, during which the atoms are free-falling in a superposition of two internal states. As a function of the probe detuning $\delta$, the fraction of atoms in each state at the output of the spectrometer oscillates in the form of a Ramsey fringe pattern, where the fringe width scales inversely with $T_R$. The central fringe location---which remains fixed for all values of $T_R$ in the absence of systematic effects---gives a precise measure of the transition frequency between the two internal states. This type of interferometer is extensively used by the atomic clock community, either with microwave transitions in the case of alkali metals \cite{Kasevich1989, Gibble1993, Santarelli1999}, or with optical single-photon transitions in the case of alkaline and rare earth metals \cite{Nicholson2012, Hinkley2013}.
Instead of using a microwave field to probe $^{39}$K, our spectroscopy measurements were carried out using two pairs of co-propagating Raman beams---one pair travelling along the upward direction with $\sigma^+$ polarization and one downward with $\sigma^-$ polarization. In this configuration, both pairs of co-propagating Raman beams are velocity-insensitive and resonant with the $\ket{F=1,m_F=0} \to \ket{F=2,m_F=0}$ clock transition, while the counter-propagating pairs are velocity-sensitive and drive only $\Delta m_F = \pm 2$ transitions. These transitions are off-resonant due to both the Doppler and Zeeman effects. Figure \ref{fig:KClock} displays optical Ramsey fringes measured as a function of the frequency difference between Raman lasers. Using $T_R = 40$ ms, we obtain a statistical uncertainty in the fractional frequency of $\Delta f/f = 4.1 \times 10^{-11}$ with a single fringe measurement---more than a 30-fold improvement over the previous best measurements at $1.3 \times 10^{-9}$ \cite{Arimondo1977}. Although free from cavity-pulling effects \cite{Sortais2000}, our measurement is sensitive to light shifts from the Raman beams \cite{Zanon-Willette2016b}. In addition, we apply a relatively large bias field of $\sim 100$ mG in order to increase the efficiency of the state preparation scheme, but this too shifts the clock transition frequency due to the quadratic Zeeman effect. A preliminary study of the dominant systematic effects indicates our accuracy is limited at the level of a few $10^{-9}$ due to Zeeman shifts, AC Stark shifts, and the calibration of our frequency source \cite{Laurent1998, Stern2009}. We estimate frequency shifts due to cold collisions \cite{Kokkelmans1997, Sortais2000} to be less than 100 mHz at our densities of $\sim 5 \times 10^9$ atoms/cm$^3$.
\section{Conclusion}
\label{sec:Conclusion}
We have demonstrated that $^{39}$K can be used for high-precision measurements thanks to a new technique to prepare a cold sample of atoms in a magnetically-insensitive state with 95\% purity. We used this sample in an atom-interferometric gravimeter, and achieved a state-of-the-art performance. Potassium atoms prepared with this technique could be ideal for new determinations of the fine structure constant \cite{Bouchendira2011}, searches for new fundamental forces \cite{Jaffe2016}, or for tests of the equivalence principle \cite{Wolf2011a, Schlippert2014a, Lefevre2017}. We also demonstrate that, in a Ramsey-type spectrometer, the ground state hyperfine splitting of $^{39}$K can be measured with a sensitivity improved by more than a factor of 30. We anticipate that new rounds of measurements at increased interrogation time could yield a precision below $10^{-12}$. At this level, an improved mitigation strategy for the present systematic limitations will be required, including a characterization of collisional frequency shifts---which have not yet been measured in $^{39}$K. Our scheme can be easily modified to prepare atoms in any other ground state magnetic sub-levels, and is particularly well suited to atoms with closely-spaced excited states, such as the isotopes of lithium and potassium.
\acknowledgments
This work is supported by the French national agencies CNES (Centre National d'Etudes Spatiales), l'Agence Nationale pour la Recherche, the D\'{e}l\'egation G\'{e}n\'{e}rale de l'Armement, the European Space Agency, IFRAF (Institut Francilien de Recherche sur les Atomes Froids), and action sp\'{e}cifique GRAM (Gravitation, Relativit\'{e}, Astronomie et M\'{e}trologie). L. Antoni-Micollier and B. Barrett thank CNES and IOGS for financial support. P. Bouyer thanks Conseil R\'{e}gional d'Aquitaine for the Excellence Chair. Finally, we wish to thank G. Santarelli for helpful discussions and M. Prevedelli for his assistance repairing a critical part of the experiment.
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
\label{sec:introduction}
Since the discovery of the first exoplanet by \citet{Mayor1995} the number of new planets detected has been growing exponentially. First from ground-based radial velocity \citep{Vogt2000,Valenti2005,Jenkins2009} and transit surveys \citep{Bakos2004,Pollacco2006} and in recent years from the CoRoT and \emph{Kepler} satellites \citep{Barge2008,Borucki2010}. Today, several thousand exoplanets are known and our understanding of the population has grown immensely. In particular, we now know that on average a late-type main-sequence star is orbited by at least one planet \citep{Batalha2014}.
Despite these advances, observational biases mean we still know only a few planets transiting both bright and early-type stars. The Multi-site All-Sky CAmeRA \citep[MASCARA,][]{Talens2017} transit survey aims to fill this gap in parameter space by monitoring all stars with magnitudes $4 < m_V < 8.4$. MASCARA consists of a northern and a southern station, each equipped with five cameras to cover the entire local sky down to airmass 2 and partially down to airmass 3. In this way MASCARA is expected to find several new transiting hot Jupiters around bright stars, some of which will be of early spectral types \citep{Snellen2012}.
MASCARA was designed to find Jupiter-like planets transiting bright stars because these allow detailed atmospheric characterization. Observations of the planet transmission spectra during transit and thermal emission at secondary eclipse, coupled with detailed atmospheric models, can be used to constrain atmospheric properties such as temperature structure, composition and the presence of clouds and hazes \citep{Madhusudhan2009,Sing2011,Deming2013,Kreidberg2014,Sing2016}. Such observations, however, are expensive if not impossible on fainter stars. In particular, the use of high-resolution transmission spectroscopy has been limited to a few bright targets such as HD\,209458 and HD\,189733 \citep{Snellen2010,Brogi2016} which were first discovered trough radial velocity observations.
In addition to finding planets around bright stars, the all-sky, magnitude-limited sample targeted by MASCARA is biased towards early type stars, which are particularly challenging to follow-up at fainter magnitudes. Studying the hot Jupiter population around hot stars is interesting for several reasons. First, these hot Jupiters will have the highest equilibrium temperatures, which is expected to have a significant impact on their atmospheric structure and composition - such as the occurrence of inversion layers due to the presence of specific molecular compounds \citep[e.g. TiO/VO][]{Hubeny2003,Fortney2008} high up in the atmospheres (e.g. \cite{Burrows2007,ODonovan2010} but see also \citet{Schwarz2015}). Second, hot Jupiters around early type stars receive more UV radiation than similar planets around late type stars. This UV radiation may drive unique chemical processes in the atmospheres of these planets \citep{Casewell2015}. Finally, comparing their orbital properties with those of hot Jupiters orbiting solar-type stars may shed light on their formation. For example, the large incidence of misaligned orbits for hot Jupiters orbiting early-type stars \citep{Winn2010,Schlaufman2010,Albrecht2012} may be linked to orbital migration processes \citep[e.g. ][]{Fabrycky2007,Nagasawa2008}.
In this paper we present the discovery of MASCARA-1\,b, the first exoplanet discovered by MASCARA. MASCARA-1\,b orbits the $m_V=8.3$ A8 star HD\,201585 with a period of $2.15~\rm{days}$. We present our observations in Section \ref{sec:observations}. Stellar and system parameters are described in Sects. \ref{sec:star} and \ref{sec:system}, and we conclude with a discussion in Sect. \ref{sec:discussion}.
\section{Observations}
\label{sec:observations}
MASCARA's northern station is located on La Palma and started science operations early 2015, producing photometry for all stars with $4 < m_V < 8.4$ at a cadence of 6.4 seconds down to airmass $\sim$3 \citep{Talens2017}. The raw MASCARA photometry is processed using a heavily modified version of the coarse decorrelation algorithm described in \citet{CollierCameron2006} in order to remove instrumental systematics and earth-atmospheric effects. Subsequently, the data are binned to a cadence of 320 seconds and further detrended by means of an empirical fit to remove both residual variations linked to the target positions on the CCDs and long-term trends in the data. Finally, a Box Least-Squares \citep[BLS,][]{Kovacs2002} transit search algorithm is used to find potential transit signals. The details of the data reduction and transit search algorithms will be discussed in an upcoming paper (Talens et al. in prep).
\begin{table}
\centering
\caption{Observations used in the discovery of MASCARA-1\,b.}
\begin{tabular}{l c c c}
Instrument & Date & $N_{\rm{obs}}$ & $t_{\rm{exp}}$ [s] \\
\hline
\hline
MASCARA & Feb. 2015 - Sept. 2016 & 32705 & 320 \\
HERMES & 15 June 2016 & 1 & 1400 \\
HERMES & 17 July 2016 & 1 & 800\\
HERMES & 18 July 2016 & 1 & 800\\
NITES & 23 July 2016 & 1092 & 10 \\
HERMES & 5 Sept. 2016 & 1 & 1100\\
HERMES & 7 Sept. 2016 & 1 & 1200\\
HERMES & 7 Sept. 2016 & 1 & 1100\\
HERMES & 9 Sept. 2016 & 1 & 1100\\
HERMES & 10 Sept. 2016 & 1 & 1000\\
HERMES & 10 Sept. 2016 & 1 & 1200 \\
HERMES & 11 Sept. 2016 & 1 & 1000\\
HERMES & 12 Sept. 2016 & 1 & 1200 \\
HERMES & 13 Sept. 2016 & 1 & 900 \\
HERMES & 14 Sept. 2016 & 1 & 1100\\
NITES & 17 Sept. 2016 & 887 & 15 \\
SONG & 30 Sept. 2016 & 28 & 600
\end{tabular}
\label{tab:datasets}
\end{table}
After running the BLS algorithm on data taken between February and December 2015, MASCARA-1, (HD\,201585, HIP\,104513) was among the most promising stars to be listed as a potential exoplanet host. We obtained high-dispersion spectra at R=85,000 from the HERMES spectograph \citep{Raskin2011} at the 1.2m Mercator telescope on La Palma taken between June and September 2016 (see Table \ref{tab:datasets}). A total of 13 spectra were taken and reduced using the HERMES pipeline\footnote{\label{note:hermes}\url{http://hermes-as.oma.be/manuals/cookbook6.0.pdf}}, allowing us to characterize the host star and constrain the companion mass from radial velocity (RV) measurements.
In addition to the MASCARA photometry, two transits were observed with the 0.4m NITES telescope \citep{McCormac2014} on La Palma. 1,092 images with 10s exposure times were obtained with a Johnson-Bessel R-band filter on July 23, 2016 and 887 images with 15s exposure times were obtained in the V-band on September 17, 2016. The telescope was defocused slightly on both nights to avoid saturation. The data were reduced in Python with CCDPROC \citep{Craig2015} using a master bias, dark and flat. Master calibrations were made using a minimum of 21 of each frame. Non-variable comparison stars were selected by hand and aperture photometry extracted using SEP \citep{Barbary2016,Bertin1996}. The shift between each defocused image was measured using the DONUTS \citep{McCormac2013} algorithm and the photometry apertures were recentered between frames. The final aperture size is chosen to minimise the RMS of the data points out of transit. The $0.66\arcsec$ pixel scale of NITES, versus $1\arcmin$ for MASCARA, allowed us to check against a faint eclipsing binary system within the MASCARA photometric aperture that could have explained the apparent transit signal.
Additional confirmation about the nature of the system, as well as a measurement of its spin-orbit angle, can come from a detection of the Rossiter-McLaughlin effect using in-transit spectroscopy \citep[e.g.][]{Zhou2016}. We therefore obtained 28 high-dispersion echelle spectra at R$\sim$90,000 using the automated 1m Hertzsprung SONG telescope \citep{Andersen2014} at Observatorio del Teide on September 30, 2016 (see Table \ref{tab:datasets}). The observations were taken with 10 minute exposure times and a slit width of 1.2 arcsec and spanned ${\sim}5$~hr in total, covering the transit and post-egress. Ingress occurred during evening twilight. Spectra were extracted from the observations following the procedure outlined in \citet{Grundahl2017}, subsequently bad pixels were removed and the spectra were normalized.
\begin{figure}
\centering
\includegraphics[width=8.5cm]{fig1a_MASCARA_Q1Q7.pdf}
\includegraphics[width=8.5cm]{fig1b_NITES_20160723.pdf}
\includegraphics[width=8.5cm]{fig1c_NITES_20160917.pdf}
\caption{Photometry and best-fit models used in the characterization of MASCARA-1\,b. (a) MASCARA photometry obtained between February 2015 and September 2016. Shown are the 320 s cadence data (grey points), the data binned by 0.01 in phase (red points) and the best-fit transit model (black line). (b) and (c) NITES transit observations taken on 23 July 2016 and 17 September 2016, respectively. Shown are the data (grey points), the data binned by 0.01 in phase (red points) and the best-fit transit model (black line).}
\label{fig:photometry}
\end{figure}
\section{Stellar parameters}
\label{sec:star}
At $m_V=8.3$, most basic properties of MASCARA-1 are available in the literature. The parallax was already measured by Hipparcos \citep{VanLeeuwen2007} to be $5.0\pm0.7~\rm{mas}$, implying a distance of $201.6\pm28.9~\rm{pc}$. Recently, the GAIA DR1 \citep{GAIA2016a} lists a parallax of $5.3\pm0.3~\rm{mas}$ ($188.7\pm10.7~\rm{pc}$), within 1$\sigma$ of the Hipparcos value. For our analysis we adopt the GAIA parallax.
\citet{McDonald2012} combined the Hipparcos parallaxes with photometry from a number of surveys, including SDSS and 2MASS, to determine the effective temperature and luminosity of over 100,000 stars, including MASCARA-1 for which they derive an effective temperature of $7554~\rm{K}$ and a bolometric luminosity of $14.98~\rm{L}_\odot$. We verified the effective temperature by fitting the wings of the hydrogen lines in the combined HERMES spectra with PHOENIX spectral models \citep{Husser2013} and find $T_{\rm{eff}} = 7600\pm150~\rm{K}$, consistent with the work done by \citet{McDonald2012}. For our further analysis, we used the values as derived by \citet{McDonald2012}, assuming uncertainties of $150~\rm{K}$ and $5~\rm{L}_\odot$ on $T_{\rm{eff}}$ and $L_*$ respectively\footnote{For $T_{\rm{eff}}$ we assume an uncertainty comparable to our own analysis. For $L_*$ the error on the distance alone implies an uncertainty of $4.3~\rm{L}_\odot$, which we round up to account for uncertainties in the total flux.} - which are not stated by the authors. We updated the luminosity for the new GAIA distance to be $13.1\pm3.0~\rm{L}_\odot$. From the bolometric luminosity and effective temperature, we derive a stellar radius of $R_* = 2.1\pm0.2~\rm{R}_\odot$. We use {\sc bagemass} \citep{Maxted2015}, a stellar evolution MCMC code, to determine the stellar mass and age from the effective temperature, metallicity and stellar density\footnote{The stellar density is obtained from the fit to the transit light curve as described in Sect. \ref{sec:photfit} using the relations of \citet{Seager2003}.}, finding $1.72\pm0.07~\rm{M}_\odot$ and $1.0\pm0.2~\rm{Gyr}$, respectively. As an extra check we also compute the stellar radius from the {\sc bagemass} output and find a value of $1.97 \pm 0.07~\rm{R}_\odot$, consistent with the radius derived from the GAIA parallax.
From the HERMES radial velocity pipeline\footnote{See footnote \ref{note:hermes}} we also obtain the stellar equatorial rotation speed $v\sin i_\star = 106.7\pm0.4~\rm{km~s}^{-1}$.
\begin{table*}
\centering
\caption{Transit model parameters and their best-fit values obtained from fits to the MASCARA data, NITES data, and the combined photometric data set.}
\begin{tabular}{l c c c c c c c}
Parameter & Symbol & Units & MASCARA & NITES & MASCARA + NITES\\
\hline
\hline
Reduced chi-square & $\chi^2_\nu$ & - & $1.98$ & $0.51$ & $1.98$ \\
Norm. MASCARA & $F_0$ & - & $1.00021\pm0.00004$ & - & $1.00032\pm0.00004$ \\
Norm. NITES & $F_1$ & - & - & $1.0000\pm 0.0002$ & $0.99962\pm0.00009$ \\
Norm. NITES & $F_2$ & - & - & $1.0002\pm 0.0002$ & $0.99975\pm0.00009$ \\
Epoch & $T_p$ & BJD &$2457097.277\pm0.002$ & $2457097.2751\pm0.0009$ & $2457097.278\pm0.002$ \\
Period & $P$ & d & $2.14879\pm0.00001$ & 2.14879 (fixed) & $2.148780\pm0.000008$ \\
Duration & $T_{14}$ & h & $4.06^{+0.08}_{-0.07}$ & $4.1^{+0.08}_{-0.06}$ & $4.05\pm0.03$ \\
Planet-to-star ratio\tablefootmark{a} & $p=R_p/R_*$ & - & $0.063^{+0.002}_{-0.001}$ & $0.078^{+0.002}_{-0.001}$ & $0.0735\pm0.0007$ \\
Impact parameter & $b$ & - & $0.2\pm0.2$ & $0.3\pm0.2$ & $0.2\pm0.1$ \\
Eccentricity & $e$ & - & $0$ (fixed) & $0$ (fixed) & $0$ (fixed) & \\
\end{tabular}
\tablefoot{
\tablefoottext{a}{The best fit value of $p$ differs significantly between fits and as we are unable to reconcile the different data we adopt $p=0.07\pm0.01$ as our final value.}
}
\label{tab:photpars}
\end{table*}
\section{System parameters}
\label{sec:system}
System parameters are derived from modelling the available photometric and spectroscopic observations. The transit shape is fitted using the available MASCARA and NITES photometric data. The mass of the planet is constrained using the HERMES radial velocity measurements, and the projected spin-orbit angle is obtained from the SONG transit data.
\subsection{Photometric transit fit}
\label{sec:photfit}
A \citet{Mandel2002} model is fitted to the photometric data taken by MASCARA using a Markov-chain Monte Carlo (MCMC) approach with the {\sc Python} packages {\sc batman} and {\sc emcee} \citep{Kreidberg2015,FM2013}. We fit a circular transit model, optimizing for the normalization $F_0$, the transit epoch $T_p$, the orbital period $P$, the transit duration $T_{14}$, the planet-to-star radius ratio $p$ and the impact parameter $b$ using uniform priors on all parameters. We employ a quadratic limb darkening law with fixed coefficients from \citet{Claret2000}, using values of $u_1=0.2609$, $u_2=0.3405$, appropriate for the host star. Subsequently, the NITES photometry is also fitted using the same model, except with the orbital period fixed to the best fit value obtained from the MASCARA photometry and separate normalization factors $F_1$ and $F_2$ for each transit. In the modelling presented here we do not include the possible effects of non-spherical geometry, which can influence $T_{14}$ and $b$, or gravity darkening, which can modify $p$. Both of these effect might play a role for a rapid rotator such as MASCARA-1. The best-fit parameter values and uncertainties, obtained from the median and the 16th and 84th percentiles of the output MCMC chains, are listed in Table \ref{tab:photpars}.
We find that all orbital parameters are consistent between the MASCARA and NITES data with exception of the planet-to-star ratio $p$, which takes the value of $0.063^{0.002}_{-0.001}$ for the MASCARA data and $0.078^{0.002}_{-0.001}$ for the NITES data. We consider the possibility that this difference in the value of $p$ is a result of blended light in the MASCARA photometry, caused by the $1\arcmin$ pixel scale of MASCARA, which would decrease the transit depth and thus the value of $p$. However, a simple calculation shows that we would need ${\sim}35\%$ blended light in the MASCARA aperture which is more than can be explained by the background sources present, which we conservatively estimate to contribute a maximum of ${\sim}20\%$ of the total light. Another possibility comes from the algorithm used to process the MASCARA data, as we are still in the process of investigating possible systematic changes introduced by our methods.
To further investigate this discrepancy we perform a joint fit to the MASCARA and NITES photometry. The reduced chi-square value for the NITES fit indicates that we are over-estimating the uncertainties on the photometry, this is plausible since we assume a conservative model for the scintillation noise, the dominant source of noise in the observations, when calculating the uncertainties. To ensure we place similar weight on both the MASCARA and NITES observations we reduce the NITES uncertainties by a factor 2 for the joint fit. The results of the joint fit are shown in Fig. \ref{fig:photometry} and the parameters are listed in Table \ref{tab:photpars}. From the joint fit we obtain $p=0.0735\pm0.0007$, significantly different from both the NITES and MASCARA values. It should also be noted that the lack of baseline in the NITES observations results in a lower normalization in the joint fit, helping to reduce any dicrepancy in the depths. We are unable at this time to explain this discrepancy and adopt a value of $p=0.07\pm0.01$, midway in between and consistent with both the MASCARA and NITES values, for the derivation of the planet radius listed in Table \ref{tab:syspars}.
\subsection{Radial velocities and planet mass}
Stellar radial velocities and their uncertainties are determined from the HERMES spectra using the standard pipeline\footnote{See footnote \ref{note:hermes}}, in which a mask of individual lines is cross-correlated with the data to obtain an average line profile. For the observations of MASCARA-1 the {\sc HermesF0Mask} was used, being the closest to the spectral type of the star of the available masks. The average line profile is subsequently fitted with a rotationally broadened line profile, the dominant source of broadening, to obtain the radial velocities and their uncertainties. The obtained uncertainties are relatively large (${\sim}300\rm{~m~s}^{-1}$) due the high equatorial rotation speed of the star. The epochs of the radial-velocity data points are phase-folded using the best-fit orbital solution derived from the photometry. A circular orbital solution was fitted to the data keeping the period and epoch fixed to the values derived above. The RV data and best-fit solution are shown in Fig. \ref{fig:RV}, and the best-fit parameters are listed in Table \ref{tab:specpars}. The amplitude of the stellar radial velocity variation is not well constrained to $K=400\pm100~\rm{m~s}^{-1}$, corresponding to a planet mass of $M_p = 3.7\pm0.9~\rm{M}_{\rm{jup}}$.
\begin{figure}
\centering
\includegraphics[width=8.5cm]{fig2_HERMES_rv.pdf}
\caption{HERMES RV measurements of MASCARA-1 obtained from July to September 2016. Shown are the data (red points) and the best-fit model (black line). The high projected rotation speed of the star, $v\sin i_\star = 106.7~\rm{km~s}^{-1}$, results in relatively large uncertainties on the RV measurements of order $300\rm{~m~s}^{-1}$.}
\label{fig:RV}
\end{figure}
\subsection{Rossiter-McLaughlin effect and projected spin-orbit angle}
To consolidate the confirmation of the planetary nature of the companion, and to determine the projection of the spin-orbit angle (obliquity), we observed the systems during a transit with the SONG telescope (See section \ref{sec:observations}). Simple Cross Correlation Functions (CCFs) are created from the obtained spectra using a stellar template appropriate for MASCARA-1. \footnote{The template was created using the interactive data language (IDL) interface SYNPLOT (I. Hubeny, private communication) to the spectrum synthesis program SYNSPEC \citep{Hubeny1995}, utilizing a Kurucz model atmosphere (see \url{http://kurucz.harvard.edu/grids.html}).} The CCFs are shown in Fig. 3 panel A (top left).
To analyze the distortions of the stellar absorption lines and measure the projected obliquity ($\lambda$) we use the code presented in \citet{Albrecht2007}, including updates presented in \citet{Albrecht2013}. In short, we create a pixelated stellar disk where each pixel is assigned a radial velocity based on stellar rotation, and micro- and macro-turbulence \citep[see e.g.][]{Gray1984}. Here the Point Spread Function (PSF) of the spectrograph is included in the width of the Gaussian function representing the micro-turbulence. We do not include convective blueshift as the CCFs do not have a high enough SNR to constrain this parameter and in turn the influence on $\lambda$ is small. The light contribution of each pixel in our model is governed by a quadratic limb darkening model. For each observation, a model stellar absorption line is created. During transit, the integration is carried out only over the surface not blocked by the companion. These model CCFs are then compared to the observed CCFs, taking into account shifts due to the orbital velocity.
\begin{table*}
\centering
\caption{Model parameters used in fitting the RV and RM measurements and their best-fit values.}
\begin{tabular}{l c c c c c c c}
Parameter & Symbol & Units & HERMES & SONG \\
\hline
\hline
Epoch\tablefootmark{a} & $T_p$ & BJD & $2457097.278$ (fixed) & $2457097.283\pm0.0022$\\
Period & $P$ & d & $2.148780$ (fixed) & $2.148780$ (fixed) \\
Duration\tablefootmark{a} & $T_{14}$ & h & - & $4.15\pm0.0030$\\
Planet-to-star ratio\tablefootmark{a} & $p=R_p/R_*$ & - & - & $0.0795\pm0.00065$ \\
Impact parameter\tablefootmark{a} & $b$ & - & - & $0.122\pm 0.012$ \\
Eccentricity & $e$ & - & $0$ (fixed) & $0$ (fixed)\\
RV amplitude\tablefootmark{a} & $K$ & m s$^{-1}$ & $400\pm100$ & $190\pm90$\\
Systemic velocity\tablefootmark{b} & $\gamma$ & km s$^{-1}$ & $11.20\pm0.08$ & $8.52\pm0.02$\\
Projected obliquity & $\lambda$ & \degr & - & $69.5\pm4$\\
Micro turbulence & $\nu$ & km s$^{-1}$ & - & $0.4 \pm 0.3$ \\
Macro turbulence & $\zeta$ & km s$^{-1}$ & - & $7.3 \pm 0.2$ \\
Projected rotation speed & $v \sin i_\star$ & km s$^{-1}$ & - & $109\pm3$\\
\end{tabular}
\tablefoot{
\tablefoottext{a}{For the SONG analysis, the best-fit parameters from the photometry and radial velocities were used to set Gaussian priors on $T_p$, $T_{14}$, $p$, $b$ and $K$.}
\tablefoottext{b}{We find a significant offset between the values for $\gamma$ derived from the HERMES and SONG data. An analysis of our methods revealed the offset likely originates from a difference in the spectral templates used in reducing the HERMES and SONG data, but does not influence our results.}
}
\label{tab:specpars}
\end{table*}
\begin{figure*}
\centering
\includegraphics[width=8.5cm]{hd201585_stellar_line_transit.pdf}
\includegraphics[width=8.5cm]{hd201585_planet_shadow.pdf}
\includegraphics[width=8.5cm]{hd201585_residuals.pdf}
\includegraphics[width=8.5cm]{hd201585_shadow_model.pdf}
\caption{Observations of MASCARA-1 during and after the transit of MASCARA-1\,b. In each panel, the solid horizontal line indicates the end of transit, and the dashed lines the end of ingress and beginning of egress. The solid black vertical lines show the best-fit value for $v \sin i_\star$. The dashed-dotted red lines indicate the mid-transit time and the rest-frame velocity of the star.
{\it Panel A:} The observed CCFs in gray scale. {\it Panel B:} The same as in Panel A, but with the best-fit model for the undisturbed stellar CCF subtracted, isolating the distortions due to the planetary transit over the rotating stellar photosphere. {\it Panel C:} Residuals after our best fitting model -- including the planetary transit -- is subtracted. {\it Panel D:} The corresponding model to the data shown in Panel B.}
\label{fig:RM}
\end{figure*}
We run an MCMC to obtain confidence intervals for $\lambda$ and $v \sin i_\star$, the parameters of interest. For this we allowed the following parameters to vary: $\lambda$, $v \sin i_\star$, macro-turbulence ($\zeta$), micro-turbulence ($\nu$), the transit epoch $T_{p}$, planet-to-star radius ratio $p$, scaled semi-major axis $a/R_*$, the cosine of the orbital inclination $\cos i$, the stellar systemic velocity $\gamma_{\rm SONG}$, the stellar RV amplitude $K$, and limb darkening parameters $u_1+u_2$. We use prior information from the photometric and radial-velocity observations for the parameters $T_{p}$, $a/R_*$, $p$ and $K$. We further used the prior information for the impact parameter $b$, and the transit duration $T_{14}$ also derived from photometry (Table~\ref{tab:photpars}), $0.6\pm0.1$ for $u_1+u_2$ \citep{Claret2013}, and uniform priors on all other parameters. Finally for each calculation of the likelihood we allow each of the observed CCFs to be offset and scaled in intensity \citep[see also][]{Albrecht2013}. This way we include the influence of the none perfect normalization of the spectra into the uncertainty interval of the final parameters. A mismatch in the continuum normalization of the observed spectrum and the template spectrum causes an offset in the ``baseline'' of the obtained CCF. In addition, the size of the CCFs depends on the SNR of the spectra, which did vary throughout the observations. In Tab.~\ref{tab:specpars} we report the results for the above parameters, including the $68\%$ confidence intervals from the analysis of the CCFs.
With an projected obliquity of $69.5\pm0.3^\circ$ we find that the rotation axis of MASCARA-1 is misaligned with respect to the orbit of the planet. This can also be seen in Fig.~\ref{fig:RM} panel B, where the planet shadow is isolated. Despite a low impact parameter the trail of the planet does not reach from $-v \sin i_\star$ to $+v \sin i_\star$, as it would be the case if the planet would travel near the stellar equator. For central transits and large misalignments the amplitude of the RM effect is reduced, as the planet travels from pole to pole, and not over the equator. If the impact parameter would have been larger, MASCARA-1\,b would have traveled largely over the receding stellar surface area. We also find that the SONG data prefer a small impact parameter ($b=0.122\pm0.012$), consistent with the photometry ($b=0.2\pm0.1$). The transit mid-point is found 7~min later than expected from the photometric ephemeris. We argue that the formal uncertainty in the projected rotation speed ($v \sin i_\star=109.0\pm0.1$~km\,s$^{-1}$) underestimates the true uncertainty in this parameter. MASCARA-1 is a fast rotator and we expect a departure from the perfect spherical shape assumed in our analysis. In addition, we also expect a significant gravity darkening. This would result in a dark band along the equator for large $i_\star$ and an apparent increase in the stellar limb darkening for low $i_\star$. We therefore argue that an uncertainty of 3\,km\,s$^{-1}$ is more appropriate for the projected rotation speed. We find that our data prefers a larger than expected limb darkening $u_1+u_2 =0.9\pm0.1$ which we attribute to the simplified physics in our model rather than to a true disagreement with the model limb darkening parameters. Indeed if our data would be of high SNR, we might be able to determine $i_\star$ and therefore the actual obliquity ($\Psi$) in this system from the RM data alone. However given the SNR in our SONG data we postpone such an analysis until such a dataset is available. We also expect the uncertainty in $\lambda$ to be about $4^\circ$, larger then the formal value quoted above, for the same reasons that we expect the $v \sin i_\star$ uncertainty to be underestimated.
\section{Discussion and Conclusions}
\label{sec:discussion}
\begin{figure}
\centering
\includegraphics[width=8.5cm]{fig5_mass_radius.pdf}
\caption{Mass-radius relationship for transiting exoplanets. MASCARA-1\,b is indicated by the star marker and the points are coloured by the theoretical equilibrium temperature.}
\label{fig:mass_radius}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=8.5cm]{fig4_host_stars.pdf}
\caption{Effective temperature versus visual magnitude for stars hosting transiting exoplanets. MASCARA-1 is indicated by the red star and systems for which the obliquity has been measured are indicated by open circles. MASCARA targets stars to the left of the dashed black line ($4<m_V<8.4$).}
\label{fig:hosts}
\end{figure}
\begin{table}
\small
\centering
\caption{Parameters describing the MASCARA-1 system, derived from the best-fit models to the photometric and spectroscopic data.}
\begin{tabular}{l c c c}
Parameter & Symbol & Value \\
\hline
\hline
Stellar parameters \\
\hline
Identifiers & & HD\,201585, HIP\,104513\\
Right Ascension & & $21^h10^m12.37^s$ \\
Declination & & $+10\degr44\arcmin19.9\arcsec$ \\
Spectral Type & & A8\\
V-band magnitude & $m_V$ & 8.3 \\
Age\tablefootmark{a} & & $1.0\pm0.2~\rm{Gyr}$\\
Effective temperature\tablefootmark{b} & $T_{\rm{eff}}$ & $7554\pm150~\rm{K}$ \\
Projected rotation speed & $v\sin i_\star$ & $109\pm4\rm{~km~s}^{-1}$ \\
Surface gravity & $\log g$ & 4\\
Metallicity & [Fe/H] & 0\\
Stellar mass\tablefootmark{a} & $M_*$ & $1.72\pm0.07~\rm{M}_\odot$ \\
Stellar radius & $R_*$ & $2.1\pm{0.2}~\rm{R}_\odot$ \\
Stellar density & $\rho_*$ & $0.33^{+0.02}_{-0.04}~\rm{g~cm}^{-3}$ \\
\hline
Planet parameters\\
\hline
Planet radius\tablefootmark{c} & $R_p$ & $1.5\pm0.3~\rm{R}_{\rm{Jup}}$ \\
Planet mass & $M_p$ & $3.7\pm0.9~\rm{M}_{\rm{Jup}}$ \\
Planet density & $\rho_p$ & $1.5\pm0.9~\rm{g~cm}^{-3}$\\
Equilibrium temperature\tablefootmark{d} & $T_{\rm{eq}}$ & $2570^{+50}_{-30}~\rm{K}$\\
\hline
System parameters\\
\hline
Epoch & $T_p$ & $2457097.278\pm0.002$ BJD \\
Period & $P$ & $2.148780\pm0.000008$ days \\
Semi-major axis & $a$ & $0.043\pm0.005~\rm{AU}$ \\
Inclination & $i$ & $87\degr{}^{+2}_{-3}$ \\
Eccentricity & $e$ & 0 (fixed) \\
Projected obliquity & $\lambda$ & $69.5\degr\pm3$ \\
\hline
\end{tabular}
\tablefoot{
\tablefoottext{a}{Computed using the {\sc BAGEMASS} code from \citet{Maxted2015}}
\tablefoottext{b}{Taken from \citet{McDonald2012}, assuming a typical uncertainty.}
\tablefoottext{c}{Assuming $p=0.07\pm0.01$.}
\tablefoottext{d}{Computed assuming uniform redistribution and a Bond albedo of zero.}
}
\label{tab:syspars}
\end{table}
Table \ref{tab:syspars} lists the final physical parameters describing the MASCARA-1 system. We find a planetary mass and radius of $3.7\pm0.9~\rm{M}_{\rm{Jup}}$ and $1.5\pm0.3~\rm{R}_{\rm{Jup}}$, albeit with a large uncertainty on the mass caused by the high spin-rotation velocity of the star. The uncertainty on $M_p$ can be reduced by obtaining more RV measurements and we have started a monitoring campaign for this purpose. MASCARA-1\,b orbits a bright A star in $2.148780\pm8\times10^{-6}$ days at a distance of $0.043\pm0.005~\rm{AU}$. The high temperature of the host star means that MASCARA-1\,b has a high equilibrium temperature of $2570^{+50}_{-30}~\rm{K}$ ($A_B=0$), making it one of the hottest gas giants known.
Figure \ref{fig:mass_radius} shows the location of MASCARA-1\,b in the planetary mass-radius diagram. From this we can see that while the radius is large it is not the most extreme case found to date, though it is the most-irradiated of the more massive ($M_p > 3~\rm{M}_{\rm{Jup}}$) hot Jupiters. MASCARA-1\,b also follows the empirical relationship between mass, radius, equilibrium temperature, host star metallicity and tidal heating from \citet{Enoch2012}. Figure \ref{fig:hosts} shows the location of MASCARA-1 compared to other stars hosting transiting exoplanets as a function of visual magnitude and effective temperature. It is clear that MASCARA-1 lies in a part of host star parameter space that has been largely unexplored to date, being a bright early-type star.
Currently there are only a few host stars with effective temperatures larger than $7000$~K for which the stellar obliquity is known (e.g. WASP-33 \cite{CollierCameron2010}, KELT-17 \cite{Zhou2016}, Kepler-13 \cite{Mazeh2012}, see also Fig. \ref{fig:hosts})\footnote{Visit \url{www.astro.keele.ac.uk/jkt/tepcat/rossiter.html} for an up-to-date list of obliquity measurements.} while some constraints exists for other similar systems such as HAT-P-57 \citep{Hartman2015}. This small group of systems does display a large spread in spin orbit alignments and MASCARA-1 does confirm this trend. So far it is not clear what causes the large obliquities. They might be caused by dynamical interactions \citep[e.g. ][]{Fabrycky2007,Nagasawa2008}. However they might also be a general feature of star formation. Proposed mechanisms include: chaotic star formation \citep{Bate2010,Thies2011}, magnetic star-disk interaction \citep{Lai2011,Foucart2011,Spalding2015}, torques due to neighbouring stars \citep{Batygin2012,Lai2014}, tidal dissipation \citep{Rogers2013a}, and internal gravity waves \citep{Rogers2013b}. We can differentiate between the different theories by measuring obliquities in systems with varying multiplicity, planet mass, orbital separation, and stellar mass and structure \citep[e.g.][]{Albrecht2013}. Detecting and characterizing exoplanets orbiting early type main sequence stars should lead to a better understanding of the environment in which planets form.
Since it orbits a bright host star MASCARA-1\,b is a particularly interesting target for atmospheric characterization via transmission spectroscopy. It is reminiscent of the WASP-33 system \citep{CollierCameron2010}, in which a $2.1~M_{\rm{Jup}}$ hot Jupiter transits a $T_{\rm{eff}}=7400~\rm{K}$ A-star. However, while follow-up observations of WASP-33 are significantly hampered by the delta-scuti variations of its host star, no such variability is detected in MASCARA-1. Assuming a temperature of $2550~\rm{K}$ and a hydrogen/helium atmosphere, MASCARA-1\,b will have an atmospheric scale height of $H=215~\rm{km}$, implying that an absorption feature which extends out to $5\times H$ will result in a transmission signal of $0.01\%$. Interestingly, the sodium feature of the host star is significantly weaker than for a solar type star, in addition to being velocity broadened to $106~\rm{km~s}^{-1}$. This will make it significantly more straightforward to isolate the planet atmospheric sodium absorption from stellar effects – e.g due to the Rossiter-McLaughlin effect \citep[e.g.][]{Snellen2008,DiGloria2015}.
The detection of MASCARA-1\,b shows that the Multi-site All-Sky CAmeRA has the potential to increase the number of transiting exoplanets suitable for high-resolution atmospheric studies \citep[e.g.][]{Snellen2010,Brogi2012} as well as expand our knowledge of planets orbiting early-type stars. The Northern MASCARA station on La Palma has now gathered over 2 years of data and the Southern station in La Silla starts observations in June 2017. We expect to find several more planets around both early- and late-type stars in the coming years.
\begin{acknowledgements}
IS acknowledges support from a NWO VICI grant (639.043.107). This project has received funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (grant agreement nr. 694513). Based on observations made with the Mercator Telescope, operated on the island of La Palma by the Flemmish Community, at the Spanish Observatorio del Roque de los Muchachos of the Instituto de Astrofísica de Canarias. Based on observations obtained with the HERMES spectrograph, which is supported by the Research Foundation - Flanders (FWO), Belgium, the Research Council of KU Leuven, Belgium, the Fonds National de la Recherche Scientifique (F.R.S.-FNRS), Belgium, the Royal Observatory of Belgium, the Observatoire de Genève, Switzerland and the Thüringer Landessternwarte Tautenburg, Germany. Based on observations made with the Hertzsprung SONG telescope operated on the island of Tenerife by the Aarhus and Copenhagen Universities in the Spanish Observatorio del Teide of the Instituto de Astrofísica de Canarias. The Hertzsprung SONG telescope is funded by the Danish National Research Foundation, Villum Foundation, and Carlsberg Foundation. This research has made use of the SIMBAD database,
operated at CDS, Strasbourg, France. This research has made use of the VizieR catalogue access tool, CDS, Strasbourg, France. We have benefited greatly from the publicly available programming language {\sc Python}, including the {\sc numpy, matplotlib, pyfits, scipy} and {\sc h5py} packages.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Kernel Adaptive Filters }
Kernel adaptive filters \cite{liu2010} are a class of autoregressive nonlinear models for time series that operate by embedding observations of the signal onto a feature space; then, the feature embeddings are combined with parameters adapted online to produce the prediction. The closed form of the KAF predictor is possible due to the use of a reproducing kernel Hilbert space (RKHS) \cite{scholkopf01} as feature space, this allows us to express inner products among features as kernel evaluations.
\subsection{Model specification}
\label{sub:model_specification}
Let us define a time series by $\{y_i\}_{i\in\mathbb{N}}$, denote the order of the filter by $d$ and the input to the filter at time $i$ by $\mathbf{x}_i = [y_{i-d},\ldots,y_{i-1}]^\top$. The first step of a KAF is to map $\mathbf{x}_i$ onto an RKHS $\mathcal{H}$ through $\phi_{\mathbf{x}_i}$ and then approximate $y_i$ linearly as
\begin{equation}
\hat{y}_i = \langle \phi_{\mathbf{x}_i} , W \rangle
\end{equation}
where $\langle \cdot, \cdot \rangle$ is the inner product in $\mathcal{H}$ and the feature weight $W\in\mathcal{H}$ is chosen according to the Representer Theorem \cite{scholkopf_representer}, which states that the weight $W$ that minimises any square loss is a linear combination of the feature samples $\phi_{\mathbf{x}_i}$, where the $\mathbf{x}_i$ are observations. In the online operation case, this corresponds to considering the entire history of the signal up to time $i$, that is, denoting the new weights $\alpha_{i,j}$, the optimal weight at time $i$ has the form $W_i = \sum_{j=1}^{i-1}\alpha_{i,j}\phi_{\mathbf{x}_j}$. Additionally, as the inner product between feature samples can be expressed as kernel evaluations \cite{scholkopf01}, we have
\begin{align}
\hat{y}_i &= \langle \phi_{\mathbf{x}_i} , \sum_{j=1}^{i-1}\alpha_{i,j}\phi_{\mathbf{x}_j} \rangle\\
&= \sum_{j=1}^{i-1}\alpha_{i,j} \langle \phi_{\mathbf{x}_i} , \phi_{\mathbf{x}_j} \rangle\\
&= \sum_{j=1}^{i-1}\alpha_{i,j} K(\mathbf{x}_i, \mathbf{x}_j)
\end{align}
where the parameters of the model $\alpha_{i,j}$ can be adapted online using e.g., LMS \cite{liu2010} or RLS \cite{engel04}. In this form, the KAF estimator is unsuitable for online operation due to the fact that the number of terms in the above summation grows unbounded, however, we can consider the so-called sparsification criteria that selects a subset of the regressors $\mathbf{x}_i$ to produce sound estimates and keep computational complexity at bay.
\subsection{Sparsification criteria}
\label{sub:sparsification_criteria}
There are three main sparsification criteria used in KAF. First, the novelty criterion \cite{platt91}, where a new sample is added to the dictionary only if it is far enough from the dictionary and if the prediction error associated to that sample was large. Second, the approximate linear dependence (ALD) criterion \cite{engel04}, which includes a new member in the dictionary if the feature sample associated to such element is \emph{approximately linearly dependent} of the feature dictionary members in the RKHS. Third, the coherence criterion \cite{richard09} which is a simplified variant of ALD that only considers the distance between the input and the closest dictionary member instead of the entire dictionary. All sparsification criteria produce a dictionary in time against which the new input is compared (instead of using all historical observations), with this choice, the new form of the KAF predictor is
\begin{align}
\hat{y}_i &= \sum_{j=1}^{i-1}\alpha_{i,j} K(\mathbf{x}_i, \mathbf{s}_j)
\end{align}
where the set $\mathcal{D}_i = \{\mathbf{s}_j\}_{N_i}$ is known as dictionary at time $i$ and the vectors $\mathbf{s}_j$ as centres.
Sparsification criteria for KAFs ensure finite-sized dictionary when the signal lies on a compact set \cite{honeine15}, however, when the signal is monotonically increasing these criteria keep including more centres in the dictionary, thus resulting in higher computational complexity due to having dictionary members that do not necessarily improve predictions.
\subsection{Illustrative example using synthetic data}
\label{sub:example}
We considered a synthetic sinewave and implemented a kernel least mean square algorithm \cite{liu08} using the novelty sparsification criterion, the motivation for this example was to train KLMS on data that lie on a compact set and therefore see that the dictionary remains bounded. The result of the implementation is shown in Fig. \ref{fig:KLMS_sine}, in the top plot the true sinewave signal (red) and the KLMS prediction (blue) are shown, together with the values of the signal that required the addition of dictionary elements in black crosses; the bottom plot shows the number of support vectors accumulated in time. From this figure we can see that the aim of the sparsification criterion chosen is fulfilled: the addition of samples to the dictionary occurs in the initial part of the experiment since the range of the data is remains fixed in time .
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\textwidth]{img/sine_KLMS.pdf}
\caption{KLMS prediction for a synthetic sinewave: Top plot shows the true and predicted signals together with the centres added and the bottom plot the accumulated number centres in time}
\label{fig:KLMS_sine}
\end{figure}
\section{Conclusions}
To overcome the increased computational complexity related to learn monotonic time series using kernel adaptive filters, we have proposed the unit-norm Gaussian kernel and show how it can be used within a kernel least mean square setting. The proposed approach only assess similarity between observed inputs and the dictionary in terms of their direction and not magnitude, thus avoiding redundancies among samples that have similar shape but different magnitude. We have illustrated the shortcomings of standard KAFs that include observations in the dictionary based on the Euclidean norm only, and have validated our proposed approach to sidestep this drawback. Through experimental validation using real-world data, we have confirmed that the unit-norm approach to KAF outperforms standard methods in terms of dictionary size (model complexity) and normalised mean square error (predictive ability).
\section{Simulations}
We validated the proposed unit-norm KLMS against the standard KLMS in the prediction of two real-world time series. The first experiment considers the sunspots time series, where the aim was to show that the unit-norm KLMS behaves well for compact-range signals just as the standard KLMS. The second experiment studied the Mauna Loa CO$_2$ concentration time series, where the size of the dictionary of the standard KLMS grows unbounded, whereas that of the proposed unit-norm KLMS does not. Both time series are available from Python's \texttt{statsmodel} datasets.
\subsection{Sunspots time series}
\label{sub:sunspots}
We considered 250 samples of the yearly sunspot time series and implemented both the proposed unit-norm and standard KLMS algorithms with the novelty criterion described in the previous sections. The results are shown in Figs. \ref{fig:KLMS_sunspots} for KLMS and \ref{fig:PKLMS_sunspots} for the unit-norm KLMS. For both figures, the title shows the normalised mean square error (NMSE) and the final dictionary size, and the top plot shows the true sunspots signal (thick red line) and the KLMS predictions (thin blue line), together with the values of the signal that required the addition of dictionary centres in black crosses. The bottom plots in both figures show the number of dictionary centres accumulated in time. We can see how the sparsification criteria worked for both kernel predictors by only including centres durign the initial parts of the experiment, since the range of the signal is compact. The peaks of the estimate for the unit-norm KLMS can be attributed to the fact that the proposed algorithm does not consider the magnitude of the input and therefore the prediction tends to grow unless it is learnt otherwise, this resulted in a slightly higher NMSE for the proposed model.
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\textwidth]{img/sunspots_KLMS.pdf}
\caption{KLMS prediction for the sunspots time series: Top plot shows the true and predicted signals together with the support vectors added and the bottom plot the accumulated number of support vectors in time}
\label{fig:KLMS_sunspots}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\textwidth]{img/sunspots_PKLMS.pdf}
\caption{Unit-norm KLMS prediction for the sunspots time series: Top plot shows the true and predicted signals together with the support vectors added and the bottom plot the accumulated number of support vectors in time}
\label{fig:PKLMS_sunspots}
\end{figure}
\subsection{Mauna Loa CO$_2$ time series}
\label{sub:maunaloa}
The Mauna Loa dataset contains the weekly concentration of CO$_2$ collected at the Mauna Loa Observatory in Hawaii, between 1958 and 2001; the series has missing values, which we replaced by the previous value for this experiment. This signal has both increasing and semi-periodic components. From Fig. \ref{fig:KLMS_maunaloa} we can see how the number of dictionary centres grows linearly with the range of the signal when using the standard KLMS, this evidences the drawback of the KLMS approach for increasing time series, even though the signal follows a fairly constant semi-periodic pattern with increasing trend.
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\textwidth]{img/MaunaLoa_KLMS.pdf}
\caption{Standard KLMS prediction for the Mauna Loa CO$_2$ time series: Top plot shows the true and predicted signals together with the support vectors added and the bottom plot the accumulated number of support vectors in time}
\label{fig:KLMS_maunaloa}
\end{figure}
Conversely, the proposed unit-norm KLMS predictor only
required two dictionary centres to predict the Mauna Loa time series with high accuracy, thus representing an improvement over the standard KLMS in terms of the normalised mean square error---see Fig. \ref{fig:PKLMS_maunaloa}. An interesting feature of the unit-norm KLMS prediction of the Mauna Loa dataset can be appreciated analysing the dictionary: Fig. \ref{fig:centres} shows the two centres considered by the algorithm, this suggests that the algorithm evaluates similarity between the input and the convex and concave parts of the signal to perform the prediction. This revealed that the algorithm learnt the repetitive behaviour of the series owing to the unit-norm representation, since the standard KLMS does not recognise repetitive patterns if they are of different magnitude.
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\textwidth]{img/MaunaLoa_PKLMS.pdf}
\caption{Unit-norm KLMS prediction for the Mauna Loa CO$_2$ time series: Top plot shows the true and predicted signals together with the support vectors added and the bottom plot the accumulated number of support vectors in time}
\label{fig:PKLMS_maunaloa}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\textwidth]{img/centres.pdf}
\caption{Dictionary centres of unit-norm KLMS for the Mauna Loa CO$_2$ time series. These elements represents the information encapsulated by the proposed unit-norm KLMS, where the algorithm only requires to assess whether the signal is in the concave or convex part of the cycle, but not its magnitude, to compute the prediction.}
\label{fig:centres}
\end{figure}
\section{Introduction}
Support vector machines (SVM) \cite{scholkopf01} are known to be a competitive, intuitive and theoretically-grounded alternative to neural networks within the Machine Learning community. Although SVM were originally envisioned to be trained online using batches of data, novel kernel methods for regression have been designed in the last decade to learn sequentially, this class is termed kernel adaptive filters (KAFs) \cite{liu2010}. The construction of KAFs has been possible by combining the properties of representation on high-dimensional feature spaces \cite{steinwart01} together with adaptive learning approaches from the Signal Processing community, using e.g. the least-mean-square (LMS) or recursive-least-square (RLS) rationales \cite{haykin08}. Due to the simplicity of their implementation and intuitive presentation, KAFs have been used in a number of applications from medicine \cite{georga16} to telecommunications \cite{van2006sliding}; moreover, KAF is an active field of research in terms of kernel design \cite{tobar_quat2,tobar_quat1,paul_quat}, automatic determination of model orders \cite{Zhao2016}, and learning approaches \cite{tobar15a}.
In the same manner that SVM and similarity-based modelling \cite{tobar_sim} operate, KAF predicts values of a time-series assessing the similarity between the observations of the signal and a dictionary of historical input-output data. This notion of similarity, given by the kernel function, is rarely studied in practice, where the standard Gaussian kernel is the \textit{de facto} alternative, and the similarity between observations and the dictionary is measured in terms of the Euclidean distance. One drawback of this choice of similarity is that when the signal moves to a region that is unknown for the dictionary, the KAF prediction reverts to zero even when it is clear that the signal cannot take that value. This makes KAF struggle to learn very simple signals, for instance, a linear function would require a monotonically-increasing number of dictionary members, since the signal is always increasing in magnitude and therefore moving away from the current dictionary.
We address this issue by using a unit-norm representation of the input to the kernel, meaning that the similarity between observations and dictionary centres is calculated as the Euclidean distance of their normalised (unit-norm) versions, or in other words, they are compared \textit{up to their magnitude}. Our hypothesis is that this will help us to acquire knowledge from one region and then extrapolate this knowledge to regions where the signal looks similar but with a different magnitude. We achieve this by proposing a novel kernel, the unit-norm Gaussian kernel, that retains the useful properties of the standard Gaussian kernel but at the same time incorporates the unit-norm comparison. We also describe a specific sparsification criteria for our model, which builds the dictionary from unit-norm data, this allows us to derive a simple and intuitive form of the proposed kernel predictor. Through illustrative examples using both synthetic and real-world datasets we show the drawbacks of current KAF methods and how these are sidestepped by the proposed unit-norm formulation of KAF.
\section{KAF with Unit-Norm Dictionary Centres}
\label{sec:a_novel_kernel_adaptive_filter}
Our aim is to overcome the unbounded growth of the dictionary of kernel adaptive filters when the range of the signal expands in time. To this end, we will populate the dictionary with normalised (unit-norm) versions of the observed samples, this way, the dictionary will not contain centres that are close-to-proportional to one another, but it will only contain centres that differ in their direction rather than magnitude. Furthermore, as the dictionary will be \emph{blind} to the magnitude of the data, we will preserve the input magnitude by amplifying the kernel evaluation by the magnitude of the input. We present our approach by proposing a novel kernel as follows.
\subsection{The unit-norm Gaussian kernel}
\label{sub:the_unit_norm_gaussian_kernel}
A standard kernel within KAF is the Gaussian kernel defined by
\begin{equation}
K_\text{G}(\mathbf{x}_1,\mathbf{x}_2) = \exp\left(\frac{-1}{2l^2} ||\mathbf{x}_1-\mathbf{x}_2||^2 \right)
\label{eq:SE}
\end{equation}
where $||\cdot||$ denotes the norm in the input space and $l>0$ is referred to as the kernel lengthscale.
We propose a novel kernel termed unit-norm Gaussian kernel (UG) defined as follows
\begin{equation}
K_\text{UG}(\mathbf{x}_1,\mathbf{x}_2) = ||\mathbf{x}_1||\exp\left(\frac{-1}{2l^2} ||\vec{\mathbf{x}}_1-\vec{\mathbf{x}}_2||^2 \right)||\mathbf{x}_2||
\label{eq:NSE}
\end{equation}
where $\vec{\mathbf{x}} = \mathbf{x}/||\mathbf{x}||$ is the unit-norm (normalised) version of $\mathbf{x}$ that preserves the direction of $\mathbf{x}$ but not its magnitude. Let us see that the proposed $K_\text{UG}$ kernel is a valid positive-definite kernel, in effect, denoting by $\phi$ the eigenfunction of the $K_\text{G}$ kernel, i.e., $K_\text{G}(\mathbf{x}_1,\mathbf{x}_2) = \langle \phi_{\mathbf{x}_1}, \phi_{\mathbf{x}_2}\rangle$, the proposed $K_\text{UG}$ can be decomposed into the inner product form
\begin{align}
K_\text{UG}(\mathbf{x}_1,\mathbf{x}_2) &=||\mathbf{x}_1||K_\text{G}(\vec{\mathbf{x}}_1,\vec{\mathbf{x}}_2)||\mathbf{x}_2|| \\
&=||\mathbf{x}_1|| \langle \phi_{\vec{\mathbf{x}}_1}, \phi_{\vec{\mathbf{x}}_2}\rangle\mathbf{x}_2|| \nonumber\\
&= \langle ||\mathbf{x}_1||\phi_{\vec{\mathbf{x}}_1}, ||\mathbf{x}_2||\phi_{\vec{\mathbf{x}}_2} \rangle\nonumber\\
&= \langle \psi_{\mathbf{x}_1}, \psi_{\mathbf{x}_2} \rangle\nonumber
\end{align}
where $\psi_{\mathbf{x}} = ||\mathbf{x}||\phi_{\vec{\mathbf{x}}}$ is the feature expansion of the proposed $K_\text{UG}$ kernel. Therefore, as $K_\text{UG}$ admits the above inner product decomposition, it is a positive definite kernel \cite{scholkopf01}.
Besides the fact that the proposed kernel aims to avoid redundancies in the dictionary by only considering unit-norm versions of the inputs, $K_\text{UG}$ in eq. \eqref{eq:NSE} poses a clear advantage in parameter setting. In the standard Gaussian kernel in eq. \eqref{eq:SE} the lengthscale parameter of the kernel is set based on both the dimension and magnitude of the input samples, meaning that for each new set of data, the lengthscale has to be hand-tuned using heuristics or computationally-demanding methods \cite{dsp2017a}. Conversely, as all the inputs to the proposed $K_\text{UG}$ kernel are unit-norm, the lengthscale parameter can be chosen based only on the dimension, in fact, by assuming that the difference $\vec{\mathbf{x}}_1-\vec{\mathbf{x}}_2$ in eq. \eqref{eq:NSE} is isotropic, the lengthscale parameter $l$ can be set to $l=l_0\sqrt{d}$, where $d$ is the dimension of the input and $l_0$ is the \textit{lengthscale per coordinate}. Experimental evaluations revealed that values for $l_0$ between 1 and 3 are suitable options for a wide range of data sets when using the proposed kernel $K_\text{UG}$.
\subsection{KLMS with the unit-norm Gaussian kernel}
\label{sub:klms_with_the_unit_norm_gaussian_kernel}
Taking advantage of the relationship between $K_\text{G}$ and $K_\text{UG}$ in eqs. \eqref{eq:SE}-\eqref{eq:NSE}, the kernel prediction of a time series using the unit-norm Gaussian kernel can be expressed in terms of the standard Gaussian kernel as
\begin{align}
\hat{y}_i &= \sum_{j=1}^{N_i}\alpha_{i,j}K_\text{UG}(\mathbf{x}_i,\mathbf{s}_j)\\
&= \sum_{j=1}^{N_i}\alpha_{i,j}||\mathbf{x}_i||K_\text{G}(\vec{\mathbf{x}}_i,\vec{\mathbf{s}}_j)||\mathbf{s}_j||\nonumber
\end{align}
Furthermore, if we only record the unit-norm versions of the input samples in the dictionary, the above expression can be further simplified into
\begin{align}
\hat{y}_i = \sum_{j=1}^{N_i}\alpha_{i,j}||\mathbf{x}_i||K_\text{G}(\vec{\mathbf{x}}_i,\vec{\mathbf{s}}_j)
\label{eq:unKLMS}
\end{align}
where recall that $\forall j\ \vec{\mathbf{s}}_j = \mathbf{s}_j$ since they are unit-norm.
Choosing the LMS update rule for the weights $\{\alpha_{i,j}\}_{i\in\mathbb{N}, j=1:N_i}$ (KLMS), that is, denoting the prediction error at time $i$ as $e_i = y_i - \hat{y}_i$, the update rule of the $j^\text{th}$ weight, with learning rate $\mu_0$, is given by
\begin{align}
\alpha_{i+1,j} &= \alpha_{i,j} - \frac{\mu_0}{2}\frac{\partial e_i^2}{\partial\alpha_{i,j}}\\
& = \alpha_{i,j} - \mu_0 e_i \frac{\partial e_i}{\partial\alpha_{i,j}}\nonumber\\
& = \alpha_{i,j} + \mu_0 \frac{\partial }{\partial\alpha_{i,j}}\left(\sum_{j=1}^{N_i}\alpha_{i,j}||\mathbf{x}_i||K_\text{G}(\vec{\mathbf{x}}_i,\vec{\mathbf{s}}_j)\right)\nonumber\\
& = \alpha_{i,j} + \mu_0 ||\mathbf{x}_i||K_\text{G}(\vec{\mathbf{x}}_i,\vec{\mathbf{s}}_j).\nonumber
\end{align}
Notice that unlike the KLMS with Gaussian kernel, the sequential correction to the weights depends of the magnitude of the input, this hinders the choice of the learning rate since the optimal choice for $\mu_0$ depends directly from the norm of the data samples and has to be tuned for each dataset---if $\mu_0$ is not chosen carefully, convergence of the KLMS is not guaranteed. As we are precisely interested in cases where the signal grows unbounded, we use the normalised version of KLMS instead, this is achieved by choosing the learning rate to be inversely proportional to the square norm of the input in eq. \eqref{eq:unKLMS}, that is,
\begin{equation}
\mu_0 = \mu\left(\epsilon + ||\mathbf{x}||^{2}\sum_{i=1}^{N_t}K_\text{G}^2(\mathbf{x},\vec{\mathbf{s}}_i)\right)^{-1}
\end{equation}
where $\epsilon>0$ is a constant that prevents the above quantity from going to infinity. Setting the new learning rate $\mu$ is now straightforward, since from the LMS convergence properties we know that convergence is guaranteed for $\mu\in(0,1)$. With this choice, the weight update rule is given by
\begin{align}
\alpha_{i+1,j} &= \alpha_{i,j} + \mu \frac{||\mathbf{x}_i||K_\text{G}(\vec{\mathbf{x}}_i,\vec{\mathbf{s}}_j)}{\epsilon + ||\mathbf{x}_i||^2\sum_{j=1}^{N_i}K_\text{G}^2(\mathbf{x}_i,\vec{\mathbf{s}}_j)},\ \mu\in (0,1)
\end{align}
Observe that, as $\epsilon$ is chosen to be as close to zero as possible, the correction term of the weight update of the proposed normalised KLMS with unit-norm Gaussian kernel resembles that of the standard Gaussian kernel but divided by $||\mathbf{x}_i||$. This has a clear interpretation, since the unit-norm Gaussian kernel evaluation in eq. \eqref{eq:NSE} is inversely proportional to the norm of the input and therefore the increments need to be bounded so the estimate does not diverge. From now on we will refer to the proposed approach as the unit-norm KLMS.
\subsection{Sparisification}
\label{sub:sparisification}
One of the differences between the proposed unit-norm Gaussian kernel in eq. \eqref{eq:NSE} and the standard one in eq. \eqref{eq:SE} is that the latter is a measure of similarity, whereas the former is not due to the proportionality to the norm of the input. For this reason, we consider the novelty sparsification criteria for unit-norm KLMS expressed in terms of the standard Gaussian kernel instead of the Euclidean distance between the unit-norm samples. Specifically, a new input sample $\mathbf{x}$ with unit-norm version $\vec{\mathbf{x}}=\mathbf{x}/||\mathbf{x}||$ and output $y$ is added to the dictionary if (i) it is far enough from the dictionary:
\begin{align}
K_\text{G}(\vec{\mathbf{x}},\mathcal{D}_i) = \max_{\vec{\mathbf{s}}_j\in\mathcal{D}_i} \ K_\text{G}(\vec{\mathbf{x}},\vec{\mathbf{s}}) < \delta_\text{dict}
\end{align}
and (ii) if the relative error associated to the prediction of $y$ given the input $\mathbf{x}$ is large enough:
\begin{align}
|e|/|y| = \frac{1}{|y|}\left|y-\sum_{j=1}^{N_i}\alpha_{i,j}K_\text{UG}(\mathbf{x},\vec{\mathbf{s}}_j)\right| > \delta_\text{pred}.
\end{align}
The reason to compute distance to the dictionary using the Gaussian kernel is that the parameter $\delta_\text{dict}$ is between 0 and 1, and that comparison with the standard Gaussian-kernel KLMS is direct.
\section*{Acknowledgments}
This work was partially supported by Conicyt projects PAI-82140061 and Basal-CMM.
\bibliographystyle{ieeetr}
|
\section*{Introduction}
Exactly solvable Busch {\it et al.} model\cite{Busch} of two ultra-cold bosons confined in a harmonic trap and interacting via contact forces was one of the milestones bringing us closer to our understanding of strongly correlated many-body systems. Although the model deals with only two particles, because of its exact solutions, it inspired many theoretical and experimental studies on collective properties of ultra-cold atoms which are far beyond a simple perturbative description \cite{Blume,Mack,Idziaszek,Astrakharchik,Werner,Idziaszek2,Stetcu,Duerr,Sowinski,Rontani,Sowinski2,March}. Particularly, exact solutions of the model were essential for studies of the Tonks-Girardeau limit of infinite repulsions between particles \cite{Girardeau,Girardeau2,Girardeau3,Kinoshita,Paredes,Murphy,Goold,Lapeyre,Yin}. The validity of the model in a wide range of interactions was finally confirmed in beautiful experiments \cite{Stoferle,Zurn} with a few ultra-cold particles.
Here, we present a wide generalization of the one-dimensional Busch {\it et al.} model to the case of two quantum particles (bosons as well as fermions) interacting via the force of a finite range. With this model and its analytical solutions, it is possible to examine easily different effects caused by the strength and the range of the mutual forces,
restoring well-known results in limiting cases and discovering unsuspected properties in the intermediate regime where the system smoothly transitions between them. Although the model studied seems to be artificial, it approximates the real inter-particle interaction much closer than an oversimplified zero-range potential. Moreover, a proposed shape of the interaction potential can be quite well engineered experimentally in systems of ultra-cold atoms where mutual interactions and shape of an external potential may be controlled with amazing accuracy \cite{Porras,Kim,Carr,Saffman,Islam}. Particularly in the context of ultra-cold Rydberg atoms, due to the Rydberg blockade phenomena, the shape of the inter-particle interaction potential is very close to the shape studied here\cite{Ryd1,Ryd2,Ryd3,Ryd4}. From this point of view, the model studied may have some importance for understanding of the general problem of two interacting quantum particles.
\section{The Model}
In the following we study properties of the system of two identical quantum particles of mass $m$ confined in a one-dimensional harmonic trap of frequency $\Omega$ and interacting via soft-core finite-range rectangular potential. The Hamiltonian of the system reads
\begin{equation} \label{Ham}
{\cal H} = -\frac{\hbar^2}{2m}\left(\frac{\partial^2}{\partial x_1^2}+\frac{\partial^2}{\partial x_2^2}\right)+\frac{m\Omega^2}{2}\left(x_1^2+x_2^2\right)+{\cal V}(x_1-x_2),
\end{equation}
where the interaction potential ${\cal V}(x)$ has a form
\begin{equation}
{\cal V}(x) = \left\{
\begin{array}{ll}
V, &\textrm{if}\,\,|x|< a\\
0, & \textrm{if}\,\, |x| \geq a
\end{array} \right.\label{inter},
\end{equation}
{\it i.e.}, the interaction energy is constant and it is non-zero only when the distance between particles is not larger than $a$. Similar model was considered previously in three dimensions \cite{DeurePhD}. Depending on the situation we consider symmetric (for bosons) or antisymmetric (for fermions) wave functions with respect to an exchange of particles' positions
\begin{equation} \label{symetria}
\Psi(x_1,x_2) = \pm \Psi(x_2,x_1).
\end{equation}
It is quite natural that in the limit of $a\rightarrow 0$ with constrain $2aV=\mathrm{const}$ one restores the Busch {\it et al.} model with delta-like contact interaction \cite{Busch}, while in the limit $V\rightarrow\infty$ an extensively studied model of hard spheres is obtained \cite{Girardeau}.
Our aim is to give a straightforward and analytical prescription for the eigenstates of the Hamiltonian \eqref{Ham} as a function of the potential depth $V$ and its range $a$. With these solutions, we examine different properties of a few of the lowest eigenstates in the bosonic and fermionic cases. In particular, we consider different single-particle system characteristics (density profile, momentum distribution) as well as inter-particle correlations reflected in a reduced single-particle density matrix.
\section{The Eigenproblem}
To find eigenstates and corresponding eigenenergies of the Hamiltonian \eqref{Ham} it is very convenient to perform standard transformation to the coordinates of the center-of-mass frame:
\begin{align} \label{NewCoord}
R =\frac{x_1+x_2}{2}, \qquad
r =x_2-x_1.
\end{align}
In these new variables the Hamiltonian \eqref{Ham} can be written as a sum of two independent single-particle Hamiltonians ${\cal H}={\cal H}_R + {\cal H}_r$:
\begin{align}
{\cal H}_R &= -\frac{\hbar^2}{4m}\frac{\mathrm{d}^2}{\mathrm{d} R^2}+{m\Omega^2R^2}
, \label{HamR} \\
{\cal H}_r &= -\frac{\hbar^2}{m}\frac{\mathrm{d}^2}{\mathrm{d} r^2}+\frac{m\Omega^2}{4}r^2+{\cal V}(r). \label{Hamr}
\end{align}
The Hamiltonian ${\cal H}_R$ has a textbook form of the harmonic oscillator and it can be diagonalized straightforwardly.
The Hamiltonian ${\cal H}_r$ has an additional term related to the interactions \eqref{inter} and the corresponding eigenequation, when written in the natural units of an external harmonic oscillator, has the form:
\begin{equation} \label{EigenEq}
\left[-\frac{\mathrm{d}^2}{\mathrm{d} r^2}+\frac{1}{4}r^2+{\cal V}(r)\right]\Phi(r) = E \Phi(r).
\end{equation}
Since the Hamiltonian ${\cal H}_r$ commutes with the operator of the parity inversion, ${\cal P}\!:r \rightarrow -r$, the eigenstates of ${\cal H}_r$ can be chosen as either even or odd functions of the relative position~$r$. They directly correspond to bosonic and fermionic statistics, respectively.
The eigenequation \eqref{EigenEq} has the form of the Weber differential equation \cite{Weber}:
\begin{equation} \label{WebberEq}
\left(-\frac{\mathrm{d}^2}{\mathrm{d}r^2}+\frac{1}{4} r^2\right) \Phi(r)=-\left(u+\frac{1}{2}\right) \Phi(r),
\end{equation}
with $u$ equal to $-E+V-1/2$ and $-E-1/2$ for $|r|<a$ and $|r|\geq a$, respectively. The Weber equation was originally studied to solve Laplace equation expressed in parabolic coordinates \cite{Weber} but it appears in different problems of mathematical physics and many of its properties are well known \cite{Merzbacher,Abramowitz}. In the case studied, when the problem is not reduced to the ordinary harmonic oscillator problem ($V\neq 0$), it is very convenient to consider two different pairs of the solutions $\{\varphi_u^{(+)}(r),\varphi_u^{(-)}(r)\}$
and $\{\phi_u^{(+)}(r),\phi_u^{(-)}(r)\}$ having appropriate symmetry under parity transformation ${\cal P}$ but different properties on the boundaries. The first pair can be expressed in terms of the confluent hypergeometric function $_1\mathbf{F}_1$ as:
\begin{align}
\varphi^{(+)}_u(r) &=\mathrm{e}^{-\frac{r^2}{4}} \, _1\mathbf{F}_1\left(\frac{u+1}{2};\frac{1}{2};\frac{r^2}{2}\right), \label{WebberEq1a} \\
\varphi^{(-)}_u(r) &=r\,\mathrm{e}^{-\frac{r^2}{4}} \, _1\mathbf{F}_1\left(\frac{u+2}{2};\frac{3}{2};\frac{r^2}{2}\right). \label{WebberEq1b}
\end{align}
These functions may be consider as appropriate solutions only in the region $|r|< a$ because they are divergent in the infinity, $r\rightarrow \pm\infty$. The second pair of solutions is expressed in terms of other confluent hypergeometric function $\mathbf{U}$:
\begin{align}
\phi^{(+)}_u(r) &= \mathrm{e}^{-\frac{r^2}{4}} \mathbf{U}\left (\frac{u+1}{2};\frac{1}{2};\frac{r^2}{2}\right), \label{WebberEq2a}\\
\phi^{(-)}_u(r) &= r\,\mathrm{e}^{-\frac{r^2}{4}} \mathbf{U}\left(\frac{u+2}{2};\frac{3}{2};\frac{r^2}{2}\right). \label{WebberEq2b}
\end{align}
\begin{figure}
\includegraphics[width=0.7\linewidth]{Fig1.ps}
\caption{Spectrum of the relative motion Hamiltonian ${\cal H}_r$ as a function of interaction strength $V$ and chosen potential range $a$. For clarity, even (odd) solutions corresponding to bosonic (fermionic) cases are plotted with solid (dashed) lines. Note that specific degeneracy between neighboring eigenenergies is established in the hard-core limit ($V\rightarrow\infty$). Energy and interaction strength $V$ are measured in units of $\hbar\Omega$. \label{Fig1}}
\end{figure}
In contrast to the first pair, these functions decay appropriately in the limit $r\rightarrow \pm\infty$ but they do not have appropriate behavior at $r=0$, {\it i.e.}, odd functions $\phi^{(-)}_u(r)$ are discontinuous, whereas even ones $\phi^{(+)}_u(r)$ have discontinuous first derivative. It means that functions $\phi^{(\pm)}_u(r)$ may be considered as appropriate solutions only in the region $|r|\geq a$. Consequently, any solution of the eigenequation \eqref{EigenEq} with energy $E_i$ (which is continuous and has a continuous first derivative in a whole space) may be constructed as following
\begin{equation} \label{solu}
\Phi^{(\pm)}_{i}(r) = {\cal N}_i^{(\pm)}\left\{ \begin{array}{ll}
A^{(\pm)}_\nu\varphi^{(\pm)}_\nu(r), & |r|<{a} \\
\phi^{(\pm)}_\mu(r), & |r|\geq{a}
\end{array}\right. ,
\end{equation}
where $\nu=-E_i+V-1/2$, $\mu= -E_i-1/2$ and ${\cal N}_i^{(\pm)}$ is a normalization coefficient of the resulting function. An additional coefficient $A^{(\pm)}_\nu$ together with the eigenenergy $E_i$ are determined by matching conditions at $|r|=a$, ensuring that the function \eqref{solu} and their first derivatives are continuous in the whole space:
\begin{align}
A_\nu^{(\pm)}\varphi_\nu^{(\pm)}(a)&=\phi_\mu^{(\pm)}(a), \label{transa}\\
\left.A_\nu^{(\pm)}\frac{\mathrm{d}}{\mathrm{d}r}\varphi_\nu^{(\pm)}(r)\right|_{r=a} &= \left.\frac{\mathrm{d}}{\mathrm{d}r}\phi_\mu^{(\pm)}(r)\right|_{r=a}. \label{transb}
\end{align}
As typical for such problems, the conditions \eqref{transa} and \eqref{transb} can be fulfilled only for some particular values of an eigenenergy $E_i$ leading directly to the quantization of the physical spectrum. In the case studied, the matching conditions \eqref{transa} and \eqref{transb} are fulfilled when eigenenergy $E_i$ is a solution of the following transcendental equation
\begin{equation}
\phi_\mu^{(\pm)}(a)\left.\frac{\mathrm{d}}{\mathrm{d}r}\varphi_\nu^{(\pm)}(r)\right|_{r=a}-
\varphi_\nu^{(\pm)}(a)\left.\frac{\mathrm{d}}{\mathrm{d}r}\phi_\mu^{(\pm)}(r)\right|_{r=a}=0, \nonumber
\end{equation}
where $\nu=-E_{i}+V-1/2$, $\mu= -E_{i}-1/2$. It directly leads to the following equations
\begin{equation} \label{App1}
2 (\nu +1) \, _1\mathbf{F}_1\left(\frac{\nu +3}{2};\frac{3}{2};\frac{a^2}{2}\right) \mathbf{U}\left(\frac{\mu +1}{2};\frac{1}{2};\frac{a^2}{2}\right)+(\mu +1) \,
_1\mathbf{F}_1\left(\frac{\nu +1}{2};\frac{1}{2};\frac{a^2}{2}\right) \mathbf{U}\left(\frac{\mu +3}{2};\frac{3}{2};\frac{a^2}{2}\right)=0,
\end{equation}
and
\begin{equation} \label{App2}
3 (\mu +2) \, _1\mathbf{F}_1\left(\frac{\nu +2}{2};\frac{3}{2};\frac{a^2}{2}\right) \mathbf{U}\left(\frac{\mu +4}{2};\frac{5}{2};\frac{a^2}{2}\right)+2 (\nu +2) \,
_1\mathbf{F}_1\left(\frac{\nu +4}{2};\frac{5}{2};\frac{a^2}{2}\right) \mathbf{U}\left(\frac{\mu +2}{2};\frac{3}{2};\frac{a^2}{2}\right)=0
\end{equation}
determining even and odd solutions, respectively.
Equations \eqref{App1} and \eqref{App2} are quite complicated but they can be solved straightforwardly with simple numerical methods. After determining eigenenergies one finds the corresponding coefficients
\begin{equation}
A_{\nu}^{(\pm)}=\frac{\phi_{\mu}^{(\pm)}(a)}{\varphi_{\nu}^{(\pm)}(a)},
\end{equation}
and thus corresponding wave functions of the relative motion \eqref{solu}.
It is worth mentioning that in the limiting case of noninteracting particles ($V=0$) standard solutions of a one-dimensional harmonic oscillator are restored. Indeed, in this particular case one finds $\mu=\nu=-E_i -1/2$ and the matching conditions \eqref{transa} and \eqref{transb} reduce to a simple demanding that eigenenergies are half-integer numbers, $E_i=i + 1/2$. In consequence, appropriate functions $\varphi^{(\pm)}_\nu(r)$ and $\phi^{(\pm)}_\mu(r)$ become equivalent and they are expressed in terms of Hermite polynomials $\mathbf{H}_i$:
\begin{equation}
\Phi_i(r) = {\cal N}_i\,\mathrm{e}^{-\frac{r^2}{4}}\,\mathbf{H}_i\left(\frac{r}{\sqrt{2}}\right).
\end{equation}
In the opposite limit of infinite repulsions ($V\rightarrow\infty$) situation is also simplified. In this case the relative wave functions \eqref{solu} must vanish in a whole range $|r|<a$, {\it i.e.}, all amplitudes $A_\nu^{(\pm)}=0$. It immediately leads to the simplified quantization condition
\begin{equation}
\phi^{(\pm)}_\mu(a)=0,
\end{equation}
and in consequence to the typical for hard-core limit degeneracy between neighboring even and odd solutions.
Having analytical solutions \eqref{solu} one can express any eigenstate of the Hamiltonian \eqref{Ham} as a simple product of two wave functions
\begin{equation} \label{EigenGeneral}
\Psi_{ij}(x_1,x_2) = \Upsilon_i\left(\frac{x_1+x_2}{2}\right)\Phi_j(x_1-x_2),
\end{equation}
where $\Upsilon_i(R)$ and $\Phi_j(r)$ are appropriate eigenstates of the center-of-mass and relative motion Hamiltonians, respectively. Although these two particular coordinates ($R$ and $r$) are completely decoupled, the wave function \eqref{EigenGeneral} cannot be written (for any finite $V$) as a product (for bosons) or single Slater determinant (for fermions) of wave functions of independent particles. This observation leads directly to non trivial quantum correlations between particles which are discussed in the following.
\section*{Spectral properties}
Many properties of the system studied can be extracted directly from the spectrum of the relative motion Hamiltonian \eqref{Hamr}. In Fig.~\ref{Fig1} we show several the lowest eigenenergies of ${\cal H}_r$ as functions of the interaction strength $V$ for different potential ranges $a$. Solid lines correspond to even wave functions (bosons) while dashed lines to odd cases (fermions). As suspected, for $V=0$ the spectrum of noninteracting particles is restored, {\it i.e.}, alternating bosonic (symmetric) and fermionic (antisymmetric) states of the relative motion have equally distributed energies $\hbar/2$, $3\hbar/2$, $5\hbar/2$, {\it etc}. For non-vanishing interactions $V\neq 0$, depending on the potential range $a$, eigenenergies vary. A rapidity of these changes crucially depends on a sign of interactions -- for the attraction is much higher than for the repulsion. It is interesting to note that for any finite range ($a\neq 0$) and sufficiently large attractions each eigenstate may have arbitrary large negative energy. Note also that, independently on the potential range $a$, energies never cross. It means that for any finite $a$ and $V$ any eigenstate of the system is not degenerated. In fact, it is a direct consequence of one-dimensionality of the relative motion Hamiltonian ${\cal H}_r$ \cite{LandauBook}.
In the particular limit of strong repulsions, the neighboring states of opposite symmetry become degenerate independently on the interaction range $a$. This observation is a direct consequence of the Bose-Fermi mapping \cite{Girardeau,Girardeau2}.
\begin{figure}
\includegraphics[width=0.5\linewidth]{Fig2.ps}
\caption{Even (bosonic) part of the spectrum of the relative motion Hamiltonian ${\cal H}_r$ as a function of rescaled interaction strength $g=2aV$ for two different values of the interaction range: $a=0.5$ (thin dashed blue line) and $a=0.15$ (thin solid red line). Along with vanishing $a$ the limiting case of contact forces $a\rightarrow 0$ (thick black line) is obtained in a wide range of interactions. Energy and rescaled interaction strength $g$ are measured in units of $\hbar\Omega$ and $\sqrt{\hbar^3\Omega/m}$, respectively. \label{Fig2}}
\end{figure}
It is clearly seen in Fig.~\ref{Fig1} that properties of even and odd solutions of the relative coordinate eigenproblem are essentially different when the potential range $a$ becomes smaller than the natural length scale of the problem (in dimensionless units equal to $1$). As it is seen, in contrast to bosonic states, energies of fermionic states (dashed lines) become more horizontal, {\it i.e.}, they are less sensitive to the interaction energy strength. This behavior is a consequence of the fact that odd functions always vanish at $r=0$, {\it i.e.}, along with decreasing $a$ the interaction energy rapidly decreases independently on interaction strength $V$. In the limit of vanishing $a$ the interaction is completely described in terms of the $s$-wave scattering which is not present between indistinguishable fermionic particles.
\begin{figure}
\includegraphics[width=0.8\linewidth]{Fig3.ps}
\caption{The density distribution \eqref{DensD} calculated for bosonic (solid red line) and fermionic (dashed blue line) ground-states for different ranges of the potential $a$ (columns) and different potential strengths $V$ (rows). Note that for a sufficiently large strength $V$ the density profile is the same for both statistics (see main text for details). The positions and the densities are measured in units of
$\sqrt{\hbar/(m\Omega)}$ and $\sqrt{m\Omega/\hbar}$, respectively.\label{Fig3}}
\end{figure}
\section*{Contact interactions limit}
Mentioned above limiting case of contact forces can be explored more precisely by considering a formal limit in which the interactions ${\cal V}(r)$ become identical with $\delta$-like potential of the form $g\delta(r)$, {\it i.e.}, in the limit $a\rightarrow 0$ with fixed product $2aV=g=\mathrm{const}$. In this limit, the problem reduces to the celebrated model of two quantum particles interacting via contact forces for which exact analytical solutions are known \cite{Busch}. To show how the limiting spectrum is restored we fix the potential range $a$ and for given limiting interaction $g$ we calculate a rescaled value of potential strength $V=g/(2a)$. In this way one obtains the spectrum of the relative motion Hamiltonian ${\cal H}_r$ for different ranges $a$ rescaled to the interaction strength $g$ of the Busch {\it et al.} model. Results of this procedure adopted to even (bosonic) eigenstates of the relative motion Hamiltonian are presented in Fig.~\ref{Fig2}. As it is seen, with decreasing potential range $a$ corresponding eigenenergies approach the results for contact interactions (thick black line). For $a=0.15$ (red solid line) an agreement is almost perfect in a wide range of interactions. These results are in qualitative agreement with previously obtained finite range corrections obtained within the Green's function approach for higher dimensionality \cite{Zinner}.
At this point, it should be noted that for any finite $a$, in contrast to the contact interactions limit, all eigenenergies become negative for sufficiently strong attractions (see Fig.~\ref{Fig1}). Only in the case of contact interactions there exists exactly one bound state of the Hamiltonian \eqref{Hamr} -- the ground-state of the bosonic system.
\section*{Single-particle quantities}
The simplest quantities which can be measured experimentally quite easily are related to single-particle properties. All of them are fully captured by the reduced single-particle density matrix which for the model studied has a form:
\begin{equation}
\Gamma(x,x') = \int_{-\infty}^\infty\!\mathrm{d}x_2\, \Psi^*(x,x_2)\Psi(x',x_2),
\end{equation}
where $\Psi(x_1,x_2)$ is a chosen two-particle state of the system. In the following we focus on the properties of the bosonic and the fermionic ground-states, {\it i.e.}, according to the notation of eq. \eqref{EigenGeneral} the states $\Psi_{00}(x_1,x_2)$ and $\Psi_{01}(x_1,x_2)$, respectively.
\begin{figure}
\includegraphics[width=0.8\linewidth]{Fig4.ps}
\caption{The distribution of the single-particle momentum \eqref{MomentumD} calculated for bosonic (solid red line) and fermionic (dashed blue line) ground-states for different ranges of the potential $a$ (columns) and different potential strengths $V$ (rows). Although, for sufficiently large strength $V$ density profiles in Fig.~\ref{Fig3} are the same for both statistics, momentum distribution not necessarily have this property. Only for large enough ranges both distributions become equal indicating appearance of crystallization (see main text for details). The momentum and the momentum distributions are measured in units of $\sqrt{m\hbar\Omega}$ and $\sqrt{1/(m\hbar\Omega)}$, respectively.\label{Fig4}}
\end{figure}
Typically, we are mostly interested not in the whole reduced single-particle density matrix $\Gamma(x,x')$ but only in its diagonal part
\begin{equation} \label{DensD}
n(x)=\Gamma(x,x),
\end{equation}
which represents a density profile of particles. Analogously, a diagonal part of its Fourier transform
\begin{equation} \label{MomentumD}
\pi(p) = {1\over 2\pi}\int_{-\infty}^\infty\mathrm{d}x\int_{-\infty}^\infty\mathrm{d}x'\,\, \Gamma(x,x')\,\mathrm{e}^{i p(x-x')/\hbar},
\end{equation}
encodes distribution of a single-particle momentum. Properties of these two simple quantities crucially depend on the range of the potential $a$. These differences are especially manifested in the cases which are beyond applicability of the Busch {\it et al.} model.
In Fig.~\ref{Fig3} and Fig.~\ref{Fig4} we show density and momentum distributions for bosonic (red solid line) and fermionic (dashed blue line) ground-states obtained for a few representative potential ranges ($a=1$, $a=1.5$, and $a=2$) and different potential strengths $V$, including hard-core limit case $V\rightarrow\infty$.
Let us recall that in the hard-core limit, the bosonic wave functions necessarily satisfy the condition $\Psi_{00}(x_{1},x_{2})=0$ on the line $x_{1}=x_{2}$, regardless of $a$ since corresponding wave functions of relative motion $\Phi(r)$ vanish at $r=0$. This observation, usually called fermionization, enables one to map the bosonic ground-state wave function to the fermionic one via the following relation $\Psi_{00}(x_{1},x_{2})=|\Psi_{01}(x_{1},x_{2})|$. In consequence, in the hard-core limit, the bosonic and fermionic ground states share not only the same energy but also have the same spatial density profiles (bottom row in Fig.~\ref{Fig3}). Note however, that this particular mapping (forced by infinite repulsions) does not necessarily mean that also momentum distributions for bosonic and fermionic ground-states are the same (bottom row in Fig.~\ref{Fig4}).
\begin{figure}
\includegraphics[width=0.5\linewidth]{Fig5.ps}
\caption{Two-particle probability density of finding particles at positions $x_1$ and $x_2$ in hard-core limit $V\rightarrow\infty$ for different values of the potential range $a$. For sufficiently large range ($a\gtrsim 2$) particles occupy exactly opposite sides of the trap. This is one of the features of the crystallization mechanism. The positions and the probability distributions are measured in units of $\sqrt{\hbar/(m\Omega)}$ and $m\Omega/\hbar$, respectively.\label{Fig5} }
\end{figure}
The situation becomes essentially different for potential ranges $a\gtrsim 2$. In these cases, not only the density profiles $n(x)$ but also the momentum distributions $\pi(p)$ of bosonic and fermionic ground-states become identical in the hard-core limit (right bottom plots in Figs.~\ref{Fig3} and \ref{Fig4}). At the same time the density profiles exhibit spatial separation into two independent peaks indicating localization of particles at opposite sides of the trap. This result is clearly understandable when two-particle density profile $\rho(x_1,x_2)=|\Psi(x_1,x_2)|^2$ is considered. As it is seen in Fig.~\ref{Fig5} the probability of finding both particles at the same side of the trap ($x_1,x_2>0$ or $x_1,x_2<0$) vanishes when $a\gtrsim 2$. It is often said that the system enters the crystallization regime where individual particles are spatially separated\cite{Deure} and behave as distinguishable parties. In consequence, any physical property of the system does not depend on a quantum statistics. However, as explained in the next section, spatial separation does not mean that individual particles can be treated as independent since non-classical correlations between them are still present.
For finite interaction strengths $V<\infty$ (see Figs.~\ref{Fig3} and \ref{Fig4}) one can observe how the quasi-fermionization is built along with increasing $V$. It is quite instructive to note that the fermionization regime is reached earlier if the interaction range $a$ is bigger. It is consistent with previous results concerning quasi-degeneracy in the spectrum of the relative motion Hamiltonian (compare to Fig.~\ref{Fig1}).
\section*{Inter-particle correlations}
As was mentioned before, the eigenstates \eqref{EigenGeneral} are always separable when written with respect to coordinates \eqref{NewCoord}. However, for any finite interaction, it cannot be written as a product with respect to positions of particles $x_1$ and $x_2$. This simple observation means that the eigenstates of the interacting system studied encode nonclassical correlations between particles. These inter-particle correlations are nicely captured by spectral properties of the reduced single-particle density matrix $\Gamma(x,x')$. It is known that the matrix $\Gamma(x,x')$ can be decomposed to its natural Schmidt orbitals
\begin{equation} \label{sd}
\Gamma(x,x') =\sum_{i}\lambda_{i} u_{i}(x)u_{i}(x'),
\end{equation}
where $u_{i}(x)$ and $\lambda_{i}$ are eigenvectors and eigenvalues of the reduced density matrix $\Gamma(x,x')$, respectively. Coefficients $\lambda_i$ have a direct interpretation of probabilities of finding a single particle in quantum states described by the corresponding orbitals and they are normalized to unity, $\sum_i \lambda_i = 1$. Let us note that in the case of fermions, due to Pauli exclusion principle \eqref{symetria}, all non-zero eigenvalues are doubly degenerated.
If both bosons occupy exactly the same orbital $u_0(x)$ then the reduced density matrix simply projects to the orbital $u_0(x)$ and particles are trivially correlated. What is less intuitive, correlations between particles are also trivial whenever particles occupy two different orbitals $u_0(x)$ and $u_1(x)$ in the way that the two-particle wave function is represented by their Slater determinant (for fermions) or permanent (for bosons). In all these cases, the state is regarded as non-entangled since correlations originate only in the quantum statistics of indistinguishable particles \cite{Schliemann,Ghirardi}. In other cases, additional correlations are forced by mutual interactions and they are directly reflected in increasing number of non-vanishing occupations $\lambda_i$ in the decomposition \eqref{sd}.
\begin{figure}
\includegraphics[width=0.9\linewidth]{Fig6.ps}
\caption{(a) First four of the largest eigenvalues of the reduced single-particle density matrix $\Gamma(x,x')$ calculated for the bosonic (solid black lines) and fermionic (dashed blue lines) ground-state in the hard-core limit $V\rightarrow\infty$ as functions of the potential range $a$. In the limit of contact interactions ($a\rightarrow 0$) fermionic ground-state does not manifest any non-trivial correlations. For large ranges ($a\gtrsim 2$) bosonic eigenvalues become degenerate and equal to appropriate fermionic ones indicating crystallization regime. Note that fermionic eigenvalues are always doubly degenerated. (b-c) The von Neumann entropy $\mathbf{S}$ as a function of interaction strength $V$ for different potential ranges $a$. Note that in the fermionic case its growth is strongly suppressed in the limit of vanishing range ($a\rightarrow 0$) as a consequence of vanishing contact interaction in this limit. In all plots the potential range $a$ and strength $V$ are measured in units of $\sqrt{\hbar/(m\Omega)}$ and $\hbar\Omega$ respectively. \label{Fig6}}
\end{figure}
In the Fig.~\ref{Fig6}a we present four the largest eigenvalues $\lambda_i$ of the reduced density matrix $\Gamma(x,x')$ calculated for bosonic (solid black lines) and fermionic (dashed blue line) ground-states, $\Psi_{00}(x_1,x_2)$ and $\Psi_{01}(x_1,x_2)$, as functions of potential range $a$ in the hard-core limit $V\rightarrow\infty$. As noted above, in the fermionic case eigenvalues are doubly degenerated. In the contact interaction limit ($a\rightarrow 0$), fermions, in contrast to bosons, become noninteracting and therefore there is no additional correlation beyond that induced by quantum statistics ($\lambda_0=\lambda_1=0.5$). In opposite limit of large ranges $a\gtrsim 2$, bosonic occupations $\lambda_i$ become doubly degenerated and equal to fermionic ones. Simultaneously, reduced single-particle matrices of bosonic and fermionic ground-states become identical. By a direct inspection of the two-particle state we found that the ground-state of the system can be written as ($\pm$ sign for bosons and fermions, respectively):
\begin{equation}
\Psi(x_1,x_2) = \sum_i \kappa_i \left[{\cal L}_{i}(x_1){\cal R}_{i}(x_2)\pm {\cal L}_{i}(x_2){\cal R}_{i}(x_1)\right], \nonumber
\end{equation}
where ${\cal L}_i(x)$ and ${\cal R}_i(x)$ are single-particle orbitals localized in left and right side of the trap constructed from the corresponding even and odd single-particle orbitals of the reduced density matrix $\Gamma(x,x')$ \cite{Koscik}:
\begin{align}
{\cal L}_i(x) &= \frac{1}{\sqrt{2}}\left[u_{2i}(x)+u_{2i+1}(x)\right], \\
{\cal R}_i(x) &= \frac{1}{\sqrt{2}}\left[u_{2i}(x)-u_{2i+1}(x)\right].
\end{align}
The construction is possible, since in this case the appropriate eigenorbitals are degenerated and any of their linear combination remains as an eigenvector of $\Gamma(x,x')$. The amplitudes $\kappa_i$ are related directly to the occupations $\lambda_i$, $\kappa_i=\sqrt{\lambda_{2i}}=\sqrt{\lambda_{2i+1}}$.
This observation is one of quite spectacular manifestations of the crystallization mentioned before. Note that the values of the dominant occupations $\lambda_0$ and $\lambda_1$ decreases with potential range $a$. It means that even in the crystallization regime particles cannot be treated as trivially correlated parties and therefore they cannot be locally described with individual well-defined orbitals.
Non-classical correlations between particles are quite well quantified by the von Neumann entropy. When occupancies $\lambda_i$ are known, it can be calculated straightforwardly as ${\mathbf S} = -\sum_i \lambda_i \log_2\lambda_i$. This measure exactly vanishes in the non-entangled product state of bosons ($\lambda_0=1$) and it is equal to 1 in the non-entangled bosonic and fermionic states ($\lambda_0=\lambda_1=0.5$). Note however that in the bosonic case, an opposite implication does not hold, {\it i.e.}, the condition $\mathbf{S}=1$ does not necessarily means that the state in non-entangled \cite{Ghirardi}. In Fig.~\ref{Fig6}b (for bosons) and Fig.~\ref{Fig6}c (for fermions) we show the dependence of the von Neumann entropy on the potential strength $V$ for different ranges $a$ calculated in the ground-state of interacting particles. It is clear that in a considered range of parameters the von Neumann entropy $\mathbf{S}$ is a monotonic function of the interaction $V$ and its growth crucially depends on the potential range $a$. It is also worth noticing that in the fermionic case and vanishing range ($a\rightarrow 0$) the von Neumann entropy remains unchanged independently on the interaction strength $V$. This is a direct consequence of vanishing contact interaction for fermionic species.
\section*{Summary}
In this paper, we present properties of the exactly solvable model of two interacting particles confined in a harmonic trap. Inter-particle forces are modeled by a square wall controlled by two independent parameters: potential range and its strength. The results enabled us to investigate and discuss different properties of the system in a whole range of parameters between limiting cases of well known Busch {\it et al.} and hard-core models. Obtained results suggest that any finite range of the inter-particle forces is directly reflected in simple quantities which can be measured experimentally. The prominent example is the many-body spectrum where, in contrast to the zero-range case, all eigenstates become unbounded from below for attractive forces. We show that also density and momenta distributions maybe strongly affected by the finite range of the potential. All these deviations from the Busch {\it et al.} model maybe examined in recent experiments on a few ultra-cold particles.
The model presented belongs to the specific class of quite realistic quantum many-body problems having exact analytical solutions \cite{UshveridzeBook,LiebBook,Korepin,SutherlandBook}. From this point of view it is not only an interesting academic example. In fact, it may serve as a first building block for constructions of the many-body ground states of larger number of interacting particles. For example, it can be used as an input for the variational Jastrow-like ansatz based on analytical solutions of two-body problems \cite{Jastrow}.
|
\section{Algorithm} \label{sec:overall-algorithm}
\tikzset
not allowed/.style={%
dotted,
very thick,
color=red
}
}
\tikzset
required/.style={%
}
}
\tikzset
optional/.style={%
dashed
}
}
\tikzset
pointO/.style={%
fill=black,regular polygon, regular polygon sides=4,inner sep=1pt
}
}
\newlength\vertSmall
\newlength\vertBig
\newlength\vertBigger
\newlength\labelGap
\newlength\coordGap
Our algorithm is a dynamic program, similar at a high level to CKY \parencite{Cocke:1969,Kasami:1966,Younger:1967}.
The states of our dynamic program (\emph{items}) represent partial parses.
Usually in CKY, items are defined as covering the $n$ words in a sentence, starting and ending at the spaces between words.
We follow \textcite{eisner:1996}, defining items as covering the $n{-}1$ spaces in a sentence, starting and ending on words, as shown in Figure~\ref{fig:alg-example}.
This means that we process each word's left and right dependents separately, then combine the two halves.
We use three types of items:
(1) a single \emph{edge}, linking two words,
(2) a continuous \emph{span}, going from one word to another, representing all edges linking pairs of words within the span,
(3) a span (as defined in 2) plus an additional word outside the span, enabling the inclusion of edges between that word and words in the span.
Within the CKY framework, the key to defining our algorithm is a set of rules that specify which items are allowed to combine.
From a bottom-up perspective, a parse is built in a series of steps, which come in three types:
(1) adding an edge to an item,
(2) combining two items that have non-overlapping adjacent spans to produce a new item with a larger span,
(3) combining three items, similarly to (2).
\tightparagraph{Example}
To build intuition for the algorithm, we will describe the derivation in Figure~\ref{fig:alg-example}.
Note, item sub-types (I, X, and N) are defined below, and included here for completeness.
\begin{figure}
\centering
\scalebox{0.8}{ \input{figures/picture-algorithm} }%
\vspace{-5mm}
\caption{\label{fig:alg-example}
An example derivation using our graph parsing deduction rules.
}
\end{figure}
\noindent
(1) We initialize with spans of width one, going between adjacent words, e.g.\@\xspace between \emph{ROOT} and \emph{We}. \\
\strut\hfill $\emptyset \; \mapsto \; I_{0,1}$\hfill\strut
\noindent
(2) Edges can be introduced in exactly two ways, either by linking the two ends of a span, e.g.\@\xspace \emph{like}--\emph{running}, or by linking one end of a span with a word outside the span, e.g.\@\xspace \emph{like}--.\@\xspace (which in this case forms a new item that has a span and an external word). \\
\strut\hfill$I_{2,3} \; \land \; like$--$running \; \mapsto \; I_{2,3}$\hfill\strut \\
\strut\hfill$I_{3,4} \; \land \; like$--.$ \; \mapsto \; X_{3,4,2}$\hfill\strut
\noindent
(3) We add a second edge to one of the items. \\
\strut\hfill$I_{1,2} \; \land \; running$--$We \; \mapsto \; X_{1,2,3}$\hfill\strut
\noindent
(4) Now that all the edges to \emph{We} have been added, the two items either side of it are combined to form an item that covers it. \\
\strut\hfill$I_{0,1} \; \land \; X_{1,2,3} \; \mapsto \; N_{0,2,3}$\hfill\strut
\noindent
(5) We add an edge, creating a crossing because \emph{We} is an argument of a word to the right of \emph{like}. \\
\strut\hfill$N_{0,2,3} \; \land \; ROOT$--$like \; \mapsto \; N_{0,2,3}$\hfill\strut
\noindent
(7) We use a ternary rule to combine three adjacent items.
In the process we create another crossing. \\
\strut\hfill$N_{0,2,3} \; \land \; I_{2,3} \; \land \; X_{3,4,2} \; \mapsto \; I_{0,6}$\hfill\strut
\subsection{Algorithm definition}
\paragraph{Notation}
Vertices are $p$, $q$, etc.
Continuous ranges are $[pq]$, $[pq)$, $(pq]$, or $(pq)$, where the brackets indicate inclusion, $[\,]$, or exclusion, $(\,)$, of each endpoint.
A span $[pq]$ and vertex $o$ that are part of the same item are $[pq.o]$.
Two vertices and an arrow indicate an edge, $\vec{pq}$.
Two vertices without an arrow are an edge in either direction, $pq$.
Ranges and/or vertices connected by a dash define a set of edges, e.g.\@\xspace the set of edges between $o$ and $(pq)$ is $o$--$(pq)$ (in some places we will also use this to refer to an edge from the set, rather than the whole set).
If there is a path from $p$ to $q$, $q$ is \textit{reachable} from $p$.
\paragraph{Item Types}
As shown in Figure~\ref{fig:alg-example}, our items start and end on words, fully covering the spaces in between.
Earlier we described three item types: an edge, a span, and a span plus an external vertex.
Here we define spans more precisely as $I$, and divide the span plus an external point case into five types differing in the type of edge crossing they contain:
\begingroup
\setlength{\columnsep}{6pt}%
\setlength{\intextsep}{3pt}%
\begin{wrapfigure}[2]{r}{0pt}
\begin{tikzpicture}
\node (p) at (0, 0) {$p$};
\node (q) at (1, 0) {$q$};
\draw (p.north) -- (q.north);
\end{tikzpicture}
\end{wrapfigure}
\noindent \textbf{$I$, Interval} A span for which there are no edges $sr : r \in (pq)$ and $s \notin [pq]$.
\noindent
\begin{wrapfigure}[2]{r}{0pt}
\begin{tikzpicture}
\node (p) at (0, 0) {\phantom{$o$}};
\node (q) at (1.0, 0) {\phantom{$o$}};
\node (o) at (1.5, 0) {$o$};
\node [pointO] at (o.north) {};
\draw (p.north) -- (q.north);
\draw [out=45,in=135] (p.north) to (o.north);
\end{tikzpicture}
\end{wrapfigure}
\noindent \textbf{$X$, Exterval} An interval and either $op$ or $oq$, where $o \notin [pq]$.
\noindent
\begin{wrapfigure}[2]{r}{0pt}
\begin{tikzpicture}
\node (p) at (0, 0) {};
\node (m1) at (0.3, 0) {};
\node (m2) at (0.6, 0) {};
\node (m3) at (0.9, 0) {};
\node (m4) at (1.2, 0) {};
\node (q) at (1.5, 0) {};
\node (o) at (2, 0) {};
\draw (p.center) -- (q.center);
\node [pointO] at (o.center) {};
\draw [out=45,in=135] (m1.center) to (o.center);
\draw [out=45,in=135] (m4.center) to (o.center);
\draw [out=45,in=135] (p.center) to (m2.center);
\draw [out=45,in=135] (m3.center) to (q.center);
\end{tikzpicture}
\end{wrapfigure}
\noindent \textbf{$B$, Both} A span and vertex $[pq.o]$, for which there are no edges $sr : r \in (pq)$ and $s \notin [pq] \cup o$.
Edges $o$--$[pq]$ may be crossed by $pq$, $p$--$(pq)$ or $q$--$(pq)$, and at least one crossing of the second and third types occurs.
Edges $o$--$(pq)$ may not be crossed by $(pq)$--$(pq)$ edges.
\begin{wrapfigure}[3]{r}{0pt}
\begin{tikzpicture}
\node (p) at (0, 0) {};
\node (m) at (0.5, 0) {};
\node (m2) at (1.0, 0) {};
\node (q) at (1.5, 0) {};
\node (o) at (2, 0) {};
\draw (p.center) -- (q.center);
\node [pointO] at (o.center) {};
\draw [out=45,in=135] (m.center) to (o.center);
\draw [out=45,in=135] (p.center) to (m2.center);
\node (rp) at (0, -0.6) {};
\node (rm) at (0.5, -0.6) {};
\node (rm2) at (1.0, -0.6) {};
\node (rq) at (1.5, -0.6) {};
\node (ro) at (2, -0.6) {};
\draw (rp.center) -- (rq.center);
\node [pointO] at (ro.center) {};
\draw [out=45,in=135] (rm2.center) to (ro.center);
\draw [out=45,in=135] (rm.center) to (rq.center);
\end{tikzpicture}
\end{wrapfigure}
\noindent \textbf{$L$, Left} Same as $B$, but $o$--$(pq)$ edges may only cross $p$--$(pq]$ edges. \\
\noindent \textbf{$R$, Right} Symmetric with $L$.
\begin{wrapfigure}[2]{r}{0pt}
\begin{tikzpicture}
\node (p) at (0, 0) {};
\node (m) at (0.5, 0) {};
\node (q) at (1.5, 0) {};
\node (o) at (2, 0) {};
\draw (p.center) -- (q.center);
\node [pointO] at (o.center) {};
\draw [out=45,in=135] (m.center) to (o.center);
\end{tikzpicture}
\end{wrapfigure}
\noindent \textbf{$N$, Neither} An interval and a vertex $[pq.o]$, with at least one $o$--$(pq)$ edge, which can be crossed by $pq$, but no other $[pq]$--$[pq]$ edges.
\endgroup
Items are further specified as described in Alg.~\ref{alg:rules}.
Most importantly, for each pair of $o$, $p$, and $q$ in an item, the rules specify whether one is a parent of the other, and if they are directly linked by an edge.
For an item $H$ with span $[ij]$, define $covered(H)$ as $(ij)$, and define $visible(H)$ as $\{i, j\}$.
When an external vertex $x$ is present, it is in $visible(H)$.
Also, call the union of multiple such sets $covered(F, G, H)$, and $visible(F, G, H)$.
\begin{algorithm*}
\input{rules-compressed}
\vspace{-1mm}
\caption{\label{alg:rules}
Dynamic program for Lock-Free, One-Endpoint Crossing, Directed, Acyclic graph parsing.
}
\end{algorithm*}
\paragraph{Deduction Rules}
To make the deduction rules manageable, we use templates to define some constraints explicitly, and then use code to generate the rules.
During rule generation, we automatically apply additional constraints to prevent rules that would leave a word in the middle of a span without a parent or that would form a cycle (proven possible below).
Algorithm~\ref{alg:rules} presents the explicit constraints.
Once expanded, these give rules that specify all properties for each item (general type, external vertex position relative to the item spans, connectivity of every pair of vertices in each item, etc).
The final item for $n$ vertices is an interval where the left end has a parent.
For parsing we assume there is a special root word at the end of the sentence.
\subsection{Properties}
\begin{definition} \label{def:1ec}
A graph is \textbf{One-Endpoint Crossing} if, when drawn with vertices along the edge of a half-plane and edges drawn in the open half-plane above, for any edge $e$, all edges that cross $e$ share a vertex. Also let that vertex be $\mathcal{P}t$($e$).
\end{definition}
Aside from applying to graphs, this is the same as \textcite{ec}'s 1-EC\@\xspace tree definition.
\begin{definition} \label{def:prob-struct}
A \textbf{Locked-Chain\@\xspace} (shown in Fig.~\ref{fig:bad-structure}) is formed by a set of consecutive vertices in order from $0$ to $N$, where $N > 3$, with edges $\{ (0, N{-}1), (1, N)\} \cup \{(i, i{+}2) \forall i \in [0, N{-}2]\}$.
\end{definition}
\begin{definition} \label{def:lock-free}
A graph is \textbf{Lock-Free} if it does not contain edges that form a Locked-Chain\@\xspace.
\end{definition}
Note that in practise, most parse structures satisfy 1-EC\@\xspace, and the Locked-Chain\@\xspace structure does not occur in the PTB\xspace when using our head rules.
\begin{theorem} \label{theorem:complete}
For the space of Lock-Free One-Endpoint Crossing graphs, the algorithm is sound, complete and gives unique decompositions.
\end{theorem}
Our proof is very similar in style and structure to \textcite{ec}.
The general approach is to consider the set of structures an item could represent, and divide them into cases based on properties of the internal structure.
Then we show how each case can be decomposed into items, taking care to ensure all the properties that defined the case are satisfied.
Uniqueness follows from having no ambiguity in how a given structure could be decomposed.
Completeness and soundness follow from the fact that our rules apply equally well in either direction, and so our top-down decomposition implies a bottom-up formation.
To give intuition for the proof, we show the derivation of one rule below.
The complete proof can be found in \textcite{Kummerfeld:PhD}.
We do not include it here due to lack of space.
We do provide the complete set of rule templates in Algorithm~\ref{alg:rules}, and in the proof of Lemma~\ref{lemma:B-case} we show that the case in which an item cannot be decomposed occurs if and only if the graph contains a Locked-Chain\@\xspace.
To empirically check our rule generation code, we checked that our parser uniquely decomposes all 1-EC\@\xspace parses in the PTB\xspace and is unable to decompose the rest.
Note that by using subsets of our rules, we can restrict the space of structures we generate, giving parsing algorithms for projective DAGs, projective trees \parencite{eisner:1996}, or 1-EC\@\xspace trees \parencite{ec}.
Also, versions of these spaces with undirected edges could be easily handled with the same approach.
\begingroup
\setlength\intextsep{0pt}
\begin{wrapfigure}[3]{r}{0pt}%
\begin{tikzpicture}%
[every fit/.style={rectangle,rounded corners,draw,inner sep=2pt}]%
\node (p) at (0, 0) {};
\node (q) [right=2cm of p] {};
\node (a) at ($(p)!0.15!(q)$) {};
\node (s) at ($(p)!0.4!(q)$) {};
\node (t) at ($(p)!0.65!(q)$) {};
\node (b) at ($(p)!0.85!(q)$) {};
\node (pText) [below=-3pt of p.center] {p\strut};
\node (qText) [below=-3pt of q.center] {q\strut};
\node (sText) [below=-3pt of s.center] {s\strut};
\node (tText) [below=-3pt of t.center] {t\strut};
\draw [thick] (p.center) -- (q.center);
\draw [thick,out=40,in=140,color=black] (p.center) to node (ps) [midway] {} (s.center);
\draw [thick,out=60,in=120,color=black] (a.center) to node (aa) [midway] {} (t.center);
\draw [thick,out=60,in=120,color=black] (s.center) to node (sc) [midway] {} (b.center);
\node () [fit=(p) (q) (ps) (aa) (pText) (qText) (sText) (tText)] {};
\end{tikzpicture}%
\end{wrapfigure}%
\tightparagraph{Derivation of rule (4) in Algorithm~\ref{alg:rules}}
This rule applies to intervals with the substructure shown, and with no parent in this item for $p$.
They have at least one $p$--$(pq)$ edge (otherwise rule 1 applies).
The longest $p$--$(pq)$ edge, $ps$, is crossed (otherwise rule 2 applies).
Let $C$ be the set of $(ps)$--$(sq)$ edges (note: these cross $ps$).
Either all of the edges in $C$ have a common endpoint $t \in (sq)$, or if $|C| = 1$ let $t$ be the endpoint in $(sq)$ (otherwise rule 6 or 7 applies).
Let $D$ be the set of $s$--$(tq)$ edges.
$|D| > 0$ (otherwise rule 3 or 5 applies).
We will break this into three items.
First, $(st)$--$(tq]$ edges would violate the 1-EC\@\xspace property and $(st)$--$[ps)$ edges do not exist by construction.
Therefore, the middle item is an Interval $[st]$, the left item is $[ps.t]$, and the right item is $[tq.s]$ (since $|C| > 0$ and $|D| > 0$).
The left item can be either an $N$ or $R$, but not an $L$ or $B$ because that would violate the 1-EC\@\xspace property for the $C$ edges.
The right item can be an $X$, $L$, or $N$, but not an $R$ or $B$ because that would violate the 1-EC\@\xspace property for the $D$ edges.
We will require edge $ps$ to be present in the first item, and not allow $pt$.
To avoid a spurious ambiguity, we also prevent the first or third items from having $st$ (which could otherwise occur in any of the three items).
Now we have broken down the original item into valid sub-items, and we have ensured that those sub-items contain all of the structure used to define the case in a unique way. \\
\endgroup
\begin{figure}
\centering
\begin{tikzpicture}
\node (p) at (0, 0) {};
\node (pText) [below=-2.0ex of p] {\footnotesize $0$\strut};
\node (sp) at (0.8, 0) {};
\node (spText) [below=-2.0ex of sp] {\footnotesize $1$\strut};
\node (m1) at (1.6, 0) {};
\node (m1Text) [below=-2.0ex of m1] {\footnotesize $2$\strut};
\node (m2) at (2.4, 0) {};
\node (m2Text) [below=-2.0ex of m2] {\footnotesize $3$\strut};
\node (m3) at (3.2, 0) {};
\node (m4) at (4, 0) {};
\node (m5) at (4.8, 0) {};
\node (m5Text) [below=-2.0ex of m5] {\footnotesize $N{-}3$\strut};
\node (m6) at (5.6, 0) {};
\node (m6Text) [below=-2.0ex of m6] {\footnotesize $N{-}2$\strut};
\node (sq) at (6.4, 0) {};
\node (sqText) [below=-2.0ex of sq] {\footnotesize $N{-}1$\strut};
\node (q) at (7.2, 0) {};
\node (qText) [below=-2.0ex of q] {\footnotesize $N$\strut};
\draw [out=15,in=165] (p.north) to (sq.north);
\draw [out=15,in=165] (sp.north) to (q.north);
\draw [out=20,in=160] (p.north) to (m1.north);
\draw [out=20,in=160] (sp.north) to (m2.north);
\draw [out=20,in=160] (m1.north) to (m3.north);
\draw [out=20,in=160] (m2.north) to (m4.north);
\draw [out=20,in=160] (m3.north) to (m5.north);
\draw [out=20,in=160] (m4.north) to (m6.north);
\draw [out=20,in=160] (m5.north) to (sq.north);
\draw [out=20,in=160] (m6.north) to (q.north);
\node [fill=white] (dots) at (3.6,0.25) {\small $\;$ . $\;$ . $\;$ .$\;$ };
\node [fill=white] (dots) at (3.6,-0.1) {\small $\;$ . $\;$ . $\;$ .$\;$ };
\draw (p.north) -- (q.north);
\end{tikzpicture}
\vspace{-5mm}
\caption{\label{fig:bad-structure} Visualization of Locked-Chain\@\xspace structures.
Note, the use of $0$ to $N$ does not imply this must span the entire sentence, these numbers are just for convenience in the definition.
}
\end{figure}
Now we will further characterize the nature of the Lock-Free restriction to the space of graphs.
\begin{lemma} \label{lemma:chain-limit}
No edge in a Locked-Chain\@\xspace in a 1-EC\@\xspace graph is crossed by edges that are not part of it.
\end{lemma}
\begin{proof}
First, note that:
$\mathcal{P}t((0, N{-}1)) = N$,
$\mathcal{P}t((1, N)) = 0$, and
$\{\mathcal{P}t((i, i{+}2)) = i{+}1 \; \forall i \in [0, N{-}2]\}$
Also, call the set $\{(i, i{+}2) \forall i \in [0, N{-}2]\}$, the \emph{chain}.
Consider an edge $e$ that crosses an edge $f$ in a Locked-Chain\@\xspace.
Let $e_{in}$ be the end of $e$ that is between the two ends of $f$, and $e_{out}$ be the other end.
One of $e$'s endpoints is at $\mathcal{P}t(f)$, and $\mathcal{P}t(e)$ is an endpoint of $f$.
There are three cases:
(i) $f = (1, N)$.
Here, $e_{out} = \mathcal{P}t(f) = 0$, and $e_{in} \in (1, N)$.
For all vertices $v \in (1, N)$ there is an edge $g$ in the chain such that $v$ is between the endpoints of $g$.
Therefore, $e$ will cross such an edge $g$.
To satisfy the 1-EC\@\xspace property, $g$ must share an endpoint with $f$, which means $g$ is either $(1, 3)$ or $(N{-}2, N)$.
In the first case, the 1-EC\@\xspace property forces $e = (0, 2)$, and in the second $e = (0, N{-}1)$, both of which are part of the Locked-Chain\@\xspace.
(ii) $f = (0, N{-}1)$, symmetrical with (i).
(iii) $f = (i, i{+}2)$, for some $i \in [0, N{-}2]$.
Here, $e_{in} = \mathcal{P}t(f) = i{+}1$.
We can assume $e$ does not cross $(0, N{-}1)$ or $(1, N)$, as those cases are covered by (i).
As in (i), $e$ must cross another edge in the chain, and that edge must share an endpoint with $f$.
This forces $e$ to be either $(i{-}1, i{+}1)$ or $(i{+}1, i{+}3)$ (excluding one or both if they cross $(0, N{-}1)$ or $(1, N)$), which are both in the Locked-Chain\@\xspace.
\end{proof}
Our rules define a unique way to decompose almost any item into a set of other items.
The exception is $B$, which in some cases can not be divided into two items (i.e.\@\xspace has no valid binary division).
\begin{lemma} \label{lemma:B-case}
A $B[ij.x]$ has no valid binary division if and only if the graph has a Locked-Chain\@\xspace.
\end{lemma}
\begin{proof}
Consider the $k$ and $l$ that give the longest $ik$ and $lj$ edges in a $B$ with no valid binary division (at least one edge of each type must exist by definition).
No vertex in $(ik)$ or $(jl)$ is a valid split point, as they would all require one of the items to have two external vertices.
Now, consider $p \in [kj]$.
If there is no edge $l_1r_1$, where $i \le l_1 < p < r_1 \le j$, then $p$ would be a valid split point.
Therefore, such an edge must exist.
Consider $l_1$, either $l_1 \in (ik)$ or there is an edge $l_2c$, where $i \le l_2 < l_1 < c \le j$ (by the same logic as for $l_1r_1$).
Similarly, either $r_1 \in (jl)$ or there is an edge $cr_2$ (it must be $c$ to satisfy 1-EC\@\xspace).
We can also apply this logic to edges $l_2c$ and $cr_2$, giving edges $l_3l_1$ and $r_1r_3$.
This pattern will terminate when it reaches $l_u \in (ik)$ and $r_v \in (jl)$ with edges $l_ul_{u-2}$ and $r_{v-2}r_v$.
Note that $k=l_{u-1}$ and $l=r_{v-1}$, to satisfy 1-EC\@\xspace.
Since it is a $B$, there must be at least two $x$--$(ij)$ edges.
To satisfy 1-EC\@\xspace, these end at $l_{u-1}$ and $r_{v-1}$.
Let $x$ be to the right (the left is symmetrical), and call $i=0$, $j=N{-}1$, and $x=N$.
Comparing with the Locked-Chain\@\xspace definition, we have all the edges except one: $0$ to $N{-}1$.
However, that edge must be present in the overall graph, as all $B$ items start with an $ij$ edge (see rules 3 and 5 in Algorithm~\ref{alg:rules}).
Therefore, if there is no valid split point for a $B$, the overall graph must contain a Locked-Chain\@\xspace.
Now, for a graph that contains a Locked-Chain\@\xspace, consider the items that contain the Locked-Chain\@\xspace.
Grouping them by their span $[ij]$, there are five valid options: $[0, N{-}1]$, $[1, N]$, $[0, N]$, ($i \leq 0 \land j > N$), and ($i < 0 \land j \geq N$).
Items of the last three types would be divided by our rules into smaller items, one of which contains the whole Locked-Chain\@\xspace.
The first two are $B$s of the type discussed above.
\end{proof}
Now we will prove that our code to generate rules from the templates can guarantee a DAG is formed.
\begin{lemma} \label{lemma:reach}
For any item $H$, $\forall v \in covered(H)$ \linebreak
$\exists u \in visible(H): v$ is reachable from $u$.
\end{lemma}
\begin{proof}
This is true for initial items because $covered(H) = \emptyset$.
To apply induction, consider adding edges and combing items.
The lemma clearly remains true when adding an edge.
Consider combining items $E$, $F$, $G$ to form $H[ij.x]$, and assume the lemma is true for $E$, $F$, and $G$ (the binary case is similar).
Since all vertices are reachable from $visible(E,F,G)$, we only need to ensure that $\forall v \in visible(E,F,G)\; \exists u \in visible(H): v$ is reachable from $u$.
The connectivity between all pairs $\{(u, v) \mid u \in visible(H), v \in visible(E,F,G)\}$ can be inferred from the item definitions, and so this requirement can be enforced in rule generation.
\end{proof}
\begin{lemma} \label{lemma:dag}
The final item is a directed acyclic graph.
\end{lemma}
\begin{proof}
First, consider acyclicity.
Initial items do not contain any edges and so cannot contain a cycle.
For induction, there are two cases:
(i) Adding an Edge $\vec{pq}$ to an item $H$:
Assume that $H$ does not contain any cycles.
$\vec{pq}$ will create a cycle if and only if $p$ is reachable from $q$.
By construction, $p$ and $q \in visible(H)$, and so the item definition contains whether $p$ is reachable from $q$.
(ii) Combining Items:
Assume that in isolation, none of the items being combined contain cycles.
Therefore, a cycle in the combined item must be composed of paths in multiple items.
A path in one item can only continue in another item by passing through a visible\@\xspace vertex.
Therefore, a cycle would have to be formed by a set of paths between visible\@\xspace vertices.
But the connectivity of every pair of visible\@\xspace vertices is specified in the item definitions.
In both cases, rules that create a cycle can be excluded during rule generation.
By induction, the items constructed by our algorithm do not contain cycles.
Together with Lemma~\ref{lemma:reach} and the final item definition, this means the final structure is an acyclic graph with all vertices reachable from vertex $n$.
\end{proof}
Next, we will show two properties that give intuition for the algorithm.
Specifically, we will prove which rules add edges that are crossed in the final derivation.
\begin{lemma} \label{lemma:no-int-cross}
An edge $ij$ added to $I[ij]$ is not crossed.
\end{lemma}
\begin{proof}
First, we will show three properties of any pair of items in a derivation (using $[ij.x]$ and $[kl.y]$).
(1) \textit{It is impossible for either $i < k < j < l$ or $k < i < l < j$}, i.e.\@\xspace items cannot have partially overlapping spans.
As a base case, the final item is an interval spanning all vertices, and so no other item can partially overlap with it.
Now assume it is true for an item $H$ and consider the rules in reverse, breaking $H$ up.
By construction, each rule divides $H$ into items with spans that are adjacent, overlapping only at their visible\@\xspace vertices.
Also, since the new items are nested within $H$, they do not overlap with any items $H$ did not overlap with.
By induction, no pair of items have partially overlapping spans.
(2) \textit{For items with nested spans ($i \le k < l \le j$), $y \in [ij] \cup \{x\}$}.
Following the argument for the previous case, the $[ij.x]$ item must be decomposed into a set of items that includes $[kl.y]$.
Now, consider how those items are combined.
The rules that start with an item with an external vertex produce an item that either has the same external vertex, or with the external vertex inside the span of the new item.
Therefore, $y$ must either be equal to $x$ or inside $[ij]$.
(3) \textit{For items without nested spans, $x \notin (kl)$}.
Assume $x \in (kl)$ for two items without nested spans.
None of the rules combine such a pair of items, or allow one to be extended so that the other is nested within it.
But all items are eventually combined to complete the derivation.
By contradiction, $x \notin (kl)$.
Together, these mean that given an interval $H$ with span $[ij]$, and another item $G$, either $\forall v \in visible(G), v \in [ij]$ or $\forall v \in visible(G), v \notin (ij)$.
Since edges are only created between visible\@\xspace vertices, no edge can cross edge $ij$.
\end{proof}
\begin{lemma} \label{lemma:ext-crossing}
All edges aside from those considered in Lemma~\ref{lemma:no-int-cross} are crossed.
\end{lemma}
\begin{proof}
First, consider an edge $ij$ added to an item $[ij.x]$ of type B, L, R, or N.
This edge is crossed by all $x$--$(ij)$ edges, and in these items $|x$--$(ij)| \ge 1$ by definition.
Note, by the same argument as Lemma~\ref{lemma:no-int-cross}, the edge is not crossed later in the derivation.
Second, consider adding $e \in \{xi, xj\}$, to $H$, an item with $[ij]$ or $[ij.x]$, forming an item $G[ij.x]$.
Note, $e$ does not cross any edges in $H$.
Let $E(F[kl.y])$ be the set of $y$--$[kl]$ edges in some item $F$.
Note that $e \in E(G)$.
We will show how this set of edges is affected by the rules and what that implies for $e$.
Consider each input item $A[kl.y]$ in each rule, with output item $C$.
Every item $A$ falls into one of four categories:
(1) $\forall f \in E(A), f$ is crossed by an edge in another of the rule's input items,
(2) $E(A) \subseteq E(C)$,
(3) $A \land kl \mapsto C$ and there are no $ky$ or $ly$ edges in $A$,
(4) $A$ contains edge $kl$ and there are no $ky$ or $ly$ edges in $A$.
Cases 2-4 are straightforward to identify.
For an example of the first case, consider the rightmost item in rule 4.
The relevant edges are $k$--$(lj]$ (by construction, $kl$ is not present).
Since the leftmost item is either an R or N, $|l$--$(ik)| \ge 1$.
Since $i < k < l < j$, all $k$--$(lj]$ edges will cross all $l$--$[ik)$ edges.
Therefore applying this rule will cross all $k$--$(lj]$ edges in the rightmost item.
Initially, $e$ is not crossed and $e \in E(G)$.
For each rule application, edges in $E(A)$ are either crossed (1 and 3), remain in the set $E(C)$ (2), or must already be crossed (4).
Since the final item is an interval and $E($Interval$) = \emptyset$, there must be a subsequent rule that is not in case 2.
Therefore $e$ will be crossed.
\end{proof}
\subsection{Comparison with \textcite{ec}} \label{sec:ec-comparison}
Our algorithm is based on \textcite{ec}, which had the crucial idea of one-endpoint crossing and a complete decomposition of the tree case.
Our changes and extensions provide several benefits:
\tightparagraph{Extension to graphs}
By extending to support multiple parents while preventing cycles, we substantially expand the space of generatable structures.
\tightparagraph{Uniqueness}
By avoiding derivational ambiguity we reduce the search space and enable efficient summing as well as maxing.
Most of the cases in which ambiguity arises in \textcite{ec}'s algorithm are due to symmetry that is not explicitly broken.
For example, the rule we worked through in the previous section defined $t \in (sq)$ when $|C| = 1$.
Picking $t \in (ps)$ would also lead to a valid set of rules, but allowing either creates a spurious ambiguity.
This ambiguity is resolved by tracking whether there is only one edge to the external vertex or more than one, and requiring more than one in rules 6 and 7.
Other changes include ensuring equivalent structures cannot be represented by multiple item types and enforcing a unique split point in $B$ items.
\tightparagraph{More concise algorithm definition}
By separating edge creation from item merging, and defining our rules via a combination of templates and code, we are able to define our algorithm more concisely.
\subsection{Algorithm Extensions}
\subsubsection{Edge Labels and Word Labels}\label{sec:labels}
Edge labels can be added by calculating either the sum or max over edge types when adding each edge.
Word labels (e.g.\@\xspace POS Tags) must be added to the state, specifying a label for each visible word ($p$, $q$ and $o$).
This state expansion is necessary to ensure agreement when combining items.
\subsubsection{Ensuring a Structural Tree is Present}
Our algorithm constrains the space of graph structures, but we also want to ensure that our parse contains a projective tree of non-trace edges.
To ensure every word gets one and only one structural parent, we add booleans to the state, indicating whether $p$, $q$ and $o$ have structural parents.
When adding edges, a structural edge cannot be added if a word already has a structural parent.
When combining items, no word can receive more than one structural parent, and words that will end up in the middle of the span must have exactly one.
Together, these constraints ensure we have a tree.
To ensure the tree is projective, we need to prevent structural edges from crossing.
Crossing edges are introduced in two ways, and in both we can avoid structural edges crossing by tracking whether there are structural $o$--$[pq]$ edges.
Such edges are present if a rule adds a structural $op$ or $oq$ edge, or if a rule combines an item with structural $o$--$[pq]$ edges and $o$ will still be external in the item formed by the rule.
For adding edges, every time we add a $pq$ edge in the $N$, $L$, $R$ and $B$ items we create a crossing with all $o$--$(pq)$ edges.
We do not create a crossing with $oq$ or $op$, but our ordering of edge creation means these are not present when we add a $pq$ edge, so tracking structural $o$--$[pq]$ edges gives us the information we need to prevent two structural edges crossing.
For combining items, in Lemma~\ref{lemma:ext-crossing} we showed that during combinations, $o$--$[pq]$ edges in each pair of items will cross.
As a result, knowing whether any $o$--$[pq]$ edge is structural is sufficient to determine whether two structural edges will cross.
\subsection{Complexity}
Consider a sentence with $n$ tokens, and let $E$ and $S$ be the number of edge types and word labels in our grammar respectively.
\tightparagraph{Parses without word or edge labels}
Rules have up to four positions, leading to complexity of $O(n^4)$.
Note, there is also an important constant--once our templates are expanded, there are 49,292 rules.
\tightparagraph{With edge labels}
When using a first-order model, edge labels only impact the rules for edge creation, leading to a complexity of $O(n^4 + E n^2)$.
\tightparagraph{With word labels}
Since we need to track word labels in the state, we need to adjust every $n$ by a factor of $S$, leading to $O(S^4 n^4 + E S^2 n^2)$.
\section{Conclusion}
We propose a representation and algorithm that cover $97.3\%$ of graph structures in the PTB\xspace.
Our algorithm is $O(n^4)$, uniquely decomposes parses, and enforces the property that parses are composed of a core tree with additional traces and null elements.
A proof of concept parser shows that our algorithm can be used to parse and recover traces.
\section*{Acknowledgments}
Thank you to Greg Durrett for advice on parser implementation and debugging, and to the anonymous reviewers for their helpful feedback.
This research was partially supported by a General Sir John Monash Fellowship and the Office of Naval Research under MURI Grant No.\@\xspace N000140911081.
\section{Results} \label{sec:impl}
\tightparagraph{Algorithm Coverage}
In Table~\ref{tab:coverage} we show the impact of design decisions for our representation.
The percentages indicate how many sentences in the training set are completely recoverable by our algorithm.
Each row shows the outcome of an addition to the previous row, starting from no traces at all, going to our representation with the head rules of \textcite{cck}, then changing the head rules, reversing null-null edges, and changing the target of edges in parallel constructions.
The largest gain comes from changing the head rules, which is unsurprising since \textcite{cck}'s rules were designed for trees (any set of rules form valid structures for trees).
\tightparagraph{Problematic Structures}
Of the sentences we do not cover, $54\%$ contain a cycle, $45\%$ contain a 1-EC\@\xspace violation, and $1\%$ contain both.
To understand these problematic sentences, we manually inspected a random sample of twenty parses that contained a cycle and twenty parses with a 1-EC\@\xspace violation (these forty are $6\%$ of all problematic parses, enough to identify the key remaining challenges).
For the cycles, eleven cases related to sentences containing variations of NP~\emph{said} interposed between two parts of a single quote.
A cycle was present because the top node of the parse was co-indexed with a null argument of \emph{said} while \emph{said} was an argument of the head word of the quote.
The remaining cases were all instances of pseudo-attachment, which the treebank uses to show that non-adjacent constituents are related \parencite{ptb-guide}.
These cases were split between use of Expletive (5) and Interpret Constituent Here (4) traces.
It was more difficult to determine trends for cases where the parse structure has a 1-EC\@\xspace violation.
The same three cases, Expletive, Interpret Constituent Here, and NP \emph{said} accounted for half of the issues.
\subsection{Implementation}
We implemented a parser with a first-order model using our algorithm and representation.
Code for the parser, for conversion to and from our representation, and for our metrics is available\footnote{
\url{https://github.com/jkkummerfeld/1ec-graph-parser}
}.
Our parser uses a linear discriminative model, with features based on \textcite{McDonald-etal:2005:Proj}.
We train with an online primal subgradient approach \parencite{Ratliff:2007} as described by \textcite{Kummerfeld-etal:2015:EMNLP}, with parallel lock-free sparse updates.
\begin{table}
\small
\centering
\vspace{2mm}
\begin{tabular}{lrr}
\hline
& \multicolumn{2}{c}{Coverage (\%)} \\
Representation & Sentences & Edges \\
\hline
\hline
Projective trees, no nulls & 26.59 & 96.27 \\
Projective trees, with nulls & 43.85 & 96.27 \\
Projective graphs & 50.60 & 96.67 \\
One-EC graphs & 71.84 & 98.31 \\
+ Head rule changes & 92.35 & 99.23 \\
+ Null reversal & 97.02 & 99.45 \\
+ Parallel construction shift & 97.31 & 99.49 \\
\hline
\end{tabular}
\caption{ \label{tab:coverage}
Training set coverage for different representations.
One-EC graphs uses our representation, but with the head rules from \textcite{cck}.
For the edge results, we only exclude edges necessary to make each parse representable (e.g.\@\xspace excluding only one edge in a cycle and counting the rest).
}
\end{table}
\tightparagraph{Loss Function}
We use a weighted Hamming distance for loss-augmented decoding, as it can be efficiently decomposed within our dynamic program.
Calculating the loss for incorrect spines and extra edges is easy.
For missing edges, we add when a deduction rule joins two spans that cover an end of the edge, since if it does not exist in one of those items it is not going to be created in future.
To avoid double counting we subtract when combining two halves that contain the two ends of a gold edge\footnote{
One alternative is to count half of it on each end, removing the need for subtraction later.
Another is to add it during the combination step.
}.
\tightparagraph{Inside--Outside Calculations}
Assigning scores to edges is simple, as they are introduced in a single item in the derivation.
Spines must be introduced in multiple items (left, right, and external positions) and must be assigned a score in every case to avoid ties in beams.
We add the score every time the spine is introduced and then subtract when two items with a spine in common are combined.
\tightparagraph{Algorithm rule pruning}
Many 1-EC\@\xspace structures are not seen in our data.
We keep only the rules used in gold training parses, reducing the set of 49,292 from the general algorithm to 627 (including rules for both adding arcs and combining items).
Almost every template in Algorithm~\ref{alg:rules} generates some unnecessary rules, and no items of type $B$ are needed.
The remaining rules still have high coverage of the development set, missing only 15 rules, each applied once (out of 78,692 rule applications).
By pruning in this way, we are considering the intersection of 1-EC\@\xspace graphs and the true space of structures used in language.
\tightparagraph{Chart Pruning}
To improve speed we use beams and cube pruning \parencite{Chiang:2007}, discarding items based on their Viterbi inside score.
We divide each beam into sub-beams based on aspects of the state.
This ensures diversity and enables consideration of only compatible items during binary and ternary compositions.
\tightparagraph{Coarse to Fine Pruning}
Rather than parsing immediately with the full model we use several passes with progressively richer structure \parencite{Goodman:1997}:
(1) Projective parsing without traces or spines, and simultaneously a trace classifier,
(2) Non-projective parsing without spines, and simultaneously a spine classifier,
(3) Full structure parsing.
Each pass prunes using parse max-marginals and classifier scores, tuned on the development set.
The third pass also prunes spines that are not consistent with any unpruned edge from the second pass.
For the spine classifier we use a bidirectional LSTM tagger, implemented in DyNet \parencite{dynet}.
\tightparagraph{Speed}
Parsing took an average of $8.6$ seconds per sentence for graph parsing and $0.5$ seconds when the parser is restricted to trees\footnote{
Using a single core of an Amazon EC2 m4.2xlarge instance (2.4 GHz Xeon CPU and 32 Gb of RAM).
}.
Our algorithm is also amenable to methods such as semi-supervised and adaptive supertagging, which can improve the speed of a parser after training \parencite{Lewis-Steedman:2014,Kummerfeld-Roesner-Dawborn-Haggerty-Curran-Clark:2010:ACL}.
\tightparagraph{Tree Accuracy}
On the standard tree-metric, we score $88.1$.
Using the same non-gold POS tags as input, \textcite{cck} score $90.9$, probably due to their second-order features and head rules tuned for performance\footnote{
Previous work has shown that the choice of head can significantly impact accuracy \parencite{schwartz-abend-rappoport:2012:PAPERS}.
}.
Shifting to use their head rules, we score $88.9$.
Second-order features could be added to our model through the use of forest reranking, an improvement that would be orthogonal to this paper's contributions.
We can also evaluate on spines and edges.
Since their system produces regular PTB\xspace trees, we convert its output to our representation and compare its results with our system using their head rules.
We see slightly lower accuracy for our system on both spines (94.0 vs.\@\xspace 94.3) and edges (90.4 vs.\@\xspace 91.1).
\begin{table}
\small
\centering
\vspace{2mm}
\begin{tabular}{lrrr}
\hline
System & P & R & F \\
\hline
\hline
\multicolumn{4}{c}{Null Elements Only} \\
\textcite{Johnson:2002} & 85 & 74 & 79 \\
\textcite{hayashi-nagata:2016} & 90.3 & 81.7 & 85.8 \\
\textcite{kato-matsubara:2016} & 88.5 & 82.1 & 85.2 \\
This work & 89.5 & 81.6 & 85.4 \\
\hline
\multicolumn{4}{c}{Null Elements and Co-indexation} \\
\textcite{Johnson:2002} & 73 & 63 & 68 \\
\textcite{kato-matsubara:2016} & 81.2 & 74.7 & 77.8 \\
This work & 74.3 & 67.3 & 70.6 \\
\hline
\end{tabular}
\caption{ \label{tab:accuracy}
Accuracy on section 23 using Johnson's metric.
}
\end{table}
\newcommand{K\&M \parencite*{kato-matsubara:2016}\xspace}{K\&M \parencite*{kato-matsubara:2016}\xspace}
\tightparagraph{Trace Accuracy}
Table~\ref{tab:accuracy} shows results using \textcite{Johnson:2002}'s trace metric.
Our parser is competitive with previous work that has highly-engineered models: Johnson's system has complex non-local features on tree fragments, and similarly \textcite[K\&M][]{kato-matsubara:2016} consider complete items in the stack of their transition-based parser.
On co-indexation our results fall between Johnson and K\&M.
Converting to our representation, our parser has higher precision than K\&M on trace edges ($84.1$ vs.\@\xspace $78.1$) but lower recall ($59.5$ vs.\@\xspace $71.3$).
One modeling challenge we observed is class imbalance: of the many places a trace could be added, only a small number are correct, and so our model tends to be conservative (as shown by the P/R tradeoff).
\section{Introduction}
Many syntactic representations use graphs and/or discontinuous structures, such as traces in Government and Binding theory and f-structure in Lexical Functional Grammar \parencite{gb,Bresnan:1982}.
Sentences in the Penn Treebank \parencite[PTB\xspace,][]{ptb} have a core projective tree structure and trace edges that represent control structures, wh-movement and more.
However, most parsers and the standard evaluation metric ignore these edges and all null elements.
By leaving out parts of the structure, they fail to provide key relations to downstream tasks such as question answering.
While there has been work on capturing some parts of this extra structure, it has generally either been through post-processing on trees
\parencite{Johnson:2002,Jijkoun:2003,Campbell:2004,Levy:2004,Gabbard:2006}
or has only captured a limited set of phenomena via grammar augmentation
\parencite{collins:1997,dienes-dubey:2003,schmid:2006,cai-chiang-goldberg:2011}.
We propose a new general-purpose parsing algorithm that can efficiently search over a wide range of syntactic phenomena.
Our algorithm extends a non-projective tree parsing algorithm \parencite{ec,ec-gp} to graph structures, with improvements to avoid derivational ambiguity while maintaining an $O(n^4)$ runtime.
Our algorithm also includes an optional extension to ensure parses contain a directed projective tree of non-trace edges.
Our algorithm cannot apply directly to constituency parses--it requires lexicalized structures similar to dependency parses.
We extend and improve previous work on lexicalized constituent representations \parencite{cck,Shen:2007,hayashi-nagata:2016} to handle traces.
In this form, traces can create problematic structures such as directed cycles, but we show how careful choice of head rules can minimize such issues.
We implement a proof-of-concept parser, scoring $88.1$ on trees in section 23 and $70.6$ on traces.
Together, our representation and algorithm cover $97.3\%$ of sentences, far above the coverage of projective tree parsers ($43.9\%$).
\section{Background}
This work builds on two areas: non-projective tree parsing, and parsing with null elements.
\textbf{Non-projectivity} is important in syntax for representing many structures, but inference over the space of all non-projective graphs is intractable.
Fortunately, in practice almost all parses are covered by well-defined subsets of this space.
For dependency parsing, recent work has defined algorithms for inference within various subspaces \parencite{Gomez-Rodriguez:2010,ec}.
We build upon these algorithms and adapt them to constituency parsing.
For constituency parsing, a range of formalisms have been developed that are mildly-context sensitive, such as CCG \parencite{Steedman:2000}, LFG \parencite{Bresnan:1982}, and LTAG \parencite{Joshi:1997}.
Concurrently with this work, \textcite{Cao-etal:2017:ACL} also proposed a graph version of \textcite{ec}'s One-Endpoint Crossing (1-EC\@\xspace) algorithm.
However, Cao's algorithm does not consider the direction of edges\footnote{
To produce directed edges, their parser treats the direction as part of the edge label.
} and so it could produce cycles, or graphs with multiple root nodes.
Their algorithm also has spurious ambiguity, with multiple derivations of the same parse structure permitted.
One advantage of their algorithm is that by introducing a new item type it can handle some cases of the Locked-Chain\@\xspace we define below (specifically, when $N$ is even), though in practise they also restrict their algorithm to ignore such cases.
They also show that the class of graphs they generate corresponds to the 1-EC\@\xspace pagenumber-2 space, a property that applies to this work as well\footnote{
This is a topological space with two half-planes sharing a boundary.
All edges are drawn on one of the two half-planes and each half-plane contains no crossings.
}.
\textbf{Parsing with Null Elements} in the PTB\xspace has taken two general approaches.
The first broadly effective system was \textcite{Johnson:2002}, which post-processed the output of a parser, inserting extra elements.
This was effective for some types of structure, such as null complementizers, but had difficulty with long distance dependencies.
The other common approach has been to thread a trace through the tree structure on the non-terminal symbols.
\textcite{collins:1997}'s third model used this approach to recover wh-traces, while \textcite{cai-chiang-goldberg:2011} used it to recover null pronouns, and others have used it for a range of movement types \parencite{dienes-dubey:2003,schmid:2006}.
These approaches have the disadvantage that each additional trace dramatically expands the grammar.
Our representation is similar to LTAG-Spinal \parencite{Shen:2007} but has the advantage that it can be converted back into the PTB\xspace representation.
\textcite{hayashi-nagata:2016} also incorporated null elements into a spinal structure but did not include a representation of co-indexation.
In related work, dependency parsers have been used to assist in constituency parsing, with varying degrees of representation design, but only for trees \parencite{hall2007hybrid,hall-nivre:2008:PaGe,kong-rush-smith:2015:NAACL-HLT,fernandezgonzalez-martins:2015:ACL-IJCNLP}.
\textcite{kato-matsubara:2016} described a new approach, modifying a transition-based parser to recover null elements and traces, with strong results, but using heuristics to determine trace referents.
\section{Parse Representation} \label{sec:representation}
Our algorithm relies on the assumption that we can process the dependents to the left and right of a word independently and then combine the two halves.
This means we need lexicalized structures, which the PTB\xspace does not provide.
We define a new representation in which each non-terminal symbol is associated with a specific word (the head).
Unlike dependency parsing, we retain all the information required to reconstruct the constituency parse.
Our approach is related to \textcite{cck} and \textcite{hayashi-nagata:2016}, with three key differences:
(1) we encode non-terminals explicitly, rather than implicitly through adjunction operations, which can cause ambiguity,
(2) we add representations of null elements and co-indexation,
(3) we modify head rules to avoid problematic structures.
\begin{figure}
\input{figures/representation-definition}
\vspace{-4mm}
\caption{\label{fig:repr2}
Parse representations for graph structures, PTB\xspace (top) and ours (bottom).
}
\end{figure}
Figure~\ref{fig:repr2} shows a comparison of the PTB\xspace representation and ours.
We add lexicalization, assigning each non-terminal to a word.
The only other changes are visual notation, with non-terminals moved to be directly above the words to more clearly show the distinction between \emph{spines} and \emph{edges}.
\tightparagraph{Spines}
Each word is assigned a spine, shown immediately above the word.
A spine is the ordered set of non-terminals that the word is the head of, e.g.\@\xspace S-VP for \emph{like}.
If a symbol occurs more than once in a spine, we use indices to distinguish instances.
\tightparagraph{Edges}
An edge is a link between two words, with a label indicating the symbols it links in the child and parent spines.
In our figures, edge labels are indicated by where edges start and end.
\tightparagraph{Null Elements}
We include each null element in the spine of its parent, unlike \textcite{hayashi-nagata:2016}, who effectively treated null elements as words, assigning them independent spines.
We also considered encoding null elements entirely on edges but found this led to poorer performance.
\begin{figure}
\begin{subfigure}[b]{0.21\textwidth}
\centering
\scalebox{0.55}{\input{figures/picture-null-to-null-flat}}
\caption{\label{fig:null-null}
Null to null
}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.27\textwidth}
\centering
\scalebox{0.55}{\input{figures/picture-gapping-flat}}
\caption{\label{fig:gapping}
Parallel Constructions
}
\end{subfigure}
\vspace{-8mm}
\caption{
Examples of syntactic phenomena.
Only relevant edges and spines are shown.
}
\end{figure}
\tightparagraph{Co-indexation}
The treebank represents movement with index pairs on null elements and non-terminals, e.g.\@\xspace *\textsubscript{1} and NP\textsubscript{1} in Figure~\ref{fig:repr2}.
We represent co-indexation with edges, one per reference, going from the null element to the non-terminal.
There are three special cases of co-indexation:
\noindent
\textbf{(1)}
It is possible for trace edges to have the same start and end points as a non-trace edge.
We restrict this case to allow at most one trace edge.
This decreases edge coverage in the training set by 0.006\%.
\noindent
\textbf{(2)}
In some cases the reference non-terminal only spans a null element, e.g.\@\xspace the WHNP in Figure~\ref{fig:null-null}.
For these we use a reversed edge to avoid creating a cycle.
Figure~\ref{fig:null-null} shows a situation where the trace edge links two positions in the same spine, which we assign with the spine during parsing.
\noindent
\textbf{(3)}
For parallel constructions the treebank co-indexes arguments that fulfill the same roles (Fig.~\ref{fig:gapping}).
These are distinct from the previous cases because neither index is on a null element.
We considered two options: add edges from the repetition to the referent (middle), or add edges from the repetition to the parent of the first occurrence (bottom).
Option two produces fewer non-1-EC\@\xspace structures and explicitly represents all predicates, but only implicitly captures the original structure.
\subsection{Avoiding Adjunction Ambiguity}
Prior work on parsing with spines has used r-adjunction to add additional non-terminals to spines.
This introduces ambiguity, because edges modifying the same spine from different sides may not have a unique order of application.
We resolve this issue by using more articulated spines with the complete set of non-terminals.
We found that $0.045\%$ of spine instances in the development set are not observed in training, though in $70\%$ of those cases an equivalent spine sans null elements is observed in training.
\subsection{Head Rules} \label{sec:rep-head}
\begin{figure}
\begin{subfigure}[b]{0.19\textwidth}
\centering
\scalebox{0.55}{\input{figures/picture-cycle-flat}}
\caption{\label{fig:cycle}
Cycle
}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.27\textwidth}
\centering
\scalebox{0.55}{\input{figures/picture-non1ec-flat}}
\caption{\label{fig:not-1ec}
Not One-Endpoint Crossing
}
\end{subfigure}
\vspace{-3mm}
\caption{Examples of problematic graph structured syntactic phenomena before our head rule changes.}
\end{figure}
To construct the spines, we lexicalize with head rules that consider the type of each non-terminal and its children.
Different heads often represent more syntactic or semantic aspects of the phrase.
For trees, all head rules generate valid structures.
For graphs, head rules influence the creation of two problematic structures:
\tightparagraph{Cycles}
These arise when the head chosen for a phrase is also an argument of another word in the phrase.
Figure~\ref{fig:cycle} shows a cycle between \emph{which} and \emph{proposed}.
We resolve this by changing the head of an SBAR to be an S rather than a Wh-noun phrase.
\tightparagraph{One-Endpoint Crossing Violations}
Figure~\ref{fig:not-1ec} shows an example, with the trace from \emph{CEO} to \emph{Page} crossing two edges with no endpoints in common.
We resolve this case by changing the head for VPs to be a child VP rather than an auxiliary.
|
\section{Introduction}
Obtaining computational models of structures of amorphous graphene is currently gaining traction, as more and more experimental methods to generate amorphous graphene structures become available. Experimentally it has been shown that amorphous graphene can be obtained by prolonged Ga-ion beam irradiation\,\cite{Zhou2010,KotakoskiNano2015}, prolonged electron-beam irradiation\,\cite{Teweldebrhan2009,KotakoskiPRL2011}, chemical vapor deposition\,\cite{Zhao2012, Li2013} and unzipping of amorphous carbon nanotubes\,\cite{Chattopadhyay2014}.
I will introduce a method which is inspired by the differential mutation (DM) algorithm, a special case of the differential evolution (DE) algorithm introduced by Price and Storn\,\cite{Storn1995,Stornbook2005}. A recent review about the DE algorithm can be found in \cite{Das2016}. The DE-algorithm is a genetic algorithm that works on the principle of mutation, cross-over and selection.
To work satisfactorily, the original DE method requires a large population size in order to have sufficient variation in the possible solution space\,\cite{Stornbook2005}. This poses a problem for structural optimization of large atomic systems due to the following reason. A population consists of multiple versions of the atomic system, each representing a different structural arrangement of the atoms.
If the configuration in 3-dimensional space is the focus of the optimization, this would require that $3\times N_p \times N$ variables need to be optimized, where $N_p$ is the number of atoms and $N$ the number of systems.
Since $N$ has to be large enough to include the possible solution space, one can easily see that such an approach might be beyond current computational limits for large systems.
To overcome this problem, I merged the DM algorithm with a thermal selection criterion inspired by simulated annealing (SA). This allows for thermally bounded increases in energy, which can lead to the breaking of bonds. Particularly during the early stages of the optimization, this allows for rearrangements of the atoms corresponding to the escape from unfavorable local minima traps.
Further, to speed up the computation, the cross-over (part of the original DE method) has been omitted as it would lead to unnecessary complications in the optimization procedure originating from the initial random placement of the atoms within the individual systems. In fact, including cross-over would require the atoms to be sorted based on their relative location in space, which then would have to be frequently updated, in order to avoid the creation of energetically unfavorable holes and overlap structures.
Developing hybrid algorithms incorporating DE and SA is relatively new and has so far only be implemented and tested on benchmark functions and circuit design.
For circuit optimization problems, Olen{\v{s}}ek et. al developed a parallel simulated annealing and differential evolution (PSADE) algorithm\,\cite{olensek2011}, which was later modified by combining the algorithm with a population based ranking (DESAPR)\,\cite{olensek2016}. Combining the DE algorithm with SA was also used in the development of
ESADE (enhanced self-adaptive differential evolution)\,\cite{Guo2014} and iSADE\,\cite{Zhao2013}, which combines a hybrid DE with SA and self-adaptive Gaussian immune operation. Tests on benchmark functions showed a very good performance and robustness of the hybrid algorithms.
However, implementation and augmentations of these methods to large scale structural optimization problems have not yet been done. Optimization problems which deal with the structural arrangement of atoms are different than the range of problems tackled with the above mentioned procedures.\\
Zacharias\,\cite{Zacharias1998} developed an algorithm that switches between a pristine genetic algorithm and a pristine simulated annealing procedure and applied it to small silicon clusters.
Applications of genetic algorithms to nanocluster and crystal structure optimization has shown tremendous successes\cite{WoodleyReview}.
Particularly in the field of crystal structure prediction Woodley and Catlow developed a genetic algorithm that only uses the knowledge of the dimensions of the unit cell\,\cite{Woodley1999, Woodley2004_1,Woodley2004_2,Woodley200984}.
Oganov used his USPEX code (USPEX: Universal Structure Predictor: Evolutionary Xtallography), which merges an evolutionary algorithm with ab-initio calculations, to predict crystal structures using supercells\,\cite{Oganov2006,GLASS2006713,Oganov2008}.
Both the methods make use of the repetitive nature of atomic configurations in crystals, as such they can achieve high quality minimum structures by using only a very limited number of atoms which are then optimizated representing unit or supercells.
However, different from crystalline materials, amorphous materials to not exhibit a long range order. Hence assumptions about repeated spatial arrangements of the atoms cannot be used and the number of atoms making up the size of the system has to be large enough in order to avoid finite size effects on the properties of the final structure.
The goal of this paper is to make the DE/DM fit for structural optimization problems of amorphous systems, which require a large number of atoms,
are high dimensional and for which the minima are degenerate originating from the indistinguishability of the atoms.
The proposed method resembles a thermally-driven differential mutation. Using an analytic expression for the potential energy stored in the configurations of the atoms and a confining volume, it is capable to reliably obtain low lying energy minima with just a very limited population size and can be used for multimodal structural optimization problems involving large
amorphous systems in real-space representation.
As will be illustrated on the optimization of amorphous graphene, the advantage of the method is that it provides with multiple distinctly different configurations which are close in energy and have very similar features in the radial distribution function, the occurrence of ring sizes, the bond angle distribution and the coordination number.
The method provides a new approach for the computational discovery of amorphous graphene structures, which can be used in further computational studies. Up-to-date only very few computational methods have been used to this effect. Amorphous graphene structures were obtained by creating Stone-Wales defects\,\cite{Stone1986} using the Wooten-Wearie-Winer\,(WWW) method\,\cite{WWWmethod}.
Starting from pristine graphene structures, Kapko\,\cite{Kapko2010}, Tuan\,\cite{Tuan2012} and Kumar\,\cite{Thorpe2012} used this method to generate two-dimensional structures. Whereas Li and Drabold\,\cite{Li2011,LiandDrabold2013}, and Mortazavi\,\cite{Mortazavi2016} obtained three-dimensional structures. Small three-dimensional amorphous graphene layers have also been obtained by Popescu\,\cite{Popescu2013}.
Structures were also obtained by cooling from high temperatures using molecular dynamics (MD) by Van Hoang\,\cite{VanHoang2015} and Kumar\,\cite{Thorpe2012} and
Holmstr\"om\,\cite{holmstrom2011} used a stochastic quenching method. Relaxation via MD and time-stamped force-bias Monte-Carlo methods were used too to observe healing in disordered amorphous graphene structures\,\cite{Bal2014}.
The article is structured in the following way. In section 2 the method is introduced. In section 3, the algorithm is then applied to generate amorphous graphene structures for systems consisting of 500 atoms and a population size of only six systems. Computational details will be given in this section and the results are compared with those from refs.\,\cite{Tuan2012,Thorpe2012,Li2011,LiandDrabold2013,Mortazavi2016,Popescu2013,VanHoang2015,holmstrom2011}.
\section{Method}
\begin{figure*}[t!]
\begin{minipage}{0.25\textwidth}
\includegraphics[width=1.75in]{figure1}
\caption{\label{fig1} Schematic drawing of the algorithm. The modified differential mutation algorithm is shown in fig.\,(\ref{fig2}) }
\end{minipage}
\hfill
\begin{minipage}{0.725\textwidth}
\includegraphics[width=5.5in]{figure2}
\caption{\label{fig2} Schematic drawing of the modified differential mutation algorithm with embedded thermal selection criteria.}
\end{minipage}
\end{figure*}
Figs.\,(\ref{fig1}) and (\ref{fig2}) show the schematics of the algorithm. At the start of the evolution sequence a set of $N$ systems $\{\zeta_1,\zeta_2,\ldots ,\zeta_N\}$ are initiated. For each system $N_p$ atoms are randomly distributed confined to some specified volume. This has to be done in such a way that the starting configuration of the atoms is different for each system. An initial temperature $T$ and the differential mutation parameters, $C_r$ and $F$, are chosen. Ideally, if known, the initial temperature should be higher than the melting point of the corresponding crystalline system to ensure sufficient mobility at the beginning of the optimization run. $C_r\in[0,1]$ influences the number of site mutations of the individual systems and $F$ determines the weight of the difference vector (see below) in mutations. $F$ and $C_r$ should be chosen such as to achieve a good acceptance/rejection ratio of the mutant trial vectors. \\
After the initialization, the differential mutation routine is started. From the $N$ systems three systems are selected of which one is the target system $\zeta_l$ and the other two $\zeta_m$ and $\zeta_n$ will be used to mutate the target system via weighted difference. This is done in the following way, from each of the two systems $\zeta_m$ and $\zeta_n$ one atom (say $r_i \in \zeta_m$ and $r_j \in \zeta_n$) is chosen randomly.
The difference in location of the atoms (i.e. $\Delta r= r_i-r_j$) is calculated and added, with the difference weight factor $F$, to the position of an atom $r_k$ from the target system $\zeta_l$, i.e. $\tilde{r}_k=r_k+F*\Delta r$. This creates a mutation on the position of atom $r_k$. The procedure is repeated following the rule, that at least the position of one atom, say $r_{\tilde{k}}$, will change. This atom is determined via random selection from all the atoms of $\zeta_k$. For the other $N_p-1$ atoms a probability criterion is used where the parameter $C_r$ determines the probability of mutation. For this purpose, for every atom ($r_k$ with $k \neq \tilde{k}$), a random number is drawn from a uniform distribution in the range zero to one. If this number is less than $C_r$ the location of the atom is mutated according to the mentioned procedure otherwise the original location of this atom is carried into the mutant system.
\begin{figure*}[t!]
\begin{minipage}{0.475\textwidth}
\includegraphics[width=3.25in]{figure3}
\caption{\label{fig3} Illustration of the algorithm at a population of 6 systems (black dots) each consisting of 20 carbon atoms. The magnifications illustrate the target system (outlined in blue) and the two systems used in determining the difference (outlined in yellow). The difference is added to the target system to create mutations on the locations of some of its atoms. }
\end{minipage}
\hfill
\begin{minipage}{0.475\textwidth}
\includegraphics[width=3.25in]{figure4}
\caption{\label{fig4} Illustration of the evolution of the population of 6 system consisting of 20 carbon atoms each. Illustrated are the systems after initialization at $T=5800K$ (black diamonds), after equilibration at the same temperature (blue) and at a lower temperature (red). The magnifications show examples of the individual systems.}
\end{minipage}
\end{figure*}
Once the mutant system $\tilde{\zeta}$ has been obtained its potential energy $\tilde{U}$ is compared to that of the target system $U$. If the difference $\Delta U=\tilde{U}-U$ is smaller or equal than zero, the mutant system is accepted as member of the next generation otherwise the thermal metropolis criterion is used. For this purpose a random number from a uniform random number distribution in the range zero to one (i.e. $rng[0,1]$) is drawn and compared to the Boltzmann factor of the change in energy. If $rng[0,1]<\exp(-\beta \Delta U)$ the mutant is accepted otherwise not, in which case the original target system is carried into the next generation. Here $\beta =1/(k_BT)$ and $k_B$ is the Boltzmann constant. This process is repeated for the remaining $N-1$ systems $\{\zeta_i\}_{i\ne k}$, specifying each of the remaining systems as target system and randomly choosing two more systems from the old generation as difference systems. The resulting population is the new population.\\
The procedure is repeated until equilibration at the specified temperature is reached at which point the temperature is lowered and the process repeated.
Equilibration is reached once the average of the energy of the individual systems does not change any longer significantly. The individual values in energy will still fluctuate within some bounds, indicative of the transitions of the systems between different energy basins. In general the upper bound of the energy fluctuations are higher at higher temperatures than at lower. The optimization routine can be stopped when either sufficient convergence in the location of the atoms of each of the systems is reached or at zero temperature. Here, sufficient convergence means that the location and energies of the single atoms in the systems $\{\zeta_1,\ldots ,\zeta_N\}$ do not change any longer significantly, i.e. the system reached the vicinity of a low energy minimum from which it can no longer escape. In this case other methods such as gradient optimization or zero temperature Monte-Carlo optimization may be used to free the systems of the temperature induced vibrational disorder.
The procedure can be summarized as follows:
\begin{enumerate}[label=(\roman*)~~~~]
\item\label{step1}{$\bullet$ choose parameters $F$, $C_r$ and starting temperature $T$}
\item\label{step2}{$\bullet$ initialize the population systems $\{\zeta_1,\zeta_2,\ldots ,\zeta_N\}$}
\item\label{step3}{{FOR $\ell =1$ to $N$ ~~~ !\# $\ell$ is the target system}
\begin{enumerate}[leftmargin=0.4cm,labelwidth=1.2cm,start=4,label=(\roman*)~~~~~~~~]
\item\label{step4}{$\bullet$ randomly select systems $\zeta_n$ and $\zeta_m$ (with $n\neq m$, $m\neq \ell$ and $n\neq \ell$)}
\item\label{step5}{$\bullet$ draw a random number $k_0$ from $\{1,2,\ldots ,N_p\}$}
\item\label{step6}{{FOR $k=1$ to $N_p$ ~~~ !\# $k$ is an atom of $\zeta_\ell$}}
\item\label{step7}{{~~~IF $k=k_0$ or $rng[0,1]\le C_r$}
\begin{enumerate}[start=8,label=(\roman*)~~~~~~~~~~~~~]
\item\label{step8}{$\bullet$ randomly select atom $r_i$ from system $\zeta_m$ and atom $r_j$ from system $\zeta_n$}
\item\label{step9}{$\bullet$ form the difference vector $\Delta r=r_i-r_j$}
\item\label{step10}{$\bullet$ mutate the location of atom $k$ using the weighted difference vector $\Delta r$, i.e. $\tilde{r}_k=r_k+F*\Delta r$}
\end{enumerate}
{~~~ELSE}
\begin{enumerate}[start=11,label=(\roman*)~~~~~~~~~~~]
\item{$\bullet$ set $\tilde{r}_k=r_k$}
\end{enumerate}
{~~~ENDIF\\
ENDFOR}}
\end{enumerate}
\begin{enumerate}[start=12,label=(\roman*)~~~~~~~~]
\item\label{step12}{$\bullet$ calculate the difference in energy between the mutant and the target system, i.e. $\Delta U=\tilde{U}-U$}
\item\label{step13}{$\bullet$ if $\Delta U\le 0$ or $rng[0,1]< exp[-\beta\Delta U]$ accept the mutant into the next generation, otherwise carry system $\zeta_l$}
\end{enumerate}
{ENDFOR}}
\end{enumerate}
\begin{enumerate}[start=14,label=(\roman*)~~]
\item\label{step15}{$\bullet$ check for equilibration; if equilibration at temperature $T$ is reached continue at next step, otherwise go to
\ref{step3}}
\item\label{step16}{$\bullet$ if $T\le T_o$ or some other convergence criteria is reached exit the program, otherwise lower the temperature and go to \ref{step3}}
\end{enumerate}
Figs.\,(\ref{fig3}) and (\ref{fig4}) illustrate the procedure at the structural optimization of amorphous graphene consisting of 20 atoms using a population size of $N=6$ systems. The systems were initialized by placing atoms at random positions in a confined volume [see insets of Fig.\,(\ref{fig3})]. The starting temperature was chosen as $T=5800K$. In order to illustrate the difference between the systems, fig.\,(\ref{fig3}) shows a plot of the average distance between nearest neighbors versus the average energy per atom at $T=5800K$. The insets illustrate step (viii) to step (x) of the procedure. From each of the two difference systems (yellow insets) two atoms were selected at random and the difference in their positions was added using the weighting factor $F$ to the atoms of the target system (blue inset). This created mutations (red atoms) on the positions of two of the target atoms (blue atoms).\\
Fig.\,(\ref{fig4}) shows plots of the average distance between nearest neighbors versus the average energy per atom for the systems at different stages during the evolution process. The black dots represent the systems directly after their initialization. The blue dots and blue insets show the systems after the initial equilibration, and the red dots represent the configurations of the systems at a much lower temperature. As can be seen, at high temperatures after equilibration the systems sample a wider range of energies than at low temperatures. This is to be expected since the available configuration space is larger at higher temperatures than at lower, allowing for more variety in the mutation. Whereas at low temperatures the systems become more and more trapped in structural arrangement that are similar in energy, but yet represent different configurations.
\section{Results}
Low energy structures of amorphous graphene were obtained using the bond-order potential introduced by Erhart and Albe\,\cite{Erhart2005}.
To speed up the computation the thermally driven differential mutation procedure was performed using a center approximation in the calculation of the effective three-body energies.
This approximation is depicted in fig.\,(\ref{nnapproxfig}). For each displacement of an atom, say atom $i$ (depicted in red), the bond angle potential was calculated taking into account only the angles with atom $i$ at their vertex (shown in red), but not the angles for which atom $i$ defines one of the sides (shown in blue). The pairwise distance dependent interactions were not modified. Since the calculation of three-body interactions is computationally the most costly, this small approximation significantly reduced the computation time. The gradient procedure was performed using the full Erhart and Albe\,\cite{Erhart2005} potential.
\begin{figure}[h!]
\includegraphics[width=1.4in]{figure5}
\caption{\label{nnapproxfig} To speed up the calculation only the angles with atom $i$ (red atom) at the vertex (drawn in red) were used for the calculation of the three-body contributions in the energy.}
\end{figure}
Six systems $\{\zeta_1 ,\ldots ,\zeta_6\}$ were initiated by placing for each of the systems 500 atoms at random positions within the confined volume of a box of dimensions $36.274 \text{\AA}\times 36.274\text{\AA} \times 1.400\text{\AA}$, corresponding to an atom surface density of $0.38 \text{\AA}^{-2}$. In x- and y-direction periodic boundary and minimum image conditions were imposed. In z-direction only periodic boundary conditions were implemented, which forces the atoms to arrange in a single layer sheet. The starting temperature of the system was chosen as $T=5800K$. This temperature value is well above the melting temperature of graphene, allowing easily for structural breaking and reformation of bonds. $C_r$ was set to $0.02$, giving an average of $11$ side mutations per generation for each of the systems. $F$ was kept variable in the range $0.001$ to $0.005$. The values for $C_r$ and $F$ were determined from short trial runs in an effort to achieve a balance between sufficient side mutations per generation, a good acceptance rate of the mutant systems and the capability to escape local traps during the initial stages of the optimization.
The temperature was lowered following an iterative exponential decay procedure $T_n=T_{n-1}\exp [-\gamma (n-1)]$, where $\gamma=0.01$ determines the rate of cooling and $n$ enumerates the cooling steps. After every change in temperature the systems were given enough trials to equilibrate, approximately $100,000$ trial displacements per atom.\\
The thermally driven DM optimization was performed until the systems were cooled to a temperature of $T= 75K$.
The obtained structures were then refined and relaxed to zero temperature using a standard gradient method\cite{Gradient2006}. For this purpose the force acting on the atoms was used to calculate their new positions according to $\vec{r}_i^k=\vec{r}_i^{k-1}+\alpha \vec{F}_i$. Where $\vec{r}_i^0$ are the positions of the atoms after the DM hybrid optimization, $\vec{F}_i$ denotes the
net force acting on atom $i$ and $\alpha$ is a factor determining the step length of the procedure. It was found that a parameter of $\alpha=0.001$ led to a sufficiently fast convergence without enabling the systems to escape their current configurations.
The gradient optimization was performed without restrictions on the z-direction, allowing the surface to freely buckle. Further, the full C-C potential\,\cite{Erhart2005} was used (i.e. without the approximation in the three-body contribution to the potential). The gradient optimization was stopped once the energy of the systems converged.
\begin{figure}[b!]
\includegraphics[width=3.25in]{figure6}
\caption{\label{rings} Occurence of n-ring structures for the different systems.}
\end{figure}
\begin{table}
\caption{Root mean square deviation of the bond angle $\theta_{rms}$ from the mean and second moment $\mu_2$ of the ring distribution of the six systems $\zeta_1,\ldots ,\zeta_6$ compared to those from refs.\cite{Kapko2010,Tuan2012,Thorpe2012,VanHoang2015}.}
\label{table2}
\begin{tabular}{|c|c|c|}
\hline
~ & $\theta_{rms}$ ($^o$) & $\mu_2$\\ \hline\hline
$\zeta_1$ & $5.50$ & $0.44$ \\
$\zeta_2$ & $5.80$ & $0.49$\\
$\zeta_3$ & $5.26$ & $0.41$ \\
$\zeta_4$ & $5.37$ & $0.46$\\
$\zeta_5$ & $5.58$ & $0.48$ \\
$\zeta_6$ & $5.67$ & $0.45$\\ \hline
Kapko\cite{Kapko2010} (WWW) & $16$ & $0.4$ \\
Tuan\cite{Tuan2012} (WWW) &
\begin{tabular}{@{}c@{}} $11.02$ \\ $18.09$ \end{tabular} &
\begin{tabular}{@{}c@{}} $0.47$\\ $0.88$ \end{tabular}\\
Kumar\cite{Thorpe2012}
\begin{tabular}{@{}c@{}} (WWW)\\ (MD) \end{tabular} &
\begin{tabular}{@{}c@{}} ~$9.62-11.95$~~\\ $9.71-12.83$ \end{tabular} &
\begin{tabular}{@{}c@{}} ~$ 0.43-0.67$~~\\ $0.45-0.65$\end{tabular} \\
Van Hoang \cite{VanHoang2015} (MD) &
\begin{tabular}{@{}c@{}} $*$\\ $*$ \end{tabular} &
\begin{tabular}{@{}c@{}} $2.475$\\ $1.919$\end{tabular}\\
\hline
\end{tabular}
\end{table}
\begin{figure*}[ht!]
\includegraphics[width=4.75in]{figure7}
\caption{\label{topview} {Topview of the systems $\zeta_1, \ldots ,\zeta_6$ labeled (a) through (f) respectively.
The bonds of the rings are colored based on the number of
ring members, i.e. 5-ring (blue), 6-ring (gray), 7-ring (green), and 8-, 9- and 4-ring (red).}}
\end{figure*}
\begin{figure*}[ht!]
\begin{minipage}{0.45\textwidth}
\includegraphics[width=3.25in]{figure8_1}
\end{minipage}
\hfill
\begin{minipage}{0.45\textwidth}
\includegraphics[width=3.25in]{figure8_2}
\end{minipage}
\caption{\label{perspective} Perspective view of the structures $\zeta_1$ and $\zeta_2$.}
\end{figure*}
The results can be seen in figs.\,(\ref{rings}) to (\ref{bondangle}) and tab.\,(\ref{table2}-\ref{table1}). Fig.\,(\ref{topview}) shows a top view of the final structures of the systems $\zeta_1,\ldots,\zeta_6$, and fig.\,(\ref{perspective}) shows, as example, the perspective view of the systems $\zeta_1$ and $\zeta_2$. It can be clearly seen that these structures represent different configurations.
The ring distribution can be seen in fig.\,(\ref{rings}), where the occurrence of the rings is plotted versus the ring size for the six systems $\{\zeta_1,\ldots ,\zeta_6\}$.
{ Similar to the structures obtained in refs.\,\cite{Thorpe2012,Tuan2012,Kapko2010,Li2011} 5-, 6- and 7-atom rings were dominating.}
In all the systems about $58-65\%$ of the rings were 6-atom rings followed by 5-atom ($18-24\%$) and 7-atom ($15-17\%$) rings.
Systems $\zeta_2$, $\zeta_4$ and $\zeta_5$ had each six 8-atom rings, giving a relative occurrence of $2\%$. The occurrence of 8-atom rings for the systems $\zeta_1$ and $\zeta_6$ was $0.4\%$ and $0.8\%$ respectively. System $\zeta_3$ did not exhibit an 8-atom ring. Only systems $\zeta_1$, $\zeta_3$, $\zeta_4$ and $\zeta_6$ had each one 9-atom ring, at $0.4\%$ occurrence.
{ Structures with up to 9-membered rings were also observed in studies by Holstr{\"o}m et al.\,\cite{holmstrom2011} and Van Hoang\,\cite{VanHoang2015}.}
A 4-atom ring could only be found for systems $\zeta_5$ and $\zeta_6$, giving an occurance of $0.4\%$. {
In studies\,\cite{Thorpe2012,Tuan2012,Kapko2010,Li2011,holmstrom2011} there were no rings reported with less than 5-members, and only Van Hoang\,\cite{VanHoang2015} reported structures of 3- and 4-membered rings with $0.020\%-0.039\%$ and $0.237\%-0.393\%$ occurance respectively. }
In all the systems the mean of the number of ring members was close to 6. Further, the variance in the ring-size distribution was between $0.41$ and $0.49$ (see tab.\,\ref{table2}).
{ This is a smaller range as reported by Tuan et al.\,\cite{Tuan2012} and Kumar et al.\,\cite{Thorpe2012}, and below the values reported by Van Hoang\,\cite{VanHoang2015}}.
{ The top view of the structures of systems $\zeta_1,\ldots ,\zeta_6$ can be seen in fig.\,(\ref{topview}).}
The different sizes are indicated by colors. In all the systems the bonds of the 6-ring structures, which formed large connected structures, are colored gray. 5-atom rings are colored blue, 7-atom rings green and 4-, 8- and 9-atom rings red.
\begin{table}
\caption{Distribution of the coordination number of the first coordiation shell of the obtained systems $\zeta_1,\ldots ,\zeta_6$, and values reported using the WWW-method, MD-cooling and stochastic quench.{\color{black} The mean coordination number of the first coordination shell is given by $\langle n_1\rangle$.}}
\label{table3}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
~ & 1 & 2 & 3 &4 &5 & 6 & \color{black}$\langle n_1\rangle$ \\ \hline\hline
$\zeta_1$ & $0$ & $0.002$ & $0.988$ & $0.010$ & $0$ & $0$ & \color{black} $3.008$\\
$\zeta_2$ & $0$ & $0.002$ & $0.988$ & $0.010$ & $0$ & $0$ & \color{black}$3.008$ \\
$\zeta_3$ & $0$ & $0.002$ & $0.988$ & $0.010$ & $0$ & $0$ & \color{black}$3.008$ \\
$\zeta_4$ & $0$ & $0.010$ & $0.988$ & $0.002$ & $0$ & $0$ & \color{black}$2.992$\\
$\zeta_5$ & $0$ & $0.008$ & $0.984$ & $0.008$ & $0$ & $0$ & \color{black}$3.000$ \\
$\zeta_6$ & $0$ & $0.004$ & $0.988$ & $0.008$ & $0$ & $0$ & \color{black}$3.004$\\ \hline
{\bf WWW-meth.} &~&~&~&~&~&~&~\\
\begin{tabular}{@{}c@{}c@{}c@{}}Kapko\cite{Kapko2010}\\ Tuan\cite{Tuan2012}\\ Kumar\cite{Thorpe2012}\\ Popescu\cite{Popescu2013} \end{tabular} & $0$ & $0$ & $1.000$ & $0$ & $0$ & $0$ & $3$ \\
{\bf MD-cooling} &~&~&~&~&~&~&~\\
Kumar \cite{Thorpe2012} & $0$ & $0.006$ & $0.991$ & $0.003$ & $0$ & $0$ & $2.997$\footnote{value calculated from distribution of coordination number}\\
Van Hoang \cite{VanHoang2015} & \begin{tabular}{@{}c@{}}$0.004$\\ $0.002$\end{tabular}&
\begin{tabular}{@{}c@{}} $0.083$ \\ $0.065$ \end{tabular} &
\begin{tabular}{@{}c@{}} $0.457$ \\ $0.689$ \end{tabular} &
\begin{tabular}{@{}c@{}} $0.356$ \\ $0.208$ \end{tabular} &
\begin{tabular}{@{}c@{}} $0.091$ \\ $0.033$ \end{tabular} &
\begin{tabular}{@{}c@{}} $0.009$ \\ $0.004$ \end{tabular} &
\begin{tabular}{@{}c@{}} $3.474^{\it a}$ \\ $3.220^{\it a}$ \end{tabular} \\
{\bf stoch.\,quench} &~&~&~&~&~&~&~\\
Holmstr{\"o}m\cite{holmstrom2011} & $0.005$ & $0.010$ & $0.985$ & $0$ & $0$ & $0$ & $2.980^a$ \\ \hline
\end{tabular}
\end{table}
Fig.\,(\ref{radial}) shows a plot of the radial distribution function $g(r)$ of the systems $\{\zeta_1 ,\ldots ,\zeta_6\}$ (lines). The bars show the radial distribution function of a hexagonal graphene structure with a bond length of $1.45$\AA. The inset shows a magnification of the range from $4$ to $7$\AA . For the calculation of $g(r)$ the length of the intervals $\Delta r$ was set to $0.1$\AA \, for $0<r<3$\AA \, and $0.2$\AA \, for $r>3$\AA.
\begin{figure}[h!]
\includegraphics[width=3.25in]{figure9}
\caption{\label{radial} The radial distribution function of the different systems from 0 to 10 \AA. The inset shows a magnification of the range $4.5$ to $8$ \AA.}
\end{figure}
With some small spread, the distribution of the radial distance between the atoms of the systems $\zeta_1,\ldots,\zeta_6$ clearly resembled the first four peaks (up to $4$\AA) in the radial distribution function of the graphene structure. From $4$ to $7$\AA \, only some of the systems showed, albeit relatively weak, the peaks of graphene. Which may indicate a weak medium range order. From $7$ to $10$\AA \, peaks were no longer distinctly visible indicative of the disordered nature of the system. Overall the six systems exhibited very similar radial distribution functions, { which are also qualitatively similar to the distributions reported in refs.\,\cite{Thorpe2012,holmstrom2011,Kapko2010,Li2011}.}
\begin{figure}[h!]
\includegraphics[width=3.25in]{figure10}
\caption{\label{bondangle} The distribution of bond angles of the different systems.}
\end{figure}
Fig.\,(\ref{bondangle}) shows the normalized distribution of bond angles of the different systems. The length of an interval was $\Delta \theta =1^o$. As can be seen, the most common angle was around $120.0^o$ corresponding to the internal angles of hexagon structures. A distinct second peak formed at $108.5^o$, corresponding to the pentagon structures [compare to fig.\,(\ref{rings})]. The peak for the heptagon angle (expected to be at $128.6^o$) was not distinct. Due to their larger size, 7-atom rings have more flexibility to bend, leading to a wider spread in the distribution of internal angles. This together with the close proximity to the hexagon angles explains why an expected peak of hexagon angles appears to be smoothed out.
The root mean square deviation of the bond angles of the systems from the mean can be seen in tab.\, (\ref{table1}). For the six systems $\{\zeta_1 ,\ldots,\zeta_6\}$ it was found to be in the range between $5.2^o$ and $5.8^o$. { These values are significantly smaller than the $\theta_{rms}$ values reported for the 2-dimensional structures in refs.\,\cite{Kapko2010,Tuan2012,Thorpe2012}. Peaks at $60^o$ and $90^o$, as reported by Holmstr{\"o}m et al.\,\cite{holmstrom2011} were not observed. However, for his 3-dimensional structure Holmstr{\"o}m also noted the highest probability to be around the $120^o$ angle, similar to the present study [see fig.(\ref{bondangle})]. }\\
The coordination numbers of the first coordination shell (counting bonds up to a radius of $2.0$\AA\,) were very close to 3.
{ They were dominated by 3 bonds per atom [see tab.\,(\ref{table2})], which can be explained by the preference of the carbon atoms in single layer sheets to bond to three neighbors. The percentage of atoms to form 2 or 4 bonds was very low, i.e. combined $\leq 1.6\%$. This is in excellent agreement with refs.\,\cite{Kapko2010,Tuan2012,Thorpe2012,Popescu2013}. It is to note that the WWW-method is essentialy a bond-rotation algorithm and hence confines the number of bonds per atom to exactly 3. Only the systems by Van Hoang\,\cite{VanHoang2015} and Holmstr\"om\,\cite{holmstrom2011} showed single coordinated atoms, and Van Hoangs\,\cite{VanHoang2015} distributions were generally wider spread.} \\
The average energies per atom $U/N_p$ of the systems were very similar, with a mean of $-7.073 eV$ and a standard deviation of $\sigma_U=0.013 eV$. See tab.\,(\ref{table1}), $\zeta_4$ had the lowest energy of $-7.089 eV$ and $\zeta_2$ the highest of $-7.050 eV$. \\
\begin{table}[t!]
\caption{Table of the average energy per atom $U/N_p$, the height $\Delta z$ and root mean square height $rms(z)$.}
\label{table1}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
~ & $\zeta_1$ & $\zeta_2$ & $\zeta_3$ & $\zeta_4$ & $\zeta_5$ & $\zeta_6$ \\ \hline\hline
$U/N_p$ (eV) & $-7.067$ & $-7.050$ & $-7.086$ & $-7.089$ & $-7.070$ & $-7.073$ \\
$\Delta z$ (\AA) & $3.75$ & $4.64$ & $4.16$ & $5.30$ & $4.40$ & $4.07$ \\
$rms(z)$ (\AA) & $0.81$ & $0.90$ & $0.79$ & $1.14$ & $0.78$ & $0.84$\\
\hline
\end{tabular}
\end{table}
The physical width of the obtained structures was calculated with respect to a plane determined by the location of all the atoms. The smallest width was found for $\zeta_1$ as $\Delta z=3.75$\AA\, and the largest for $\zeta_4$ as $\Delta z=5.30$\AA . The root mean square value of the width was found to be between $0.78$\AA\, ($\zeta_5$) and $1.14$\AA\, ($\zeta_4$).
Puckering in single layer amorphous graphene sheets has also been observed in refs.~\cite{Li2011,Popescu2013,Mortazavi2016,LiandDrabold2013,holmstrom2011}.
The systems in refs.\,\cite{Kapko2010,Tuan2012,Thorpe2012} were confined to two dimensions, which prevented the structures to pucker.
\\
Overall, while representing different structural arrangements of the atoms, the systems obtained with the described optimization procedure showed very similar properties.
\section{Conclusion}
A temperature driven differential mutation method was introduced.
The method combines a differential mutation algorithm with a thermal selection criteria
for real-space optimization of amorphous systems using an anlytical expression for the energy stored in the atomic configurations.
The inclusion of the temperature in the selection criteria allows for the breaking of existing bonds in order to rearrange the atoms and form new bonds. This corresponds to a climb over energy barriers, where the systems temporarily are allowed to take energetically less favorable configurations. This in turn allows the systems to escape local minima.
The method is capable to deal with structural optimization of amorphous materials using only a small population size consisting of the different systems. It reliably obtains low energy structures corresponding to very low minima on the energy landscape.\\
The method was tested on the structural optimization of amorphous graphene, starting from unbiased randomly selected locations of the atoms. The population consisted of 6 systems $\{\zeta_1,\ldots ,\zeta_6\}$, each with $500$ atoms.
The results showed that, while being microscopically distinct different structures, the overall properties of the obtained systems were very similar. No distinct differences between $\zeta_1,\zeta_2,\ldots$ and $\zeta_6$ could be observed in the radial distribution function, the distribution of bond angles, average energy per atoms and coordination number. Some minor difference in the distribution of rings could be observed. However, it is to be expected that if the size of the systems were to increase these differences might also vanish.
|
\section{Introduction}
Proper configuration of software is a particularly challenging issue in both research and industry. Interactions between design decisions have effects on performance and functionality that are difficult to predict. The observation that automated algorithm configuration and parameter tuning tools can simplify this task has led to a new software development paradigm: \textit{Programming by Optimization}~(PbO)~\cite{DBLP:journals/cacm/Hoos12}. Development in the PbO paradigm consists of specifying large design spaces of program component implementations: the onerous task of determining which components work best in a given use case is achieved via automated analysis tools and optimization heuristics.
The standard Programming by Optimization tools operate on design spaces specified in a specialized extension of a target programming language, transformed into the target language by a specialized \emph{weaver} tool. The optimization choices over the combined design space are made by a separate algorithm configuration tool, which has historically been applied to the resulting executable program.
The weaver-based architecture severely limits the applicability of Programming by Optimization. The reliance on markup extensions hinders the adoption of PbO for existing code bases. In addition, the external configuration tools have to operate on the executable via a brittle textual interface (command line arguments), which introduces significant overhead and makes on-line optimization difficult. Despite these shortcomings, no new alternative to weaver tools has been introduced since the initial PbO proposal.
We developed \textsc{ContainAnt}, a software library for Programming by Optimization that addresses these limitations by replacing syntactic extensions and weavers with the Dependency Injection architectural pattern \cite{prassana:dep}. By exploiting a structural correspondence between Dependency Injection and context-free grammars, we obtain a new class of grammar-based evolutionary heuristics suitable for automated algorithm configuration. We determined that these new heuristics significantly outperform existing configuration algorithms on several common configuration tasks and optimization problems, both in terms of solution quality and execution speed (in one case reducing the optimization time from four~hours to 46~seconds).
This paper discusses the theory and implementation of the \textsc{ContainAnt} library and its grammar-based heuristics. Sections~\ref{ssec:di}~and~\ref{ssec:pbo} introduce Dependency Injection and review the existing work on Programming by Optimization. Section~\ref{ssec:bnf} describes the theoretical correspondence between Dependency Injection and optimization problems over context-free grammars, while Section~\ref{ssec:heur} gives novel heuristics for solving the resulting grammatical optimization problems using genetic algorithms and ant colony techniques. The remainder of the paper analyzes five different experiments used to evalute the performance of the \textsc{ContainAnt} heuristics.
\subsection{Dependency Injection} \label{ssec:di}
Object-oriented software provides functionality via multiple interdependent components. Software engineering efforts to handle the problems of dependency instantiation and reference acquisition between these components has led to the widespread adoption of a new type of middleware library, the so-called Dependency Injection (DI) container \cite{prassana:dep}. The term ``Dependency Injection'' was coined by Fowler \cite{FowlerIoC} in 2004 and DI containers have seen increasingly widespread use over the last decade, with popular frameworks including the Java$^\mathtt{TM}$ Spring framework\footnote{\url{http://projects.spring.io/spring-framework} } and Google Guice\footnote{\url{https://github.com/google/guice} }.
Software written using DI inherently exposes highly structured configuration parameters: components are configured by searching over the space of dependencies, without modifying the source code of the components themselves. The traditional operation of a DI container is to perform the wiring between the constructors of dependent objects (also known as the `object graph') by consulting a configuration object or file that contains a list of bindings between abstract types and their constructor arguments. The container then selects a target class and greedily supplies the dependencies to a suitable constructor of the target class. At this point, it is worth noting a significant limitation of some popular DI containers (e.g.\ Guice): configuration is not possible if the object graph contains ambiguities such as a choice of multiple subtypes of an abstract class. As described in detail in Section
\ref{ssec:heur}, the optimization based approach of \textsc{ContainAnt} removes this limitation.
\subsection{Related Work}
\label{ssec:pbo}
In their seminal work on Programming~by~Optimization, Hoos \textit{et al.}~\cite{DBLP:journals/cacm/Hoos12} delineated five levels of PbO,
ranging in sophistication from tuning the exposed parameters of an application (Level~$0$) to the use of evidence-based methods for exploring large design spaces as the driving activity for the software design process (Level~$4$).
To realize the higher levels of PbO, they introduced the concept of a PbO-enhanced language: a superset of an existing programming language (e.g.\ PbO-\textsc{Java} or PbO-\textsc{C}) which includes constructs for declaring the possible design choices for parameters and blocks of code. The code written in this markup language is translated into the target language via a syntactic transformation performed by a specialized \emph{PbO weaver} tool, reminiscent of a macro preprocessor.
The optimized choices (as determined over the combined design spaces on a set of training cases) are made by an external automatic configuration tool. Configuration optimizers have been proposed that use various heuristics, e.g.\ iterated local search \cite{Hutter:2009:PAA:1734953.1734959}, genetic algorithms \cite{Ansotegui:2009:GGA:1788994.1789011} and iterated racing \cite{irace}.
A notable achievement of PbO is the development and use of the \textsc{SMAC} configuration tool \cite{HutHooLey11-smac} to improve upon the state-of-the-art in SAT solving by tuning parameters of the \textsc{Spear} SAT solver. While the majority of configuration optimizers are model-free, \textsc{SMAC} alternates between building a regression model to predict configuration performance and gathering additional performance data based on this model. The regression model is obtained via \emph{random forests}, a method which is known to perform well on categorical variables and also allows quantification of uncertainty.
Search Based Software Engineering (SBSE) \cite{Harman:2012:SSE:2379776.2379787} is the application of heuristic search to various aspects of the software development process, with a strong historical emphasis on software testing. Much recent interest within SBSE has focused on `embedded adaptivity' \cite{Harman:2014:GIA:2593929.2600116}, i.e.\ allowing software developers to delegate the configuration/generation of specified aspects of program functionality to heuristic search procedures \cite{Burles:2015:EDI:2739482.2768423}. Such SBSE activity is therefore strongly aligned with the previously stated goals of PbO, but often with emphasis on a generation process which can respond dynamically to changes in the operating environment of the program. Previous work in this area includes \textsc{Gen-O-Fix} \cite{GenofixTR} and ECSELR \cite{Yeboah-Antwi:2015:EAS:2739482.2768425},
both of which are embedded monitor systems that support search via Evolutionary Computation. \textsc{Templar} and \textsc{Polytope} are two alternative approaches to software component generation: \textsc{Templar} \cite{TemplarSwan2015} provides a `top-down' framework for orchestrating one or more `variation points' generated by Genetic Programming, while \textsc{Polytope} \cite{Swan:2017:evoApplications} uses methods from datatype generic programming to support the `bottom up' generation of individual variation points in source code.
Since Dependency Injection containers already automate a non-trivial part of the Software Engineering process, they provide a natural entry point for the application of heuristic methods from SBSE.
\section{Grammatical Optimization} \label{ssec:bnf}
Backus-Naur Form (BNF) is a widely adopted syntax for describing context-free languages. For the sake of technical convenience (the ability to have different rewrite rules with identical bodies), we present a slight variation of the usual notion, the labeled~BNF~formalism introduced by Forsberg~and~Ranta~\cite{forsberg:lbnf}. Formally, such a grammar $\mathcal{G}$ consists of the following components:
\begin{itemize}
\item A set of \textbf{terminal symbols}~$\mathcal{G}_T$. These are the literals or words that make up the language.
\item A pointed set of \textbf{non-terminal symbols}~$\mathcal{G}_N$, with a distinguished \textbf{start symbol} $\mathtt{s} \in \mathcal{G}_N$. These categorize the sub-expressions of the language.
\item A set of \textbf{rewrite rules}~$\mathcal{G}_R$. Normally, each rewrite rule has the form $(a,\overline{b})$ where $a \in \mathcal{G}_N$ and $\overline{b}$ is a sequence of symbols from $\mathcal{G}_T \cup \mathcal{G}_N$. Since we are dealing with labeled BNF, rewrite rules have the form $(\ell, a,\overline{b})$ where $\ell$ is a unique \textbf{label}, the \textbf{left-hand side}~$a$ is a non-terminal and the \textbf{right-hand side}~$\overline{b}$ is a finite sequence of symbols from $\mathcal{G}_T \cup \mathcal{G}_N$.
\end{itemize}
At this point, it is customary to introduce the notion of sentence: a sequence of terminal symbols obtained from the start~symbol~$s$ by applying a sequence of rewrite rules, i.e.\ by replacing non-terminals with the right-hand sides of the corresponding rewrite rules.
Such a sequence of rules can be represented as a rooted tree known as a \emph{derivation tree}. A grammar is unambiguous if each of its sentences has a unique corresponding derivation tree.
In practice, the actual sentences of the language turn out to be immaterial from the perspective of a grammatical optimization problem, so it is simpler to work from a direct definition of derivation trees. Thus we ignore the underlying sentences altogether and inductively define a derivation tree of sort $x \in \mathcal{G}_N$ to consist of the following data:
\begin{itemize}
\item A rewrite rule of the form $(\ell, x, \overline{b})$,
\item A derivation tree of sort $a \in \mathcal{G}_N$ for every non-terminal symbol~$a$ in the sequence~$\overline{b}$.
\end{itemize}
From here on, all derivation trees are assumed to have sort~$\mathtt{s}$~(the start symbol of the grammar $\mathcal{G}$). The set of all such derivation trees is denoted $D(\mathcal{G})$.
Grammars can be specified by listing their rewrite rules in the following format:
\begin{verbatim}
Label. <LHS> ::= RHS
\end{verbatim}
where angled brackets are used to distinguish between terminals and non-terminals.
\noindent
Two elementary examples follow:
\subsubsection{Binary Strings}
The grammar of binary strings is given by:
\begin{align*}
G_T &= \left\{\mathtt{0}, \mathtt{1}, \mathtt{e} \right\}, \\
G_N &= \left\{\mathtt{s} \right\}, \\
G_R &= \left\{ (\mathtt{0},\mathtt{s},\mathtt{0s}), (\mathtt{1},\mathtt{s},\mathtt{1s}), (\mathtt{e},\mathtt{s},\mathtt{e}) \right\}.
\end{align*}
Using the shorthand defined above, the rewrite rules could also be written as
\begin{verbatim}
0. <s> ::= 0 <s>
1. <s> ::= 1 <s>
e. <s> ::= e
\end{verbatim}
Derivation trees for this grammar correspond to finite sequences of binary digits (with \texttt{e} being the terminating character). A grammar of strings over any given finite alphabet can be defined analogously.
\subsubsection{Finite Sets}
Any finite set~$S$ gives rise to a grammar by~setting
\begin{align*}
G_T &= S, \\
G_N &= \left\{\mathtt{s} \right\}, \\
G_R &= \left\{ (x,\mathtt{s},x)\:|\: x \in S \right\}.
\end{align*}
The derivation trees of this grammar are in bijective correspondence with elements of the set~$S$.
\subsection{Problem Statement}
An instance of the \textbf{grammatical optimization problem} is given by the following data:
\begin{itemize}
\item A grammar $\mathcal{G}$ and
\item An \textbf{objective~function} $f: D(\mathcal{G}) \rightarrow \mathbb{R}$ defined on the derivation trees of the grammar $\mathcal{G}$.
\end{itemize}
Without loss of generality, we assume that our goal is \textit{maximizing} the objective function, i.e.~solving the grammatical optimization problem consists of finding a globally optimal derivation tree:
$$x^* = \mathrm{arg} \mkern-18mu \max_{x \in D(\mathcal{G})\ \ \ \ \ } \mkern-18mu f(x).$$
The definition above is extremely general: indeed, \textit{every} discrete optimization problem can be reduced to the grammatical optimization problem over the grammar of binary strings.
\subsection{Semantics}
One can reduce an optimization problem instance with candidate solutions~$S$ and objective function~$f: S \rightarrow \mathbb{R}$ to an instance of the grammatical optimization problem by giving an encoding grammar $\mathcal{G}$ and a surjective function
$$k: D(\mathcal{G}) \mathrel{\mathrlap{\rightarrow}\mkern1mu\rightarrow} S$$
with surjectivity ensuring that every candidate solution is described by at least one sentence of the language.
To forbid \emph{ad-hoc} encodings (e.g.\ the~encoding of any discrete optimization problem into the grammar of binary strings discussed above), one should think of the function~$k$ as giving a \textit{semantics} to the sentences of the language defined by the grammar~$\mathcal{G}$. From here on, we demand that the semantics be compositional: the meaning of a derivation tree should be given in terms of the meanings of its parts (direct subtrees). The compositionality requirement provides a formal counterpart to the intuitive desideratum that the structure of the grammar be related to the structure of the search space~$S$, without ruling out any interesting grammatical representations.
We will shortly see that both dependency injection and the algorithm configuration problem have sensible, compositional representations as instances of the grammatical optimization problem. What's more, the same holds for many problems of interest in both continuous and combinatorial optimization.
\subsection{Rosetta Stone}
Analyzing the process of dependency injection leads to a powerful ``dictionary'' correlating the terminology of grammars with the terminology of object-oriented programming.
If the goal is to instantiate an object of some given class $C$, one first has to find a constructor of $C$ (if the class has no constructors, instantiation is impossible). In turn, the selected constructor will expose zero or more classes as dependencies. If the selected constructor $c()$ has no dependencies, the object can be instantiated directly by calling $c()$. However, if there are one or more dependencies $D_1, D_2, \dots$, one has to recursively instantiate objects $d_1,d_2,\dots$ compatible with the given classes before calling $c(d_1,d_2,\dots)$ to instantiate an object of class~$C$.
Now, let $\mathcal{G}$ be a grammar. To construct a derivation tree of some given sort $\mathtt{s} \in \mathcal{G}_N$, one starts by choosing a rewrite rule with left-hand side $\mathtt{s}$ (no suitable tree can exist in the absence of such a rule). If the right-hand side of the chosen rewrite rule contains no non-terminals, the construction is finished. However, if the right-hand side contains one or more non-terminals $n_1,n_2,\dots \in \mathcal{G}_N$, one has to recursively construct a derivation tree for each sort $n_i$ before constructing the derivation tree for the target sort $\mathtt{s}$.
The structure of the algorithms for dependency injection and derivation tree construction (Algorithms~\ref{alg:rosetta-object}~and~\ref{alg:rosetta-grammar}) turn out to be nigh-identical. This suggests an analogy between dependency injection and grammatical optimization, with classes corresponding to non-terminals, constructors corresponding to rewrite rules and constants corresponding to terminals. Thus, the grammatical rewrite rule corresponding to the constructor (\textsc{Java} syntax)
\begin{align*}
\mathrm{T}\ ctor(\mathrm{T}1\ a1,\ \mathrm{T}2\ a2,\dots)
\end{align*}
under this assignment is simply
\begin{verbatim}
ctor. <T> ::= ctor <T1> <T2> [...]
\end{verbatim}
With this correspondence in mind, we can now recast dependency injection as a grammatical decision/optimization problem.
Given a grammar $\mathcal{G}$, deciding whether $D(\mathcal{G})=\emptyset$ amounts to solving a dependency injection problem. The correspondence gives rise to a semantics assigning the constructed object to each derivation tree of the grammar. In the sequel, this is referred to as the \textit{usual semantics}.
\begin{algorithm}
\caption{Class instantiation using Dependency Injection}
\lstset{language=Pascal,
basicstyle=\footnotesize
}
\begin{lstlisting}
function instantiate(t: Class)
for c in t.constructors
{try to construct each argument recursively}
for i := 0 to c.arguments.length
args[i] := instantiate(classOf(a))
end for
{if recursive calls succeed}
if !args.contains(null)
return c(args) {call constructor}
end if
{else try the next constructor}
end for
return null
end function
\end{lstlisting}
\label{alg:rosetta-object}
\end{algorithm}
\begin{algorithm}
\caption{Recursive Derivation Tree Construction}
\lstset{language=Pascal,
basicstyle=\footnotesize
}
\begin{lstlisting}
function construct(t: Sort)
for c in t.rewriteRules
{try to construct each subtree recursively}
for i := 0 to c.nonterminals.count
subtrees[i] := construct(sortOf(i))
end for
{if recursive calls succeed}
if !subtrees.contains(null)
return Tree(c,subtrees)
end if
{else try the next rewrite rule}
end for
return null
end function
\end{lstlisting}
\label{alg:rosetta-grammar}
\end{algorithm}
\section{Heuristics} \label{ssec:heur}
\textsc{ContainAnt} is, first and foremost, a Dependency Injection library.
In order to be as widely applicable as Programming~by~Optimization, the default heuristics of \textsc{ContainAnt}~cannot be problem-specific: they have to operate at the level of problem descriptions. Metaheuristics that only exist as nature-inspired metaphors or informal algorithm templates (i.e.\ without the ability to automatically transform a problem specification into a working implementation) are insufficient for this. These requirements leave us with a rather small class of suitable metaheuristics, which we now describe.
\subsection{Genetic Programming: \textsc{GrEvo}}
Incorporating context-free grammars into genetic programming was proposed by Ryan~\textit{et al.}~\cite{ryan:grevo}. Their seminal work on~\textit{Grammatical~Evolution} allowed the elimination of the closure requirement, a major drawback of untyped Genetic Programming, which required all functions to be able to accept as input the outputs of all other functions. The genotypes are numerical sequences, translated into sentences of a BNF grammar using the mapping of Algorithm~\ref{alg:grevo-geno}. Transcribed into the derivation tree formalism of Section~\ref{ssec:bnf}, the genotypes encode the choice of rewrite rule at each recursive step of the derivation tree construction~(Algorithm~\ref{alg:rosetta-object}).
\begin{algorithm}
\caption{\textsc{GrEvo} Genotype-Phenotype Mapping}
\lstset{language=Pascal,
basicstyle=\footnotesize
}
\begin{lstlisting}
function getPhenotype(g: List[Int], t: Sort)
{choose rewrite rule based on genotype}
c := t.rewriteRules[g.head]
g := g.tail
for i := 0 to c.nonterminals.count
subtrees[i] := getPhenotype(g, sortOf(i))
end for
return Tree(c,subtrees)
end function
\end{lstlisting}
\label{alg:grevo-geno}
\end{algorithm}
Since Grammatical~Evolution allows the generation of syntactically correct sentences in an arbitrary language, its implementations are not tied to any specific problem, and are able to operate on any formal grammar specification. Grammatical~Evolution remains the most popular metaheuristic of its kind, generally outperforming derivate algorithms such as Grammatical Swarm~\cite{oneill:swarm}.
The \textsc{ContainAnt} distribution includes an implementation of the Grammatical Evolution metaheuristic with fixed-length genotypes for solving the grammatical optimization problem. This implementation is henceforth called~\textsc{GrEvo}. The performance analysis~(Section~\ref{ssec:anal}) shows that some characteristics of \textsc{GrEvo}, such as its premature convergence and poor locality, make it suboptimal for tackling the grammatical optimization problem. This limitation motivates the novel grammar-based heuristic introduced below.
\subsection{Ant Programming: \textsc{GrAnt}}
Ant colony optimization methods are the main alternative to Genetic Programming for the automated production of computer programs via stochastic search. Ant Programming based on BNF grammars has been investigated by Keber~and~Schuster~\cite{keber:gap} under the name Generalized Ant Programming (GAP) in the context of option pricing, and later by Salehi-Abari~and~White~\cite{salehi:egap} for general automatic programming~(EGAP). The development of these heuristics led to what has been called an ``up-hill battle'' between the two methods, while genetic programming was found to be statistically superior to~EGAP~\cite{salehi:uphill}.
Here, we describe a novel ant colony algorithm~(\textsc{GrAnt}) for solving the grammatical optimization problem that significantly outperforms Grammatical~Evolution on diverse optimization problems. The new heuristic is based on the MIN-MAX Ant System\cite{stutzle:mmas}, but differs from previous Ant Programming algorithms on two key points:
\begin{enumerate}
\item The pheromone levels (associated with rewrite rules) are bounded between a minimum and maximum pheromone value. However, the maximum is treated as a soft bound that can be changed by specific events over the course of the search.
\item Each ant constructs a complete derivation tree in a depth-first, targeted fashion (cf.~EGAP'a use of partial sentences and non-terminals).
\end{enumerate}
\textsc{GrAnt}~(Algorithm~\ref{alg:grant}) maintains a pheromone table, holding a pheromone level lying between a hard minimum level $\tau_{min}$, and a soft maximum $\tau_{max}$ for every rewrite rule.
A search iteration begins with each ant constructing a derivation tree of the target sort.The construction proceeds by recursively choosing rewrite rules using simple pheromone-proportional selection. The fitness of the constructed trees is calculated, pheromones are updated by applying evaporation. The iteration-best ant is allowed to deposit pheromones by adding the fitness value to the pheromone level of each rewrite rule used in the derivation. If the iteration-best fitness ever exceeds $\tau_{max}$, then $\tau_{max}$ is updated to the higher value. The motivation for this behavior is assigning more weight to pheromone increases caused by finding fit solutions vs. pheromone buildup caused by repeatedly exploring an area of the search space. As an additional benefit, this eliminates the need for normalizing the amount of pheromones on the edges (shaking). Upon reaching the stopping condition, the algorithm returns the overall best solution found.
\begin{algorithm}
\caption{\textsc{GrAnt} Heuristic}
\lstset{language=Pascal,
basicstyle=\footnotesize
}
\begin{lstlisting}
function grant(t: Sort)
while (! stopped)
solv := construct(p,t) {wlog 1 ant}
iter := fitnessOf(solv)
evaporatePheromone()
for rule in solv
addPheromone(rule, iter)
end for
{update max pheromone}
if iter > tau_max then
tau_max := iter
end if
{update best solution}
if fitnessOf(best) > iter then
best := solv
end if
end while
return best
end function
{recursive path construction}
function construct(p: Pheromones, t: Sort)
{pheromone-proportional rule selection}
c := p.select(rewriteRules(target))
{construct subtree for each non-terminal
of the selected rule}
for i := 0 to c.nonterminals.count
subtrees[i] := construct(p, sortOf(i))
end for
return Tree(c, subtrees)
end function
\end{lstlisting}
\label{alg:grant}
\end{algorithm}
\section{Implementation}
\textsc{ContainAnt} is implemented as a Dependency Injection library for the \textsc{Scala} programming language. The statically typed, object-oriented nature of \textsc{Scala} makes it well-suited for Dependency Injection, and its run-time reflection facilities tremendously simplify the \textsc{ContainAnt} architecture. Moreover, \textsc{Scala} runs on the Java Virtual Machine, allowing the library to work with code bases written in any JVM language (including \textsc{Clojure} and \textsc{Java}).
\textsc{ContainAnt}'s job is assembling objects and object graphs. In effect, the library takes over object instantiation. Instead of using the \texttt{new} keyword with a constructor to instantiate classes, the programmer requests an instance of a given class from \textsc{ContainAnt} (\texttt{ContainAnt create[ClassName]}). The container then heuristically determines what to build by resolving dependencies, choosing appropriate constructors and wiring everything together.
To take advantage of the heuristic capabilities, the programmer has to supply an objective function. With the exception of this objective function, the configuration of \textsc{ContainAnt} is modeled on Google's popular \textsc{Guice} dependency injection library. The programmer provides a \texttt{Module} (a plain object implementing a marker trait) containing the constructors and helper functions to be used during Dependency Injection. If the software to be optimized uses Dependency Injection, these modules will already be present, ready to be used by \textsc{ContainAnt}. This is in strict contrast with the weaver~approach to Programming by Optimization: weaver rules are not present in programs that were not designed with the corresponding PbO toolset in mind.
\textsc{ContainAnt} parses module specifications using \textsc{Scala}'s reflection capabilities, turning the Dependency Injection problem into a grammatical optimization instance. Our analysis~(Section~\ref{ssec:anal}) indicates that the default \textsc{GrAnt} search heuristic suffices to solve many optimization and algorithm configuration problems without problem-specific tuning. This means that using \textsc{ContainAnt} does not require the practitioner to deal with grammars, or even being aware of the heuristics working ``under the hood''.
Since \textsc{ContainAnt} acts like an ordinary dependency injection container, taking over the instantiation of objects and resolution of dependencies, it need not distinguish between off-line and on-line~adaptive optimization: the distinction can be made by using an embedded wrapper to select between `construct on first use' or dynamic/periodic reconstruction \cite{Burles:2015:EDI:2739482.2768423}.
There are no major obstacles to turning the container into a drop-in replacement for \textsc{Guice} by implementing the complete \textsc{Guice} API, thus making PbO immediately available to hundreds of enterprise software projects. This is possibly the most important application of the correspondence detailed in Section~\ref{ssec:bnf}, and the main future target of \textsc{ContainAnt}~development.
\section{Case Studies}
To demonstrate the general behavior of \textsc{ContainAnt}~and~\textsc{SMAC} \cite{HutHooLey11-smac}, and to compare the performance of their heuristics, we implemented two classical optimization problems (Branin function, Subset Sum) and three algorithm configuration problems ($D$-ary heaps, skiplists and syntax highlighting). For comparison purposes, one problem of each class was also implemented for use with \textsc{SMAC}. In this section, we offer a detailed look at each problem, followed by a performance comparison showing that \textsc{GrAnt}~significantly outperform the other heuristics in all but one of these problems.
\subsection{Classical Problems}
\subsubsection{Branin Function}
In this first case study, we compare \textsc{ContainAnt} with \textsc{SMAC} on a global optimization problem. The goal is to minimize the value of the \textit{Branin function} on a given bounded subset of the Euclidean plane.
The Branin function (introduced by Dixon~and~Szeg\H{o} in their traditional optimization test suit~\cite{dixon:global}) has long been a popular benchmark for continuous optimization heuristics. The function has the form
\begin{align*}
\mathrm{branin}(x_1,x_2) =& \left( \left(x_2 - \frac{5.1}{4\pi^2}\right) x_1^2 + \frac{5}{\pi}x_1 - 6 \right)^2 + \\
& 10\left( 1 - \frac{1}{8\pi} \right) \cos(x_1) + 10
\end{align*}
with the domain restricted so that $x_1 \in [-5, 10]$ and $x_2 \in [0, 15]$. There are three global minima on this domain, each with value $\sim 0.397 = 2.48^{-1}$.
The Branin function provides an ideal context for comparing the behavior and the performance of \textsc{SMAC} and \textsc{ContainAnt}, since the \textsc{SMAC}~distribution already includes a configuration for optimizing the Branin function in one of the default example scenarios.
There are many practical techniques for representing a continuous solution space as a BNF grammar. The most intuitive way is including a sufficiently fine ``uniform grid'' of constants from the domain as terminals of the grammar. Alternatively, the grammar of binary strings presented in Section~\ref{ssec:bnf} can represent every dyadic fraction in a compact interval. Dyadic fractions form a dense subset of the interval and provide arbitrary-precision approximations to any given number. We decided to go with the former, more intuitive grammar for this experiment. The result is a large grammar with many terminals, but one that aligns very well with \textsc{SMAC}'s solution representation, thereby ensuring that both heuristics explore search spaces of the same size, which leads to a completely fair comparison.
\subsubsection{Subset Sum}
Given a finite set of integers $S \subseteq \mathbb{Z}$ and target~number~$c \in \mathbb{Z}$, is there a subset~$I \subseteq S$ such that
$$\sum_{i \in I} i = c\:?$$
Known as ``subset sum'', this is one of the ur-examples of an NP-complete decision problem. Recast as an optimization problem, we will attempt to maximize the function $f(I) = \left|c - \sum I\right|^{-1}$ with $f(I) = 2$ if $\sum I = c$. We use two subset sum benchmark instances (P01 and P03) from Burkardt's Scientific Computing Dataset~\cite{burkardt:data} for this case study.
The grammar for this instance consists of the finite grammar generated by the numbers in~$S$, along with the following generic rewrite rules for constructing sets of numbers:
\begin{verbatim}
empty. <Set> ::= empty
add. <Set> ::= add <Int> <Set>
\end{verbatim}
with the obvious compositional semantics
\begin{align*}
k(\mathtt{empty}) &= \emptyset \\
k(\mathtt{add}\:y\:z) &= \left\{k(y)\right\} \cup k(z)
\end{align*}
Notice that the argument-passing system of \textsc{SMAC} would not be capable of supplying arguments of this complexity. The experiment is limited to the \textsc{ContainAnt} heuristics, with 100 runs and the heuristics capped at 1000 objective function evaluations.
\subsection{Programming by Optimization}
\subsubsection{D-ary Heaps}\label{ssec:dheap}
A \emph{min-heap} (resp.~max-heap) structure is a rooted tree in which every node has a value larger (smaller) than the value of its parent. A $D$-ary heap is a heap structure built on a complete $D$-ary tree. The familiar binary heaps are $D$-ary heaps with $D=2$. General $D$-ary heaps allow faster key update operations than the binary case --- $O\left(\log_D n\right)$ vs. $O\left(\log_2 n\right)$. This makes $D$-ary min-heaps (resp.~max-heaps) appropriate for algorithms where decrease (increase) operations are more common than minimum (maximum) extraction.
Generalizing the binary case, the underlying tree can always be implemented as an array, with the children of the $i$th node placed at indices $iD + 1, iD + 2, \dots, iD + D$. This implementation strategy improves cache efficiency and enables random access. There is a performance trade-off, however: the array will eventually fill up, triggering an expensive resize operation.
A $D$-ary heap data structure implemented with arrays has three parameters: the initial size of the array, the expansion factor of the resize operation, and (of course) the arity~$D$. The optimal values of these parameters depend on the expected number of values to be stored in the structure, as well as the expected distribution of decrease/increase and minimum/maximum extraction operations.
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{fig-dheap}
\caption{A ternary min-heap and its array representation.}
\label{fig:dheap}
\end{figure}
The optimization of $D$-ary heaps was implemented by~Hoos~and~Hsu as a test instance for the original Programming~by~Optimization proposal. The original code is written in an extended dialect of the \textsc{Java} programming language, designed for use with a PbO~weaver. The weaver-specific declarations have to be factored out into constructor arguments - a mere three lines of changes, one for each parameter described above. The resulting standard \textsc{Java} is directly usable by \textsc{ContainAnt}.
The grammar for the data structure configuration problem consists of the constructor for the dynamic heap class as the only proper rewrite rule; there are classes and constants for heaps, their arities, expansion factors and initial sizes, all of them equipped with their usual semantics. The objective function counts the number of accesses to the underlying array (with each resize operation counting as two accesses for each index, in line with the usual amortized analysis for array lists) under a given test load. Evaluating the objective function for this task is very expensive, so the experiment is limited to 10 runs, with the heuristics capped at 1000 objective function evaluations.
\subsubsection{Skiplists}
Skiplists are a probabilitistic alternative to balanced binary search trees \cite{pugh:skiplist}. Skiplists are essentially ordered linked lists where each node may contain multiple forward links. In the familiar linked list, a node consists of a value (a piece of data) and a link to the next node. Nodes in a skiplist contain a whole hierarchy of links, each one pointing to a farther subsequent node than the one below it. These auxiliary links provide an ``express lane'' for navigating the structure and can be exploited to implement all three dictionary operations (insertion, lookup and deletion of values) with logarithmic expected time complexity. Thus, the performance of skiplists is comparable to that of balanced binary search trees.
Skiplists are parametrized by two numeric values: the transition probability $p\in \mathbb{R}$ and the maximal height of the hierarchy~$h\in \mathbb{N}$. To find a given value $v$ in a skiplist, start by following the highest level links of the hierarchy, advancing until either $v$ is encountered, or the value of the next node is greater than $v$. In the latter case, continue the search by following links one level down in the hierarchy.
To insert a given value $v$ into a skiplist, start by finding its location using the method described above. Create a node for storing $v$. Now, generate a uniform random real $x \in [0,1]$ and link the newly created node to its neighbors in level~$\ell$ of the hierarchy if and only if $p^k > \ell$. When $p=0.5$, one can intuitively think of this process as a series of coin flips. If you get heads, you link the node to its neighbors on level $\ell$ of the hierarchy, then repeat the procedure on level $\ell+1$. If you get tails or reach the maximum height $\ell = h$, the insertion operation ends.
Instead of having a fixed parameter~$p$, where the probability of inserting a value into level $k$ of the hierarchy is always $\frac{1}{p^{-k}}$, one can consider a more general skiplist architecture, where this probability is given by $\frac{1}{P_k}$, where~$P: \mathbb{N} \rightarrow \mathbb{R}_+$ is an arbitrary monotone sequence. In the experiment, we will focus on three different types of sequences:
\begin{itemize}
\item \textbf{Geometric}: $P_k = a^k$ for some $a > 1$,
\item \textbf{Arithmetic}: $P_k = a + k$ for some $a > 0$ and
\item Sums of the previous two types.
\end{itemize}
Hence, our skiplists will have two parameters: the maximum height~$h$, and the probability sequence~$P$. The expected time complexity of lookups is independent of the distribution of the values~\cite{motwani:randomized}. However, the optimal choices of the parameters~$P$~and~$h$ do depend on the expected number of items to be stored in the skiplist. Skiplists are often stored in a distributed fashion, where the optimal configuration may further depend on variables such as network latency, giving rise to an on-line data structure configuration problem.
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{fig-skiplist}
\caption{Skiplist with a three-layer hierarchy.}
\label{fig:skiplist}
\end{figure}
\textsc{ContainAnt} is readily able to solve this parameter tuning problem --- indeed, we have already evaluated this capability on a much larger search space in Section~\ref{ssec:dheap}. However, we can use optimization to explore a more interesting search space by considering a generalized variant of skiplists.
Let $\mathtt{p}$ denote the (non-terminal corresponding to) the class of integer sequences. The grammar for this data structure configuration problem has a rewrite rule corresponding to the constructor of the skiplist class, as well as three special rewrite rules for constructing the probability sequences:
\begin{verbatim}
geom. <Prob> ::= geom <Double>
arit. <Prob> ::= arit <Double>
sum. <Prob'> ::= sum <Prob> <Prob>
\end{verbatim}
The compositional semantics assigns
\begin{align*}
k(\mathtt{geom}\:y) &= (1,k(y),k(y)^2,k(y)^3,\dots) \\
k(\mathtt{arit}\:y) &= (1,1+k(y),1+2k(y),\dots) \\
k(\mathtt{sum}\:y\:z) &= k(y) \oplus k(z)
\end{align*}
where the symbol $\oplus$ denotes the termwise sum of two sequences. As in the other grammars, there are constructors for skiplists and constants for the numerical parameters, all of them equipped with their usual semantics. This shows that the grammatical approach can conveniently represent sophisticated search spaces that would be difficult and sometimes impossible to describe via \textsc{SMAC}'s text-based configuration files.
The objective function fills the skiplist structure with 1000 random values, and performs 100 random lookups, measuring the total number of comparisons performed. All heuristics are capped at 100 objective function evaluations. The search is fast enough to make 100 runs of the experiment feasible.
\subsubsection{Syntax Highlighting}\label{ssec:syntax}
This final case study serves to showcase a practical use case for Programming~by~Optimization in general and \textsc{ContainAnt} in particular: the creation of software with search-based ``dynamic adaptive'' features. Our minimal example is a syntax highlighter that automatically adjusts itself to different display environments. The potential applications include battery-saving color schemes compatible across different devices (using the technique of Burles~\textit{et al.}~\cite{burles:energy} to incorporate energy consumption into the objective function) and schemes that remain readable when transplanted to different environments (e.g. embedded into social media or displayed by the fixed background color ``webview'' of a mobile application).
\textsc{Agda} is an increasingly popular dependently typed programming language designed by Ulf~Norell \cite{norell:agda}. The \textsc{Agda}~compiler can generate documentation web pages which include the navigable, syntax-highlighted source code of the compiled software. Unfortunately, the default color scheme for the syntax highlighting is unreadable on dark backgrounds, which causes problems when embedding the generated documentation into a larger website.
Our test program generates a readable color scheme for \textsc{Agda} documentation given a target background color as input. The program consists of little more than a naive fitness function quantifying the readability of a color scheme by penalizing low contrast and by rewarding color schemes based around a small number of complementary colors. All of the search is relegated to either~\textsc{SMAC}~or~\textsc{ContainAnt}. The former requires a configuration file with 27 categorical variables, each with 27 options. In addition, about 100 lines of boilerplate code had to be written for handling command line arguments and interfacing with~\textsc{SMAC}. For \textsc{ContainAnt}, the grammar specification, consisting of the constructors for the \texttt{ColorScheme} and \texttt{RGBValue} classes, takes 37 lines altogether. The heuristics are capped at 1000 objective function evaluations.
\subsection{Analysis}\label{ssec:anal}
All experiments were performed on the following system:
\begin{itemize}
\item \textbf{CPU}: Intel Xeon E5-2676 clocked at 2.40GHz with 30~MB Level 3 cache,
\item \textbf{RAM}: 1019280k total,
\item \textbf{Swap}: disabled,
\item \textbf{JVM version}: 1.8.0\_121.
\end{itemize}
The data and code that actually conducted this analysis are published in the companion GitHub repository\footnote{ \url{https://github.com/zaklogician/ContainAnt} } of the article.
The \textsc{ContainAnt} implementation is deterministic, and the repository bundles a convenient build script, allowing anyone to execute the same analysis and replicate/duplicate our results.
The performance of the heuristics was compared on three variables:
\begin{enumerate}
\item The mean quality (\textit{avg}) achieved by the best~of~run solution returned by the heuristic, averaged over all runs.
\item The optimum quality (\textit{max}) achieved by the best~of~run solution returned by the heuristic, taken over all runs.
\item The variance\footnote{Important for on-line optimization, where the heuristics will be run a large number of times. A technique with high mean but low variance may well lose out to another technique with lower mean but high variance over a large number of runs.} (\textit{var}) of the quality achieved by the best~of~run solutions, taken over all runs.
\end{enumerate}
The significance of the differences between the performance of the top heuristics is checked using the nonparametric protocol of Wineberg~and~Christensen~\cite{wineberg:stat}. The final $p$-values are reported in Table~\ref{tbl:results}.
Each experiment is performed with a fixed number of runs (that number depending on the case study, as explained in the respective subsections). Our goal is to pick the technique that achieves the solution of the highest quality possible, given a single run with a fixed budget of ``computational effort''. To ensure fair comparison, we need to limit the number of objective function evaluations identically for all heuristics. For the constructive heuristics (random search and~\textsc{GrAnt}), this can easily be achieved by capping the number of iterations. For \textsc{GrEvo}, the number of evaluations depends only on the population size and the number of generations, allowing us to limit the number of evaluations by capping the product of these two parameters. \textsc{SMAC} has a mechanism for imposing this cap directly via the configuration file.
The \textsc{GrEvo} heuristic has some tunable (hyper)-parameters, including population size and the number of generations. We hand-selected the best-performing ratio of these parameters from the set $\left\{ (100:10), (40:25), (25:40), (10:100) \right\}$ separately for each case study. \textsc{ContainAnt} is capable of tuning the hyper-parameters of its own heuristics. In principle \textsc{ContainAnt} could be used as its own hyper-heuristic to self-improve \textsc{GrEvo}. We experimented with these capabilities during the early days of development. However, we abandoned this avenue once evidence emerged that significant improvement to these parameters would not be possible within the constraints of the case studies (see~the~paragraph dedicated to \textsc{GrEvo} below).
Table~\ref{tbl:results} summarizes the results achieved by~\textsc{SMAC} and the \textsc{ContainAnt}~heuristics on all five case studies presented above.
\begin{table*}[!t]
\centering
\caption{Performance of~\textsc{SMAC}, \textsc{ContainAnt}~heuristics and random search on the five case studies.}
\label{tbl:results}
\begin{tabular}{lllll}
\textbf{Branin:} & {\ul \textbf{GrAnt}} & {\ul \textbf{GrEvo}} & {\ul \textbf{Rand.}} & {\ul \textbf{SMAC}} \\
max: & 2.48 & 1.55 & 2.45 & 2.48 \\
avg: & 1.80 & 0.87 & 1.37 & 1.47 \\
var: & 0.24 & 0.08 & 0.32 & 0.35 \\
p: & \textless.001 & & &\\
\textbf{Subset Sum P02:} & {\ul \textbf{GrAnt}} & {\ul \textbf{GrEvo}} & {\ul \textbf{Rand.}} & {\ul \textbf{SMAC}} \\
max: & 2.00 & 1.00 & 1.00 & ~- \\
avg: & 0.91 & 0.65 & 0.03 & ~- \\
var: & 0.45 & 0.22 & 0.02 & ~- \\
p: & \textless.001 & & &\\
\textbf{Subset Sum P03:} & {\ul \textbf{GrAnt}} & {\ul \textbf{GrEvo}} & {\ul \textbf{Rand.}} & {\ul \textbf{SMAC}} \\
max: & 2.00 & 0.00 & 0.00 & ~- \\
avg: & 0.38 & 0.00 & 0.00 & ~- \\
var: & 0.62 & 0.00 & 0.00 & ~- \\
p: & \textless.001 & & &\\
\textbf{DHeap:} & {\ul \textbf{GrAnt}} & {\ul \textbf{GrEvo}} & {\ul \textbf{Rand.}} & {\ul \textbf{SMAC}} \\
max: & 46801 & 46801 & 46801 & ~- \\
avg: & 46801 & 46752 & 46594 & ~- \\
var: & 0 & 10671 & 52919 & ~- \\
p: & 0.168 & & &\\
\textbf{Skiplist:} & {\ul \textbf{GrAnt}} & {\ul \textbf{GrEvo}} & {\ul \textbf{Rand.}} & {\ul \textbf{SMAC}} \\
max: & 0.33 & 0.25 & 0.27 & ~- \\
avg: & 0.28 & 0.25 & 0.25 & ~- \\
var: & 0.01 & 0.00 & 0.00 & ~- \\
p: & \textless.001 & & &\\
\textbf{Syntax H. Blue:} & {\ul \textbf{GrAnt}} & {\ul \textbf{GrEvo}} & {\ul \textbf{Rand.}} & {\ul \textbf{SMAC}} \\
max: & 37.82 & 38.02 & 37.57 & ~- \\
avg: & 34.36 & 34.50 & 32.18 & ~- \\
var: & 5.85 & 8.73 & 7.46 & ~- \\
p: & 0.663 & & &\\
\textbf{Syntax H. Yellow:} & {\ul \textbf{GrAnt}} & {\ul \textbf{GrEvo}} & {\ul \textbf{Rand.}} & {\ul \textbf{SMAC}} \\
max: & 34.92 & 34.68 & 33.85 & 34.44 \\
avg: & 31.51 & 32.33 & 29.22 & 30.92 \\
var: & 4.02 & 3.30 & 5.67 & 5.22 \\
p: & 0.082 & & &
\end{tabular}
\end{table*}
\subsubsection*{GrAnt}
The \textsc{GrAnt} heuristic significantly outperformed all others in the majority of experiments. The only exception is the syntax~highlighting study, where \textsc{GrEvo} systematically had the highest nominal mean. However, hypothesis testing reveals that the differences are not significant. \textsc{GrAnt} is the only heuristic to perform equally well across both combinatorial~optimization and algorithm~configuration problems, and the only one to find globally optimal solutions to both the Branin~function and both subset sum instances.
\subsubsection*{GrEvo}
The poor performance of the \textsc{GrEvo} heuristic, consistent across parameter settings, is crying out for an explanation. Our investigation suggests that the main culprit may be early loss of diversity (visible in the Skiplist~study, where the algorithm converges in a mere five generations), caused by the fact that the first few elements of the genome have a disproportionately high influence on the phenotype in Grammatical~Evolution~\cite{rothlauf:locality}. Increasing the population size is not possible without moving beyond the strict computational bounds of our case studies, rendering~Grammatical~Evolution unsuitable for many real-time applications. Solving this issue could be an avenue of further research.
\subsubsection*{SMAC}
As expected, the quality of the results returned by~\textsc{SMAC} significantly outperformed random search in all cases. However, the average quality lingered beneath that of~\textsc{GrAnt} in the case of the Branin function (although the best solution for the Branin function was globally optimal) and beneath both \textsc{ContainAnt} heuristics in the algorithm configuration case. Another major issue is speed: \textsc{SMAC} spends over four hours on latter problem, while the \textsc{ContainAnt} heuristics finish both in 46~seconds.
\section{Conclusion}
Dependency Injection can be used to improve the existing weaver-based Programming by Optimization tools. We have described a library that implements several grammatical optimization metaheuristics, including a novel Ant Programming approach. The library provides better support for Programming by Optimization than specialized language extensions and weaver tools, while doing away with several limitations such as difficulties with on-line optimization.
Furthermore, regarding Dependency Injection as an instance of a grammatical optimization problem leads to a whole new class of heuristics for automatic algorithm configuration. The proposed grammatical Ant Programming heuristic \textsc{GrAnt} significantly outperforms existing algorithms on five problems of interest, in one case reducing a four hour long \textsc{SMAC} optimization task to 46 seconds while significantly improving on the solution quality.
Programming by Optimization libraries can act as drop-in replacement for existing Dependency Injection containers, making PbO immediately applicable to a large number of enterprise software projects. The development of \textsc{ContainAnt} in this direction is a promising target of future work.
\bibliographystyle{plain}
|
\section{Introduction}
The spin down of cool stars ($M_*\lesssim 1.3M_{\sun}$) is a complex function of mass and age, as shown by the increasing number of rotation period measurements for large stellar populations \citep{barnes2003rotational, irwin2009ages, barnes2010simple, agueros2011factory, meibom2011color, mcquillan2013measuring, bouvier2014angular, stauffer2016rotation, 2017ApJ...835...16D}. Observed properties of these stars show a wide range of mass loss rates, coronal temperatures, field strengths and geometries, which all connect with stellar rotation to control the loss of angular momentum (\citealp{reiners2012radius}; \citealp{gallet2013improved}; \citealp{van2013fast}; \citealp{brown2014metastable}; \citealp{matt2015mass}; \citealp{gallet2015improved}; \citealp{amard2016rotating}; \citealp{blackman2016minimalist}; See et al. in prep). Despite the wide range of interlinking stellar properties an overall trend of spin down with an approximately Skumanich law is observed at late ages; $\Omega_* \propto \tau^{-0.5}$ \citep{skumanich1972time, soderblom1983rotational}.
For Sun-like stars on the main sequence, the spin-down process is governed primarily by their magnetised stellar winds which remove angular momentum over the star's lifetime. \cite{parker1958dynamics} originally posited that stellar winds must exist due to the thermodynamic pressure gradient between the high temperature corona and interplanetary space. Continued solar observations have constrained theoretical models for the solar wind to a high degree of accuracy \citep{van2014alfven, usmanov2014three, oran2015steady}. Recent models of the solar wind are beginning to accurately reproduce the energetics within the corona and explain the steady outflow of plasma into the Heliosphere (e.g. \citealp{grappin1983dependence}; \citealp{van2010data}; \citealp{pinto2016flux}). The wind driving is now known to be much more complex than a thermal pressure gradient, with authors typically heating the wind through the dissipation of Alfv\'en waves in the corona. Other cool stars are observed with x-ray emissions indicating hot stellar coronae like that of the Sun \citep{rosner1985stellar, hall2007activity, wright2004chromospheric, wolk2005stellar}. Similar stellar winds and wind heating mechanisms are therefore expected to exist across a range of Sun-like stars. Assuming equivalent mass loss mechanisms, results from the Solar wind are incorporated into more general stellar wind modelling efforts \citep[e.g.][]{cohen2014grid, alvarado2016simulating}.
Detailed studies of wind driving physics remain computationally expensive to run, so are usually applied on a case-by-case basis. How applicabile the heating physics gained from modelling the Solar wind is to other stars still in question. With the reliability of such results even for the global properties of a given star in question, large parameter studies with simpler physics remain useful. A more general method can allow for parametrisations which are more appropriate to the variety of stellar masses and rotation periods found in observed stellar populations. Parker-type solutions remain useful for this due to their simplicity and versatility \citep{parker1965dynamical, mestel1968magnetic, sakurai1990magnetohydrodynamic, keppens1999numerical}. In these solutions, wind plasma is accelerated from the stellar surface and becomes transonic at the sonic surface. With the addition of magnetic fields the wind also become trans-alfv\'enic, i.e faster than the Alfv\'en speed, at the Alfv\'en surface. \cite{weber1967angular} showed for a one-dimensional magnetised wind that the Alfv\'en radius represented a lever arm for the spin-down torque. Since the introduction of this result, many researchers have produced scaling laws for the Alfv\'en radius (\citealp{mestel1984angular}; \citealp{kawaler1988angular}; \citealp{matt2008accretion}; \citealp{matt2012magnetic}; \citealp{ud2009dynamical}; \citealp{pinto2011coupling}; \citealp{reville2015effect}; Pantolmos. in prep) all of which highlight the importance of the magnetic field strength and mass loss rate in correctly parametrising a power law dependence. In such formulations, the mass loss rate is incourporated as a free parameter as the physical mechanisms which determines it are not yet completely understood. Measuring the mass loss rate from Sun-like stars is particularly difficult due to the wind's tenuous nature and poor emission. \cite{wood2004astrospheres} used Lyman-$\alpha$ absorption from the interaction of stellar winds and their local interstellar medium to measure mass loss rates, but the method is model-dependent and only available for a few stars. Theoretical work from \cite{cranmer2011testing} predicts the mass loss rates from Sun-like stars, but it is uncertain if the physics used within the model scales correctly between stars. Therefore, parameter studies where the mass loss rate is an unknown parameter are needed.
In addition to the mass loss rate, the angular momentum loss rate is strongly linked with the magnetic properties of a given star. Frequently researchers assume the dipole component of the field to be the most significant in governing the global wind dynamics \citep[e.g. ][]{ustyugova2006propeller, zanni2009mhd, gallet2013improved, cohen2014grid, gallet2015improved, matt2015mass, johnstone2015stellar}. Zeeman Doppler Imaging (ZDI) studies \citep[e.g.][]{morin2008large, petit2008toroidal, fares2009magnetic, vidotto2014stellar, jeffers2014e, see2015energy, see2016connection, folsom2016evolution, hebrard2016modelling, see2016studying}, provide information on the large scale surface magnetic fields of active stars. Observations have shown stellar magnetic fields to be much more complex than simple dipoles, containing combinations of many different field modes. ZDI is a topographic technique typically decomposes the field at the stellar surface into individual spherical harmonic modes. The 3D field geometry can then be recovered with field extrapolation techniques using the ZDI map as an inner boundary. Several studies have considered how these observed fields affect the global wind properties. Typically used to determine an initial 3D field solution, then a magnetohydrodynamics code evolves this initial state in time until a steady state solution for the wind and magnetic field geometry is attained \citep[e.g.][]{vidotto2011understanding, cohen2011dynamics, garraffo2016space, reville2016age, alvarado2016simulating, nicholson2016temporal, do2016magnetic}. These works are less conducive to the production of semi-analytical formulations, as the principle drivers of the spin-down process are hidden within complex field geometries, rotation and wind heating physics.
A few studies show systematically how previous torque formulations depend on magnetic geometry using single modes. \cite{reville2015effect} explored thermally driven stellar winds with dipolar, quadrupolar and octupolar field geometries. They concluded that higher order field modes produce a weaker torque for the same field strength and mass loss, which is supported by results from \cite{garraffo2016missing}. Despite these studies and works like them, only one study has systematic scaled the mass loss rate for a mixed field geometry field \citep{strugarek2014modelling}. However, the aforementioned studies of the angular momentum loss from Sun-like stars have yet to address the systematic addition of individual spherical harmonic field modes.
Mixed geometry fields are observed within our closest star, the Sun, which undergoes a 11 year cycle oscillating between dipolar and quadrupolar field modes from cycle minimum to maximum respectively \citep{derosa2012solar}. Observed Sun-like stars also exhibit a range of spherical harmonic field combinations. Simple magnetic cycles are observed using ZDI, both HD 201091 \citep{saikia2016solar} and HD 78366 \citep{morgenthaler2012long} show combinations of the dipole, quadrupole and octupole field modes oscillating similarly to the solar field. Other cool stars exist with seemingly stochastic changing field combinations \citep{petit2009polarity, morgenthaler2011direct}. Observed magnetic geometries all contain combinations of different spherical harmonic modes with a continuous range of mixtures, it is unclear what impact this will have on the braking torque.
In this study we will investigate the significance of the dipole field when combined with a quadrupolar mode. We focus on these two field geometries, which are thought to contribute in anti-phase to the solar cycle and perhaps more generally to stellar cycles in cool stars. Section 2 covers the numerical setup with a small discussion of the magnetic geometries for which we develop stellar wind solutions. Section 3 presents the main simulation results, including discussion of the qualitative wind properties and field structure, along with quantitative parametrisations for the stellar wind torque. Here we also highlight the dipole's importance in the braking, and introduce an approximate scaling relation for the torque. Finally in Section 4 we focus on the magnetic field in the stellar wind, first a discussion of the overall evolution of the flux, then a discussion of the open flux and opening radius within our simulations. Conclusions and thoughts for further work can then be found in Section 5. The Appendix contains a short note on the wind acceleration profiles of our wind solutions.
\section{Simulation Method}
\subsection{Numerical Setup}
This work uses the magnetohydrodynamics (MHD) code PLUTO \citep{mignone2007pluto, mignone2009pluto}, a finite-volume code which solves Riemann problems at cell boundaries in order to calculate the flux of conserved quantities through each cell. PLUTO is modular by design, capable of interchanging solvers and physics during setup. The present work uses a diffusive numerical scheme, the solver of Harten, Lax, and van Leer, HLL \citep{einfeldt1988godunov}, which allows for greater numerical stability in the higher strength magnetic field cases. The magnetic field solenoidality condition ($\nabla\cdot{\bf B}=0$) is maintained using the Constrained Transport method (See \cite{toth2000b} for discussion).
The MHD equations are solved in a conservative form, with each equation relating to the conservation of mass, momentum and energy, plus the induction equation for magnetic field,
\begin{eqnarray}
\frac{\partial\rho}{\partial t} + \nabla\cdot\rho{\bf v} &=& 0,\\
\frac{\partial {\bf m}}{\partial t} + \nabla\cdot({\bf mv - BB +I}p_T) &=& \rho{\bf a},\\
\frac{\partial E}{\partial t} + \nabla\cdot((E+p_T){\bf v - B(v\cdot B})) &=& {\bf m \cdot a} ,\\
\frac{\partial{\bf B}}{\partial t} + \nabla \cdot ({\bf vB - Bv}) &=& 0 .
\end{eqnarray}
Here $\rho$ is the mass density, $\bf v$ is the velocity field, $\bf a$ is the gravitational acceleration, $\bf B$ is the magnetic field\footnote{The PLUTO code operates with a factor of $1/\sqrt{4\pi}$ absorbed into the normalisation of B. Tabulated parameters are given in cgs units with this factor incorporated.}, $p_T=p+B^2/8\pi$ is the combined thermal and magnetic pressure, $\bf m$ is the momentum density given by $\rho{\bf v}$ and $E$ is the total energy density. The energy of the system is written as $E = \rho\epsilon +{\bf m}^2/(2\rho) + {\bf B}^2/2$, with $\epsilon$ representing the internal energy per unit mass of the fluid. $\bf I$ is the identity matrix. A polytropic wind is used for this study, such that the closing equation of state takes the form $\rho\epsilon = p/(\gamma -1)$ where $\gamma$ represents the polytropic index.
We assume the wind profiles to be axisymmetric and solve the MHD equations using a spherical geometry in 2.5D, i.e. our domain contains two spatial dimensions ($r, \theta$) but allows for 3D axisymetric solutions for the fluid flow and magnetic field using three vector components ($r, \theta, \phi$). The domain extends from one stellar radius ($R_*$) out to $60R_*$ with a uniform grid spacing in $\theta$ and a geometrically stretched grid in $r$, which grows from an initial spacing of $0.01R_*$ to $1.08R_*$ at the outer boundary. The computational mesh contains $N_r\times N_{\theta}=256\times 512$ grid cells. These choices allow for the highest resolution near the star, where we set the boundary conditions that govern the wind profile in the rest of the domain.
Initially a polytropic parker wind \citep{parker1965dynamical, keppens1999numerical} with $\gamma=1.05$ fills the domain, along with a super-imposed background field corresponding to our chosen magnetic geometry and strength. During the time-evolution, the plasma pressure, density, and poloidal components of the magnetic field ($B_r,B_\theta$) are held fixed at the stellar surface, whilst the poloidal components of the velocity ($v_r,v_\theta$) are allowed to evolve in response to the magnetic field (the boundary is held with $dv_r/dr=0$ and $dv_{\theta}/dr=0$). We then enforce the flow at the surface to be parallel to the magnetic field ($\bf v || B$). The star rotates as a solid body, with $B_{\phi}$ linearly extrapolated into the boundary and $v_{\phi}$ set using the stellar rotation rate $\Omega_{*}$,
\begin{equation}
v_{\phi}=\Omega_{*}{r sin\theta}+\frac{\bf v_p \cdot B_p}{|{\bf B_p|}^2}B_{\phi},
\end{equation}
where the subscript ``p'' denotes the poloidal components ($r,\theta$) of a given vector. This condition enforces an effective rotation rate for the field lines which, in steady state ideal MHD, should be equal to the stellar rotation rate and conserved along field lines \citep{zanni2009mhd, reville2015effect}. This ensures the footpoints of the stellar magnetic field are correctly anchored into the surface of the star. The final boundary conditions are applied to the outer edges of the simulation, a simple outflow (zero derivative) is set at $60R_*$ allowing for the outward transfer of mass, momenta and magnetic field, along with an axisymmetric condition along the rotation axis ($\theta=0$ and $\pi$). Due to the supersonic flow properties at the outer boundary and its large radial extent compared with the location of the fast magnetosonic surface, any artefacts from the outer boundary cannot propagate upwind into the domain.
The code is run, following the MHD equations above, until a steady state solution is found. The magnetic fields modify the wind dynamics compared to the spherically symmetric initial state, with regions of high magnetic pressure shutting off the radial outflow. In this way, the applied boundary conditions allow for closed and open regions of flow to form (e.g \citealp{washimi1993thermo}; \citealp{keppens2000stellar}), as observed within the solar wind. In some cases of strong magnetic field small reconnection events are seen, caused by the numerical diffusivity of our chosen numerical scheme. Reconnection events are also seen in Pantolmos \& Matt (in prep) and discussed within their Appendix. We adopt a similar method for deriving flow quantities in cases exhibiting periodic reconnection events. In such cases, once a quasi-steady state is established a temporal average of quantities such as the torque and mass loss are used.
Inputs for the simulations are given as ratios of characteristic speeds which control key parameters such as the wind temperature ($c_s/v_{esc}$), field strength ($v_A/v_{esc}$) and rotation rate ($v_{rot}/v_{kep}$). Where $c_s=\sqrt{\gamma p/\rho}$ is the sound speed at the surface, $v_A=B_*/\sqrt{4\pi\rho}$ is the Alfv\'en speed at the north pole, $v_{rot}$ is the rotation speed at the equator, $v_{esc}=\sqrt{2GM_*/R_*}$ is the surface escape speed and $v_{kep}=\sqrt{GM_*/R_*}$ is the keplerian speed at the equator. In this way, all simulations represent a family of solutions for stars with a range of gravities. As this work focuses on the systematic addition of dipolar and quadrupolar geometries, we fix the rotation rate for all our simulations. \cite{matt2012magnetic} showed that the non-linear effects of rotation on their torque scaling can be neglected for slow rotators. They defined velocities as a fraction of the breakup speed,
\begin{equation}
f=\frac{v_{rot}}{v_{kep}}\bigg|_{r=R_*, \theta=\pi/2}=\frac{\Omega_*R_*^{3/2}}{(GM_*)^{1/2}}.
\end{equation}
The Alfv\'en radius remains independent of the stellar spin rate until $f\approx0.03$, after which the effects of fast rotation start to be important. For this study a solar rotation rate is chosen ($f=4.46\times10^{-3}$), which is well within the slow rotator regime. We set the temperature of the wind with $c_s/v_{esc}=0.25$, higher than $c_s/v_{esc}=0.222$ used previosuly in \cite{reville2015effect}. This choice of higher sound speed drives the wind to slightly higher terminal speeds, which are more consistent with observed solar wind speeds. Each geometry is studied with 10 different field strengths controlled by the input parameter $v_A/v_{esc}$, which is defined here with the Alfv\'en speed on the stellar north pole (see following Section). Table \ref{Parameters} lists all our variations of $v_A/v_{esc}$ for each geometry.
\begin{figure*}
\includegraphics[width=\textwidth]{f1.eps}
\caption{Initial magnetic configurations for a dipolar field, quadrupolar field and two mixed cases (red, green, magenta and blue for the dipole fractions of 100\%, 50\%, 10\% and purely quadrupole respectively). Mixed cases have the dominant pure field geometry over-plotted in dashed colour. The combined fields add in the northern hemisphere and subtract in southern hemisphere because they belong to opposite field symmetry families. With as much as half the field strength in the quadrupole, shown in green, the topology of the field is still dominated by the dipole field.}
\label{RstudyField}
\end{figure*}
Due to the use of characteristic speeds as simulation inputs, our results can be scaled to any stellar parameters. For example, using solar parameters, the wind is driven by a coronal temperature of $\approx$1.4MK and our parameter space covers a range of stellar magnetic field strengths from 0.9G to 87G over the pole. Changing these normalisations will modify this range.
\subsection{Magnetic Field Configuration}
Within this work, we consider magnetic field geometries that encompass a range of dipole and quadrupole combinations with different relative strengths. We represent the mixed fields using the ratio, $\mathcal{R}_{dip}$, of dipolar field to the total combined field strength.
In this study the magnetic fields of the dipole and quadrupole are described in the formalism of \cite{gregory2010magnetic} using polar field strengths,
\begin{eqnarray}
B_{r,dip}(r,\theta) &=& B_{*}^{l=1} \left(\frac{R_*}{r}\right)^{3} \cos\theta, \\
B_{\theta,dip}(r,\theta) &=& \frac{1}{2} B_{*}^{l=1} \left(\frac{R_*}{r}\right)^{3} \sin\theta, \\
B_{r,quad}(r,\theta) &=& \frac{1}{2} B_{*}^{l=2} \left(\frac{R_*}{r}\right)^{4} (3\cos^2\theta -1) ,\\
B_{\theta,quad}(r,\theta) &=& B_{*}^{l=2} \left(\frac{R_*}{r}\right)^{4} \cos\theta\sin\theta.
\end{eqnarray}
The total field, comprised of the sum of the two geometries,
\begin{equation}
{\bf B}(r,\theta) = {\bf B}_{dip}(r,\theta) + {\bf B}_{quad}(r,\theta),
\end{equation}
where the total polar field $B_*=B_*^{l=1}+B_*^{l=2}$, is controlled by the $\mathcal{R}_{dip}$ parameter,
\begin{equation}
\mathcal{R}_{dip}=\frac{B_{r,dip}}{B_{r,dip}+B_{r,quad}}\bigg|_{r=R_*,\theta=0}=\frac{B_*^{l=1}}{B_*}.
\label{R_dip}
\end{equation}
This work considers aligned magnetic moments such that $\mathcal{R}_{dip}$ ranges from 1 to 0, corresponding to all the field strength in the dipolar or quadrupolar mode respectively. As with $v_A/v_{esc}$, $\mathcal{R}_{dip}$ is calculated at the north pole. This sets the relative strengths of the dipole and quadrupole fields,
\begin{equation}
B_{*}^{l=1}=\mathcal{R}_{dip}B_{*}, \qquad B_{*}^{l=2}=(1-\mathcal{R}_{dip})B_{*},
\end{equation}
Alternative parametrisations are commonly used in the analysis of ZDI observations and dynamo modelling. These communities use the surface averaged field strengths, $\langle |B| \rangle$, or the ratio of magnetic energy density ($E_m\propto B^2$) stored within each of the dipole and quadrupole field modes at the stellar surface. During the solar magnetic cycle, values of $B^2_{quad}/B^2_{dip}$ can range from $\approx10-100$ at solar maximum to $\approx10^{-2}$ at solar minimum \citep{derosa2012solar}. A transformation from our parameter to the ratio of energies is simply given by:
\begin{equation}
\frac{B^2_{quad}}{B^2_{dip}}=\frac{2}{3}\frac{(1-\mathcal{R}_{dip})^2}{\mathcal{R}_{dip}^2},
\end{equation}
where the numerical pre-factor accounts for the integration of magnetic energy in each mode over the stellar surface.
\begin{figure*}
\includegraphics[width=\textwidth]{f2.png}
\caption{Logarithm of density normalised by the surface value for dipolar, quadrupolar and mixed magnetic fields for cases 7, 27, 57, 67 (see Table \ref{Parameters}). The winds are initialised using the same initial polytropic parker wind solution with $\gamma=1.05$ and $c_s/v_{esc}=0.25$. Stellar rotation rate and magnetic field strength are set with $f=4.46\times10^{-3}$ and $v_A/v_{esc}=3.0$. The Alfv\'en and sonic Mach surfaces are shown in blue and black respectively, in addition the fast and slow magnetosonic surfaces are indicated with dot-dash and dashed white lines. A transition from one to two streamers is seen with increasing quadrupolar field (decreasing $\mathcal{R}_{dip}$), and the two combined field cases exhibit the top bottom asymmetry from the field addition and subtraction.}
\label{RstudyExample}
\end{figure*}
Initial field configurations are displayed in Figure \ref{RstudyField}. The pure dipolar and quadrupolar cases are shown in comparison to two mixed cases ($\mathcal{R}_{dip}=0.5, 0.1$). These combined geometry fields add in one hemisphere and subtract in the other. This effect is due to the different symmetry families each geometry belongs to, with the dipole's polarity reversing over the equator unlike the equatorially symmetric quadrupole. Continuing the use of ``primary'' and ``secondary'' families as in \cite{mcfadden1991reversals} and \cite{derosa2012solar}, we refer to the dipole as primary and quadrupole as secondary. The fields are chosen such that they align in polarity in the northern hemisphere. This choice has no impact on the derived torque or mass loss rate due to the symmetry of the quadrupole about the equator. Either aligned or anti-aligned, these fields will always create one additive hemisphere and one subtracting; swapping their relative orientations simply switches the respective hemispheres. This is in contrast to combining dipole \& octupole fields, where the aligned and anti-aligned cases cause subtraction at the equator or poles respectively (\citealp{gregory2016multipolar}; Finley \& Matt. in prep).
Figure \ref{RstudyField} indicates that even with equal quadrupole and dipole polar field strengths, $\mathcal{R}_{dip}=0.5$, the overall dipole topology will remain. In this case the magnetic energy density in the dipolar mode is 1.5 times greater than the quadrupolar mode and with the more rapid radial decay of the quadrupolar field, this explains the overall dipolar topology. A higher fraction of quadrupole is required to produce a noticeable deviation from this configuration, which is shown at $\mathcal{R}_{dip}=0.1$. More than half of the parameter space that we explore lies in the range where the energy density of the quadrupole mode is greater than that of the dipole ($B^2_{quad}/B^2_{dip}>1.0$). For this study both the pure dipolar and quadrupolar fields are used as controls (both of which were studied in detail within \cite{reville2015effect}), and 5 mixed cases parametrised by $\mathcal{R}_{dip}$ values ($\mathcal{R}_{dip}$ = 0.8, 0.5, 0.3, 0.2, 0.1). We include $\mathcal{R}_{dip}=0.8$ to demonstrate the dominance of the dipole at higher values. Each $\mathcal{R}_{dip}$ value is given a unique identifying colour which is maintained in all figures throughout this paper. Table \ref{Parameters} contains a complete list of parameters for all cases, which are numbered by increasing $v_A/v_{esc}$ and quadrupole fraction.
\section{Simulation results}
\subsection{Morphology of the Field and Wind Outflow}
\begin{sidewaysfigure*}
\vspace*{-9cm} \includegraphics[width=\textwidth]{f3.png}
\caption{Simulation results for the lowest $\mathcal{R}_{dip}$ values 0.3, 0.2 and 0.1 (top, middle and bottom respectively), coloured by poloidal wind speed, with field lines in white. The current sheets are indicated by dashed lines, whose colour corresponds to their $\mathcal{R}_{dip}$ value in future figures. The streamer configuration is modified by both changes to the field strength and mixing ratio. Increased field strength or $\mathcal{R}_{dip}$ value tends to revolve the southern hemisphere streamer towards the south pole. The Alfv\'en surfaces have been coloured to show the flux of angular momentum normal to the surface [Units normalised by: $8\times10^{-6}\rho_*v_{kep}R_*$]. The average Alfv\'en radius, $\langle R_A \rangle$, from equation (\ref{R_A}) is shown in dashed grey. The sonic surface and opening radius are shown in solid black and dashed red respectively. The morphology and properties of the lower field cases are nearly indistinguishable, with only slight differences to the streamer locations. The reduction in torque with increasing quadrupolar fraction can be visually seen by moving down the grid. The most dipolar field sits in the top right panel and the most quadrupolar in the bottom left, these models are chosen to emphasis the transition in field dominance. }
\label{streamers}
\end{sidewaysfigure*}
Figure \ref{RstudyField} shows the topological changes in field structure from the addition of dipole and quadrupole fields. It is evident in these initial magnetic field configurations that the global magnetic field becomes asymmetric about the equator for mixed cases, as does the magnetic boundary condition which is maintained fixed at the stellar surface. It is not immediately clear how this will impact the torque scaling from \cite{reville2015effect}, who studied only single geometries.
Results for these field configurations using our PLUTO simulations are displayed in Figure \ref{RstudyExample}. The dipole and quadrupole cases are shown in conjunction with the mixed field cases, $\mathcal{R}_{dip}=0.5, 0.1$. The Figure displays for a comparable value of polar magnetic field strength, the different sizes of Alfv\'en surface that are produced. The mixed magnetic geometries modify the size and morphology of the Alfv\'en and sonic surfaces. Due to the slow rotation, the fast and slow magnetosonic surfaces are co-located with the sonic and Alfv\'en surfaces (the fast magnetosonic surface being always the larger of the two surfaces).
The field geometry is found to imprint itself onto the stellar wind velocity with regions of closed magnetic field confining the flow creating areas of co-rotating plasma, referred to as deadzones \citep{mestel1968magnetic}. Steady state wind solutions typically have regions of open field where a faster wind and most of the torque is contained, along with these deadzone(s) around which a slower wind is produced. Similarly to the solar wind, slower wind can be found on the open field lines near the boundary of closed field \citep{feldman2005sources, riley2006comparison, fisk1998slow}. Observations of the Sun reveal the fast wind component emerging from deep within coronal holes, typically over the poles, and the slow wind component originating from the boundary between coronal holes and close field regions. Due to the polytropic wind used here, we do not capture the different heating and acceleration mechanisms required to create a true fast and slow solar-like wind \citep[as seen with the Ulysses spacecraft e.g.][]{mccomas2000solar, ebert2009bulk}. Our models produce an overall wind speed consistent with slow solar wind component, which we assume to represent the average global flow. More complex wind driving and coronal heating physics are required to recover a multi-speed wind, as observed from the Sun \citep{cranmer2007self, pinto2016flux}.
Figure \ref{streamers} displays a grid of simulations with a range of magnetic field strengths and $\mathcal{R}_{dip}=0.3, 0.2, 0.1$ values ($B^2_{quad}/B^2_{dip}$ ranges from 3.6 to 54; values consistent with the solar cycle maximum), where the mixing of the fields plays a clear role in the changing dynamics of the flow. Regions of closed magnetic field cause significant changes to the morphology of the wind. A single deadzone is established on the equator by the dipole geometry whereas the quadrupole creates two over mid latitudes. Mixed cases have intermediate states between the pure regimes. Within our simulations the deadzones are accompanied by streamers which form above closed field regions and drive slower speed wind than from the open field regions. The dynamics of these streamers, their location and size are an interesting result of the changing topology of the flow.
The dashed coloured lines within Figure \ref{streamers} show where the field polarity reverses using $B_r=0$, which traces the location of the streamers. The motion of the streamers through the grid of simulations is then observed. With increasing quadrupole field, the single dipolar streamer moves into the northern hemisphere and with continued quadrupole addition a second streamer appears from the southern pole and travels towards the northern hemisphere until the quadrupolar streamers are recovered both sitting at mid latitudes. This motion can also be seen for fixed $\mathcal{R}_{dip}$ cases as the magnetic field strength is decreased. For a given $\mathcal{R}_{dip}$ value the current sheets sweep towards the southern hemisphere with increased polar field strength, in some cases (36 and 38) moving onto the axis of rotation. This is the opposite behaviour to decreasing the $\mathcal{R}_{dip}$ value, i.e. the streamer configuration is seen to take a more dipolar morphology as the field strength is increased. Additionally within Figure \ref{streamers}, for low field strengths each $\mathcal{R}_{dip}$ produces a comparable Alfv\'en surface with very similar morphology, all dominated by the quadrupolar mode.
\subsection{Global Flow Quantities}
Our simulations produce steady state solutions for the density, velocity and magnetic field structure. To compute the wind torque on the star we calculate $\Lambda$, a quantity related directly to the angular momentum flux ${\bf F_{AM}}=\Lambda\rho{\bf v}$ \citep{keppens2000stellar},
\begin{equation}
\Lambda(r,\theta)=rsin\theta\bigg(v_{\phi}-\frac{B_{\phi}}{\rho}\frac{|{\bf B_p}|^2}{{\bf v_p \cdot B_p}}\bigg).
\end{equation}
Within axisymmetric steady state ideal MHD, $\Lambda$ is conserved along any given field line. However we find variations from this along the open-closed field boundary due to numerical diffusion across the sharp transition in quantities found there. The spin-down torque, $\tau$, due to the transfer of angular momentum in the wind is then given by the area integral,
\begin{equation}
\tau=\int_A\Lambda\rho{\bf v} \cdot d{\bf A},
\end{equation}
where $A$ is the area of any surface enclosing the star. For illustrative purposes, Figure \ref{streamers} shows the Alfv\'en surface coloured by angular momentum flux (thick multi-coloured line), which is seen to be strongly focused around the equatorial region. The angular momentum flux is calculated normal to the Alfv\'en surface,
\begin{equation}
\frac{d\tau}{dA}=\Lambda\rho{\bf v}\cdot {\bf \hat{A}}={\bf F_{AM}}\cdot {\bf \hat{A}},
\end{equation}
where ${\bf \hat{A}}$ is the normal unit vector to the Alfv\'en surface. The mass loss rate from our wind solutions is calculated similarly to the torque,
\begin{equation}
\dot{M}=\int_A\rho{\bf v} \cdot d{\bf A}.
\end{equation}
Both expressions for the mass loss and torque are evaluated using spherical shells of area $A$ which are outside the closed field regions. This allows for the calculation of an average Alfv\'en radius (which is cylindrical from the rotation axis) in terms of the torque, mass flux and rotation rate,
\begin{equation}
\langle R_A\rangle=\sqrt{\frac{\tau}{\dot{M}\Omega_*}}.
\label{R_A}
\end{equation}
\begin{figure*}
\includegraphics[width=\textwidth]{f4.eps}
\caption{Parameter space explored in terms of $\Upsilon$, $\Upsilon v_{esc}/\langle v(R_A)\rangle$ and $\mathcal{R}_{dip}$. Five mixed geometries are explored along with pure cases of both dipole and quadrupole geometries. Colours for each $\mathcal{R}_{dip}$ value are used throughout this work. The black line indicates $\Upsilon_{crit}$, equation (\ref{UP_CRIT}). The formula for predicting the torque exhibits a quadrupolar scaling for $\Upsilon$ and $\mathcal{R}_{dip}$ values below the line, and dipolar above (See Section 3.4).}
\label{PS}
\end{figure*}
\begin{table*}
\caption{Input Parameters and Results from the 70 Simulations}
\label{Parameters}
\center
\setlength{\tabcolsep}{1pt}
\begin{tabular}{cccccccc|cccccccc}
\hline\hline
Case & $\mathcal{R}_{dip}$ & $v_A/v_{esc}$ & $\langle R_A\rangle/R_*$ & $\Upsilon$ & $R_o/R_*$ & $\Upsilon_{open}$ & $\langle v(R_A)\rangle/v_{esc} $ & Case & $\mathcal{R}_{dip}$ & $v_A/v_{esc}$ & $\langle R_A\rangle/R_*$ & $\Upsilon$ & $R_o/R_*$ & $\Upsilon_{open}$ & $\langle v(R_A)\rangle/v_{esc} $ \\ \hline
1 & 1 & 0.1 & 3.06 & 11.1 & 1.31 & 294 & 0.123 & 36 & 0.3 & 2 & 5.66 & 2930 & 2.61 & 2040 & 0.242 \\
2 & 1 & 0.3 & 4.19 & 73.2 & 1.88 & 819 & 0.183 & 37 & 0.3 & 3 & 6.76 & 6850 & 3.01 & 3460 & 0.283 \\
3 & 1 & 0.5 & 5.05 & 192 & 2.33 & 1450 & 0.221 & 38 & 0.3 & 6 & 9.41 & 31200 & 3.8 & 8840 & 0.360 \\
4 & 1 & 1 & 6.88 & 773 & 2.95 & 3550 & 0.287 & 39 & 0.3 & 12 & 13 & 137000 & 5.05 & 21600 & 0.432 \\
5 & 1 & 1.5 & 8.56 & 1880 & 3.41 & 6530 & 0.334 & 40 & 0.3 & 24 & 15.7 & 360000 & 6.18 & 37300 & 0.476 \\
6 & 1 & 2 & 10 & 3660 & 3.8 & 9970 & 0.367 & 41 & 0.2 & 0.1 & 2.43 & 10.7 & 1.2 & 120 & 0.078 \\
7 & 1 & 3 & 12.6 & 9280 & 4.54 & 18100 & 0.414 & 42 & 0.2 & 0.3 & 2.96 & 72.4 & 1.54 & 245 & 0.109 \\
8 & 1 & 6 & 18.2 & 43900 & 6.07 & 47000 & 0.463 & 43 & 0.2 & 0.5 & 3.33 & 190 & 1.76 & 368 & 0.129 \\
9 & 1 & 12 & 25.1 & 178000 & 8 & 109000 & 0.544 & 44 & 0.2 & 1 & 4.04 & 729 & 2.1 & 701 & 0.163 \\
10 & 1 & 24 & 29.6 & 452000 & 9.75 & 180000 & 0.543 & 45 & 0.2 & 1.5 & 4.61 & 1630 & 2.39 & 1070 & 0.187 \\
11 & 0.8 & 0.1 & 2.51 & 11.2 & 1.2 & 245 & 0.114 & 46 & 0.2 & 2 & 5.09 & 2930 & 2.56 & 1480 & 0.205 \\
12 & 0.8 & 0.3 & 3.89 & 73.5 & 1.76 & 651 & 0.168 & 47 & 0.2 & 3 & 5.92 & 6840 & 2.9 & 2390 & 0.240 \\
13 & 0.8 & 0.5 & 4.64 & 192 & 2.1 & 1120 & 0.203 & 48 & 0.2 & 6 & 7.93 & 31600 & 3.58 & 5890 & 0.301 \\
14 & 0.8 & 1 & 6.19 & 751 & 2.73 & 2620 & 0.261 & 49 & 0.2 & 12 & 10.4 & 129000 & 4.54 & 13500 & 0.392 \\
15 & 0.8 & 1.5 & 7.6 & 1780 & 3.12 & 4690 & 0.305 & 50 & 0.2 & 24 & 12.6 & 359000 & 5.56 & 24500 & 0.439 \\
16 & 0.8 & 2 & 8.88 & 3390 & 3.46 & 7210 & 0.339 & 51 & 0.1 & 0.1 & 2.44 & 10.5 & 1.2 & 121 & 0.079 \\
17 & 0.8 & 3 & 11.1 & 8660 & 4.14 & 13100 & 0.386 & 52 & 0.1 & 0.3 & 2.95 & 71.3 & 1.54 & 243 & 0.110 \\
18 & 0.8 & 6 & 16.3 & 41700 & 5.67 & 35000 & 0.463 & 53 & 0.1 & 0.5 & 3.29 & 188 & 1.76 & 358 & 0.129 \\
19 & 0.8 & 12 & 22.9 & 183000 & 7.77 & 84500 & 0.531 & 54 & 0.1 & 1 & 3.92 & 722 & 2.16 & 652 & 0.164 \\
20 & 0.8 & 24 & 27.9 & 475000 & 9.07 & 147000 & 0.560 & 55 & 0.1 & 1.5 & 4.41 & 1620 & 2.44 & 964 & 0.190 \\
21 & 0.5 & 0.1 & 2.63 & 11.4 & 1.14 & 168 & 0.095 & 56 & 0.1 & 2 & 4.81 & 2890 & 2.61 & 1290 & 0.208 \\
22 & 0.5 & 0.3 & 3.38 & 74.1 & 1.54 & 407 & 0.140 & 57 & 0.1 & 3 & 5.53 & 6840 & 2.9 & 2050 & 0.244 \\
23 & 0.5 & 0.5 & 3.94 & 191 & 1.82 & 674 & 0.169 & 58 & 0.1 & 6 & 7.13 & 33900 & 3.52 & 4850 & 0.317 \\
24 & 0.5 & 1 & 5.11 & 736 & 2.33 & 1500 & 0.223 & 59 & 0.1 & 12 & 8.96 & 149000 & 4.31 & 10300 & 0.376 \\
25 & 0.5 & 1.5 & 6.11 & 1660 & 2.67 & 2510 & 0.259 & 60 & 0.1 & 24 & 10.5 & 452000 & 5.16 & 17700 & 0.408 \\
26 & 0.5 & 2 & 7.03 & 3050 & 2.95 & 3740 & 0.289 & 61 & 0 & 0.1 & 2.47 & 10.2 & 1.2 & 127 & 0.081 \\
27 & 0.5 & 3 & 8.65 & 7500 & 3.46 & 6670 & 0.334 & 62 & 0 & 0.3 & 2.98 & 70.3 & 1.59 & 256 & 0.113 \\
28 & 0.5 & 6 & 12.6 & 36600 & 4.6 & 17900 & 0.407 & 63 & 0 & 0.5 & 3.33 & 185 & 1.82 & 377 & 0.134 \\
29 & 0.5 & 12 & 18.3 & 172000 & 6.3 & 46000 & 0.464 & 64 & 0 & 1 & 3.96 & 715 & 2.22 & 682 & 0.168 \\
30 & 0.5 & 24 & 23 & 485000 & 7.49 & 83300 & 0.519 & 65 & 0 & 1.5 & 4.44 & 1600 & 2.5 & 1010 & 0.196 \\
31 & 0.3 & 0.1 & 2.46 & 11 & 1.14 & 124 & 0.077 & 66 & 0 & 2 & 4.83 & 2890 & 2.67 & 1350 & 0.214 \\
32 & 0.3 & 0.3 & 3.04 & 73.4 & 1.48 & 268 & 0.109 & 67 & 0 & 3 & 5.54 & 6950 & 2.95 & 2150 & 0.252 \\
33 & 0.3 & 0.5 & 3.46 & 191 & 1.71 & 420 & 0.130 & 68 & 0 & 6 & 6.98 & 34900 & 3.63 & 4910 & 0.326 \\
34 & 0.3 & 1 & 4.3 & 733 & 2.1 & 870 & 0.171 & 69 & 0 & 12 & 8.46 & 158000 & 4.43 & 9970 & 0.390 \\
35 & 0.3 & 1.5 & 5.03 & 1630 & 2.39 & 1420 & 0.215 & 70 & 0 & 24 & 9.65 & 584000 & 5.16 & 16400 & 0.421 \\
\hline
\vspace{0.05cm}
\end{tabular}
\end{table*}
Throughout this work, $\langle R_A\rangle$ is used as a normalised torque which accounts for the mass loss rates which we do not control. Values of the average Alfv\'en radius are tabulated within Table \ref{Parameters}. $\langle R_A \rangle$ is shown in Figure \ref{streamers} using a grey vertical dashed line. For each case, the cylindrical Alfv\'en radius is offset inwards of the maximum Alfv\'en radius from the simulation, a geometrical effect as this corresponds to the average cylindrical $R_A$ and includes variations in flow quantities as well. Exploring Figure \ref{streamers}, the motion of the deadzones/current sheets have little impact on the overall torque. For example, no abrupt increase in the Alfv\'en radius is seen from case 34 to 36 (where the southern streamer is forced onto the rotation axis) compared to cases 44 and 46. The torque is instead governed by the magnetic field strength in the wind which controls the location of the Alfv\'en surface.
We parametrise the magnetic and mass loss properties using the ``wind magnetisation'' defined by,
\begin{equation}
\Upsilon=\frac{B^2_*R_*^2}{\dot{M}v_{esc}},
\end{equation}
where $B_*$ is the combined field strength at the pole. Previous studies that used this parameter defined it with the equatorial field strength (e.g. \citealp{matt2008accretion}; \citealp{matt2012magnetic}; \citealp{reville2015effect}; Pantolmos \& Matt. in prep). We use polar values unlike previous authors due to the additive property of the radial field at the pole, for aligned axisymmetric fields. Note that selecting one value of the field on the surface will not always produce a value which describes the field as a whole. The polar strength works for these aligned fields, but will easily break down for un-aligned fields and anti-aligned axisymmetric odd $l$ fields, thus it suits the present study, but a move away from this parameter in future is warranted.
During analysis, the wind magnetisation, $\Upsilon$, is treated as an independent parameter that determines the Alfv\'en radius $\langle R_A\rangle$ and thus the torque, $\tau$. We increase $\Upsilon$ by setting a larger $v_A/v_{esc}$, creating a stronger global magnetic field. Table \ref{Parameters} displays all the input values of $\mathcal{R}_{dip}$ and $v_A/v_{esc}$ as well as the resulting global outflow properties from our steady state solutions, which are used to formulate the torque scaling relations within this study. Figure \ref{PS} displays all 70 simulations in $\Upsilon-\mathcal{R}_{dip}$ space. Cases are colour-coded here by their $\mathcal{R}_{dip}$ value, a convention which is continued throughout this work.
\subsection{Single Mode Torque Scalings}
\begin{table*}
\caption{Best Fit Parameters to equations (\ref{UP_OLD}) and (\ref{UP_VA})}
\label{fitValues}
\center
\begin{tabular}{ccc|ccc}
\hline\hline
Topology($l)$ & $K_s$ & $m_s$ &$K_l$ &$m_l$& $m_{l,th}(l)$ \\ \hline
Dipole ($1$) & $1.49\pm0.03$ & $0.231\pm0.003$ &$0.92\pm0.04$ &$0.258\pm0.005$ &0.250 \\
Quadrupole ($2$) & $1.72\pm0.03$ & $0.132\pm0.003$ &$1.11\pm0.04$&$0.156\pm0.004$&0.167 \\
\hline
\end{tabular}
\end{table*}
\begin{figure*}
\includegraphics[width=\textwidth]{f5.eps}
\caption{Average Alfv\'en radius vs wind magnetisation for all cases. Simulations are marked with colour-coded circles indicating their $\mathcal{R}_{dip}$ value. Left: Solid lines show the fit of dipole (red) and quadrupole (blue) to equation (\ref{UP_OLD}). Dashed lines show the dipolar component fit, equation (\ref{UP_DIPOLE}). Right: Solid lines show the analytic solution of dipole (red) and quadrupole (blue) to equation (\ref{UP_VA}) with $K_l=1$. Dashed lines show the dipolar component fit from equation (\ref{UP_DIPOLE2}), dependent on only the value of the field order $l$, unlike in the $\Upsilon$ space. }
\label{RstudyUpsilon}
\end{figure*}
The efficiency of the magnetic braking mechanism is known to be dependent on the magnetic field geometry. This has been previsously shown for single mode geometries \citep[e.g.][]{reville2015effect, garraffo2016missing}. We first concider two pure gemetries, dipole and quadrupole, using the formulation from \cite{matt2008accretion},
\begin{equation}
\frac{\langle R_A\rangle}{R_*} = K_{s} \Upsilon^{m_{s}},
\label{UP_OLD}
\end{equation}
where $K_s$ and $m_s$ are fitting parameters for the pure dipole and quadrupole cases, using the surface field strength. Here we empirically fit $m_s$; the interpretation of $m_s$ is discussed in \cite{matt2008accretion}, \cite{reville2015effect} and Pantolmos \& Matt (in prep), where it is determined to be dependant on magnetic geometry and the wind acceleration profile. The Appendix contains further discussion of the wind acceleration profile and its impact on this power law relationship.
The left panel of Figure \ref{RstudyUpsilon} shows the Alfv\'en radii vs the wind magnetisations for all cases (colour-coded with their $\mathcal{R}_{dip}$ value). Solid lines show scaling relations for dipolar (red) and quadrupolar (blue) geometries, as first shown in \cite{reville2015effect}. We calculate best fit values for $K_s$ and $m_s$ for the dipole and quadrupole, tabulated in Table \ref{fitValues}. Values here differ due to our hotter wind ($c_s/v_{esc}=0.25$ than their $c_s/v_{esc}=0.222$), using polar $B_*$, and we do not account for our low rotation rate. As previously shown, the dipole field is far more efficient at transferring angular momentum than the quadrupole. In this study we concider the effect of combined geometries, within Figure \ref{RstudyUpsilon} these cases lie between the dipole and quadrupole slopes, with no single power law of this form to describe them.
Pantolmos \& Matt (in prep) have shown the role of the velocity profile in the power law dependence of the torque. In our simulations, the acceleration of the flow from the base wind velocity to its terminal speed is primarily governed by the thermal pressure gradient, however magnetic topologies can all modify the radial velocity profile (as can changes in wind temperature, $\gamma$, and rapid rotation, not included in our study). Effects on the torque formulations due to these differences in acceleration can be removed via the multiplication of $\Upsilon$ with $v_{esc}/\langle v(R_A)\rangle$. In their work, the authors determine the theoretical power law dependence, $m_{l,th}=1/(2l+2)$, from one-dimensional analysis. In this formulation the slope of the power law is controlled only by the order of the magnetic geometry, $l$, which is $l=1$ and $l=2$ for the dipole and quadrupole respectively,
\begin{equation}
\frac{\langle R_A\rangle}{R_*} = K_l\bigg[\Upsilon\frac{ v_{esc}}{\langle v(R_A)\rangle}\bigg]^{m_l},
\label{UP_VA}
\end{equation}
where $K_l$ and $m_l$ are fit parameters to our wind solutions, tabulated in Table \ref{fitValues}. The value of $\langle v(R_A)\rangle$ is calculated as an average of the velocity at all points on the Alfv\'en surface in the meridional plane. \footnote{It could be argued that this should be weighted by the total area of the Alfv\'en surface, but for simplicity we calculate the un-weighted average. }
Equation (\ref{UP_VA}) is able to predict accurately the power law dependence for the two pure modes using the order of the spherical harmonic field, $l$. We show this in the right panel of Figure \ref{RstudyUpsilon}, where the Alfv\'en radii are plotted against the new parameter, $\Upsilon v_{esc}/\langle v(R_A)\rangle$. A similar qualitative behaviour is shown to the scaling with $\Upsilon$ in the left panel. Using the theoretical power law dependencies, the dipolar (red) and quadrupolar (blue) slopes are plotted with $m_{l,th}=1/4$ and $m_{l,th}=1/6$ respectively. Using a single fit constant $K_l=1$ for both sloes within this figure shows good agreement with the simulation results.
More accurate values of $K_l$ and $m_l$ are fit for each mode independently. These values produce a better fit and are compared with the theoretical values in Table \ref{fitValues}. The mixed simulations show a similar qualitative behaviour to the plot against $\Upsilon$.
Obvious trends are seen within the mixed case scatter. A saturation to quadrupolar Alfv\'en radii values for lower $\Upsilon$ and $\mathcal{R}_{dip}$ values is observed, along with a power law trend with a dipolar gradient for higher $\Upsilon$ and $\mathcal{R}_{dip}$ values. This indicates that both geometries play a role in governing the lever arm, with the dipole dominating the braking process at higher wind magnetisations.
\subsection{Broken Power Law Scaling For Mixed Field Cases}
\begin{figure}
\includegraphics[width=0.5\textwidth]{f6.eps}
\caption{Average Alfv\'en radius vs the dipolar wind magnetisation. Considering only the dipolar field strength, we produce a single power law for the Alf\'ven radius, equation (\ref{UP_DIPOLE}). Our wind solutions are shown to agree well with dipole prediction in most cases. Disagreement at low $\Upsilon_{dip}$ and $\mathcal{R}_{dip}$ values are explained by the quadrupolar slopes, shown in coloured dashed lines.}
\label{Upsilon_dip}
\end{figure}
Observationally the field geometries of cool stars are, at large scales, dominated by the dipole mode with higher order $l$ modes playing smaller roles in shaping the global field. It is the global field which controls the spin-down torque in the magnetic braking process. Higher order modes (such as the quadrupole) decay radially much faster than the dipole and as such they have a reduced contribution to setting the Alfv\'en speed at distances larger than a few stellar radii.
We calculate $\Upsilon_{dip}$, which only takes into account the dipole's field strength,
\begin{equation}
\Upsilon_{dip}=\bigg(\frac{B^{l=1}_*}{B_{*}}\bigg)^2\frac{B^2_{*}R_*^2}{\dot{M}v_{esc}}=\mathcal{R}_{dip}^2\Upsilon.
\end{equation}
Taking as a hypothesis that the field controlling the location of the Alfv\'en radius is the dipole component, a power law scaling using $\Upsilon_{dip}$ can be constructed in the same form as Matt \& Pudritz (2008),
\begin{equation}
\frac{\langle R_A\rangle}{R_*} = K_{s,dip}[\Upsilon_{dip}]^{m_{s,dip} } = K_{s,dip}[\mathcal{R}_{dip}^2\Upsilon]^{m_{s,dip} }.
\label{UP_DIPOLE}
\end{equation}
Substitution of the dipole component into equation (\ref{UP_VA}) similarly gives,
\begin{equation}
\frac{\langle R_A\rangle}{R_*} = K_{l,dip}\bigg[\mathcal{R}_{dip}^2\Upsilon_{}\frac{ v_{esc}}{\langle v(R_A)\rangle}\bigg]^{m_{l,dip}},
\label{UP_DIPOLE2}
\end{equation}
where $K_{s,dip}$, $m_{s,dip}$, $K_{l,dip}$, and $m_{l,dip}$ will be parameters fit to simulations.
A comparison of these approximations can be seen in Figure \ref{RstudyUpsilon}, where equations (\ref{UP_DIPOLE}) (left panel) and (\ref{UP_DIPOLE2}) (right panel) are plotted with dashed lines for all the $\mathcal{R}_{dip}$ values used in our simulations. Mixed cases which lie above the quadrupolar slope are shown to agree with the dashed-lines in both forms. Such cases are dominated by the dipole component of the field only, irrespective of the quadrupolar component.
The role of the dipole is even more clear in Figure \ref{Upsilon_dip} where only the dipole component of $\Upsilon$ is plotted for each simulation. The solid red line in Figure \ref{Upsilon_dip}, given by equation (\ref{UP_DIPOLE}), shows agreement at a given $\mathcal{R}_{dip}$ with deviation from this caused by a regime change onto the quadrupolar slope (shown in dashed colour).
The behaviour of our simulated winds, despite using a combination of field geometries, simply follow existing scaling relations with this modification. In general, the dipole ($\Upsilon_{dip}$) prediction shows good agreement with the simulated wind models, except in cases where the Alfv\'en surface is close-in to the star. In these cases, the quadrupole mode still has magnetic field strength able to control the location of the Alfv\'en surface. Interestingly, and in contrast to the dipole-dominated regime, the quadrupole dominated regime behaves as if all the field strength is within the quadrupolar mode. This is visible within Figure \ref{RstudyUpsilon} for low values of $\Upsilon$ and $\mathcal{R}_{dip}$.
The mixed field $\langle R_A \rangle$ scaling can be described as a broken power law, set by the maximum of either the dipole component or the pure quadrupolar relation. With the break in the power law given by $\Upsilon_{crit}$,
\begin{equation}
\frac{\langle R_A \rangle}{R_*}=\left\{
\begin{array}{@{}ll@{}}
K_{s,dip}[\mathcal{R}_{dip}^2\Upsilon]^{m_{s,dip}}, & \text{if}\ \Upsilon>\Upsilon_{crit}(\mathcal{R}_{dip}), \\
K_{s,quad}[\Upsilon]^{m_{s,quad}}, & \text{if}\ \Upsilon\leq\Upsilon_{crit}(\mathcal{R}_{dip})
\end{array}\right.
\label{BrokenPowerLaw}
\end{equation}
where $\Upsilon_{crit}$ is the location of the intercept for the dipole component and pure quadrupole scalings,
\begin{equation}
\Upsilon_{crit}(\mathcal{R}_{dip})=\bigg[\frac{K_{s,dip}}{K_{s,quad}}\mathcal{R}_{dip}^{2m_{s,dip}} \bigg]^{\frac{1}{m_{s,quad}-m_{s,dip}}}.
\label{UP_CRIT}
\end{equation}
The solid lines in Figure \ref{PS} show the value of $\Upsilon_{crit}$, equation (\ref{UP_CRIT}), diving the two regimes. Specifically, the solutions above the solid black line behave as if only the dipole component ($\Upsilon_{dip}$) is governing the Alfv\'en radius.
Transitioning from regimes is not perfectly abrupt. Therefore producing an analytical solution for the mixed cases which includes this behaviour would increase the accuracy for stars near the regime change. E.g. we have formulated a slightly better fit, using a relationship based on the quadrature addition of different regions of field. However it provides no reduction to the error on this simpler form and is not easily generalised to higher topologies. For practical purposes, the scaling of equation (\ref{BrokenPowerLaw}) and (\ref{UP_CRIT}) predict accurately the simulation torque with increasing magnetic field strength for a variety of dipole fractions. We therefore present the simplest available solution, leaving the generalised form to be developed within future work.
\begin{figure*}
\includegraphics[width=\textwidth]{f7.png}
\caption{Magnetic flux vs Radial distance for all cases studied within this work, compared with analytical predictions. Solid grey lines show the 10 simulation fields for each field geometry. Solutions of equation (\ref{phi}) for the potential field magnetic flux are shown in black solid lines for each $\mathcal{R}_{dip}$ value. In each case, the flux of dipole and quadrupole components using a potential field are plotted with dashed red and blue respectively, equations (\ref{quad_flux}) and (\ref{dip_flux}). Each simulation matches the potential field flux, until the wind pressures open the field to a constant flux. The open flux radii are displayed as grey circles. The lower right panel shows a comparison of each potential field flux decay along with the opening radii for each case (i.e. the solid black lines and grey circles from the other panels), colour-coded to the value of $\mathcal{R}_{dip}$.}
\label{RstudyRaRo}
\end{figure*}
\section{The impact of geometry on the magnetic flux in the wind}
\subsection{Evolution of the Flux}
The magnetic flux in the wind is a useful diagnostic tool. The rate of the stellar flux decay with distance is controlled by the overall magnetic geometry. We calculate the magnetic flux as a function of radial distance by evaluating the integral of the magnetic field threading closed spherical shells, where we take the absolute value of the flux to avoid field polarity cancellations,
\begin{equation}
\Phi(r)=\oint_r|{\bf B} \cdot d{\bf A}|.
\label{phi}
\end{equation}
Considering the initial potential fields of the two pure modes this is simply a power law in field order $l$,
\begin{equation}
\Phi(r)_P=\Phi_*\bigg(\frac{R_*}{r}\bigg)^l,
\label{potential}
\end{equation}
where $l=1$ dipole and $l=2$ quadrupole, we denote the flux with ``$P$'' for the potential field. Figure \ref{RstudyRaRo} displays the flux decay of all values of $v_A/v_{esc}$ for each $\mathcal{R}_{dip}$ value, grey lines. The behaviour is qualitatively identical to that observed within previous works \citep[e.g.][]{schrijver2003asterospheric, johnstone2010modelling, vidotto2014m, reville2015effect}, where the field decays as the potential field does until the pressure of the wind forces the field into a purely radial configuration with a constant magnetic flux, referred to as the open flux. The power law dependence of equation (\ref{potential}) indicates for higher $l$ mode magnetic fields, the decay will be faster. We therefore expect the more quadrupolar dominated fields studied in this work to have less open flux.
In the case of mixed geometries a simple power law is not available for the initial potential configurations, instead we evaluate the flux using equation (\ref{phi}), where $\bf B$ is the initial potential field for each mixed geometry. This allows us to calculate the radial evolution of the flux for a given $\mathcal{R}_{dip}$ which we compare to the simulated cases. Figure \ref{RstudyRaRo} shows the flux normalised by the surface flux versus radial distance from the star. For each $\mathcal{R}_{dip}$ value, the magnetic flux decay of the potential field (black solid line) is shown with the different strength $v_A/v_{esc}$ simulations (grey solid lines). A comparison of the flux decay for all potential magnetic geometries is available in the bottom right panel showing, as expected, the increasingly quadrupolar fields decaying faster.
In this study we control $v_A/v_{esc}$ which, for a given surface density, sets the polar magnetic field strength for our simulations. The stellar flux for different topologies and the same $B_*$ will differ and must be taken into account in order to describe the dipole and quadrupolar components (dashed red and blue) in Figure \ref{RstudyRaRo}. We plot the magnetic flux of the potential field quadrupole component alone in dotted blue for each $\mathcal{R}_{dip}$ value,
\begin{equation}
\Phi(r)_{P,quad}=(1-\mathcal{R}_{dip})\Phi_{*,quad}\bigg(\frac{R_*}{r}\bigg)^2,
\label{quad_flux}
\end{equation}
and similarly the potential field dipole component of the magnetic flux,
\begin{equation}
\Phi(r)_{P,dip}=\mathcal{R}_{dip}\Phi_{*,dip}\bigg(\frac{R_*}{r}\bigg),
\label{dip_flux}
\end{equation}
where in both equations the surface flux of a pure dipole/quadrupole ($\Phi_{*,dip}$, $\Phi_{*,quad}$) field is required to match our normalised flux representation.
Due to the rapid decay of the quadrupolar mode, the flux at large radial distances for all simulations containing the dipole mode is described by the dipolar component. The quadrupole component decay sits below and parallel to the potential field prediction for small radii, becoming indistinguishable for the lowest $\mathcal{R}_{dip}$ values as the flux stored in the dipole is decreased. Importantly for small radii, simulations containing a quadrupolar component are dominated by the quadrupolar decay following a $l=2$ power law decay, which can be seen by shifting the blue dashed line upwards to intercept $\Phi/\Phi_*=1$ at the stellar surface.
This result for the flux decay is reminiscent of the broken power law description for the Alfv\'en radius in Section 3.4. The field acts as a quadrupole using the total field for small radii and the dipole component only for large radii. There is a transition between these two regimes that is not described by either approximation. But is shown by the potential solution in solid black.
\subsection{Topology Independent Open Flux Formulation}
The magnetic flux within the wind decays following the potential field solution closely until the magnetic field geometry is opened by the pressures of the stellar wind and the field lines are forced into a nearly radial configuration with constant flux, shown in Figure \ref{RstudyRaRo} for all simulations. The importance of this open flux is discussed by \cite{reville2015effect}. These authors showed a single power law dependence for the Alfv\'en radius, independent of magnetic geometry, when parametrised in terms of the open flux, $\Phi_{open}$,
\begin{equation}
\Upsilon_{open}=\frac{\Phi_{open}^2/R_*^2}{\dot{M}v_{esc}},
\end{equation}
which, ignoring the effects of rapid rotation, can be fit with,
\begin{equation}
\frac{\langle R_A\rangle}{R_*} = K_o[\Upsilon_{open}]^{m_o},
\label{UP_OPEN_OLD}
\end{equation}
where, $m_o$ and $K_o$ are fitting parameters for the open flux formulation.
\begin{table*}
\caption{Open Flux Best Fit Parameters to equations (\ref{UP_OPEN_OLD}) and (\ref{UP_OPEN})}
\label{fitValues_open}
\center
\begin{tabular}{ccccc}
\hline\hline
Topology($l)$ & $K_o$& & $m_o$ \\ \hline
Dipole ($1$) & $0.37\pm0.05$ && $0.360\pm0.006$ \\
Quadrupole ($2$) & $0.62\pm0.01$ & &$0.283\pm0.002$ \\
\hline
& $K_c$ & $K_{c,th}$& $m_c$ & $m_{c,th}$ \\ \hline
Topology Independent & $0.08\pm0.03$&$ 0.0796$&$0.471\pm0.003$&0.500\\
\hline
\end{tabular}
\end{table*}
\begin{figure*}
\includegraphics[width=\textwidth]{f8.eps}
\caption{Left: Average Alfv\'en radius vs open flux magnetisation for all cases. Fits to equation (\ref{UP_OPEN_OLD}) are shown for the dipole ($\mathcal{R}_{dip}=1$) and quadrupole ($\mathcal{R}_{dip}=0$) fields. The geometry of the field is shown to influence the scaling relation, due to differences in the wind acceleration. Right: Average Alfv\'en radius vs open flux magnetisation accounting for the acceleration profile using work done by Pantolmos \& Matt (in prep). The fit of equation (\ref{UP_OPEN}) is shown to reduce the scatter for all simulations. A systematic discrepancy is still seen from the single power law with changing geometry.}
\label{RstudyUpsilonOpen}
\end{figure*}
Using the open flux parameter, Figure \ref{RstudyUpsilonOpen} shows a collapse towards a single power law dependence as in \cite{reville2015effect}. However our wind solutions show a systematic difference in power law dependence from dipole to quadrupole. On careful inspection of the result from Figure 6 of \cite{reville2015effect}, the same systematic trend between their topologies and the fit scaling is seen. \footnote{A choice in our parameter space may have made this clearer to see in Figure \ref{RstudyUpsilonOpen}, due to the increased heating and therefore larger range of acceleration allowing the topology to impact the velocity profile.} We calculate best fits for each pure mode separately i.e. the dipole and quadrupole, tabulated in Table \ref{fitValues_open}.
Pantolmos \& Matt (in prep) find solutions for thermally driven winds with different coronal temperatures, from these they find the wind acceleration profiles of a given wind to very significantly alter the slope in $R_A$-$\Upsilon_{open}$ space. From this work our trend with geometry indicates that each geometry must have a slightly different wind acceleration profile. This is most likely due to difference in the super radial expansion of the flux tubes for each geometry, which is not taken into account with equation (\ref{UP_OPEN_OLD}). The field geometry is imprinted onto the wind as it accelerates out to the Alfv\'en surface. As such, this scaling relation is not entirely independent of topology. Further details on the wind acceleration profile within our study is available in the Appendix. Pantolmos (in prep) are able to include the effects of acceleration in their scaling through multiplication of $\Upsilon_{open}$ with $v_{esc}/\langle v(R_A)\rangle$. The expected semi-analytic solution from Pantolmos \& Matt (in prep) is given,
\begin{equation}
\frac{\langle R_A\rangle}{R_*} = K_c\bigg[\Upsilon_{open}\frac{ v_{esc}}{\langle v(R_A)\rangle}\bigg]^{m_c},
\label{UP_OPEN}
\end{equation}
where the fit parameters are derived from one-dimensional theory as constants, $K_{c,th}=1/4\pi$ and $m_{c,th}=1/2$.
We are able to reproduce this power law fit of $\Upsilon_{open}$ with the wind acceleration effects removed, on the right panel of Figure \ref{RstudyUpsilonOpen}. Including all simulations in the fit, we arrive at values of $K_c=1.01K_{c,th}\pm0.07$ and $m_c=0.942m_{c,th}\pm0.009 $ for the constants of proportionality and power law dependence. However a systematic difference is still seem from one $\mathcal{R}_{dip}$ value to another. More precise fits can be found for each geometry independently, but the systematic difference appearing in the right panel implies a modification to our semi-analytic formulations is required to describe the torque fully in terms of the open flux.
Here we show the scaling law from \cite{reville2015effect} is improved with the modification from Pantlomos (in prep). This formulation is able to describe the Alfv\'en radius scaling with changing open flux and mass loss. However with the open flux remaining an unknown from observations and difficult to predict, scaling laws that incorporate known parameters (such as those of equations (\ref{BrokenPowerLaw}) and (\ref{UP_CRIT})) are still needed for rotational evolution calculations.
\subsection{The Relationship Between the Opening and Alfv\'en Radii}
\begin{figure}
\includegraphics[width=0.5\textwidth]{f9.eps}
\caption{Alfv\'en radii vs opening radii for all simulated cases. Black dashed line represent $R_A/R_o=3.2$ and $1.7$. Different geometries have a changing relationship between the torque lever arm and the opening radius of the field.}
\label{RstudyRaRoAll}
\end{figure}
The location of the field opening is an important distance. It is both critical for determining the torque and for comparison to potential field source surface (PFSS) models \citep[][]{altschuler1969magnetic}, which set the open flux with a tunable free parameter $R_{ss}$. The opening radius, $R_o$, we define is the radial distance at which the potential flux reaches the value of the open flux ($\Phi_P(R_o)=\Phi_{open}$). This definition is chosen because it relates to the 1D analysis employed to describe the power law dependences of our torque scaling relations. Specifically, a known value of $R_o$ allows for a precise calculation of the open flux (a priori from the potential field equations), which then gives the torque on the star within our simulations. The physical opening of the simulation field takes place at slightly larger radii than this with the field becoming non-potential due to its interaction with the wind (which explains why the closed field regions seen in Figure \ref{streamers} typically extend slightly beyond $R_o$). A similar smooth transition is produced with PFSS modelling.”
$R_o$ is marked for each simulation in Figure \ref{RstudyRaRo} and again for comparative purposes in the bottom right panel. It is clear that smaller opening radii are found for lower $\mathcal{R}_{dip}$ cases. Due to their more rapidly decaying flux, they tend to have a smaller fraction of the stellar flux remaining in the open flux. From the radial decay of the magnetic field, the open flux and opening radii are observed to be dependent on the available stellar flux and topology. Pantolmos \& Matt (in prep) have recently shown these to also be dependent on the wind acceleration profile. This complex dependence makes it difficult to predict the open flux for a given system.
Our simulations produce values for the average Alfv\'en radius, $\langle R_A\rangle$, and the opening radius, $R_o$, for the 7 different geometries studied. It is interesting to consider the relative size of these radii as they both characterise key dynamic properties for each stellar wind solution. For all cases shown in Figure \ref{streamers}, the opening radii are plotted in dashed red, allowing for the relative size to be compared with the cylindrical Alfv\'en radius, shown in dash grey. With increasing magnetic field strength ($\Upsilon$), both radii are seen to grow from case to case, however with increasing $\mathcal{R}_{dip}$, the cylindrical Alfv\'en radius generally grows faster than the opening radius. To quantify this, Figure \ref{RstudyRaRoAll} shows a plot of the Alfv\'en radii vs the opening radii for all cases. Linear trends of $R_A/R_o=3.2$ and $1.7$ are indicated with dashed lines. For each $\mathcal{R}_{dip}$ value, the relationship between the Alfv\'en and opening radius ($\langle R_A\rangle/R_o$) is seen to systematically decrease with increasing higher order field component. In all cases, for small radii a shallower slope is observed which then steepens with increasing radial extent.
The dependence of the Alfv\'en radius and opening radius on field geometry and magnetisation is a constraint on PFSS models, which are readily used with ZDI observations as a less computationally expensive alternative to MHD modelling \citep{jardine1999potential, jardine2002global, dunstone2008first, cohen2010coronal, johnstone2010modelling, rosen2015first, reville2015solar}. PFSS models are a useful tool, however require the source surface radius, $R_{ss}$, as an input. Authors often set a source surface and change the geometry and strength of the field freely \citep{fares2010searching, see2015energy, see2016studying}. We find however for a given $\mathcal{R}_{dip}$ value there exists a differing relation for the opening radius, as we define it here, to the Alfv\'en radius and magnetisation. These trends are observed to continue for higher $l$ mode fields (Finley \& Matt. in prep), with $\langle R_A\rangle/R_o$ decreasing overall with increased field complexity. As such, our results confirm that the opening radius should not remain fixed when changing geometries or increasing the wind magnetisation. We find the relationship of $\langle R_A\rangle/R_o$ to change in both cases. With fixed magnetisation, the opening radius should move towards the star for higher order fields to maintain a constant thermal driving. Maintaining the opening radius whilst increasing the field complexity infers that the wind has a reduced acceleration. Similarly with increased wind magnetisation the opening radius should move further from the star. The value of $R_o$ as we have defined it, is directly related to the source surface radius, and for a given magnetic geometry, the two should scale approximately together. For example, for a dipole field, comparing our definition of $R_o$ to the PFSS model shows that $R_{ss}$ equals an approximately constant value of 3/2 $R_o$. Thus conclusions made about the opening radii are constraints on future PFSS modelling.
A method for predicting $R_o$ within our simulations remains unknown, however it is understood that $R_o$ is key to predicting the torque from our simulated winds. We do however find the ratio of $\langle R_A\rangle/R_o$ to be roughly constant for a given geometry, deviations from which may be numerical or suggest additional physics which we do not explore here.
\section{Conclusion}
We undertake a systematic study of the two simplest magnetic geometries, dipolar and quadrupolar, and for the first time their combinations with varying relative strengths. We parametrise the study using the ratio, $\mathcal{R}_{dip}$, of dipolar to total combined field strength, which is shown to be a key variable in our new torque formulation.
We have shown that a large proportion of the magnetic field energy needs to be in the quadrupole for any significant morphology changes to be seen in the wind. All cases above 50\% dipole field show a single streamer and are dominated by dipolar behaviour. Even in cases of small $\mathcal{R}_{dip}$ we observe the dipole field to be the key parameter controlling the morphology of the flow, with the quadrupolar field rapidly decaying away for most cases leaving the dipole component behind. For smaller field strengths the Alfv\'en radii appears close to the star, where the quadrupolar field is still dominant, and thus a quadrupolar morphology is established. Increasing the fraction of quadrupolar field strength allows this behaviour to continue for larger Alfv\'en radii.
The morphology of the wind can be concidered in the context of star-planet or disk interactions. Our findings suggest that the connectivity, polarity and strength of the field within the orbital plane depend in a simple way on the relative combination of dipole and quadrupole fields. Different combinations of these two field modes change the location of the current sheet(s) and the relative orientation of the stellar wind magnetic field with respect to any planetary or disk magnetic field. Asymmetries such as these can modify the poynting flux exchange for close-in planets \citep{strugarek2014diversity} or the strength of magnetospheric driving and geomagnetic storms on Earth-like exoplanets. \cite{cohen2014magnetospheric} use observed magnetic fields to simulate the stellar wind environment surrounding the planet hosting star EV Lac. They calculate the magnetospheric joule heating on the exoplanets orbiting the M dwarf, finding significant changes to atmospheric properties such as thickness and temperature. Additionally, transient phenomena in the Solar wind such as coronal mass ejections are shown to deflect towards streamer belts \citep{kay2013forecasting}. This has been applied to mass ejections around M dwarfs stars \citep{kay2014implications}, and could similarly be applied here using knowledge of the streamer locations from our model grid.
If the host star magnetic field can be observed and decomposed into constituent field modes, containing dominant dipole and quadrupole components, a qualitative assessment of the stellar wind environment can be made. We find the addition of these primary and secondary fields to create an asymmetry which may shift potentially habitable exoplanets in and out of volatile wind streams. Observed planet hosting stars such as $\tau$ Bootis have already been shown to have global magnetic fields which are dominated by combinations of these low order field geometries \citep{donati2008magnetic}. With further investigation it is possible to qualitatively approximate the conditions for planets in orbit of such stars. For dipole and quadrupole dominated host stars with a given magnetic field strength our grid of models provide an estimate of the location of the streamers and open field regions.
Within this work we build on the scaling relations from, \cite{matt2012magnetic}, \cite{reville2015effect} and Pantolmos \& Matt (in prep). We confirm existing scaling laws and explore a new mixed field parameter space with similar methods. From our wind solutions we fit the variables, $K_{s,dip}$, $m_{s,dip}$, $K_{s,quad}$ and $m_{s,quad}$ (see Table \ref{qandms}), which describe the torque scaling for the pure dipole and quadrupole modes. From the 50 mixed case simulations, we produce an approximate scaling relation which takes the form of a broken power law, as a single power law fit is not available for the mixed geometries cases in $\Upsilon$ space.
For low $\Upsilon$ and low dipole fraction, the Alfv\'en radius behaves like a pure quadrupole,
\begin{equation}
\tau = K_{s,quad} \dot{M} \Omega_*R_*^2[ \Upsilon]^{2m_{s,quad}} ,
\end{equation}
\begin{equation}
= K_{s,quad} \dot{M}^{1-2m_{s,quad}} \Omega_*R_*^{2+4m_{s,quad}}\bigg[\frac{(B_{*})^2}{v_{esc}}\bigg]^{2m_{s,quad}}.
\end{equation}
At higher $\Upsilon$ and dipole fractions, the torque is only dependent on the dipolar component of the field,
\begin{equation}
\tau = K_{s,dip} \dot{M} \Omega_*R_*^2[ \Upsilon_{dip}]^{2m_{s,dip}} ,
\end{equation}
\begin{equation}
= K_{s,dip} \dot{M}^{1-2m_{s,dip}} \Omega_*R_*^{2+4m_{s,dip}}\bigg[\frac{(B^{l=1}_{*})^2}{v_{esc}}\bigg]^{2m_{s,dip}}.
\end{equation}
The later formulation is used when the Alfv\'en radius of a given dipole \& quadrupole mixed field is greater than the pure quadrupole case for the same $\Upsilon$, i.e. the maximum of our new formula or the pure quadrupole. We define $\Upsilon_{crit}$ to separate the two regimes (see Figure \ref{PS}).
The importance of the relative radial decay of both modes and the location of the opening and Alfv\'en radii appear to play a key role, and deserve further follow up investigation. This work analytically fits the decay of the magnetic flux, but a parametric relationship for the field opening remains uncertain. The relation of the relative sizes of the Alfv\'en and opening radii are found to be dependent on geometry, which can be used to inform potential field source surface modelling, where by the source surface must be specified when changing the field geometry.
Paper II includes the addition of octupolar field geometries, another primary symmetry family which introduces an additional complication in the relative orientation of the octupole to the dipole. It is shown however, that the mixing of any two axisymmetric geometries will follow a similar behaviour, especially if each belongs to different symmetry families (Finley \& Matt. in prep). The lowest order mode largely dominates the dynamics of the torque until the Alfv\'en radii and opening radii are sufficiently close to the star for the higher order modes to impact the field strength.
\acknowledgments
Thanks for helpful discussions and technical advice from Georgios Pantolmos \& Matt, Victor See, Victor R\'eville, Sasha Brun and Claudio Zanni.
This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 682393).
We thank Andrea Mignone and others for the development and maintenance of the PLUTO code.
Figures within this work are produced using the python package matplotlib \citep{hunter2007matplotlib}.
|
\section{Introduction}
\vspace{-0.2cm}
Over the last 15 years, there has been a significant growth in the number of institutional and subject repositories storing research content. However, each repository on its own is of limited use, as the key value of repositories comes from being able to search, recommend and analyse content across this distributed network. While these repositories have been established to store primarily research papers, they contain, in fact, a variety of document types, including theses and slides. Services operating on the content from across this repository network should be able to distinguish between document types based on the supplied metadata.
However, metadata inconsistencies are making this very difficult. As we show later in the study, \textasciitilde$62\%$ of documents in repositories do not have associated metadata describing the document type. Moreover, when document type is specified, it is typically not done using an interoperable vocabulary.
Consequently, digital library aggregators like CORE \cite{DBLP:journals/dlib/KnothZ12}, OpenAIRE \cite{rettberg2012openaire} and BASE \cite{summann2006bielefeld} face the challenge of offering seamless SR systems over poor quality metadata supplied by thousands of providers. We hypothesise that by understanding the document type, we can increase user engagement in these services, for example, by means of filtering or re-ranking SR systems results.
In this paper, we develop a novel and highly scalable system for automatic identification of research papers, slides and theses. By applying this identification system, we analyse the logs of CORE' SR systems to see if we can find evidence of users preferring specific document type(s) over others.
The contributions of the paper are:
\begin{itemize}[label=\textbullet]
\item Presenting a lightweight, supervised classification approach for detecting \textit{Research}, \textit{Slides} and \textit{Thesis}, based on a small yet highly predictive set of features extracted from textual descriptors of (scientific) articles, reaching an F1-score of ${96.2\%}$ with the random forest classifier.
\item A publicly exposed and annotated dataset\cite{aristotelis_petr_2017} of approximately ${11.5k}$ of documents for the sake of comparison and reproducibility.
\item Proposing a modified CTR metric, balanced QTCTR, to analyse historical SR systems' logs to evaluate user engagement with the proposed content types in digital library systems, showing our users' inclination towards research and theses over slides.
\end{itemize}
The rest of the paper is organised as follows. Firstly, we discuss related work, followed by the presentation of our current data state. Secondly, we outline our approach and present
results of the classification approach and the analysis of current user engagement using our modified CTR metrics. Finally, we end with a discussion before concluding the paper.
\vspace{-0.3cm}
\section{Related work}
\vspace{-0.2cm}
The library community holds traditionally metadata records as a key enabler for resource discovery. Systems, such as BASE and WorldCat\footnote{https://www.worldcat.org/}, have been almost solely relying on metadata in their search services until today. But as such approach, as opposed to services indexing the content, cannot guarantee metadata validity, completeness and quality, nor can achieve acceptable recall \cite{DBLP:journals/dlib/KnothZ12}, some have started to believe that aggregative digital libraries have failed due to the interoperability issues facing OAI-PMH data providers. In fact, \cite{poynder2016CNI} specifically argues that the fact that BASE and OpenAIRE do not (or cannot) distinguish between document types of the records they harvest makes them ``not as effective as users might assume''.
While automatic document categorisation using structural and content features has been previously widely studied \cite{sebastiani2002machine, qi2009web, ghosh2008combining}, little work has been done on the issue of document type categorisation in the context of digital libraries until the recent study Caragea et al. \cite{caragea2016document}. They experimented with (1) \textit{bag-of-words}, (2) document \textit{URL tokens} and (3) document \textit{structural features} to classify academic documents into several types. Their set of 43 manually engineered \textit{structural features} have shown significant performance gain over conventional \textit{bag-of-words} models in these highly diverse data collections.
Unlike previous work in standard approaches to text categorisation, summarised in \cite{aphinyanaphongs2014comprehensive}, we use a subset of file and text specific characteristics, selectively gathered from \cite{caragea2016document}. The reduced dimensionality, as a result of the subset's minimal size, allows for scalable integration in ingestion pipelines of SR systems. In addition to the previous work, our study is to our knowledge the first to understand whether the integration of these document type classification systems can lead to more effective user engagement in SR systems.
\vspace{-0.3cm}
\section{Data - current state}
\vspace{-0.2cm}
CORE is a global service that provides access to millions of (open access) research articles aggregated from thousands of OA repositories and journals at a full text level. CORE offers several services including a search engine, a recommendation system, an API for text-miners and developers as well as some analytical services. As of April 2017, CORE provides access to over $70$ million metadata records and $6$ million full texts aggregated from $2,461$ data providers. From the available metadata descriptors, a directly available field to categorise records, at a certain extent, is the \textit{dc:subjects} field. While mostly available, currently $92\%$ of cases, only a small minority contain clear descriptions of the document type. More specifically, \textasciitilde$\num[round-precision=1]{30.03062}$ of records are marked as \verb|article|, \textasciitilde$7.3\%$ are marked as \verb|thesis| and $0\%$ as \verb|slides|. This means that we do not have any type document type indication for \textasciitilde${62\%}$ of our data.
\begin{table}
\centering
\vspace{-1\baselineskip}
\begin{tabular}{lr}
\toprule
{Term name} & {Term frequency}
\\
\midrule
article & \num{0.1365669182}\\
info:eu-repo/semantics/article & \num{0.0866117754}\\
journal articles & \num{0.03849475448}\\
thesis & \num{0.020513418}\\
info:ulb-repo/semantics/openurl/article & \num{0.0016581145}\\
info:eu-repo/semantics/doctoralthesis & \num{0.01059937032}\\
info:eu-repo/semantics/bachelorthesis & \num{0.01010210787}\\
\midrule
\caption{Most popular terms found in the \textbf{dc:subjects} field with >1\% occurrence}
\label{term:freq}%
\end{tabular}
\end{table}
Table \ref{term:freq} lists the top re-occurring terms that are most indicative of the three document types we are interested in. This provides empirical evidence of the poor adoption of interoperable document type descriptors across data providers. Finally, from the \textasciitilde$6$ million full text entries that CORE contains, ${8.5}$ million unique \textit{dc:subjects} field terms are currently recorded (one record can contain multiple subjects fields).
\vspace{-0.3cm}
\section{Approach}
\vspace{-0.2cm}
While one approach to address the problem of poor or missing document type descriptors can be to create guidelines for data providers, we believe this approach is slow, unnecessarily complex and does not scale. Instead, we aim to develop an automated system that infers the document type from the full text.
The assumptions we make for this study follow several observations on the textual features of documents stored in CORE:
\begin{itemize}[label=\textbullet]
\item \textbf{F1: Number of authors}: The more authors involved in a study, the more likely a document is a research paper as opposed to slides or thesis.
\item \textbf{F2: Total words}: These were tokenised from the parsed text content using the \textit{nltk}\cite{bird2006nltk} package. Intuitively, the lengthier a document is, in terms of total written words and amount of pages, the more likely it is a thesis.
\item \textbf{F3: Number of pages}: Research papers tend to have a fewer number of pages compared to theses and slides.
\item \textbf{F4: Average words per page}: Calculated as $\frac{\mbox{\#total words}}{\mbox{\#total pages}}$. The fewer words written per page on average, the more likely the document type is \textit{slides}.
\end{itemize}
We extract F2-F4 from their respective \verb|pdf| files with pdfMiner\cite{shinyama2015pdfminer}. F1 is extracted from the supplied metadata. We then apply one of the classifiers, described later in Section \ref{evaluation}, to predict the document type given these features.
\vspace{-0.3cm}
\section{Experiments}
\vspace{-0.2cm}
\subsection{Data Sample}
\vspace{-0.2cm}
\label{data-sample}
Our first goal was to create a sufficiently large ground truth dataset. Data labelling took place with a rule-based method applied to the CORE dataset. More specifically, we used a set of regular expressions on the {dc:subjects} field and the document's title as follows:
\begin{itemize}[label=\textbullet]
\item Subjects fields for which entries include the keyword ``thesis'' or ``dissertation'' were labelled as \textit{Thesis}.
\item Subjects fields for which entries do \textbf{not} include the keyword ``thesis'' or ``dissertation'' and their title does \textbf{not} include the keyword ``slides'' or ``presentation'' were labelled as \textit{Research}.
\item Subject fields for which entries do \textbf{not} include the keyword ``thesis'' or ``dissertation'' and their title includes the keyword ``slides'' or ``presentation'' were labelled as \textit{Slides}.
\end{itemize}
While this rule-based labelling process produced a sufficiently large number of samples for the \emph{Research} and \emph{Thesis} classes, it has not yielded a satisfactory sample size for the \textit{Slides} class. To address this issue, we have mined \verb|pdf|s and metadata from SlideShare\footnote{https://www.slideshare.net/} using their openly accessible API.
We wanted the total size of the sample to satisfy two criteria, a confidence level of $95\%$ at a confidence interval of $1\%$. The equation to calculate the necessary size of the data sample is:
\begin{equation}
n=\frac{Z^2\hat{p}(1-\hat{p})}{c^2}
\end{equation}
where, $Z$ is the $Z$ score, $\hat{p}$ is the percentage probability of picking a sample and $c$ is the desired confidence interval. Given a $Z$ score of $1.96$ for a $95\%$ confidence level, a confidence interval of $0.01$ and a sample proportion $p$ of $0.5$ (used as it is the most conservative and will give us the largest sample size calculation), this equation yields \textasciitilde${9.6k}$ samples.
We have gathered these $9.6k$ samples and additionally extended the dataset by $20\%$ to form a validation set, resulting in $11.5k$ samples. To produce a sample with a representative balance of classes, we limited slides to take up to $10\%$ of the final dataset, $55\%$ for research and the remaining $35\%$ for theses entries. We also ensured that all the \verb|pdf|s in the data sample are parsable by pdfminer.
Finally, we addressed the issue of missing values for feature F1, which SlideShare did not provide in over $97\%$ of cases, by applying multivariate imputations \cite{buuren2011mice}. To improve our knowledge of the feature distributions prior to applying the imputations for the \textit{Slides} class, we relied on extra data from Figshare\footnote{https://figshare.com/}.
To visualise the dimensionality and data variance in the resulting dataset, we have produced two and three dimensional projections of our data, using techniques introduced by \cite{maaten2008visualizing}. On small datasets ($<\num{100}k$ data points) these do not require much tuning of hyper-parameters and, out of manual inspection from a limited range of hyper-parameters, we decided to use \verb|perplexity| of $30$ and a \verb|theta| of $0.5$. As Figure \ref{tsne:viz} suggests, there is sufficient evidence of data sparsity.
\begin{figure}[t
\vspace{-1\baselineskip}
\subfloat[]{{\includegraphics[width=\textwidth/2]{tsne_big_letters.png} }}%
\subfloat[]{{\includegraphics[width=\textwidth/2]{tsne_3d.png} }}%
\caption{Data variance visualisation using (a) two and (b) three dimensional projections on the corresponding principal components.}%
\label{tsne:viz}%
\end{figure}
\subsection{Feature Analysis and Model Selection}
\vspace{-0.2cm}
\label{evaluation}
We have experimented with: Random Forest (RF), Gaussian Naive Bayes (GNB), $k$ Nearest Neighbours ($k$NN), Adaboost with Decision trees (Adaboost) and linear kernel Support Vector Machines (SVM).
We followed a standard 10-fold cross-validation approach to evaluate the models with an extra 20\% of the data left aside for model validation. The class balance discussed was preserved in each fold evaluation by applying stratified splits on both test and validation sets, simulating a representative distribution of categories in the CORE dataset. All features used were compared against their normalised and log-scaled counterparts to check for any possible performance improvements. We have also optimised for a small range of hyper-parameters for each machine learning algorithm using parameter sweeps, recording the best achieved performance for each algorithm class. The evaluation results are presented in Table \ref{train:test:results}.
Two baseline models have been used to assess the improvement brought by the machine learning classifiers. The approaches used are:
\begin{itemize}[label=\textbullet]
\item \textbf{Baseline 1:} Random class assignment with probability weights corresponding to the dataset's class balance.
\item \textbf{Baseline 2:} A rule-based approach based on statistically drawn thresholds for each feature and class respectively, using the upper $0.975$ and lower $0.025$ quantiles.
\end{itemize}
An analysis was carried out on the assembled dataset to form Baseline 2, based on feature distributions' percentiles. Distributions from the sample dataset largely followed a right skewed normal distribution (Figure \ref{qq:plots}), proving such a model should be a suitable candidate to evaluate against. To avoid overfitting, outliers were removed using Tukey's method \cite{tukey1949comparing}, which was preferred due to its independence on the data distribution, omitting values outside of the range:
\begin{equation}
(Q1 - 1.5*IQR) > Y > (Q3 + 1.5*IQR)
\end{equation}
where, $Y$ is the set of acceptable data points, $Q1$ is the lower quartile, $Q3$ is the upper quartile and $\textit{IQR = Q3 - Q1}$ is the interquartile range.
\begin{table}[H]
\vspace{-1\baselineskip}
\begin{center}
\begin{tabular}{l@{\hspace{1em}}lllll}
\toprule
\multirow{2}{*}{Feature} &
\multicolumn{5}{c}{Document Type}\\
& {Research} & {} & {Slides} & {} & {Thesis}\\
\midrule
F1 & $1 \le x \le 5$ & {} & $1 \le x \le 8$ & {} & ==1\\
F2 & $\num[round-precision=0]{1226.825} \le x \le \num[round-precision=0]{19151.425}$ & {} & $\num[round-precision=0]{93.6} \le x \le \num[round-precision=0]{7339.8}$ & {} & $\num{15184}\le x \le \num{210720}$\\
F3 & $3 \le x \le 41$ & {} & $1 \le x \le \num[round-precision=0]{74.575}$ & {} & $47 \le x \le 478$\\
F4 & $\num[round-precision=0]{208.2297} \le x \le \num[round-precision=0]{926.8950}$ & {} & $\num[round-precision=0]{8.0625} \le x \le \num[round-precision=0]{722.9375}$ & {} & $\num[round-precision=0]{197.7846} \le x \le \num[round-precision=0]{529.9571}$\\
\midrule
\end{tabular}
\caption{Percentile thresholds (upper $0.975$ and lower $0.025$ quantiles) for Baseline 2, following outlier removal.}
\label{perc:thresh}
\end{center}
\end{table}
The acquired thresholds for Baseline 2 are listed in Table \ref{perc:thresh}. To assign a particular example a document type $t$, all its features must fall within the boundaries specified. When this method fails, we assign the majority class (\textit{Research}).
\begin{figure}[t
\vspace{-1\baselineskip}
\includegraphics[width=\textwidth]{all_distributions_normal_qqplots_12k_no_titles_imped.png}
\caption{Normal Q-Q Plots split by document type and feature.}
\label{qq:plots}%
\end{figure}
\subsection{Results}
\vspace{-0.2cm}
The evaluation results, presented in Table \ref{train:test:results}, show that all our models outperform the baselines by a large margin. However, baseline 2 demonstrates a perhaps surprisingly good performance on this task. Random forest and Adaboost are the top performers achieving about $0.96$ in F1-score on both the test and validation sets. While we cannot distinguish which model is better at the $95\%$ confidence level and $1\%$ confidence interval, see Section \ref{data-sample}, we decided to productionise random forest due to the model's simplicity.
Figure \ref{pvr:viz} shows a breakdown of the final precision/recall performances according to the assigned document type. This indicates that a particularly significant improvement of the machine learning models over the baselines is achieved on the \textit{Slides} class. However, as only about $10\%$ of documents in the dataset are slides, the baselines are not so much penalised for these errors in the overall results.
\begin{table
\vspace{-1\baselineskip}
\centering
\begin{tabular}{l@{\hspace{1em}}lcccccccc}
\toprule
\multirow{2}{*}{} &
\multicolumn{7}{c}{Algorithm}\\
& {Measure} & {RF} & {GNB} & {kNN} & {Adaboost} & {SVM} & {Baseline 1} & {Baseline 2}\\
\midrule
\multirow{3}{*}{\specialcell{Test\\Results}} & Precision & \bfseries \num{0.9623378} & \num{0.9430801} & \num{0.9494912} & \num{0.9580391} & \num{0.8967811} & \num{0.4925689} & \num{0.5687876}\\
{} & Recall & \bfseries \num{0.9622894} & \num{0.9414282} & \num{0.9497192} & \num{0.9569404} & \num{0.8932870} & \num{0.3269561} & \num{0.4761838}\\
{} & F1-score & \bfseries \num{0.9623000} & \num{0.9416430} & \num{0.9495535} & \num{0.9573248} & \num{0.8695493} & \num{0.3269561} & \num{0.5154115}\\
\midrule
\multirow{3}{*}{\specialcell{Validation\\Results}} & Precision & \num{0.9566711} & \num{0.9356236} & \num{0.94531} & \bfseries \num{0.9607432} & \num{0.8434624} & \num{0.5572325} & \num{0.6361684}\\
{} & Recall & \num{0.9553265} & \num{0.9338488} & \num{0.9454467} & \bfseries \num{0.9604811} & \num{0.8741409} & \num{0.4570447} & \num{0.6564814}\\
{} & F1-score & \num{0.9557796} & \num{0.9337341} & \num{0.9452725} & \bfseries \num{0.9605947} & \num{0.8310864} & \num{0.4570447} & \num{0.5945136}\\
\midrule
\end{tabular}
\caption{Test and validation set results on weighted evaluation metrics across all algorithms.}
\label{train:test:results}
\end{table}
\begin{figure
\vspace{-1\baselineskip}
\includegraphics[width=\textwidth]{pvr_all_revised_b1_and_b2.png}
\caption{Precision versus Recall for all algorithms on the test set split by class.}
\label{pvr:viz}%
\end{figure}
To evaluate the importance of individual features, a \textit{post-hoc} analysis was carried out. We fitted the models of our selected algorithms with a \textit{single} feature group at a time. In this scenario, we have recorded high precision performances. Individual feature contributions do not vary widely, except in the case of F4 and the overall performance of the SVM classifier. F1-3 are the most predictive features. We list our findings in Table \ref{indiv:feat:perf}.
\begin{table}[t
\vspace{-1\baselineskip}
\begin{center}
\begin{tabular}{lSSSSSSSSS}
\toprule
\multirow{2}{*}{Features} &
\multicolumn{9}{c}{Average Weighted F1-score}\\
& {RF} & {} & {GNB} & {} & {kNN} & {} & {Adaboost} & {} & {SVM}\\
\midrule
Only: F2 & \num{0.8825247} & {} & \num{0.7660922} & {} & \num{0.8701925} & {} & \bfseries \num{0.8839478} & {} & \num{0.1868345}\\
Only: F3 & \num{0.8436393} & {} & \num{0.8412492} & {} & \num{0.8413556} & {} & \num{0.8424147} & {} & \bfseries \num{0.8441373}\\
Only: F1 & \bfseries \num{0.8007190} & {} & \num{0.6819066} & {} & \num{0.8007190} & {} & \bfseries \num{0.8007190} & {} & \num{0.6440705}\\
Only: F4 & \bfseries \num{0.7035729} & {} & \num{0.4744616} & {} & \num{0.6918964} & {} & \num{0.7017862} & {} & \num{0.3505943}\\
\bottomrule
\specialcell{All features\\(RF)} & \bfseries \num{0.9623000} & {} & \num{0.9416430} & {} & \num{0.9495535} & {} & \num{0.9573248} & {} & \num{0.8695493}\\
\end{tabular}
\end{center}
\caption{Classifiers' performance with individual feature groups across all algorithms on the test set in descending order, based on their contribution.}
\label{indiv:feat:perf}
\end{table}
\vspace{-0.3cm}
\section{Can the model help improve user engagement in SR systems?}
\vspace{-0.2cm}
We applied the random forest model to classify existing content in CORE. Joining the document type information with CORE's SR systems' user logs, enabled us to analyse document type user preferences in CORE's SR systems.\footnote{It should be noted that as CORE provides thumbnails on its SR results pages, users get an idea of the document type prior to accessing it.} We followed the intuition that if we can find that users prefer clicking in SR results on one document type over another, this will provide the argument for using document type information in SR systems to better serve the needs of these users.
A traditional metric to measure the popularity of a link is the Click-Through Rate (CTR), measured as:
\begin{equation}
CTR_T = \frac{|Clicks|}{|Impressions|}
\end{equation}
However, we cannot use CTR directly to assess whether people are more likely to click on certain document types than others in the SR system results. This is because we serve, on average, ${\num[round-precision=1]{66.68907}\%}$ \textit{Research}, ${\num[round-precision=1]{27.17887}\%}$ \textit{Thesis} and ${\num[round-precision=1]{6.132052}\%}$ \textit{Slides} impressions across our SR engines. Consequently, the CTR metric would be biased towards the \textit{Slides} class. This is due to the fact that when an action is made on an impression set, the class most represented in the set will benefit from this action on average the least. Put differently, this is accounted to the class imbalance.
To address this problem, we extend CTR to put \textit{impression equality} into perspective with the following process. We group impressed items in sets $Q$, reflecting the documents served following a query submission (in case of the recommender, the query is a document with respect to which we recommend)\footnote{The number of impressions generated in response to a query can vary across queries. In our case, it can be from zero to ten for search and from zero to five for the recommender.}. We assign to each impression set a type $q_t$ based on the types of document(s) clicked in the results list. In case multiple clicks to distinct document types are made in response to a query, we generate multiple impression sets derived from it, each assigned to one of them.
We then calculate the \textit{Query Type Click-Through Rate} $(QTCTR)$ as a fraction of the number of queries which resulted in a click to a given document type over the number of all queries:
\begin{equation}
QTCTR=\frac{|Q_T|}{|Q|}
\end{equation}
$QTCTR$ tells us the absolute proportion of queries that result in clicking on a particular document type. We can regularise/normalise $QTCTR$ to reflect the imbalance of impression types, forming the \textit{Regularised Query Type Click-Through Rate} $(RQTCTR)$. We include impression sets with no interaction in this calculation.
\begin{equation}
RQTCTR=\frac{|Q_T|}{|Q|} * \frac{|Impressions_T|}{|Impressions|}
\end{equation}
\begin{table}[t
\vspace{-1\baselineskip}
\begin{center}
\begin{tabular}{lllcccc||cccccl}
\toprule
\multirow{3}{*}{Metric} &
\multirow{3}{*}{Engine} &
\multicolumn{8}{c}{Impression set positions}\\
{} & {} & {} & \multicolumn{4}{c}{Any position} & \multicolumn{3}{c}{Top position}\\
{} & {} & {} & {} & Research & Slides & Thesis & Research & Slides & Thesis\\
\midrule
\multirow{2}{*}{QTCTR} &
Search & {} & {} & \num[round-precision=5]{0.13685276} & \num[round-precision=5]{0.01878269} & \bfseries \num[round-precision=5]{0.32357637} & \bfseries \num[round-precision=5]{0.038179391} & \num[round-precision=5]{0.003885447} & \num[round-precision=5]{0.018287103}\\
{} & Recommender & {} & {} & \bfseries \num[round-precision=5]{0.0067483264} & \num[round-precision=5]{0.0007438155} & \num[round-precision=5]{0.0036051059} & \bfseries \num[round-precision=5]{0.0048185582} & \num[round-precision=5]{0.0004610628} & \num[round-precision=5]{0.0020431410}\\
\midrule
\multirow{2}{*}{RQTCTR} &
Search & {} & {} & \num[round-precision=5]{0.081859422} & \num[round-precision=5]{0.001416012} & \bfseries \num[round-precision=5]{0.100611446} & \bfseries \num[round-precision=5]{0.0228372653} & \num[round-precision=5]{0.0002929208} & \num[round-precision=5]{0.0056861132}\\
{} & Recommender & {} & {} & \bfseries \num[round-precision=5]{0.0048783667} & \num[round-precision=5]{0.0000346020} & \num[round-precision=5]{0.0007927883} & \bfseries \num[round-precision=5]{3.483337e-03} & \num[round-precision=5]{2.144846e-05} & \num[round-precision=5]{4.493011e-04}\\
\midrule
\end{tabular}
\end{center}
\caption{Modified click-through rate metrics performance on CORE's SR systems.}
\label{ctr:results}
\vspace*{-\baselineskip}
\end{table}
The $QTCTR$ and $RQTCTR$ values from the CORE's SR systems, for the three different document types, are presented in Table \ref{ctr:results}. The shows that there is noteworthy difference in preference for \textit{Research} type documents and \textit{Thesis} over \textit{Slides} by an order of one magnitude. This is true for clicks generated on any document in an impression set and when the click was on top positioned document. The $QTCTR$ results also reveal that many people in CORE are looking for theses. We believe this is due to the fact that CORE is one of the few systems (in not the only one) that aggregates theses from thousands of repositories at a full-text level.
\vspace{-0.3cm}
\section{Scalability analysis}
\vspace{-0.2cm}
There exists a linear relationship between the number of features ($N$) and prediction latency \cite{sklearn:efficiency}, expressed with the complexity of $O(N*M)$, where $M$ are the number of instances. The low number of features and model complexity, with our deployed model having $<10$ trees and $<5$ maximum nodes for each, the latency amounts to slightly over $\num{0.0001}$ seconds per prediction\footnote{This excludes network overhead from the API call and the feature extraction process.}. Due to CORE's continuously ongoing repository harvesting processes, the minimal feature extraction requirements will allow for new additions to be streamlined immediately after their processing, in comparison with the latency associated with the feature extraction process expected from \cite{caragea2016document}. This indicates the high scalability of our approach and applicability across millions of documents.
\vspace{-0.3cm}
\section{Future work}
\vspace{-0.2cm}
In promoting the current solution within CORE's systems, and making it accessible to users worldwide, we aim to:
\begin{itemize}[label=\textbullet]
\item Expose document type classification models as a service, with online model updating, through CORE's public API.
\item Boost \textit{Research} documents in our SR engines and negatively boost \textit{Slides} to aid faster retrieval of preferred content.
\item Evaluate the shift of user engagement as a direct effect of such changes in our services and adjusting our search/recommendation strategies accordingly.
\item Enhance user engagement analysis by cross-validation of our observations here metrics such as the \textit{dwell time}, a metric proven to be less unaffected by position, caption or other form of bias in SR results \cite{kim2014modeling}.
\item Extend the model in further iterations to also discern between sub-types of the \textit{Research} and \textit{Slides} classes, such as theoretical, surveys, use case or seminal research papers as well as slides corresponding to conference papers and lecture/course slides respectively.
\end{itemize}
\section{Conclusions}
We have presented a new scalable method for detecting document types in digital libraries storing scholarly literature achieving $0.96$ F1-score. We have integrated this classification system with the CORE digital library aggregator. This enabled us to analyse the SR system logs of to assess whether users prefer certain document types. Using a our Regularised Query Type Click-Through Rate (RQTCTR) metric, we have confirmed our hypothesis that the document type can contribute in finding a viable solution to improving user engagement.
\section*{Acknowledgements}
This work has been partly funded by the EU OpenMinTeD project under the H2020-EINFRA-2014-2 call, Project ID: 654021. We would also like to acknowledge the support of Jisc for the CORE project.
\bibliographystyle{unsrt}
|
\section{Introduction}
\label{S-I}
Random walks with long-memory arose naturally in applied mathematics, theoretical physics, computer sciences
and econometrics. One of them is the so-called elephant random walk (ERW). It is a one-dimensional discrete-time random
walk on integers, which has a complete memory of its whole history. It was introduced in 2004 by Sch\"utz and
Trimper \cite{Schutz04} in order to investigate the long-term memory effects in non-Markovian random walks.
It was referred to as the ERW in allusion to the famous saying that elephants can remember where they have been.
\\ \vspace{-1ex}\par
A wide range of literature is available on the asymptotic behavior of the ERW and its extensions
\cite{Baur16},\cite{Boyer14},\cite{Col17},\cite{Cressoni13},\cite{Cressoni07},\cite{Da13},\cite{Kumar10},\cite{Kursten16}.
However, many things remain to be done. The goal of this paper is to answer to several natural questions on the ERW.
\\ \vspace{-1ex}\par
One of them concerns the influence of the memory parameter $p$ on the almost sure asymptotic behavior of the ERW.
Quite recently, Baur and Bertoin \cite{Baur16} and independently Coletti, Gava and Sch\"utz \cite{Col17}
have studied the asymptotic normality of the ERW in the diffusive regime $p<3/4$ as well as in the critical regime $p=3/4$.
However, very few results are available on the almost sure asymptotic behavior of the ERW in the regime $p\leq 3/4$.
We shall fill the gap by proving a quadratic strong law as well as a law of iterated logarithm for
the ERW in the regime $p\leq 3/4$.
\\ \vspace{-1ex}\par
Another key question concerns the limiting distribution of the ERW in the superdiffusive regime $p> 3/4$.
Initially, it was suggested by Sch\"utz and Trimper \cite{Schutz04} that, even in the superdiffusive regime,
the ERW has a Gaussian limiting distribution. Later, it was conjectured by Da Silva et al. \cite{Da13}
that this limiting distribution is not Gaussian, see also \cite{Col17},\cite{Par06}.
One can observe that the analytical study of \cite{Da13} is not sufficient to prove the non-gaussianity of the
limiting distribution. We shall provide a rigorous mathematical proof that the limiting distribution is not Gaussian.
Starting from the symmetric initial condition, we will also show that this limiting distribution is sub-Gaussian.
\\ \vspace{-1ex}\par
Baur and Bertoin \cite{Baur16} extensively used the connection to P\'{o}lya-type urns \cite{Harris15}
as well as two functional limit theorems for multitype branching processes due to Janson \cite{Janson04}, see also \cite{Chauvin11}.
Our strategy is totally different as it relies on the theory of martingales. To be more precise, we shall make use of the strong
law of large numbers and the central limit theorem for martingales \cite{Duflo97},\cite{HallHeyde80} as well as
the law of iterated logarithm for martingales \cite{Stout70},\cite{Stout74}. We strongly believe that our approach could be successfully
extended to ERW with stops \cite{Cressoni13},\cite{Harbola14}, to amnesiac ERW \cite{Cressoni07}, as well as to multi-dimensional ERW
\cite{Cressoni13},\cite{Lyu17}.
\\ \vspace{-1ex}\par
The paper is organized as follows. In Section \ref{S-E}, we introduce the exact ERW and the martingale
we shall extensively make use of. Section \ref{S-MR} is devoted to the main results of the paper.
We establish the almost sure asymptotic behavior as well as the asymptotic normality of the ERW
in the diffusive and critical regimes. Moreover, in the superdiffusive regime,
we provide the first rigorous mathematical proof that the limiting distribution of the ERW is not Gaussian.
Our martingale approach is described in Appendix A, while all technical proofs of Section \ref{S-MR}
are postponed to Appendices B and C.
\ \vspace{-2ex} \\
\section{The elephant random walk}
\label{S-E}
The one-dimensional ERW is defined as follows. The random walk starts at the origin at time zero,
$S_0=0$. At time $n=1$, the elephant moves to the right with probability $q$ and to the left with probability
$1-q$ where $q$ lies between zero and one. Hence, the position of the elephant at time $n=1$ is given by
$S_1=X_1$ where $X_1$ has a Rademacher $\mathcal{R}(q)$ distribution.
Afterwards, at any time $n \geq 1$, we choose uniformly at random an interger $k$ among the previous times
$1,\ldots,n$, and we define
\begin{equation*}
X_{n+1} = \left \{ \begin{array}{ccc}
+X_{k} &\text{ with probability } & p, \vspace{2ex}\\
-X_{k} &\text{ with probability } & 1-p ,
\end{array} \nonumber \right.
\vspace{2ex}
\end{equation*}
where the parameter $p \in [0,1]$ is the memory of the ERW.
Then, the position of the ERW is given by
\begin{equation}
\label{POSERW}
S_{n+1}=S_{n}+X_{n+1}.
\end{equation}
\\ \vspace{-3ex}\par
In order to understand well how the elephant moves, it is straightforward to see that
for any time $n \geq 1$, $X_{n+1} = \alpha_{n} X_{\beta_n}$
where $\alpha_n$ and $\beta_n$ are two independent discrete random variables where
$\alpha_n$ has a Rademacher $\mathcal{R}(p)$ distribution while $\beta_n$ is uniformly distributed over
the integers $\{1,\cdots,n\}$. Moreover, $\alpha_n$ is independent of $X_1,\ldots,X_n$.
\\ \vspace{-1ex}\par
Let $(\mathcal{F}_n)$ be the increasing sequence of $\sigma$-algebras,
$\mathcal{F}_n=\sigma(X_1,\ldots,X_n)$.
For any time $n \geq 1$, we clearly have
\begin{equation}
\label{CEX}
\mathbb{E}[X_{n+1} | \mathcal{F}_{n}] = \mathbb{E}[\alpha_{n}] \times \mathbb{E}[X_{\beta_n} | \mathcal{F}_{n}]=(2p-1)\frac{S_n}{n} \hspace{1cm} \text{a.s.}
\end{equation}
which, together with \eqref{POSERW}, implies that
\begin{equation}
\label{CES1}
\mathbb{E}[S_{n+1} | \mathcal{F}_{n}] = \gamma_{n} S_{n} \hspace{1cm} \text{where} \hspace{1cm}
\gamma_n=\Bigl( \frac{n+2p-1}{n} \Bigr).
\end{equation}
Moreover,
$$
\prod_{k=1}^n \gamma_k =\frac{\Gamma(n+2p)}{\Gamma(n+1)\Gamma(2p)}
$$
where $\Gamma$ stands for the Euler gamma function. Therefore, let $(M_n)$ be the sequence of random variables defined,
for all $n \geq 0$, by $M_n = a_n S_n$ where $a_1=1$ and, for all $n\geq 2$,
\begin{equation}
\label{DEFAN}
a_n= \prod_{k=1}^{n-1} \gamma_k^{-1} = \frac{\Gamma(n)\Gamma(2p)}{ \Gamma(n+2p-1)}.
\end{equation}
Since $a_n = \gamma_n a_{n+1}$, we clearly deduce from \eqref{CES1} that for any time $n \geq 1$,
\begin{equation*}
\mathbb{E}[ M_{n+1} | \mathcal{F}_{n}] = M_{n} \hspace{1cm} \text{a.s.}
\end{equation*}
In other words, the sequence $(M_n)$ is a multiplicative real martingale. Our strategy is to make use of
the martingale $(M_n)$ in order to deduce the asymptotic behavior of $(S_n)$.
\section{Main results.}
\setcounter{equation}{0}
\label{S-MR}
\subsection{The diffusive regime}
Our first result concerns the almost sure convergence of the ERW in the diffusive regime where $0 \leq p <3/4$.
\begin{thm}
\label{T-ASCVG-DR}
We have the almost sure convergence
\begin{equation}
\label{T-ASCVG-DR1}
\lim_{n \rightarrow \infty} \frac{S_n}{n}=0 \hspace{1cm} \text{a.s.}
\end{equation}
\end{thm}
\noindent
We focus our attention on the almost sure rates of convergence of the ERW.
\begin{thm}
\label{T-ASCVGRATES-DR}
We have the quadratic strong law
\begin{equation}
\label{T-ASCVG-DR2}
\limsup_{n \rightarrow \infty} \frac{1}{\log n} \sum_{k=1}^n \Bigl(\frac{S_k}{k}\Bigr)^2=\frac{1}{3-4p} \hspace{1cm} \text{a.s.}
\end{equation}
In addition, we also have the law of iterated logarithm
\begin{eqnarray}
\limsup_{n \rightarrow \infty} \Bigl(\frac{1}{2 n \log \log n}\Bigr)^{1/2} S_n & = &
-\liminf_{n \rightarrow \infty} \Bigl(\frac{1}{2 n \log \log n}\Bigr)^{1/2} S_n \nonumber \\
& = & \frac{1}{\sqrt{3-4p}} \hspace{1cm} \text{a.s.}
\label{T-ASCVG-DR3}
\end{eqnarray}
In particular,
\begin{equation}
\limsup_{n \rightarrow \infty} \frac{S_n^2}{2 n \log \log n}= \frac{1}{3-4p} \hspace{1cm} \text{a.s.}
\label{T-ASCVG-DR4}
\end{equation}
\end{thm}
\noindent
Our next result is devoted to the asymptotic normality of the ERW in the diffusive regime $0 \leq p <3/4$.
\begin{thm}
\label{T-AN-DR}
We have the asymptotic normality
\begin{equation}
\label{T-AN-DR1}
\frac{S_n}{\sqrt{n}} \build{\longrightarrow}_{}^{{\mbox{\calcal L}}} \mathcal{N} \Bigl(0, \frac{1}{3-4p} \Bigr).
\end{equation}
\end{thm}
\begin{rem} One can observe that the additional term $(2q-1)n^{2p-1}/\Gamma(2p)$
is useless in Theorem 2 of \cite{Col17}. Moreover, in the particular case $p=1/2$, one find again
\begin{equation*}
\frac{S_n}{\sqrt{n}} \build{\longrightarrow}_{}^{{\mbox{\calcal L}}} \mathcal{N} (0,1).
\end{equation*}
\end{rem}
\subsection{The critical regime}
Hereafter, we investigate the critical regime where the memory parameter $p=3/4$.
\begin{thm}
\label{T-ASCVG-CR}
We have the almost sure convergence
\begin{equation}
\label{T-ASCVG-CR1}
\lim_{n \rightarrow \infty} \frac{S_n}{\sqrt{n} \log n}=0 \hspace{1cm} \text{a.s.}
\end{equation}
\end{thm}
\noindent
The almost sure rates of convergence of the ERW are as follows
\begin{thm}
\label{T-ASCVGRATES-CR}
We have the quadratic strong law
\begin{equation}
\label{T-ASCVG-CR2}
\limsup_{n \rightarrow \infty} \frac{1}{\log \log n} \sum_{k=2}^n \Bigl(\frac{S_k}{k \log k}\Bigr)^2=1 \hspace{1cm} \text{a.s.}
\end{equation}
In addition, we also have the law of iterated logarithm
\begin{eqnarray}
\limsup_{n \rightarrow \infty} \Bigl(\frac{1}{2 n \log n \log \log \log n}\Bigr)^{1/2} S_n & = &
-\liminf_{n \rightarrow \infty} \Bigl(\frac{1}{2 n \log n \log \log \log n}\Bigr)^{1/2} S_n \nonumber \\
& = & 1 \hspace{1cm} \text{a.s.}
\label{T-ASCVG-CR3}
\end{eqnarray}
In particular,
\begin{equation}
\limsup_{n \rightarrow \infty} \frac{S_n^2}{2 n \log n \log \log \log n}= 1 \hspace{1cm} \text{a.s.}
\label{T-ASCVG-CR4}
\end{equation}
\end{thm}
\noindent
One can observe a very unusual rate of convergence in the law of iterated logarithm.
Our next result deals with the asymptotic normality of the ERW in the critical regime $p=3/4$.
\begin{thm}
\label{T-AN-CR}
We have the asymptotic normality
\begin{equation}
\label{T-AN-CR1}
\frac{S_n}{\sqrt{n \log n}} \build{\longrightarrow}_{}^{{\mbox{\calcal L}}} \mathcal{N} (0,1).
\end{equation}
\end{thm}
\begin{rem} As before, the additional term $(2q-1)\sqrt{n}/\Gamma(3/2)$
is useless in Theorem 2 of \cite{Col17}.
\end{rem}
\subsection{The superdiffusive regime}
Finally, we focus our attention on the more complicated superdiffusive regime where
$3/4 < p \leq 1$.
\begin{thm}
\label{T-ASCVG-SR}
We have the almost sure convergence
\begin{equation}
\label{T-ASCVG-SR1}
\lim_{n \rightarrow \infty} \frac{S_n}{n^{2p-1}}=L \hspace{1cm} \text{a.s.}
\end{equation}
where $L$ is a non-degenrate random variable. This convergence
also holds in $\mathbb{L}^4$, which means that
\begin{equation}
\label{T-ASCVG-SR2}
\lim_{n \rightarrow \infty} \mathbb{E}\Bigl[ \Bigl| \frac{S_n}{n^{2p-1}} -L \Bigr|^4 \Bigr]=0.
\end{equation}
\end{thm}
\ \vspace{-2ex} \\
\begin{rem}
One can observe that the first three moments of $S_n$ where previously calculated in \cite{Da13} in the
special case $q=1$. However, the analytical study of \cite{Da13} is not sufficient to evaluate the
moments of $L$.
\end{rem}
\begin{thm}
\label{T-MOM-SR}
The first four moments of $L$ are given by
\begin{eqnarray}
\label{MOML1}
\mathbb{E}[L] &=& \frac{2q-1}{\Gamma(2p)},\\
\label{MOML2}
\mathbb{E}[L^2] &=& \frac{1}{(4p-3)\Gamma(2(2p-1))}, \\
\label{MOML3}
\mathbb{E}[L^3] &=& \frac{2p(2q-1)}{(2p-1)(4p-3)\Gamma(3(2p-1))}, \\
\label{MOML4}
\mathbb{E}[L^4] &=& \frac{6(8p^2-4p-1)}{(8p-5)(4p-3)^2\Gamma(4(2p-1))}.
\end{eqnarray}
\end{thm}
\ \vspace{-2ex} \\
\begin{rem} Our last result provides the first rigorous mathematical proof that, in the
superdiffusive regime, the limiting distribution $L$ of the ERW is not Gaussian.
As a matter of fact, denote by $\mu$ and $\sigma^2$ the mean value and the
variance of $L$, $\mu=\mathbb{E}[L]=2q-1$ and $\sigma^2=\mathbb{E}[(L-\mu)^2]$.
Moreover, let $\alpha$ and $\kappa$ be the skewness and the kurtosis of $L$,
respectively defined by
$$
\alpha=\frac{\mathbb{E}[(L-\mu)^3]}{\sigma^3}
\hspace{1cm}\text{and}\hspace{1cm}
\kappa=\frac{\mathbb{E}[(L-\mu)^4]}{\sigma^4}.
$$
In the special case $q=1/2$, the skewness $\alpha=0$, while the kurtosis
$$
\kappa=\frac{6(8p^2-4p-1)(\Gamma(2(2p-1)))^2}{(8p-5)\Gamma(4(2p-1))}.
$$
It is not hard to see that $\kappa$ is a decreasing function of $p$ such that,
for all $3/4 < p \leq 1$, $1 \leq \kappa < 3$. It means that $L$ has a sub-Gaussian distribution.
Furthermore, if $q=1$, $\alpha>0$ and if $q=0$, $\alpha<0$.
Finally, the kurtosis $\kappa$ shares the same value for $q=1$ or $q=0$, which can be smaller, greater or
equal to $3$.
\end{rem}
\section*{Appendix A \\ The martingale approach}
\renewcommand{\thesection}{\Alph{section}}
\renewcommand{\theequation}{\thesection.\arabic{equation}}
\setcounter{section}{1}
\setcounter{equation}{0}
We already saw that the sequence $(M_n)$ given, for all $n \geq 0$, by $M_n=a_nS_n$, is a multiplcative real martingale.
Moreover, for any $n \geq 1$, $X_n$ is a binary random variables taking values in $\{+1,-1\}$.
Consequently, $|S_n| \leq n$, which implies that $(M_n)$ is locally square integrable.
The martingale $(M_n)$ can be rewritten in the additive form
\begin{equation}
\label{DECMN}
M_n=\sum_{k=1}^n a_{k} \varepsilon_{k}
\end{equation}
since its increments $\Delta M_n= M_n-M_{n-1}$ satisfy
$\Delta M_{n}= a_{n}S_{n}-a_{n-1}S_{n-1}=a_{n} \varepsilon_{n}$
where $\varepsilon_{n}=S_{n}-\gamma_{n-1}S_{n-1}$. The predictable quadratic variation \cite{Duflo97}
associated with
$(M_n)$ is given by $\langle M \rangle_0=0$ and, for all $n\geq 1$,
\begin{equation}
\label{DEFIP}
\langle M \rangle_n=\sum_{k=1}^n \mathbb{E}[ \Delta M_k^2|\mathcal{F}_{k-1}].
\end{equation}
We immediately obtain from \eqref{CES1} that
$\mathbb{E}[\varepsilon_{n+1} | \mathcal{F}_{n}]=0$. Moreover, it follows from \eqref{POSERW} together with \eqref{CEX} that
\begin{equation}
\label{CES2}
\mathbb{E}[S_{n+1}^2 | \mathcal{F}_{n}] = \mathbb{E}[ S_n^2 +2 S_n X_{n+1} +1 | \mathcal{F}_{n}]=1+ ( 2 \gamma_n -1)S_{n}^2
\hspace{1cm} \text{a.s.}
\end{equation}
Consequently, as $\mathbb{E}[\varepsilon_{n+1}^2 | \mathcal{F}_{n}]= \mathbb{E}[ S_{n+1}^2 | \mathcal{F}_{n}] - \gamma_n^2S_n^2$,
we deduce from \eqref{CES2} that, for all $n\geq 1$,
\begin{eqnarray}
\mathbb{E}[\varepsilon_{n+1}^2 | \mathcal{F}_{n}]
& = & 1 + ( 2 \gamma_n -1)S_{n}^2 - \gamma_n^2S_n^2 = 1 - (\gamma_n -1)^2 S_n^2 \hspace{1cm} \text{a.s.} \nonumber\\
& = & 1 - (2p-1)^2 \Bigl(\frac{S_n}{n}\Bigr)^2 \hspace{1cm} \text{a.s.}
\label{CEEPS2}
\end{eqnarray}
By the same token,
\begin{eqnarray}
\mathbb{E}[\varepsilon_{n+1}^4 | \mathcal{F}_{n}]
& = & 1 -3(\gamma_n -1)^4S_n^4 +2 (\gamma_n -1)^2 S_n^2 \hspace{1cm} \text{a.s.} \nonumber\\
& = & 1 - 3(2p-1)^4 \Bigl(\frac{S_n}{n}\Bigr)^4 + 2(2p-1)^2 \Bigl(\frac{S_n}{n}\Bigr)^2 \hspace{1cm} \text{a.s.}
\label{CEEPS4}
\end{eqnarray}
On the one hand, if $p=1/2$, $\mathbb{E}[\varepsilon_{n+1}^2 | \mathcal{F}_{n}]=1$ and $\mathbb{E}[\varepsilon_{n+1}^4 | \mathcal{F}_{n}]=1$ a.s.
On the other hand, we obtain from \eqref{CEEPS2} and \eqref{CEEPS4}
the almost sure upper bounds
\begin{equation}
\label{UBCEEPS}
\sup_{n \geq 0} \mathbb{E}[\varepsilon_{n+1}^2 | \mathcal{F}_{n}] \leq 1 \hspace{1cm} \text{and} \hspace{1cm}
\sup_{n \geq 0} \mathbb{E}[\varepsilon_{n+1}^4 | \mathcal{F}_{n}] \leq \frac{4}{3}.
\end{equation}
Hereafter, we deduce from \eqref{DECMN}, \eqref{DEFIP} and \eqref{CEEPS2} that
\begin{equation}
\label{CALIP}
\langle M \rangle_n=\sum_{k=1}^n a_k^2 - (2p-1)^2 \zeta_n
\hspace{1cm} \text{where}
\hspace{1cm}
\zeta_n = \sum_{k=1}^{n-1} a_{k+1}^2 \Bigl(\frac{S_k}{k}\Bigr)^2.
\end{equation}
The asymptotic behavior of the martingale $(M_n)$ is closely related to the one of
$$
v_n=\sum_{k=1}^n a_k^2= \sum_{k=1}^{n} \Bigl( \frac{\Gamma(k) \Gamma(2p) }{\Gamma(k+2p-1)} \Bigr)^2.
$$
Via standard results on the asymptotic behavior of the Euler gamma function, we have three regimes.
In the diffusive regime where $0 \leq p <3/4$,
\begin{equation}
\label{CVGVNDR}
\lim_{n\rightarrow \infty} \frac{v_{n}}{n^{3-4p}} = \ell \hspace{1cm} \text{where} \hspace{1cm}
\ell=\frac{(\Gamma(2p))^2}{3-4p}.
\end{equation}
In the critical regime where $p=3/4$,
\begin{equation}
\label{CVGVNCR}
\lim_{n\rightarrow \infty} \frac{v_{n}}{\log n} = \frac{\pi}{4}.
\end{equation}
In the superdiffusive regime where
$3/4 < p \leq 1$, $v_n$ converges to the finite value
\begin{equation}
\label{CVGVNSR}
\lim_{n\rightarrow \infty} v_n = \sum_{k=0}^{\infty} \Bigl( \frac{\Gamma(k+1) \Gamma(2p) }{\Gamma(k+2p)} \Bigr)^2
= \sum_{k=0}^{\infty} \frac{(1)_k\,(1)_k\,(1)_k} {(2p)_k\, (2p)_k\,k!} = {}_{3}F_2\Bigl( \begin{matrix}
{1,1,1}\\
{2p,2p}\end{matrix} \Bigl| 1\Bigr)
\end{equation}
where, for any $a\in \mathbb{R}$, $(a)_k=a(a+1)\cdots(a+k-1)$ for $k\geq 1$, $(a)_0=1$ stands for the Pochhammer symbol and ${}_{3}F_2$ is the
generalized hypergeometric function defined by
\begin{eqnarray*}
{}_{3}F_2 \Bigl( \begin{matrix}
{a,b,c}\\
{d,e}\end{matrix} \Bigl|
{\displaystyle z}\Bigr)
=\sum_{k=0}^{\infty}
\frac{(a)_k\,(b)_k\,(c)_k}
{(d)_k\,(e)_k\, k!} z^k.
\end{eqnarray*}
\section*{Appendix B \\ Proofs of the almost sure convergence results}
\renewcommand{\thesection}{\Alph{section}}
\renewcommand{\theequation}{\thesection.\arabic{equation}}
\setcounter{section}{2}
\setcounter{equation}{0}
\vspace{-3ex}
\subsection*{}
\begin{center}
{\bf B.1. The diffusive regime.}
\end{center}
\ \vspace{-4ex}\\
\noindent{\bf Proof of Theorem \ref{T-ASCVG-DR}.}
First of all, we focus our attention on the proof of the almost sure convergence
\eqref{T-ASCVG-DR1}. We already saw from \eqref{CALIP} that $\langle M \rangle_n \leq v_n$.
Moreover, it follows from \eqref{CVGVNDR} that, in the diffusive regime, $v_n$
increases to infinity with an arithmetic speed $n^{3-4p}$.
Then, we obtain from the strong law of large numbers for martingales given e.g. by Theorem 1.3.24 of \cite{Duflo97} that
\begin{equation*}
\vspace{-1ex}
\lim_{n\rightarrow \infty} \frac{M_{n}}{v_{n}} = 0 \hspace{1cm} \text{a.s.}
\end{equation*}
which implies that
\begin{equation}
\label{CVGMN}
\lim_{n\rightarrow \infty} \frac{M_{n}}{n^{3-4p}} = 0 \hspace{1cm} \text{a.s.}
\end{equation}
This convergence is not sharp enough to prove \eqref{T-ASCVG-DR1}.
However, thanks to the last part of Theorem 1.3.24,
we also have the almost sure rate of convergence in \eqref{CVGMN}
\begin{equation*}
\label{CVGRATEMN}
\vspace{-1ex}
\frac{M_{n}^2}{v_n} = O(\log v_n) \hspace{1cm} \text{a.s.}
\end{equation*}
which ensures that
\begin{equation}
\label{CVGRATEMN}
\frac{M_{n}^2}{n^{3-4p}} = O(\log n) \hspace{1cm} \text{a.s.}
\end{equation}
Hereafter, as $M_n=a_nS_n$, it clearly follows from \eqref{CVGRATEMN} that
\begin{equation*}
\frac{S_{n}^2}{n} = O(\log n) \hspace{1cm} \text{a.s.}
\end{equation*}
which immediately leads to \eqref{T-ASCVG-DR1}.
\hfill $\videbox$\\
\noindent{\bf Proof of Theorem \ref{T-ASCVGRATES-DR}.}
Denote by $f_{n}$ the explosion coefficient associated with the martingale
$(M_{n})$ given, for all $n \geq 1$, by
\begin{equation*}
f_{n} = \frac{a_n^2}{v_n}.
\end{equation*}
We clearly obtain from \eqref{CVGVNDR} that $f_n$ converges to zero. Moreover, it follows from
the almost sure convergence \eqref{T-ASCVG-DR1} together with \eqref{CEEPS2} that
\begin{equation*}
\lim_{n \rightarrow \infty} \mathbb{E}[\varepsilon_{n+1}^2 | \mathcal{F}_{n}]=1 \hspace{1cm} \text{a.s.}
\end{equation*}
Furthermore, we already saw in \eqref{UBCEEPS} that
\begin{equation*}
\sup_{n \geq 0} \mathbb{E}[\varepsilon_{n+1}^4 | \mathcal{F}_{n}] \leq \frac{4}{3} \hspace{1cm} \text{a.s.}
\end{equation*}
Consequently, by virtue of the quadratic strong law for martingales given e.g. in Theorem 3 of \cite{Bercu04},
\begin{equation*}
\lim_{n\rightarrow \infty}
\frac{1}{\log v_n} \sum_{k=1}^{n} f_{k} \Bigl( \frac{M_{k}^2}{v_k} \Bigr) = 1 \hspace{1cm} \text{a.s.}
\end{equation*}
which implies, via \eqref{CVGVNDR}, that
\begin{equation}
\label{LFQ-DR1}
\lim_{n\rightarrow \infty}
\frac{1}{\log n} \sum_{k=1}^{n} \frac{a_k^2 M_{k}^2}{v_k^2} = (3-4p) \hspace{1cm} \text{a.s.}
\end{equation}
Therefore, as $M_n=a_nS_n$ and $n^2 a_n^4$ is equivalent to $(3-4p)^2v_n^2$, we find
from \eqref{LFQ-DR1} that
\begin{equation}
\label{LFQ-DR2}
\lim_{n\rightarrow \infty}
\frac{1}{\log n} \sum_{k=1}^n \Bigl(\frac{S_k}{k}\Bigr)^2=\frac{1}{3-4p} \hspace{1cm} \text{a.s.}
\end{equation}
which completes the proof the quadratic strong law \eqref{T-ASCVG-DR2}. We shall now proceed to the proof of the law of iterated logarithm
given by \eqref{T-ASCVG-DR3}. In order to apply the law of iterated logarithm for martingales due to
Stout \cite{Stout74}, see also Corollary 6.4.25 in \cite{Duflo97}, it is only necessary to verify that
\begin{equation}
\label{CONDLIL}
\sum_{n=1}^{+\infty} \frac{a_{n}^4}{v_n^2} < +\infty.
\end{equation}
This is clearly satisfied since $a_n^4v_n^{-2}$ is equivalent to $(3-4p)^2n^{-2}$
and, as is well-known,
\begin{equation*}
\sum_{n=1}^{+\infty} \frac{1}{n^2} =\frac{\pi^2}{6}.
\end{equation*}
Hence, we find from the law of iterated logarithm for martingales that
\begin{eqnarray}
\limsup_{n \rightarrow \infty} \Bigl(\frac{1}{2 v_n \log \log v_n}\Bigr)^{1/2} M_n & = &
-\liminf_{n \rightarrow \infty} \Bigl(\frac{1}{2 v_n \log \log v_n}\Bigr)^{1/2} M_n \nonumber \\
& = & 1 \hspace{1cm} \text{a.s.}
\label{LIL-MG-DR}
\end{eqnarray}
As previously seen, the identity $M_n\!=\!a_nS_n$ together with \eqref{LIL-MG-DR} immediately lead to
\begin{eqnarray*}
\limsup_{n \rightarrow \infty} \Bigl(\frac{1}{2 n \log \log n}\Bigr)^{1/2} S_n & = &
-\liminf_{n \rightarrow \infty} \Bigl(\frac{1}{2 n \log \log n}\Bigr)^{1/2} S_n \nonumber \\
& = & \frac{1}{\sqrt{3-4p}} \hspace{1cm} \text{a.s.}
\end{eqnarray*}
which completes the proof of Theorem \ref{T-ASCVGRATES-DR}.
\hfill $\videbox$\\
\vspace{-5ex}
\subsection*{}
\begin{center}
{\bf B.2. The critical regime.}
\end{center}
\ \vspace{-4ex}\\
\noindent{\bf Proof of Theorem \ref{T-ASCVG-CR}.}
We shall proceed as in the proof of Theorem \ref{T-ASCVG-DR}. It follows from \eqref{CVGVNCR}
that, in the critical regime where $p=3/4$, $v_n$ increases slowly to infinity with a
logarithmic speed $\log n$. We obtain from Theorem 1.3.24 of \cite{Duflo97} that
\begin{equation}
\label{CVGRATEMNCR}
\frac{M_{n}^2}{\log n} = O(\log \log n) \hspace{1cm} \text{a.s.}
\end{equation}
However, we deduce from \eqref{DEFAN} that
\begin{equation}
\label{CVGANCR}
\lim_{n \rightarrow \infty} n a_n^2 = \frac{\pi}{4}.
\end{equation}
Hence, as $M_n=a_nS_n$, we find from \eqref{CVGRATEMNCR} together with \eqref{CVGANCR} that
\begin{equation*}
\frac{S_{n}^2}{n\log n} = O( \log \log n) \hspace{1cm} \text{a.s.}
\end{equation*}
which immediately implies \eqref{T-ASCVG-CR1}.
\hfill $\videbox$\\
\ \vspace{-1ex}\\
\noindent{\bf Proof of Theorem \ref{T-ASCVGRATES-CR}.} The proof of Theorem
\ref{T-ASCVGRATES-CR} is left to the reader as it follows essentially the same lines as the one of
Theorem \ref{T-ASCVGRATES-DR}.
\hfill $\videbox$\\
\ \vspace{-1ex}\\
\vspace{-5ex}
\subsection*{}
\begin{center}
{\bf B.3. The superdiffusive regime.}
\end{center}
\ \vspace{-4ex}\\
\noindent{\bf Proof of Theorem \ref{T-ASCVG-SR}.}
In the superdiffusive regime $3/4<p\leq 1$, we already saw from \eqref{CVGVNSR}
that $v_n$ converges to a finite value. Hence, as $\langle M \rangle_n \leq v_n$, we deduce
from \eqref{DECMN} together with Theorem 1.3.15 of \cite{Duflo97}, the almost sure convergence
\begin{equation}
\label{CVGMNSR}
\lim_{n \rightarrow \infty} M_n = M
\hspace{1cm} \text{where} \hspace{1cm}
M=\sum_{k=1}^\infty a_k \varepsilon_k.
\end{equation}
Moreover, we obtain from \eqref{DEFAN} that
\begin{equation}
\label{CVGANSR}
\lim_{n \rightarrow \infty} n^{2p-1} a_n = \Gamma(2p).
\end{equation}
Consequently, as $M_n=a_nS_n$, \eqref{T-ASCVG-SR1} immediately follows from
\eqref{CVGMNSR} and \eqref{CVGANSR}. One can observe that
$$
L=\frac{1}{\Gamma(2p)} \sum_{k=1}^\infty a_k \varepsilon_k.
$$
It only remains to prove convergence \eqref{T-ASCVG-SR2}. For that purpose, it is only necessary to show that
the martingale $(M_n)$ is bounded in $\mathbb{L}^4$.
It is not hard to see that $(M_n)$ is bounded in $\mathbb{L}^2$. As a matter of fact, since $M_{n+1}^2=(a_{n+1} \varepsilon_{n+1}+M_{n})^2$,
it follows from \eqref{UBCEEPS} that
$$\mathbb{E}[M_{n+1}^2| \mathcal{F}_{n}]=a_{n+1}^2 \mathbb{E}[\varepsilon_{n+1}^2| \mathcal{F}_{n}]+M_{n}^2 \leq
a_{n+1}^2 +M_{n}^2 \hspace{1cm} \text{a.s.}
$$
Taking expectation on both sides, we get that $\mathbb{E}[M_{n+1}^2] \leq a_{n+1}^2 +\mathbb{E}[M_{n}^2]$ leading to
\begin{equation}
\label{BL2}
\mathbb{E}[M_n^2] \leq \sum_{k=1}^n a_k^2.
\end{equation}
Consequently, we obtain from \eqref{CVGVNSR} together with \eqref{BL2} that
\begin{equation}
\sup_{n \geq 1} \mathbb{E}[M_n^2] \leq {}_{3}F_2\Bigl( \begin{matrix}
{1,1,1}\\
{2p,2p}\end{matrix} \Bigl| 1\Bigr) < \infty.
\label{MBL2}
\end{equation}
By the same token, as $M_{n+1}^4=(a_{n+1} \varepsilon_{n+1}+M_{n})^4$, we clearly have
\begin{equation}
\mathbb{E}[M_{n+1}^4 | \mathcal{F}_{n}] = \sum_{\ell=0}^4 \binom{4}{\ell} a_{n+1}^\ell M_{n}^{4-\ell} \mathbb{E}[\varepsilon_{n+1}^\ell | \mathcal{F}_{n}].
\label{CEM4}
\end{equation}
On the one hand, we already saw from \eqref{UBCEEPS} that
\begin{equation*}
\mathbb{E}[\varepsilon_{n+1}^2 | \mathcal{F}_{n}] \leq 1 \hspace{1cm} \text{and} \hspace{1cm}
\mathbb{E}[\varepsilon_{n+1}^4 | \mathcal{F}_{n}] \leq \frac{4}{3} \hspace{1cm} \text{a.s.}
\end{equation*}
On the other hand, as in \eqref{CEEPS2}, we also have
\begin{eqnarray}
\mathbb{E}[\varepsilon_{n+1}^3 | \mathcal{F}_{n}]
& = & 2(\gamma_n -1)^3S_n^3 -2 (\gamma_n -1)^2 S_n \hspace{1cm} \text{a.s.} \nonumber\\
& = & 2(2p-1) \frac{S_n}{n} \Bigl( (2p-1)^2 \Bigl(\frac{S_n}{n}\Bigr)^2 -1 \Bigr) \hspace{1cm} \text{a.s.}
\label{CEEPS3}
\end{eqnarray}
Therefore, as $M_n=a_nS_n$ and $|S_n| \leq n$, we obtain from \eqref{CEEPS3} that
$$
M_n \mathbb{E}[\varepsilon_{n+1}^3 | \mathcal{F}_{n}] \leq 0 \hspace{1cm} \text{a.s.}
$$
Consequently, it follows from \eqref{CEM4} that
$$
\mathbb{E}[M_{n+1}^4 | \mathcal{F}_{n}] \leq \frac{4}{3} a_{n+1}^4 + 6 a_{n+1}^2 M_n^2 + M_n^4
\hspace{1cm} \text{a.s.}
$$
Taking expectation on both sides, we find that
$$
\mathbb{E}[M_{n+1}^4] \leq \frac{4}{3} a_{n+1}^4 + 6 a_{n+1}^2 \mathbb{E}[M_n^2] + \mathbb{E}[M_n^4]
$$
leading, via \eqref{BL2}, to
$$
\mathbb{E}[M_n^4] \leq \frac{4}{3} \sum_{k=1}^n a_k^4 + 6 \sum_{k=1}^n a_k^2 \mathbb{E}[M_{k-1}^2]
\leq 6\Bigl(1+\sum_{k=1}^n a_k^2 \Bigr) \sum_{k=1}^n a_k^2.
$$
Finally, we can deduce from \eqref{CVGVNSR} that
\begin{equation}
\sup_{n \geq 1} \mathbb{E}[M_n^4] < \infty,
\label{MBL4}
\end{equation}
which completes the proof of Theorem \ref{T-ASCVG-SR}.
\hfill $\videbox$\\
\ \vspace{-1ex}\\
The proof of Theorem \ref{T-MOM-SR} relies on the following well-known lemma on sums
of ratio of gamma functions.
\begin{lem}
\label{L-RG}
For any non-negative real numbers $a$ and $b$ such that $b\neq a+1$
and for all $n\geq 1$, we have
\begin{equation}
\label{SUMGAMMA}
\sum_{k=1}^n \frac{\Gamma(k+a)}{\Gamma(k+b)}= \frac{\Gamma(n+a+1)}{(b-a-1)\Gamma(n+b)}
\left(\frac{\Gamma(n+b) \Gamma(a+1)}{\Gamma(n+a+1)\Gamma(b)} -1\right).
\end{equation}
\end{lem}
\noindent{\bf Proof of Theorem \ref{T-MOM-SR}.}
Denote $\alpha=2p-1$, $\beta=2q-1$,
$$
L_n=\frac{M_n}{\Gamma(2p)}
\hspace{1.5cm} \text{and} \hspace{1.5cm}
L=\frac{M}{\Gamma(2p)}.
$$
It follows from convergence \eqref{T-ASCVG-SR2} that for any integer $d=1,\ldots,4$
\begin{equation}
\label{LIMLD}
\lim_{n\rightarrow \infty} \mathbb{E}[L_n^d]=\mathbb{E}[L^d].
\end{equation}
First of all, we already saw that for all $n\geq 1$, $\mathbb{E}[S_{n+1}|\mathcal{F}_n]=\gamma_n S_n$ a.s. Consequently,
$$\mathbb{E}[S_{n+1}]=\gamma_n \mathbb{E}[S_n]=\Bigl(\frac{n+\alpha}{n}\Bigr)\mathbb{E}[S_n],$$
which leads to
\begin{equation}
\label{MOMSN1}
\mathbb{E}[S_n]=\prod_{k=1}^{n-1} \Bigl(\frac{k+\alpha}{k}\Bigr) \mathbb{E}[S_1]=
\frac{\beta \Gamma(n+\alpha)}{\Gamma(n) \Gamma(\alpha +1)}=
\frac{\beta}{a_n}.
\end{equation}
Hence, we immediately get from \eqref{MOMSN1} that
$$
\mathbb{E}[L_n]=\mathbb{E}[L]=
\frac{\beta}{\Gamma(\alpha +1)}
=\frac{2q-1}{\Gamma(2p)}.
$$
Next, taking expectation on both sides of \eqref{CES2}, we obtain that for all $n\geq 1$,
$$
\mathbb{E}[S_{n+1}^2] =1+ ( 2 \gamma_n -1)\mathbb{E}[S_{n}^2]=1+ \Bigl(\frac{n+2\alpha}{n}\Bigr)\mathbb{E}[S_n^2]
$$
which implies that
\begin{eqnarray*}
\mathbb{E}[S_n^2] & = &\frac{\Gamma(n+2\alpha)}{\Gamma(n)\Gamma(2\alpha +1)}\left(1+ \sum_{k=1}^{n-1}
\frac{\Gamma(k+1)\Gamma(2\alpha +1)}{\Gamma(k+2\alpha+1)}\right), \\
& = &
\frac{\Gamma(n+2\alpha)}{\Gamma(n)}\sum_{k=1}^n \frac{\Gamma(k)}{\Gamma(k+2\alpha)}.
\end{eqnarray*}
Therefore, we deduce from identity \eqref{SUMGAMMA} with $a=0$ and $b=2\alpha$ that
\begin{equation}
\label{MOMSN2}
\mathbb{E}[S_n^2]=\frac{n}{2\alpha -1}\left( \frac{\Gamma(n+2\alpha)}{\Gamma(n+1)\Gamma(2\alpha)} -1\right).
\end{equation}
Hence, we obtain from
\eqref{MOMSN2} that
\begin{equation*}
\mathbb{E}[L_n^2]=\frac{a_n^2 \mathbb{E}[S_n^2]}{(\Gamma(2p))^2}= \frac{n}{2\alpha -1} \left( \frac{\Gamma(n)}{\Gamma(n+\alpha)}\right)^2 \left( \frac{\Gamma(n+2\alpha)}{\Gamma(n+1)\Gamma(2\alpha)} -1\right)
\end{equation*}
which ensures, via \eqref{LIMLD} with $d=2$, that
$$
\lim_{n\rightarrow \infty} \mathbb{E}[L_n^2]=\mathbb{E}[L^2]=
\frac{1}{(2\alpha -1)\Gamma(2\alpha)}
=\frac{1}{(4p-3)\Gamma(2(2p-1))}.
$$
Furthermore, as $S_{n+1}^3=(S_n+X_{n+1})^3$, we obtain that
$$
\mathbb{E}[S_{n+1}^3| \mathcal{F}_n] = (3 \gamma_n -2)S_{n}^3 + (\gamma_n+2)S_n \hspace{1cm}\text{a.s}
$$
Consequently,
$$
\mathbb{E}[S_{n+1}^3] =
\Bigl(\frac{3n+\alpha}{n}\Bigr)\mathbb{E}[S_n]
+\Bigl(\frac{n+3\alpha}{n}\Bigr)\mathbb{E}[S_n^3].
$$
Hence, it follows from tedious but straighforward calculations that
\begin{equation}
\label{MOMSN3SIGN}
\mathbb{E}[S_n^3] =\frac{\beta\Gamma(n+3\alpha)}{\Gamma(n)\Gamma(3\alpha + 1)}\left(1+
\frac{3\Gamma(3\alpha +1)}{\Gamma(\alpha +1)}\xi_n\right)
\end{equation}
where
$$
\xi_n= \sum_{k=1}^{n-1} \Bigl(k+\frac{\alpha}{3}\Bigr)
\frac{\Gamma(k+\alpha)}{\Gamma(k+3\alpha+1)}
=\sum_{k=1}^{n-1}
\frac{\Gamma(k+\alpha+1)}{\Gamma(k+3\alpha+1)}
-\frac{2\alpha}{3}\sum_{k=1}^{n-1}
\frac{\Gamma(k+\alpha)}{\Gamma(k+3\alpha+1)}.
$$
However, we infer from \eqref{SUMGAMMA} with $a=\alpha$ or $a=\alpha+1$ and $b=3\alpha+1$ that
$$
\xi_n=\frac{1}{2\alpha-1} \left(
\frac{\Gamma(\alpha+2)}{\Gamma(3\alpha+1)} -
\frac{\Gamma(n+\alpha+1)}{\Gamma(n+3\alpha)}\right)
-\frac{1}{3}\left(
\frac{\Gamma(\alpha+1)}{\Gamma(3\alpha+1)} -
\frac{\Gamma(n+\alpha)}{\Gamma(n+3\alpha)}\right).
$$
Therefore, we obtain from \eqref{MOMSN3SIGN} that
\begin{eqnarray}
\mathbb{E}[S_n^3] & = &\frac{\beta\Gamma(n+3\alpha)}{\Gamma(n)\Gamma(3\alpha + 1)}\left(
\frac{3(\alpha +1)}{2\alpha -1}
-\frac{\Gamma(3\alpha+1)\Gamma(n+\alpha)}{\Gamma(\alpha+1)\Gamma(n+3\alpha)}\frac{3n+\alpha +1}{2\alpha -1}
\right), \nonumber\\
& = & \frac{\beta}{(2\alpha -1)\Gamma(n)}\left(
\frac{3(\alpha +1)\Gamma(n+3\alpha)}{\Gamma(3\alpha + 1)}
-\frac{\Gamma(n+\alpha)}{\Gamma(\alpha+1)}(3n+\alpha +1)
\right).
\label{MOMSN3}
\end{eqnarray}
Consequently, we obtain from
\eqref{MOMSN3} that
\begin{equation*}
\mathbb{E}[L_n^3]
\frac{\beta(\Gamma(n))^2}{(2\alpha -1)(\Gamma(n+\alpha))^3}
\left(
\frac{3(\alpha +1)\Gamma(n+3\alpha)}{\Gamma(3\alpha + 1)}
-\frac{\Gamma(n+\alpha)}{\Gamma(\alpha+1)}(3n+\alpha +1)
\right)
\end{equation*}
which leads, via \eqref{LIMLD} with $d=3$, to
$$
\lim_{n\rightarrow \infty} \mathbb{E}[L_n^3]=\mathbb{E}[L^3]=
\frac{\beta(\alpha+1)}{\alpha(2\alpha -1)\Gamma(3\alpha)}
=\frac{2p(2q-1)}{(2p-1)(4p-3)\Gamma(3(2p-1))}.
$$
By the same token, since $S_{n+1}^4=(S_n+X_{n+1})^4$, we obtain that
$$
\mathbb{E}[S_{n+1}^4] =
1+2\Bigl(\frac{3n+2\alpha}{n}\Bigr)\mathbb{E}[S_n^2]
+\Bigl(\frac{n+4\alpha}{n}\Bigr)\mathbb{E}[S_n^4],
$$
which implies that
\begin{equation}
\label{MOMSN4PN}
\mathbb{E}[S_n^4] =\frac{\Gamma(n+4\alpha)}{\Gamma(n)\Gamma(4\alpha + 1)}\left(1+
\frac{\Gamma(4\alpha +1)}{(2\alpha -1)\Gamma(2\alpha)}\bigl(P_n+6Q_n-6R_n\bigr)\right)
\end{equation}
where
\begin{eqnarray*}
P_n & = & (2\alpha -1)\Gamma(2\alpha) \sum_{k=1}^{n-1}
\frac{\Gamma(k+1)}{\Gamma(k+4\alpha+1)}, \\
Q_n & = & \sum_{k=1}^{n-1}
\Bigl(k+\frac{2\alpha}{3}\Bigr)
\frac{\Gamma(k+2\alpha)}{\Gamma(k+4\alpha+1)}, \\
R_n & = & \Gamma(2\alpha) \sum_{k=1}^{n-1}
\Bigl(k+\frac{2\alpha}{3}\Bigr)
\frac{\Gamma(k+1)}{\Gamma(k+4\alpha+1)}.
\end{eqnarray*}
We make use once again of identity \eqref{SUMGAMMA} with appropriate values of
$a$ and $b$, to find that
\begin{equation}
\mathbb{E}[S_n^4] = \frac{\Gamma(n+4\alpha)}{\Gamma(n)\Gamma(4\alpha + 1)}\left(
\frac{24\alpha(2\alpha(\alpha +1)-1)}{(2\alpha -1)^2 (4\alpha -1)}
- \frac{\Gamma(4\alpha+1)}{(2\alpha -1)^2 \Gamma(n+4\alpha)}\zeta_n \right)
\label{MOMSN4}
\end{equation}
where
$$
\zeta_n=\frac{2(3n+2(\alpha+1))}{\Gamma(2\alpha)}\Gamma(n+2\alpha)-
\frac{3n(4\alpha-1)+2(2\alpha^2+1)}{(4\alpha-1)}\Gamma(n+1).
$$
Finally, we deduce from \eqref{LIMLD} with $d=4$ that
\begin{eqnarray*}
\lim_{n\rightarrow \infty} \mathbb{E}[L_n^4] & = &\mathbb{E}[L^4]=
\frac{24\alpha(2\alpha(\alpha +1)-1)}{(2\alpha -1)^2 (4\alpha -1)\Gamma(4\alpha+1)}
= \frac{6(2\alpha(\alpha +1)-1)}{(2\alpha -1)^2 (4\alpha -1)\Gamma(4\alpha)}, \\
& = &\frac{6(8p^2-4p-1)}{(8p-5)(4p-3)^2\Gamma(4(2p-1))},
\end{eqnarray*}
which completes the proof of Theorem \ref{T-MOM-SR}.
\hfill $\videbox$\\
\vspace{-2ex}
\section*{Appendix C \\ Proofs of the asymptotic normality results}
\renewcommand{\thesection}{\Alph{section}}
\renewcommand{\theequation}{\thesection.\arabic{equation}}
\setcounter{section}{3}
\setcounter{equation}{0}
\vspace{-2ex}
\subsection*{}
\begin{center}
{\bf C.1. The diffusive regime.}
\end{center}
\ \vspace{-4ex}\\
\noindent{\bf Proof of Theorem \ref{T-AN-DR}.}
We shall make use of the central limit theorem for martingales given e.g. by Corollary 2.1.10 of \cite{Duflo97},
to establish the asymptotic normality \eqref{T-AN-DR1}. It follows from \eqref{T-ASCVG-DR1},
\eqref{CALIP} and \eqref{CVGVNDR} that
\begin{equation}
\label{CVG-IP-DR}
\lim_{n\rightarrow \infty}
\frac{\langle M \rangle_{n}}{v_n} = 1 \hspace{1cm} \text{a.s.}
\end{equation}
Hereafter, it only remains to prove that $(M_n)$ satisfies Lindeberg's condition, that is, for all $\varepsilon > 0$,
\begin{equation}
\label{LINDEBERG}
\frac{1}{v_n} \sum_{k=1}^{n} \mathbb{E} \Bigl[ | \Delta M_{k} |^2 \mathrm{I}_{| \Delta M_{k}|
\geq \varepsilon \sqrt{v_n}} \bigl| \mathcal{F}_{k-1} \Bigr] \build{\longrightarrow}_{}^{{\mbox{\calcal P}}} 0.
\end{equation}
We obtain from \eqref{UBCEEPS} that for all $\varepsilon > 0$,
\begin{eqnarray*}
\frac{1}{v_n} \sum_{k=1}^{n} \mathbb{E} \Bigl[ | \Delta M_{k} |^2 \mathrm{I}_{| \Delta M_{k}| \geq \varepsilon \sqrt{v_n}} \bigl| \mathcal{F}_{k-1} \Bigr]
& \leq & \frac{1}{\varepsilon^2v_n^2} \sum_{k=1}^{n} \mathbb{E} \left[ | \Delta M_{k} |^4 | \mathcal{F}_{k-1} \right], \\
& \leq & \sup_{1 \leq k \leq n } \mathbb{E}[\varepsilon_{k}^4 | \mathcal{F}_{k-1}] \frac{1}{\varepsilon^2v_n^2} \sum_{k=1}^{n} a_k^4, \\
& \leq & \frac{4}{3\,\varepsilon^2v_n^2} \sum_{k=1}^{n} a_k^4.
\end{eqnarray*}
However, we deduce from \eqref{CONDLIL} together with Kronecker's lemma that
\begin{equation*}
\lim_{n \rightarrow \infty} \frac{1}{v_n^2} \sum_{k=1}^{n} a_{k}^4=0,
\end{equation*}
which ensures that Lindeberg's condition is satisfied. Therefore, we can conclude from the central limit theorem
for martingales that
\begin{equation}
\label{CLTMN-DR}
\frac{1}{\sqrt{v_n}} M_{n} \build{\longrightarrow}_{}^{{\mbox{\calcal L}}} \mathcal{N} (0, 1).
\end{equation}
As $M_n=a_n S_n$ and $\sqrt{n}a_n$ is equivalent to $\sqrt{v_n(3-4p)}$, we find from
\eqref{CLTMN-DR} that
\begin{equation*}
\frac{S_n}{\sqrt{n}} \build{\longrightarrow}_{}^{{\mbox{\calcal L}}} \mathcal{N} \Bigl(0, \frac{1}{3-4p} \Bigr),
\end{equation*}
which is exactly what we wanted to prove.
\hfill $\videbox$\\
\vspace{-5ex}
\subsection*{}
\begin{center}
{\bf C.2. The critical regime.}
\end{center}
\ \vspace{-4ex}\\
\noindent{\bf Proof of Theorem \ref{T-AN-CR}.}
We shall now proceed to the proof of the asymptotic normality \eqref{T-AN-CR1}
making use, once again, of the central limit theorem for martingales given e.g. by Corollary 2.1.10 of \cite{Duflo97}.
We have from \eqref{T-ASCVG-CR1}, \eqref{CALIP} and \eqref{CVGVNCR} that
\begin{equation}
\label{CVG-IP-CR}
\lim_{n\rightarrow \infty}
\frac{\langle M \rangle_{n}}{v_n} = 1 \hspace{1cm} \text{a.s.}
\end{equation}
Moreover, it follows from \eqref{CVGVNCR} and \eqref{CVGANCR} that
$a_n^2v_n^{-1}$ is equivalent to $(n\log n)^{-1}$. However, it is well-known that
\begin{equation*}
\sum_{n=1}^{+\infty} \frac{1}{n^2(\log n)^2} < \infty.
\end{equation*}
Consequently, we clearly have
\begin{equation}
\label{CONDLILCR}
\sum_{n=1}^{+\infty} \frac{a_{n}^4}{v_n^2} < +\infty.
\end{equation}
We already saw in the proof of Theorem \ref{T-AN-DR} that \eqref{CONDLILCR} is enough
to check that $(M_n)$ satisfies Lindeberg's condition. Consequently, we can deduce from the
central limit theorem for martingales that
\begin{equation}
\label{CLTMN-CR}
\frac{1}{\sqrt{v_n}} M_{n} \build{\longrightarrow}_{}^{{\mbox{\calcal L}}} \mathcal{N} (0, 1).
\end{equation}
Hence, as $M_n=a_n S_n$ and $\sqrt{n \log n}a_n$ is equivalent to $\sqrt{v_n}$, we find from
\eqref{CLTMN-CR} that
\begin{equation*}
\frac{S_n}{\sqrt{n \log n}} \build{\longrightarrow}_{}^{{\mbox{\calcal L}}} \mathcal{N} (0,1),
\end{equation*}
which achieves the proof of Theorem \ref{T-AN-CR}.
\hfill $\videbox$\\
\vspace{-4ex}
\bibliographystyle{acm}
|
\section{Stellar activity}
Stellar activity is a collective term for a variety of magnetic phenomena observed in cool stars, i.e.\ stars with outer convective envelopes (spectral types mid-F to mid-M) or stars that are fully convective (mid-M and later). Manifestations of magnetic activity include the presence of flares, coronal mass ejections, chromospheres and coronae, starspots, and faculae.
All of these phenomena are ultimately driven by the stellar rotation through the magnetic dynamo \citep{Parker1955}. The differential rotation of a star, both latitudinally and radially, causes not only a long-term inversion of the global magnetic field polarity, but also the localized phenomena in the stellar atmosphere which make up the individual facets of magnetic activity. Understanding the evolution of stellar rotation, from the formation of stars through the gigayears of their lifetime, is therefore fundamental to our understanding of stellar magnetic activity.
One important factor of the rotational evolution of cool stars the the spin-down that occurs due to magnetic braking \citep{Schatzman1962}. This happens because cool stars shed an ionized stellar wind, which moves away from the star along the stellar magnetic field lines. Finally, it decouples from the magnetic field, and at this moment the angular momentum is carried out of the system by the stellar wind. This continuous loss of angular momentum causes a spin-down of the star over time, which can be studied observationally \citep{Barnes2003, Barnes2010, Meibom2015, vanSaders2016}. Typically, stars set out on the main sequence with short rotation periods of half a day to a few days, and spin down to long periods of ca.\ 30 days over a few gigayears in the case of the Sun, and much longer rotation periods of the order of 100 days for low-mass stars on the main-sequence at old ages \citep{Irwin2011}. Studying slow rotation of stars is observationally challenging, because the main observables of rotation (rotational broadening of spectral lines and photometric variability due to star spots on the stellar surface) provide only weak signatures in the slow-rotation regime.
As rotation slows down, the stellar magnetic activity decreases. While this qualitatively makes sense, since (differential) rotation is the motor for stellar activity, the physical details of this are not fully understood. For example, it is not clear how the rotational period of a star and the presence and duration of activity cycles (i.e.\ the 11-year activity cycle of the Sun) are related. Still, the overall effects of activity, such as coronal X-ray emission \citep{Guedel1997, Preibisch2005, Telleschi2005}, chromospheric line emission such as the Ca II H and K lines and the H alpha line \citep{Skumanich1972, Noyes1984, Mamajek2008, Reiners2012}, and photospheric variability \citep{Bastien2014, Stelzer2016} can be related to stellar rotation and also directly to stellar age.
However, magnetic braking is not the only physical effect that influences stellar rotation and activity over time, which brings us to tidal interaction.
\section{Tidal interaction}
Whenever we have two astronomical objects in close proximity to each other, tides start to play a role. For our topic of interest, examples for relevant systems are: a close binary system consisting of two stars, a star with a massive planet in a close orbit, or a planet-moon system. Tides cause a deformation of the involved bodies due to the gravitational force acting on them, and due to their rotation around their common center of mass. There are three main observational effects of tides in such systems: alignment of the spin axes perpendicular to the orbital plane; synchronization, meaning that over time the rotational periods of the bodies and the orbital period become equal; and circularization, meaning that bodies in an eccentric orbit slowly lose their eccentricity and adopt a circular orbit \citep{Zahn2008, Mathis2009}.
For our topic of stellar activity, the synchronization effect is the most relevant one. Therefore we take a more detailed look at how this effect plays out in different combinations of orbital and rotational periods. Let us assume a system of two bodies A and B in a close-in orbit, as depicted in Fig.\ \ref{tidal}. Both objects get deformed by tides, and here we focus on what happens to the central body (A) due to the tides.
Assume that A has a longer rotational period than the orbital period of B. B raises a tidal bulge on A, and because B moves faster on its orbit than A rotates (in terms of angular velocity), the tidal bulge on A will lag behind (see Fig.\ \ref{tidal} left side). The gravitational pull of B on this bulge will therefore induce a tidal torque, and pull A into a somewhat faster rotation. Angular momentum is transferred from the orbit of B to the spin of A. Since the total angular momentum of the system is conserved, the semi-major axis of B decreases as the angular momentum of its orbit decreases, meaning B spirals slowly closer to A.
In the opposite configuration, where A has a shorter rotational period than the orbit of B, angular momentum is transferred into the other direction. The tidal bulge on A runs ahead, and gets ``pulled back'' by B as it does so (see Fig.\ \ref{tidal} right side). A therefore slows down, and the angular momentum of the orbit of B increases. This causes B to move to a larger semi-major axis. This is actually what happens in the Earth-Moon system, where Earth's rotation period of one day is shorter than the Moon's orbital period of ca.\ 27.3~days: the Earth's rotation slows down, and the Moon moves outwards over time. This effect has a magnitude of ca.\ 38~mm increase of the Moon's semi-major axis per year, measured through laser reflectors left on the Moon during the Apollo program \citep{Chapront2002}.
\begin{figure}
\includegraphics[width=\textwidth]{1.jpg}
\caption{Example of tidal interaction in a two-body system. In the example to the left, the rotation period of the central object (A) is longer than the orbital period of the smaller body (B), in the example to the right the rotation period of the central object is shorter. Typically, for star-planet systems with Hot Jupiters and relatively old host stars, the left example is representative of the orbital and rotational periods observed.}
\label{tidal}
\end{figure}
\section{Activity in stellar binaries}
For binary stars in close orbits, the tidal synchronization has strong observable effects on their stellar activity. After the stellar rotation of both stars has synchronized with the orbital period of the binary, it stays locked to this period. Even though the stars still lose some angular momentum due to the stellar wind they expel, they are kept at a high rotation rate due to this tidal locking, i.e.\ over long time scales the orbital distance of the stellar binary decreases as angular momentum is lost \citep{Stepien1995}. For the activity, this means that even though the stars in a binary may have a relatively old age, they are still rotating at a period of a few days and are similarly magnetically active as a single star of that rotation period systems.
In addition to the magnetic activity from the rotation of the individual stars, there can be interactions of the stellar magnetic fields with each other, such as magnetic loops connecting the two stars. This can lead to further magnetic activity effects \citep{Siarkowski1996, Peterson2010}.
The activity of stars in close binaries has been investigated thoroughly in various activity observables. Flare rates of M dwarfs in close pairs with white dwarfs have been found to be higher than for single M dwarfs \citep{Morgan2016}, and also their ambient activity, i.e.\ the overall activity level outside of time-resolved flares, which is thought to be a superposition of smaller-scale activity events, is high \citep{Morgan2012}. An interesting observation is that in systems with somewhat larger orbital distances of a few AU, where tidal synchronization should not be relevant, there is still an elevated activity observed \citep{Meibom2007}; this may be due to differences in the stellar formation and a different rotation period with which these moderate-distance binaries set out on the main sequence. Large-distance binaries (semi-major axes of several hundred AU) do not show this effect. But generally, activity indicators are found to be high for tidally interacting close binaries with periods of a few days \citep{Schrijver1991}.
\section{Activity in planet-hosting stars}
Thinking of a star-planet system as a scaled-down version of stellar binaries, with a very small mass ratio of the components, lets one expect that there may be relevant tidal effects as well. One active line of research in the exoplanetary field is the study of how exoplanetary orbits evolve over time, and how quickly exoplanets may spiral into their host stars \citep{Penev2011, Jackson2010}. The tidal quality factor of stars, which specifies how quickly the kinetic energy of tidal deformations and waves is dissipated, is not well constrained yet by current theories and observations, and requires further study \citep{Zahn2008, Ogilvie2007, Penev2011}. A ``smoking gun'' of an inspiraling exoplanet in the form of an actual measurement of a decreasing orbital period is yet to be found.
Concerning the tidal effects on the activity of the host star, initial theoretical studies were performed early-on. The two main interaction scenarios were identified as tidal interaction and magnetic interaction \citep{Cuntz2000}. Magnetic interaction is expected to follow scenarios either similar to loop interactions in close binaries, or similar to the Jupiter-Io unipolar inductor interaction. For tidal interaction, both the general tidal spin-up (or spin-down) of a star and activity effects due to a tidally induced increased turbulence in the outer convection layer of the star were proposed.
The observational search for star-planet interactions has been challenging. Initial detections of magnetic star-planet interaction were reported for two out of 13 stars with Hot Jupiters, where the chromospheric emission in the Ca~II lines was observed to modulate with the planetary orbital period, not the stellar rotation period \citep{Shkolnik2005}. Later observational campaigns of those targets, however, showed that during those later epochs a modulation with the stellar rotation period was present \citep{Shkolnik2008, PoppenhaegerLenz2011}. Other magnetic effects like flare triggering or hot spots in the stellar chromosphere and corona were expected from theoretical investigations \citep{Lanza2008, Cohen2009}. For the Hot Jupiter host HD~189733, several small flares in X-rays and the UV were observed during the time shortly after the secondary transit \citep{Pillitteri2011}. As the orbit of that planet is circular and not eccentric, it is not obvious why a certain phase of the orbit should show preferential flaring (as opposed to consistent flaring during one full half of the orbit when a stellar hot spot would be visible). Later observations and modelling suggested that a plasma trail of infalling material from the planet onto the star may be the source of the high-energy emission, with the largest viewing cross-section shortly after the secondary transit \citep{Pillitteri2014}. Another possible magnetic interaction effects has been reported for the system HD~17156, which hosts a Jupiter in a strongly eccentric orbit. The system showed elevated X-ray emission during two periastron passages of the planet, and low X-ray emission during two apoastron passages \citep{Maggio2015}. This may be analogous to colliding magnetospheres observed for some young binary stars in eccentric orbits \citep{Getman2011, Getman2016}.
Systematic investigations of stellar activity in larger samples of planet-hosting stars have been performed. While initial studies observed a trend of stars with close-in and massive planets to be more active than stars with small and far-away planets \citep{Kashyap2008}, these trends have simultaneously a large scatter over the whole sample \citep{Poppenhaeger2010} and can in part be traced back to selection effects from the efficiency of planet-detection methods for active and inactive stars \citep{Poppenhaeger2011}. Some effects, especially for extremely close and massive planets, seem to be still present when strictly controlling for the spectral type of the sample stars \citep{Miller2015}; however, not all Hot Jupiters necessarily have an active host star \citep{Poppenhaeger2009, Miller2012, Pillitteri2014WASP18}.
\begin{figure}
\includegraphics[width=\textwidth]{2.jpg}
\caption{Examples of several wide binary systems where one star hosts a known planet, observed in the X-ray band with \textit{Chandra} and \textit{XMM-Newton}.}
\label{spi}
\end{figure}
One important point is that even if planets may increase the stellar activity through some form of star-planet interaction, the magnetic braking of the star due to its stellar wind is going on at the same time \citep{Penev2012}. Taking into acocunt the age of a star-planet system is therefore crucial in order to distinguish whether a star is active because it is relatively young (and would therefore be active no matter if there was a planet or not), or if the star is actually old and is only active because it has been influenced by its planet. Unfortunately, ages for single field stars with ages over a gigayear are hard to estimate \citep{Soderblom2010review}. One way around this problem is using wide stellar binaries in which one of the stars hosts a known planet. In a wide stellar binary, the two stars will have the same age, and their activity levels should be similar (after adjusting for differences due to stellar mass). If the planet-hosting star has a much higher activity level than the companion star, one can deduce that the high activity level is not due to youngness of the system, but due to a planetary influence. In a sample of 18 such systems (see some examples in Fig.\ \ref{spi}), the stars for which a strong tidal influence is expected from their planet preferentially display higher activity levels than their companion stars (\citet{Poppenhaeger2014}, Poppenhaeger et al.\ submitted). This effect is absent for stars with planets that are not expected to have a strong tidal influence on their host stars.
Systematic effects on stellar activity can have important consequences for exoplanets: since the atmospheric mass loss of exoplanets is thought to be driven by X-ray and extreme UV irradiation \citep{Lecavelier2004, Sanz-Forcada2010}, an elevated stellar activity level can lead to higher evaporation rates for planets. Indeed, extended planetary atmospheres and/or active atmospheric escape have been observed for several exoplanets \citep{Vidal-Madjar2003, Lecavelier2010, Poppenhaeger2013, Bourrier2013, Kulow2014, Ehrenreich2015}. For small exoplanets, such evaporation may lead to the total loss of their atmosphere \citep{Lopez2013, Poppenhaeger2012}. Especially for habitability considerations, such as for planets in the habitable zones around M dwarfs, this is an important concern \citep{Segura2010}. M dwarfs can produce frequent flares even at older ages \citep{Guedel2004, RobradePoppenhaeger2010, Davenport2016}. From a stellar perspective this is particularly interesting in the fully convective M dwarf regime, where a different dynamo than in the solar case needs to be present due to the lack of a stellar radiative core, and different models have been developed to investigate the possible magnetic field structures for these stars \citep{Browning2008, Yadav2015}. Especially since a habitable-zone exoplanet has been detected for the nearest neighbor of the Sun \citep{Anglada-Escude2016}, the fully convective M dwarf Proxima Centauri, investigations of stellar activity and its impact on exoplanet habitability will continue to be a prime concern for studying near-by exoplanets.
\section{Conclusion}
Magnetic activity is not only an interesting stellar phenomenon, but also an important topic for exoplanets. Tidal influences on stellar activity are well-known in stellar binaries, and there is some observational evidence accumulating that also massive planets in close-in orbits can influence the stellar activity. Further investigations into the observational magnitudes of tidal effects as well as into stellar tidal quality factors will be necessary for understanding of these systems.
|
\section{Introduction}
Collective Adaptive Systems (CAS) consist of a large number of entities with
decentralised control and varying degrees of complex autonomous behaviour. They form the basis of many modern {\em smart city} critical infrastructures. Consequently, their design requires support from
formal methods and scalable automatic tools based on solid mathematical foundations.
In~\cite{LLM15a,LLM14}, Latella et al. presented a scalable mean-field model-checking procedure for verifying bounded Probabilistic Computation Tree Logic (PCTL,~\cite{HaJ94}) properties of an individual\footnote{The technique can be applied also to a finite selection of individuals; in addition, systems with several distinct types of individuals can be dealt with; for the sake of simplicity, in the present paper we consider systems with many instances of a single individual only and we focus in the model-checking a single individual in such a context.} in the context of a system consisting of a large number of interacting objects. The model-checking procedure is implemented in the tool \FlyFast{\footnote{{\tt http://j-sam.sourceforge.net/}}}.
The procedure performs on-the-fly, mean-field, approximated model-checking based on the idea of fast simulation, as introduced in~\cite{BMM07}. More specifically,
the behaviour of a generic agent with $S$ states in a system with a {\em large} number $N$ of instances of the agent at given step (i.e. time) $t$
is approximated by $\otm(\boldsymbol{\mu}(t))$ where $\otm(\vct{m})$ is
the $S \times S$ probability transition matrix of
an (inhomogeneous) DTMC and $\boldsymbol{\mu}(t)$ is a vector of size $S$ approximating
the mean behaviour of (the rest of) the system at $t$; each element of $\boldsymbol{\mu}(t)$
is associated with a distinct state of the agent, say $C$, and gives an approximation of the fraction of instances of the agent that are in state $C$ in the global system, at step $t$.
Note that such an approximation is a {\em deterministic} one, i.e.
$\boldsymbol{\mu}$ is a {\em function} of the step $t$ (the exact behaviour of the rest of the
system would instead be a {\em large} DTMC in turn); note furthermore, that the above transition matrix
does not depend on $N$~\cite{LLM15a,LLM14}.
Recently, modelling and programming languages have been proposed specifically for autonomic computing systems and CAS~\cite{De+15,Bo+15}. Typically, in such frameworks, a system is composed of a set of independent {\em components} where a component is a process equipped also with a set of {\em attributes} describing features of the component. The attributes of a component can be {\em updated} during
its execution so that the association between attribute {\em names} and attribute {\em values}
is maintained in the dynamic {\em store} of the component. Attributes can be used in {\em predicates} appearing in language constructs for component interaction. The latter is thus typically modelled using
{\em predicate-based} output/input {\em multicast}, originally proposed in~\cite{Lat83}, and
playing a fundamental role in the interaction schemes of languages like SCEL~\cite{De+15} and \textsc{Carma}{}~\cite{Bo+15}.
In fact, predicate-based communication can be used by components to dynamically organise themselves
into ensembles and as a means to dynamically select partners for interaction. Furthermore,
it provides a way for representing component features, like for instance component location in space,
which are fundamental for systems distributed in space, such as CAS~\cite{LoH16}.
In~\cite{CLM16a} we proposed a front-end modelling language for \FlyFast{} that provides constructs for dealing with {\em components} and {\em predicate-based interaction}; in the sequel, the
language---which has been inspired by \textsc{Carma}{---} will be referred to as \Abf{,} which stands for for
Predicate-based Interaction for \FlyFast{.}
Components interact via predicate-based communication. Each component consists of a behaviour, modelled as a DTMC-like agent, like in \FlyFast{,} and a set of attributes. The attribute
name-value correspondence is kept in the current store of the component.
Actions are {\em predicate based multi-cast} output and input primitives; predicates are defined over attributes.
Associated to each action there is also an (atomic) probabilistic store-update.
For instance, assume components have an attribute named $\mathsf{loc}$ which takes values in the set of points
of a space, thus recording the current location of the component. The following action models a multi-cast via channel $\alpha$ to all
components in the same location as
the sender, making it change location randomly:
$\alpha^*[\mathsf{loc}=\my.\mathsf{loc}]\tuple{}\mathtt{Jump}$. Here $\mathtt{Jump}$ is assumed to randomly update the store and, in particular attribute $\mathsf{loc}$. The computational model is {\em clock-synchronous}, as in \FlyFast{,} but at the component level.
In addition, each component is equipped with a local {\em outbox}. The effect of an output action
$\alpha^*[\pi_r]\tuple{}\sigma$ is to deliver output label $\alpha\tuple{}$ to the local outbox,
together with the predicate $\pi_r$, which (the store of) the receiver components will be required to satisfy, as
well as the current store of the component executing the action; the current store
is updated according to update $\sigma$. Note that output actions are {\em non-blocking}
and that successive output actions of the same component overwrite its outbox.
An input action $\alpha^*[\pi_s]()\sigma$ by a component
will be executed with a probability which is proportional to the {\em fraction} of all those
components whose outboxes currently contain the label $\alpha\tuple{}$, a predicate $\pi_r$
which is satisfied by the component, and a store which satisfies in turn
predicate $\pi_s$. If such a fraction is zero, then the input action will not take place (input is blocking),
otherwise the action takes place, the store of the component is updated via $\sigma$, and its outbox
cleared.
\paragraph{Related Work}
CAS are typically {\em large} systems, so that the formal analysis of models for such systems
hits invariantly the state-space explosion problem. In order to mitigate this problem, the so called `on-the-fly' paradigm
is often adopted (see e.g.~\cite{Courcoubetis1992,BCG95,Hol04,GnM11}).
In the context of probabilistic model-checking several on-the-fly approaches have been proposed, among which~\cite{DIMTV04}, \cite{LLM14a} and~\cite{Ha+09}. In~\cite{DIMTV04}, a probabilistic model-checker is shown for the
{\em time bounded} fragment of PCTL. An on-the-fly approach for {\em full} PCTL model-checking is proposed in~\cite{LLM14a} where, actually, a specific {\em instantiation} is presented of an algorithm which is {\em parametric} with respect to the specific probabilistic processes modelling language and logic, and their specific semantics.
Finally, in~\cite{Ha+09} an on-the-fly approach is used for detecting a maximal relevant search depth in an infinite state space and then a {\em global} model-checking approach is used
for verifying bounded Continuous Stochastic Logic (CSL)~\cite{Az+00,Ba+03} formulas in a continuous time setting on the selected subset of states.
An on-the-fly approach by itself however, does not solve the challenging scalability problems that arise in truly large parallel systems, such as CAS.
To address this type of scalability challenges in probabilistic model-checking, recently, several approaches have been proposed. In~\cite{He+04,Gu+06} approximate probabilistic model-checking is introduced. This is a form of statistical model-checking that consists in the generation of random executions of an {\em a priori} established maximal length~\cite{LaL16}. On each execution the property of interest is checked and statistics are performed over the outcomes. The number of executions required for a reliable result depends on the maximal error-margin of interest. The approach relies on the analysis of individual execution traces rather than a full state space exploration and is therefore memory-efficient. However, the number of execution traces that may be required to reach a desired accuracy may be large and therefore time-consuming. The approach works for general models, i.e.
models where stochastic behaviour can also be non Markovian and that do not necessarily model populations of similar objects.
On the other hand, the approach is not independent from the number of objects involved.
As recalled above, in~\cite{LLM14} a scalable model-checking algorithm is presented that is based on mean-field approximation, for the verification of time bounded PCTL properties of an individual in the context of a system consisting of a large number of interacting objects. Correctness of the algorithm with respect to exact probabilistic model-checking has been proven in~\cite{LLM14} as well. Also this algorithm is actually an instantiation of the above mentioned parametric algorithm for (exact) probabilistic model-checking~\cite{LLM14a}, but the algorithm is instantiated on
(time bounded PCTL and) the {\em approximate}, mean-field, semantics of a population process modelling language.
It is worth pointing out that \FlyFast{} allows users to perform simulations of their system models and to analyse the latter using
their {\em exact} probabilistic semantics and {\em exact} PCTL model-checking. In addition, the tool
provides {\em approximate} model-checking for bounded PCTL, using the model semantics based on mean-field.
The work of Latella et al.~\cite{LLM14} is based on mean-field approximation in the {\em discrete time} setting;
approximated mean-field model-checking in the {\em continuous time} setting has been presented in the literature as well, where
the deterministic approximation of the global system behaviour is formalised as an initial value problem using a set of { differential}
equations.
Preliminary ideas on the exploitation of mean-field convergence in continuous time for model-checking were informally sketched in~\cite{Ko+12}, but no model-checking algorithms were presented. Follow-up work on the above mentioned approach can be found in~\cite{Ko+13} which relies on earlier results on fluid model-checking by Bortolussi and Hillston~\cite{BoH12b}, later published in~\cite{BoH15}, where a {\em global CSL} model-checking procedure is proposed for the verification of properties of a selection of individuals in a population, which relies on fast simulation results.
This work is perhaps closest related to~\cite{LLM14,LLM15a};
however their procedure exploits mean-field convergence and fast
simulation~\cite{DaN08,GaG10} in a {\em continuous} time setting---using a set of { differential} equations---rather than in a discrete time setting---where an inductive definition
is used. Moreover, that approach is based on an {\em interleaving} model of computation, rather than
a clock-synchronous one; furthermore, a {\em global } model-checking approach,
rather than an on-the-fly approach is adopted; it is also worth noting that the treatment of nested formulas,
whose truth value may change over time,
turns out to be much more difficult in the interleaving, continuous time, global model-checking approach
than in the clock-synchronous, discrete time, on-the-fly one.
\Abf{} has been originally proposed in~\cite{CLM16a}, where the complete formal, exact probabilistic, semantics
of the language have been defined. The semantics definition consists of three transition rules---one for
transitions associated with output actions, one for those associated with input actions, and one for transitions
to be fired with residual probability. The rules induce a transition relation among component states
and compute the relevant probabilities. From the component transition relation, a component one-step
transition probability matrix is derived, the elements of which may depend on the fractions of the components
in the system which are in a certain state. The system-wide one-step
transition probability matrix is obtained by product---due to independence assumptions---using the
above mentioned component probability matrix and the actual fractions in the current system global state.
In~\cite{CLM16a} a translation of \Abf{} to the model specification language
of \FlyFast{} has also been presented which makes \Abf{} an additional front-end for \FlyFast{} extending its applicability
to models of systems based on predicate-based interaction. In the above mentioned paper, correctness of the translation
has been proved as well. In particular, it has been shown that the probabilistic semantics of any \Abf{} model
are isomorphic to those of the translation of the model. In other words, the transition probability matrix
of (the DTMCs of) the two models is the same. A companion translation of bounded PCTL formulas
is also defined~\cite{CLM16a} and proven correct.
The notion of the outbox used in \Abf{} is reminiscent of the notion of the {\em ether} in PALOMA~\cite{FeH14}
in the sense that the collection of all outboxes together can be thought of as a kind
of ether; but such a collection is intrinsically distributed among the components so that
it cannot represent a bottleneck in the execution of the system neither a singularity point in the
deterministic approximation.
We are not aware of other proposals, apart from~\cite{CLM16a}, of probabilistic process languages,
equipped both with standard, DTMC-based, semantics and with mean-field ones,
that provide a predicate-based interaction framework, and
that are fully supported by a tool for probabilistic simulation, exact and mean-field model-checking.
We conclude this section recalling that
mean-field/fluid procedures are based on {\em approximations} of the
global behaviour of a system. Consequently, the techniques should be considered as {\em complementary}
to other, possibly more accurate but often not as scalable, analysis techniques for CAS, primarily those based on stochastic simulation, such as
statistical model-checking.
In this paper we present some details of \Abf{,} a translation to \FlyFast{} which simplifies that proposed in~\cite{CLM16a} and an approach to model reduction
based on probabilistic bisimulation for Inhomogeneous DTMCs. In Section~\ref{FFaAbfe}
we briefly present the main ingredients of the \Abf{} syntax and informal semantics, and
we recall those features of \FlyFast{} directly relevant for understanding
the translation of \Abf{} to the \FlyFast{} input language proposed in~\cite{CLM16a}.
A revised and simplified version of the translation is described in Section~\ref{Art}.
In Section~\ref{sec:stpLanguage} we introduce a simplified language for the definition of transition-probabilities in \Abf{} that allows us to define in Section~\ref{sec:Bisimilarity} a model reduction procedure of the translation result, based on a notion of bisimulation for the kind of IDTMCs of interest, introduced in Section~\ref{sec:Bisimilarity} as well. An example of application of the
procedure is presented in Section~\ref{RedEx}. Some conclusions are drawn in Section~\ref{Concl}.
A formal proof of decidability of the cumulative probability test for state-space reduction based on
bisimulation is provided in the Appendix.
\section{Summary on \Abf{} and \FlyFast{}}
\label{FFaAbfe}
In the following we present the main ingredients of \Abf{} and the features of \FlyFast{} relevant for the present paper.
\subsection{\Abf{}}
A \Abf{} system model specification $\Upsilon = (\os_{\Upsilon}, F_{\Upsilon}, \boldsymbol{\Sigma_0})^{(N)}$
is a triple where $F_{\Upsilon}$ is the set of relevant function definitions (e.g. store updates, auxiliary constants and functions),
$\os_{\Upsilon}$ is a set of state defining equations, and
$\boldsymbol{\Sigma_0}$ is the initial system state (an $N$-tuple
of component states, each of which being a 3-tuple $(C,\gamma,O)$ of agent state $C$,
store $\gamma$ and outbox $O$).
We describe the relevant details below referring to~\cite{CLM16a} for the formal definition probabilistic semantics
of the language.%
\noindent
The \Abf{} type system consists of floating point values and operations, as in \FlyFast{},
plus simple enumeration types for attributes, declared according to the syntax
$\attype\; \mathtt{<name>} \; \enum\; \mathtt{<id-list>}$.
$\mathtt{<id-list>}$ is a finite list of identifiers.
Of course, attributes can also take floating point values.
\begin{figure}
\noindent
{\footnotesize
$
\attype\; \mathtt{Space}\; \enum\; \mathtt{A,B,C,D};\\
\vdots\\
\const\; \mathtt{H} = 0.6;\\
\const\; \mathtt{L = 1- H};\\
\const\; \mathtt{Hdiv2 = H/2};\\
\const\; \mathtt{Ldiv2 = L/2};\\
\vdots\\
\mathbf{attribute} \; \mathsf{loc} : \mathtt{Space};\\
\vdots\\
\func\; \mathtt{Hr}(x:\mathtt{Space}): \mathtt{Space};\;
x\;
\endfunc;\\
\func\; \mathtt{N}(x:\mathtt{Space}): \mathtt{Space};\;
\mathbf{case} \; x \; \mathbf{of} \; \mathtt{A:A;B:B;C:B;D:A}\;
\endfunc;\\
\func\; \mathtt{S}(x:\mathtt{Space}): \mathtt{Space};\;
\mathbf{case} \; x \; \mathbf{of} \; \mathtt{A:D;B:C;C:C;D:D}\;
\endfunc;\\
\func\; \mathtt{E}(x:\mathtt{Space}): \mathtt{Space};\;
\mathbf{case} \; x \; \mathbf{of} \; \mathtt{A:A;B:A;C:D;D:D}\;
\endfunc;\\
\func\; \mathtt{W}(x:\mathtt{Space}): \mathtt{Space};\;
\mathbf{case} \; x \; \mathbf{of} \; \mathtt{A:B;B:B;C:C;D:C}\;
\endfunc;\\
\vdots\\
\func\; \mathtt{pHr}(x:\mathtt{Space}): \mathbf{float};\;
\mathbf{case} \; x \; \mathbf{of} \; \mathtt{A: H;B: L; C:H;D: L}\;
\endfunc;\\
\func\; \mathtt{pN}(x:\mathtt{Space}): \mathbf{float};\;
\mathbf{case} \; x \; \mathbf{of} \; \mathtt{A:0;B:0;C:Ldiv2;D:Hdiv2}\;
\endfunc;\\
\func\; \mathtt{pS}(x:\mathtt{Space}): \mathbf{float};\;
\mathbf{case} \; x \; \mathbf{of} \; \mathtt{A:Ldiv2;B:Hdiv2;C:0;D:0}\;
\endfunc;\\
\func\; \mathtt{pE}(x:\mathtt{Space}): \mathbf{float};\;
\mathbf{case} \; x \; \mathbf{of} \; \mathtt{A:0;B:Hdiv2;C:Ldiv2;D:0}\;
\endfunc;\\
\func\; \mathtt{pW}(x:\mathtt{Space}): \mathbf{float};\;
\mathbf{case} \; x \; \mathbf{of} \; \mathtt{A:Ldiv2;B:0;C:0;D:Hdiv2}\;
\endfunc;\\
\vdots\\
\mathbf{update} \; \mathtt{Jump}\\
\my.\mathsf{loc} := \mathtt{Hr}(\my.\mathsf{loc}) \; \with \; \mathtt{pHr}(\my.\mathsf{loc});\\
\my.\mathsf{loc} := \mathtt{N}(\my.\mathsf{loc}) \; \with \; \mathtt{pN}(\my.\mathsf{loc});\\
\my.\mathsf{loc} := \mathtt{S}(\my.\mathsf{loc}) \; \with \; \mathtt{pS}(\my.\mathsf{loc});\\
\my.\mathsf{loc} := \mathtt{E}(\my.\mathsf{loc}) \; \with \; \mathtt{pE}(\my.\mathsf{loc});\\
\my.\mathsf{loc} := \mathtt{W}(\my.\mathsf{loc}) \; \with \; \mathtt{pW}(\my.\mathsf{loc})\\
\endupdate
$
}
\caption{A fragment of $F_{SI}$.\label{exa:F:SEIR}}
\end{figure}
In Figure~\ref{exa:F:SEIR} the attribute type
$\mathtt{Space}$ is defined that consists of four values $\mathtt{A,B,C,D}$ modelling four locations.
Some auxiliary constants are defined, using the $\const$ construct inherited from \FlyFast{:}
$\const \; \mathtt{<name>} \, = \, \mathtt{<value>}$.
A \Abf{} store update definition has the following syntax\footnote{In~\cite{CLM16a} a slightly different syntax for store updates has been used.}:
\noindent
{\small
$
\mathbf{update}{} \; upd\\
\my.a_1:= e_{11}, \ldots , \my.a_k:= e_{k1} \with{} \; p_1;\\
\vdots\\
\my.a_1:= e_{1n}, \ldots , \my.a_k:= e_{kn} \with{} \; p_n\\
\endupdate
$\\
}
\noindent
where
$upd$ is the update name (unique within the system model specification),
$a_1, \ldots, a_k$ are the attribute names of the component,
$e_{11}, \ldots, e_{kn}$ and $p_1, \ldots, p_n$ are attribute/store-probability
expressions respectively, with syntax defined according to the grammars
$
e::= v_a \; \sep \; c_a \; \sep \; \my.a \; \sep \; fn_a(e_1,\ldots,e_m)
$ and
$
p::= v_p \; \sep \; c_p \; \sep \; fn_p(e_1,\ldots,e_m).
$
In the above definition of
attribute expressions $v_a$ is an attribute value (drawn from finite set ${\cal V}$ of attribute values),
$c_a$ is an attribute constant in ${\cal V}$ defined using the $\const$;
$a \in \SET{a_1, \ldots, a_k}$ is an attribute name
and $fn_a$ is an attribute function defined by the user in $F_{\Upsilon}$, which,
when applied to attribute expressions
$e_1,\ldots,e_m$ returns an attribute value;
the syntax for such function definitions $\mathit{afd}$ is given below:\\
\noindent
{\small
$
\mathit{afd} ::= \func \; fn_a (x_1:T1, \ldots, x_m:Tm) :T; \mathit{afb} \; \endfunc\\
$
$\mathit{afb} ::= e \; \sep
\mathbf{case} \; (x_1, \ldots, x_m) \; \mathbf{of}
(v_{a_{11}}, \ldots, v_{a_{m1}}): e_{1};
(v_{a_{12}}, \ldots, v_{a_{m2}}): e_{2};
\ldots
(v_{a_{1k}}, \ldots, v_{a_{mk}}): e_{k}$\\
}
\noindent
where $fn_a$ is the name of the attribute function, $x_1:T1, \ldots, x_m:Tm$ are its parameters and their relative types,
$T$ is the type of the result of $fn_a$; $e$, $e_{i}$ are attribute-expressions and $v_{a_{ij}}$ are attribute-values.
In Figure~\ref{exa:F:SEIR} attribute functions $\mathtt{N,S,E,W}$ are defined for
North, South, East, and West, such that $\mathtt{Space}$ models the Cartesian space with four quadrants:
$\mathtt{A=N(D)=E(B)}$, $\mathtt{B=N(C)=W(A)}$, and so on, as shown diagrammatically in Figure~\ref{exa:SI} {\em right}. Function $\mathtt{Hr}$ is the identity on $\mathtt{Space}$.
In the definition of store-probability expressions $v_p \in (0,1]$,
$c_p$ is a store-probability constant in $(0,1]$ defined using
the \FlyFast{} $\const$ construct, and $fn_p$ is a store-probability function
defined by the user in $F_{\Upsilon}$, which,
when applied to attribute expressions
$e_1,\ldots,e_m$ returns a probability value
The syntax for store-probability function definitions $\mathit{pfd}$ is
similar to that of attribute functions:\\% in the obvious way:\\
\noindent
{\small
$
\mathit{pfd} ::= \func \; fn_p (x_1:T1, \ldots, x_m:Tm) : \mathbf{float}; \mathit{pfb} \; \endfunc\\
$
$\mathit{pfb} ::= p \; \sep
\mathbf{case} \; (x_1, \ldots, x_m) \; \mathbf{of}
(v_{a_{11}}, \ldots, v_{a_{m1}}): p_{1};
(v_{a_{12}}, \ldots, v_{a_{m2}}): p_{2};
\ldots
(v_{a_{1k}}, \ldots, v_{a_{mk}}): p_{k}$\\
}
\noindent
where $fn_p$ is the name of the store-probability function, the result type is $\mathbf{float}$ (actually the range $[0,1]$)
$x_1:T1, \ldots, x_m:Tm$ are its parameters and their relative types,
$p$, $p_{i}$ are store-probability expressions and $v_{a_{ij}}$ are attribute-values.
In any store update definition it must be guaranteed that the values of $p_1 \ldots p_n$ sum up\footnote{In this version of the translation we allow only flat updates, i.e. the specific probability of each combination of values assigned
to the attributes must be given explicitly. Other possibilities could be defined using {\em combinations}
of (independent) probability distributions.} to $1$. The informal meaning is clear. The store update will make
attributes $a_1, \ldots, a_k$ take the values of $e_{1i}, \ldots, e_{ki}$ respectively with probability equal to the value of $p_i$.
In Figure~\ref{exa:F:SEIR} store-probability functions $\mathtt{pHr, pN, pS, pE, pW}$ are defined
that give the probabilities of not moving ($\mathtt{pHr}$), or of jumping to North ($\mathtt{pN}$), South ($\mathtt{pS}$), East ($\mathtt{pE}$), West ($\mathtt{pW}$), as functions of the current location.
\begin{example}
\label{exa}
A simplified version of the behaviour of the epidemic process discussed in~\cite{CLM16a}
is shown in Figure~\ref{exa:SI} {\em left}\footnote{We focus only on those features that are most relevant
for the present paper. In~\cite{CLM16a} also other features are shown like, e.g. the use of
(predicate-based) input actions, which are not the main subject of this paper.}.
In Figure~\ref{exa:F:SEIR} we show a fragment of $F_{SI}$ defining store update $\mathtt{Jump}$ together with the relevant type, constant and function definitions as introduced above. The component has just one attribute, named $\mathsf{loc}$, with values in $\mathtt{Space}$. The effect of $\mathtt{Jump}$ executed by a component
in which $\mathsf{loc}$ is bound to quadrant $\ell$ is to leave the value of $\mathsf{loc}$ unchanged with probability $\mathtt{pHr}(\ell)$, change it to the quadrant North of $\ell$ with probability
$\mathtt{pN}(\ell)$, and so on. Note that $\mathtt{H} > \mathtt{L}$ and this implies that
higher probability is assigned to $\mathtt{A}$ and $\mathtt{C}$ and low probability to $\mathtt{B}$ and $\mathtt{D}$.
This is represented in Figure~\ref{exa:SI} {\em right} where higher probability locations are shown in green and
lower probability ones are shown in red; moreover, the relevant probabilities are represented as arrows ($\mathtt{H/2}, \mathtt{L/2}$)
or self-loops ($\mathtt{H}, \mathtt{L}$).
A susceptible (state $\mathtt{S}$) component becomes infected (state $\mathtt{I}$) via
an $\mathtt{inf}$ action which takes place with probability equal to the fraction of components in the system which are currently infected (i.e. $\mathtt{frc(I)}$); it remains in state $\mathtt{S}$ via the self-loop
labelled by action $\mathtt{nsc}$, with probability $\mathtt{frc(S)=1- frc(I)}$. An infected node (state $\mathtt{I}$)
may recover, entering state $\mathtt{S}$ with action $\mathtt{rec}$ and probability $\mathtt{ir}$;
while infected, it keeps executing action $\mathtt{inf}$, with probability $\mathtt{ii}$.
Note that,
for the sake of simplicity, we use only {\em internal} actions, modelled by means of
{\em output} actions with predicate false ($\bot$). We assume that in the initial global state
all outboxes are non-empty; each contains the initial store of the specific component (i.e., its initial location), predicate $\bot$ and the empty tuple $\tuple{}$).
\end{example}
\begin{figure}
\begin{center}
\scriptsize
\begin{minipage}{2in}
$
\begin{array}{l l l }
\texttt{S} & \dfas
& \textsf{frc}\,(\texttt{I}):: \texttt{inf}^* [\bot]\tuple{} \texttt{Jump}. \texttt{I +}\\
&& \textsf{frc}\,(\texttt{S}):: \texttt{nsc}^* [\bot]\tuple{}\texttt{Jump}. \texttt{S}\\\\
\texttt{I} & \dfas
& ii:: \texttt{inf}^*[\bot]\tuple{}\texttt{Jump}. \texttt{I +}\\
&& ir:: \texttt{rec}^*[\bot]\tuple{}\texttt{Jump}. \texttt{S}\\\\\\\\\\\\\\\\
\end{array}
$
\end{minipage}
\hspace{0.5in}
\resizebox{1.5in}{!}{\includegraphics{FourQuad}}
\end{center}\vspace{-1in}
\caption{$SI$, a behavioural model.\label{exa:SI}}
\end{figure}
\noindent
A \Abf{} state defining equation has the following (abstract) form:
$
C:= \sum_{j \in J} [g_j]p_j::act_j.C_j
$
where either $[g_j]p_j$ is the keyword $\mathbf{rest}$ or:
\begin{itemize}
\item $g_j$ is a boolean expression $b$ which may depend on the current store, but not on the
current occupancy measure vector:
$
b::= \top \; \sep \; \bot \; \sep \; e \, \underline{\bowtie} \, e \; \sep \; \neg b \; \sep \; b \mbox{$\;\;\wedge\;\;$} b
$ and
$
e::= v_a \; \sep \; c_a \; \sep \my.a
$
where $\top$ ($\bot$) denotes the constant $\mathtt{true}$ ($\mathtt{false}$),
$\underline{\bowtie}\, \in \SET{\geq,>,\leq,<}$,
$v_a$ is an attribute value (drawn from finite set ${\cal V}$ of attribute values),
$c_a$ is an attribute constant in ${\cal V}$ defined using
the \FlyFast{} $\const$ construct, and $a$ is the name of an attribute of the component.
\item $p_j$ is a transition probability expression
$
p::= v_p \, \sep \, c_p \, \sep \, \textsf{frc}\,(C) \, \sep \, \textsf{frc}\,(\pi)\, \sep \, \prod_{i \in I} \; p_i \, \sep \, \sum_{i \in I} p_i \, \sep \, 1-p
$, for finite $I$,
where $v_p \in (0,1]$, $c_p$ a constant in $(0,1]$ defined via the $\const$ construct,
and $\pi$ is defined as $b$ above, but where expressions $e$ can also be attribute names $a$
(i.e. $e::= v_a \; \sep \; c_a \; \sep \; \my.a \; \sep \; a$);
$\textsf{frc}\,(C)$ is the fraction of components {\em currently} in state $C$ over the total number $N$;
similarly, $\textsf{frc}\,(\pi)$ is the fraction of components the {\em current}
store of which satisfies $\pi$, over the total number $N$. Note that it must be guaranteed
that $\prod_{i \in I} p_i \leq 1$ and $\sum_{i \in I} p_i \leq 1$.
\item $act_j$ can be an output action $\alpha^*[\pi]\tuple{}\sigma$ or an input action $\alpha^*[\pi]()\sigma$,
where $\pi$ is as above and $\sigma$ is the
name of a store update. Note that in the case of an input action, $\pi$ refers to the store
of the partner component in the {\em previous} step of the computation.
\end{itemize}
If $[g_j]p_j=\mathbf{rest}$, then $act_j$ must be an output action $\alpha^*[\pi]\tuple{}\sigma$,
to be executed with the residual probability.
\subsection{\FlyFast{}}
\FlyFast{} accepts a specification $\tuple{\os,A,\vct{C_0}}^{(N)}$ of a model of a system consisting of the clock-synchronous product of $N$ instances of a probabilistic agent. The states of the DTMC-like agent model are specified by a set of state-defining equations
$\os$. The (abstract) form of a state defining equation is the following
$
C:= \sum_{i=1}^r a_i.C_i$
where $a_i \in \act{}$---the set of \FlyFast{} {\em actions}---$C,C_i \in \sc{}$---the set of \FlyFast{} {\em states}---and, for $i,j=1,\ldots,r$ $a_i\not=a_j$ if $i\not=j$; note that $C_i=C_j$ with $i\not=j$ is allowed instead\footnote{The concrete \FlyFast{} syntax is:
{\tt state C$\{$a$\_$1.C$\_1$ + a$\_2$.C$\_2$ \ldots a$\_$r.C$\_$r$\}$}.}.
Each action has a probability assigned by means of
an action probability function definition in $A$ of the form $a:: exp$ where $exp$ is an expression consisting of constants
and $\textsf{frc}\,(C)$ terms. Constants are floating point
values or names associated to such values using the construct $\const \; \mathtt{<name>} \, = \, \mathtt{<value>}$;
$\textsf{frc}\,(C)$ denotes the element associated to state $C$
in the current occupancy measure vector\footnote{The occupancy measure vector is a vector
with as many elements as the number of states of an individual agent; the
element associated to a specific state gives the fraction of the subpopulation currently in that state over the size of the overall population. The occupancy measure vector is a compact representation of the system global state.}. So, strictly speaking, $\os$ and $A$ characterise an inhomogeneous DTMC whose probability matrix $\otm(\vct{m})$
is a function of the occupancy measure vector $\vct{m}$ such that for each pair of states $C,C'$,
the matrix element $\otm(\vct{m})_{C,C'}$ is the probability of jumping from $C$ to $C'$ given the current occupancy measure vector $\vct{m}$. Letting $\sc{\os}$ be the set
of states of the agent, with $|\sc{\os}|=S$,
and $\us{S} = \SET{(m_1,\ldots,m_S) | m_1+ \ldots + m_S=1}$ denote the unit simplex of dimension $S$, we have
$\otm: \us{S} \times \sc{\os} \times \sc{\os} \rightarrow [0,1]$. Matrix $\otm$ is generated directly from the
input specification $\tuple{\os,A,\vct{C_0}}^{(N)}$; the reader interested in the details of how to derive
$\otm$ is referred to~\cite{LLM14,LLM15a}.
Auxiliary function definitions can be specified in $A$. The initial state $\vct{C_0}$ is a vector of size $N$ consisting of
the initial state of each individual object.
Finally, note that in matrix $\otm(\vct{m})$ the information on specific actions is lost, which is common in PCTL/DTMC based approaches; furthermore, we note that, by construction, $\otm(\vct{m})$ does not depend on $N$ (see~\cite{LLM14,LLM15a} for details).
\section{A revised translation}
\label{Art}
As in~\cite{CLM16a}, we define a translation such that,
given a \Abf{} system specification $\Upsilon = ( \os_{\Upsilon}, F_{\Upsilon}, \boldsymbol{\Sigma_0})^{(N)}$,
the translation returns the \FlyFast{} system specification $\tuple{\os,A,\vct{C_0}}^{(N)}$
preserving probabilistic semantics.
The predicate-based \FlyFast{} front-end
is then completed with a simple translation at the PCTL level, for which we refer to~\cite{CLM16a}.
The system model specification translation consists of two phases.
In the first phase, each action in the input system model specification $\Upsilon$ is annotated with an identifier
which is unique within the specification. We let $\aleph(\Upsilon)$ denote the resulting specification.
These annotations will make action names unique specification-wide thus eliminating
complications which may arise from multiple occurrences of the same action, in particular when leading to the same state (see~\cite{CLM16a} for details).
Of course, these annotations are disregarded in the probabilistic semantics,
when considering the interaction model of components. In other words, an
output action $\alpha\tuple{}$ in outbox $(\gamma, \pi, \alpha\tuple{})$ must match with any input action
$\alpha()$ even if $\alpha\tuple{}$ would actually correspond to $(\alpha,\iota)^*[\pi]\tuple{}$ and $\alpha()$
would actually correspond to $(\alpha,\eta)^*[\pi']()$.
Apart from this detail, the probabilistic semantics as defined in~\cite{CLM16a} remain unchanged.
The second phase is defined by the translation algorithm defined in Figure~\ref{alg:trans},
which is a revised and simplified version of that presented in~\cite{CLM16a} and is applied to $\aleph(\Upsilon)$.
We let ${\cal I}(\aleph(\Upsilon))$ denote the result of the translation, namely the pure \FlyFast{} system
specification $\tuple{\os,A,\vct{C_0}}^{(N)}$.
%
We recall here some notation from~\cite{CLM16a}. We let $\sc{\os_{\Upsilon}}$ denote the
set of states of $\Upsilon$; $\Gamma_{\os_{\Upsilon}}$ is the set of all stores defined
over the attributes of $\Upsilon$---a store is a finite mapping from the attributes of
the component to a finite set of values ${\cal V}$, thus $\Gamma_{\os_{\Upsilon}}$ is finite---and ${\cal O}_{\os_{\Upsilon}}$ the finite set of all outboxes of $\Upsilon$.
A $\Upsilon$ {\em component-state} is a triple
$(C,\gamma, O) \in \sc{\os_{\Upsilon}} \times \Gamma_{\os_{\Upsilon}} \times {\cal O}_{\os_{\Upsilon}}=\Omega_{\os_{\Upsilon}}$.
If the component-state is the target of a transition modelling the execution of an {\em output} action,
then $O=(\gamma', \pi, \alpha\tuple{})$, where $\gamma'$ is the store of the (component-state) source of the transition,
$\pi$ is the predicate used in the action---actualised with $\gamma'$---and $\alpha\tuple{}$ the actual message sent by the action.
If, instead, the component-state is the target of a transition for an {\em input}
action, then $O=\tuple{}$, i.e. the empty outbox. Note that the set of component states of $\aleph(\Upsilon)$ is identical to that of $\Upsilon$.
Also the set of all stores of $\aleph(\Upsilon)$ is the same as
that of $\Upsilon$.
In the algorithm of Figure~\ref{alg:trans} by $t*t'$ we mean the {\em syntactical} term representing
the product of terms $t$ and $t'$; the notation is extended to $\mathsf{PROD}\SET{t | \mathsf{cond}(t)}$, denoting the {\em syntactical} product $t_1 * \ldots * t_n$
if $\SET{t|\mathsf{cond}(t) = \mbox{tt}}=\SET{t_1, \ldots, t_n}\not=\emptyset$
and $1$ otherwise. Similarly, $\mathsf{SUM}\SET{t | \mathsf{cond}(t)}$ denotes the {\em syntactical}
sum $t_1 + \ldots + t_n$ if $\SET{t|\mathsf{cond}(t) = \mbox{tt}}=\SET{t_1, \ldots, t_n}\not=\emptyset$
and $0$ otherwise.
The translation algorithm uses a few auxiliary functions which we briefly discuss below:
\begin{itemize}
\item ${\cal I}_{\sc{}}: \Omega_{\os_{\Upsilon}} \rightarrow \sc{}$ is a total injection which maps every component state of $\aleph(\Upsilon)$ to a distinct state of ${\cal I}(\aleph(\Upsilon))$; we recall that $\sc{}$ denotes the set of state names of \FlyFast{} models.
\item ${\cal I}_{\act{}}: (\sc{\os_{\Upsilon}} \times \Gamma_{\os_{\Upsilon}}) \times (\Lambda_{\os_{\Upsilon}} \times I_{\aleph}) \times \Omega_{\os_{\Upsilon}} \rightarrow \act{}$ is a total injection where, as in~\cite{CLM16a}, $\Lambda_{\os_{\Upsilon}}$ is the set of action labels of $\Upsilon$ and
$I_{\aleph}$ is the set of unique identifiers used in the first phase of the translation. We recall that
$\act{}$ is the set of action names of \FlyFast{.}
The mapping of actions is a bit more delicate because we have to respect
\FlyFast{} static constraints and, in particular, we have to avoid multiple
probability function definitions for the same action. A first source of potential violations (i.e. multiple
syntactical occurrences of the same action) has been removed by action annotation in the first phase of the translation.
A second source is the fact that the same action can take place in different contexts (for example
with different stores) or leading to different target component states (maybe with different probabilities).
To that purpose, we could distinguish different occurrences of the same action in different transitions, each characterised
by its source component-state and its target component-state in $\Omega_{\os_{\Upsilon}}$. In practice, since
an action of a component cannot be influenced by the current outbox of the component,
it is sufficient to restrict the first component of the domain from $\Omega_{\os_{\Upsilon}} $ to $(\sc{\os_{\Upsilon}} \times \Gamma_{\os_{\Upsilon}})$.
\item The interpretation functions defined in Figure~\ref{semint}, namely those depending
on stores only (and not on occupancy measure vectors); we assume $\semint{\mathbf{E_L}}{\cdot}_{\gamma}$
extended to $\semint{\mathbf{E_L}}{fn}_{\gamma}$ for defined function $fn$, in the standard way.
In Figure~\ref{semint} $\beta_{\Upsilon}$ denotes the constant to value bindings
generated by the $\const$ construct in the input model specification $\Upsilon$, whereas
store update $upd$ is defined as above.
\item The translation function ${\cal I}_{{\cal P}}$ for transition probability expressions $p_j$, defined in Figure~\ref{tprobexp}.
\end{itemize}
\begin{figure}[h!]
$
\begin{array}{l c l}
\semint{\mathbf{E_L}}{\top}_{\gamma} & = &\mbox{tt}\\
\semint{\mathbf{E_L}}{\bot}_{\gamma} & = &\mbox{ff}\\
\semint{\mathbf{E_L}}{e_1 \, \underline{\bowtie} \, e_2}_{\gamma} & = &
\semint{\mathbf{E_L}}{e_1}_{\gamma} \, \underline{\bowtie} \, \semint{\mathbf{E_L}}{e_2}_{\gamma} \\
\semint{\mathbf{E_L}}{\neg b}_{\gamma} & = &\neg \semint{\mathbf{E_L}}{b}_{\gamma}\\
\semint{\mathbf{E_L}}{b_1 \mbox{$\;\;\wedge\;\;$} b_2}_{\gamma} & = &
\semint{\mathbf{E_L}}{b_1}_{\gamma} \mbox{$\;\;\wedge\;\;$} \semint{\mathbf{E_L}}{b_2}_{\gamma}\\
\semint{\mathbf{E_L}}{v_a}_{\gamma} & = & v_a\\
\semint{\mathbf{E_L}}{c_a}_{\gamma} & = & \beta_{\Upsilon}(c_a)\\
\semint{\mathbf{E_L}}{v_p}_{\gamma} & = & v_p\\
\semint{\mathbf{E_L}}{c_p}_{\gamma} & = & \beta_{\Upsilon}(c_p)\\
\semint{\mathbf{E_L}}{a}_{\gamma} & = & a\\
\semint{\mathbf{E_L}}{\my.a}_{\gamma} & = & \gamma(a)
\end{array}
$\\
$
\begin{array}{l c l}
\semint{\mathbf{E_L}}{fn_a(e_1,\ldots,e_m)}_{\gamma} & = &
\semint{\mathbf{E_L}}{fn_a}_{\gamma} (\semint{\mathbf{E_L}}{e_1}_{\gamma}, \ldots, \semint{\mathbf{E_L}}{e_m}_{\gamma})\\
\semint{\mathbf{E_L}}{fn_p(e_1,\ldots,e_m)}_{\gamma} & = &
\semint{\mathbf{E_L}}{fn_p}_{\gamma} (\semint{\mathbf{E_L}}{e_1}_{\gamma}, \ldots, \semint{\mathbf{E_L}}{e_m}_{\gamma})
\end{array}
$\\\\
$
\begin{array}{l c l}
\semint{\mathbf{E_U}}{upd}_{\gamma} & = &
\lambda \gamma' . \mathtt{dom}(\gamma')\not=\SET{a_1,\ldots, a_k} \rightarrow 0;\\
& & \hspace{0.22in}
\gamma'(a_1) = \semint{\mathbf{E_L}}{e_{11}}_{\gamma} \mbox{$\;\;\wedge\;\;$} \ldots \mbox{$\;\;\wedge\;\;$} \gamma'(a_k) = \semint{\mathbf{E_L}}{e_{k1}}_{\gamma}
\rightarrow \semint{\mathbf{E_L}}{p_1}_{\gamma};\\
& & \hspace{0.22in} \vdots\\
& & \hspace{0.22in}
\gamma'(a_1) = \semint{\mathbf{E_L}}{e_{1n}}_{\gamma} \mbox{$\;\;\wedge\;\;$} \ldots \mbox{$\;\;\wedge\;\;$} \gamma'(a_k) = \semint{\mathbf{E_L}}{e_{kn}}_{\gamma}
\rightarrow \semint{\mathbf{E_L}}{p_n}_{\gamma};\\
& & \hspace{0.22in}
\mathbf{otherwise} \; \rightarrow \; 0
\end{array}
$\\\\
$
\begin{array}{l c l}
\semint{\mathbf{E_R}}{\top}_{\gamma} & = &\mbox{tt}\\
\semint{\mathbf{E_R}}{\bot}_{\gamma} & = &\mbox{ff}\\
\semint{\mathbf{E_R}}{e_1 \, \underline{\bowtie} \, e_2}_{\gamma} & = &
\semint{\mathbf{E_R}}{e_1}_{\gamma} \, \underline{\bowtie} \, \semint{\mathbf{E_R}}{e_2}_{\gamma} \\
\semint{\mathbf{E_R}}{\neg b}_{\gamma} & = &\neg \semint{\mathbf{E_R}}{b}_{\gamma}\\
\semint{\mathbf{E_R}}{b_1 \mbox{$\;\;\wedge\;\;$} b_2}_{\gamma} & = &
\semint{\mathbf{E_R}}{b_1}_{\gamma} \mbox{$\;\;\wedge\;\;$} \semint{\mathbf{E_R}}{b_2}_{\gamma}\\
\semint{\mathbf{E_R}}{v_a}_{\gamma} & = & v_a\\
\semint{\mathbf{E_R}}{c_a}_{\gamma} & = & \beta_{\Upsilon}(c_a)\\
\semint{\mathbf{E_R}}{a}_{\gamma} & = & \gamma(a)\\
\end{array}
$\\
\caption{\label{semint} Interpretation functions relevant for the translation}
\end{figure}
\begin{figure}
$
\begin{array}{l c l}
{\cal I}_{{\cal P}}(v_p)_{\gamma} & = & v_p\\
{\cal I}_{{\cal P}}(c_p)_{\gamma} & = & \beta_{\Upsilon}(c_p)\\
{\cal I}_{{\cal P}}(\textsf{frc}\,(C))_{\gamma} & = &
\mathsf{SUM}\SET{\textsf{frc}\, ({\cal I}_{\sc{}}((C',\gamma',O')))\;|\;(C',\gamma',O') \in \Omega_{\os_{\Upsilon}} \mbox{ and } C' = C}\\
{\cal I}_{{\cal P}}(\textsf{frc}\,(\pi))_{\gamma} & = &
\mathsf{SUM}\SET{\textsf{frc}\, ({\cal I}_{\sc{}}((C', \gamma',O')))\; |\; (C',\gamma',O') \in \Omega_{\os_{\Upsilon}} \mbox{ and }
\semint{\mathbf{E_R}}{\semint{\mathbf{E_L}}{\pi}_{\gamma}}_{\gamma'}=\mbox{tt}}\\
{\cal I}_{{\cal P}}(\prod_{i \in I} \; p_i)_{\gamma} & = & \mathsf{PROD}\SET{{\cal I}_{{\cal P}} (p_i)_{\gamma} \; | \; i \in I}\\
{\cal I}_{{\cal P}}(\sum_{i \in I} \; p_i)_{\gamma} & = & \mathsf{SUM}\SET{{\cal I}_{{\cal P}} (p_i)_{\gamma} \; | \; i \in I}
\end{array}
$
\caption{\label{tprobexp} Transition probability expressions translation function definition}
\end{figure}
\begin{figure}[!h]
\fbox{
\parbox{6.1in
{
\scriptsize
\noindent
For each state equation
$
C:= \sum_{j \in J} [g_j]p_j::act_j.C_j
$
in $ \os_{\Upsilon}$:
\begin{enumerate}
\item\label{T:OUT}
For each {\em output} action $(\alpha,\iota)^*[\pi]\tuple{}\sigma=act_k$ with $k \in J$ and $[g_k]p_k\not=\mathbf{rest}$,\\
for each $\gamma \in \Gamma_{\os_{\Upsilon}}$
s.t. $\semint{\mathbf{E_L}}{g_k}_{\gamma}=\mbox{tt}$
and $(C,\gamma, O) \in \Omega_{\os_{\Upsilon}}$ for some $O \in {\cal O}_{\os_{\Upsilon}}$,
for each $\gamma' \in \Gamma_{\os_{\Upsilon}}$ s.t.
$(C_k,\gamma',(\gamma, \semint{\mathbf{E_L}}{\pi}_{\gamma}, \alpha\tuple{})) \in \Omega_{\os_{\Upsilon}}$ and $\semint{\mathbf{E_U}}{\sigma}_{\gamma}(\gamma') >0$,
let
$\xi = {\cal I}_{\act{}}((C,\gamma),(\alpha\tuple{},\iota),(C_k,\gamma',(\gamma,\semint{\mathbf{E_L}}{\pi}_{\gamma},\alpha\tuple{})))$ be a fresh new action in the \FlyFast{} model specification
${\cal I}(\aleph(\Upsilon))= \tuple{\os,A,\vct{C_0}}^{(N)}$ and add the following action probability function definition in $A$:
$\xi::\semint{\mathbf{E_U}}{\sigma}_{\gamma}(\gamma') * {\cal I}_{{\cal P}}(p_k)_{\gamma}$.
Moreover, for each outbox $O \in {\cal O}_{\os_{\Upsilon}}$ s.t.
$(C,\gamma, O) \in \Omega_{\os_{\Upsilon}}$, the following summand is
added to the equation in $\os$ for state ${\cal I}_{\sc{}}((C,\gamma,O))$:
$
\xi. \,
{\cal I}_{\sc{}}((C_k,\gamma',(\gamma,\semint{\mathbf{E_L}}{\pi}_{\gamma},\alpha\tuple{})));
$
\item\label{T:IN}
For each {\em input} action $(\alpha,\iota)^*[\pi]()\sigma = act_k$, with $k \in J$ and $[g_k]p_k\not=\mathbf{rest}$,\\
for each $\gamma \in \Gamma_{\os_{\Upsilon}}$
s.t. $\semint{\mathbf{E_L}}{g_k}_{\gamma}=\mbox{tt}$
and $(C,\gamma, O) \in \Omega_{\os_{\Upsilon}}$ for some $O \in {\cal O}_{\os_{\Upsilon}}$,
for each $\gamma' \in \Gamma_{\os_{\Upsilon}}$ s.t.
$(C_k,\gamma',\tuple{}) \in \Omega_{\os_{\Upsilon}}$ and $\semint{\mathbf{E_U}}{\sigma}_{\gamma}(\gamma') >0$,\\
let
$\xi = {\cal I}_{\act{}}((C,\gamma), (\alpha(),\iota),(C_k,\gamma',\tuple{}))$,
be a fresh new action in the \FlyFast{} model specification
${\cal I}(\aleph(\Upsilon))= \tuple{\os,A,\vct{C_0}}^{(N)}$ and add the following action probability function definition in $A$:\\
$\xi:: \semint{\mathbf{E_U}}{\sigma}_{\gamma}(\gamma')* {\cal I}_{{\cal P}}(p_k)_{\gamma}*$\\
\mbox{ } \hspace{1in}$*\mathsf{SUM}\{\textsf{frc}\, ({\cal I}_{\sc{}}(\Sigma))| \Sigma = (C'', \gamma'',(\overline{\gamma},\overline{\pi},\alpha\tuple{}))
\in \Omega_{\os_{\Upsilon}} \mbox{$\;\;\wedge\;\;$}$\\
\mbox{ } \hspace{2.2in}$ \mbox{$\;\;\wedge\;\;$} \semint{\mathbf{E_R}}{\overline{\pi}}_{\gamma}=\semint{\mathbf{E_R}}{\semint{\mathbf{E_L}}{\pi}_{\gamma}}_{\overline{\gamma}}=\mbox{tt}\}$.\\
Moreover, for each outbox $O \in {\cal O}_{\os_{\Upsilon}}$ s.t.
$(C,\gamma, O) \in \Omega_{\os_{\Upsilon}}$, the following summand is
added to the equation in $\os$ for state ${\cal I}_{\sc{}}((C,\gamma,O))$:
$
\xi. \,
{\cal I}_{\sc{}}((C_k,\gamma',\tuple{}));
$\medskip
\item\label{T:RESTOUT}
If there exists $k\in J$ s.t. $[g_k]p_k = \mathbf{rest}$, and $act_k=(\alpha,\iota)^*[\pi]\tuple{}\sigma$,
for each $\gamma \in \Gamma_{\os_{\Upsilon}}$
s.t. $(C,\gamma, O) \in \Omega_{\os_{\Upsilon}}$ for some $O \in {\cal O}_{\os_{\Upsilon}}$,
let $A_{\gamma}$ be the set of probability function definitions which has been
constructed in steps (\ref{T:OUT}) and (\ref{T:IN}) above.
Let $q_{\gamma}$ be defined by $q_{\gamma} = (1 - \mathsf{SUM}\SET{q | \zeta :: r * q \in A_{\gamma}})$.
For all $\gamma' \in \Gamma_{\os_{\Upsilon}}$ s.t.
$(C_k,\gamma',(\gamma,\semint{\mathbf{E_L}}{\pi}_{\gamma},\alpha\tuple{})) \in \Omega_{\os_{\Upsilon}}$,
let
$\xi =
{\cal I}_{\act{}}((C,\gamma), (\alpha,\iota)\tuple{},(C_k,\gamma',(\gamma,\semint{\mathbf{E_L}}{\pi}_{\gamma},\alpha\tuple{})))\in \Omega_{\os_{\Upsilon}}$,
be a fresh new action in the \FlyFast{} model specification
${\cal I}(\aleph(\Upsilon))= \tuple{\os,A,\vct{C_0}}^{(N)}$ and add
the following action probability function definition in $A$:
$\xi :: \semint{\mathbf{E_U}}{\sigma}_{\gamma}(\gamma')* q_{\gamma}$.\\
Moreover, for each outbox $O \in {\cal O}_{\os_{\Upsilon}}$ s.t.
$(C,\gamma, O) \in \Omega_{\os_{\Upsilon}}$, the following summand is
added to the equation in $\os$ for state ${\cal I}_{\sc{}}((C,\gamma,O))$:
$
\xi.\,
{\cal I}_{\sc{}}((C_k,\gamma',(\gamma,\semint{\mathbf{E_L}}{\pi}_{\gamma},\alpha\tuple{})));
$\medskip
\item No other action probability function definition and transition is included and
the initial state $\vct{C_0}$ of ${\cal I}(\Upsilon)$ is defined as
$\vct{C_0}={\cal I}_{\sc{}}(\boldsymbol{\Sigma_0})$.
\end{enumerate}
$ $
\caption{The translation algorithm\label{alg:trans}}
\end{figure}
\noindent
Output actions are dealt with in step $1$ of the algorithm of Figure~\ref{alg:trans}.
Let us consider, for example,
$(\texttt{inf},1)^*[\bot]\tuple{}\texttt{Jump}$ in the definition of state {\tt S}
in Figure~\ref{exa:SI} (assuming annotations are integer values and the action has been annotated with $1$). We know that the possible values for locations are $\texttt{A,B,C,D}$,
so that the set of all stores is $\SET{\mathsf{loc}} \rightarrow \SET{\texttt{A,B,C,D}}$. The algorithm
generates 12 actions\footnote{Diagonal jumps are not contemplated in the model; technically this comes from the
actual probability values used in the definition of $\texttt{Jump}$.
}. Let us focus on the action $\xi$ associated to local position $\texttt{A}$ (i.e. $\gamma= [\mathsf{loc} \mapsto \texttt{A}]$) and possible next position $B$ (i.e. $\gamma'= [\mathsf{loc} \mapsto B]$); the algorithm will generate the \FlyFast{} probability function definition $\xi :: \mathtt{pW}(A)*{(\textsf{frc}\,(\mathtt{I1})+\ldots+\textsf{frc}\,(\mathtt{In}))}$\footnote{Here we assume that
${\cal I}_{\sc{}}(\SET{(C,\gamma,O) \in \Omega_{\os_{\Upsilon}}| \,C={\tt I}})=\SET{\mathtt{I1},\ldots,\mathtt{In} } \subset {\cal S}$.} as well as a transition leading to (a state which is the encoding, via ${\cal I}_{\sc{}}$, of) the component state with ${\tt I}$ as (proper) state, store $\gamma'$, and
outbox $(\gamma, \bot, \texttt{inf}\tuple{})$. Since the action is not depending on the current outbox,
in practice a copy of such a transition is generated {\em for each} component state
sharing the same proper state ${\tt S}$ and the same store $\gamma$.
The translation scheme for input actions is defined in case $2$ and is similar, except that
one has also to consider the sum of the fractions of the possible partners.
The translation of the $\mathbf{rest}$ case is straightforward. Note that for
every $\zeta :: r * q \in A_{\gamma}$, $r$ is a probability value associated to a store update;
since any store update characterizes a probability distribution
over stores, assuming the range of such a distribution is
$\SET{r_1,\ldots,r_n}$ if $\zeta_i :: r_i * q \in A_{\gamma}$, then also
$\zeta_j :: r_j * q \in A_{\gamma}$ for all $j=1,\ldots,n$, $j\not=i$ with
$\sum_{i=1}^n r_j = 1$. Thus the remaining probability is
$q_{\gamma} = (1 - \mathsf{SUM}\SET{q | \zeta :: r * q \in A_{\gamma}})$, where $q$ is either a
term ${\cal I}_{{\cal P}}(p_j)_{\gamma}$, with $p_j$ occurring in a summand of the state defining equation (see step \ref{T:OUT}), or a term ${\cal I}_{{\cal P}}(p_j)_{\gamma} * \mathsf{SUM}\SET{ \textsf{frc}\, ({\cal I}_{\sc{}}(\Sigma)) | \ldots }$ (see step \ref{T:IN}).
It worth pointing out here that the translation of Figure~\ref{alg:trans} is essentially the same as that
presented in~\cite{CLM16a}, when the latter is applied to the sublanguage of \Abf{} where one requires
that each action occurs at most once. The annotations performed in the
first phase of the translation ensure that this requirement is fulfilled;
as we noted above, these annotations are purely syntactical and are disregarded at the semantics level.
We also recall that in probabilistic, pure DTMC process language semantics, actions are in the end dropped and,
for each pair of states, the cumulative probability of such actions is assigned to the single transition from
one of the states to the other one. Consequently, correctness of the translation, proved in~\cite{CLM16a},
is preserved by the simplified version presented in this paper.
We note that in the algorithm sets $\Omega_{\os_{\Upsilon}}$, $ \Gamma_{\os_{\Upsilon}}$ and
${\cal O}_{\os_{\Upsilon}}$ are used.
Of course, an alternative approach could be one which considers only the set
$\overline{\Omega_{\os_{\Upsilon}}}$ of component states which are {\em reachable}
from a given initial component state and, consequently, the sets
$\overline{\Gamma_{\os_{\Upsilon}}}$ and $\overline{{\cal O}_{\os_{\Upsilon}}}$
of {\em used} stores and outboxes. In this way, the size of the resulting \FlyFast{}
model specification would be smaller (for example in terms of number of states).
On the other hand, this approach might require recompilation for each model-checking session
starting from a different initial component state.
\section{A simplified language for Bisimulation-based optimisation}
\label{sec:stpLanguage}
In this section we consider a simplified language for transition probability expressions appearing in
state defining equations that will allow us to perform
bisimulation based optimisation of the result $\tuple{\os,A,\vct{C_0}}^{(N)}$.
The restricted syntax for transition probability expressions $p$ we use in this section is the following:
$
p::= e_p \, \sep \, e_p \cdot\textsf{frc}\,(C)\, \sep \, e_p \cdot\textsf{frc}\,(\pi)
$ and
$
e_p ::= v_p \, \sep \, c_p
$
where $v_p$ and $c_p$ and $\pi$ are defined as in Section~\ref{FFaAbfe}.
By inspection of the \FlyFast{} translation as defined in Section~\ref{Art},
and recalling that the set $\sc{\os}$ of the states of the resulting \FlyFast{} model,
ranged over by $z, z_i, \ldots$, has cardinality $S$,
it is easy to see that the probability action definition in the result of a translation of a generic {\em output} action
is either of the form $\xi :: k$, or it is of the form
$\xi :: k*\mathsf{SUM}\SET{\textsf{frc}\,(z_i) | i \in I}$ where $k$ is a \FlyFast{} constant. Moreover, if $p$ was of the form $e_p \cdot\textsf{frc}\,(C)$, then
index set $I \subseteq \SET{1,\ldots, S}$ identifies those states in $\sc{\os}$
that represent (via ${\cal I}_{\sc{}}$) component states with proper local state $C$;
if instead, $p$ was of the form $e_p \cdot\textsf{frc}\,(\pi)$, then $I \subseteq \SET{1,\ldots, S}$
identifies those states in $\sc{\os}$ that represent (via ${\cal I}_{\sc{}}$) component states
with a store satisfying $\pi$ in the relevant store.
At the \FlyFast{} semantics level, recalling that $\textsf{frc}\,(z_i)$ is exactly
the $i$-th component $m_i$ of the occupancy measure vector $\vct{m}=(m_1, \ldots, m_S)$
of the model, we can rewrite\footnote{With a little notational abuse
using $k$ also as the actual value in $[0,1]$ of the \FlyFast{} constant $k$.} the above as $k$ or $k\cdot \sum_{i \in I}m_i$.
Similarly, the probability action definition in the result of a translation of a generic {\em input} action $(\alpha,\iota)^*[\pi']()$
(executed in local store $\gamma'$)
will necessarily be of the form
$k\cdot \left(\sum_{j \in I'} m_j\right)$
or of the form $k\cdot\left(\sum_{j \in I} m_j\right)\cdot \left(\sum_{j \in I'} m_j\right)$,
for index sets $I$ as above and $I'$ as follows:\\
$
I'=\{i \in \SET{1,\ldots, S} | \exists C, \gamma, \overline{\gamma}, \overline{\pi}, \, s.t.
$\\
\mbox{ }\hspace{0.7in}
$
z_i = {\cal I}_{\sc{}}((C,\gamma,(\overline{\gamma},\overline{\pi},\alpha\tuple{})))\mbox{$\;\;\wedge\;\;$}
\semint{\mathbf{E_R}}{\overline{\pi}}_{\gamma'}=
\semint{\mathbf{E_R}}{\semint{\mathbf{E_L}}{\pi'}_{\gamma'}}_{\overline{\gamma}}=\mbox{tt}
\}\\
$
An immediate consequence of using the above mentioned restricted syntax for the probability function definitions
is that, letting
$\otm: \us{S} \times \sc{} \times \sc{} \rightarrow [0,1]$
be the transition probability matrix for the \FlyFast{} translation of a model specification,
we have that $\otm(m_1,\ldots,m_S)_{z,z'}$ is a polynomial function of degree at most $2$
in variables $m_1, \ldots, m_S$.
\section{Bisimilarity and State-space Reduction}
\label{sec:Bisimilarity}
The following definition generalises standard probabilistic bisimilarity for state labelled DTMCs to the case in which transition probabilities are {\em functions} instead of constant values.
\begin{definition}
\label{bis}
For finite set of states $\sc{}$, with $|\sc{}|=S$, let $\otm: \us{S} \times \sc{} \times \sc{} \rightarrow [0,1]$ and,
for $z \in \sc{} $ and $Q \subseteq \sc{}$, write
$\otm(\vct{m})_{z,Q}$ for $\sum_{z' \in Q}\otm(\vct{m})_{z,z'}$.
Let furthermore ${\cal L}:\sc{} \rightarrow 2^{AP}$ be a state-labelling function,
for a given set $AP$ of atomic propositions.
An equivalence relation $R \subseteq \sc{} \times \sc{}$
is called a {\em bisimulation relation} if and only if
$z_1\, R \; z_2$ implies: (i) ${\cal L}(z_1) = {\cal L}(z_2)$ and (ii) $\otm(\vct{m})_{z_1,Q} = \otm(\vct{m})_{z_2,Q}$,
for all $\vct{m} \in \us{S}$ and $Q \in \sc{}/R$. The {\em bisimulation equivalence} on
$\sc{}$ is the largest bisimulation relation $R \subseteq \sc{} \times \sc{}$.
\end{definition}
We point out that the notion of bisimilarity does {\em not}
introduce any approximation, and consequently error, in a model and related analyses. Bisimilarity is
only a way for abstracting from details that are irrelevant for the specific analyses of interest. In particular,
it is useful to remark that bisimilarity preserves also state labels, which are directly related
to the atomic propositions of logic formulas for which model-checking is performed. Actually,
it is well known that probabilistic bisimilarity coincides with PCTL equivalence, i.e. the equivalence induced on system states
by their satisfaction of PCTL formulas, for finitely branching systems~\cite{BaK08}.
Note that $\otm(m_1,\ldots,m_S)_{z_1,Q} = \otm(m_1,\ldots,m_S)_{z_2,Q}$
for all $(m_1,\ldots,m_S) \in \us{S}$ is in general not decidable. If instead we consider
only transition probability matrices as in Section~\ref{sec:stpLanguage},
we see that each side of the above equality is a polynomial function of degree at most $2$
in variables $m_1, \ldots, m_S$ and
one can define a normal form for
the polynomial expressions in $m_1, \ldots, m_S$ supported by an ordering relation
on the variable names (e.g. $m_1\prec \ldots \prec m_S$) and get expressions of the general form
$
\left(\sum_{i=1}^{S} \sum_{j \geq i}^{S} h_{ij}\cdot m_i\cdot m_j\right) +
\left(\sum_{i=1}^{S}h_{i}\cdot m_i \right) +
h
$
for suitable $h_{ij},h_{i},h$. Actually, such expressions can always be rewritten
in the form
$
\left(\sum_{i=1}^{S} \sum_{j \geq i}^{S} u_{ij}\cdot m_i\cdot m_j\right) + u
$
for suitable $u_{ij},u$ since, recalling that $\sum_{i=1}^{S} m_i = 1$, we get
$
\sum_{i=1}^{S}h_{i}\cdot m_i =
\left(\sum_{i=1}^{S} m_i\right)\cdot \left(\sum_{i=1}^{S}h_{i}\cdot m_i\right)
$
which, by simple algebraic manipulation, yields an expression of the following form:
$
\left(\sum_{i=1}^{S} \sum_{j \geq i}^{S} u'_{ij}\cdot m_i\cdot m_j\right)
$; finally, we get
$
\left(\sum_{i=1}^{S} \sum_{j \geq i}^{S} u_{ij}\cdot m_i\cdot m_j\right) + u
$
by letting $u_{ij} = h_{ij} + u'_{ij}$ and $u=h$.
The following proposition thus establishes decidability of
$\otm(m_1,\ldots,m_S)_{z_1,Q} = \otm(m_1,\ldots,m_S)_{z_2,Q}$
for all $(m_1,\ldots,m_S) \in \us{S}$ for transition probability matrices as in Section~\ref{sec:stpLanguage}:
\begin{proposition}\label{prop1}$ $\\
Let
$
A(m_1,\ldots,m_S)=
\left(\sum_{i=1}^{S} \sum_{j \geq i}^{S} a_{ij}\cdot m_i\cdot m_j\right) +
a
$
and
$
B(m_1,\ldots,m_S)=
\left(\sum_{i=1}^{S} \sum_{j \geq i}^{S} b_{ij}\cdot m_i\cdot m_j\right) +
b
$
with $a_{ij}, b_{ij},
a, b \in \mathbb{R}$,
where $m_1,\ldots,m_S$ are variables taking values over $\mathbb{R}_{\geq 0}$
with $\sum_{i=1}^S m_i = 1$.
The following holds:
$
(\forall m_1,\ldots,m_S. A(m_1,\ldots,m_S) = B(m_1,\ldots,m_S))
\Leftrightarrow
(
(\forall i,j = 1, \ldots, S$ with $i \leq j. a_{ij} = b_{ij}) \wedge
a=b
).
$
\end{proposition}
The above results can be used for reduction of the state-space of the individual agent, i.e.
the resulting \FlyFast{} model specification, after the application of the
translation described in Section~\ref{Art}, by using for instance the
standard probabilistic relational coarsest set partition problem algorithm
(see e.g. \cite{HuT92}, page 227)
with slight obvious
modifications due to the presence of state-labels and the need of symbolic computation capabilities required for checking (degree 2) polynomial expressions
equality.\footnote{For instance, on page 227 of~\cite{HuT92}, line 12,
$v(x,S) = v(y,S)$ should be replaced with $L(x) = L(y) \wedge v(x,S) = v(y,S)$ and
in line 13, $v(x,S) \not= v(y,S)$ should be replaced with $L(x) \not= L(y) \vee v(x,S) \not= v(y,S),$ in order to take state labels into consideration as well. Of course $v(x,S)$ ($L$, respectively)is to be intended as $\otm(\vct{m})_{z,Q}$ (${\cal L}$, respectively), using the notation we introduced above for Bisimilarity.
}
It is worth mentioning that state aggregation via bisimilarity is effective only if there is
some sort of compatibility between (i) state labelling---and, consequently, the specific PCTL atomic propositions one uses---and (ii) the way probabilities are assigned to transitions---and, consequently, the cumulative probabilities to equivalence classes. We will come back on this issue in the following section.
\section{Example}
\label{RedEx}
\begin{figure}
\centerline{
\resizebox{1.5in}{!}{\includegraphics{OrizTwoQuad}}
}
\caption{$SI$ in two quadrants \label{pic:TwoQuad}}
\end{figure}
The application of the translation to the specification of Example~\ref{exa} generates an agent model with 8 states, say
$\sc{\os} = \SET{SA, SB, SC, SD, IA, IB, IC, ID}$\footnote{Actually the agent resulting from the
translation of Figure~\ref{alg:trans} has a higher number of states due to the different
possibilities for outbox values. Many of these states are unreachable from the initial
state since the agent has no input action and we assume an initial unreachable state pruning has been performed.} with associated IDTMC probability transition matrix as
shown in Figure~\ref{ex:PTM} where
$m_{xy}$ represents the fraction of objects currently in state $xy$ for $x\in \SET{S,I}$
and $y\in \SET{A,B,C,D}$---i.e. the components in state $x$ and with $\mathsf{loc}=y$ in the original
specification of Fig~\ref{exa:SI}---so that
$\mathbf{m}=(m_{SA}, m_{SB}, m_{SC}, m_{SD}, m_{IA}, m_{IB}, m_{IC}, m_{ID})$ is the
occupancy measure vector. In Figure~\ref{ex:PTM}, functions $\phi_S$ and $\phi_I$ are used as abbreviations in the obvious way:
$\phi_S(\mathbf{m}) = m_{SA} + m_{SB} + m_{SC} + m_{SD}$ and
$\phi_I(\mathbf{m}) = m_{IA} + m_{IB} + m_{IC} + m_{ID}$. Let us assume now that
we are interested in checking PCTL formulas on the model of Fig~\ref{exa:SI} which distinguish
components located in $A$ or $C$ from those located in $B$ or $D$, and those
in state $S$ from those in state $I$, that is we consider
atomic propositions $Sh,Ih,Sl$ and $Il$ and a labelling ${\cal L}$ such that
${\cal L}(SA)={\cal L}(SC)=\SET{Sh}$, ${\cal L}(IA)={\cal L}(IC)=\SET{Ih}$,
${\cal L}(SB)={\cal L}(SD)=\SET{Sl}$, and ${\cal L}(IB)={\cal L}(ID)=\SET{Il}$.
\def\phi_S(\mathbf{m}){\phi_S(\mathbf{m})}
\def\phi_I(\mathbf{m}){\phi_I(\mathbf{m})}
\begin{figure}
\scriptsize
$$
\begin{array}{ c | c | c | c | c | c | c | c | c |}
& SA & SB & SC & SD & IA & IB & IC & ID\\\hline
SA & H\phi_S(\mathbf{m}) & \frac{L}{2}\phi_S(\mathbf{m}) & 0 & \frac{L}{2}\phi_S(\mathbf{m}) &H\phi_I(\mathbf{m}) & \frac{L}{2}\phi_I(\mathbf{m}) & 0 & \frac{L}{2}\phi_I(\mathbf{m})\\\hline
SB &\frac{H}{2}\phi_S(\mathbf{m}) & L\phi_S(\mathbf{m}) & \frac{H}{2}\phi_S(\mathbf{m}) & 0 & \frac{H}{2}\phi_I(\mathbf{m}) & L\phi_I(\mathbf{m}) & \frac{H}{2}\phi_I(\mathbf{m}) &0\\\hline
SC & 0 & \frac{L}{2}\phi_S(\mathbf{m}) & H\phi_S(\mathbf{m}) & \frac{L}{2}\phi_S(\mathbf{m}) & 0 & \frac{L}{2}\phi_I(\mathbf{m}) & H\phi_I(\mathbf{m}) & \frac{L}{2}\phi_I(\mathbf{m})\\\hline
SD & \frac{H}{2}\phi_S(\mathbf{m}) & 0 & \frac{H}{2}\phi_S(\mathbf{m}) & L\phi_S(\mathbf{m}) & \frac{H}{2}\phi_I(\mathbf{m}) & 0 & \frac{H}{2}\phi_I(\mathbf{m}) & L\phi_I(\mathbf{m}) \\\hline
IA & H ir & \frac{L}{2} ir & 0 & \frac{L}{2} ir & H ii & \frac{L}{2} ii & 0 & \frac{L}{2} ii \\\hline
IB & \frac{H}{2} ir & L ir & \frac{H}{2} ir & 0 & \frac{H}{2} ii & L ii & \frac{H}{2} ii & 0\\\hline
IC & 0 & \frac{L}{2} ir & H ir & \frac{L}{2} ir & 0 & \frac{L}{2} ii & H ii & \frac{L}{2} ii\\\hline
ID & \frac{H}{2} ir & 0 & \frac{H}{2} ir & L ir & \frac{H}{2} ii & 0 & \frac{H}{2} ii & L ii \\\hline
\end{array}
$$
\caption{IDTMC transition probability matrix $\otm(\vct{m})$, for $\vct{m}$ in $\us{8}$.\label{ex:PTM}}
\end{figure}
Consider relation $R$ on $\sc{\os}$ defined as
$
R = I_{{\sc{\os}}} \cup \SET{(SA,SC),(SB,SD),(IA,IC),(IB,ID)} \cup \SET{(SC,SA),\\(SD,SB),(IC,IA),(ID,IB)}
$
where $I_{{\sc{\os}}}$ is the identity relation on $\sc{\os}$.
It is very easy to show that $R$ is a bisimulation according to Definition~\ref{bis}. Clearly $R$ is an equivalence relation and
its quotient $\sc{\os}/R$ is the set $\SET{Q_{Sh},Q_{Sl},Q_{Ih},Q_{Il}}$ with
$Q_{Sh}=\SET{SA,SC},Q_{Sl}= \SET{SB,SD},Q_{Ih}=\SET{IA,IC},Q_{Il}=\SET{IB,ID}$.
In addition, for all $z_1, z_2 \in \sc{\os}$, whenever $z_1 \, R \, z_2$, we have
${\cal L}(z_1)={\cal L}(z_2)$ and $\otm(\vct{m})_{z_1,Q}=\otm(\vct{m})_{z_2,Q}$ for
all $Q \in \sc{\os}/R$ and for all $\vct{m}$, as one can easily check; clearly, $R$ is also
the largest bisimulation relation on $\sc{\os}$.
The relationship between the two occupancy measure vectors is:
$m_{{Q_{Sh}}} = m_{SA} + m_{SC}$,
$m_{{Q_{Sl}}} = m_{SB} + m_{SD}$,
$m_{{Q_{Ih}}} = m_{IA} + m_{IC}$, and
$m_{{Q_{Il}}} = m_{IB} + m_{ID}$.
We can thus use the reduced IDTMC defined by matrix $\widehat{\otm}(m_{{Q_{Sh}}},m_{{Q_{Sl}}},m_{{Q_{Ih}}},m_{{Q_{Il}}})$
shown in Figure~\ref{ex:RPTM}. It corresponds to the \FlyFast{} agent specification $\widehat{\os}$ given in Figure~\ref{RedFFS}. In a sense, the high probability locations $A$ and $C$, in the new model,
have collapsed into a single one, namely $h$ and the low probability ones ($B$ and $D$) have
collapsed into $l$, as shown in Figure~\ref{pic:TwoQuad}. We point out again the correspondence
between the symmetry in the space jump probability on one side and the definition of the state labelling function on the other side. Finally, note that a coarser labelling like, e.g.
${\cal L}'(SA)={\cal L}'(SC)={\cal L}'(IA)={\cal L}'(IC)=\SET{h}$,
${\cal L}'(SB)={\cal L}'(SD)={\cal L}'(IB)={\cal L}'(ID)=\SET{l}$
would make the model collapse into one with only two states, $Q_h$ and $Q_l$,
with probabilities $H:Q_h \rightarrow Q_h, H:Q_l \rightarrow Q_h$,
$L:Q_h \rightarrow Q_l$ and $L:Q_h \rightarrow Q_l$ where only the location would be
modelled whereas information on the infection status would be lost.
\begin{figure}
\scriptsize
\begin{verbatim}
action QSh_inf_QIh: H*(frc(QIh)+frc(QIl)); action QSh_nsc_QSh: H*(frc(QSh)+frc(QSl));
action QSh_inf_QIl: L*(frc(QIh)+frc(QIl)); action QSh_nsc_QSl: L*(frc(QSh)+frc(QSl));
action QSl_inf_QIh: H*(frc(QIh)+frc(QIl)); action QSl_nsc_QSh: H*(frc(QSh)+frc(QSl));
action QSl_inf_QIl: L*(frc(QIh)+frc(QIl)); action QSl_nsc_QSl: L*(frc(QSh)+frc(QSl));
action QIh_inf_QIh: H*ii; action QIh_rec_QSh: H*ir;
action QIh_inf_QIl: L*ii; action QIh_rec_QSl: L*ir;
action QIl_inf_QIh: H*ii; action QIl_rec_QSh: H*ir;
action QIl_inf_QIl: L*ii; action QIl_rec_QSl: L*ir;
state QSh{QSh_inf_QIh.QIh + QSh_inf_QIl.QIl + QSh_nsc_QSh.QSh + QSh_nsc_QSl.QSl}
state QSl{QSl_inf_QIh.QIh + QSl_inf_QIl.QIl + QSl_nsc_QSh.QSh + QSl_nsc_QSl.QSl}
state QIh{QIh_inf_QIh.QIh + QIh_inf_QIl.QIl + QIh_rec_QSh.QSh + QIh_rec_QSl.QSl}
state QIl{QIl_inf_QIh.QIh + QIl_inf_QIl.QIl + QIl_rec_QSh.QSh +QIl_rec_QSl.QSl}
\end{verbatim}
\caption{Reduced agent specification $\widehat{\os{}}$\label{RedFFS}}
\end{figure}
\begin{figure}
\scriptsize
$$
\begin{array}{ c | c | c | c | c |}
& Q_{Sh} & Q_{Sl} & Q_{Ih} & Q_{Il} \\\hline
Q_{Sh} & H\cdot(m_{{Q_{Sh}}}+m_{{Q_{Sl}}}) & L \cdot (m_{{Q_{Sh}}}+m_{{Q_{Sl}}}) & H \cdot (m_{{Q_{Ih}}}+m_{{Q_{Il}}}) & L \cdot (m_{{Q_{Ih}}}+m_{{Q_{Il}}}) \\\hline
Q_{Sl} & H\cdot(m_{{Q_{Sh}}}+m_{{Q_{Sl}}}) & L \cdot (m_{{Q_{Sh}}}+m_{{Q_{Sl}}}) & H \cdot (m_{{Q_{Ih}}}+m_{{Q_{Il}}}) & L \cdot (m_{{Q_{Ih}}}+m_{{Q_{Il}}}) \\\hline
Q_{Ih} & H \cdot ir & L\cdot ir & H\cdot ii & L\cdot ii \\\hline
Q_{Il} & H\cdot ir & L\cdot ir & H\cdot ii & L\cdot ii \\\hline
\end{array}
$$
\caption{IDTMC transition probability matrix function $\widehat{\otm}(\vct{m})$, for $\vct{m}$ in $\us{4}$.\label{ex:RPTM}}
\end{figure}
\section{Conclusions}
\label{Concl}
\Abf{}~\cite{CLM16a} is a language for a predicate-based front-end of \FlyFast{,}
an on-the-fly mean-field model-checking tool. In this paper we presented a simplified version of
the translation proposed in~\cite{CLM16a} together with an approach for
model reduction that can be applied to the result of the translation.
The approach is
based on probabilistic bisimilarity for inhomogeneous DTMCs.
An example of application of the procedure has been shown. The implementation of
a compiler for \Abf{} mapping the language to \FlyFast{} is under development as an add-on
of \FlyFast{.}
We plan to apply the resulting extended tool to more as well as more complex models, in order to
get concrete insights on the practical applicability of the framework and on the actual limitations
imposed by the restrictions necessary for exploiting bisimilarity-based state-space reduction.
Investigating possible ways of relaxing some of such restrictions will also be an interesting
line of research.\\$ $\\
\emph{Acknowledgments}
Research partially funded by EU Project
n.~600708 \emph{A Quantitative Approach to Management and Design of
Collective and Adaptive Behaviours} (QUANTICOL).
\bibliographystyle{eptcs}
|
\section{Introduction}
Navigation is a commonly encountered task by autonomous agents that need to reach a destination or, more generally, to find a solution to a problem, where the solution is a sequence of instructions that transition a system from one state to another. This task is often performed when the agent does not have precise information about her current location. Examples of such agents are self-navi\-gating missiles, self-driving cars, and robotic vacuum cleaners.
Figure~\ref{intro-example figure} depicts an example $T_0$ of a transition system. This system consists of eight states $a,\dots,h$ represented by the vertices of the graph. The agent cannot distinguish state $a$ from state $b$ and state $c$ from state $d$, which is denoted by dashed lines connecting the indistinguishable states. The directed edges of the graph represent transitions that the system can make and the labels on these edges represent the instructions that the agent must give to do this. For example, if in state $a$ the agent executes instruction $0$, then the system transitions into state $g$, if, instead, the agent executes instruction $1$, the system transitions into state $e$. Although in this paper we consider non-deterministic transition systems where the execution of the same instruction can transition the system into one of the several states, for the sake of simplicity the transition system $T_0$ is deterministic.
\begin{figure}[ht]
\begin{center}
\vspace{0mm}
\scalebox{.7}{\includegraphics{Images/intro-example.pdf}}
\caption{Transition System $T_0$.}\label{intro-example figure}
\end{center}
\vspace{0mm}
\end{figure}
Note that, in system $T_0$ the agent can navigate from state $a$ to state $c$ by using instruction $1$ in state $a$ and the same instruction $1$ again in state $e$. However, a different sequence of instructions is required to reach state $c$ from state $b$. As the agent cannot distinguish state $a$ from state $b$, in state $a$ she does not know which instructions to use to accomplish her goal. Moreover, if the agent does reach state $c$, she cannot verify that the task is completed, because she cannot distinguish state $c$ from state $d$. For this reason, in this paper instead of navigation between states, we consider navigation between equivalence classes of states with respect to the indistinguishability relation. For example, the agent can navigate from class $[a]=\{a,b\}$ to class $[c]=\{c,d\}$ by using instruction $1$ in each state she passes.
\subsubsection{Perfect Recall}
In order to achieve a goal, the agent would need to follow a certain strategy that must be stored in her memory. We assume that the strategy is permanently stored (``hardwired") in the memory and cannot be changed during the navigation. For example, a robotic vacuum cleaner might be programmed to change direction when it encounters a wall, to make a circle when the dirt sensor is triggered, and to follow a straight path otherwise. A crucial question for us is if the vacuum cleaner can remember the walls and the dirty spots it has previously encountered. In other words, we distinguish an agent that can keep track of the classes of states she visited and the instructions she used from an agent who only knows her current state. We say that in the former case the agent has {\em perfect recall} and in the later she does not.
A strategy of an agent without perfect recall can only use information available to her about {\em the current state} to decide which instruction to use. In other words, a strategy of such an agent is a function that maps classes of indistinguishable {\em states} into instructions. A strategy of an agent with perfect recall can use information about {\em the history} of previous transitions to decide which instruction to use. In other words, a strategy of such an agent is a function that maps classes of indistinguishable {\em histories} into instructions. We call the former {\em memoryless} strategies and the later {\em recall} strategies.
In theory, a robotic vacuum cleaner without perfect recall is only equipped with read-only memory to store the strategy. A theoretical
robotic vacuum with perfect recall in addition to read-only memory that contains the strategy also has an unlimited read-write memory that contains logs of all previous transitions. In practice, the most popular brand of robotic vacuum cleaners, Roomba, is only using read-write memory to store information, such as a cleaning schedule, that is not used for navigation. This means that Roomba is using a memoryless strategy. The other popular robotic vacuum cleaner, Neato, is scanning the room before cleaning to use this information for navigation purposes. Although, of course, Neato only has a read-write memory of a limited size, the navigation strategy used by Neato is an example of a recall strategy.
\subsubsection{Examples}
Transition system $T_0$ has a recall strategy to navigate from class $[a]$ to class $[e]$. This might be surprising, because the same strategy must work to navigate from both state $a$ and state $b$ of class $[a]$ to the only state $e$ of class $[e]$. Yet, one would expect to use instruction $1$ in state $a$ and instruction $0$ in state $b$ to get to $e$. The recall strategy to navigate from class $[a]$ to class $[e]$ consists in first using instruction $0$ once no matter what the starting state is, and then using instruction $1$ until state $e$ is reached. When this strategy is used starting from a state in class $[a]$, the system first transitions into state $g$, then into state $a$, and then finally into state $e$. This is a recall strategy because it uses a different instruction in state $a$ depending if the system has already visited a state in class $[g]$ or not.
To show that there is no memoryless strategy to navigate from class $[a]$ to class $[e]$ note that any such strategy will have to use the same instruction $i_0$ at every visit to states $a$ and $b$. If $i_0=0$, then when the navigation starts in state $b$, the system stays ``locked" in set of states $\{a,g,b\}$ and never reaches state $e$. Similarly, if $i_0=1$, then when the navigation starts in state $b$, the system stays ``locked" in set of states $\{b,f,d,h\}$ and never reaches state $e$. Thus, there is no memoryless strategy to navigate from class $[a]$ to class $[e]$.
In some situation, even when there is a path between appropriate states, there might be no memoryless strategy and no recall strategy. For example, transition system $T_0$ has no recall strategy to navigate from class $[c]$ to class $[g]$. Indeed, if such a recall strategy exists, it would have to use the same instruction $i_0$ when the system {\em starts} in either state $c$ or state $d$. Suppose $i_0=0$. Thus, if the system {\em starts} in state $c$ it is ``locked" in state $h$. Assume now that $i_0=1$. Hence, if the system {\em starts} in state $d$ it is ``locked" in state $h$. Therefore, there is no recall strategy from class $[c]$ to class $[g]$.
\begin{table}[ht]
\centering
\begin{tabular}{ c | c c c c c c}
& $\{a,b\}$ & $\{c,d\}$ & $\{e\}$ & $\{f\}$ & $\{g\}$ & $\{h\}$\\ \hline
$\{a,b\}$ & m & m & r & r & m & r$^\dag$ \\
$\{c,d\}$ & - & m & - & - & - & r$^\dag$ \\
$\{e\}$ & m & m & m & r$^\dag$ & m & m \\
$\{f\}$ & m & m & r$^\dag$ & m & m & m \\
$\{g\}$ & m & m & m & m & m & m$^\dag$ \\
$\{h\}$ & - & - & - & - & - & m\\
\end{tabular}
\caption{Navigability between classes in system $T_0$.}\label{s r table}
\end{table}
Table~\ref{s r table} shows when memoryless strategies and recall strategies between classes of states of transition system $T_0$ exist. In this table, letter ``m" at the intersection of row $x$ and column $y$ denotes the existence of a memoryless strategy from class $x$ to class $y$. Letter ``r" denotes the existence of a recall strategy, but no memoryless strategy, and dash ``-" denotes that neither memoryless strategy nor recall strategy exist. Symbol $\dag$ marks the cases that we have found to be interesting to think about.
\subsubsection{Navigability between Sets}
As we have seen, there is a recall strategy, but no memoryless strategy, to navigate from class $[a]$ to class $[e]$. One can similarly show that there is no memoryless strategy to navigate from class $[a]$ to class $[f]$. However, if the goal is to navigate from class $[a]$ to either class $[e]$ or to class $[f]$, then there is a memoryless strategy to do this. Indeed, consider a memoryless strategy that uses instruction $1$ in every state. This strategy can transition the system from a state of class $[a]$ to a state of a class in set $\{[e],[f]\}$.
Thus, navigability between {\em sets of classes} can not be reduced to navigability between {\em classes}. For this reason, in this paper we study properties of navigability between sets of classes. If there is a strategy to navigate from a set of classes $A$ to a set of classes $B$, then we write $A\rhd B$. It will be clear from the context if we refer to the existence of a memoryless strategy or a recall strategy.
\subsubsection{Universal Properties of Navigability}
In the examples above we talked about properties of navigability for the transition system $T_0$. In the rest of this paper we study universal properties of navigability between sets of classes that are true in all transition systems. An example of such a property is {\em reflexivity}: $A\rhd B$, where $A\subseteq B$. This property is true for both memoryless and recall strategies because absolutely {\em any} strategy can be used to navigate from a subset to the whole set. In fact, in this case the goal is achieved before the navigation even starts.
Another example of a property of navigation which is universally true for both memoryless and recall strategies is {\em augmentation}: $A\rhd B \to (A\cup C) \rhd (B\cup C)$. It says that if there is a strategy to navigate from set $A$ to set $B$, then there is a strategy to navigate from set $A\cup C$ to set $B\cup C$.
An example of a property which is universally true for recall strategies, but is not universally true for memoryless strategies is {\em transitivity}: $A\rhd B\to(B\rhd C\to A\rhd C)$. It states that if there is a strategy to navigate from set $A$ to set $B$ and a strategy to navigate from set $B$ to set $C$, then there is a strategy to navigate from set $A$ to set $C$. To see that this property is not universally true for memoryless strategies, note that, in transition system $T_0$, memoryless strategy that always uses instruction $0$ can be used to navigate from set $\{[a]\}$ to set $\{[g]\}$ and memoryless strategy that always uses instruction $1$ can be used to navigate from set $\{[g]\}$ to set $\{[e]\}$. At the same time, as we have shown earlier, there is no memoryless strategy to navigate from set $\{[a]\}$ to set $\{[e]\}$.
In this paper we show that reflexivity, augmentation, and transitivity principles form a sound and complete logical system that describes all universal properties of navigability by recall strategies. These are the three principles known in database theory as Armstrong's axioms~\cite[p.~81]{guw09}, where they give a sound and complete axiomatization of functional dependency~\cite{a74}. We also give a sound and complete axiomatization of universal properties of navigability by memoryless strategies. It consists of the reflexivity and augmentation principles mentioned above as well as the {\em monotonicity} principle $(A\cup C)\rhd B\to A\rhd B$. The latter principle is true for the recall strategies as well, but it is provable from Armstrong's axioms.
\subsubsection{Literature Review}
Most of the existing literature on logical systems for reasoning about strategies is focused on modal logics. Logics of coalition power were developed by \cite{p01illc,p02}, who also proved the completeness of the basic logic of coalition power. Pauly's approach has been widely studied in literature~\cite{g01tark,vw05ai,b07ijcai,sgvw06aamas,abvs10jal,avw09ai,b14sr}. Alternative logical system were proposed by \cite{mn12tocl}, \cite{w15lori,w17synthese}, and \cite{lw17icla}.
\cite{ahk02} introduced Alternating-Time Temporal Logic (ATL) that combines temporal and coalition modalities. \cite{vw03sl} proposed to combine ATL with epistemic modality to form Alternating-Time Temporal Epistemic Logic. A completeness theorem for a logical system that combines coalition power and epistemic modalities was proven by \cite{aa12aamas}.
The notion of a strategy that we consider in this paper is much more restrictive than the notion of strategy in the works mentioned above. Namely, we assume that the strategy must be based only on the information available to the agent. This is captured in our setting by requiring the strategy to be the same in all indistinguishable states or histories. This restriction on strategies has been studied before under different names. \cite{ja07jancl} talk about ``knowledge to identify and execute a strategy", \cite{jv04fm} discuss ``difference between an agent knowing that he has a suitable strategy and knowing the strategy itself". \cite{v01ber} calls such strategies ``uniform". \cite{nt17aamas} use the term ``executable strategy". \cite{nt17tark} proposed a complete trimodal logical system describing an interplay between distributed knowledge, uniform strategic power modality, and standard strategic power modality for achieving a goal by a coalition in one step. \cite{fhlw17ijcai} developed a complete logical system in a single-agent setting for uniform strategies to {achieve} a goal in multiple steps. \cite{nt17aamas} developed a similar system for maintaining a goal in multi-agent setting. Our contribution is different from all of the above papers by being the first to propose complete logical systems for recall strategies and memoryless strategies.
\subsubsection{Paper Outline}
In the next section we define transition systems and the syntax of our logical systems. This section applies equally to recall and memoryless strategies. The rest of the paper is split into two independent sections. The first of them proves the soundness and the completeness of Armstrong's axioms for navigability under recall strategies and the second gives an axiomatization for memoryless strategies.
\section{Syntax and Semantics}
In this section we formally define the language of our logical system, the notion of a transition system, and the related terminology. In the introduction, relation $\rhd$ was viewed as a relation between equivalence classes of a given transition system. Thus, our language depends on these classes and changes from transition system to transition system. In order to have a single language for all transition systems we introduce a fixed finite set of ``views" $V$, whose elements act as names of the equivalence classes in any given transition system.
\begin{definition}\label{Phi}
$\Phi$ is the minimal set of formulae such that
\begin{enumerate}
\item $A\rhd B\in\Phi$ for all nonempty\footnote{If one allows sets $A$ and $B$ to be empty, most of the proofs in this paper will remain unchanged, but both logical systems will need an additional axiom $\neg(A\rhd \varnothing)$ for each nonempty set $A$.} sets $A,B\subseteq V$,
\item $\neg\phi,\phi\to\psi\in \Phi$ for all formulae $\phi,\psi\in \Phi$.
\end{enumerate}
\end{definition}
Each transition system specifies a mapping $*$ of views into equivalence classes of states. Transitions between states under an instruction $i$ are captured by a transition function $\Delta_i$.
\begin{definition}\label{transition system}
$(S,\sim,*,I,\{\Delta_i\}_{i\in I})$ is a transition system, if
\begin{enumerate}
\item $S$ is a set of states,
\item $\sim$ is an equivalence (indistinguishability) relation on $S$,
\item $*$ is a function from $V$ to $S/\!\sim$,
\item $I$ is an arbitrary nonempty set of ``instructions",
\item $\Delta_i$ maps set $S$ into nonempty subsets of $S$ for each $i\in I$.
\end{enumerate}
\end{definition}
We write $a^*$ instead of $*(a)$, where $a\in V$. An example of a transition system is system $T_0$ depicted in Figure~\ref{intro-example figure}.
\begin{definition}\label{history}
A finite sequence $w_0,i_1,w_1,\dots,i_n,w_n$, where $n\ge 0$,
is called a history if
\begin{enumerate}
\item $w_k\in S$ for each $k$ such that $0\le k\le n$,
\item $i_k\in I$, for each $k$ such that $1\le k\le n$,
\item $w_k\in \Delta_{i_k}(w_{k-1})$,
for each $k$ such that $1\le k\le n$.
\end{enumerate}
\end{definition}
For example, sequence $g,1,a,1,e,1,c,0,h$ is a history for system $T_0$.
The set of all histories is denoted by $H$.
\begin{definition}
History $h=w_0,i_1,w_1,\dots,i_n,w_n$ is indistinguishable from history $h'=w'_0,i_1,w'_1,\dots,i_n,w'_n$ if $w_k \sim w'_k$ for each $k$ such that $0\le k\le n$.
\end{definition}
For example, histories $a,0,g$ and $b,0,g$ are indistinguishable in transition system $T_0$. Indistinguishability of histories of $h$ and $h'$ is denoted by $h\approx h'$. The equivalence class of history $h$ with respect to this equivalence relation is denoted by $\ldbrack h \rdbrack$. Equivalence class of a state $w$ with respect to equivalence relation $\sim$ is denoted by $[w]$.
\begin{lemma}\label{histories approx lemma}
If $w_0,i_1,w_1,\dots,w_n\approx w'_0,i_1,w'_1,\dots,w'_n$, then $[w_k]=[w'_k]$ for each $k\le n$. \hfill $\boxtimes\hspace{1mm}$}% \linebreak
\end{lemma}
\begin{definition}\label{strategy}
A memoryless strategy is a function from set $S/\!\sim$ to set $I$. A recall strategy maps set $H/\!\approx$ to set $I$.
\end{definition}
We write $s[w]$ and $s\ldbrack h\rdbrack$ instead of $s([w])$ and $s(\ldbrack h\rdbrack)$.
\begin{definition}\label{path}
An infinite sequence $w_0,i_1,w_1,i_2,w_2\dots$ is called
a path under a memoryless strategy $s$ if for each $k\ge 1$
\begin{enumerate}
\item $w_0,i_1,w_1,i_2,w_2\dots,w_{k-1}\in H$,
\item $i_{k}=s[w_{k-1}]$.
\end{enumerate}
\end{definition}
\begin{lemma}\label{path exists}
For any history $w_0,i_1,w_1,\dots, i_n,w_n$ and any memoryless strategy $s$, if $i_{k}=s[w_{k-1}]$ for each $k$ such that $1\le k\le n$, then there are states $w_{n+1},w_{n+2},\dots$ and instructions $i_{n+1},i_{n+2},\dots$ such that sequence $w_0,i_1,w_1,\dots, i_n,w_n,i_{n+1},w_{n+1},i_{n+2},w_{n+2},\dots$ is a path under strategy $s$.
\end{lemma}
\begin{proof} Elements $i_{n+1},w_{n+1},i_{n+2},w_{n+2},\dots$ can be constructed recursively because (a) there is a state $w_{k+1}\in\Delta_{i_{k+1}}(w_k)$ for any state $w_{k}$ and any $i_{k+1}\in I$ by item 5 of Definition~\ref{transition system}; (b) $I\neq\varnothing$ by item 4 of Definition~\ref{transition system}.
\end{proof}
\begin{definition}\label{perfect path}
An infinite sequence $w_0,i_1,w_1,i_2,w_2\dots$ is called
a path under a recall strategy $s$ if for each $k\ge 1$
\begin{enumerate}
\item $w_0,i_1,w_1,i_2,w_2\dots,w_{k-1}\in H$,
\item $i_{k}=s\ldbrack w_0,i_1,w_1,i_2,w_2\dots,w_{k-1}\rdbrack$.
\end{enumerate}
\end{definition}
\begin{definition}\label{*set}
$A^*=\{a^*\;|\;a\in A\}$, for all sets $A\subseteq V$.
\end{definition}
\begin{definition}\label{path set}
For a given memoryless strategy or recall strategy $s$, let $Path_s(A)$ be the set of all paths $w_0,i_1,w_1,i_2,w_2\dots$ under $s$ such that $[w_0]\in A^*$.
\end{definition}
\begin{definition}\label{visit set}
Let set $Visit_s(B)$ be the set of all paths $w_0,i_1,w_1,i_2,w_2\dots$ under $s$ such that $[w_k]\in B^*$ for some $k\ge 0$.
\end{definition}
We write $Visit(B)$ instead of $Visit_s(B)$ when value of $s$ is clear from the context.
\section{Navigation with Recall Strategies}
In this section we show that Armstrong's axioms give a complete axiomatization of navigability between sets of classes with recall strategies. We start with a formal semantics of navigability relation $\rhd$ under recall strategies.
\begin{definition}\label{perfect sat}
$T\vDash A\rhd B$ if $Path_s(A)\subseteq Visit(B)$ for some recall strategy $s$ of transition system $T$.
\end{definition}
\subsubsection{Axioms}
The axioms of the logical system that we consider in this section are the tautologies in language $\Phi$ and the following additional principles known as Armstrong's axioms \cite[p.~81]{guw09}:
\begin{enumerate}
\item Reflexivity: $A\rhd B$, where $A\subseteq B$,
\item Augmentation: $A\rhd B \to (A\cup C)\rhd (B\cup C)$,
\item Transitivity: $A\rhd B\to (B\rhd C \to A\rhd C)$.
\end{enumerate}
We write $\vdash\phi$ if formula $\phi$ is provable from these axioms using the Modus Ponens inference rule. We write $X\vdash\phi$ if $\phi$ is provable using a set of additional axioms $X$.
\subsubsection{Soundness} In this section we prove soundness of the above axioms with respect to perfect recall semantics.
\begin{theorem}\label{perfect soundness}
If $\vdash\phi$, then $T\vDash\phi$ for every system $T$.
\end{theorem}
The soundness of Armstrong's axioms with respect to perfect recall semantics is established in Lemma~\ref{perfect reflexivity}, Lemma~\ref{perfect augmentation}, and Lemma~\ref{perfect transitivity} below. We start with technical notions that we need to prove Lemma~\ref{perfect transitivity}.
\begin{definition}\label{truncation}
For any history $w_0,i_1,\dots,i_n,w_n$ and any set $B\subseteq V$, let truncation $(w_0,i_1,\dots,i_n,w_n)|_B$ be
\begin{enumerate}
\item $w_0,i_1,\dots,i_n,w_n$, if $n=0$ or $[w_0]\in B^*$,
\item $(w_1,i_2,\dots,i_n,w_n)|_B$, otherwise.
\end{enumerate}
\end{definition}
For transition system $T_0$, if $B=\{b\}$ and $b^*=[e]$, then truncation $(g,1,a,1,e,1,c,0,h)|_B$ is history $(e,1,c,0,h)$.
\begin{lemma}\label{histories truncate lemma}
If $h_1\approx h_2$, then $h_1|_B\approx h_2|_B$.\hfill $\boxtimes\hspace{1mm}$}% \linebreak
\end{lemma}
\begin{definition}\label{circ}
For any recall strategies $s_1$ and $s_2$, any set $B\subseteq V$, and any history $h=w_0,i_1,\dots,w_n$, let
$$
(s_1\circ_B s_2)\ldbrack h\rdbrack=
\begin{cases}
s_1\ldbrack h\rdbrack, & \mbox{ if } \forall k\le n\,([w_k]\notin B^*) ,\\
s_2\ldbrack h|_B\rdbrack, & \mbox{ otherwise}.
\end{cases}
$$
\end{definition}
In other words, ``composition" strategy $s_1\circ_B s_2$ follows strategy $s_1$ until a state $w$ is reached such that $[w]\in B^*$. Once this happens, the memory of all prior states and instructions is erased and strategy $s_2$ is executed as if navigation started from state $w$.
\begin{lemma}
Operation $s_1\circ_B s_2$ is well-defined.
\end{lemma}
\begin{proof}
See Definition~\ref{truncation}, Lemma~\ref{histories approx lemma}, and Lemma~\ref{histories truncate lemma}.
\end{proof}
\begin{definition}\label{truncation of paths}
For any path $w_0,i_1,w_1,i_2,\dots$ and any set $B\subseteq V$, let truncation $(w_0,i_1,w_1,i_2,\dots)|_B$ be
\begin{enumerate}
\item path $w_0,i_1,w_1,i_2,\dots$, if $[w_0]\in B^*$,
\item path $(w_1,i_2,w_2,i_3\dots)|_B$, otherwise.
\end{enumerate}
\end{definition}
\begin{lemma}\label{truncate on B lemma}
For any recall strategies $s_1,s_2$ and any sets $A,B\subseteq V$, if path $\pi\in Path_{s_1\circ_B s_2}(A)\cap Visit(B)$, then $\pi|_B\in Path_{s_2}(B)$. \hfill $\boxtimes\hspace{1mm}$}% \linebreak
\end{lemma}
\begin{lemma}\label{strange lemma}
$Path_{s_1\circ_B s_2}(A)\subseteq Visit(B)$
if $Path_{s_1}(A)\subseteq Visit(B)$, for any recall strategies $s_1,s_2$ and $A,B\subseteq V$.
\end{lemma}
\begin{proof}
Assume that there is a path $w_0,i_1,w_1,\dots$ in $Path_{s_1\circ_B s_2}(A)$ such that $w_0,i_1,w_1,\dots\notin Visit(B)$. Thus, $i_k = {(s_1\circ_B s_2)}\ldbrack w_0,i_1,w_1,\dots, w_{k-1}\rdbrack$ for each $k\ge 1$ and $[w_0]\in A^*$ by Definition~\ref{path set} and Definition~\ref{perfect path}. Also, $[w_k]\notin B^*$ for each $k\ge 0$ by Definition~\ref{visit set}. Hence, by Definition~\ref{circ}, we have $i_k = {s_1}\ldbrack w_0,i_1,w_1,\dots, w_{k-1}\rdbrack$ for each $k\ge 1$. Thus, $w_0,i_1,w_1,\dots\in Path_{s_1}(A)$ by Definition~\ref{perfect path} and Definition~\ref{perfect path}. Therefore, $w_0,i_1,w_1,\dots\in Visit(B)$ by the assumption $Path_{s_1}(A)\subseteq Visit(B)$ of the lemma.
\end{proof}
\begin{lemma}\label{apply s2}
For any recall strategies $s_1,s_2$ and $A,B\subseteq V$, if $Path_{s_1\circ_B s_2}(A)\subseteq Visit(B)$, then $(Path_{s_1\circ_B s_2}(A))|_B\subseteq Path_{s_2}(B)$.
\end{lemma}
\begin{proof}
Consider any path $\pi\in Path_{s_1\circ_B s_2}(A)$. Then, $\pi\in Visit(B)$ because $Path_{s_1\circ_B s_2}(A)\subseteq Visit(B)$. Hence, $\pi|_B\in Path_{s_2}(B)$ by Lemma~\ref{truncate on B lemma}.
\end{proof}
\begin{lemma}\label{less is more}
$Path_s(A)\subseteq Visit(C)$, if $(Path_s(A))|_B\subseteq Visit(C)$,
for any $s$ and any $A,B,C\subseteq V$.\hfill $\boxtimes\hspace{1mm}$}% \linebreak
\end{lemma}
\begin{lemma}\label{path properties}
For any recall strategy $s$ and any $A,B,C\subseteq V$,
\begin{enumerate}
\item\label{line 1} $Path_s(A)\subseteq Path_s(B)$ if $A\subseteq B$,
\item\label{line 2} $Path_s(A)\subseteq Visit(A)$,
\item\label{line 3} $Path_s(A\cup C)=Path_s(A) \cup Path_s(C)$,
\item\label{line 4} $Visit(B) \cup Visit(C) = Visit(B\cup C)$.
\end{enumerate}
\end{lemma}
\begin{proof}
See Definition~\ref{perfect path} and Definition~\ref{visit set}.
\end{proof}
The next three lemmas prove soundness of Armstrong's axioms with respect to the perfect recall semantics.
\begin{lemma}\label{perfect reflexivity}
$T\vDash A\rhd B$, where $A\subseteq B$.
\end{lemma}
\begin{proof}
By Definition~\ref{transition system}, set $I$ is nonempty, and thus it contains at least one instruction $i_0$. Let $s$ be a recall strategy such that $s\ldbrack h\rdbrack=i_0$ for each history $h$. By item~\ref{line 1} and item~\ref{line 2} of Lemma~\ref{path properties},
$
Path_s(A)\subseteq Path_s(B)\subseteq Visit(B).
$
Therefore, $T\vDash A\rhd B$ by Definition~\ref{perfect sat}.
\end{proof}
\begin{lemma}\label{perfect augmentation}
If $T\vDash A\rhd B$, then $T\vDash(A\cup C)\rhd(B\cup C)$.
\end{lemma}
\begin{proof}
By Definition~\ref{perfect sat}, the assumption $T\vDash A\rhd B$ implies that there exists a recall strategy $s$ such that $Path_s(A)\subseteq Visit(B)$. Thus, by item~\ref{line 3}, item~\ref{line 2}, and item~\ref{line 4} of Lemma~\ref{path properties},
\begin{eqnarray*}
Path_s(A\cup C)& = & Path_s(A) \cup Path_s(C) \\
&\subseteq & Visit(B) \cup Path_s(C) \\
&\subseteq & Visit(B) \cup Visit(C)\\
&\subseteq & Visit(B\cup C).
\end{eqnarray*}
Therefore, $T\vDash(A\cup C)\rhd(B\cup C)$ by Definition~\ref{perfect sat}.
\end{proof}
\begin{lemma}\label{perfect transitivity}
If $T\vDash A\rhd B$ and $T\vDash B\rhd C$, then $T\vDash A\rhd C$.
\end{lemma}
\begin{proof}
By the assumption $T\vDash A\rhd B$ and Definition~\ref{perfect sat}, we have $Path_{s_1}(A)\subseteq Visit(B)$ for some recall strategy $s_1$. Similarly, the assumption $T\vDash B\rhd C$ implies that there exists a recall strategy $s_2$ such that $Path_{s_2}(B)\subseteq Visit(C)$.
By Lemma~\ref{strange lemma}, statement $Path_{s_1}(A)\subseteq Visit(B)$ implies $Path_{s_1\circ_B s_2}(A)\subseteq Visit(B)$.
Thus, by Lemma~\ref{apply s2},
$$(Path_{s_1\circ_B s_2}(A))|_B\subseteq Path_{s_2}(B).$$
Hence, we have $(Path_{s_1\circ_B s_2}(A))|_B\subseteq Visit(C)$ because of $Path_{s_2}(B)\subseteq Visit(C)$.
Thus, by Lemma~\ref{less is more}, $Path_{s_1\circ_B s_2}(A)\subseteq Visit(C)$. Therefore, $T\vDash A\rhd C$ by Definition~\ref{perfect sat}.
\end{proof}
\subsubsection{Completeness}
In the rest of this section we prove the completeness of Armstrong's axioms with respect to the perfect recall semantics. We start by defining a canonical transition system $T(X)=(V\cup \{\circlearrowleft\},=,*,I,\{\Delta_i\}_{i\in I})$ for an arbitrary maximal consistent set of formulae $X\subseteq \Phi$. The set of states of this transition system consists of a single state for each view, plus one additional state that we denote by symbol $\circlearrowleft$. Informally, the additional state is a sink or a ``black hole" state from which there is no way out. State $h$ in transition system $T_0$ depicted in Figure~\ref{intro-example figure} is an example of a black hole state. Note that the indistinguishability relation on the states of the canonical transition system is equality relation $=$. That is, the agent has an ability to distinguish any two different states in the system. The fact that equality is suitable as an indistinguishability relation for the canonical transition system with perfect recall is surprising. The indistinguishability relation for the canonical transition system for memoryless strategies, discussed in the next section, is different from equality.
Each view $v\in V$ is also a state in the canonical transition system. The equivalence class of state $v$ consists of the state itself: $[v]=\{v\}$. We define $v^*$ to be class $[v]$.
\begin{lemma}\label{* iff}
$u\in A$ iff $[u]\in A^*$ for each view $u\in V$ and each set $A\subseteq V$. \hfill $\boxtimes\hspace{1mm}$}% \linebreak
\end{lemma}
Informally, if set $X$ contains formula $A\rhd B$, then we want the canonical transition system $T(X)$ to have a recall strategy to navigate from set $A^*=\{[a]\;|\;a\in A\}=\{\{a\}\;|\;a\in A\}$ to set $B^*=\{[b]\;|\;b\in B\}=\{\{b\}\;|\;b\in B\}$. It turns out that it is sufficient to have just a single instruction that transitions the system from any state in set $A$ to a state in set $B$. We denote this instruction by pair $(A,B)$.
\begin{definition}\label{perfect canonical I}
$I=\{(A,B)\;|\; A\rhd B \in X\}$.
\end{definition}
Recall that assumption $A\rhd B\in X$ requires sets $A$ and $B$ to be nonempty due to Definition~\ref{Phi}.
As discussed above, for any instruction $(A,B)$ we define the nondeterministic transition function $\Delta_{(A,B)}$ to transition the system from a state in $A$ to a state in $B$. If used outside of set of states $A$, instruction $(A,B)$ transitions the system into black hole state $\circlearrowleft$:
\begin{definition}\label{perfect canonical delta}
$$\Delta_{(A,B)}(w)=
\begin{cases}
B, & \mbox{ if $w\in A$},\\
\{\circlearrowleft\}, & \mbox{ otherwise}.
\end{cases}
$$
\end{definition}
This concludes the definition of the transition system $T(X)$.
Next, for any recall strategy $s$ and any set of states $G\subseteq V$, we define a family of sets of states $\{G^s_n\}_{n\ge 0}$. Informally, set $G^s_n$ is the set of all states from which strategy $s$ ``draws" the system into set $G$ after at most $n$ transitions. For any history $h=w_0,i_1,w_1,\dots,i_n,w_n$, by $hd(h)$ we mean the state $w_n$.
\begin{definition}\label{perfect Gn}
For any recall strategy $s$ and any nonempty set $G\subseteq V$, let chain of $G^s_0\subseteq G^s_1\subseteq\dots\subseteq V\cup\{\circlearrowleft\}$ be defined as
\begin{enumerate}
\item $G^s_0=G$,
\item $G^s_{n+1}=G^s_{n}\cup\left\{hd(h)\;\middle|\; h\in H \mbox{ and }\Delta_{s\ldbrack h\rdbrack}(hd(h))\subseteq G^s_{n}\right\}$ for all $n\ge 0$.
\end{enumerate}
\end{definition}
Note that this definition, in essence, has an existential quantifier over history $h$. Thus, informally, strategy $s$ is allowed to ``manipulate" the history in order to ``draw" the system into set $G$.
\begin{lemma}\label{no holes in G}
$\circlearrowleft\;\notin G^s_n$ for each $n\ge 0$.
\end{lemma}
\begin{proof}
We prove this statement by contradiction. Let $n$ be the smallest non-negative integer number such that $\circlearrowleft\;\in G^s_n$. Note that $n\neq 0$ because $G^s_0=G\subseteq V$ by Definition~\ref{perfect Gn}. Thus, $n>0$. Hence, by Definition~\ref{perfect Gn}, there exists a history $h$ such that $hd(h)=\;\circlearrowleft$ and
$
\Delta_{s\ldbrack h\rdbrack}(\circlearrowleft)\subseteq G^s_{n-1}.
$
Note that $\Delta_{s\ldbrack h\rdbrack}(\circlearrowleft)=\{\circlearrowleft\}$ by Definition~\ref{perfect canonical delta}. Therefore, $\circlearrowleft\;\in G^s_{n-1}$, which contradicts the choice of integer $n$.
\end{proof}
\begin{definition}\label{perfect Ginfty}
$G^s_\infty=\bigcup_{k\ge 0} G^s_k.$
\end{definition}
We now prove properties of the family of sets $\{G^s_n\}_{n\ge 0}$ that are needed to finish the proof of the completeness.
\begin{lemma}\label{g to Gn-1}
$X\vdash \{g\}\rhd G^s_{n-1}$ for each integer $n\ge 1$ and each state $g\in G^s_n$.
\end{lemma}
\begin{proof}
By Definition~\ref{perfect Gn}, assumptions $n\ge 1$ and $g\in G^s_n$ imply that there is a history $h$ such that $hd(h)=g$ and
\begin{equation}\label{delta g}
\Delta_{s\ldbrack h\rdbrack}(g)\subseteq G^s_{n-1}.
\end{equation}
Furthermore, by Definition~\ref{perfect canonical I}, there are nonempty sets $A,B\subseteq V$ such that $s\ldbrack h \rdbrack=(A,B)$.
\noindent{\em Case I}: $g\notin A$. Then, $\Delta_{s\ldbrack h\rdbrack}(g)=\{\circlearrowleft\}$ by Definition~\ref{perfect canonical delta}. Hence, $\{\circlearrowleft\}\subseteq G^s_{n-1}$ by equation~(\ref{delta g}), which contradicts Lemma~\ref{no holes in G}.
\noindent{\em Case II}: $g\in A$. Then, $\vdash \{g\}\rhd A$ by the Reflexivity axiom. At the same time, $X\vdash A\rhd B$ by Definition~\ref{perfect canonical I}. Thus, by the Transitivity axiom, $X\vdash \{g\}\rhd B$.
Assumption $g\in A$ implies $\Delta_{s\ldbrack h\rdbrack}(g)=B$, by Definition~\ref{perfect canonical delta}. Thus, $B\subseteq G^s_{n-1}$ due to equation~(\ref{delta g}). Hence, $\vdash B\rhd G^s_{n-1}$ by the Reflexivity axiom.
Finally, statements $X\vdash \{g\}\rhd B$ and $\vdash B\rhd G^s_{n-1}$ by the Transitivity axiom imply, using Modus Ponens inference rule twice, that $X \vdash \{g\}\rhd G^s_{n-1}$.
\end{proof}
\begin{lemma}\label{all the a}
For any $n\ge 1$ and any views $a_1,\dots,a_n\in V$, if $X\vdash \{a_k\}\rhd B$ for each $k\le n$, then $X\vdash \{a_1,\dots,a_n\}\rhd B$.
\end{lemma}
\begin{proof}
We prove this statement by induction on $n$. In the base case, $X\vdash \{a_1\}\rhd B$ due to the assumption of the lemma.
By the induction hypothesis, $X\vdash \{a_1,\dots,a_{n-1}\}\rhd B$. Thus, by the Augmentation axiom,
\begin{equation}\label{friday}
X\vdash \{a_1,\dots,a_{n-1},a_n\}\rhd B\cup \{a_n\}.
\end{equation}
At the same time, $X\vdash \{a_n\}\rhd B$ by the assumption of the lemma. Hence, by the Augmentation axiom $X\vdash B\cup\{a_n\}\rhd B$. Thus, $X\vdash \{a_1,\dots,a_{n-1},a_n\}\rhd B$ by statement~(\ref{friday}) and the Transitivity axiom.
\end{proof}
\begin{lemma}\label{Gn to Gn-1}
$X\vdash G^s_n\rhd G^s_{n-1}$ for each $n\ge 1$.
\end{lemma}
\begin{proof}
The statement of the lemma follows from Lemma~\ref{g to Gn-1} and Lemma~\ref{all the a}.
\end{proof}
\begin{lemma}\label{Gn G}
$X\vdash G^s_n\rhd G$ for each $n\ge 0$.
\end{lemma}
\begin{proof}
We prove this statement by induction on integer $n$. In the base case, due to Definition~\ref{perfect Gn}, it suffices to show that $\vdash G\rhd G$, which is an instance of the Reflexivity axiom.
For the induction step, note that $X\vdash G^s_n\rhd G^s_{n-1}$ by Lemma~\ref{Gn to Gn-1}. At the same time, $X\vdash G^s_{n-1}\rhd G$ by the induction hypothesis. Hence, $X\vdash G^s_n\rhd G$ by the Transitivity axiom.
\end{proof}
\begin{lemma}\label{infty = n}
There is $n\ge 0$, such that $G^s_\infty=G^s_n$.
\end{lemma}
\begin{proof}
Since $G^s_0\subseteq G^s_1\subseteq G^s_2\subseteq \dots\subseteq V\cup\{\circlearrowleft\}$ and set $V$ is finite, there must exist an integer $n$ such that $G^s_n=\bigcup_{k\ge 0} G^s_k$. Therefore, $G^s_n=G^s_\infty$ by Definition~\ref{perfect Ginfty}.
\end{proof}
\begin{lemma}\label{Delta Ginfity}
Set $(\Delta_{s\ldbrack h\rdbrack}(hd(h)))\setminus G^s_\infty$ is non-empty for each history $h$ such that $hd(h)\notin G^s_\infty$.
\end{lemma}
\begin{proof}
Suppose $\Delta_{s\ldbrack h\rdbrack}(hd(h))\subseteq G^s_\infty$ for some history $h$. It suffices to show that $hd(h)\in G^s_\infty$. Indeed, by Lemma~\ref{infty = n} there is $n$ such that $G^s_\infty=G^s_n$. Thus, $\Delta_{s\ldbrack h\rdbrack}(hd(h))\subseteq G^s_n$. Hence, $hd(h)\in G^s_{n+1}$ by Definition~\ref{perfect Gn}. Then, $hd(h)\in G^s_\infty$ by Definition~\ref{perfect Ginfty}.
\end{proof}
\begin{lemma}\label{history extend}
For any positive integer $k$ and any history $w_0,i_1,w_1,\dots,w_{k-1}$, if $w_{k-1}\notin G^s_\infty$, then there is a state $w_{k}\notin G^s_\infty$ such that $w_0,i_1,w_1,\dots,w_{k-1},i_k,w_k$ is a history, where $i_k=s\ldbrack w_0,i_1,w_1,\dots,w_{k-1}\rdbrack$.
\end{lemma}
\begin{proof}
By Lemma~\ref{Delta Ginfity}, set $(\Delta_{i_k}(w_{k-1}))\setminus G^s_\infty$ is not empty. Let $w_k$ be any state such that
$w_k\in \Delta_{i_k}(w_{k-1})$ and $w_k\notin G^s_\infty$. Then, $w_0,i_1,w_1,\dots,w_{k-1},i_k,w_k$ is a history by Definition~\ref{history}.
\end{proof}
\begin{lemma}\label{path exists not in G infty}
For each state $w_0\notin G^s_\infty$ there exists a path $w_0,i_1,w_1,\dots$ under recall strategy $s$ such that $w_k\notin G^s_\infty$ for each $k\ge 0$.
\end{lemma}
\begin{proof}
Note that single-element sequence $w_0$ is a history by Definition~\ref{history}. Due to Lemma~\ref{history extend}, there is an infinite sequence $w_0,i_1,w_1,\dots$ such that for each integer $k\ge 1$,
\begin{enumerate}
\item $w_0,i_1,w_1,\dots,w_{k-1}$ is a history,
\item $i_k=s\ldbrack w_0,i_1,w_1,\dots,w_{k-1}\rdbrack$,
\item $w_k\notin G^s_\infty$.
\end{enumerate}
By Definition~\ref{perfect path}, sequence $w_0,i_1,w_1,\dots$ is a path under recall strategy $s$.
\end{proof}
\begin{lemma}\label{perfect vdash to vDash}
If $X\vdash A\rhd B$, then $T(X)\vDash A\rhd B$.
\end{lemma}
\begin{proof}
Assumption $X\vdash A\rhd B$ implies $(A,B)\in I$ by Definition~\ref{perfect canonical I}. Consider recall strategy $s$ such that $s\ldbrack h\rdbrack=(A,B)$ for each class of histories $\ldbrack h \rdbrack$. Consider any path $w_0,i_1,w_1,\dots$ under recall strategy $s$ where $[w_0]\in A^*$. Then, $w_0\in A$ by Lemma~\ref{* iff}.
By Definition~\ref{perfect sat} and Definition~\ref{visit set} it suffices to show that $[w_1]\in B^*$. Indeed, $i_1=(A,B)$ by choice of recall strategy $s$. Thus, $\Delta_{i_1}(w_0)=\Delta_{(A,B)}(w_0)=B$ by Definition~\ref{perfect canonical delta} and due to the assumption $w_0\in A$.
By Definition~\ref{perfect path}, sequence $w_0,i_1,w_1$ is a history. Hence, $w_1\in \Delta_{i_1}(w_0)$ by Definition~\ref{history}.
Thus, $w_1\in \Delta_{i_1}(w_0)=B$. Then, $[w_1]\in B^*$ by Lemma~\ref{* iff}.
\end{proof}
\begin{lemma}\label{perfect vDash to vdash}
If $T(X)\vDash E\rhd G$, then $X\vdash E\rhd G$.
\end{lemma}
\begin{proof}
By Definition~\ref{perfect sat}, assumption $T(X)\vDash E\rhd G$ implies $Path_s(E)\subseteq Visit(G)$ for some recall strategy $s$.
\noindent{\em Case I:} $E\subseteq G^s_\infty$. By Lemma~\ref{infty = n} there exists an integer $n\ge 0$, such that $E\subseteq G^s_n$. Thus, $\vdash E\rhd G^s_n$ by the Reflexivity axiom. At the same time, $X\vdash G^s_n\rhd G$ by Lemma~\ref{Gn G}. Therefore, $X\vdash E\rhd G$ by the Transitivity axiom.
\noindent{\em Case II:} $E\nsubseteq G^s_\infty$. Then, there is an element $w_0\in E$ such that $w_0\notin G^s_\infty$. Thus, by Lemma~\ref{path exists not in G infty} there is a path $\pi=w_0,i_1,w_1,\dots$ under recall strategy $s$ such that $w_k\notin G^s_\infty$ for all $k\ge 0$. Hence, $w_k\notin G$ for all $k\ge 0$ because $G=G^s_0\subseteq G^s_\infty$ by Definition~\ref{perfect Gn} and Definition~\ref{perfect Ginfty}.
Thus, $[w_0]\in E^*$ and $[w_k]\notin G^*$ for all $k\ge 0$ by Lemma~\ref{* iff}. Then, states in path $\pi\in Path_s(E)$ by Definition~\ref{path set} and $\pi\notin Visit(G)$ by Definition~\ref{visit set}. Therefore, $Path_s(E)\nsubseteq Visit(G)$, which contradicts the choice of strategy $s$.
\end{proof}
\begin{lemma}\label{perfect induction}
$X\vdash\phi$ iff $T(X)\vDash\phi$ for each $\phi\in \Phi$.
\end{lemma}
\begin{proof}
We prove this lemma by induction on the structural complexity of $\phi$. The base case follows from Lemma~\ref{perfect vdash to vDash} and Lemma~\ref{perfect vDash to vdash}. The induction case follows from the maximality and the consistency of set $X$ in the standard way.
\end{proof}
We are now ready to state and prove the completeness theorem for the recall strategies.
\begin{theorem}\label{perfect completeness}
If $T\vDash\phi$ for every system $T$, then $\vdash\phi$.
\end{theorem}
\begin{proof}
Suppose $\nvdash\phi$. Let $X$ be a maximal consistent set containing formula $\neg\phi$. Thus, $T(X)\vDash\neg\phi$ by Lemma~\ref{perfect induction}. Therefore, $T(X)\nvDash\phi$.
\end{proof}
\section{Navigation with Memoryless Strategies}
In this section we give a sound and complete axiomatization of navigability under memoryless strategies. We start by modifying Definition~\ref{perfect sat} to refer to memoryless strategies instead of recall strategies:
\begin{definition}\label{imperfect sat}
$T\vDash A\rhd B$ if $Path_s(A)\subseteq Visit(B)$ for some memoryless strategy $s$ of a transition system $T$.
\end{definition}
\subsubsection{Axioms}
The logical system for memoryless strategies is the same as for recall strategies with the exception that the Transitivity axiom is replaced by the following principle:
\begin{enumerate}
\item[3.] Monotonicity: $A'\rhd B \to A\rhd B$, where $A\subseteq A'$.
\end{enumerate}
This principle can be derived from Armstrong's axioms.
\begin{theorem}\label{imperfect soundness}
If $\vdash\phi$, then $T\vDash\phi$ for every system $T$.
\end{theorem}
\begin{proof}
Soundness of the Reflexivity axiom and the Augmentation axiom is similar to the case of perfect recall, see Theorem~\ref{perfect soundness}. Soundness of the Monotonicity axiom follows from $Path_s(A)\subseteq Path_s(A')$, where $A\subseteq A'$.
\end{proof}
\subsubsection{Completeness}
In the rest of this section we prove completeness of our logical system with respect to the memoryless semantics. First, we define a canonical transition system $T(X)=(S,\sim, *, I,\{\Delta_i\}_{i\in I})$ for an arbitrary maximal consistent set of formulae $X\subseteq \Phi$.
Like in the perfect recall case, the canonical system has one state for each view and an additional ``black hole" state $\circlearrowleft$. Unlike the previous construction, the new canonical transition system has more additional states besides state $\circlearrowleft$. Drawing on our original intuition of a transition system as a maze, we think about these new states as ``wormholes". For any sets of states $A$ and $B$ in the maze there is a wormhole state $w(A,B)$ that can be used to travel one-way from set $A$ to set $B$. Then, $S=V\cup\{\circlearrowleft\}\cup\{w(A,B)\;|\;A,B\subseteq V\}$.
The agent can distinguish any two different non-wormhole states, but she can not distinguish wormholes. In other words, each non-wormhole state $v\in V\cup \{\circlearrowleft\}$ forms its own indistinguishability class $[v]=\{v\}$, while all wormholes belong to the same single indistinguishability class of wormholes. Like in the perfect recall case, for each $v\in V$, we define $v^*$ to be class $[v]$.
\begin{lemma}\label{imperfect * iff}
$u\in A$ iff $[u]\in A^*$ for each view $u\in V$ and each set $A\subseteq V$. \hfill $\boxtimes\hspace{1mm}$}% \linebreak
\end{lemma}
Like in the canonical model for the perfect recall case, for sets $A,B\subseteq V$ such that $X\vdash A\rhd B$, we introduce an instruction $(A,B)$ that can be used to navigate from set $A$ to set $B$. Unlike the perfect recall case, we introduce such an instruction only if sets $A$ and $B$ are disjoint. This is an insignificant technical restriction that we use to simplify the proof of Lemma~\ref{imperfect vDash to vdash}.
\begin{definition}\label{imperfect canonical I}
$$I=\{(A,B)\;|\;X\vdash A\rhd B \mbox{ and } A\cap B=\varnothing\}.$$
\end{definition}
Recall that assumption $X\vDash A\rhd B$ implies that sets $A$ and $B$ are nonempty due to Definition~\ref{Phi}.
In the perfect recall case, instruction $(A,B)$ can be used to transition the system {\em directly} from a state in set $A$ to a state in set $B$. In our case, this transition happens via the wormhole state $w(A,B)$. In other words, when instruction $(A,B)$ is invoked in a state from set $A$, the system transitions into state $w(A,B)$. When the same instruction is invoked in $w(A,B)$, the system transitions into a state in $B$.
\begin{definition}\label{imperfect canonical Delta}
$$
\Delta_{(A,B)}(u)=
\begin{cases}
\{w(A,B)\}, & \mbox{if $u\in A$},\\
B, & \mbox{if $u=w(A,B)$},\\
\{\circlearrowleft\}, & \mbox{otherwise}.
\end{cases}
$$
\end{definition}
\begin{lemma}\label{imperfect vdash to vDash disjoint}
If $X\vdash A\rhd B$ and sets $A$ and $B$ are disjoint, then $T(X)\vDash A\rhd B$.
\end{lemma}
\begin{proof}
Assumptions $X\vdash A\rhd B$ and $A\cap B=\varnothing$ imply that $(A,B)\in I$, by Definition~\ref{imperfect canonical I}. Consider a memoryless strategy $s$ such that $s(x)=(A,B)$ for each class $x$. By Definition~\ref{imperfect sat}, it suffices to show that $Path_s(A)\subseteq Visit(B)$.
Consider any $w_0,i_1,w_1,\dots\in Path_s(A)$. Then, by Definition~\ref{path set}, sequence $w_0,i_1,w_1,\dots$ is a path under strategy $s$ such that $[w_0]\in A^*$. Hence, $w_0\in A$ by Lemma~\ref{imperfect * iff}. Thus,
\begin{equation}\label{Delta w0}
\Delta_{(A,B)}(w_0)=\{w(A,B)\}
\end{equation}
by Definition~\ref{imperfect canonical Delta}.
At the same time, $w_1\in\Delta_{i_1}(w_0)$ by Definition~\ref{history}. Hence, $w_1\in\Delta_{s[w_0]}(w_0)$ by Definition~\ref{path}. Thus, $w_1\in\Delta_{(A,B)}(w_0)$ by the choice of strategy $s$. Then, $w_1=w(A,B)$ by equation~(\ref{Delta w0}). Hence,
\begin{equation}\label{Delta w1}
\Delta_{(A,B)}(w_1)=B
\end{equation}
by Definition~\ref{imperfect canonical Delta}.
Similarly, $w_2\in\Delta_{i_2}(w_1)$ by Definition~\ref{history}. Hence, $w_2\in\Delta_{s[w_1]}(w_1)$ by Definition~\ref{path}. Thus, $w_2\in\Delta_{(A,B)}(w_1)$ by the choice of strategy $s$. Then, $w_2\in B$ by equation~(\ref{Delta w1}). Hence, $[w_2]\in B^*$ by Lemma~\ref{imperfect * iff}. Therefore, $w_0,i_1,w_1,i_2,w_2,\dots\in Visit(B)$ by Definition~\ref{visit set}.
\end{proof}
\begin{lemma}\label{imperfect vdash to vDash}
If $X\vdash A\rhd B$, then $T(X)\vDash A\rhd B$.
\end{lemma}
\begin{proof}
Suppose that $X\vdash A\rhd B$.
If $A\setminus B\neq \varnothing$. Thus, $X\vdash A\setminus B\rhd B$ by the Monotonicity Axiom. Hence, $T(X)\vDash A\setminus B\rhd B$ by Lemma~\ref{imperfect vdash to vDash disjoint}. Therefore, $T(X)\vDash A\rhd B$ due to the soundness of the Augmentation axiom, see Theorem~\ref{imperfect soundness}.
If $A\setminus B=\varnothing$, then $A\subseteq B$. Therefore, $T(X)\vDash A\rhd B$ due to the soundness of the Reflexivity axiom.
\end{proof}
Recall that all wormhole states belong to a single indistinguishability class of wormholes. For any memoryless strategy $s$, let $(A_s,B_s)$ be the instruction assigned by strategy $s$ to the class of wormholes. Once strategy $s$ is fixed, the states of the canonical transition system can be partitioned into five groups: set $A_s$, set $V\setminus A_s$, the single element set $\{\circlearrowleft\}$ containing the black hole state, the single element set $\{w(A_s,B_s)\}$ containing the wormhole state $w(A_s,B_s)$, and the set $\{w(C,D)\;|\; (C,D)\neq (A_s,B_s)\}$ of all other wormholes. Definition~\ref{imperfect canonical Delta} restricts transitions under strategy $s$ that are possible between these five groups of states. For example, from set $V\setminus A_s$ one can transition either into set $\{\circlearrowleft\}$ or into set $\{w(C,D)\;|\; (C,D)\neq (A_s,B_s)\}$. The arrows in Figure~\ref{trap figure} show all possible transitions between these five groups of states allowed under Definition~\ref{imperfect canonical Delta}. These five groups of states can be further classified into ``above the line" and ``below the line" states, as shown. Notice that once the system transitions into one of the ``below the line" states, it is trapped there and it will never be able to transition under the memoryless strategy $s$ into an ``above the line" state.
\begin{figure}[ht]
\begin{center}
\vspace{-0mm}
\scalebox{.7}{\includegraphics{Images/trap.pdf}}
\caption{Transitions under memoryless strategy $s$.}\label{trap figure}
\end{center}
\vspace{0mm}
\end{figure}
\begin{lemma}\label{imperfect vDash to vdash}
If $T(X)\vDash E\rhd G$, then $X\vdash E\rhd G$.
\end{lemma}
\begin{proof}
If $E\subseteq G$, then $\vdash E\rhd G$ by the Reflexivity axiom. In the rest of the proof we suppose that there is $e_0\in E\setminus G$. By Definition~\ref{imperfect sat}, assumption $T(X)\vDash E\rhd G$ implies that there exists a memoryless strategy $s$ such that $Path_s(E)\subseteq Visit(G)$.
First, we show that $E\setminus G\subseteq A_s$. Suppose there is a view $e_1\in E$ such that $e_1\notin G$ and $e_1\notin A_s$. By Definition~\ref{history}, single-element sequence $e_1$ is a history. Thus, by Lemma~\ref{path exists}, there is a path $\pi\in Path_s(E)$ that starts with state $e_1$. Since path $\pi$ starts in a state from set $V\setminus A_s$, all non-initial states of this path are ``below the line", see Figure~\ref{trap figure}. Hence, neither of the states in path $\pi$ belong to set $G$, because $e_1\notin G$ and none of the states ``below the line" are in set $G$ either. Thus, $\pi\in Path_s(E)$ and $\pi\notin Visit(G)$, which contradicts the choice of strategy $s$. Therefore, $E\setminus G\subseteq A_s$. In particular $e_0\in A_s$.
Second, we show that $s[e_0]=(A_s,B_s)$. Suppose that $s[e_0]=(C,D)$, where $(C,D)\neq (A_s,B_s)$. If $e_0\notin C$, then $\pi=e_0,(C,D),\circlearrowleft,s[\circlearrowleft],\circlearrowleft,s[\circlearrowleft],\dots$ is a path under strategy $s$ by Definition~\ref{path} and Definition~\ref{imperfect canonical Delta}. Note that $\pi\notin Visit(G)$ because $e_0\notin G$ and $\pi\in Path_s(E)$. This contradicts $Path_s(E)\subseteq Visit(G)$. Similarly, if $e_0\in C$, then sequence $\pi=e_0,(C,D),w(C,D),(A_s,B_s),\circlearrowleft,s[\circlearrowleft],\dots$ is a path such that $\pi\notin Visit(G)$ and $\pi\in Path_s(E)$, which again contradicts $Path_s(E)\subseteq Visit(G)$.
Third, we prove that $B_s\subseteq G$. Suppose that there is a state $b_0\in B_s\setminus G$. By Definition~\ref{history}, Definition~\ref{imperfect canonical Delta}, and the choice of instruction $(A_s,B_s)$, sequence $e_0,(A_s,B_s),w(A_s,B_s),(A_s,B_s),b_0$ is a history. Thus, by Lemma~\ref{path exists}, there is a path $\pi\in Path_s(E)$ that starts as $e_0,(A_s,B_s),w(A_s,B_s),(A_s,B_s),b_0$. Assumption $b_0\in B_s\setminus G$ implies that $b_0\in V\setminus A_s$ because sets $A_s$ and $B_s$ are disjoint by Definition~\ref{imperfect canonical I}. Thus, path $\pi$ after state $b_0$ contains only states ``below the line", see Figure~\ref{trap figure}, none of which are in set $G\subseteq V$. Recall also that $e_0,b_0\notin G$ by the choice of states $e_0$ and $b_0$. Thus, $\pi\in Path_s(E)$ and $\pi\notin Visit(G)$, which again contradicts $Path_s(E)\subseteq Visit(G)$.
Note that $X\vdash A_s\rhd B_s$ by Definition~\ref{imperfect canonical I}. Hence, $X\vdash (A_s\cup G)\rhd (B_s\cup G)$ by the Augmentation axiom. Thus, $X\vdash (A_s\cup G)\rhd G$ because $B_s\subseteq G$. At the same time, $E\setminus G\subseteq A_s$ implies that $E\subseteq A_s\cup G$. Therefore, $X\vdash E\rhd G$ by the Monotonicity axiom.
\end{proof}
\begin{lemma}\label{imperfect induction}
$X\vdash\phi$ iff $T(X)\vDash\phi$ for each $\phi\in \Phi$.
\end{lemma}
\begin{proof}
We prove this lemma by induction on the structural complexity of $\phi$. The base case follows from Lemma~\ref{imperfect vdash to vDash} and Lemma~\ref{imperfect vDash to vdash}. The induction case follows from the maximality and the consistency of set $X$ in the standard way.
\end{proof}
We are now ready to state and prove the completeness theorem for memoryless strategies.
\begin{theorem}\label{imperfect completeness}
If $T\vDash\phi$ for every system $T$, then $\vdash\phi$.
\end{theorem}
\begin{proof}
Suppose $\nvdash\phi$. Let $X$ be a maximal consistent set containing formula $\neg\phi$. Thus, $T(X)\vDash\neg\phi$ by Lemma~\ref{imperfect induction}. Therefore, $T(X)\nvDash\phi$.
\end{proof}
\section{Conclusion}
In this paper we have shown that the properties of navigability under perfect recall strategies are exactly those described by Armstrong's axioms for functional dependency in database theory. In the absence of perfect recall, the Transitivity axiom is no longer valid, but it could be replaced by the Monotonicity axiom.
\bibliographystyle{aaai}
|
\section{Introduction: Implications from the recent astonishing observational discoveries}
Star formation is one of the most crucial processes in the Universe.
This is because star formation can influence the evolution of the Universe over a wide range of physical scales.
For instance, the presence of stars can affect the formation and evolution of galaxies through the so-called stellar feedback.
Another example may be that star formation can presumably trigger the subsequent formation of planetary systems.
It has long been suggested that star and planet formation are intimately coupled with each other (see \citealt{hayashi81}).
This coupling possibly occurs via the formation and evolution of circumstellar disks
that emerge as a consequence of the conservation of angular momentum in a star formation stage (\citealt{li14}).
The disks can also serve as the birth place of planetary systems in the subsequent planet formation stage (\citealt{benz14}).
It is paramount that the recent, unprecedented high spatial resolution ALMA observations taken toward a young stellar object (YSO), HL Tauri,
have partially confirmed such a coupling between stars and planets (\citealt{alma15});
the observations reveal that nearly concentric multiple gaps are present in the observed dust continuum emission originating from the disk.
While an ultimate origin of such gaps is still under active investigation,
one of the most intriguing hypotheses is that unseen, forming (proto)planets would be responsible for the formation of the gaps (e.g., \citealt{akiyama16}).
Furthermore, the recent rapid accumulation of extrasolar planets, also known as exoplanets, has suggested that
planetary systems around the main-sequence stars are ubiquitous in the Universe (\citealt{winn15}).
This observational result also provides additional supportive evidence that there is an intimate coupling between star and planet formation.
Based on these recent astonishing observational discoveries,
it is in almost no doubt that planet formation is a natural outcome of star formation.
It is, however, important to point out that these observations would readily pose the following questions:
when does (a first step of) planet formation occur in the context of star formation?
And what processes/quantities would be ideal to explore an intimate coupling between star and planet formation?
These questions are interesting because HL Tauri is currently regarded as a quite young ($\lesssim 1$ Myr), Class I-II YSO.
In other words, if the observed gaps would be caused by unseen planets,
then the ALMA observations suggest that planet formation should have taken place much earlier than expected before.
And if this would be the case, there might be a possibility that
one can trace the formation history of planets back to the initial condition of star formation,
by focusing on certain physical processes/quantities.
To address these questions, we have developed a couple of projects,
one of which is to observationally investigate dust growth in Class 0 YSOs (\citealt{li17}),
the other of which is to theoretically model the circumstellar disk around HL Tauri,
with the emphasis on magnetic fields threading the disk (\citealt{hasegawa17}).
We will discuss below that dust growth and magnetic fields are the important ingredients
to examine a vital connection between star and planet formation.
\section{An observational study about dust growth in Class 0 YSOs}
\begin{table*}
\caption{List of our targets and the resultant value of $\beta$}
\centering
\label{table1}
\begin{tabular}{llclc}
\hline
Source Name & Region & Distance (pc) & Class & Best Fit ($\beta$) \\
\hline
L1527 & Taurus & 140 & 0/I & 0.30 \\
L1448CN & Perseus & 240 & 0 & 0.58 \\
NGC1333 IRAS4 A1 & Perseus & 240 & 0 & 0.51 \\
NGC1333 IRAS4 A2 & Perseus & 240 & 0 & 1.23 \\
NGC1333 IRAS2A & Perseus & 240 & 0 & 1.20 \\
Barnard 1b N & Perseus & 240 & 0 & 1.53 \\
Barnard 1b S & Perseus & 240 & 0 & 2.02 \\
L1157 & Isolated & 325 & 0 & 0.82 \\
Serpens FIRS 1 & Serpens & 415 & 0/I & 1.74 \\
\hline
\end{tabular}
\end{table*}
\begin{figure*}
\vskip -0.5cm
\centering
\includegraphics[angle=0,width=11cm]{fig1a.pdf}
\includegraphics[angle=0,width=10cm]{fig1b.pdf}
\caption{An example of the visibility data (top) and the result of SED fitting (bottom) for L1157.}
\label{fig1}
\vspace{-0.5cm}
\end{figure*}
Here, we introduce our project about dust growth in Class 0 YSOs (\citealt{li17}).
The main motivation of this project is to identify when dust growth, a first step of planet formation, occurs in an early stage of star formation.
In order to proceed, we utilize radio, interferometric observations covering submillimeter to cm wavelengths,
and perform spectral energy distribution (SED) fitting based on a modified black-body radiation formalism.
Longer wavelength observations are required to examine whether large ($\sim$ 1mm) dust grains are present
in the high density, star-forming region.
While collisional growth of dust grains can be possible in such a high density region,
the thermal emission from large dust grains becomes optically thin(ner) only at submillmeter wavelengths or longer there.
In principle, at least 3 different frequency band observations are needed in SED fitting
to simultaneously determine the fundamental properties of dust grains such as the dust mass, the dust size, and the dust temperature.
We make use of the abundant archival data in order to collect a good quality of data.
More specifically, we retrieve the archive data of (J)VLA, SMA, CARMA, and ALMA.
In this project, we attempt to undertake a survey type of studies.
This is because
the previous studies reported a discrepancy between observations and theory;
the previous (sub)millimeter observations suggest that
large ($\sim$ 1mm) dust grains would be needed to fully reproduce the observed value of the mm opacity spectral index
(the so-called $\beta$) for some of class 0/I YSOs (e.g., \citealt{jorgensen07,ricci10}).
Theoretical studies, however, indicate that the presence of such large grains would be very unlikely in these systems
(e.g., \citealt{ormel09}; \citealt{hirashita13}).
This is simply because these systems are in general short-lived and a relatively low density,
both of which prevent collisional growth of dust particles.
Thus, it is important to systematically investigate a large number of targets
to address whether large dust grains exist in Class 0 YSOs.
Based on our close retrieval of the rich archive data,
we find that 9 well-studied Class 0 YSOs can meet the purpose of our SED fitting (see Table \ref{table1}).
All the targets are chosen from nearby star-forming regions due to the (low) angular resolution of SMA archive data.
In addition to the requirement that 3 different wavelength observations should be available for SED fitting (see above),
we request that the selected targets exhibit at least one significant detection at $>$ 3 mm wavelength observations
where the dust thermal emission becomes optically thinner.
Furthermore, Spitzer 24 and 27 $\mu$m MIPS data are also included
in order to cover a high frequency wing in the SED plot.
Once a list of our targets are complied and the standard data reduction is performed,
we develop and carry out a systematic data analysis procedure
to derive the dust properties in Class 0 YSOs from SED fitting.
This procedure consists of three steps.
The first step is to dynamically bin the azimuthally-averaged visibility amplitude as a function of $uv$ distance
to properly weight a collection of observations that were conducted in different weather conditions and for different science purposes.
This automatized binning allows us to obtain a good quality of data that can have a good value of the signal-to-noise ratio (SNR)
while keeping the spatial information.
The second step is to decompose the resultant visibility amplitude into two components,
one of which is a spatially compact, disk-like component,
the other of which is a spatially extended, envelope component.
This decomposition is motivated by a current picture of star formation (e.g., \citealt{li14}).
The final step is to perform SED fitting using a modified black body formalism.
Based on our preliminary results, only the compact component can be utilized for our SED fitting due to its good data quality.
Figure \ref{fig1} shows an example of our results for L1157.
The top panel depicts the automatically binned, visibility amplitude as a function of $uv$ distance,
and the bottom one is for the result of our SED fitting.
We find that the compact component for this target is well represented by a straight line
where the visibility amplitude is constant (the top panel).
This indicates that the compact, disk-like structure is not resolved by the observations at all, and is regarded as a point source.
On the other hand, the extended component is well fitted by a Gaussian profile, and is viewed as a blob in the image domain.
For the SED fitting (the bottom panel),
our results show that a best fit is obtained
when the resultant value of $\beta$ is about 0.8 for the compact component.
This suggests that large ($\sim$ mm) dust grains should be present in this target.
In order to verify whether or not this result would be applicable for other YSOs,
we analyze the rest of our targets using the same procedure.
Our results are summarized in Table \ref{table1} (see the Best Fit ($\beta$)).
One immediately notices that dust growth (a low value of $\beta(<1)$) very likely takes place in roughly half of our targets,
which supports the previous observational results.
On the contrary, our results suggest that the other half of our targets have a high value of $\beta(>1)$,
which tends to be consistent with the results of theoretical studies.
Thus, our SED fitting implies that the degree of grain growth may depend on the properties of YSOs,
which may be different for different targets.
How are our results and interpretations reasonable?
One of the advantages of our method is its simplicity,
which permits a systematic analysis of the archive data for a number of targets.
It is nonetheless obvious that such simplicity is established by imposing a number of (implicit) assumptions and approximations.
One of the most serious assumptions is that the dust properties (the dust mass, the dust size, and the dust temperature)
are all represented by single quantities for the compact component.
In other words, there is no spatial information for them.
In fact, our derived value of $\beta$ should be regarded as the mean value for the compact component.
It is important to realize that
all the dust quantities are indeed a function of the distance from the central source.
As an example, the dust temperature would be much higher in the inner part of the compact component than the outer part.
Furthermore, different wavelength observations generally detect the emission originating from different regions due to the optical depth effect.
Thus, much higher spatial resolution observations, fine-spaced SED sampling, and/or more detailed radiative transfer modeling
would be demanded to fully address the presence of large ($\sim$ mm) dust grains in Class 0 YSOs.
In addition, it would be useful to investigate this using another method such as a chemistry model (\citealt{harada17}).
A more complete SED modeling and discussion can be found in \citealt{li17}.
\begin{figure*}[!ht]
\centering
$\begin{array}{cc}
\includegraphics[angle=0,width=8.5cm]{fig2a.pdf}
\includegraphics[angle=0,width=8.5cm]{fig2b.pdf}
\end{array}$
\caption{Our semi-analytical formulae for the accretion stress ($\bar{W}^{\rm AD}_{r\phi}$),
the wind stress ($\bar{W}^{\rm AD}_{z\phi}$), and the local component of disk turbulence ($\alpha^{\rm AD}_{\rm turb}$) as a function of the plasma $\beta_z$ (left),
and the resultant disk properties for the case that $\beta_z =80$ (right).
On the right panel, the disk accretion rate, the gas surface density profile, and the value of $\alpha_{\rm turb}$ are plotted as a function of the distance from the central star,
from top to bottom.}
\label{fig2}
\end{figure*}
\section{A theoretical modeling of the magnetized circumstellar disk around HL Tauri}
We now discuss another project in which a theoretical modeling of the circumstellar disk around HL Tauri is developed (\citealt{hasegawa17}).
The main motivation of this project is to demonstrate how magnetic fields threading disks would be useful
to make a vital connection between star and planet formation.
As described above, the ALMA observations of HL Tauri are revolutionizing a current picture of star and planet formation.
While one of the most striking features in the observations is the nearly concentric multiple gaps in the dust continuum emission,
we here focus on the interplay between magnetically driven disk accretion and the vertical distribution of dust particles in the disk.
It has been suggested that HL Tauri can probably be classified as a young $(\lesssim 1$ Myr), Class I-II YSO.
In fact, the previous observations taken toward HL Tauri reveal a number of characteristic features of YSOs
such as molecular outflow and the presence of an envelope (\citealt[references herein]{akiyama16}).
Furthermore, some observations and modeling indicate that
the disk accretion rate onto the central star is about $10^{-7}-10^{-6}$ M$_{\odot}$ yr$^{-1}$ (\citealt{hayashi93,beck10}).
When the famous $\alpha-$prescription is adopted (\citealt{shakura73}),
this accretion rate corresponds to $\alpha_{\rm acc}$ which is $10^{-2}-10^{-1}$.
Note that $\alpha$ is labeled as $\alpha_{\rm acc}$ to explicitly denote that this $\alpha$ is derived from the disk accretion rate.
Such a high value of $\alpha_{\rm acc}$ may be reasonable for HL Tauri due to its young age.
At the same time, the recent ALMA observations provide another opportunity to conduct a detailed radiative transfer modeling.
\citet{pinte16} have undertaken this and found that the vertical distribution of dust particles is very thin ($\lesssim 1$ AU at the semimajor axis of $\sim$ 100 AU).
Based on their modeling,
the required value of $\alpha$ (defined as $\alpha_{\rm turb}$) should be as small as a few $10^{-4}$
in order to realize such a high degree of vertical dust settling.
When turbulence is the main driver of disk accretion, it is generally expected that $\alpha_{\rm acc} \simeq \alpha_{\rm turb}$,
since most of disk turbulence would be more or less local processes.
For the HL Tauri case, however,
it is obvious that there is more than one order magnitude difference between $\alpha_{\rm acc}$ and $\alpha_{\rm turb}$.
In order to reconcile this big mismatch,
we develop a simple, but physically motivated disk model.
In this model, it is assumed that
disk accretion onto the central star is driven not only by local turbulence (which is adopted in many of disk models),
but also by magnetically induced disk winds.
The recent numerical simulations show that disk winds can be launched at disk surfaces and remove a substantial amount of the disk angular momentum vertically
under the condition that disks are threaded by relatively strong, vertical magnetic fields (e.g., \citealt{suzuki09,simon13}).
Disk winds would play an important role in the evolution especially when ambipolar diffusion, a non-ideal MHD effect, is properly taken into account (e.g., \citealt{bai11}).
Practically, we first develop semi-analytical formulae for the accretion stress ($\bar{W}^{\rm AD}_{r\phi}$),
the wind stress ($\bar{W}^{\rm AD}_{z\phi}$), and the local component of disk turbulence ($\alpha^{\rm AD}_{\rm turb}$),
as a function of the plasma $\beta_z$.
Note that $\bar{W}^{\rm AD}_{r\phi} \propto \alpha_{acc}$.
Also, beware of $\beta_z$ which is different from the opacity index ($\beta$) that appeared in Section 2.
These formulae are obtained, so that they can fit the numerical simulations very well (\citealt[][see the left panel of Figure \ref{fig2}]{simon13}).
We then parameterize the plasma $\beta_z$ to find out a disk configuration that can account for the HL Tauri disk,
under the assumption of steady disk accretion.
Figure \ref{fig2} (the right panel) shows the resultant disk properties for the case that $\beta_z = 80$.
From top to bottom, the disk accretion rate, the gas surface density profile, and the value of $\alpha_{\rm turb}^{\rm AD}$ are plotted
as a function of the distance from the central star.
As a reference, the dust surface density distribution (denoted by $\Sigma_d^{\rm Kwon}$) that is derived from CARMA observations (\citealt{kwon11}),
and the locations of gaps (denoted by the hatched regions) detected by ALMA observations (\citealt{alma15}) are both included in the middle.
Our results show that when magnetically induced disk winds are taken into account,
the winds contribute predominantly to the disk accretion rate (see the top).
This occurs because the wind stress ($\bar{W}^{\rm AD}_{z\phi}$) can be scaled as $\alpha_{acc} r / h$,
where $r$ is the distance from the central star, $h$ is the pressure scale height, and $r/h \gg 1$.
For the gas distribution (the middle),
we find that the gas surface density becomes relatively low.
This is a direct outcome that the wind can remove the disk angular momentum significantly.
It is interesting that the resultant value of the gas surface density may be comparable to that of dust,
especially at the locations of observed gaps.
Finally, our results (the bottom) indicate that disk wind models can reproduce a low value of $\alpha_{\rm turb}$
that is suggested for the HL Tauri disk (see the hatched region).
This becomes possible thanks to the winds that can establish efficient disk accretion onto the central star
without a high level of disk turbulence.
Thus, our modeling suggests that magnetically induced disk winds can achieve a disk configuration
where the disk accretion rate is high due to the winds
while the local turbulence is weak enough to realize a high degree of vertical dust setting.
How can our current, simplified modeling be improved?
There might be at least 3 ways to sophisticate our model.
First, one can adopt more realistic formulae for $\bar{W}^{\rm AD}_{r\phi}$, $\bar{W}^{\rm AD}_{z\phi}$, and $\alpha^{\rm AD}_{\rm turb}$,
following the more recent work such as \citet{bai16,suzuki16}.
Second, the vertical distribution of dust particles can be computed directly
from the dust size and magnetically driven disk turbulence, for a given value of the plasma $\beta_z$ (e.g., \citealt{zhu15}).
Third, a global, time-dependent, 1D disk model can be developed
by implementing the above two improvements.
Such a disk model would be useful to examine how magnetic fields are related to the observed multiple gaps in the HL Tauri disk,
and hence ultimately serve as a powerful tool to explore an intimate coupling between star and planet formation.
A more complete discussion can be found in \citealt{hasegawa17},
where some improvements are already undertaken.
\section{Summary and future prospects}
We have discussed two projects that have been developed to address when a first step of planet formation occurs in star-forming environments
and what processes/quantities are useful to make a vital link between star and planet formation.
As described above, our current projects can provide only a partial answer to these questions.
In our future work, we will undertake much higher spatial resolution observations and/or more detailed modeling
to obtain a more comprehensive understanding
of how star and planet formation take place together and of how these two processes are intimately coupled with each other.
\section*{Acknowledgments}
This research was carried out at Jet Propulsion Laboratory, California Institute of Technology under a contract with NASA.
Y.H. is supported by JPL/Caltech.
|
\section{\label{sec1}Introduction}
Graphene-based nanostructures have attracted a wide attention owing to their several interesting electronic and transport properties \cite{novo,neto,zhang,vp,jun,lin,du} for a decade. Unconventional quantum Hall effect \cite{novo,zhang,vp}, half metallicity \cite {jun,lin}, high carrier mobility \cite{du}, such interesting features make graphene as promising candidates for nanoelectronics and spintronics applications. The recent fabrication of freestanding graphene nanoribbons (GNRs) \cite{meyer,moro} has generated renewed interest in carbon-based materials with exotic properties. GNRs are basically a single strips of graphene. The electronic properties \cite{fujita,saito} of graphene nanoribbons depend on the geometry of the edges and lateral width of the nanoribbons \cite{nakada}. According to the edge termination type, mainly there are two kinds of GNR, namely armchair graphene nanoribbon (AGNR) and zigzag graphene nanoribbon (ZGNR).
Since the edges play an important role in determining the electronic properties of GNR, they offer a variety of possibilities for tunable electronic properties, such as edge modulation by inorganic atoms, molecules or radicals \cite{wang,hzheng,song,hzeng}, application of transverse electric fields \cite{xh}, adsorption or doping of atoms or molecules \cite{hz,longo,chan,av,dw,kan,sevin,rigo,brito} etc. Metal atoms adsorbed onto graphene sheets also represent a new way for the development of new electronic or spintronic devices. The electronic, structural, and magnetic properties of transition metals (TM) on graphene sheets \cite{chan,av,dw} and graphene nanoribbons (GNR) \cite{kan,sevin,rigo,brito} have been studied extensively, which are mostly based on ab-initio density-functional theory (DFT). The spin dependent transport in GNRs in presence of Rashba SOC has been investigated in some cases, such as spin filtering effect in zigzag GNR \cite{liu}, possible spin polarization directions for GNR with Rashba SOC \cite{chico}, effects of spatial symmetry of GNR on spin polarized transport \cite{qzhang} etc.
Among the metal adatoms, the study of GNRs in presence of transition metals warrants some special attention since TM serve as important catalysts for the synthesis of graphite, CNT, GNR etc. Since TM catalysts (particularly iron) is a common impurity in the graphite \cite{pe}, graphene layers fabricated from graphite are likely to have these impurities. Longo et al. have showed that the behaviour of Fe atom in a GNR is magnetic, in contrast to the behavior found in graphene \cite{av,ejg}. Mao et al. \cite{mao} showed that adsorption of Fe on graphene makes graphene metallic and generates 100\% spin polarization. Basically the study of TM adsorption on graphene \cite{cao,nk}
shows possible applications in the realization of graphene-based electronic and spintronic devices. Motivated by the above studies, it will be highly desirable to explore the spintronic behaviour of TM adsorption on graphene and graphene-based structures.
In the present work, we explore the charge and spin transport properties of ZGNR decorated by magnetic adatoms. For this, we consider a case where TM (particularly Fe) are adsorbed onto ZGNR. The resulting broken structural symmetry gives rise to a Rashba spin-orbit interaction (RSOI) and the hybridization between the carbon $\pi$ state and the 3$d$-shell states of magnetic adatoms generates a macroscopic exchange field \cite{jiang}. A hall conductivity, $\sigma^H_{xy}$ of magnitude $2e^2/h$ is observed for the case where the Fermi energy lies in the bulk gap. First principle calculations report a bulk gap of almost 5.5 meV in Fe adsorbed GNRs, which should be possible to verify in experiments \cite{jiang}. Further, in order to avoid any complicacy, such as adatom-adatom interaction, spin-spin correlation etc., we consider very small concentration of magnetic adatoms, so that any interactions other than RSOI and exchange field will not be present.
We organize our paper as follows. In the following section, we present, for completeness, the theoretical formalism leading to the expressions for the two terminal charge and spin polarized conductances and four terminal longitudinal and spin Hall conductances using the well known Landauer-B\"{u}ttiker formula. After that we include an elaborate discussion of the results. Here, we have tried to resolve few queries, how the spin polarized conductance behaves in the two terminal case, whether there is any similarity between the two terminal spin polarized conductance and four terminal spin Hall conductance. We have also included an interesting comparison for the charge conductance properties for the case of two and four terminal ZGNR.
\section{\label{sec2}Theoretical formulation and model}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.3\textwidth]{fig1.eps}
\caption{(Color online) Schematic view of a two terminal graphene nanoribbon. The black and white circles represent the A and B sublattices of graphene. The brown circles are the magnetic adatoms. The green circles are the affected site due to magnetic adatoms. The black lines surrounding the magnetic atoms correspond to nearest neighbour hopping and Rashba SOC. Rest of the black lines contain only nearest neighbour hopping. The leads are attached at both ends, which are denoted by red color and are semi-infinite in nature. The leads are free of any kind of SOC. $N_x$ and $N_y$ are the length and width of the nanoribbon respectively.}
\label{setup}
\end{center}
\end{figure}
To begin with we describe the geometry of the system and make our notations clear. We consider a graphene sheet adsorbed with magnetic atoms, which induces Rashba SOI and generates an exchange field. The effective tight-binding Hamiltonian for graphene with such adatoms is given by \cite{jiang},
\begin{eqnarray}
H&=& -
t\sum\limits_{\langle ij\rangle}c_i^{\dagger} c_j +
i\lambda_R\sum\limits_{\langle ij\rangle\in\mathcal{R}}c_i^{\dagger} \left( {\vec{\sigma}} \times {\bf\hat{d}}_{ij}\right)_z c_j \nonumber \\ &+&
\lambda_{EB}\sum\limits_{i\in \mathcal{R}} c_i^{\dagger}\sigma_zc_i
\label{h2}
\end{eqnarray}
where $c_i^{\dagger}=\left(c_{i\uparrow}^{\dagger} \quad c_{i\downarrow}^{\dagger}\right)$ is the creation operator of electrons at site $i$. The first term is the nearest neighbour hopping term, with a hopping strength, $t=2.7$ eV. The second term is the nearest neighbour Rashba term which explicitly violates $z\rightarrow -z$ symmetry. This term is induced by the adatoms residing on the set of hexagons $\mathcal{R}$ that are inhabited by the magnetic adatoms. The third term is the exchange bias that (as shown in Fig.\ref{setup}) originates due to magnetic adatoms.
\subsection{Two terminal (2T) GNR: formulation of charge and spin polarized conductances}
The zero temperature conductance, $G$ which
denotes the charge transport measurements, is related with
the transmission coefficient as in \cite {land_cond,land_cond2},
\begin{equation}
G = \frac{e^2}{h} T(E)
\end{equation}
The Transmission coefficient can be calculated via \cite{caroli,Fisher-Lee},
\begin{equation}
T = \text{Tr}\left[\Gamma_R {\cal G}_R
\Gamma_L {\cal G}_A\right]
\end{equation}
${\cal G}_{R(A)}$ is the retarded (advance) Green's function. $\Gamma_{L(R)}$ are the coupling matrices representing the
coupling between the central region and the left (right) lead. They are
defined by the relation \cite{dutta},
\begin{equation}
\Gamma_{L(R)} = i\left[\Sigma_{L(R)} -
(\Sigma_{L(R)})^\dagger\right]
\end{equation}
Here $\Sigma_{L(R)}$ is the retarded self-energy associated with the left (right) lead. The self-energy contribution is computed
by modeling each terminal as a semi-infinite perfect wire \cite{nico}.
We define the spin polarized conductance as,
\begin{equation}
G^s_\alpha = \frac{\hbar}{2e}\left[\frac{I_R^\alpha}{V_L-V_R}\right]
\end{equation}
$I_R^\alpha \;(\alpha=x,y,z)$ is the spin current flowing through right lead and $V_{L/R}$ is the potential at the left/right lead.
The spin polarized current can be calculated using \cite{chang,roche},
\begin{equation}
I_R^\alpha = \frac{e^2}{h} \text{Tr}\left[\hat{\sigma}_\alpha\Gamma_R G_R
\Gamma_L G_A\right](V_R - V_L)
\end{equation}
where, $\alpha=x,y,z$ and $\sigma$ denote the Pauli matrices.
Fig.\ref{setup} shows the geometry used for the calculations of charge and spin polarized conductances. Fig.\ref{setup} is the setup corresponds to ZGNR. The length and width of these systems can be determined as shown in the given figure. In Fig.\ref{setup}, the width is, $N_y = 12$ and the length is, $N_x = 21$. Thus we can denote the zigzag setup by $N_x$Z-$N_y$A $=$ 17Z-12A. This nomenclature for denoting the dimensions of the nanoribbon will be followed throughout the paper.
The black and white circles stand for the A and B sublattices of graphene. The brown circles are the magnetic adatoms. The green circles are the affected site due to magnetic adatoms. The black lines surrounding the adatoms correspond to the nearest neighbour hopping and the Rashba SOC. Rest of the black lines denote only nearest neighbour hopping. The leads are semi-infinite in nature, attached at both ends and are denoted by red color. The leads are considered to describes by a pure tight binding graphene lattice and hence are free of any kind of SOC.
\subsection{Four terminal (4T) GNR: formulation of longitudinal and spin Hall conductances}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.3\textwidth]{fig2.eps}
\caption{(Color online) Schematic view of a four terminal graphene nanoribbon. The black and white circles represent the A and B sublattices of graphene. The brown circles are the magnetic adatoms. The green circles are the affected sites due to magnetic adatoms. The black lines surrounding the Au atoms correspond to nearest neighbour hopping and Rashba SOC. Rest of the black lines contain only nearest neighbour hopping. The leads are attached at the four sides, which are denoted by red color and are semi-infinite in nature. The leads are free of any kind of SOC.}
\label{setup4}
\end{center}
\end{figure}
In order to observe the spin Hall conductance, a charge current is allowed to flow between terminal 1 and 2, and a spin current is observed to flow along the transferred direction of the rectangular sample, that is between terminals 3 and 4 as shown in Fig.\ref{setup4}.
In case of four terminal device, the longitudinal and spin Hall conductances are defined as follows,
\begin{equation}
G = \frac{I^q_2}{V_1-V_2}\quad G_{SH}^\alpha = \frac{\hbar}{2e}\left[\frac{I_3^\alpha}{V_1-V_2}\right]
\end{equation}
where, $I^q_2$ is the charge current flowing through terminal 2 and $I_3^\alpha \;(\alpha=x,y,z)$ is the spin current polarized in a particular direction $\alpha$ and flowing through terminal 3. $V_i$ is the potential at the $i$-th lead.
Following the Landauer-B\"{u}ttiker formula \cite{land_cond,land_cond2}, the charge and spin currents can be calculated from the following expression \cite{chang,roche},
\begin{equation}
I^{\alpha}_p = \frac{e^2}{h} \sum\limits_q \text{Tr}\left[\hat{\sigma}_\alpha\Gamma_q {\cal G}_R
\Gamma_p {\cal G}_A\right](V_p - V_q)
\label{current}
\end{equation}
where, $\hat{\sigma}_\alpha=(\sigma_0,\sigma_x,\sigma_y,\sigma_z)$. $\sigma_0$ is a $2\times2$ identity matrix and $\sigma_x,\sigma_y,\sigma_z$ are the Pauli matrices. $\sigma_0$ in Eq.\ref{current} gives the usual charge current, while the Pauli matrices yield the spin currents polarized in different directions ($x$, $y$ and $z$).
Since leads 3 and 4 are voltage probes, $I^q_3=I^q_4=0$. On the other
hand, as the currents in various leads depend only on voltage
differences among them, we can set one of the voltages to zero without
any loss of generality. Here we set $V_2 = 0$.
\section{\label{sec3}Results and Discussions}
We have studied the charge and spin conductance properties in ZGNR decorated by magnetic adatoms and compared the results for two and four terminal devices.
Before embarking on the results, we briefly describe the values of different parameters used in our calculation. We set the hopping term, $t=2.7$ eV. Throughout our work, we take the ZGNR setup as 89Z-48A for the two terminal case. While for the four terminal device it is 85Z-52A. All the energies are measured in unit of $t$. The charge conductance is measured in units of $\frac{e^2}{\hbar}$ and the spin polarized conductance is measured in units of $\frac{e}{4\pi}$ for the 2T case. The longitudinal and spin Hall conductances are measured in units of $\frac{e^2}{\hbar}$ and $\frac{e}{4\pi}$ respectively for the 4T case. Also the lattice constant, $a$ is taken to be unity. All the measurable quantities are averaged over 50 independent random-adatom configurations for different adatom concentrations, $n_{ad}$. In this work, we have considered three different adatom concentrations, namely, $n_{ad} = 0.025,0.05$ and $0.1$. We have checked that 50 configurations are adequate in the present context, especially, since the adatom densities considered here are small. For most of our numerical calculations we have used KWANT \cite{kwant}.
\subsection{Two terminal}
The behaviour of charge conductance, $G$ for two terminal case is studied as a function of the Fermi energy, $E$ as shown in Fig.\ref{2t_cond} for two different cases, $\lambda_R=0$ in Fig.\ref{2t_cond}(a) and $\lambda_R=0.2$ in Fig.\ref{2t_cond}(b). The strength of the exchange field is kept fixed at $\lambda_{EB}=0.18$. Corresponding to these parameters, the insulating bulk states are clearly discernible from the chiral conducting edge states \cite{jiang}. The variation of $G$ as a function of the Fermi energy is quite similar for two different strengths of the Rashba SOI. $G$ is symmetric around the zero of the Fermi energy and shows a nice necklace-type pattern for both the case. However, the behaviour of $G$ close to zero of the Fermi energy is different for the two different values of $\lambda_R$. In order to visualize the behaviour of $G$ close to zero of the Fermi energy, we plot $G$ for a small range of the Fermi energy as shown in the inset in Fig.\ref{2t_cond}. In the absence of Rashba SOI, $G$ always stays finite. On the other hand, in presence of Rashba SOI, $G$ tends to vanish values in the vicinity of $E=0$ as we increase the adatom concentration. It identically vanishes at $n_{ad} = 0.1$. The $2e^2/h$ plateau in pristine graphene in presence of Rashba SOI is believed to be a signature of the quantum spin Hall insulating phase protected by the time reversal symmetry. In presence of the exchange bias, the time reversal symmetry is explicitly violated leading to the disappearance of the plateau and giving rise to features of an ordinary insulator.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.35\textwidth]{fig3a.eps}\quad\includegraphics[width=0.35\textwidth]{fig3b.eps}
\caption{(Color online) The two terminal charge conductance, $G$ is plotted as a function of the Fermi energy, E (a) in the absence of Rashba SOI and (b) in the presence of Rashba SOI with strength, $\lambda_R = 0.2$. The strength of the exchange field is fixed in both the cases at $\lambda_{EB} =0.18$. To visualize the behaviour of $G$ near the zero of the Fermi energy, we plot the variation of $G$ for a small range of the Fermi energy as shown in the insets.}
\label{2t_cond}
\end{center}
\end{figure}
We did not show the fluctuations in the charge conductance in Fig.\ref{2t_cond} in order to avoid cluttering of data. However, the fluctuations in the charge conductance, $\Delta G$ is plotted as a function of the Fermi energy separately as shown in Fig.\ref{2t_dcond} for clarity. Fig.\ref{2t_dcond}(a) shows the variation of $G$ in the absence of Rashba SOI and Fig.\ref{2t_dcond}(b) for a Rashba strength, $\lambda_R = 0.2$. In both cases, fluctuations increases with increasing the adatom concentration. Rashba SOI suppresses the fluctuations in the charge conductance as seen from Fig.\ref{2t_dcond}(b). The spike-like behaviour in $\Delta G$ is due to the finite number of open channels in the leads and may be due to the spin precision effect \cite{sudin,lsheng}. For clarity, we have taken only two different adatom concentrations, namely $n_{ad}=$ 0.025 (denoted by blue color) and 0.1 (denoted by green color).
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.35\textwidth]{fig4a.eps}\\\includegraphics[width=0.35\textwidth]{fig4b.eps}
\caption{(Color online) The fluctuations in the charge conductance, $\Delta G$ is plotted as a function of the Fermi energy, $E$ for (a) $\lambda_R = 0$ and (b) $\lambda_R=0.2$. Only two different adatom concentrations are taken, namely, $n_{ad} = 0.025$ and 0.1 for clarity.}
\label{2t_dcond}
\end{center}
\end{figure}
It is expected that in the absence of Rashba SOI, the exchange field generates only the $z$-component of the spin polarized conductance, since the exchange field contains only the $z$-component of the Pauli spin matrix (see Eq.\ref{h2}). However, the inclusion of the Rashba SOI generates all the three components of the spin polarized components.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.3\textwidth]{fig5a.eps}\\\includegraphics[width=0.3\textwidth]{fig5b.eps}
\caption{(Color online) (a-c)The two terminal spin polarized conductance, $G^s_z$ is plotted as a function of the Fermi energy in the absence of Rashba SOI for three different adatom concentrations, namely $n_{ad} =$0.025, 0.05 and 0.1. (d-f) Their corresponding fluctuations are plotted as a function of $E$.}
\label{2t_gsz}
\end{center}
\end{figure}
Fig.\ref{2t_gsz}(a-c) shows the variation of the $z$-component of the spin polarized component, $G^s_z$ (in units of $e/4\pi$) as a function of the Fermi energy, $E$ in the absence of Rashba SOI for three different adatom concentrations. $G^s_z$ is antisymmetric as a function of the Fermi energy owing to the electron-hole symmetry \cite{chico,moca}. Though we consider very dilute adatom concentrations, $G^s_z$ acquires quite a large value in the absence of Rashba SOI. The spike-like nature as discussed before is due to the finite number of open channels in the leads. The fluctuations in $G^s_z$ denoted by $\Delta G^s_z$ is plotted as a function of the Fermi energy, $E$ as shown in Fig.\ref{2t_gsz}(d-f). $\Delta G^s_z$ is symmetric about $E=0$. However, in the absence of Rashba SOI, $\Delta G^s_z$ is more or less independent of the adatom concentration.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.3\textwidth]{fig6a.eps}\\\includegraphics[width=0.3\textwidth]{fig6b.eps}
\caption{(Color online) In presence of Rashba SOI, the three components of the spin polarized conductance (a) $G^s_x$, (b) $G^s_y$ and $G^s_z$ are plotted as a function of the Fermi energy. (d-f) Their fluctuations are plotted as a function of $E$. }
\label{2t_gsz_r}
\end{center}
\end{figure}
The behaviour of all the three components of spin polarized conductance as a function of $E$ in presence of Rashba SOI is plotted in Fig.\ref{2t_gsz_r}(a-c). The magnitude of the spin polarized conductance increases with increasing the adatom concentration. The $z$-component, namely $G^s_z$ has a higher magnitude than the other two components of the spin polarized conductance. However, all of them are antisymmetric about $E=0$ as they should be. The magnitude of $G^s_z$ in presence of Rashba SOI is lower than the values corresponding to that of in the absence of Rashba SOI. The spike-like feature is mostly prominent in the behaviour of $G^s_z$ and a little in $G^s_y$. The behaviour of the corresponding fluctuations in the spin polarized conductance is shown in Fig.\ref{2t_gsz_r}(d-f). As the adatom concentration is increased, the fluctuations also increase. But the most interesting feature about the fluctuations is that all the three components, that is $\Delta G^s_i (i=x,y,z)$ have a unique behaviour as elaborated in the following. Even if the different components of the spin polarized conductance have different variations as a function of $E$ and have different orders of magnitude for a particular strength of the exchange field in presence of Rashba SOI, yet the orders of magnitude of their fluctuations are almost same. Further, as usual the fluctuations increases with increasing adatom concentration.
The inclusion of magnetic adatoms generates the exchange field and as a result, the time reversal symmetry is broken. Hence the system will no longer have a Kramer's doublet. This phenomena can be justified through the density of states (DOS) for different species, that is DOS for up and down spins. We denote the total density of states by DOS, the density of states coming from spin up electron by UDOS and that from spin down electron by DDOS. We also define the difference between the UDOS and DDOS by DiffDOS.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.3\textwidth]{fig7a.eps}\\\includegraphics[width=0.3\textwidth]{fig7b.eps}
\caption{(Color online) The density of states (DOS) for the spin up and spin down electrons are plotted as a function of the Fermi energy, $E$ (a) for $\lambda_R = 0$ and (d) for $\lambda_R = 0.2$. The difference between DOS$\uparrow$ and DOS$\downarrow$ are plotted as a function of $E$ (b) for $\lambda_R = 0$ and (e) for $\lambda_R = 0.2$. The behaviour of total DOS is plotted as a function of the Fermi energy (c) for $\lambda_R = 0$ and (e) for $\lambda_R = 0.2$. for clarity we consider only a single adatom concentration, $n_{ad} = 0.1$.}
\label{2t_dos}
\end{center}
\end{figure}
The variation of the DOS as a function of the Fermi energy is hence plotted in Fig.\ref{2t_dos} for a particular strength of the exchange field, $\lambda_{EB}=0.18$. Fig.\ref{2t_dos}(a) shows the behaviour of UDOS and DDOS in the absence of Rashba SOI and Fig.\ref{2t_dos}(d) in presence of Rashba SOI. It is observed that UDOS is antisymmetric with respect to DDOS as a function of the Fermi energy. The difference between UDOS and DDOS, that is DiffDOS is plotted in Fig.\ref{2t_dos}(b) and Fig.\ref{2t_dos}(e) in the absence and presence of Rashba SOI respectively. In both cases UDOS $-$ DDOS is antisymmetric about $E=0$. The total DOS, that is the sum of DOS due to spin up and spin down electrons is plotted in Fig.\ref{2t_dos}(c) and Fig.\ref{2t_dos}(f). DOS is symmetric about $E=0$.
We can summarize the observations noted from the DOS data, as shown by the following set of properties,
\begin{eqnarray}
\text{UDOS}(E) &=& \text{DDOS}(-E) \nonumber\\
\text{DiffDOS}(E) &=& -\text{DiffDOS}(-E) \\
\text{DOS}(E) &=& \text{DOS}(-E) \nonumber
\label{dos}
\end{eqnarray}
In order to have a deeper look, we have also shown the local charge and spin currents for a fixed Fermi energy. Since the left lead acts as an input to the system, we are interested in the local charge and spin currents which have originated due to the left lead.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.23\textwidth]{fig8a.eps}\includegraphics[width=0.23\textwidth]{fig8b.eps}
\caption{(Color online) (a) The local charge current, $J_0$ and (b) the $z$-component of the local spin current are shown in the absence of Rashba SOI. Both the figures are obtained for the adatom concentration $n_{ad} = 0.1$ and we set the Fermi energy at $E = -0.09$. Figure (b) shows presence of some centers around which $J_z$ whirls. }
\label{2t_current}
\end{center}
\end{figure}
Fig.\ref{2t_current}(a) shows the nature of the local charge current, $J_0$ in the absence of Rashba SOI for a fixed value of the Fermi energy, namely $E=-0.09$. Clearly the local charge current is flowing between the left to the right lead without any distortion. In case of spin currents, there are three components, namely $J_x,J_y,J_z$. Since in the absence of Rashba SOI only $z$-component of the local spin current exits, we have calculated $J_z$ as shown in Fig.\ref{2t_current}(b). The number of paths between left and right leads are less compared to the local charge current. Few of the paths are even circling around certain points signaling a vortex like behaviour.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.2\textwidth]{fig9a.eps}\quad\includegraphics[width=0.2\textwidth]{fig9b.eps}
\includegraphics[width=0.2\textwidth]{fig9c.eps}\quad\includegraphics[width=0.2\textwidth]{fig9d.eps}
\caption{(Color online) (a) The local charge current $J_0$, (b) the $x$-component of the local spin current, $J_x$, (c) the $y$-component of the local spin current, $J_y$ and (d) the $z$-component of the local spin current, $J_z$ are shown in the presence of Rashba SOI with strength $\lambda_R = 0.2$. The strength of the exchange field is $\lambda_{EB} = 0.18$. We set the Fermi energy at $E = -0.09$ and the adatom concentration in the present case is $n_{ad} = 0.1$. Centers similar to Fig.\ref{2t_current}(b) have been noticed in the $J_x$, $J_y$ and $J_z$ plots.}
\label{2t_current_lr}
\end{center}
\end{figure}
Since the inclusion of Rashba SOI generates the two other components of the spin polarized conductance, namely the $x$ and $y$ components, it will be meaningful to study the local spin currents of these components. Existence of all of these components should be beneficial for spintronic applications of magnetic adatom decorated GNRs.
For this, we set, as before, the strength of the Rashba SOI at $\lambda_R=0.2$ and fixed the energy at $E=-0.09$. The local charge and spin currents are shown in Fig.\ref{2t_current_lr}. The local charge current, $J_0$ is again flowing between left and right leads as shown in Fig.\ref{2t_current_lr}(a), which reveals that $J_0$ is almost independent of the Rashba SOI which is understandable. Fig.\ref{2t_current_lr}(b),(c) and (d) show respectively $J_x$, $J_y$ and $J_z$. If we recall Fig.\ref{2t_gsz_r}, where $G^s_i (i=x,y,z)$ are plotted as a function of the Fermi energy, the $x$-component of the spin polarized conductance has the lower magnitude compared to the other two components. This can be explained from Fig.\ref{2t_current_lr}(b). Here we see that the number of clear paths between left and right leads are very less compared to the paths corresponding to $J_y$ and $J_z$. As a result, $G^s_x$ has lesser magnitude than $G^s_y$ and $G^s_z$.
One point should be mentioned here that all the plots in Fig.\ref{2t_current} and Fig.\ref{2t_current_lr} are obtained for a single configuration corresponding to $n_{ad}=0.1$. Another configuration or taking an average over several configurations may change the plots, but the qualitative feature will remain same.
\subsection{Four terminal}
Having emphasized upon the conductance characteristics of a 2T GNR, it is useful to compare and contrast with respect to the 4T devices.
To begin with the results in case of four terminal device, we shall remind ourselves that the setup for measuring the longitudinal conductance $(G)$ and spin Hall conductance $(G_{SH})$. As shown in Fig.\ref{setup4}, an electric current is allowed to pass between terminal 1 and 2, and the longitudinal conductance is measured. Terminals 3 and 4 are the voltage probes, hence there will be no charge current flowing through them. The spin Hall conductance is measured between terminals 3 and 4.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.3\textwidth]{fig10a.eps}\\\includegraphics[width=0.3\textwidth]{fig10b.eps}
\caption{(Color online) The four terminal charge conductance, $G$ is plotted as a function of the Fermi energy, $E$ (a) in the absence of Rashba SOI and (b) in the presence of Rashba SOI with strength, $\lambda_R = 0.2$. The strength of the exchange field is fixed in both the cases at $\lambda_{EB} =0.18$. To visualize the behaviour of $G$ near the zero of the Fermi energy, we plot the variation of $G$ for a small range of the Fermi energy as shown in the inset.}
\label{4t_cond}
\end{center}
\end{figure}
The behaviour of the charge conductance, $G$ for a 4T case is studied as a function of the Fermi energy ($E$) as shown in Fig.\ref{4t_cond} for two different cases as earlier, namely, $\lambda_R=0$ in Fig.\ref{4t_cond}(a) and $\lambda_R=0.2$ in Fig.\ref{4t_cond}(b). The strength of the exchange field is kept fixed as in case of two terminal case, that is at $\lambda_{EB}=0.18$. $G$ is symmetric around the zero of the Fermi energy and shows a nice necklace-type pattern for both the case. However, the behaviour of $G$ close to zero of the Fermi energy is different for two different values of $\lambda_R$. We plot $G$ for a small range of the Fermi energy in the vicinity of $E=0$ as shown in the inset in Fig.\ref{4t_cond}. In the absence of Rashba SOI, $G$ always remain non-zero. On the other hand, in presence of Rashba SOI, $G$ tends to have lower values near $E=0$ as we increase the adatom concentration. One can expect an insulating behaviour about $E=0$ in presence of Rashba SOI if we increase the adatom concentration beyond $n_{ad} = 0.1$. There is a crucial difference between the 2T and 4T devices with regard to the plateau at $2e^2/h$ in the vicinity of the zero of the Fermi energy implying the existence of a quantum spin Hall phase. As discussed earlier, there is a fairly flat plateau for a 2T GNR without magnetic adatoms, which is visibly absent for a 4T GNR (see Fig.\ref{4t_cond} and the insets). This is another minor, where in a 2T device, we have observed $G$ to completely vanish at $n_{ad}=0.1$, while in the 4T GNR, it would eventually vanish for $n_{ad} > 0.1$.
The fluctuations in the charge conductance, $\Delta G$ is plotted in Fig.\ref{4t_dcond}(a) in the absence of Rashba SOI and in Fig.\ref{4t_dcond}(b) in presence of Rashba SOI. In each of the plots, $\Delta G$ falls off as we come close to $E=0$, from either side of the band and again increases near $E=0$. The fluctuations increases as we increase the adatom concentration. By comparison the two plots, we observe that the inclusion of Rashba SOI causes larger fluctuations than the corresponding case where Rashba SOI is absent.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.3\textwidth]{fig11a.eps}\\\includegraphics[width=0.3\textwidth]{fig11b.eps}
\caption{(Color online) The fluctuations in the charge conductance, $\Delta G$ is plotted as a function of the Fermi energy, $E$ for (a) $\lambda_R = 0$ and (b) $\lambda_R=0.2$. Three different adatom concentrations are taken, namely, $n_{ad} = 0.025$, 0.05 and 0.1. }
\label{4t_dcond}
\end{center}
\end{figure}
Same as that of the spin polarized conductance, the spin Hall conductance has also three components, namely $G_{SH}^x$, $G_{SH}^y$ and $G_{SH}^z$. Since the $x$ and $y$-component of the spin Hall conductance are absent in the absence of Rashba SOI, we have shown only the variation of $G_{SH}^z$ in units of $e/4\pi$ as a function of the Fermi energy as shown in Fig.\ref{4t_gsz}(a-c)and corresponding fluctuations in Fig.\ref{4t_gsz}(d).
$G_{SH}^z$ is plotted for three different adatom concentrations separately for better clarity. The behaviour of $G_{SH}^z$ is antisymmetric about $E=0$ and the magnitude (irrespective of the sign factor) increases as the adatom density is enhanced. The maximum values of $G_{SH}^z$ are observed on either side of the zero of the Fermi energy and close to $E=0$, the value decreases. The corresponding fluctuations are shown in a single plot in Fig.\ref{4t_gsz}(d). As expected, with increasing the adatom concentration, $\Delta G_{SH}^z$ increases. Another important point to be noted here is that there is a finite region about the zero energy where $G_{SH}^z$ and $\Delta G_{SH}^z$ are strictly zero irrespective of the adatom concentration which is owing to the single channel transmission \cite{chico,liu}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.3\textwidth]{fig12a.eps}\\\includegraphics[width=0.3\textwidth]{fig12b.eps}
\caption{(Color online) (a-c)The four terminal spin polarized conductance, $G^s_z$ is plotted as a function of the Fermi energy, $E$ in the absence of Rashba SOI for three different adatom concentrations. (d) Their corresponding fluctuations are plotted as a function of $E$.}
\label{4t_gsz}
\end{center}
\end{figure}
The effect of the inclusion of Rashba SOI on the three components of the spin Hall conductance is observed in Fig.\ref{4t_gsz_lr}(a-c). All of them are antisymmetric about $E=0$. Their magnitude increases with increasing the adatom concentration. Though the behaviour of the different components of the spin Hall conductance are different in nature as a function of the Fermi energy, their order of magnitude is similar. The fluctuations in the spin Hall conductance are plotted in Fig.\ref{4t_gsz_lr}(d-f) as a function of the Fermi energy. as observed the case in the absence of Rashba SOI, in the present case, $G_{SH}^x$, $G_{SH}^y$ and $G_{SH}^z$ have similar variations with $E$, that is they increase with increasing adatom concentration. further, their orders of magnitude are also same.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.3\textwidth]{fig13a.eps}\\\includegraphics[width=0.3\textwidth]{fig13b.eps}
\caption{(Color online) In presence of Rashba SOI, the three components of the spin polarized conductance (a) $G^s_x$, (b) $G^s_y$ and $G^s_z$ are plotted as a function of the Fermi energy, $E$. (d-f) Their fluctuations are plotted as a function of $E$. }
\label{4t_gsz_lr}
\end{center}
\end{figure}
The DOS plots in the absence of Rashba SOI are shown in Fig.\ref{4t_dos}(a-c) and in the presence of Rashba SOI in Fig.\ref{4t_dos}(d-f). UDOS and DDOS are antisymmetric to each other as we have observed in case of two terminal case. the magnitude of UDOS and DDOS are higher at higher values of $E$ as shown in Fig.\ref{4t_dos}(a) and Fig.\ref{4t_dos}(d) and in both the cases, that is, in presence and absence of Rashba SOI, their magnitudes seem to be independent of the strength of Rashba SOI. The difference between UDOS and DDOS, DiffDOS (= UDOS - DDOS) is antisymmetric in nature as a function of the Fermi energy and has somewhat larger oscillatory behaviour in presence of RSOI than that in the absence of RSOI. Also the total DOS is symmetric about $E=0$. Basically all the three quantities are in good agreement with Eq.\ref{dos} both in the absence and presence of Rashba SOI.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.3\textwidth]{fig14a.eps}\\\includegraphics[width=0.3\textwidth]{fig14b.eps}
\caption{(Color online) The four terminal density of states (DOS) for the spin up and spin down electrons are plotted as a function of the Fermi energy (a) for $\lambda_R = 0$ and (d) for $\lambda_R = 0.2$. The difference between DOS$\uparrow$ and DOS$\downarrow$ are plotted as a function of $E$ (b) for $\lambda_R = 0$ and (e) for $\lambda_R = 0.2$. The behaviour of total DOS is plotted as a function of the Fermi energy (c) for $\lambda_R = 0$ and (e) for $\lambda_R = 0.2$. for clarity we consider only a single adatom concentration, $n_{ad} = 0.1$.}
\label{4t_dos}
\end{center}
\end{figure}
Finally, the local charge current and the $z$-component of the local spin current are shown in Fig.\ref{4t_current}(a) and Fig.\ref{4t_current}(a) respectively in the absence of Rashba SOI. For the local charge current, it is observed that as though the transverse leads are voltage probes, the charge current is flowing between lead 1 and lead 3. In another word, the charges are trying to accumulate at the transverse edges of the sample and as a result less charge current flows between terminals 1 and 2. This in turn reduces the charge conductance in case of a 4T device than the 2T case. The $z$-component of the local spin current is clearly flowing from terminals 1 to 3 and between terminal 1 to 4. As a result the non-zero $G_{SH}^z$ occurs due to presence of the exchange field.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.23\textwidth]{fig15a.eps}\includegraphics[width=0.23\textwidth]{fig15b.eps}
\caption{(Color online) (a) For the four terminal case, the local charge current, $J_0$ and (b) the $z$-component of the local spin current are shown in the absence of Rashba SOI. Both the figures are obtained for the adatom concentration $n_{ad} = 0.1$ and we set the Fermi energy at $E = -0.09$.}
\label{4t_current}
\end{center}
\end{figure}
The local charge and spin currents are shown in Fig.\ref{4t_current_lr} in presence of Rashba SOI. The local charge current, $J_0$ flows between the left and right leads (terminals 1 and 2 respectively) as shown in Fig.\ref{4t_current_lr}(a). We observe that $J_0$ is almost independent of the Rashba SOI. This explains why the charge conductance is independent of the Rashba SOI as shown in Fig.\ref{4t_cond}(a) and Fig.\ref{4t_cond}(b).
Fig.\ref{4t_current_lr}(b), (c) and (d) show respectively local currents, namely $J_x$, $J_y$ and $J_z$. We observe that the magnitude of the $z$-component of the spin Hall conductance (see Fig.\ref{4t_gsz}(a-c)) is one order magnitude lower than any one of the components of SHC in presence of Rashba SOI, which can be explained as following. From Fig.\ref{4t_current_lr}(b-c), the number of paths between terminal 1 and either terminals 3 or 4 are more than the absence of Rashba SOI case as shown in Fig.\ref{4t_current}(b). As a result, the amount of spin current will be less in the absence of Rashba SOI.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.2\textwidth]{fig16a.eps}\quad\includegraphics[width=0.2\textwidth]{fig16b.eps}
\includegraphics[width=0.2\textwidth]{fig16c.eps}\quad\includegraphics[width=0.2\textwidth]{fig16d.eps}
\caption{(Color online) (a) The local charge current $J_0$, (b) the $x$-component of the local spin current, $J_x$, (c) the $y$-component of the local spin current, $J_y$ and (d) the $z$-component of the local spin current, $J_z$ are shown in the presence of Rashba SOI with strength $\lambda_R = 0.2$. The strength of the exchange field is $\lambda_{EB} = 0.18$. We set the Fermi energy at $E = -0.09$ and the adatom concentration in the present case is $n_{ad} = 0.1$.}
\label{4t_current_lr}
\end{center}
\end{figure}
\section{\label{sec6}Conclusion}
In the present work we have studied the behaviour of the charge and spin transport properties in graphene nanoribbon with magnetic adsorbates both in case of two terminal and a four terminal GNR. Specifically for the two terminal case, we study the charge and spin polarized conductance and for the four terminal case the spin Hall conductance. In all the cases we found that the charge conductance is symmetric about the zero of the Fermi energy, while the spin polarized conductance (for two terminal case) and spin Hall conductance (for four terminal case) are antisymmetric about the zero of the Fermi energy.The fluctuations of the charge and spin conductances show systematic behaviour, that is they increase with increasing adatom concentration. We also study the DOS behaviour and their behaviour are alike in both two and four terminal cases. The local charge current is found to be independent of the strength of Rashba SOI, while the three different components of the local spin currents are sensitive to Rashba SOI that is generated by the magnetic adatoms. Further the $z$-component of the spin polarized conductance for 2T GNR is larger by approximately a factor of 5 compared to its 4T counterpart, while the other two components are nearly same for the 2T and 4T devices. Of course a 4T device permits observations of a spin Hall conductance, which is absent in the case of a 2T GNR. Moreover the conductance properties of a 4T setup has lesser fluctuations.
The increase in the components of the spin polarized conductances in magnetic decorated GNRs with Rashba SOI signal a larger spin current flowing in the sample and hence must have greater utility as possible spintronic devices.
\setcounter{secnumdepth}{0}
\section{ACKNOWLEDGMENTS}
SG gratefully acknowledges a research fellowship from MHRD, Govt. of India.
SB thanks SERB, India for financial support under the grant F. No: EMR/2015/001039.
|
\section{Introduction}
\label{section:introduction}
Ever since Anfinsen's pioneering experiment involving the denaturation and
renaturation of bovine ribonuclease \cite{anfinsen},
it has long been held that the native state of a protein is that conformation
which has minimum free energy (MFE) among all possible conformations.
For that reason, both the
Monte Carlo algorithm \cite{metropolis:MonteCarlo} and the
Gillespie algorithm \cite{gillespieStochasticSimulation1} have
been used in biopolymer folding studies to estimate
{\em kinetic folding time}, even when {\em detailed balance} does not
hold.\footnote{Definitions of Monte Carlo
algorithm, Gillespie algorithm, detailed balance, mean first passage time,
master equation, etc. are given later in the paper.}
Since a number of different measures
of folding time have been described for RNA secondary structure formation
\cite{Schmitz.jmb96,flamm,Wolfinger:04a,Senter.po12}, the goal of the
current paper is to clarify the relations between these measures, to
point out the possibly low correlation between mean first passage time
(MFPT) when computed by the Monte Carlo algorithm versus
Gillespie algorithm, and to prove an asymptotic result that
precisely relates Monte Carlo trajectory time with Gillespie trajectory time.
We hope that this clarification will be useful to those users who
perform computational experiments to estimate RNA folding kinetics.
\subsubsection*{Protein folding kinetics using Monte Carlo and/or Gillespie}
For protein to be biologically useful, its
amino acid sequence must satisfy two requirements: (1)
a {\em thermodynamic} requirement that the sequence have a unique,
thermodynamically stable native structure, and (2) a {\em kinetic}
requirement that the denatured polypeptide chain be capable of reaching
the native state within reasonable time (milliseconds to seconds) under
appropriate solution conditions.
In \cite{Sali.n94,karplus:JMolecBiol94}
a simple Monte Carlo folding experiment was designed for a 27 bead
hetero-polymer on a cubic lattice using hydrophobic contact potentials.
Defining a sequence to fold rapidly provided its MFPT is sufficiently small,
the authors concluded that a necessary and sufficient condition
for a sequence to fold rapidly in their model is the existence of
a pronounced energy minimum with substantial gap between the energy of
the native state and that of the most stable misfolded structure.
A succinct summary of the theoretical work in
\cite{Sali.n94,karplus:JMolecBiol94}
is that the thermodynamic requirement entails the kinetic requirement.
In \cite{Abkevich.pnas96,Abkevich.psb97} Monte Carlo hetero-polymer
folding experiments were performed to show how prebiotic proteins might
have evolved to fold rapidly in a primordial soup in order to survive
hydrolysis. Hetero-polymer sequences were {\em evolved} by applying
a Metropolis criterion to the acceptance of pointwise mutations, depending
on whether the mutation increased MFPT. One of the conclusions from this
work was that the kinetic requirement entails the thermodynamic requirement.
More recently, the notion of {\em Markov state model} (MSM) has been introduced
to analyze protein folding kinetics by statistical analysis of a number of
molecular dynamics (MD) trajectories -- see for instance
\cite{Swope,Bowman.m09,Huang.psb10,Weber.jctc11,Bowman.mmb14,Harrigan.bj17}.
To construct a Markov state model, microstates are first formed by applying
a clustering algorithm to conformations taken at fixed time intervals
(shapshots) from MD trajectories by using a measure of
structural similarity such as RMSD. States of the Markov state model
are then defined to be kinetically reachable macrostates constructed from
microstates, where transition probabilities are defined by counting the
relative number of transitions between macrostates. The kinetics of folding
are then analyzed by computing MFPT to reach the macrostate of the MSM
which contains the native state. Provided that the number of states
is reasonably small, the MFPT can be computed analytically from the
{\em fundamental matrix} as done in \cite{Weber.jctc11}. Analytic computations
of MFPT are much more efficient and accurate than repeated simulations
of the Monte Carlo algorithm -- however, both methods are theoretically
equivalent. In contrast, in some papers such as \cite{Zhang.pnas02},
instead of computing the MFPT of the Markov state model, transition rates
are computed from which the {\em master equation} is solved to determine
the {\em equilibrium time}, i.e. time necessary for
molecular {\em equilibrium}. Provided that the number of states is
reasonably small, it is possible to solve the master equation analytically
by spectral decomposition; otherwise, the Gillespie algorithm
can be used to stochastically estimate the equilibrium time.
In \cite{Levy.ps13}, a formula (involving the integral of an appropriate time
correlation function) is given to estimate the
equilibrium time among unfolded states of a protein having
two state kinetics.
\subsubsection*{RNA folding kinetics using Monte Carlo and/or Gillespie}
In \cite{Schmitz.jmb96}, the {\em folding time} for an RNA sequence is
defined to be the sum of reciprocals of the rate constants
$k = k_0 \cdot \exp(-\Delta E^{\ddag}/RT)$ in each step of a folding
trajectory, where $k_0$ is a rate calibration constant and $\Delta E^{\ddag}$
is the activation energy of adding or removing a base pair to the current
structure, using the Turner free energy model \cite{Turner.nar10}.
The resultant folding time differs from MFPT by additionally accounting
for the transition rates between trajectory steps. A similar notion of
folding time is adopted in \cite{Geis.jmb08}, where additionally
basins of attraction are estimated by assigning a collection of
sampled secondary structures to a locally optimal structure by a greedy
procedure.
In \cite{Senter.po12}, mean first passage time was estimated by
the average number of Monte Carlo steps, taken over 50 runs, to fold the
empty secondary structure into the MFE structure. The authors determined
that for the selenocysteine (SECIS-1) family RF00031 from
the Rfam database \cite{Nawrocki.nar15},
the Pearson correlation coefficient is $0.48436$
(p-value $1.8 \cdot 10^{-4}$) between the logarithm of MFPT for a given RNA
sequence and the standard deviation of the number of base pairs taken
over the ensemble of all secondary structures of that sequence.
After publication of \cite{Senter.po12}, computational experiments
performed with the Gillespie algorithm in place of the Monte Carlo algorithm
by using {\tt Kinfold} \cite{flamm} suggested that there is no
such correlation between log MFPT and standard deviation of the number of
base pairs of secondary structures in the ensemble (data not shown).
This observation raises the question of which method to use for RNA
folding kinetics.
In \cite{flamm} the program {\tt Kinfold} is described, now part of
Vienna RNA Package \cite{Lorenz.amb11}, which implements
the Gillespie algorithm for RNA secondary structure formation for the
Turner energy model \cite{Turner.nar10}. In that paper,
{\em folding time} is defined to be the {\em first passage time} from
the empty structure to a given target structure, so that the MFPT is
expected folding time. In \cite{Aviram.amb12}, an increase in
speed of {\tt Kinfold} is reported, by modifying the source code to
incorporate memoization. In \cite{Dykeman.nar15}, the program
{\tt KFOLD} is described, which also implements the Gillespie algorithm,
uses the Turner energy model, and performs
the same elementary step base pair addition/removal transitions
as does {\tt Kinfold}; however, {\tt KFOLD} achieves a
remarkable speed-up over {\tt Kinfold} by exploiting the fact that many of
the base-pair addition/deletion moves and their corresponding
rates do not change between each step in the simulation. As in
\cite{flamm}, folding time is defined in \cite{Dykeman.nar15} to be the
{\em first passage time} from the empty structure to a given target structure.
Instead of determining the mean first passage time for RNA secondary
structure formation by using the Monte Carlo algorithm or the
Gillespie algorithm, an alternative approach is to compute
the {\em equilibrium time} by solving the master equation or by
simulating the Gillespie algorithm until equilibrium has been achieved.
In \cite{Wolfinger:04a} the method {\tt Treekin} is described, where
macrostates are defined by {\em basins of attraction}, and
population occupancy curves are computed over time for each macrostate.
A similar approach is described in \cite{Senter.jmb15}, where the
authors use a different method of constructing macrostates.
\subsubsection*{Relation between MFPT for Markov chain versus process}
In this paper, we consider {\em mean first passage time} (MFPT) of
random walks in a finite, discrete-time {\em Markov chain} $\mathbb{M}_1$,
with transition probabilities given by
equation~(\ref{eqn:transitionProb1}) in the next section.
We also consider MFPT of trajectories in the finite, continuous-time
{\em Markov process} $\mathbb{M}_2$,
with transition rates given by equation~(\ref{eqn:transitionProb2}) in the
next section. Mean first passage
time for the Markov chain $\mathbb{M}_1$ is computed by repeated simulations of
the Metropolis Monte Carlo algorithm, while
that for the Markov process $\mathbb{M}_2$ is computed by repeated
simulations of the Gillespie algorithm.
Although distinct, both methods to compute MFPT have been described
in the literature -- see summary in the previous two sections.
There appears to be a tacit and sometimes explicit assumption
that the kinetics results are essentially equivalent
\cite{amatoRecombRNA,Tang.jmb08}.
Indeed, mean first passage times for protein and/or RNA folding have been
computed by matrix inversion and/or Monte Carlo simulations over a Markov
chain \cite{Klemm08funnelsin,Tang.jmb08,Bowman.pnas10,Huang.psb10,Weber.jctc11,Levy.ps13}, while the master equation and/or Gillespie simulations have
been used in \cite{flamm,Wolfinger:04a,Aviram.amb12,Dykeman.nar15,Xu.pnas16}.
Of particular interest is the construction of Markov state models from
molecular dynamics folding trajectories for a protein or RNA molecule,
followed by either mean first passage time computed by matrix inversion
\cite{Weber.jctc11} versus equilibrium time computed by solving the
master equation \cite{Xu.pnas16}.
This leads to the question: what is the relation between
mean first passage time computed by the Monte Carlo algorithm versus
the Gillespie algorithm?
The answer to this question is trivial in the case that
the Markov chain is $N$-regular, i.e. that each state has degree $N$
($N$ neighbors).
In this case, we show
the MFPT obtained by the time-driven Monte Carlo algorithm is
equal to $N$ times that of the Gillespie algorithm.
In the context of RNA secondary structures of a
given sequence RNA, this
suggests that MFPT determined by the time-driven Monte Carlo algorithm
might be approximately equal to $\langle N \rangle$ times the MFPT
determined by the Gillespie algorithm, where $\langle N \rangle$
denotes the Boltzmann expected node {\em degree}
(expected number of structural neighbors), formally defined by
\begin{eqnarray}
\label{eqn:networkDegree}
\langle N \rangle &=& \sum_{s} N(s) \cdot P(s) =
\sum_{s} N(s) \cdot \frac{\exp(-E(s)/RT)}{Z}
\end{eqnarray}
where the sum is taken over all secondary structures $s$ of the RNA
sequence, $N(s)$ is the number of neighbors of $s$, i.e. structures that
can be obtained from $s$ by the addition or removal of a single base pair,
$E(s)$ is the free energy of structure $s$ using the Turner energy model
\cite{Turner.nar10}, $R$ the universal gas constant, $T$ absolute
temperature, and $Z = \sum_{s} \exp(-E(s)/RT)$ denotes the partition function.
However, we show that this is not the case,
and can only conclude that the Monte Carlo and Gillespie algorithms
return loosely correlated mean first passage times for RNA secondary
structure folding. On the other hand,
we prove that asymptotically, the {\em expected time} for a $K$-transition
Monte Carlo trajectory is equal to that for a $K$-transition
Gillespie trajectory multiplied by $\langle N \rangle$.
A rapidly emerging area of synthetic biology concerns the
computational design of synthetic RNA \cite{Choi.z14,Dotu.nar15}
by using inverse folding software
\cite{Taneda.aabc11,Zadeh.jcc11,GarciaMartin13,EsmailiTaheri.bb15}.
It seems clear that the next step in synthetic RNA design
will be to control the kinetics of RNA folding. Due to the differences
between MFPT computations with the Monte Carlo and Gillespie algorithms
shown in this paper, we suggest that the Gillespie algorithm and
related population occupancy curves determined by solution of the
{\em master equation}
constitute a more appropriate approach to macromolecular folding
kinetics \cite{Xu.pnas16}, rather than the Monte Carlo algorithm and
related computation of mean first passage time by matrix inversion
\cite{Bowman.pnas10}. In the context of synthetic RNA design, we advocate
the computation of MFPT using matrix inversion for a coarse-grained model
\cite{Senter.jcb15}, as an efficient, initial screen of
potential candidate, followed by the slower but more accurate coarse-grain
method {\tt Treekin} \cite{flamm,wolfingerStadler:kinetics}, followed by
repeated simulations using {\tt KFOLD} \cite{Dykeman.nar15}.
\subsubsection*{Plan of the paper}
In Section~\ref{section:definitions},
we provide some background on RNA secondary structure,
Markov chains and Markov processes.
Section~\ref{section:HTandET} describes three versions of the Monte Carlo
algorithm (discrete-time time-driven, discrete-time event-driven and
continuous-time event-driven) and the Gillespie algorithm, and illustates
these algorithms with a simple example. Readers familiar with Markov chains,
Markov processes, Monte Carlo algorithm and the Gillespie algorithm should
read the pseudocode of Algorithms 1-4 (given in Figures 1-4) and otherwise
skip Sections~\ref{section:definitions},\ref{section:HTandET} and proceed
directly to Section~\ref{section:results}. That section
presents a proof of a new theorem that Monte Carlo trajectory
time asymptotically equals Gillespie trajectory time multiplied
by the expected degree. RNA secondary structure folding simulations using
the Monte Carlo algorithm and the Gillespie algorithm are given, which
provide a computational illustration of the theorem. Despite the close
asymptotic relation between Monte Carlo and Gillespie trajectory time,
Section~\ref{section:results} shows that there is no such relation
between Monte Carlo and Gillespie mean first passage time for RNA
secondary structure folding, but only a loose correlation.
Section~\ref{section:discussion} presents some discussion and
concluding remarks. The Appendix presents computations that show
that detailed balance does not hold for the Markov chain of secondary
structures of an RNA sequence.
\section{Background}
\label{section:definitions}
In this section, we describe some definitions concerning RNA secondary
structures, expected degree of a secondary structure network,
Markov chains and Markov processes. We use the notation
$\mathbb{R}$ for the set of real numbers,
$\mathbb{R}^{\geq 0}$ for the set of non-negative real numbers,
and $\mathbb{N}$ for the set of natureal numbers $0,1,2,\ldots$.
\subsection{Background on RNA secondary structure}
A secondary structure for an RNA nucleotide sequence
$\aseq = a_1,\dots,a_n$ is a set $s$ of Watson-Crick or wobble
base pairs $(i,j)$, containing neither base triples nor pseudoknots.
More formally we have the following definition.
\begin{definition}
\label{def:secStr}
A secondary structure for a given RNA nucleotide sequence
$a_1,\dots,a_n$ is a set $s$ of base pairs $(i,j)$, where $1 \leq i < j \leq n$,
such that:
\begin{enumerate}
\item
if $(i,j)\in s$ then
$a_i,a_j$ form either a Watson-Crick (AU,UA,CG,GC) or
wobble (GU,UG) base pair,
\item
if $(i,j)\in s$ then $j-i>\theta=3$ (a steric constraint
requiring that there be at least $\theta=3$ unpaired bases between
any two positions that are paired),
\item
if $(i,j)\in s$ then for all $i' \ne i$ and $j' \ne j$,
$(i',j) \not\in s$ and $(i,j') \not\in s$ (nonexistence of base triples),
\item
if $(i,j)\in s$ and $(k,\ell)\in s$, then
it is not the case that $i<k<j<\ell$ (nonexistence of pseudoknots).
\end{enumerate}
\end{definition}
The Turner energy model \cite{Turner.nar10} is an additive energy
model, where enthalpies, entropies and free energies
for stacked base pairs, hairpins, bulges, internal loops, etc.
are derived by least-squares fitting of experimental data
from UV-absorbance (so-called {\em optical melting}) experiments,
following the pioneering work of Tinoco \cite{tinoco:reviewInBook}.
In this model, there is no energy contribution for a base pair; in contrast,
stacked base pairs contribute negative, stabilizing free energy, while
loop regions contribute positive, destabilizing free energy due to entropic
loss. For instance, from melting experiments at $37^{\circ}$ C,
Turner's rules assign stacking free energy of $–2.08 \pm 0.06$ and
enthalpy of $–10.48 \pm 1.24$ to the stacked pair
$\begin{array}{ll}
\mbox{$5'$-{\tt CU}-$3'$}\\
\mbox{$3'$-{\tt GA}-$5'$}\\
\end{array}$
The minimum free energy (MFE) secondary structure can be computed in time
that is cubic with respect to the length of an input RNA sequence; this is
done by {\em dynamic programming} following the Zuker algorithm
\cite{zukerStiegler}, as implemented in publicly available software, such
as the Vienna RNA Package \cite{Lorenz.amb11}.
Secondary structures can be depicted in several equivalent manners.
For instance, the sequence and dot-bracket representation for
a type III hammerhead ribozyme
from Peach Latent Mosaic Viroid (PLMVd) AJ005312.1/282-335
is given by
\begin{quote}
\bf
\begin{verbatim}
GAUGAGUCUGUGCUAAGCACACUGAUGAGUCUAUGAAAUGAGACGAAACUCAUA
.((((((.(((((...))))).......((((........))))...)))))).
\end{verbatim}
\end{quote}
The left panel of Figure~\ref{fig:Vienna185andVienna207} displays the
equivalent usual presentation of the same structure, computed by {\tt RNAfold}
from the Vienna RNA Package \cite{Lorenz.amb11} using the Turner 1999
energy parameters. This minimum free energy (MFE) structure agrees with the
consensus structure from the Rfam database \cite{Gardner.nar11}. The right
panel of Figure~\ref{fig:Vienna185andVienna207} depicts the network of all
secondary structures for the 11-nt RNA sequence {\tt GCGCGCGCGCG}, containing
27 nodes and 42 edges, where nodes are secondary structures in dot-bracket
notation, and edges are indicated between each two structures $s,t$ in which
$t$ is obtained by adding or removing a single base pair to/from $s$.
The move set $MS_1$ [resp. $MS_2$] consists of adding or removing
[resp. adding, removing or shifting \cite{flamm,Clote.po15}] a base pair
(in this paper, we mostly concentrate on the $MS_1$ move set).
{\em Mean first passage time} (MFPT) for $\mathcal{M}_1$
[resp. $\mathcal{M}_2$] is the average first passage time from
a designated initial state $x_0$ to a designated final state $x_{\infty}$,
which can be approximated by averaging over a number of runs of the
Monte Carlo algorithm [resp. Gillespie algorithm]. In the context of
RNA secondary structure folding kinetics, the initial structure
$x_0$ is often taken to be the empty structure and the target structure
$x_{\infty}$ is often taken to be the MFE structure -- see
Figure~\ref{fig:Vienna185andVienna207}) where the empty structure is
indicated by a green circle, and the MFE structure by an orange circle.
The contribution of
this paper is to show that asymptotic trajectory times for Monte Carlo and
Gillespie algorithms are related, but that mean first passage times appear
to be only loosely correlated.
\subsection{Background on Markov chains and processes}
Suppose that $X=\{1,\ldots,n\}$
is a finite set of states, and for each state $x$ there
is an associated energy $E(x)$. Suppose that $G$ is a graph with vertex
set $X$ and an arbitrary, but fixed edge set. $G$ could be a complete
graph with edges between any two
distinct states, or $G$ could be the network of secondary structures
for a given RNA sequence,
whose vertices are the exponentially \cite{zukerSankoff}
many secondary structures
of the sequence, and whose edges are between any two structures $x,y$ where
$y$ is obtained from $x$ by removing or adding a single base pair -- i.e.
$y$ is obtained by a move from $MS_1$ from $x$, or equivalently,
$x,y$ have base pair distance \cite{moulton} of $1$. For state
$x \in X$, let $N_x$ ambiguously denote either the number of immediate
neighbors of $x$, or the set of immediate neighbors of $x$ in graph
$G$; i.e. those $y \in X$ such that there is an edge from $x$ to $y$
belonging to $G$. Let
$\mathbb{M}_1$ be a discrete Markov chain with set $X$ of states, having
transition probability matrix ${\bf P} = (p_{x,y})$ given by
\begin{eqnarray}
\label{eqn:transitionProb1}
p_{x,y} = \left\{
\begin{array}{ll}
\frac{1}{N_x} \cdot \min\left(1,\exp(-\frac{E(y)-E(x)}{RT})\right)
&\mbox{if $y \in N_x$}\\
1 - \sum_{u \in N_x} p_{x,u} &\mbox{if $x=y$}\\
0 &\mbox{if $y \not\in N_x$, $y \ne x$.}
\end{array} \right.
\end{eqnarray}
In the context of biomolecular folding (especially in {\em Markov state models}
\cite{Swope,Huang.psb10,Pande.m10,Bowman.pnas10,Levy.ps13}),
the transition matrix is used to describe protein and RNA folding kinetics.
Let $\mathbbp(t)$ be the time-dependent {\em population occupancy} row
vector, where $\mathbbp_i(t)$ is the probability that the molecule is
in state $i$ at (discrete) time $t \in \mathbb{N}$ --
here, state $i$ represents a particular
molecular conformation in a discretized space of all possible conformations.
Then
\begin{eqnarray}
\label{eqn:kineticsMarkovChain}
\mathbbp(t) &=& \mathbbp(0) \cdot\mathbbP^t
\end{eqnarray}
Let $\mathbb{M}_2$ be a continuous Markov process with set $X$ of states,
having rate matrix ${\bf Q} = (q_{x,y})$ given by
\begin{eqnarray}
\label{eqn:transitionProb2}
q_{x,y} = \left\{
\begin{array}{ll}
\min\left(1,\exp(-\frac{E(y)-E(x)}{RT})\right)
&\mbox{if $y \in N_x$}\\
- \sum_{u \in N_x} q_{x,u} &\mbox{if $x=y$}\\
0 &\mbox{if $y \not\in N_x$, $y \ne x$}.
\end{array} \right.
\end{eqnarray}
In the context of biomolecular folding, the rate matrix is usually
used in the context of the {\em master equation}, also called the
{\em Kolmogorov forward equation}, defined as follows. As before,
if $\mathbbp(t)$ denotes the time-dependent {\em population occupancy} row
vector, where $\mathbbp_i(t)$ is the probability that the molecule is
in state $i$ at (continuous) time $t \in \mathbb{R}$, then the master equation
is the following system of ordinary differential equations
\begin{eqnarray*}
\frac{d \mathbbp_1(t)}{dt} &=&\sum_{j \ne 1} \left( q_{j,1} \mathbbp_j(t) -
q_{1,j} \mathbbp_1(t)
\right)\\
\frac{d \mathbbp_2(t)}{dt} &=&\sum_{j \ne 2} \left( q_{j,2} \mathbbp_j(t) -
q_{2,j} \mathbbp_2(t)
\right)\\
\cdots & &\\
\frac{d \mathbbp_n(t)}{dt} &=&\sum_{j \ne n} \left( q_{j,n} \mathbbp_j(t) -
q_{n,j} \mathbbp_n(t)
\right)\\
\end{eqnarray*}
Since $q_{i,i} = - \sum_{j \ne i} q_{i,j}$, this set of
equations is equivalent to the matrix differential equation
\begin{eqnarray}
\label{eqn:populationOccupancyEquation}
\frac{d \mathbbp(t)}{dt} &=& \mathbbp(t) \cdot \mathbbQ
\end{eqnarray}
Let $\mathcal{M_J}$ denote the {\em embedded Markov chain}, also called
{\em jump chain}, corresponding to the Markov process $\mathbb{M}_2$,
having state space $X$ and transition probability matrix $S$ given by
\begin{eqnarray}
\label{eqn:jumpChain0}
s_{x,y} = \left\{
\begin{array}{ll}
\frac{q_{x,y}}{\sum_{z \ne x} q_{x,z}}
&\mbox{if $y \in N_x$}\\
0 &\mbox{if $x=y$ or $y \not\in N_x$}.
\end{array} \right.
\end{eqnarray}
The {\em stationary} probability distribution
${\bf p}^* = (p^*_1,\ldots,p^*_n)$ [resp.
${\bf q}^* = (q^*_1,\ldots,q^*_n)$]
for a Markov chain [resp. Markov process]
with transition probability matrix $P=(p_{x,y})$ [resp. rate matrix
$Q=(q_{x,y})$] must satisfy
${\bf p}^* {\bf P} = {\bf p}^*$ [resp. ${\bf q}^* {\bf Q} = {\bf 0}$].
It is well-known that
the stationary distribution exists and is unique for any finite, irreducible,
aperiodic Markov chain \cite{cloteBackofen}
Detailed balance is said to hold, if for all states $x,y \in X$,
$p^*_x p_{x,y} = p^*_y p_{y,x}$ [resp. $q^*_x q_{x,y} = q^*_y q_{y,x}$].
Define the Boltzmann probabilities
\begin{eqnarray}
\label{eqn:stationaryBoltzmann}
\pi_{x} = \frac{\exp(-E(x)/RT)}{Z}
\end{eqnarray}
where the partition function $Z$ is defined by
$Z = \sum_{k=1}^{n} \exp(-E(k)/RT)$.
It is trivial to verify that detailed balance holds for the rate
matrix $Q$ defined in equation~(\ref{eqn:transitionProb2}). From detailed
balance, it is easy to verify that the Boltzmann distribution is the
stationary distribution for the Markov process defined by
equation~(\ref{eqn:transitionProb2}); indeed, if $x,y$ are neighbors
and $E(y)>E(x)$, then $q_{x,y} = \exp(-(E(y)-E(x))/RT)$ and
$\pi_x q_{x,y} = \exp(-E(y)/RT) = \pi_y q_{y,x}$. However, as shown in
the Appendix, the Markov chain $\mathbb{M}_1$ does not generally
satisfy detailed balance for the network of RNA secondary structures.
Using Hasting's trick \cite{hastingsTrick}, where in equation
(\ref{eqn:transitionProb1}) $p_{x,y}$ is redefined by
$p_{x,y} = \frac{1}{N_x} \cdot \min\left(1,\exp(-\frac{E(y)-E(x)}{RT}\cdot
\frac{N_x}{N_y})\right)$ for $y \in N_x$, we obtain a Markov chain satisfying
detailed balance. Moreover the stationary distribution for the Hastings
Markov chain is easily shown to be the Boltzmann distribution from
equation~(\ref{eqn:stationaryBoltzmann}). Nevertheless, this modification
comes at a computational cost, since $N_y$ must be determined for each
neighbor $y$ of $x$. It is perhaps for this reason that Hasting's trick is
not used in some kinetics simulations from the literature, as
in the Monte Carlo time-driven algorithm for protein
folding \cite{Sali.n94} and in Markov state models defined from
long molecular dynamics (MD) trajectories \cite{Bowman.pnas10}.
\section{Monte Carlo and Gillespie algorithms}
\label{section:HTandET}
In this section, we explore the relation between MFPT for a discrete time
Markov chain with transition probability matrix (\ref{eqn:transitionProb1})
and that of the continuous time Markov process with
rate matrix (\ref{eqn:transitionProb2}). Consider
Algorithms~\ref{algo:1}, \ref{algo:2}, \ref{algo:3} and \ref{algo:4}.
The MFPT of the Markov chain $\mathbb{M}_1$ can be
approximated by taking the average over many
runs of Algorithm~\ref{algo:1}, as done to show that proteins fold quickly exactly
when there is a large energy gap between the energy of the native state and
that of the next lowest energy misfolded state \cite{Sali.n94}.
Time-driven Monte Carlo Algorithm \ref{algo:1}
is equivalent to event-driven Monte
Carlo Algorithm \ref{algo:2}, since the probability of leaving state $x$ at any given
time is geometrically distributed with success probability $\Phi_1(x)$, the
probability of leaving state $x$.
By replacing the geometric distribution by the continuous exponential
distribution with the same mean, Monte Carlo Algorithms 2 and 3 are equivalent
(see Figure~\ref{fig:waitingTimeHistogramGeomertricAndExponentialDistribution}
for a comparison of relative histograms of samples from geometric
and exponential distributions having the same mean).
For a given state $x$, the {\em probability} $\Phi_1(x)$ of leaving state $x$ is
equal to the reciprocal of the number of neighbors $N_x$ of $x$ times
the {\em flux} $\Phi_2(x)$ of leaving $x$.
Nevertheless the {\em jump probability} $s_{x,y}$ is identical in Algorithms
2,3,4 since $p_{x,y}/\Phi_1(x)$ is equal to $q_{x,y}/\Phi_2(x)$. It follows
that Algorithms 1,2,3,4 have identical probabilities of visiting exactly the
same states in a trajectory, the only difference being that the expected waiting
time in any state $x$ differs by the factor $N_x$ of the number of neighbors
of $x$.
The previous discussion shows that the only difference between the
MFPT of a discrete Markov chain and that of the related continuous time Markov
process lies in the fact that the time increment $\Delta t$ in the former
is sampled from a geometric distribution with mean
$N_x/\sum_{y \in N_x} \min\left(1,\exp(-\frac{E(y)-E(x)}{RT})\right)$
while the time increment $\Delta t$ in the latter
is sampled from a exponential distribution with mean
$1/\sum_{y \in N_x} \cdot$ $\min\left(1,\exp(-\frac{E(y)-E(x)}{RT})\right)$.
Hence, if the value $N_x$ equals a fixed
constant $N$ for each state $x$, then the Markov chain MFPT equals
$N$ times the Markov process MFPT, as shown in the following example.
\medskip
\noindent
\begin{example}
\label{example1}
Let $X = \{ 0,1,\ldots,2n-1 \}$ be the set of $2n$ states,
where the energy of state $k$ is given by
\begin{eqnarray}
\label{eqn:energyOfStatesToyMarkovChain}
E(k) = \left\{ \begin{array}{ll}
-3 &\mbox{if $k=0$}\\
-2 &\mbox{if $k>0$ and $k$ is even}\\
-1 &\mbox{if $k$ is odd.}\\
\end{array} \right.
\end{eqnarray}
The left panel of
Figure~\ref{fig:comparisonFourMCalgorithms} depicts the $6$-node
network $X$, where $n=3$.
Let $Adj(i,j)$ be the indicator function for whether $i,j$ are adjacent in
a circle of size $2n$; i.e.
\begin{eqnarray}
\label{eqn:adjacencyOfCircle}
Adj(i,j) &=& I[ j \equiv (i \pm 1) \bmod 2n].
\end{eqnarray}
Since the underlying graph is $2$-regular, $N_x=2$ for each state $x$, and
the transition probability matrix of $\mathbb{M}_1$ is defined by
\begin{eqnarray}
\label{fig:transitionProbMatrixToyMarkovChain}
p_{i,j} = \left\{ \begin{array}{ll}
0.5 \min(1,\exp(-(E(j)-E(i)))) &\mbox{if $Adj(i,j)$}\\
1-p_{i,i+1 \bmod 2n}-p_{i,i-1 \bmod 2n} &\mbox{if $i=j$}\\
0 &\mbox{otherwise}\\
\end{array} \right.
\end{eqnarray}
while the rate matrix of $\mathbb{M}_2$ is defined by
\begin{eqnarray}
\label{fig:rateMatrixToyMarkovChain}
q_{i,j} = \left\{ \begin{array}{ll}
\min(1,\exp(-(E(j)-E(i)))) &\mbox{if $Adj(i,j)$}\\
-p_{i,i+1 \bmod 2n}-p_{i,i-1 \bmod 2n} &\mbox{if $i=j$}\\
0 &\mbox{otherwise}\\
\end{array} \right.
\end{eqnarray}
\end{example}
\subsubsection*{Relation between $\mathcal{M}_1$ and $\mathcal{M}_2$}
Given that the definition of the
transition probability matrix ${\bf P} = (p_{x,y})$
of the Markov chain $\mathcal{M}_1$, as given in
equation~(\ref{eqn:transitionProb1}), is similar to that of the
rate matrix ${\bf Q} = (q_{x,y})$ of the Markov process $\mathcal{M}_2$,
as given in equation~(\ref{eqn:transitionProb2}), it is natural to
ask what the relation is between ${\bf P}$ and ${\bf Q}$. The answer is
that there is no relation. For the Markov chain $\mathcal{M}_1$,
equation~(\ref{eqn:kineticsMarkovChain}) states that the population
occupancy vector at time $t=1$ satisfies
$\mathbbp(1) = \mathbbp(0) \cdot\mathbbP$.
In contrast, for the Markov process $\mathcal{M}_2$,
equation~(\ref{eqn:populationOccupancyEquation}) states that the population
occupancy vector at time $t=1$ satisfies
$\mathbbp(1) = \mathbbp(0) \cdot \exp(\mathbbQ)$,
since the solution of the master equation
(\ref{eqn:populationOccupancyEquation}) is
\begin{eqnarray}
\label{eqn:solutionMasterEquation}
\mathbbp(t) &=& \mathbbp(0) \cdot \exp(\mathbbQ \cdot t)
\end{eqnarray}
For instance, if we set $n=2$ in Example~\ref{example1}, then the transition
probability matrix is
\begin{align*}
\mathbbP &= \left( \begin{array}{llllllll}
0.8647& &0.0677& &0.0000& &0.0677\\
0.5000& &0.0000& &0.5000& &0.0000\\
0.0000& &0.1839& &0.6321& &0.1839\\
0.5000& &0.0000& &0.5000& &0.0000\\
\end{array} \right)
\end{align*}
while the rate matrix is
\begin{align*}
\mathbbQ &= \left( \begin{array}{llllllll}
-0.2707& &0.1353& &0.0000& &0.1353\\
1.0000& &-2.0000& &1.0000& &0.0000\\
0.0000& &0.3679& &-0.7358& &0.3679\\
1.0000& &0.0000& &1.0000& &-2.0000\\
\end{array} \right)
\end{align*}
and the exponential of the rate matrix is
\begin{align*}
\exp(\mathbbQ) &= \left( \begin{array}{llllllll}
0.8299& &0.0566& &0.0569& &0.0566\\
0.4183& &0.1983& &0.3205& &0.0629\\
0.1547& &0.1179& &0.6095& &0.1179\\
0.4183& &0.0629& &0.3205& &0.1983\\
\end{array} \right)
\end{align*}
Clearly, the value $\mathbbp(1) = \mathbbp(0) \cdot\mathbbP$ obtained by
equation~(\ref{eqn:kineticsMarkovChain}) is unequal to the value
$\mathbbp(1) = \mathbbp(0) \cdot \exp(\mathbbQ)$ obtained by
equation~(\ref{eqn:populationOccupancyEquation}).
\subsubsection*{Mean first passage time for Example~\ref{example1}}
Using the 20 state Markov chain obtained in Example~\ref{example1} by
setting $n=10$,
the time to reach the minimum energy state $0$ was computed for the
(1) time-driven Monte Carlo Algorithm~\ref{algo:1},
(2) event-driven Monte Carlo Algorithm~\ref{algo:2}
with geometrically distributed waiting times,
(3) event-driven Monte Carlo Algorithm~\ref{algo:3}
with exponentially distributed waiting times,
(4) Gillespie's Algorithm~\ref{algo:4}. Times for Algorithms
\ref{algo:1}, \ref{algo:2} and \ref{algo:3}
were divided by $2$, since
the number $N_x$ of neighbors of each state $x$ is $2$, while times for
Algorithm~\ref{algo:4} were reported as computed.
The average time was taken over 10,000 separate
runs of each algorithm, and then histograms were produced for 1000 repetitions
of each of the 10,000 runs. It follows by the central limit theorem that
each histogram is approximately normal; moreover, the mean and standard
deviation for each histogram is reported as follows:
(1) $\mu=92.94$, $\sigma=0.76$,
(2) $\mu=92.97$, $\sigma=0.75$,
(3) $\mu=92.96$, $\sigma=0.75$,
(4) $\mu=92.92$, $\sigma=0.78$.
The right panel of Figure~\ref{fig:comparisonFourMCalgorithms}
displays superimposed histograms of for the expected number
of moves obtained on average for Algorithms 1,2,3,4
for a state set $X$ with $20$ states $0,1,\ldots,19$.
\section{Results}
\label{section:results}
\subsection{Expected trajectory time}
Throughout this section, we assume that $\mathbb{M}_1$ [resp. $\mathbb{M}_2$] is
a finite, irreducible, aperiodic Markov chain [resp. Markov process] whose
transition probabilities $p_{x,y}$ [resp. transition rates $q_{x,y}$]
satisfy equation~(\ref{eqn:transitionProb1})
[resp. equation~(\ref{eqn:transitionProb2})]. Moreover, we assume that
$\mathbb{M}_1$ and $\mathbb{M}_2$ have the same underlying set
$X=\{1,\ldots,n\}$ of states, and that each state $x \in X$
is labeled by the same energy $E(x)$ in both $\mathbb{M}_1$ and $\mathbb{M}_2$.
The main result of this section is that asymptotically, for large values of $K$,
the expected time taken by a $K$-transition
trajectory\footnote{Each step of Algorithms
\ref{algo:2}, \ref{algo:3} and \ref{algo:4} involves a transition from
a current state to a distinct state; however, each step of the time-driven
Monte Carlo Algorithm~\ref{algo:1} does not necessarily involve a transition
to a new state,
especially if the current state $x$ has low energy, so that $p_{x,x}$ may be
large. For this reason, we use the term $K$-transition trajectory.}
for each of the Monte Carlo
Algorithms \ref{algo:1}, \ref{algo:2}, \ref{algo:3} is equal to $\mu$
multiplied by the expected time for a $K$-transition trajectory of the
Gillespie Algorithm~\ref{algo:4}, where $\mu$ denotes the
expected degree as defined in equation~(\ref{eqn:networkDegree}); i.e. the
expected number of neighbors, given by
\[
\sum_{x \in X} \frac{\exp(-E(x)/RT)}{Z} \cdot N_x
\]
Let Algorithm 1$^{\dag}$, 2$^{\dag}$, 3$^{\dag}$, 4$^{\dag}$ respectively
denote Algorithm 1,2,3,4 where line 3 of each algorithm is replaced by
the line ``{\bf \tt while {\sc true}}''; i.e. there is no condition on the
while loop, so the algorithms do not terminate.
In Theorem~\ref{thm1},
we establish that for any given RNA sequence, as the number
$K$ of trajectory steps approaches infinity, the trajectory time of
Algorithm 3$^{\dag}$ equals that of Algorithm 4$^{\dag}$ multiplied by
the expected network degree $\mu$. In the previous section, we showed
that Algorithms 1,2,3 are equivalent; it follows that
Algorithms 1$^{\dag}$, 2$^{\dag}$, 3$^{\dag}$ are also equivalent. Thus
Theorem~\ref{thm1} directly relates (asymptotic)
{\em trajectory time} of Monte Carlo time-driven and event-driven
algorithms with that of Gillespie's algorithm.
In order to give a formal statement of this result,
we need to provide some definitions.
\begin{definition}[$K$-step trajectory and expected trajectory time]
\label{def:KstepTrajectoryAndExpectedTrajectoryTime}
A $K$-step trajectory (sometimes called $K$-transition trajectory)
of Monte Carlo Algorithm 3$^{\dag}$ [resp. Gillespie Algorithm 4$^{\dag}$]
is a sequence $x_0,x_1,\ldots,x_K$ of states in $\{1,\ldots,n\}$,
where $x_i \ne x_{i+1}$ for $i=0,\ldots,K$, although
$x_i = x_j$ can occur for $|j-i| \geq 2$. Let
$\tau^K_3$ [resp. $\tau^K_4$] denote the random variable whose value is
the sum of the time increments $\Delta t$ in line 11 of
Algorithm 3$^{\dag}$ [resp. Algorithm 4$^{\dag}$].
Denote the {\em expected time} for $K$ steps of Algorithm 3$^{\dag}$
[resp. Algorithm 4$^{\dag}$] by $E[\tau^K_3]$ or
$\langle \tau^K_3 \rangle$ [resp. $E[\tau^K_4]$ or
$\langle \tau^K_4 \rangle$],
where the expectation is taken over all stochastically
sampled $K$-step trajectories (lines 13-18) and sampled time increments
(line 11) of Algorithm 3$^{\dag}$ [resp. Algorithm 4$^{\dag}$].
\end{definition}
\noindent
\begin{theorem}
\label{thm1}
Let $\mu = \sum_{x \in X} \frac{\exp(-E(x)/RT)}{Z} \cdot N_x$
denote the Boltzmann expected network degree, where
$N_x$ denotes the degree of state $x$. Then
$\lim_{K \rightarrow \infty}
\frac{E[\tau^K_1]}{E[\tau^K_2]} = \mu$.
\end{theorem}
\medskip
\noindent
{\bf \sc Proof:} For each state $x \in X$, define the
Boltzmann probability $\pi_{x}$ is defined by
$\pi_{x} = \frac{\exp(-E(x)/RT)}{Z}$, where
the partition function $Z$ is defined by
$\sum\limits_{x \in X} \exp(-E(x)/RT)$.
For state $x \in X$, define the {\em probability} of leaving $x$
[resp. {\em flux} out of $x$], denoted $\Phi_1(x)$ [resp. $\Phi_2(x)$] by
\begin{eqnarray}
\label{eqn:rateFlux}
\Phi_1(x) = \sum_{y \ne x} p_{x,y} = \frac{\Phi_2(x)}{N_x}\\
\label{eqn:probabilityFlux}
\Phi_2(x) = \sum_{y \ne x} q_{x,y} = N_x \cdot \Phi_1(x)
\end{eqnarray}
Recall that the transition probability matrix $S$
of the {\em embedded chain}, also called {\em jump chain},
$\mathbb{M}_J$ is defined in equation~(\ref{eqn:jumpChain0}), and
hence satisfies
\begin{eqnarray}
\label{eqn:jumpChain}
s_{x,y} = \left\{
\begin{array}{ll}
\frac{q_{x,y}}{\Phi_2(x)}
= \frac{p_{x,y}}{\Phi_1(x)}
&\mbox{if $y \in N_x$}\\
0 &\mbox{if $x=y$ or $y \not\in N_x$}.
\end{array} \right.
\end{eqnarray}
A $K$-step trajectory of either Algorithm~3$^{\dag}$ or 4$^{\dag}$ corresponds
to a random walk on the jump chain with states $1,\ldots,n$ and transition
matrix $S$ as defined in equation~(\ref{eqn:jumpChain}).
Since the rate matrix
$Q$, defined in equation~(\ref{eqn:transitionProb2}),
for the Markov process $\mathbb{M}_2$ corresponding to Algorithm~4$^{\dag}$
trivially satisfies detailed balance, so does the transition matrix $S$
for the jump chain, defined in equation~(\ref{eqn:jumpChain}).
\medskip
\noindent
{\sc Claim:} Define the row vector
$s^* = (s^*_1,\ldots,s^*_n)$ by
\begin{eqnarray}
\label{eqn:stationaryJumpChain}
s^*_x &=&
\frac{q^*_x \Phi_2(x)}{\sum_{y \in \mathbb{SS}({\bf a})} q^*_y \Phi_2(y)}.
\end{eqnarray}
Then $s^* S = s^*$, hence $s^*$ is the (unique) stationary distribution for
the jump chain $\mathbb{M}_J$.
\smallskip
\noindent
{\sc Proof:}
For fixed state $y \in \{1,\ldots,n\}$, the $y$th coordinate of the row vector
$s^* S$ satisfies
\begin{eqnarray*}
(s^* S)_y &=&
\sum_{x=1}^n s^*_x \cdot s_{x,y} = \sum_{x \ne y} s^*_x \cdot s_{x,y} \\
&=&
\sum_{x \ne y}
\left( \frac{q^*_x \Phi_2(x)}{\sum_{z \in \mathbb{SS}({\bf a})} q^*_z \Phi_2(z)}
\right) \cdot \frac{q_{x,y}}{\Phi_2(x)}\\
&=&
\sum_{x \ne y}
\frac{q^*_x q_{x,y} }{\sum_{z \in \mathbb{SS}({\bf a})} q^*_z \Phi_2(z)}
=
\sum_{x \ne y}
\frac{q^*_y q_{y,x} }{\sum_{z \in \mathbb{SS}({\bf a})} q^*_z \Phi_2(z)} \\
&=&
q^*_y \cdot \left( \sum_{x \ne y} q_{y,x} \right)
\frac{1}{\sum_{z \in \mathbb{SS}({\bf a})} q^*_z \Phi_2(z)} \\
&=&
\frac{q^*_y \Phi_2(y)}{\sum_{z \in \mathbb{SS}({\bf a})} q^*_z \Phi_2(z)}
=
s^*_y
\end{eqnarray*}
Note that line 2 follows by definition of $s^*_x$ and $s_{x,y}$;
line 3 follows by detailed balance of $q^*_x q_{x,y} = q^*_y q_{y,x}$;
line 4 follows by factoring out terms that do not depend on $x$, and
line 5 follows by the definition of $\Phi_2(y)$.
An equivalent, more intuitive statement of
equation~(\ref{eqn:stationaryJumpChain}) is that $s^*_y$ equals the
the Boltzmann probability $q^*_y$ of state $y$, times the flux
$\Phi_2(y)$ out of $y$, divided by the {\em expected flux} with respect to
the Boltzmann distribution; i.e. $s^*_y$ is the ratio of the Boltzmann
weighted flux out of $y$ with respect to the expected flux out of any
state, where expectation is taken over all states. This proves
the claim.
If the number $K$ of steps in the trajectory is large, then the expected
number of occurrences of each state $x$ is $s^*_x K$. The expected time to
leave state $x$ in Algorithm~3$^{\dag}$ [resp. Algorithm~4$^{\dag}$] is
$\frac{1}{\Phi_1(x)}$ [resp. $\frac{1}{\Phi_2(x)}$], since waiting times
are exponentially distributed. It follows that the ratio of
the expected time for a $K$-step trajectory of Algorithm~3$^{\dag}$ divided by
the expected time for a $K$-step trajectory of Algorithm~4$^{\dag}$ equals:
\begin{eqnarray*}
\frac{E[\tau^K_1]}{E[\tau^K_2]} &=&
\frac{\sum_{x=1}^n \frac{s^*_x K}{\Phi_1(x)}}
{\sum_{x=1}^n \frac{s^*_x K}{\Phi_2(x)}} =
\frac{\sum_{x=1}^n \frac{s^*_x N_x}{\Phi_2(x)}}
{\sum_{x=1}^n \frac{s^*_x}{\Phi_2(x)}}\\
&=&
\frac{ \sum_{x=1}^n \frac{q^*_x \Phi_2(x)}{\sum_z q^*_z \Phi_2(z)} \cdot
\frac{N_x}{\Phi_2(x)}}
{ \sum_{x=1}^n \frac{q^*_x \Phi_2(x)}{\sum_z q^*_z \Phi_2(z)} \cdot
\frac{1}{\Phi_2(x)}} \\
&=&
\frac{ \sum_{x=1}^n q^*_x N_x }
{ \sum_{x=1}^n q^*_x } =
\frac{ \sum_{x=1}^n \pi_x N_x }
{ 1 } \\
&=&
\sum_{x=1}^n \frac{\exp(-E(x)/RT) \cdot N_x}{Z} = \mu({\bf a}).
\end{eqnarray*}
This completes the proof of Theorem~\ref{thm1}. $\blacksquare$
\medskip
\noindent
\begin{corollary}
\label{cor1}
The same proof shows that mean recurrence time for
Markov chain $\mathbb{M}_1$ is equal to $\mu$ times the mean recurrence time
for Markov process $\mathbb{M}_2$. (See \cite{feller1} for the definition
of mean recurrence time.)
\end{corollary}
\subsection{Illustrative examples from RNA}
For a fixed RNA sequence ${\bf a}=a_1,\ldots,a_m$, denote the
set of all secondary structures for ${\bf a}$ by $\mathbb{SS}({\bf a})$.
Define the Markov chain $\mathbb{M}_1$ [resp. Markov process $\mathbb{M}_2$]
for the set $\mathbb{SS}({\bf a})$ of states, where transitions
$s \rightarrow t$ occur between secondary structures $s,t$ that differ by base
pair distance of $1$. $\mathbb{M}_1$ [resp. $\mathbb{M}_2$] has
probability matrix $P=(p_{x,y})$ [resp. rate matrix $Q=(q_{x,y})$]
defined by equation~(\ref{eqn:transitionProb1})
[resp. equation (\ref{eqn:transitionProb2})], where $E(x)$ denotes the
free energy of secondary structure $x$ with respect to the Turner energy
model \cite{Turner.nar10}. Theorem~\ref{thm1}
then implies that for trajectories of
length $K$, for large $K$, the expected trajectory time
$E[\tau^K_1]$ for each of Algorithm 1$^{\dag}$, Algorithm 2$^{\dag}$, and
Algorithm 3$^{\dag}$ is approximately $\mu$ multiplied by the trajectory time
$E[\tau^K_2]$ for Algorithm 4$^{\dag}$, where $\mu$ is the expected
degree for the network $\mathbb{SS}({\bf a})$ of secondary structures for
RNA sequence ${\bf a} = a_1,\ldots,a_n$.
Although the set $\mathbb{SS}({\bf a})$ is generally of size exponential
in $n$ (see \cite{steinWaterman,Fusy.jmb12}), we recently described a cubic
time dynamic programming algorithm \cite{cloteJCC2015} to
compute the {\em expected network degree} of $\mathbb{S}({\bf a})$
\[
\mu({\bf a}) = \sum\limits_{x \in \mathbb{SS}({\bf a})}
\frac{\exp(-E(x)/RT)}{Z} \cdot N_x
\]
where $N_x$ is the set of secondary structures $y$ having base pair distance
$1$ to $x$.
We illustrate Theorem~\ref{thm1} in the following computational experiments.
For a given RNA sequence ${\bf a}$, let $x_0$ denote the empty secondary
structure containing no base pairs. If algorithms
C${^\dag}$ and D${^\dag}$ begin from the same initial state $x_0$ and
use the same pseudo-random number generator seed, then the algorithms
are said to be {\em synchronized}, or to generate {\em synchronized
trajectories}; if no random seed is set, then the algorithms are said to be
{\em unsynchronized}.
If Algorithms 3${^\dag}$ and 4${^\dag}$ are synchronized, then clearly
each algorithm visits exactly the
same states in the same order.
The following computational experiment is described. Given an RNA sequence,
${\bf a} = a_1,\ldots,a_n$, we determine the minimum free energy (MFE)
structure $s^*$ by Zuker's algorithm \cite{zukerStiegler} as implemented in
Vienna RNA Package \cite{Lorenz.amb11}. For the 32 nt fruA SECIS element
{\tt CCUCGAGGGG AACCCGAAAG GGACCCGAGA GG},
for which the target minimum free energy
structure is reached generally in less than one thousand moves,
we analyzed a trajectory of one million moves. For {\em synchronized} runs of
Algorithm~3$^{\dag}$ [resp. Algorithm~4$^{\dag}$] the trajectory time
was approximately $98.16168 \cdot 10^9$
[resp. $9.82994 \cdot 10^9$]
expected number of neighbors
$\mu({\bf a}) = 10.00146$, and the ratio of Algorithm~3$^{\dag}$
trajectory time divided by $\mu({\bf a})$
is $9.81474 \cdot 10^9$ -- close to the Algorithm~4$^{\dag}$ trajectory
time.
For {\em nonsynchronized} runs of
Algorithm~3$^{\dag}$ [resp. Algorithm~4$^{\dag}$] the trajectory time
was approximately $98.03194 \cdot 10^9$
[reps. $9.78867 \cdot 10^9$],
and the ratio of Algorthm C$^{\dag}$ trajectory time divided by
$\mu({\bf a})$
is $9.80177 \cdot 10^9$ -- very close to Algorithm~4$^{\dag}$ trajectory
time.
Similar validation of Theorem~\ref{thm1} was shown in
second computational experiment, where a
500 thousand step trajectory was
generated for the
76 nt ala-tRNA from {\em Mycoplasma mycoides} with Sprinzl ID RA1180
(tRNAdb ID tdbR00000006) \cite{Juhling.nar09} using Algorithms
C$^{\dag}$ and D$^{\dag}$. A final illustration of Theorem~\ref{thm1}
is given in the right panel of
Figure~\ref{fig:comparisonFourMCalgorithms}, which displays a scatter plot,
for both synchronized and unsynchronized runs of one million steps for the
Monte Carlo Algorithm~3$^{\dag}$ and the Gillespie Algorithm~4$^{\dag}$ for
forty 20 nt randomly generated RNA sequences, each of whose MFE
secondary structure has free energy $E < -2.5$ kcal/mol, each
having at most 2500 secondary structures, and each
with expected compositional frequency of 0.25 for each of A,C,G,U.
Figure~\ref{fig:RNAsuboptNumNbors} shows the distribution for the number of
neighbors
the 76 nt transfer RNA (RA1180 from tRNAdb 2009 \cite{Juhling.nar09}),
with mean $29.29$ and standard deviation $4.14$, as well as that
for the 56 nt spliced leader RNA from {\em Leptomonas collosoma}
(a known conformational switch), with mean
$70.05$ and standard deviation of $33.84$. Although
we can exactly compute the expected network degree in
seconds \cite{cloteJCC2015},
the only current method to obtain an approximation of the distribution is
by sampling using {\tt RNAsubopt} \cite{Wuchty.b99} from the
Vienna RNA Package \cite{Lorenz.amb11}, a computation requiring many hours.
It follows that the RNA examples illustrating Theorem~\ref{thm1}
could not have been
given without use of the dynamic programming algorithm to compute expected
network degree \cite{cloteJCC2015}.
\medskip
\begin{remark}
\label{remark1}
In our examples of RNA secondary structure trajectory
time, we considered elementary step transitions between
secondary structures $x,y$, in which $y$ is obtained from $x$ by addition or
removal of a single base pair. However, Theorem~\ref{thm1}
is equally applicable
for more general move sets between RNA secondary structures, including
{\em shift moves} \cite{flamm} and even the addition and removal of entire
stems \cite{Isambert.pnas00}. In each of these cases, it is easily established
that the underlying Markov chain [resp. Markov process] is finite, irreducible
and aperiodic, so that Theorem~\ref{thm1} applies.
\end{remark}
\begin{remark}
\label{remark2}
The program {\tt Kinfold} \cite{flamm} is an implementation of the Gillespie
algorithm, which supports both move set $MS_1$, consisting of adding or
removing a single base pair, and move set $MS_2$, consisting of adding,
removing, or {\em shifting} a base pair. Since we found it difficult to
modify the source code of {\tt Kinfold} to implement the Monte Carlo algorithm,
we implemented the Monte Carlo and Gillespie algorithms for RNA secondary
structure networks in a C-program {\tt mc.c}, available at
\url{http://bioinformatics.bc.edu/clotelab/RNAexpNumNbors/}.
In \cite{Clote.po15}, we generalized
the algorithm from \cite{cloteJCC2015} to efficiently
compute the expected network
degree for RNA secondary structure networks with respect to move set
$MS_2$. At the present time, however, our program {\tt mc.c} does not
support shift moves.
\end{remark}
\subsection{Monte Carlo and Gillespie MFPT}
Here, we show that unless the network is regular (each node has the
same degree), there is no simple analogue of Theorem~\ref{thm1}
to relate the mean first passage time (MFPT) for the Monte Carlo Algorithm~3
and the Gillespie Algorithm~4 for RNA secondary structure folding kinetics.
Algorithms 3$^{\dag}$ and 4$^{\dag}$ differ
from Algorithms 3 and 4, in that the former algorithms do not terminate
computation upon visiting the minimum free energy structure $x_{\infty}$.
This modification allowed us to establish Theorem~\ref{thm1}.
In Example~\ref{example1} from Section~\ref{section:HTandET},
we described the computation of mean first passage time
(MFPT) using each of Algorithms 1,2,3,4 for a Markov chain $\mathbb{M}_1$
[resp. Markov process $\mathbb{M}_2$], having states
$2n$ states $0,1,\ldots,2n-1$, where energy $E(x)$ of state $x$ is defined
in equation~(\ref{eqn:energyOfStatesToyMarkovChain}), and transition probability
$p_{x,y}$ of moving from state $x$ to state $y \in \{ x-1 \bmod 2n, x+1
\bmod 2n\}$ is defined in
equation~(\ref{fig:transitionProbMatrixToyMarkovChain}).
Since the underlying graph is $2$-regular,
Figure~\ref{fig:comparisonFourMCalgorithms} indeed shows
that MFPT for each of
Monte Carlo Algorithms 1,2,3 is twice the MFPT of Gillespie Algorithm 4.
By slightly modifying the topology of Example~\ref{example1},
we obtain a non-regular network with states $0,\ldots,2n-1$,
for which all states have degree $2$, with the exception of
the (terminal) minimum energy state $0$ and the last state $2n-1$. For
this example, described in the following,
there appears to be no simple relation between the MFPT of Monte
Carlo Algorithms 1,2,3 with that of Gillespie Algorithm 4.
\medskip
\noindent
\begin{example}
\label{example2}
Define Markov chain $\mathbb{M}_1$ [resp. Markov process
$\mathbb{M}_2$] as in Example~\ref{example1},
with the sole exception that state $0$ is
no longer connected to state $2n-1$; i.e. the indicator
function $Adj(i,j)$ for whether $i,j$ are adjacent is redefined by
\begin{eqnarray}
\label{eqn:adjacencyOfCircleLinearized}
Adj(i,j) &=& \left\{ \begin{array}{ll}
1 &\mbox{if $0 \leq i < 2n-1$ and $j=i+1$}\\
1 &\mbox{if $0 \leq j < 2n-1$ and $i=j+1$}\\
0 &\mbox{otherwise}
\end{array} \right.
\end{eqnarray}
Transition probabilities [resp. rates] are defined by
equation~(\ref{eqn:transitionProb1})
[resp. equation~(\ref{eqn:transitionProb2})] where we set $RT=1$, hence
\begin{eqnarray}
\label{fig:transitionProbMatrixToyMarkovChainBis}
p_{i,j} = \left\{ \begin{array}{ll}
0.5 \cdot \min(1,\exp(-(E(j)-E(i))))
&\mbox{if $Adj(i,j), 1 \leq i,j < 2n-1$}\\
\min(1,\exp(-(E(j)-E(i))))
&\mbox{if $i=0,j=1$ or $i=2n-1,j=2n-2$}\\
1-p_{i,(i+1 \bmod 2n)}-p_{i,(i-1 \bmod 2n)} &\mbox{if $i=j$}\\
0 &\mbox{otherwise}\\
\end{array} \right.
\end{eqnarray}
while the rate matrix of $\mathbb{M}_2$ remains unmodified, defined by
\begin{eqnarray}
\label{fig:rateMatrixToyMarkovChainBis}
q_{i,j} = \left\{ \begin{array}{ll}
\min(1,\exp(-(E(j)-E(i)))) &\mbox{if $Adj(i,j)$}\\
-p_{i,(i+1 \bmod 2n)}-p_{i,(i-1 \bmod 2n)} &\mbox{if $i=j$}\\
0 &\mbox{otherwise}\\
\end{array} \right.
\end{eqnarray}
\end{example}
Let $n=10$, so that there are $20$ states $0,\ldots,19$, and let the
initial state $x_0 = 10$, as in Example~\ref{example1}.
The time to reach the minimum energy state $0$ was computed for the
time-driven Monte Carlo Algorithm~1, event-driven Monte Carlo Algorithm~2
with geometrically distributed waiting times, event-driven Monte Carlo
Algorithm~3 with exponentially distributed waiting times and
the Gillespie Algorithm~4. The average time was taken over 10,000 separate
runs of each algorithm, and histograms were produced for 1000 repetitions
of each of the 10,000 runs. The mean and standard
deviation for each histogram is reported as follows:
(1) $\mu = 520.93$, $\sigma = 16.99$,
(2) $\mu = 520.52$, $\sigma=16.97$,
(3) $\mu = 520.67$, $\sigma=17.00$,
(4) $\mu = 265.21$, $\sigma=8.62$.
Figure~\ref{fig:comparisonFourMCalgorithmsBis} displays the relative
histograms for this data, where for comparison purposes, first passage
times for Gillespie Algorithm~4 are multiplied by $2$. The uniform
expected number of neighbors, or network degree, computed over the
the collection of 19 non-terminal states $x \ne 0$, is
$1.947$,
while the Boltzmann expected number of neighbors is
$1.710$.
In contrast, the ratio of the average $520.706$ of the mean first passage
times computed by Monte Carlo Algorithms 1,2,3 divided by the mean first passage
time $265.209$ for the Gillespie Algorithm 4 is equal to $1.963$.
The p-value for the 2-tailed T-test for equality of Monte Carlo
MFTP $520.706$ with Gillespie MFPT $265.21$ times {\em uniform} network degree
$1.947$ is $1.93149 \cdot 10^{-11}$. It follows that Monte Carlo MFPT is
statistically different from Gillespie MFPT times the expected degree of
the network in the case of Example~\ref{example2}
(for either uniform or Boltzmann probability).
\subsection{Folding trajectories for the 10-mer GGGGGCCCCC}
\label{section:foldingTrajectoriesFor10mer}
Theorem~\ref{thm1} implies that IF there is {\em no termination} condition
for Algorithms 3 and 4 (obtained by replacing line 3 by
{\bf while {\sc true}}), so that every secondary structure of the 10-mer
GGGGGCCCCC may be visited multiple times (including the MFE structure),
THEN for a sufficiently large number of algorithm steps $K$, the
total time for the trajectory of Algorithm 3 is (asymptotically) equal
to the total time for the trajectory of Algorithm 4 times the {\em expected
network degree}. We now show that this is \underline{false}
for mean first passage
times; i.e. if Algorithms 3 and 4 terminate upon reaching the MFE
structure, then there is no such relation between their trajectory times,
which correspond to the first passage time from the empty initial structure
to the MFE structure. Before proceeding, we need some notation.
\begin{definition}[Expected convergence time]
\label{def:MFEtrajectoryAndExpectedTrajectoryTime}
Given an RNA sequence ${\bf a} = a_1,\ldots,a_n$,
let $\tau^{\mbox{\tiny MFE}}_3({\bf a})$ [resp.
$\tau^{\mbox{\tiny}{MFE}}_4({\bf a})$]
denote the random variable whose value is
the sum of the time increments $\Delta t$ in line 11 of
Algorithm~3 [resp. Algorithm~4] until the algorithm converges,
where initial state $x_0$ is the empty secondary structure
of {\bf a} and final state $x_{\infty}$ is the MFE structure of ${\bf a}$.
Let $E[\tau^{\mbox{\tiny MFE}}_3({\bf a})]$ or
$\langle \tau^{\mbox{\tiny MFE}}_3({\bf a}) \rangle$
[resp. $E[\tau^{\mbox{\tiny MFE}}_4({\bf a})]$ or
$\langle \tau^{\mbox{\tiny MFE}}_4({\bf a}) \rangle$]
denote the expected convergence time of Algorithm~3
[resp. Algorithn~4], or in other words, the \underline{mean first passage time}
to fold the RNA sequence {\bf a} using Monte Carlo Algorithm~3
[resp. Gillespie Algorithm~4].
\end{definition}
For the 10 nt RNA sequence GGGGGCCCCC, we ran Algorithms 3 and 4 to compute
the trajectory time for 100,000 distinct, {\em synchronized} folding
trajectories, each trajectory starting from the empty initial structure
$x_0 = \emptyset$ and terminating upon reaching
the minimum free energy (MFE) structure
$x_{\infty} = \{ (1,10),(2,9),(3,8) \}$\footnote{Recall that secondary
structures may be represented as a set of base pairs.}
with dot-bracket notation
$\op \op \op \bullet \bullet \bullet \bullet \cp \cp \cp$.
Figure~\ref{fig:relativeHistogramAlgoCandDandCminusDegTimesD}a
[resp.
Figure~\ref{fig:relativeHistogramAlgoCandDandCminusDegTimesD}b]
depicts the relative histogram for the 100,000 trajectory times of
Algorithm~3 [resp. 4] -- note that the first passage times
appear to be exponentially distributed, although the fit is not good.
Since the folding experiments were
synchronized, for each $i=1,\ldots,100,000$, the number of steps taken by
each of Algorithms \ref{algo:3} and \ref{algo:4}
were identical, as were the secondary structures
visited in each step, so that the only difference between
Algorithms \ref{algo:3} and \ref{algo:4}
consisted in the incremental times $\Delta t$ determined in line 11 of
each algorithm, as well as in the total trajectory time in line 20 of each
algorithm.
\subsubsection*{Expected degree with respect to the Boltzmann distribution}
Using the software from \cite{cloteJCC2015}, we find that {\em Boltzmann}
expected number of neighbors $\langle N \rangle$, as defined in
equation~(\ref{eqn:networkDegree}), for the
set of all 62 secondary structures of the 10-mer GGGGGCCCCC is
$\langle N \rangle = 3.031162 \approx 3.03$.
Figure~\ref{fig:relativeHistogramAlgoCandDandCminusDegTimesD}c
depicts the relative histogram for the pairwise differences
$T_c - \langle N \rangle \cdot T_d$.
Let the null hypothesis $H_0$ assert that
trajectory time $T_c$ for Algorithm~3
to reach the MFE structure is equal to $\langle N \rangle$ multiplied by
the trajectory time $T_d$ for Algorithm~4
to reach the MFE structure; i.e. $H_0$ is the assertion that
$T_c = 3.031162 \cdot T_d$.
The p-value for $n=100,000$
paired values $(T_c, \langle N \rangle \cdot T_d)$ is $9.0923E-107$, since
the mean $\overline{x}_{dif}$
of paired differences $T_c- \langle N \rangle \cdot T_d$
satisfies
$\overline{x}_{dif} =
1277.20$ $= \sum_{i=1}^n \frac{T_c[i]-\langle N \rangle \cdot T_d[i]}{n}$,
the standard deviation $s_{dif}$
of the paired differences satisfies $s_{dif}=18379.89$,
the test statistic $t = 21.97 = \frac{\overline{x}_{dif}}{s_{dif}/\sqrt{n}}$,
and the number $df$ of degrees of freedom $n-1=99,999$.
For $\alpha=0.05$, to compute the $1-\alpha = 95\%$ confidence interval for
$T_c-\langle N \rangle \cdot T_d$, we determine the margin of error
$E = t_{\alpha/2} \cdot \frac{s_{dif}}{\sqrt{n}} \approx
1.645 \cdot 58.12 = 95.60$, yielding a confidence interval of
$(\overline{x}_{dif}-E, \overline{x}_{dif}+E) = (1181.59,1372.80)$.
Under the null hypothesis,
$\overline{x}_{dif}$ is asserted to be $0$, which does not belong
to the confidence interval. Even repeating the computation with
$\alpha=10^{-100}$, we still observe that $0$ does not belong to the
corresponding confidence interval $(39.33,2515.06)$.
Since Algorithm~3 first passage times are greater than expected
network degree time Algorithm~4 first passage times, it must be that
a disproportionately large set of the secondary structures visited in
folding trajectories of the 10-mer GGGGGCCCCC have {\em larger} degree
(a larger number of neighboring structures that can be reached by the
addition or removal of a single base pair) than the network average.
This suggests that the (identical) folding trajectories of
Algorithms~\ref{algo:3} and \ref{algo:4}
do {\em not} visit mainly low energy structures, although our detailed
analysis of one 10 nt RNA sequence can hardly be representative.
Note as well that the paired
times $T_c$ and $\langle N \rangle \cdot T_d$ are poorly correlated.
Since first passage times are not normally distributed, we compute
the Spearman correlation of $0.439397$; although Pearson correlation should
not be used for non-normal distributions, the Pearson correlation is
$0.445950$.
Analysis of the same data after removal of outliers (e.g. computing the
10\%-trimmed mean, etc.) even strengthen the conclusions just reached
(data not shown).
It follows that the analogue of Theorem~\ref{thm1} for
trajectories that terminate when reaching the MFE
does not hold -- moreover, the correlation is low for the paired trajectory
times for synchronized Algorithms~\ref{algo:3} and \ref{algo:4}
to reach the MFE structure.
Since tremendous computational resources would be required to perform
a similar analysis for longer RNA sequences, we restrict our attention to
the correlation between the synchronized and non-synchronized first
passage times for
Algorithms~\ref{algo:3} (continuous-time event-driven Monte Carlo)
and \ref{algo:4} (Gillespie)
-- see Tables~\ref{table:evan1000corr} and
\ref{table:DividingByExpNumNborsDoesNotHelp}.
\subsubsection*{Expected degree with respect to the uniform distribution}
Our software \cite{cloteJCC2015} also computes that the {\em uniform}
expected number of neighbors $\langle N \rangle = 3.548387 \approx 3.55$
for the set of all 62 secondary structures of the 10-mer GGGGGCCCCC,
defined by equation~(\ref{eqn:networkDegree}) where the free energy $E(s)$
of every secondary structure is defined to be zero, or equivalently by setting
probability $P(s) = \frac{1}{Z} = \frac{1}{62}$, where $Z$ now denotes the the
total number of secondary structures of GGGGGCCCCC. We now re-analyze the
100,000 first passage times of Algorithms~\ref{algo:3} and \ref{algo:4}.
Recall that the null hypothesis $H_0$ asserts that
trajectory time $T_c$ for Algorithm~3
to reach the MFE structure is equal to $\langle N \rangle$ multiplied by
the trajectory time $T_d$ for Algorithm~4
to reach the MFE structure, except that we now use uniform expected
number of neighbors;
i.e. $H_0$ is the assertion that $T_c = 3.548387 \cdot T_d$.
The p-value for $n=100,000$
paired values $(T_c, \langle N \rangle \cdot T_d)$ is $3.42447E-09$.
For $\alpha=0.05$, to compute the $1-\alpha = 95\%$ confidence interval for
$T_c-\langle N \rangle \cdot T_d$, we determine the 95\%
confidence interval of
$(\overline{x}_{dif}-E, \overline{x}_{dif}+E) = (-478.01,-269.87)$.
The 95\% confidence interval for an independent run of
Algorithms~\ref{algo:3} and \ref{algo:4}
to generate another set of 100,000 first passage times was
$(-389.30, -362.03)$.
It follows that the uniform probability analogue of Theorem~\ref{thm1} for
trajectories that terminate when reaching the MFE
does not hold.
Since Algorithm~3 first passage times are now {\em less} than the
{\em uniform} expected number of neighbors multiplied by
Algorithm~4 first passage times, it must be that a
a disproportionately large set of the secondary structures which are visited in
folding trajectories of the 10-mer GGGGGCCCCC have {\em lower} degree
(a lower number of neighboring structures that can be reached by the
addition or removal of a single base pair) than the network average.
This suggests that the (identical) folding trajectories of
Algorithms~\ref{algo:3} and \ref{algo:4}
do {\em not} visit mainly high energy structures (having few base pairs,
hence potentially a larger number of neighbors), although our detailed
analysis of one 10 nt RNA sequence can hardly be representative.
Finally, we note that the ratio of the mean first passage time
$\langle T_c \rangle$ [resp. $\langle T_d \rangle$],
taken over $n=100,000$ runs of Algorithm~3 [resp. Algorithm~4] is
$\frac{\langle T_c \rangle}{\langle T_d \rangle} \approx 3.43$,
a value that appears somewhat closer to uniform expected network degree
of $3.548387$ than the Boltzmann expected network degree of $3.031162$.
\subsubsection*{Correlation analysis for 20-mers}
In the case of RNA secondary structure folding kinetics,
it appears that Gillespie MFPT times expected network
degree is even less correlated with Monte Carlo MFPT
than unaltered Gillespie MFPT.
Table~\ref{table:evan1000corr} shows the Pearson correlation between the
averages, taken over 1000 runs, of the MFPT for each of 1000 20 nt RNAs,
when computed by the Monte Carlo (MC) method
(Algorithm~3$^{\dag}$), synchronized (syn) [resp. not synchronized (ns)]
with Gillespie's (G) method (Algorithm~4$^{\dag}$). Additionally, we
performed two repetitions of the unsyncronized runs, designated experiment
A and B in the table. For these data, the MFPT computed by
Monte Carlo is highly correlated with that computed by Gillespie:
G (syn) has correlation of 0.85702 with MC (syn) and a correlation of
0.81030 with MC (ns). Surprisingly, G (syn) has a somewhat {\em lower}
orrelation of 0.85381, for MC* (syn), obtained by dividing the
MC (syn) time for each sequence by its expected number of neighbors;
similarly, G (syn) has a somewhat {\em lower} correlation of 0.80559
for MC* (ns), obtained by dividing the MC (ns) time for each sequence
by its expected number of neighbors. In other words, despite the fact that
the ratio of Monte Carlo trajectory time over Gillespie trajectory time
equals the expected number of neighbors for sufficiently long trajectories
as proved in Theorem~\ref{thm1},
there is no such relation between Monte Carlo MFPT
and Gillespie MFPT, presumably since the mean first passage time is
generally reached within $K$ steps, where $K$ is too small for the
asymptotic effects of Theorem~\ref{thm1} to apply.
\begin{table*}
\begin{tabular}{|l|llllll|}
\hline
& {\mbox{\small MC (syn)}} &{\mbox{\small MC (ns A)}} &{\mbox{\small MC (ns B)}} &{\mbox{\small G (syn)}} &{\mbox{\small G (ns A)}} &{\mbox{\small G (ns B)}}\\
\hline
{\mbox{\small MC (syn)}} &1.000000 &0.994405 &0.981810 &0.857019 &0.871860 &0.880024\\
{\mbox{\small MC (ns A)}} &0.994405 &1.000000 &0.989948 &0.810297 &0.830360 &0.839469\\
{\mbox{\small MC (ns B)}} &0.981810 &0.989948 &1.000000 &0.783331 &0.809399 &0.814758\\
{\mbox{\small G (syn)}} &0.857019 &0.810297 &0.783331 &1.000000 &0.996953 &0.996666\\
{\mbox{\small G (ns A)}} &0.871860 &0.830360 &0.809399 &0.996953 &1.000000 &0.998148\\
{\mbox{\small G (ns B)}} &0.880024 &0.839469 &0.814758 &0.996666 &0.998148 &1.000000\\
\hline
\end{tabular}
\caption{Correlation shown by folding time averages,
taken over 1000 runs, for 1000
20 nt random sequences taken from the benchmarking set of
\cite{Senter.jcb15}. Abbreviations are as follows.
MC (sym): synchronized Monte Carlo Algorithm~3$^{\dag}$;
MC (ns A): batch A for nonsynchronized Monte Carlo Algorithm~3$^{\dag}$;
MC (ns B): batch B for nonsynchronized Monte Carlo Algorithm~3$^{\dag}$;
G (sym): synchronized Gillespie Algorithm~4$^{\dag}$;
G (ns A): batch A for nonsynchronized Gillespie Algorithm~4$^{\dag}$;
G (ns B): batch B for nonsynchronized Gillespie Algorithm~4$^{\dag}$.
}
\label{table:evan1000corr}
\end{table*}
\begin{table*}
\begin{tabular}{|l|llllll|}
\hline
& {\mbox{\small MC(syn)}} &{\mbox{\small MC(ns)}} &{\mbox{\small G(syn)}} &{\mbox{\small G(ns)}} &{\mbox{\small MC*(syn)}} &{\mbox{\small MC*(ns)}}\\
\hline
{\mbox{\small MC(syn)}} &1.000000 &0.994405 &0.857019 &0.871860 &0.985542 &0.978662\\
{\mbox{\small MC(ns)}} &0.994405 &1.000000 &0.810297 &0.830360 &0.981410 &0.985471\\
{\mbox{\small G(syn)}} &0.857019 &0.810297 &1.000000 &0.996953 &0.853810 &0.805594\\
{\mbox{\small G(ns)}} &0.871860 &0.830360 &0.996953 &1.000000 &0.871133 &0.827705\\
{\mbox{\small MC*(syn)}} &0.985542 &0.981410 &0.853810 &0.871133 &1.000000 &0.994563\\
{\mbox{\small MC*(ns)}} &0.978662 &0.985471 &0.805594 &0.827705 &0.994563 &1.000000\\
\hline
\end{tabular}
\caption{Correlation shown by folding time averages,
taken over 1000 runs, for 1000
20 nt random sequences taken from the benchmarking set of
\cite{Senter.jcb15}. Abbreviations are as in Table~\ref{table:evan1000corr},
with the addition that MC* (syn) [resp. MC* (ns)] is produced by
dividing the Monte Carlo (Algorithm~3$^{\dag}$) folding time by the expected
sequence connectivity for synchronized [resp. nonsynchronized] computations.
}
\label{table:DividingByExpNumNborsDoesNotHelp}
\end{table*}
\section{Discussion}
\label{section:discussion}
In this paper, we have compared Markov chains and related Markov processes
by considering four closely related Algorithms 1,2,3,4.
If the underlying graph for the Markov chain $\mathbb{M}$ is $N$-regular,
so that each state of $\mathbb{M}$ has exactly $N$ neighbors, then
it follows that on average, mean first passage time (MFPT)
computed by each of the Monte Carlo Algorithms 1,2 and 3 equals $N$
multiplied by the MFTP computed by Gillespie Algorithm 4. Although the
Markov chain of RNA secondary structures of a given RNA sequence is
generally not
$N$-regular for any $N$, the total time along a Monte Carlo trajectory
is asymptotically equal to the Boltzmann expected number of neighbors
$\langle N \rangle = \sum_{s} \frac{\exp(-E(s)/RT)}{Z} \cdot N(s)$
multiplied by
the total time along a Gillespie trajectory, as proved in Theorem~\ref{thm1}.
Computational experiments on several RNAs confirm this result,
provided that trajectories are sufficiently long to exhibit
asymptotic properties of Markov chains. Our code {\tt mc.c}
for Algorithms 3 and 4, is written in C and makes calls to the
function {\tt energy\_of\_structure()} from {\tt libRNA.a} of
Vienna RNA Package \cite{Lorenz.amb11}. This program, along with
C programs to compute the expected network degree for RNA secondary
structures with move set $MS_1$ (base pair addition or deletion)
\cite{cloteJCC2015} or with move set $MS_2$ (base pair addition, deletion
or shift) \cite{Clote.po15}, is publicly available at
\url{http://bioinformatics.bc.edu/clotelab/RNAexpNumNbors}.
Since Anfinsen's pioneering experimental result on the folding of bovine
ribonuclease \cite{anfinsen}, it is widely accepted that the
native state of a biomolecule is a free energy minimum.
In the literature on biomolecular folding, there is sometimes
a tacit assumption that kinetics simulations
using the Monte Carlo and Gillespie algorithms yield comparable results.
Indeed, in \cite{amatoRecombRNA} Tang et al. assert that
``We demonstrate with two different RNA that the different analysis methods
(ME, MC, MMC) produce comparable results and can be used
interchangeably.\footnote{ME stands for the {\em Master Equation}, meaning the
computation of the time-dependent population occupancy vector
$\mathbbp(t)$ by solution of the master equation
$\frac{d \mathbbp(t)}{dt} = \mathbbp(t) \cdot Q$,
where $Q$ is the rate matrix. MC stands for Monte Carlo simulation, and
MMC stands for {\em Map-based Monte Carlo} simulation, a method inspired
by probabilistic road map methods from robotics, where a constant $k$
many closest neighboring secondary structures are selected for each
RNA secondary structure from a sampled collection of modest size.
Subsequently Monte Carlo simulation is applied to the resulting network,
which is much smaller than the network of all secondary structures.
The correctness of the authors' assertion is due uniquely to the fact
that the roadmap sampling network is {\em $k$-regular}.}''
The results of this paper suggest that one should not make tacit
assumptions concerning folding kinetics simulations using Monte Carlo and
Gillespie algorithms, when different nodes in the network have different
numbers of neighbors, as in the case for RNA secondary structures.
If computing the mean first passage time requires sufficiently long
trajectories, then we would expect Monte Carlo MFPT to approximately equal
Gillespie MFPT times the expected number of neighbors. However, for a
slight modification of Example~\ref{example1}
described in Example~\ref{example2}, as well as for
a benchmarking set of 1000 20 nt RNAs, each of which has at most 2,500
secondary structures, no such relation was found. Even worse, for the
set of 1000 RNAs, the
correlation between Monte Carlo MFPT and Gillespie MFTP times
the expected number of neighbors was found to be {\em lower}
than the correlation without its consideration. This result suggests
that number of trajectory
steps necessary to reach the minimum free energy
structure may be too small to see the asymptotic relation expected by
Theorem~\ref{thm1}. We conclude that RNA secondary structure folding
may occur {\em faster} than the time scale required for Theorem~\ref{thm1},
a type of {\em Levinthal paradox} \cite{levinthal:1968} which suggests that
{\em folding pathways} could be encoded in RNA sequences, or that
RNA secondary structure formation could follow either a
{\em kinetic funnel model} \cite{Bryngelson.p95} or
a {\em kinetic hub model} \cite{Bowman.pnas10}.
Next, we argue that macromolecular folding kinetics are better captured
by Gillespie's Algorithm 4, than Monte Carlo Algorithms 1,2,3.
For all but pathological
RNA sequences, it is the case that detailed balance
$\pi_x \cdot p_{x,y} = \pi_y \cdot p_{y,x}$ does not hold for
the Boltzmann distribution $\pi$,
as shown by considering $x$ to be the empty structure and $y$ to be a structure
containing exactly one base pair. Since the Boltzmann distribution is
not necessarily the stationary distribution, this argument does not
imply that detailed balance does not hold; however in
Section~\ref{section:foldingTrajectoriesFor10mer}
the (non-Boltzmannian) stationary distribution is computed for a
tiny RNA sequence, for which it is shown that detailed balance does not
hold.
Since the stationary distribution for the Markov chain underlying
each of the Monte Carlo Algorithms 1,2,3 is
not necessarily the Boltzmann distribution, one might instead
consider modified versions of these algorithms defined as follows.
Algorithm 1$^{\ddag}$ is obtained from Algorithm 1$^{\dag}$ by replacing
the expression in line 9 by
\[
\frac{\pi_y}{\pi_x} \cdot \frac{N_x}{N_y}
\]
or equivalently by
\[
\exp\left( \frac{-(E(y)-E(x))}{RT} \right) \cdot \frac{N_x}{N_y}.
\]
Similarly,
Algorithm 2$^{\ddag}$ [resp. 3$^{\ddag}$]
is obtained from Algorithm 2$^{\dag}$ [resp. 3$^{\dag}$]
by replacing the expression in line 9 by
\[
\frac{1}{N_x} \cdot \min\big(1, \frac{\pi_y}{\pi_x} \cdot \frac{N_x}{N_y}
\big)
\]
or equivalently
\[
\frac{1}{N_x} \cdot \min\big(1,
\exp\left( \frac{-(E(y)-E(x))}{RT} \right) \cdot \frac{N_x}{N_y} \big).
\]
This ensures that the new Markov chain underlying each of the Algorithms
1$^{\ddag}$, 2$^{\ddag}$, 3$^{\ddag}$ is Markov chain $\mathbb{M}_H$,
whose transition probability matrix $P_H = (p^H_{x,y})$ is defined by
Hastings' trick, i.e.
\begin{align}
\label{eqn:hastingsTrick}
p^{H}_{x,y} &= \frac{1}{N_x} \min(1, \exp(-(E(y)-E(x))/RT) \cdot N_x/N_y)
\end{align}
However, the embedded matrix (jump matrix) for
$M_H$ would then be different for the Markov process $\mathbb{M}_2$
underlying the Gillespie algorithm. Although we would have ensured that
the stationary distribution for each of the Algorithms
1$^{\ddag}$, 2$^{\ddag}$, 3$^{\ddag}$, and 4 is the Boltzmann distribution,
we would no longer have trajectory times related as in Theorem~\ref{thm1}.
By Theorem~1, either the Monte Carlo algorithm or the Gillespie algorithm
may be used for statistical analysis of trajectories (frequency of visitation
of states, etc.). However in the context of RNA secondary structure folding
time to the minimum free energy structure, the Boltzmann distribution
is the stationary distribution {\em only} for the Gillespie Algorithm 4
and for the Monte Carlo Algorithms 2,3 modified by Hastings'
trick using transition probabilities from equation~(\ref{eqn:hastingsTrick}).
Moreover, the minimum free energy structure is {\em targeted} only if the
Boltzmann distribution is the stationary distribution, where by {\em targeted},
we mean that the MFE structure has the highest probability, and that
$\lim\limits_{n \rightarrow \infty} Pr[X_n=x] = \pi_x$, where $X_n$ is the
random variable for the secondary structure at step $n$ and $\pi_x$ is
the Boltzmann probability of structure $x$. However, there is a
serious computational cost for the Hastings-modified Monte Carlo algorithm;
indeed, determination of the update to state $x$ requires the computation
of free energies of all neighbors of not only current state $x$, but
as well of all neighbors in $N_y$ of all states $y \in N_x$. For repeated
RNA kinetics simulations, this cost could be prohibitive.
We argue that the Gillespie algorithm should be used for RNA secondary
structure folding kinetics, and that specifically one should compute
the population occupancy vector $\mathbbp(t)$ determined by solution of the
matrix differential equation $\frac{d \mathbbp(t)}{dt} = \mathbbp(t)
\cdot Q$, for rate matrix $Q$ as in \cite{Zhang.pnas02,Xu.pnas16},
rather than the mean first passage time by use of the fundamental matrix
or by matrix inversion \cite{meyerMFPT} as done for
Markov state models in \cite{Bowman.pnas10,Huang.psb10}. For
synthetic design of RNA molecules \cite{Zadeh.jcc11,Dotu.nar15},
we advocate fast MFPT computation using coarse-grained models
\cite{wolfingerStadler:kinetics,Senter.jcb15} to select design candidates
for subsequent scrutiny by more accurate methods such as {\tt KFOLD}.
\section{Acknowledgements}
This research was begun during a visit to California Institute of
Technology, Free University of Berlin and the Max Planck Institute
for Molecular Genetics. For discussions, P.C. would like
to thank Frank No{\'e}, Knut Reinert, Martin Vingron
and Marcus Weber (Berlin) and Niles Pierce, Erik Winfree (Pasadena).
The research was funded by a Guggenheim Fellowship, Deutscher Akademischer
Austauschdienst (DAAD), and by National Science Foundation grant DBI-1262439.
Any opinions, findings,
and conclusions or recommendations expressed in this material are
those of the authors and do not necessarily reflect the views of the
National Science Foundation.
\hfill\break \clearpage
\begin{figure*}
\begin{algorithm}[\bf Discrete-time time-driven MC algorithm for MFPT]
\label{algo:1}
\hfill\break
\medskip
\begin{small}
\mverbatim
1. procedure Metropolis(initialState $x_0$, targetState $x_{\infty}$)
2. $x = x_0$; time $t=0$
3. while $x \ne x_{\infty}$ \{
4. choose random $y \in N_x$
5. if $E(y)<E(x)$ //greedy step
6. x = y //update x
7. else //Metropolis step
8. choose random $z \in (0,1)$
9. if $z < \exp\left(-\frac{E(y)-E(x)}{RT}\right)$
10. x = y //update x
11. t = t + 1 //update time regardless of whether x is modified
12. \} //end of while loop
13. return final state $x$, total time $t$
|mendverbatim
\end{small}
\end{algorithm}
\caption{Discrete-time time-driven Monte Carlo (MC) algorithm to
estimate mean first passage time for a given Markov chain, whose
transition probabilities are given by equation~(\ref{eqn:transitionProb1}).
Since the Boltzmann probabilities are defined by $p_x^*=\exp(-E(x)/RT)/Z$,
$p_y^*=\exp(-E(y)/RT)/Z$, it follows that $p_y^*/p_x^* = \exp(-(E(y)-E(x))/RT)$,
so that the probability that state $x$ is modified to $y$ in
lines 5-10 is precisely $p_{x,y} = 1/N_x \cdot \min(1,\exp(-(E(y)-E(x))/RT))$.
}
\label{fig:algo1}
\end{figure*}
\begin{figure*}
\begin{algorithm}[\bf Discrete-time event-driven MC algorithm for MFPT]
\label{algo:2}
\hfill\break
\medskip
\begin{small}
\mverbatim
1. procedure Metropolis(initialState $x_0$, targetState $x_{\infty}$)
2. $x=x_0$; time $t=0$
3. while $x \ne x_{\infty}$ \{
4. $\Phi_1=0$; //probability of leaving current state $x$
5. for all $y \in N_x$ do
6. $p_{x,y} = \frac{1}{N_x} \min\left(1,\exp(-\frac{E(y)-E(x)}{RT})\right)$
7. $\Phi_1 += p_{x,y}$
8. for all $y \in N_x$ do $s_{x,y} = \frac{p_{x,y}}{\Phi_1}$ //jump probabilities
9. sample $\Delta t \thicksim Geom(\frac{1}{\Phi},\sqrt{\frac{1-\Phi}{\Phi^2}})$ where $\Phi=\Phi_1$
10. t += $\Delta t$ //update time
11. choose random $z_2 \in (0,1)$
12. //sample $y\in N_x$, using roulette wheel as in the following
13. sum = 0
14. for $y \in N_x$
15. sum = sum + $s_{x,y}$
16. if $z_2 \leq sum$ then
17. x=y; break //use roulette wheel
18. \} //end of while loop
19. return final state $x$, total time $t$
|mendverbatim
\end{small}
\end{algorithm}
\caption{Discrete-time event-driven Monte Carlo (MC) algorithm to
estimate mean first passage time for a given Markov chain, whose
transition probabilities are given by equation~(\ref{eqn:transitionProb1}).
Since the probability that state $x$ is modified to $y$ in
lines 5-10 of Algorithm~\ref{algo:1} equals
$p_{x,y} = 1/N_x \cdot \min(1,\exp(-(E(y)-E(x))/RT))$, it follows that
the probability that $x$ will be modified is $\sum_{z \in N_x} p_{x,z}$
which equals $1-p_{x,x}$. The probability that $x$ will be modified in
the $k$th iteration of the while loop of Algorithm~\ref{algo:1}
(but not before) is exactly $p_{x,x}^{k-1} \cdot (1-p_{x,x})$; i.e.
equivalent to the probability that the heads is obtained in the
$k$th coin flip (but not before) for a coin whose heads probability is
$(1-p_{x,x})$, given as $\Phi$ in line 7 of Algorithm~\ref{algo:2}.
It follows that the number of steps $\Delta t$ before
the state $x$ is modified is geometrically distributed with parameter
$\Phi=(1-p_{x,x})$; since the mean [resp. standard deviation] for a
geometrically distributed random variable with parameter $\Phi$ is
$\frac{1}{\Phi}$ [resp. $\sqrt{\frac{1-\Phi}{\Phi^2}}$] we denote the
distribution by $Geom(\frac{1}{\Phi},\sqrt{\frac{1-\Phi}{\Phi^2}})$.
Time increments $\Delta t$ are sampled from this geometric distribution,
hence the expected time increment in line 9 of Algorithm~\ref{algo:2}
is identical to the expected time increment before state $x$ is modified
in lines $4-10$ of Algorithm~\ref{algo:1}, so both algorithms are
equivalent.
}
\label{fig:algo2}
\end{figure*}
\begin{figure*}
\begin{algorithm}[\bf Continuous-time event-driven MC algorithm for MFPT]
\label{algo:3}
\hfill\break
\medskip
\begin{small}
\mverbatim
1. procedure hybridMC(initialState $x_0$, targetState $x_{\infty}$)
2. $x=x_0$; time $t=0$
3. while $x \ne x_{\infty}$ \{
4. $\Phi_1=0$ //$\Phi_1$ is probability of leaving $x$
5. for all $y \in N_x$ do
6. $p_{x,y} = \frac{1}{N_x} \min\left(1,\exp(-\frac{E(y)-E(x)}{RT})\right)$
7. $\Phi_1 += p_{x,y}$
8. for all $y \in N_x$ do $s_{x,y} = \frac{p_{x,y}}{\Phi_1}$ //jump probabilities
9. //sample $\Delta t \thicksim Exp(\frac{1}{\Phi},\frac{1}{\Phi})$ where $\Phi=\Phi_1$
10. choose random $z_1 \in (0,1)$
11. $\Delta t$ = $-\frac{1}{\Phi_1} \ln(z_1)$ //sample time increment
12. t += $\Delta t$ //update time
13. choose random $z_2 \in (0,1)$
14. sum = 0
15. for $y \in N_x$
16. sum += $s_{x,y}$
17. if $z_2 \leq sum$ then
18. x=y; break //use roulette wheel
19. \} //end of while loop
20. return final state $x$, total time $t$
|mendverbatim
\end{small}
\end{algorithm}
\caption{Continuous-time event-driven Monte Carlo (MC) algorithm, to
estimate mean first passage time for a given Markov chain, whose
transition probabilities are given by equation~(\ref{eqn:transitionProb1}).
The geometically distibuted time increment $\Delta t$ from line 9 of
Algorithm~\ref{algo:2} is replaced in line 9 of the current algorithm
by the exponentially distributed time increment having the same expectation
$\frac{1}{\Phi}$. As the standard deviatioin of the exponential distribution
with parameter $1/\Phi$ is also $1/\Phi$, we denote this distribution by
$Exp(\frac{1}{\Phi},\frac{1}{\Phi})$ -- this allows the reader to immediately
see that the time increment $\Delta t$ in line 9 of both algorithms has
the same mean and approximately the same standard deviation (especially if
$\Phi$ is small).
}
\label{fig:algo3}
\end{figure*}
\begin{figure*}
\begin{algorithm}[\bf Gillespie's algorithm for MFPT]
\label{algo:4}
\hfill\break
\medskip
\begin{small}
\mverbatim
1. procedure Gillespie(initialState $x_0$, targetState $x_{\infty}$)
2. $x = x_0$; time $t=0$
3. while $x \ne x_{\infty}$ \{
4. $\Phi_2=0$ //$\Phi_2$ is flux (not probability) out of $x$
5. for all $y \in N_x$ do
6. $q_{x,y} = \min\left(1,\exp(-\frac{E(y)-E(x)}{RT})\right)$
7. $\Phi_2 += q_{x,y}$
8. for all $y \in N_x$ do $s_{x,y} = \frac{q_{x,y}}{\Phi_2}$ //jump probabilities
9. //sample $\Delta t \thicksim Exp(\frac{1}{\Phi},\frac{1}{\Phi})$ where $\Phi=\Phi_2$
10. choose random $z_1 \in (0,1)$
11. $\Delta$ t = $-\frac{1}{\Phi_2} \ln(z_1)$ //sample time increment
12. t += $\Delta$ t //update time
13. choose random $z_2 \in (0,1)$
14. sum = 0
15. for $y \in N_x$
16. sum += $s_{x,y}$
17. if $z_2 \leq sum$ then
18. x=y; break //use roulette wheel
19. \} //end of while loop
20. return final state $x$, total time $t$
|mendverbatim
\end{small}
\end{algorithm}
\caption{Continuous-time event-driven Gillespie algorithm, to
estimate mean first passage time for a given Markov {\em process}, whose
transition {\em rates} are given by equation~(\ref{eqn:transitionProb2}).
The only difference between this algorithm and Algorithm~\ref{algo:3}
is that in line 6, the transition rate $q_{x,y}$ is missing the
factor $1/N_x$ present in transition probability $p_{x,y}$, and so
the value $\Phi$ from line 9 is smaller in the current algorithm than
in Algorithm~\ref{algo:3} by the factor $N_x$. It immediately follows that
Algorithms~\ref{algo:3} and \ref{algo:4} are identical in a regular
network where $N_x = N$ for all states $x$; moreover, it is suggestive
(but incorrect for MFPT) that the values $\Phi$ from line 9 of Algorithms
\ref{algo:3} and \ref{algo:4} might be related by the Boltzmann expected
number $\langle N_x \rangle$ of neighbors.
}
\label{fig:algo4}
\end{figure*}
\begin{figure*}
\includegraphics[width=0.35\linewidth]{FIGURES/figure5A}
\hspace{1cm}
\includegraphics[width=0.65\linewidth]{FIGURES/figure5B}
\caption{{\em (Left)} Minimum free energy secondary (MFE) structure of
the 54 nt Peach Latent Mosaic Viroid (PLMVd) AJ005312.1/282-335,
colored by positional entropy -- see \cite{Huynen.jmb97} for the definition
of positional entropy. The MFE structure was computed by {\tt RNAfold} using
Vienna RNA Package 1.8.5, employing the Turner 1999 free energy
parameters. This structure is identical to the consensus structure from
the Rfam database \cite{Gardner.nar11}, while the structure produced by
Vienna RNA Package 2.1.9, employing the Turner 2004 free energy
parameters consists of two external stem-loops. This example shows the
sometimes dramatic difference in
MFE structure computation due solely to slightly different
free energy parameters \cite{Turner.nar10}.
{\em (Right)} Network of secondary structures for the 11-nt RNA sequence
{\tt GGCCGGCCGGC}, where edges are indicated between every two structures
whose base pair distance is $1$ (i.e. move set $MS_1$).
The expected degree for this network having 27 nodes and 42 edges is
$3.111111$ with respect to the uniform distribution
and $3.074169$ with respect to the Boltzmann distribution.
The MFE structure of {\tt GCGCGCGCGCG}, depicted in the orange circle,
has free energy $-2.6$ kcal/mol. Mean first passage time for secondary
structure folding is the average first passage time from the empty initial
structure (green circle) to the MFE structure (orange circle), where
edge probabilities [rates] for the Monte Carlo [resp. Gillespie] algorithm
is given by equation~(\ref{eqn:transitionProb1})
[resp. equation~(\ref{eqn:transitionProb2})].
}
\label{fig:Vienna185andVienna207}
\end{figure*}
\begin{figure*}[!ht]
\begin{center}
\includegraphics[width=.45\linewidth]{FIGURES/figure6A}
\includegraphics[width=.45\linewidth]{FIGURES/figure6B}
\caption{{\em (Left)} Superimposition of histograms for the geometric and
exponential distributions. In each case, using the method described in
the text, 10,000 points were sampled from the geometric (blue squares)
[resp. exponential (black circles)]
distribution having mean 10 and standard deviation $\sqrt{90} \approx
9.4868$ [resp. mean 10]. For the geometric distribution, this corresponds
to the waiting time before obtaining a heads, where the probability of
flipping a heads is $0.1$. For the 10,000 points sampled from the
geometric distribution, the sample mean
is 9.995400 and the sample standard deviation is 9.413309. For the 10,000
points sampled from the exponential distribution, the sample mean is
9.989983 and the sample standard deviation is 9.950308.
{\em (Right)} Superimposition of the probability [resp. probability
density] of the geometric [resp. exponential] distribution having mean
of 10, as computed by Mathematica.
Graphics produced with Mathematica.
}
\label{fig:waitingTimeHistogramGeomertricAndExponentialDistribution}
\end{center}
\end{figure*}
\begin{figure*}[!ht]
\begin{center}
\includegraphics[width=0.45\linewidth]{FIGURES/figure7A}
\includegraphics[width=0.45\linewidth]{FIGURES/figure7B}
\caption{
{\em (Left)} Topology of Markov chain/process from Example~\ref{example1},
consisting of $2n$ states $0,1,\ldots,2n-1$,
each having exactly two neighbors. State $0$ has energy
$-3$, non-zero even [resp. odd] states have energy $-2$ [resp. $-1$].
The figure illustrates the topology for $n=3$,
Example~\ref{example1} in the text considers $n=10$.
{\em (Right)}
Superimposed histograms for Monte Carlo and Gillespie first passage times
for Example~\ref{example1}, taken over 1000 computational experiments,
each experiment consisting of taking the average
{\em first passage time} in 10,000 repetitions of
each of the following algorithms:
(1) time-driven Monte Carlo algorithm (blue squares with filled area),
(2) event-driven Monte Carlo algorithm with geometrically distributed waiting
times (black circles),
(3) event-driven Monte Carlo algorithm with exponentially distributed waiting
times (green squares),
(4) Gillespie's algorithm (red circles). Histograms for a single experiment
for each algorithm, consisting of 10,000 separate runs, resembled
an exponential distribution, of course, but the averages of each experiment
are approximately normally distributed (central limit theorem).
}
\label{fig:comparisonFourMCalgorithms}
\end{center}
\end{figure*}
\begin{figure*}[!ht]
\centering
\includegraphics[width=.45\linewidth]{FIGURES/figure8A}
\includegraphics[width=.45\linewidth]{FIGURES/figure8B}
\caption{Distribution of the Boltzmann-weighted
number of {\em neighbors}, where the number of neighbors of
structure $s$ is the number of structures $t$ obtained from $s$ by
removal or addition of a single base pair.
Relative frequency at position $x$
is defined as the sum, taken over all {\em sampled}
structures $s$ of degree $x$, of the Boltzmann probability
$\exp(-E(s)/RT)/Z^*$,
where $Z^*$ is the sum of Boltzmann factors of all {\em sampled} structures.
This value approximates the true Boltzmann probability
$\exp(-E(s)/RT)/Z$, where the partition function $Z$ is the sum of
Boltzmann factors of {\em all} structures.
{\em (Left)} Distribution of the number of neighbors for transfer RNA
(RA1180 from tRNAdb 2009 \cite{Juhling.nar09}), having expected number
of structures
$\langle N \rangle = 29.29 \pm 4.14$. Using {\tt RNAsubopt -e 12}
\cite{Lorenz.amb11}, all structures were generated, whose free energy is within
12 kcal/mol of the minimum free energy ($110,572$ structures), and the
ratio $Z^*/Z$ of the sum of corresponding Boltzmann factors divided by
the partition function is $0.9990$.
{\em (Right)} Distribution of the number of neighbors
for spliced leader RNA, an RNA conformational switch, from the
trypanosome {\em Leptomonas collosoma}, having expected number of
structures $\langle N \rangle = 70.05 \pm 33.84$.
A total of 57,803 structures were generated using {\tt RNAsubopt -e 10},
and the ratio $Z^*/Z$ of the sum of corresponding Boltzmann factors divided by
the partition function is $0.9992$.
}
\label{fig:RNAsuboptNumNbors}
\end{figure*}
\begin{figure*}[!ht]
\begin{center}
\includegraphics[width=0.45\linewidth]{FIGURES/figure9A}
\includegraphics[width=0.45\linewidth]{FIGURES/figure9B}
\caption{
{\em (Left)}
Superimposed histograms for Monte Carlo and Gillespie first passage times
for Example~\ref{example2}, taken over 1000 computational experiments,
each experiment consisting of taking the
average {\em first passage time} in 10,000 repetitions of
each of the algorithms:
(1) time-driven Monte Carlo algorithm (blue squares with filled area),
(2) event-driven Monte Carlo algorithm with geometrically distributed waiting
times (black circles),
(3) event-driven Monte Carlo algorithm with exponentially distributed waiting
times (green squares),
(4) Gillespie's algorithm (red circles).
The computations were performed exactly as in
Figure~\ref{fig:comparisonFourMCalgorithms} -- the topology of
Example~\ref{example2} differs from that of Example~\ref{example1} only
in that there is no edge between the minimum energy state $0$ and the
last state $2n-1$, i.e. state 19 in Example~\ref{example2}.
{\em (Right)}
Ratio $\frac{\tau^K_1}{\tau^K_2}$ of trajectory time
for $K$ steps ($K=1$ million) in Monte Carlo and Gillespie algorithms,
plotted as a function of
the expected number of neighbors (expected degree) for 40 randomly generated
20 nt RNAs. This figure illustrates Theorem 1, which states that the
ratio $\frac{E(\tau^K_1)}{E(\tau^K_2)}$ of expected trajectory time for
$K$-step trajectories using Monte Carlo Algorithm 3$^{\dag}$ over the
expected trajectory time of $K$-step trajectories using Gillespie's Algorithm
4$^{\dag}$ equals the expected degree of the network of RNA secondary
structures, as $K$ increases to infinity. Both synchronized and unsynchronized
trajectories are plotted.
}
\label{fig:comparisonFourMCalgorithmsBis}
\end{center}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.3\linewidth]{FIGURES/figure10A}
\quad
\includegraphics[width=0.3\linewidth]{FIGURES/figure10B}
\includegraphics[width=0.3\linewidth]{FIGURES/figure10C}
\caption{Relative histograms for first passage times for the continuous-time,
event-driven Monte Carlo algorithm and the Gillespie algorithm.
(a) Probability distribution for Algorithm~3 trajectory time $T_c$ from the
empty initial structure to the MFE structure, with
mean of $10953.60$, standard deviation of $17765.47$,
maximum of $221,692.29$, and minimum of $0.41$.
Red line shows the
exponential distribution with parameter $\lambda = \frac{1}{10953.60} \approx
9.13 \times 10^{-5}$.
(b) Probability distribution for Algorithm~4 trajectory time $T_d$ from the
empty initial structure to the MFE structure, with
mean of $3192.31$, standard deviation of $5654.83$,
maximum of $72,473.93$, and minimum of $0.075$.
Red line shows the
exponential distribution with parameter $\lambda = \frac{1}{3192.31} \approx
3.13 \times 10^{-4}$
(c) Probability distribution for time $T_c - \langle N \rangle \cdot T_d$,
where $\langle N \rangle = 3.031162 \approx 3.03$ is the expected number
of neighbors for the network of all 62 secondary structures for the 10-nt
sequence GGGGGCCCCC. This distribution has
mean of $1277.20$, standard deviation of $18,379.89$,
maximum of $204,747.05$, and minimum of $-199,785.20$.
See text for a proof that with overwhelming statistical significance we have
the strict inequalities
${\langle N \rangle}_b \cdot T_d < T_c < {\langle N \rangle}_u \cdot T_d$,
where ${\langle N \rangle}_b$ [resp. ${\langle N \rangle}_u$] denotes the
Boltzmann [resp. uniform] expected number of neighbors.
}
\label{fig:relativeHistogramAlgoCandDandCminusDegTimesD}
\end{figure*}
\hfill\break \clearpage
\begin{footnotesize}
\bibliographystyle{plain}
|
\section{Introduction}
The richness of many astrophysical processes ranging from supernova remnants and diffuse galactic magnetic fields to the cosmic microwave background (CMB) can be revealed through careful measurement of polarized emission \cite{Tinbergen96}.
In fact, it is often the distribution of polarization over an extended source that contains the physical information of interest. For such observations, a measurement platform must provide a sufficiently stable calibration for the consistent interpretation of the signal over the angular and temporal scales of interest.
In practice, these measurements can require high accuracy, the emission may be only slightly polarized, and instrument effect may contaminate the measurement; these challenges can be addressed through the use of modulation.
In the millimeter and submillimeter wavelength range, polarization modulation can be achieved by placing a device in the optical path capable of varying the incoming polarized light in a known way while leaving the overall total intensity unchanged.
Examples of such devices include half-wave plates (HWP)\cite{Jones88, Platt91, Leach91, Johnson07} and variable-delay polarization modulators (VPMs) \cite{Krejny08, Chuss2012, Harrington2018}.
When combined with polarization sensitive detectors, such an optical system is capable of distinguishing between potentially larger unpolarized continuum flux and the targeted polarized signal.
Polarization modulation has proven critical for far infrared and millimeter polarimetry, which includes the study of magnetic field structure, synchrotron emission, polarized thermal dust emission, and the CMB \cite{Krejny08, Wiesemeyer14, Johnson07, Takakura17, Harrington16, Kusaka14}.
In this work, we present the design for a new reflective half-wave plate (RHWP) modulator of sufficient size to be suitable for use as a front-end polarization modulator for study of the CMB polarization at large angular scales. The design breaks new ground in terms of diameter, flatness, and system control\cite{Williams10, Wiesemeyer14}.
Many design choices, e.g., the wave plate diameter, the modulation frequency, and the targeted wavelength are easily adopted to many different systems.
As a specific goal, the design is optimized to fit in the 90 GHz telescope of the Cosmology Large Angular Scale Surveyor (CLASS) \cite{Harrington2018}.
While the device is of broad general interest, using the 90 GHz CLASS telescope for an initial demonstration has many attractive features.
For example, the telescope is already designed for a reflective front-end modulator such that no additional relay optics are required for integration, the calibration and optical performance of the CLASS telescopes are well understood and therefore commissioning can specifically target the RHWP, and using CLASS will enable a unique and important comparison of systematic and efficiency performance metrics between the CLASS VPM and this new RHWP modulator.
The paper is organized as follows. Section \ref{modulator_model} introduces a simple model describing how the RHWP can be used as a modulator. Section \ref{mechanical} describes the design of the wave plate and its expected mechanical performance.
Section \ref{drive} describes how the modulator motion is controlled and gives preliminary performance from laboratory measurements.
Finally we provide conclusions in Section \ref{conclusion}.
\section{Operation of a reflective wave plate modulator}\label{modulator_model}
As a motivation for this instrument, a simplified model that demonstrates the concept of how the RHWP operates is presented.
A more complete description can be found in the literature\cite{Wiesemeyer14}, and a more complete electromagnetic model will be included in future work.
To describe the optical operation of the RHWP, we use Mueller calculus for the description of polarized light \cite{Shurcliff66}.
The objective of a half-wave plate is to effectively introduce half a wavelength of path length difference between orthogonal linear polarization states.
Regardless of the physical system used to implement such a transformation, the effect in the Stokes basis $(I, Q, U, V)$ can be represented via the Mueller matrix
\[
M_{WP} = \left(\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & -1 & 0 \\
0 & 0 & 0 & -1
\end{array}\right),
\]
where the chosen Stokes basis is oriented such that the wave plate ``fast'' axis is aligned with that of the $Q$ basis. In this orientation, the half-wave delay is introduced between the linear polarization states that define $Q$ such that the sign of $U$ is flipped; thus, the $-1$ in the $UU$ element of $M_{WP}$. Rotation of the wave plate in a fixed lab coordinate system can then be expressed as
\begin{equation}
\label{mueller_wp}
M_{WP}(\theta) = R(-\theta) \cdot M_{WP} \cdot R(\theta) =
\left(\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & \cos 4 \theta & \sin 4 \theta & 0 \\
0 & \sin 4 \theta & -\cos 4 \theta & 0 \\
0 & 0 & 0 & -1
\end{array}\right),
\end{equation}
where $R(\theta)$ is the Stokes basis rotation matrix:
\[
R(\theta) = \left(\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & \cos (2 \theta) & \sin (2 \theta) & 0 \\
0 & -\sin (2 \theta) & \cos (2 \theta) & 0 \\
0 & 0 & 0 & 1
\end{array}\right).
\]
When followed by a detector sensitive to linear polarization, the polarization signal is then modulated at \emph{four} times the rotation rate of the RHWP.
To operate as a modulator, $\theta$ can be varied in a known way for subsequent analysis of the polarized component.
A good review that places the half-wave plate modulator in context with alternative modulation strategies can be found in Figure 1 of Chuss et al.~(2012)\cite{Chuss12b}.
To achieve this effect in a reflective wave plate, note that the incoming and outgoing signals are easily distinguished by operating at nonzero incident angle, $\alpha$.
In this configuration, the polarization state of an incoming signal can be split with a reflective wire array polarization filter. Light with polarization along the wires is reflected while the orthogonal state is transmitted; see Figure \ref{fig:wp_cartoon}.
In the limit where the wires are much smaller than the wavelength, the relative phase between the two reflected states is given by
\begin{equation}
\label{eq:phase}
\phi = \frac{4 \pi d}{\lambda} \cos{\alpha},
\end{equation}
where $d$ is the distance between the wires and the mirror and $\lambda$ is the wavelength of the signal.
The distance $d$ and the angle $\alpha$ are chosen so that $\phi = \pi$, i.e., the signal reflected by the mirror has traveled an additional half a wavelength.
In this design, the angle of incidence is set to $\alpha=22.2^\circ$, and the nominal central wavelength is $\lambda=3.26$ mm to be consistent with the CLASS 90 GHz telescope \cite{Harrington2018}.
Following Equation \ref{eq:phase}, the wire--mirror spacing is set to be $d=0.88$ mm.
In practice, observations occur over some finite bandwidth, and the total impact of modulation is the integrated effective delay experienced by each frequency in that band \cite{Wiesemeyer14}.
\begin{figure}[t]
\centering
\includegraphics[width=4in]{wp_cartoon.jpg} \caption{Schematic of a reflective wave plate. The wire--mirror distance, $d$, is fixed to incorporate a half-wave path length difference between signals polarized parallel to the wires and those orthogonal to the wires. Modulation is achieved through rotation of the full wire and mirror assembly.}
\label{fig:wp_cartoon}
\end{figure}
\section{Wave Plate Mechanical Design}\label{mechanical}
The architecture for this RHWP is flexible and has the ability of being tailored for use in a broad range of frequencies and optical systems. From Equation \ref{eq:phase}, one needs only adjust wire--mirror distance to operate in a different frequency band.
Packaging and illumination profile concerns can also be addressed by using a diameter for the modulator that is suitable for the particular optical system.
Since this particular design is optimized for testing on the CLASS 90 GHz telescope, the diameter is set to match the current modulator for that telescope, 60 cm.
This choice ensures consistent edge illumination to the operating telescope optics and avoids concerns from interference in packaging.
Broadly speaking, the wave plate consists of three components: a wire array, a stiff support plate, and a flat mirror; see Figure \ref{fig:assem}, Box A. Part of the innovative structure of this design is that the stiff support structure is decoupled from the precision flat requirements of the mirror.
This relaxes the stiffness requirements of the wire support and allows the entire assembly to be lighter weight and more modular.
The stiff support plate holds a uniformly-spaced wire array.
Between the wires and support plate, a mirror is held with 127 threads per inch fine adjustment screws -- AJS127-0.5H.
The mirror is augmented with a raised rim around its perimeter with height equal to the desired wire--mirror separation $d$.
Full contact between the mirror rim and wires is the only requirement for specification of the phase delay.
This means that the mirror rim also functions as a metrology flat for defining the plane of the wires. Once in contact, the surface normal vectors for the wire array and the mirror are both parallel, and alignment with the axis of rotation of the RHWP is achieved through fine adjustment of the mirror.
By design, the contact between the wires and the mirror ensures the wires are moved in concert with the mirror during such an adjustment.
The full RHWP modulator system concept is shown in Figure \ref{fig:assem}.
In addition to the wave plate itself, a rotary drive system suitable for stable motion control and position registration is included.
The drive is a critical aspect of the full modulator; the wave plate itself is insufficient.
Details of the drive will be discussed further in \mbox{Section \ref{drive}.}
In the following subsections the construction of the wire array, \mbox{\ref{sub_wires}}, the targeted wire tension, \mbox{\ref{sub_wireT}}, the design and performance of the support plate, \mbox{\ref{sub_plate}}, and the design and performance of the mirror, \mbox{\ref{sub_mirror}}, are described.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{system_fig_used.jpg}
\caption{{\it Box A.} A cross section schematic for the RHWP concept.
Wires are bonded along the perimeter of the support plate. Precision screws can position the mirror with a raised rim into contact with the wire array. {\it Box B.} Rear view of the support plate.
The support plate uses five stiffening ribs running parallel to the wire direction to support the bending moment from the wire tension on the plate's opposite side. The plate is mounted to the rotation drive through the central ring.
{\it Background.} Cross section view of the model of the rotating half-wave plate system.
The wires are schematically illustrated with the sheet of thin bars over the mirror. The actual wires would not be visible in this scale. The full wave plate assembly is mounted on an interface bearing. The opposite side of the bearing holds the drive motor and absolute rotary encoder.}
\label{fig:assem}
\end{figure}
\subsection{Wire grid fabrication}\label{sub_wires}
To make the wire array, fine-gauge wire is wrapped onto a cylinder following Novak et al.\cite{Novak89}. The circumference defines the length of the wires.
After securing the wires to removable bars running the length of the cylinder, the wires are cut between the bars, and the wire sheets are unwrapped from the cylinder and transferred to an intermediate stretching frame. Once brought to the desired tension, the wires are transferred to their final support structure.
This precision wire grid polarizer fabrication procedure builds upon the work of others\cite{Payne78, Costley77, Lahtinen99} and has been extended to larger grids\cite{Voellmer2008}.
The PIPER balloon-borne CMB experiment significantly improved the large mandrel wrapping technique\cite{Chuss2014}, which was further adapted for the CLASS VPMs\cite{Harrington2018}.
Following the same reasoning as used for previous large grids\cite{Voellmer2008}, 50 $\mu$m diameter tungsten wire is used for the array.
For increased electrical conductivity, the tungsten is plated with a $\sim 2.5$ $\mu$m thick layer of copper, using a nickel adhesion layer.
Assuming a bulk resistivity of $1.72 \times 10^8$ $\Omega\cdot$m at 300 K, the penetration scale of the field into the wire plating is $\approx0.21\mu$m at a wavelength of 3 mm.
Thus the wire plating is expected to be sufficiently thick to determine the emission properties of the wire.
A wire spacing of 175 $\mu$m is targeted.
This wire diameter to spacing ratio strikes a reasonable balance between the inductance for the polarization that is parallel to the wires and the capacitance of the perpendicular polarization.
This broadly optimizes the performance of the wire grid as a quasioptical polarization diplexer\cite{Chuss2012}.
We briefly summarize the wrapping procedure in this paper; a more complete description has been described in previous literature\cite{Harrington2018}.
An 8-inch diameter mandrel is used into which two bars are sunk. These bars protrude slightly above the surface of the mandrel and provide locations for the eventual attachment points of each wire. The first step in the wrapping process is to cut grooves in these mandrel bars that will register each wire with the proper spacing relative to its neighbors.
These grooves are cut into the bars using an end mill set to cut at a 45$^\circ$ angle. In this configuration, as the mandrel is turned, the end mill cuts `V'-shaped grooves in the mandrel bars at the desired wire spacing. This process takes place on a customized CNC milling machine, and the rotation of the axis is synced to the axial motion of the mandrel to set this spacing.
After cutting the grooves in the bars, the wire is wrapped around the mandrel. The wire is fed from its spool through a wire tensioner to maintain approximately constant tension during the wrapping process and avoid sudden tension spikes that might break the wire. The synchronized motion of the CNC is utilized here to ensure that the wires are wrapped to lay in the grooves produced by the end mill. Images of the mandrel bars and wrapping process are shown in Figure~\ref{fig:gridwrap}.
After wrapping is complete, Stycast 2850FT epoxy (with Catalyst 23LV) is used to secure the wires to the mandrel bars. The wires are then cut between the mandrel bars, and the wires are ready to be unwrapped.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{SPIEgridwrap2.jpg}
\caption{{\it Left.} A close up image of the wires after being wrapped on the mandrel. The two mandrel bars are seen running vertically, and the wires are crossing horizontally. {\it Right.} The custom setup for wrapping the wires on the mandrel is shown. The mandrel is mounted on an auxiliary rotary stage. As the stage rotates, the wire is drawn from a supply spool onto the mandrel.}
\label{fig:gridwrap}
\end{figure}
Once unwrapped from the mandrel, the bars supporting the wires are transferred to a stiff frame capable of loading the wires with the desired level of tension \cite{Voellmer2008}.
Rather than mounting the mandrel bars against flat registration pulling bars in the tensioning frame, the frame includes curved bars against which the mandrel bars are pulled.
In this way, the relative tension of the wires across the array is tailored to meet the targets described in Section \ref{sub_wireT} below. The wires are then transferred to the support frame, described in Section \ref{sub_plate}.
\subsection{Wire tension}\label{sub_wireT}
Previous measurements of the sky from Cerro Toco suggest that a polarization modulation frequency of \mbox{10 Hz} is an adequate target to elevate the polarization signal above the $1/f$ knee of low frequency noise \cite{Harrington21, Kusaka14}.
To avoid excitation of vibration in the wires, which would degrade the uniformity of the phase delay definition, the fundamental frequency, $f$, of the wires is required to be much higher than $10~\mathrm{Hz}$.
For this wire grid, an innovative strategy is adopted of varying the wire tension in accordance with the length of the wire in a way that the resonance frequency remains roughly constant across the face of the wave plate.
Using this strategy, the overall tension on the ring supporting the wires is considerably less than if the wire tension was kept constant across the array.
Under less tension, the support structure requires less mass to limit deformation to a given level.
The tension of each wire is a function of its position in the array. Assuming a fixed resonance frequency $f$, the tension is given as
\begin{equation}\label{eq:wire_tension}
\begin{aligned}
T(x)&=4f^2m(x)L_0(x)\\
&=16f^2\rho\pi r^2(R_0^2-x^2)
\end{aligned}
\end{equation}
where $L_0(x)=2\sqrt{R_0^2-x^2}$ and $m(x)=\rho\pi r^2L_0(x)$ are the length and mass of the wire as a function of its $x$ position; directions are defined by the coordinate system shown in Figure \ref{fig:platedeform}. $R_0=30~\mathrm{cm}$ is the radius of the mirror; $\rho=19.3~\mathrm{g/cm^3}$ is the density of tungsten, and $r=25~\mathrm{\mu m}$ is the radius of the wire. Figure \ref{fig:wireTx} shows the wire tension $T(x)$ for different fundamental frequencies.
Based on a suite of simulations to predict the overall deflection of the support plate, see Section \ref{sub_plate}, a total load from the wires on the support plate of $1000~\mathrm{N}$ was found to have a good balance between a high wire resonance frequency and a reasonably light-weight support plate.
The form of the total tension includes two parts: a $980~\mathrm{N}$ component that is proportional to $R_0^2-x^2$, cf. Equation \ref{eq:wire_tension}, and an added $20~\mathrm{N}$ component proportional to $x^2$ to prevent the tension from going to $0$ at the edges of the mirror.
Without the offset, wires close to the edges could become loose after the support plate deforms under the tension from the wires.
The black curve in \mbox{Figure \ref{fig:wireTx}} illustrates the targeted initial tension of the wires across the array while being held with the tensioning frame.
\begin{figure}[t]
\centering
\includegraphics[width=0.48\linewidth]{WireTensionPosition.pdf}
\caption{Wire tension as a function of its $x$ position, for different fundamental frequencies. The black thick line is the target initial tension, which includes a $980~\mathrm N$ component that is proportional to $(R_0^2-x^2)$, and a $20~\mathrm N$ component that is proportional to $x^2$. The smallest fundamental frequency on the targeted curve is $93.4~\mathrm Hz$.}
\label{fig:wireTx}
\end{figure}
The initial tension profile from the tensioning frame is slightly relaxed once the wires are bonded to the support plate perimeter and the load on the tensioning frame is released.
The differences in tension, and resulting changes in resonance frequency, are sufficiently small that pre-loading the support plate can be omitted. Incorporating this tension change after release, the resonance frequency of the wires is found to be $ >91.5~\mathrm{Hz}$ across the entire array, as seen in Figure \ref{fig:wiredTdf}.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth]{WireTensionRelease.pdf}
\caption{\textit{Upper left}: the initial (black) and released (blue) wire tension.
\textit{Upper right}: The fundamental frequency corresponding to the initial (black) and released (blue) wire tension. The lower panels are showing the difference between initial and released quantities. The smallest resonance frequency on the released curve is $91.5~\mathrm{Hz}$. The small variations in the difference result from the positions of the stiffening ribs backing the support plate.}
\label{fig:wiredTdf}
\end{figure}
\subsection{Design of the support plate}\label{sub_plate}
The support plate is designed to hold the wires of the RHWP. While the final position of the wires is determined by the raised rim of the mirror, the support plate is designed to limit the need of the mirror flattening correction.
From this motivation, five ribs are added along the side of the plate opposite the wires.
By aligning the ribs to support against the bending moment from the wires, the plate deformation is reduced without adding much mass; see Figure \ref{fig:assem}, Box B.
The number of ribs, their locations, and their heights were chosen to minimize the mass of the plate while keeping the deformation in the $z$-direction less than $\pm 100~\mathrm{\mu m}$.
This limit is sufficiently small to ensure the flattening ring can further reduce residual deformations to a small fraction of a wavelength.
From this initial result, small adjustments were made to improve manufacturability of the design, add margin to the total deformation, and properly interface with other parts of the RHWP system.
These modifications reduce the deflection from the initial 200 $\mu$m range to only $\sim 34$ $\mu$m at the expense of a modest increase in mass.
For a different application requiring even less deformation or mass, the ribs could be further optimized by making them taller, different heights, or configured with T-shaped cross sections.
Figure \ref{fig:platedeform} shows the finite element analysis (FEA) simulation results of the support plate deformation from the initial tension of the wires.
A non-uniformly-distributed $1000~\mathrm{N}$ tension, see Figure \ref{fig:wireTx}, is applied on the surface of the support plate perimeter rim.
The $\mathrm{UY}$ plot, upper right, indicates that the tension on the wires will decrease after the plate deforms, which implies these results represent the upper limit on the deformation amplitudes.
\begin{figure}[t]
\centering
\includegraphics[width=0.48\linewidth]{PlateDeform-dX.pdf}
\includegraphics[width=0.48\linewidth]{PlateDeform-dY.pdf}
~\\~\\
\includegraphics[width=0.48\linewidth]{PlateDeform-dZ.pdf}
\includegraphics[width=0.48\linewidth]{PlateDeform-res.pdf}
\caption{The FEA simulated plate displacements along $x$-axis, labeled UX (\textit{top-left}), $y$-axis, labeled UY (\textit{top-right}), $z$-axis, labeled UZ (\textit{bottom-left}), and the total amplitude, labeled URES (\textit{bottom-right}).
The wire initial tension is applied to the top surface of the plate edge ring, and its amplitude and direction are represented by the arrows arround the circumference of the plate in each panel.
The total force applied on the upper/lower half-ring is $1000~\mathrm{N}$.
The coordinate system is shown in the center of each panel, in which the $z$-axis is pointing towards the reader.
Most of the deformation comes from the $z$-axis displacements with a range of $65~\mathrm{\mu m}$, which is comparable to the wire diameter.}
\label{fig:platedeform}
\end{figure}
Using $\mathrm{UX}$, upper left, the impact on the wire spacing of the array is estimated.
The wire pitch is found to change by less than $+17.6~\mathrm{nm}$, close to the edge, and $-8.5~\mathrm{nm}$, close to the center; this is a negligible fraction of the ($175~\mathrm{\mu m}$) target pitch.
The change in wire tension and resonance frequency $f$ is estimated by considering $\mathrm{UY}$; the resulting upper limits of the final tension and resonance frequency are shown with the blue curves in the upper panels of \mbox{Figure \ref{fig:wiredTdf}}. The smallest $f$ after wires are released is $91.5~\mathrm{Hz}$ -- much higher than the $10~\mathrm{Hz}$ modulation frequency.
\subsection{Design of the flattening ring mirror}\label{sub_mirror}
As described above, the mirror piece integrates the flat mirror with a delay defining raised rim.
Furthermore, the entire mirror structure is designed to be stiff enough to resist the force from pressing against the wires and drawing them into a flat configuration.
This section reviews the results of FEA simulations of the final shape of the mirror while in contact with the wires.
The force on the mirror perimeter is determined by using the $\mathrm{UZ}$ deformation results; see Figure \ref{fig:platedeform} lower left.
When pushing the mirror towards the wires, the flattening ring will first touch the wires closer to the left and right edge.
To make the wires flat, the mirror must move forward for at least the range of the $z$-direction deformation.
The mirror will then support the projection of wire tension along the $z$-axis.
When computing the projected force on the mirror rim, in addition to the $\mathrm{UZ}$ result, a uniformly distributed $25~\mathrm{\mu m}$ displacement is added as an operational buffer.
The projected force on the face of the plate is shown in the left panel of \mbox{Figure \ref{fig:mirrordeform}}.
This corresponds to a total of $22.5~\mathrm{N}$ force normal to the top surface of the mirror rim.
The simulated mirror deformation along the $z$-axis is shown in the right panel of Figure \ref{fig:mirrordeform}.
The displacement range across the whole mirror is $29.7~\mathrm{\mu m}$.
This result assumes an \emph{overestimated} deformation of the support plate, since the assumed wire load was from the initial wire tension and not from the released tension.
The support plate is therefore expected to deform \emph{less} than shown in Figure \ref{fig:platedeform}, and the mirror will be required to support \emph{less} load along its rim.
Given this estimation method, this final flatness prediction represents a conservative estimate of the mirror deformation. At this level, the final surface is more likely to be dominated by machining tolerances, expected to be $< \pm 20 \mu$m.
Depending on the application, the mirror plate could be made with a larger overall thickness -- potentially with lightweighting features -- to achieve a flatness level consistent with the project objectives.
For initial testing on the CLASS telescope, the edges are under-illuminated; a 30 cm diameter circle centered on the RHWP is a reasonable estimate of the illuminated region \cite{Eimer12}. In this region, see the magenta circle in Figure \ref{fig:mirrordeform}, the displacement range is from $-5.3~\mathrm{\mu m}$ to $2.1~\mathrm{\mu m}$.
\begin{figure}[t]
\centering
\includegraphics[width=0.4\linewidth]{WireTz.pdf}\quad
\includegraphics[width=0.55\linewidth]{MirrorDeform.pdf}
\caption{\textit{Left}: The $z$-axis projection of the wire tension on the mirror flattening ring.
\textit{Right}: The FEA simulated $z$-axis deformation of the mirror.
A total of $22.5~\mathrm{N}$ force that obeys the distribution shown in the left panel is applied on the mirror flattening ring.
The $z$-axis displacement range across the whole mirror is $29.7~\mathrm{\mu m}$, but the range within the illuminated region (inside magenta circle) is from $-5.3~\mathrm{\mu m}$ to $2.1~\mathrm{\mu m}$. The visible trefoil deformation pattern results from the selected location of the three mounting points.}
\label{fig:mirrordeform}
\end{figure}
As mentioned above, the mirror is held using three high precision screws.
Each screw contacts a V-groove along a block mounted to the rear of the mirror.
The three grooves are aligned with the center of the mirror.
The V-blocks and the tip of the adjustment screws are of similar hardness stainless steel alloys to avoid potential binding or pitting between these mating surfaces.
The mirror is held fast against the V-blocks by two extension springs neighboring each adjustment screw.
With this method of mounting, the external stress from the support plate and any thermo-mechanical stress from temperature changes in the operating environment cannot be transferred to the mirror itself.
\iffalse
\begin{itemize}
\item Lower limit on the fundamental frequency, and how does the tension on the wire rely on the fundamental frequency. \textcolor{cyan}{Figure: tension - wire position, for different fundamental frequencies.} Remember added extra tension on each wire.
\item Explain the design of the plate, with fins at the back side. Fins: reduce the deformation of the plate under then tension from the wires, and reduce weight. \textcolor{cyan}{Figure: plate alone, cross-section or back view, with simple wire on. Or refer to the last figure.}
\item Plate deformation under the tension wire, $x$-axis (wire pitch), $y$-axis (tension release) and $z$-axis (mirror deformation). \textcolor{cyan}{Figure: 3 plate deformations.}
\begin{itemize}
\item wire pitch change, $x$-axis. Give some upper limit, read from the previous figure directly.
\item wire fundamental frequency change, $y$-axis. \textcolor{cyan}{Figure: profile of $dy$, corresponding tension change, and fundamental frequency change.}
\item details on mirror deformation, $z$-axis.
\begin{itemize}
\item \textcolor{cyan}{Figure: cross-section of plate, wire, mirror; for explaining how the profile of force is computed.}
\item \textcolor{cyan}{Figure: profile of $dz$, profile of the force (with changed tension) on the mirror flattening ring, mirror $z$-axis deformation.}
\end{itemize}
\end{itemize}
\item Figure of the half wave plate system; how the mirror is held and made adjustable. How to bring the mirror into the alignment.
\end{itemize}
\fi
\section{RHWP Drive}\label{drive}
Our initial operational plan for the RHWP modulator is to execute rotation using a single direction, constant angular velocity motion.
In reality, the wave plate will be mounted in a telescope system executing scanning maneuvers and being buffeted by winds.
In order to isolate the rotation from these external factors, a servo drive system with sufficient bandwidth and control loop speed to produce stable output motion has been developed.
The drive system, including position readout, is a critical component for the success of the modulator. Without precision phase control and accurate position knowledge, demodulation of the data products can become cumbersome at best and possibly contaminated by significant systematic errors and instabilities.
Therefore the drive system is a \emph{critical} part of the modulator system itself, and the current configuration for our drive is described here. One interesting feature of the reflective wave plate approach is that the drive system naturally lives behind the wave plate mirror. Not only does this help separate drive electronics from the optical path, reducing risk from radio frequency interference, but it also enables the use of conventional bearings, motors, and encoders for essentially arbitrary diameter devices.
Due to the requirement for both velocity and position control, a programmable servo drive and a brushless AC servo motor are used to position the RHWP.
The motor is a 16 pole direct-drive servo motor from Kollmorgen, DH062M-12-1310.
This motor is able to control a load with several hundred times the moment of inertia of the armature without the need for gear reduction, which provides the advantage of being able to incorporate the motor, load, and position readout encoder all on one shaft.
The servo drive is an ABB/Baldor MicroFlex e100,
and the controlling computer is a PICMG 1.3 single board computer running the VxWorks real time operating system mounted in a mixed PCI/PCI Express passive backplane.
The servo loop runs at 20 Hz, taking as input the commanded position and velocity one time step in the future along with the current commanded acceleration.
When running at a constant velocity, the commanded position is simply incremented by the velocity times the time step, and the commanded acceleration is zero.
The loop first queries the servo drive for its current motor encoder position and velocity along with various other quantities such as current, voltage, following error, etc.
This query uses the Baldor Immediate Command Mode protocol (ICM), which provides the ability to include these queries into a signal TCP packet, and receive the data packed into one packet as well.
Typical turn around time is about 2.2 ms over 100 Mbps Ethernet -- more than adequate for the 50 ms servo loop time step.
The loop then calls a profiler, which uses these quantities to calculate the required velocity and acceleration for the next time step.
These are then packed into an ICM packet and sent to the drive.
The position readout encoder is a Heidenhain RCN 2581 angle encoder with a system accuracy of $\pm 2$ seconds of arc.
This encoder uses an Endat serial data interface with a resolution of 28 bits per revolution for the absolute position readout.
It also has 1 volt peak-to-peak analog sine/cosine incremental signals with a resolution of 16384 cycles per revolution.
The absolute position is read out by a PCI Express card on the control computer.
The readout is done at $\sim 201$ Hz and is activated by an interrupt sent by the same clock that triggers the detector data reads -- ensuring synchronous recording for later demodulation.
The interrupt handler first latches the time on a PCI GPS card also installed in the control computer.
It then latches the encoder signal and passes off to a lower priority interrupt service routine, which reads the clock and encoder, differences the encoder position to derive velocity, gathers all of the additional data read from the servo drive at 20 Hz, and then packs all of this into a data structure written to a ring buffer.
A third task running at 1 Hz empties the ring buffer and then writes the data to disk.
Our current test setup uses the full system shown in Figure \ref{fig:assem}, except the wave plate is replaced by a metal disk replicating the expected 0.627 Nm moment of inertia of the RHWP. To minimize risk from pulse width modulation motor drive currents, a three phase electromagnetic interference filter with $\sim 250$ kHz cutoff is added.
While the motor has an internal Heidenhain ECN 113 rotary encoder, we use the sine/cosine outputs from the higher precision RCN 2581 for motor position feedback.
This increases the feedback resolution by a factor of 8 and significantly reduces velocity error due to the much better linearity of the higher precision encoder.
Laboratory testing in front of a cold receiver with transition edge sensor bolometers using a similar setup showed no discernible interference in the relatively high noise environment of the room.
Velocity control of this preliminary lab test is demonstrated to 0.006\% at the 2.5 Hz rotational velocity of the RHWP.
\section{Conclusion}\label{conclusion}
Front-end polarization modulation is a powerful tool for distinguishing a faint polarized signal from a much larger unpolarized background while protecting against systematic errors such as instrument polarization.
In this work, the design of a new \mbox{60 cm} clear aperture reflective wave plate has been demonstrated.
The architecture is flexible and can be scaled for a wide range of applications.
In the prototype, the parallelism of polarizer and mirror are anticipated to meet our need through the constraints in the design.
Similarly, the deformation of the mirror in its final configuration is estimated to be limited to less than \mbox{29 $\mu$m.}
A prototype drive system has been constructed in the lab using a realistic mass model, and early tests demonstrate the planned control system is feasible and enables a stable rotational frequency, normally \mbox{2.5 Hz}, with an acceptably small error.
\acknowledgments
We acknowledge the National Science Foundation Division of Astronomical Sciences for their support.
The RHWP modulator is being developed with support from Grant Numbers 2034400 and 2109311. We further acknowledge the very generous support of Jim and Heather Murren (JHU A\&S ’88), Matthew Polk (JHU A\&S Physics BS ’71), David Nicholson, and Michael Bloomberg (JHU Engineering ’64). Zhilei Xu is supported by the Gordon and Betty Moore Foundation through grant GBMF5215 to the Massachusetts Institute of Technology.
|
\section*{Acknowledgements}
\ifdefined\isaccepted
We thank Philip Sperl and researchers at the Fraunhofer Institute for Applied and Integrated Security and at the Center for Human-Compatible AI for helpful discussions.
We thank P.~A.~M. Casares for helpful feedback on earlier drafts of this paper.
We also thank our anonymous reviewers for feedback on this paper.
Work supported in part by the Berkeley Existential Risk Initative and Centre for Effective Altruism Long Term Future Fund.
\else
Removed for double blind review.
\fi
\section{Adding a Communication Channel to Simple Push}
\label{sec:apdx-comm}
\begin{figure}[h]
\begin{subfigure}[b]{0.5\linewidth}
\centering
\includegraphics[width=0.8\linewidth]{img/attack-push-no-comm-selfplay-0.pdf}
\caption{Agressor.}
\label{fig:no_comm_0}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}
\centering
\includegraphics[width=0.8\linewidth]{img/attack-push-no-comm-selfplay-1.pdf}
\caption{Defender.}
\label{fig:no_comm_1}
\end{subfigure}
\caption{Average return of agent controlling (a) agressor and (b) defender in Simple Push \emph{without} communication. The adversary, in red, fails to outperform the self-play baseline, in black.}%
\label{fig:sp_no_comm}
\end{figure}
Our initial experiments in the \emph{Simple Push} environment did not lead to adversarial policies that were capable of outperforming their victims (See Figures~\ref{fig:no_comm_0}, \ref{fig:no_comm_1}).
Inspired by the cooperative environments explained in \citet{lowe2017multi}, we add a communication channel: this channel allows each agent to observe a one-hot coded action taken by the other agent.
This communication channel has no other effect on environment dynamics, and agents' reward does not depend on the contents of the communication channel.
The size of the communication channel essentially represents the number of tokens either agent can use to communicate with the other.
Because this setting is competitive, there is no reason for an agent to provide information in the communication channel which would be beneficial to the opponent.
Therefore, an optimal policy should simply ignore the \enquote{messages} sent by the opponent.
However, this channel increases the dimensionality and offers an adversary the possibility to learn messages that might \enquote{confuse} a (sub-optimal) victim.
In a small ablation on communication channels supporting 10, 25, 50 and 100 tokens, we find that adversarial policies are successful with 50 or more tokens and unsuccessful at less than 25.
We also find that the number of timesteps until convergence, as well as general instability during training increases with higher sizes.
For further experiments we use a communication channel of 50 tokens to allow for fast training while still providing an environment in which adversarial policies are possible.
\section{Conclusion}
We introduced the defense by PBRL as a method to reduce exploitability of RL policies.
Our results show an increase in zero-shot robustness against new adversaries compared to self-play training.
However, some self-play victims are naturally robust and PBRL comes with an increased computational cost.
We find the size of the population necessary depends on the environment used, and larger populations can increase overall robustness.
This work suggests that increasing diversity during training can lead to improved robustness and contributes toward the goal of making agents less exploitable.
Supplementary material available at \hyperref[]{\url{https://reducing-exploitability.github.io}}, source code available at \hyperref[]{\url{https://github.com/HumanCompatibleAI/reducing-exploitability}}.
\section{Discussion and Future Work}
In this work we evaluated our hardened PBRL victim agent against adversarial policies trained with similar numbers of timesteps of experience as the victim agent.
However, PBRL \emph{also} requires training $n$ opponents for this many timesteps -- so the \emph{total} number of timesteps and computational resources is $n+1$ times greater.
We believe this overhead is often tolerable.
First, defenders may be able to limit the number of timesteps an attacker can train against the victim for, such as if access to the policy is behind a rate-limited API.
Second, the number of opponents $n$ can sometimes be quite small -- just $n=2$ suffices for Simple Push.
Finally, defenders often have significant computational resources: while PBRL is unlikely to prevent an attack from a sophisticated adversary like a nation state, it may be enough to defeat many low-resource attacks.
Nonetheless, reducing this computational overhead is an important direction for future work.
For example, can we obtain similar performance with fewer opponents if we train them to be maximally diverse from one another?
In addition, \emph{if} additional compute resources are available, our approach allows a defender to \emph{make use} of them.
Once an agent has converged, using additional compute to continue training in self-play is usually of no use.
However, convergence is not sufficient for robustness -- as illustrated by the existence of adversarial policies.
Our approach enables a purposeful use for additional computing power.
A possible explanation for adversarial policies could be that neural networks of the fairly small sizes used in RL are not able to represent more robust policies.
The fact that our experiments in Laser Tag have not hit diminishing returns of population size regarding the performance of the protagonist shows that changing the training setup can improve model robustness for a fixed model architecture.
Effectively PBRL is a way to improve a policy given a fixed neural network size when continuing to train will not provide additional benefits (due to convergence).
A key open question is how the number of opponents $n$ required for robustness scales with the complexity of the environment.
PBRL will scale poorly if the required population size is proportional to the size of the state space: in more complex environments each opponent will take longer to train \emph{and} more opponents will be required.
But a priori it seems likely that $n$ may depend more on the number of high-level strategies in the environment.
This is only loosely related to the dimensionality of the state space.
For example, some simple matrix games have high strategic complexity, while some high-dimensional video games have only a handful of sensible strategies.
Our evaluation uses the original adversarial policy attack of \citet{gleave_adversarial_2020}, which we established was strong enough to exploit unhardened victims in these environments.
However, it is possible that alternative attacks would be able to exploit even our hardened victim.
We hope to see iterative development of stronger attacks and defences, similar to the trend in adversarial examples more broadly.
\section{Background}
Following \citet{gleave_adversarial_2020}, we model the tasks as Markov games $M = (\mathcal{S}, (\mathcal{A}_\alpha, \mathcal{A}_\nu), \mu, \mathcal{T}, \gamma, (R_\alpha, R_\nu))$ \citep{Littman94markovgames} with state set $\mathcal{S}$, action sets $\mathcal{A}_i$, initial state distribution $\mu$, transition dynamics $\mathcal{T}$, discount factor $\gamma$ and reward functions $R_i$.
The index $i \in \{\alpha,\nu\}$ represents the adversarial ($\alpha$) and victim ($\nu$) agent respectively.
We focus on two-player, zero-sum games as they have a clear competitive setup and obvious adversarial goals.
Consequently $R_{\alpha} = -R_{\nu}$.
Notably, the victim policy trained in a zero-sum game will be \emph{harder} to exploit than those in a positive-sum setting, where they may have learned to cooperate with other agents (in an exploitable fashion) to maximize their overall reward.
The attacker has grey-box access to the victim.
That is, they can train against a fixed victim, but cannot directly inspect its weights.
The attacker has no additional capabilities to manipulate the victim or the environment.
\section{PBRL Defense}
We would like to find a Nash equilibrium $(\pi_{\nu}, \pi_{\alpha})$, which for zero-sum games corresponds to the minimax solutions:
We take the $\arg \min_{\pi_\nu}\max_{\pi_\alpha}$ of the expectation $\mathbb{E}\left[\sum_{t=0}^\infty \gamma^tR_{\alpha}(S^{(t)},A_\alpha^{(t)}, A_\nu^{(t)}, S^{(t+1)}) \mid \pi_\alpha,\pi_\nu\right],
$
with random variables sampled from $S^{(0)} \sim \mu$, $A_i^{(t)} \sim \pi_i(\cdot \mid S^{(t)})$, and $S^{(t+1)} \sim \mathcal{T}(S^{(t)}, A_{\alpha}^{(t)}, A_{\nu}^{(t)})$.
However, as previously discussed self-play may not find the Nash equilibrium.
Yet the self-play policies are nonetheless often highly capable -- against their self-play opponent.
We therefore conjecture that self-play gets stuck in a \emph{local} Nash equilibrium:
it converges but is not globally optimal.
Notably, adversarially training against some $\pi_{\alpha}$ trained against $\pi_{\nu}$ might just cause $\pi_{\nu}$ to move to a new local Nash equilibrium that is robust to $\pi_{\alpha}$, but not unseen adversary $\pi_{\alpha}'$.
We therefore propose using population based reinforcement learning \citep[PBRL;][]{jaderberg_human-level_2019}.
We train an agent, the \emph{protagonist}, for robustness by training it against a population of $n$ opponents $\pi_{o_i}$.
By jointly optimizing against multiple opponents we increase the coverage of the space of opponent policies.
Since an adversary $\pi_\alpha'$ optimizes in a similar way as the opponents it is likely to be close to one of the opponent policies $\pi_{o_i}$.
Moreover, given sufficient diversity in opponents it may be easier for the protagonist to learn a policy close to global Nash than to learn $n$ strategies that overfit to each opponent.
Each of the $n$ opponents has identical architecture and training objective, differing only in the seed used to randomly initialize their network.
We alternate between training the opponents against a fixed protagonist, and a protagonist against all fixed opponents.
All agents (opponent and protagonist) are trained for the same total number of timesteps.
This alleviates a factor for diminishing returns with higher population sizes, which \citet{vinitsky2020robust} encountered when training adversaries such that \emph{in aggregate} they receive as many timesteps as the main agent.
This allows us to make use of an order of magnitude larger population sizes with populations containing tens of opponents.
The total number of training timesteps for all policies is $n+1$ times the number of timesteps the protagonist is trained for.
When logging timesteps for PBRL training, we report the number of timesteps the protagonist agent trains, since this is the relevant metric for protagonist training.
Note that this means the compute necessary for training PBRL is $n+1$ times higher than self-play at the same number of training steps (although PBRL is more parallelizable).
\def\rllib/{\texttt{RLlib}}
We train all policies -- self-play, PBRL and adversarial -- using Proximal Policy Optimization \citep[PPO;][]{schulman_proximal_2017}. %
PPO is widely used and has achieved good results with self-play in complex environments \citep{bansal_emergent_2018}.
Furthermore, prior work on adversarial policies in higher-dimensional two-agent environments uses PPO \citep{gleave_adversarial_2020}.
We use the PPO implementation in \rllib/~\citep{rllib}, from the \texttt{ray} library \citep{ray}, due to its support for multi-agent environments and parallelizing RL training.
\section{Related Work}
\citet{gleave_adversarial_2020} introduced the adversarial policy threat model, and the first attack: using RL to train an adversary against a fixed victim.
\citet{wu_adversarial_2021} added an auxiliary term to reward the victim paying attention to the adversary.
They apply this attack to the original environments \citep{bansal_emergent_2018}, and a new environment: Roboschool Pong.
\citet{guo_adversarial_2021} develop a different attack for semi-competitive games, exploiting agents in the original environments and Starcraft 2.
Comparatively little attention has been given to defenses relative to attacks.
\citet{gleave_adversarial_2020} attempted to harden the victim by fine-tuning it against a fully trained adversary.
The hardened victim was robust to that particular adversarial policy -- but it was still vulnerable to repeating the attack.
Furthermore, the hardened victim achieved lower performance against the original, non-adversarial opponent.
\citet{guo_adversarial_2021} use the same approach and find that fine-tuning on adversaries from a stronger attack can provide robustness against adversaries from weaker attacks.
However, the authors don't evaluate the robustness of the hardened victim to an attack targeting itself rather than the original victim.
One prior defense is \emph{Adversarially Robust Control} \citep[ARC;][]{kuutti2021arc}.
They consider the semi-competitive setting of autonomous driving.
They find that imitation-learned policies are vulnerable to adversarial vehicles trained to cause collisions -- even when the adversary is limited to only cause preventable collisions.
To improve robustness they fine-tune the imitation policies against an ensemble of adversaries that train concurrently with the main policy.
Since autonomous driving is semi-competitive, an optimal policy against adversaries might fare poorly against regular agents, so \citeauthor{kuutti2021arc} add an auxiliary loss to keep the fine-tuned policy similar to the imitation policy.
In contrast, we focus on the more challenging zero-sum setting which self-play was designed to work with.
\citet{vinitsky2020robust} formulate the single-agent robust RL problem as a zero-sum two-agent problem:
The adversary learns to apply noise perturbations to the environment dynamics trying to minimize the performance of the robust agent in the environment.
To improve robustness the authors train the control agent against a population of these adversaries and find this increases robustness against overfitting to particular adversaries, increases robustness against an out-of-distribution held out test set and is less brittle than the domain randomization baseline.
Our work considers a more general attack setting:
Instead of an adversarial formulation of a single-agent control setting, we analyze two-agent symmetric and asymmetric games.
It seems likely that these kinds of two-player games, have a higher diversity of possible adversarial, possibly long-term, strategies compared to adversaries that control limited perturbations of transition dynamics.
Lower strategic diversity could be one possible reason why \citeauthor{vinitsky2020robust} encounter diminishing returns at fairly low sizes of adversarial populations.
Additionally, we alleviate one factor for diminishing returns with higher population sizes:
Training all policies, including adversaries, with a balanced number of timesteps (see Section~\ref{sec:defense}) eliminates the problem that adversaries in larger populations have less opportunity to train.
However, this approach does come with an increase in computational effort.
An additional key difference is that while \citeauthor{vinitsky2020robust} investigate robustness under swaps of adversaries from different training runs, our threat model consists of an attacker that optimizes performance against the specific policy whose robustness we evaluate, i.e. our evaluation is for zero-shot robustness against new adversaries as opposed to adversaries from separate seeds.
Self-play is designed to solve zero-sum games -- so it is natural to ask why it is vulnerable to such exploits in the first place.
Indeed fictitious play, where each agent plays the best response to the opponent's historically observed strategy, is known to find the Nash equilibrium for two-player, zero-sum games \citep{robinson_fictious_1951}.
Self-play is related to fictitious play, and has been shown to also converge to Nash equilibria provided that RL converges to a best response asymptotically \citep{heinrich_selfplay_2015}.
However, RL may never converge to a best response.
A major culprit is non-transitivity: \citet{balduzzi_openended_2019} show that in games like rock-paper-scissors, self-play may get stuck in a cycle.
Even in a transitive game, deep RL algorithms may never converge to a best response if they are stuck in a local minimum, or if the policy network cannot represent the optimal policy.
\section{Experiments}
\label{sec:experiments}
We evaluate the PBRL defense in two low-dimensional environments, described in Section~\ref{sec:env}.
In Section~\ref{sec:attack} we confirm that baseline self-play policies are vulnerable to attack.
To the best of our knowledge, these are the lowest-dimensional environments in which an adversarial policy has been found.
Finally, in Section~\ref{sec:defense} we find that PBRL improves robustness against adversarial policies, and explore the relationship with population size.
Unless otherwise noted, in all experiments we train 5 seeds of victim policies.
We attack each victim using 3 seeds of adversaries for a total of 15 adversaries.
Unless omitted for legibility, 95\% confidence intervals are shown as shaded regions for training curves and bars in bar plots.
\subsection{Environments}
\label{sec:env}
\begin{figure}
\begin{subfigure}[b]{0.5\linewidth}
\centering
\includegraphics[width=0.6\linewidth]{img/laser_tag_env.png}
\caption{Laser Tag}
\label{fig:envs:laser}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}
\centering
\fbox{
\input{txt/laser_tag_env_tikz.tex}
}
\caption{Simple Push}
\label{fig:envs:push}
\end{subfigure}
\caption{Illustrations of the (a)~Laser Tag and (b)~Simple Push environments.
Laser Tag is a symmetric grid-world game, Simple Push is continuous control and asymmetric.
It consists of \emph{aggressor} and \emph{defender} agents, as well as target and decoy landmarks.
}
\label{fig:envs}
\end{figure}
We evaluate in two low-dimensional, two-player zero-sum games illustrated in Figure~\ref{fig:envs}: \emph{Laser Tag} and \emph{Simple Push}.
\emph{Laser Tag} is a symmetric game with incomplete information \citep{lanctot_unified_2017}.
The players see 17 spaces in front, 10 to the sides, and 2 spaces behind their agent.
The two agents move on a grid world and get points for tagging each other with a light beam.
Obstacles block movement and beams.
We make the environment zero-sum by also subtracting a point from the tagged player.
\emph{Simple Push} is a continuous environment introduced by \citet{mordatch2017emergence} and released with \citet{lowe2017multi}.
The environment is asymmetric, with one agent the \emph{aggressor} and the other the \emph{defender}.\footnote{Note, that the notion of \emph{aggressor} and \emph{defender} in this environment is orthogonal to \emph{adversary} and \emph{victim} in the sense of adversarial policies.}
The environment contains two randomly placed landmarks.
Only the defender knows which of these is the true target, the other landmark acts as a \enquote{decoy} for the aggressor.
The aggressor receives positive reward based on the defenders distance to the true target.
Subtracted from this is a relative penalty, based on its own distance.
Unlike vanilla Simple Push, where the defender's rewards are solely based on its own distance, we make the environment zero-sum by giving the defender the negative of the aggressor's reward.
Initial experiments, whose training curves can be found in Figure~\ref{fig:sp_no_comm} of the appendix,
showed that the attack from \citet{gleave_adversarial_2020} does not find an adversarial policy in vanilla Simple Push.
As Simple Push is very low dimensional (a two-dimensional continuous control task), we develop a variant with a \enquote{cheap talk} communication channel (see Section~\ref{sec:apdx-comm} in the appendix) that increases the dimensionality but does not otherwise change the dynamics.
This channel extends the action and observation spaces allowing agents to send one-hot coded tokens to each other.
With this channel we successfully find adversarial policies, supporting the hypothesis mentioned by \citeauthor{gleave_adversarial_2020}, that increased dimensionality increases the chance to find adversarial policies.
Consequently, we perform experiments in Simple Push with a one-hot coded communication channel of 50 tokens.
See Section~\ref{sec:apdx-comm} of the appendix for details on the communication channel.
\begin{figure*}[h!]
\centering
\begin{minipage}[]{0.33\linewidth}
\centering
\includegraphics[width=0.94\columnwidth]{img/attack-laser-selfplay.pdf}
\caption{Training curve of adversaries in \emph{Laser Tag}. After fewer than 15 million timesteps of training on average a successful adversarial policy can be found. However variance is high and adversary performance deteriorates against some strong victims over time.}
\label{fig:lt_attack_sp}
\end{minipage}%
\quad%
\begin{minipage}[]{0.64\linewidth}
\begin{subfigure}[t]{0.5\linewidth}
\centering
\includegraphics[width=\linewidth]{img/attack-push-selfplay-0.pdf}
\caption{Aggressor return.}
\label{fig:sp_agressor}
\end{subfigure}%
\begin{subfigure}[t]{0.5\linewidth}
\centering
\includegraphics[width=\linewidth]{img/attack-push-selfplay-1.pdf}
\caption{Defender return.}
\label{fig:sp_defender}
\end{subfigure}
\vspace{4px}
\caption{\emph{Simple Push} average return of adversary controlling aggressor (left) and defender (right) \emph{with} 50-token communication channel.
The black dotted line marks the return achieved by the self-play training baseline controlling the respective agent at the end of training.
Adversary controlling aggressor leads to weak attack, defender leads to strong attack.}
\label{fig:sp}
\end{minipage}
\end{figure*}
\subsection{Self-play Baseline}
\label{sec:attack}
Before evaluating our PBRL defense, we first consider the exploitability of the self-play baseline by the standard adversarial policy attack from \citet{gleave_adversarial_2020}.
We find self-play policies in both Laser Tag and Simple Push to be vulnerable.
However, in Simple Push the attack only succeeds when there is a communication channel.
In Laser Tag adversarial policies can be found, however variance is high and some self-play victims are hard to attack.
These results add nuance to the conclusion from \citet{gleave_adversarial_2020} that adversarial policies are easier to find in higher-dimensional environments.
Our environments are significantly lower-dimensional than those considered by \citeauthor{gleave_adversarial_2020}, suggesting the minimum dimensionality for attack is smaller than previously believed.
However, the fact that policies are only vulnerable in Simple Push given a communication channel, and the high variance in the victim robustness in Laser Tag, supports the overall claim that dimensionality is an important mediator for exploitability.
\paragraph{Laser Tag.}
\label{sec:attack-laser}
We train self-play policies in the symmetric Laser Tag environment for 25 million timesteps.
This should be adequate to produce a strong policy, as the paper introducing the environment \citep{lanctot_unified_2017} trained self-play for only 3 million timesteps.
Figure~\ref{fig:lt_attack_sp} shows the average return of the adversaries trained against these victims.
We train adversaries for 50 million timesteps, twice as many as the victims, in order to reason about the adversaries' behavior given more compute.
Since the game is symmetric, an agent with a return above zero outperforms its opponent.
Successful adversarial policies can be found: if the adversary were to stop training once it outperforms the victim, on average fewer than 15 million timesteps are necessary for a successful attack. %
The loose confidence interval suggests high variability in different seeds.
Some of the trained victims are robust while others are not.
On average attacker performance deteriorates after 20 million timesteps which suggests an instability in adversary training.
\paragraph{Simple Push.}
\label{sec:attack-push}
Since Simple Push has an asymmetric observation space, we train self-play using a separate policy for either player.
We train the agents for 25 million timesteps, which we expect to be more than sufficient given the 625,000 timesteps used in prior work \cite{lowe2017multi}.
While prior work used MADDPG, not PPO, in exploratory experiments we find PPO to perform comparably to MADDPG.%
Again we train adversaries for 50 million timesteps, twice as many as the victims, in order to reason about the adversaries' behavior given more compute.
Figure~\ref{fig:sp_defender} shows when the adversary controls the defender the adversary achieves almost twice the return on average (in red) as the self-play baseline (in black) at 25 million timesteps, which is the point where adversary and victim trained for the same number of timesteps.
By contrast, Figure~\ref{fig:sp_agressor} shows that when controlling the aggressor the adversary needs around 20 million timesteps just to match the victim.
Return after training the adversary for twice as many timesteps as the victim is only slightly higher than the baseline.
These results suggest defender policies in Simple Push are more robust to adversarial policies than aggressor policies.
We conjecture this asymmetry is due to the defender having more information than the aggressor: it knows the target landmark.
Consequently, %
the aggressor needs to observe the opponent to learn the true target landmark.
The defender could exploit this and perform movements that fool a victim aggressor.
Due to the stronger nature of the attack controlling the defender, we focus our defense in the upcoming section on adversaries that control the defender agent.
\subsection{PBRL Defense}
\label{sec:defense}
\begin{figure*}[h!]
\centering
\begin{subfigure}[t]{0.43\linewidth}
\includegraphics[width=0.99\linewidth]{img/attack-laser-pbt-curves.pdf}
\caption{Averages of adversary returns.}
\label{fig:curve-laser-pbt}
\end{subfigure}%
\begin{subfigure}[t]{0.57\linewidth}
\includegraphics[width=0.254\linewidth]{img/PBRL-20.pdf}%
\includegraphics[width=0.199\linewidth]{img/PBRL-40.pdf}%
\includegraphics[width=0.199\linewidth]{img/PBRL-60.pdf}%
\includegraphics[width=0.199\linewidth]{img/PBRL-80.pdf}%
\includegraphics[width=0.15\linewidth]{img/legend.png}%
\caption{Averages with confidence intervals.}
\label{fig:curve-laser-pbt-ci}
\end{subfigure}
\caption{Training curves for adversaries attacking in \emph{Laser Tag}.
We measure exploitability by comparing how many timesteps the adversary needs to train for until it outperforms the victim, i.e. achieves larger than 0 return. We compare the self-play baseline and policies PBRL-$n$, trained with population size $n$.
With increasing population size the number of timesteps the adversary needs increases, suggesting that exploitability decreases with higher population sizes.
However, we also find self-play is more robust than policies trained using PBRL against adversaries that continue training, suggesting that there might be more instability when training an adversary against self-play. Figures in (b) show the confidence intervals for returns of adversaries attacking PBRL and self-play policies from (a).}
\end{figure*}
In this section, we evaluate the effectiveness of our PBRL defense by trying to exploit PBRL-trained policies.
We find some improvement in robustness relative to the self-play baseline in both environments.
In Laser Tag larger populations increase the number of timesteps needed to find the first adversarial policy -- at the cost of requiring more computational resources.
In Simple Push adversaries don't outperform self-play baselines and there are no significant improvements with more than $n=2$ opponents.
We evaluate our defense by attacking the PBRL-trained protagonist agent, and compare to the attacks on the baseline self-play policies from Section~\ref{sec:attack}.
We focus on the relative performance of the adversary compared to what the victims achieve on-distribution. %
We continue training adversaries up to 50 million timesteps against the same fixed 25-million-timestep victim to evaluate how many timesteps are needed to attack more robust victims.
\paragraph{Laser Tag.}
To explore the impact of population size we train policies with $n=20$, $40$, $60$, and $80$ opponents in Laser Tag.
Figure~\ref{fig:curve-laser-pbt} shows the average return of adversaries attacking these hardened protagonists.
We find that using PBRL increases robustness: finding an adversary that achieves higher than 0 reward takes more timesteps on average.
While fairly noisy, generally the number of timesteps needed to outperform the victim -- when return crosses over the zero line -- increases with increasing population size.
An adversary attacking a protagonist hardened against a population of size 80 needs to train on almost double the timesteps as the self-play victim.
However, the relative difference in timesteps is smaller the higher the population size.
Although the adversarial policy was trained for up to double the number of timesteps as the protagonist agent, note that PBRL used $80$ times as much compute for every timestep of the protagonist, since it had to train the opponents for the same number of timesteps.
Additionally, we find that adversaries that continue to train eventually do outperform PBRL victims on average, whereas average performance against self-play victims \emph{decreases} over time.
When evaluating $95\%$ confidence intervals (Figure~\ref{fig:curve-laser-pbt-ci}), %
variance of adversaries attacking self-play increases over time, which suggests adversary training to be less stable when attacking self-play as opposed to attacking PBRL.
Self-play seems to converge to policies of widely varying robustness, whereas in PBRL variance of attackers is lower.
\paragraph{Simple Push.}
\label{sec:defense-push}
We focus on making the \emph{aggressor} agent more robust in Simple Push, as Section~\ref{sec:attack-push} showed that the defender self-play policy is already relatively robust to attack.
Consequently, the protagonist controls the \emph{aggressor} and adversaries the \emph{defender} agent.
We use PBRL to train against $n=$~2, 4, 8 and 16 opponents.
Since the environment is not balanced, as a baseline we judge performance by comparing to the returns at the end of training.
The adversary is successful if it achieves higher results against the victim than the PBRL opponents did on average at the end of training against that same victim policy.
\begin{figure}[h]
\centering
\includegraphics[width=0.85\columnwidth]{img/attack-push-pbt-1.pdf}
\caption{PBRL-trained agents (green) are significantly more robust to adversarial attack than the self-play baseline (red), and attain similar return to the baseline that plays against a \emph{non-adversarial} agent (black dotted line).
}
\label{fig:push_pbt}
\end{figure}
Figure~\ref{fig:push_pbt} shows the returns at 25 million timesteps.
Since baselines in different setups could converge to different returns, we calculate separate baseline thresholds for each of the 4 settings (in addition to the self-play baseline from Figure~\ref{fig:sp}), marked by the dashed black line.
Notably, the values all policies converge to differ by less than 3\%.
The PBRL-trained agent is significantly more robust than the self-play policy, in red.
In fact, the PBRL protagonist achieves similar return \emph{under a zero-shot attack} as the self-play policy does \emph{against its self-play opponent} (the dashed threshold).
\begin{figure}[h]
\centering
\includegraphics[width=0.9\columnwidth]{img/attack-push-pbt-curves.pdf}
\caption{Return of adversaries attacking all PBRL-trained agents (number of opponents $n = 2, 4,8,16$), averaged over all 60 adversaries (15 per $n$). Victims are all similarly robust in Simple Push, adversaries need at least as many timesteps as the victim in order to outperform it.}
\label{fig:push_pbt_curve}
\end{figure}
Figure~\ref{fig:push_pbt_curve} shows the training curve, when training adversaries for up to 50 million timesteps.
Since there is no discernible difference in the 4 PBRL settings, we average over these for a total of 60 adversarial policies.
When the adversary trains \emph{twice as much} as the PBRL protagonist victim, the attack outperforms the victim, but is not particularly strong.
Although PBRL is significantly more robust than self-play (effectively PBRL with $n=1$ opponents), perhaps surprisingly there is little benefit from using more than $n=2$ opponents.
In particular, there is no clear decrease in robustness when using $n$ as low as 2, which is the lowest PBRL settings that does not degenerate to self-play.
This is in contrast to Laser Tag, which saw large differences in robustness depending on population size.
This is likely due to different environmental dynamics.
In Simple Push there are only a handful of high-level strategies that one can pursue.
By contrast, the Laser Tag environment allows for more variation in the details of possible strategies, making it harder to achieve full robustness.
Additionally, it is possible that the intervention of slightly increasing dimensionality, by adding a \enquote{cheap-talk} channel, can be circumvented with minimally higher diversity during training.
\section{Introduction}
The discovery of \emph{adversarial examples} \citep{szegedyIntriguing2014} showed that capable image classifiers can be fooled by inputs that are easily classified by a human.
Reinforcement learning (RL) policies have been shown to also be vulnerable to adversarial inputs \cite{huangAdversarial2017, kosDelving2017}.
However, attackers often are not able to directly perturb inputs of a victim agent.
\citet{gleave_adversarial_2020} instead model the attacker as controlling an adversarial agent in an environment shared with the victim.
This adversarial agent does not have any special powers, but can \emph{indirectly} influence the victim's observations by taking actions in the world.
\citeauthor{gleave_adversarial_2020} train such \emph{adversarial policies} against a fixed victim policy using RL, and are able to exploit state-of-the-art victim policies in zero-sum, two-agent, simulated robotics environments \cite{bansal_emergent_2018}.
This vulnerability is surprising given that self-play has produced policies that can defeat the world champions in Go \cite{silverMasteringGameGo2016a} and Dota~2 \cite{openai_dota_2019}.
Critically, zero-sum games are \emph{naturally} adversarial, with self-play being akin to adversarial training:
so we might expect self-play policies to be naturally robust.
Yet training an adversary for 3\% of the timesteps the original self-play procedure trained for, \citet{gleave_adversarial_2020} manage to exploit the victims.
The resulting adversarial policies exhibit counterintuitive behaviour such as falling over and curling into a ball in a game where the goal is to prevent the opponent from getting by.
This would be a weak strategy against most agents, but is able to destabilize the victim.
Similar to adversarial examples, these policies showcase failure modes that affect a model, but which a human would most likely be unaffected by.
Such attacks are a critical danger to deep RL policies in high-stakes settings where there may be adversaries, such as autonomous driving or automated financial trading.
We propose using population based reinforcement learning \citep[PBRL;][]{jaderberg_human-level_2019} to train an agent against a diverse population of opponents unlike traditional RL approaches, which train one fixed policy per player.
PBRL is a variation of population based training \citep[PBT;][]{jaderberg2017population}, which jointly optimizes a population of models and their hyperparameters for improved convergence, adapted for RL.
Whereas self-play trains an agent to be robust against \emph{itself}, PBRL with a sufficient number of opponents will force an agent to be robust against a wide range of strategies.
We conjecture that this will have similar benefit as adversarial training \cite{goodfellow2015} for classification models.
We evaluate PBRL as a defense in two simple two-player, zero-sum games.
We find that the self-play baseline can on average be exploited by an adversary using less than 60\% as many timesteps as self-play.
The PBRL trained policy is more robust: more timesteps are needed until an initial adversarial policy can be found.
We make three key contributions. First, we introduce PBRL as an end-to-end %
robust training method for deep RL. Second, we evaluate this empirically, decreasing exploitability of the victim. Third, we investigate how attributes of the environment (such as dimensionality) and algorithm (such as population size) influence robustness.
|
\section{Introduction}
Speech Emotion Recognition (SER) is a task aiming to understand the underlying emotional information in speech. In recent years, there has been a surge of interest in SER, because emotion conveys crucial information during human-computer interaction. However, two obstacles usually hinder the application of SER: the imbalance and scarcity issues of the SER dataset. We can further describe the issues from two perspectives.
Firstly, SER datasets have a common issue that the speech data distributions are non-uniform or highly skewed among different emotion classes. Speech utterances labeled as "Neutral" are much more frequent than those labeled as "Happy, Angry, etc.". The data scarcity of various emotion classes results in highly imbalanced datasets. To alleviate the data imbalance issue, a common way is to generate synthetic data through data augmentation techniques. Authors in \cite{Sahu2018OnES} use vanilla Generative Adversarial Networks (GANs) \cite{Goodfellow2014GenerativeAN} to generate synthetic openSMILE acoustic features \cite{Eyben2013RecentDI} for the training of the SER task. In \cite{Bao2019CycleGANBasedES}, the authors employ a CycleGAN \cite{Zhu2017UnpairedIT} to transfer openSMILE feature vectors from one emotion class to another. Instead of openSMILE features, some approaches do augmentation on low-level acoustic features (e.g. Mel-Spectrogram), which is an advantage for other tasks. Approaches in \cite{Rizos2020StarganFE, He2021AnIS} apply a starGAN \cite{Choi2018StarGANUG} to convert speech utterances of Neutral emotion to other emotions. Authors in \cite{Chatziagapi2019DataAU} modify a GAN to perform augmentation for imbalanced data and they empirically demonstrate that this GAN can improve the SER performance under imbalanced datasets.
Another important aspect to consider the scarcity issue is the availability of data for various existing spoken languages. There are very few data resources available for many languages, which is a significant barrier to the research and application of SER for these languages. One way to address this problem is by training an SER model on one or more languages with enough data (source), combined with very few training utterances of other languages (target). The idea is to use the knowledge gained from source languages to improve the performance on low-resource target languages \cite{Pan2010ASO}. Such SER training strategy that spans different languages, i.e. cross-lingual SER, has been widely studied \cite{Sagha2016CrossLS, Latif2018TransferLF, Goel2020CrossLC}. Although adding source languages improves the SER performance on target languages, there is still room to improve the cross-lingual SER task.
In this paper, We propose an augmentation strategy, which can be used to address the data imbalance and scarcity problem. One component of our augmentation approach is a GAN, whose generator is employed for augmentation. The generator is inspired by \cite{Tamkin2021ViewmakerNL}, where authors successfully perform automatic augmentation for multiple downstream tasks. To stabilize the GAN or improve its performance, additional auxiliary objectives are usually introduced during training \cite{Chen2019SelfSupervisedGV, Yamamoto2020ParallelWA, Jeong2021TrainingGW}. Thus, we apply a representation learner based on triplet loss to learn emotion representations from the original data. And the learned knowledge of the representation learner is used to increase and stabilize the augmentation performance for the SER task.
We summarize our contributions as follows: 1) We propose a GAN-based augmentation for the SER task. 2) We demonstrate that our augmentation approach can largely improve the SER performance with a highly imbalanced dataset (+8\% recall with our augmentation). 3) Moreover, for the cross-lingual SER task, our augmentation improves the performance for three different target languages with only around 50 utterances.
\section{Method}
The proposed SER data augmentation method is designed with two augmentation requirements: preserving the emotion class and providing rich variance. This means that an augmented utterance should have the same emotion label as the original one while also differing sufficiently from it. Our system setup and corresponding loss functions reflect these requirements.
\subsection{Model Components} \label{sec_Model_Description}
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{architecture.pdf}
\caption{\textbf{TOP}: The training pipeline of our GAN. \textbf{MIDDLE}: The training pipeline of our representation learner. \textbf{BOTTOM}: The loss to guide the augmentor to preserve the emotion information.}
\label{architecture}
\vspace{-0.3cm}
\end{figure}
Our approach consists of a GAN and a representation learner. The generator in our GAN augment the Mel-Spectrograms. The representation learner enables the GAN to present high augmentation variance and stabilize the augmentation.
The GAN is shown on the top of Fig. \ref{architecture}, it contains one generator and one discriminator, while we denote the inputted Mel-Spectrograms
as $\boldsymbol{X}$. In the rest of the paper, we refer to the generator as the augmentor since it is used to perform augmentation on Mel-Spectrograms.
The augmentor employs an encoder to map a Mel-Spectrogram $\boldsymbol{X}_1$ into a latent vector $z$, given a Gaussian noise $\boldsymbol{n}$. Then, $z$
is fed to the decoder to generate a tensor $P$ with the same shape as $\boldsymbol{X}_1$. To avoid excessive removal of features from the original Mel-Spectrogram,
we constrain the decoder output $P$ with an $l_1$-norm operation as used in \cite{Tamkin2021ViewmakerNL}. We then use a hyper-parameter $\epsilon$ to
control the augmentation intensity and obtain the augmented Mel-Spectrogram $\boldsymbol{\hat{X}}_1$. For brevity, in this paper, we denote our
augmentor as $AUG(\boldsymbol{X})$. Finally, we feed the discriminator $D(\boldsymbol{X})$ with $\boldsymbol{X}=\boldsymbol{\hat{X}}_1$ or $\boldsymbol{X}_2$, where $\boldsymbol{X}_2$ is another
sample from the original dataset. $D(\boldsymbol{X})$ aims to determine whether the input is augmented or not. Overall, we apply the non-saturating loss \cite{Goodfellow2014GenerativeAN} to train our GAN,
\begin{equation}
\mathcal{L}_{GAN} = \underset{AUG}{\text{min}} \underset{D}{\text{max}}
[
\text{log} D(\boldsymbol{X}) +
\text{log}( -D(AUG( \boldsymbol{X})) )
].
\label{loss_GAN}
\end{equation}
To achieve better augmentation, we further implement a representation learner. Its target is to learn meaningful emotion representations that can later be used to improve our augmentor. In contrast to a classification loss, a triplet loss is applied to train more discriminative emotion representations \cite{Harvill2019RetrievingSS, Huang2018SpeechER}. As illustrated in the middle of Fig. \ref{architecture}, our representation learner maps the Mel-Spectrogram $\boldsymbol{X}$ into an emotion representation vector $r$.
The objective is to attract the representations of $\boldsymbol{X}_{A}$ (Anchor) and $\boldsymbol{X}_{P}$ (Positive), whose emotion classes are the same, while repelling representations of $\boldsymbol{X}_{A}$ and $\boldsymbol{X}_{N}$ (Negative), since they are from different emotion classes. We define the triplet loss of our representation learner as,
\begin{equation}
\mathcal{L}_{REP} = \text{max}(\text{dist}(r_{A},r_{P}) - \text{dist}(r_{A},r_{N})+\beta, 0),
\label{loss_REP}
\end{equation}
where dist($\cdot,\cdot$) is $l_1$ distance, and $\beta$ is the margin of the triplet loss. The margin is a hyper-parameter that defines how far away the dissimilarities should be.
The representation learner is then employed to guide the augmentor to preserve the emotion class after the augmentation. Thus, we introduce the loss $\mathcal{L}_{EMO}$ to penalize the augmentor for generating an output that is of different emotion class to the input. The setup is shown at the bottom of Fig. \ref{architecture}. As we can see, we freeze the representation learner, and an augmented Mel-Spectrogram $\boldsymbol{\hat{X}}_{P}$ is used to supersede $\boldsymbol{X}_{P}$ as one of the inputs for the triplet loss. The motivation is to use the knowledge of the representation learner to guide the augmentor to preserve emotional information. We define the loss as,
\begin{equation}
\begin{multlined}
\mathcal{L}_{EMO} =
\text{max}(\text{dist}(r_{A}, \hat{r}_{P}) - \text{dist}(r_{A},r_{N})+\beta, 0).
\end{multlined}
\label{loss_EMO}
\end{equation}
Then, we define the model loss as,
\begin{equation}
\begin{multlined}
\mathcal{L}_{Model} =
w_{g}\mathcal{L}_{GAN} + w_{r}\mathcal{L}_{REP} + w_{e}\mathcal{L}_{EMO},
\end{multlined}
\label{loss_M}
\end{equation}
where $\mathcal{L}_{Model}$ are weighted by $w_{g}$, $w_{r}$ and $w_{e}$.
\subsection{Auxiliary Objectives for Augmentation} \label{sec_Auxiliary}
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth]{losses.pdf}
\caption{\textbf{TOP}: The auxiliary loss to guide the augmentor to provide augmentation variance. \textbf{BOTTOM}: The auxiliary loss to balance emotion preservation and augmentation variance.}
\label{losses}
\vspace{-0.4cm}
\end{figure}
To achieve better augmentation, we apply the representation learner to provide auxiliary objectives during training. We first expect the augmentor can provide augmentation variance. In other words, for the same Mel-Spectrogram, the augmentaor can produce various augmented versions. Otherwise, identical mapping or barely perceivable augmentation of the augmentaor might not be helpful for the SER task. Thus, as shown on the top of Fig. \ref{losses}, for the same input $\boldsymbol{X}$, we first obtain two augmented Mel-Spectrograms $\boldsymbol{\hat{X}_{aug1}}$ and $\boldsymbol{\hat{X}_{aug2}}$ based on different noise $n$ and augmentation intensity $\epsilon$. We then send these two augmented Mel-Spectrograms into the representation learner to acquire representations $r_{aug1}$ and $r_{aug2}$. We freeze the representation learner as well in order to utilize its knowledge of the original data. Our target is to repel these two representations, which can force the augmentor to generate various augmented versions whose representations are dissimilar, even from the same input. We define the auxiliary variance loss as,
\begin{equation}
\begin{multlined}
\mathcal{L}_{VAR} =
\text{dot}(\hat{r}_{aug1}, \hat{r}_{aug2}).
\end{multlined}
\label{loss_VAR}
\end{equation}
We enforce the dissimilarity by minimizing the dot product (\textup{dot($\cdot$, $\cdot$)}) between the representations from two augmented Mel-Spectrograms.
However, our initial experiments showed that $\mathcal{L}_{VAR}$ leads to unstable augmentation, we observe this issue by t-SNE visualisation \cite{van2008visualizing}. As shown on the left in Fig. \ref{tsne}, each point depicts the emotion representation $r$ derives from Mel-Spectrogram in the dataset IEMOCAP \cite{busso2008iemocap}. For each emotion class, we randomly pick 30 speech utterances and augment each sample 3 times with different noise and augmentation intensities. In the figure, the augmented representations from different emotions are not distinct and do not form clusters. The reason is that to meet the augmentation variance loss $\mathcal{L}_{VAR}$, the augmentor might perform arbitrarily to generate augmented versions whose representations are excessively far from original data, which means the emotion class information is damaged. Such behavior is against the loss $\mathcal{L}_{EMO}$ (Eq. \ref{loss_EMO}, aiming to preserve the emotion information), however, since the dataset is small, the augmentor is less penalized by the original data.
To enable our augmentor to preserve the emotion information while providing variant augmentations, we need to balance $\mathcal{L}_{EMO}$ (Eq. \ref{loss_EMO}) and $\mathcal{L}_{VAR}$ (Eq. \ref{loss_VAR}). Rather than doing tedious fine-tuning of the weights of these two losses, we achieve the balance by introducing another auxiliary loss as shown at the bottom of Fig. \ref{losses}. As we can see, another triplet loss is applied, but the positive and the negative inputs are both augmented. The motivation is to prevent the representations of the augmented Mel-Spectrograms from tangling together, by constraining the augmentor not only with original data, but also with augmented versions themselves. In other words, the augmented versions should be close to original versions if they belong to the same emotion, while be far from other augmented version when they from different emotion classes. Therefore, the triplet loss is defined as,
\begin{equation}
\mathcal{L}_{BAL} = \text{max}(\text{dist}(r_{A}, \hat{r}_{P}) - \text{dist}(r_{A},\hat{r}_{N})+\beta, 0).
\label{loss_BAL}
\end{equation}
As we can observe from the t-SNE visualisation on the right of Fig. \ref{tsne}, with the addition of the balancing loss, the augmented representations have a clear boundary between different classes. We conduct ablation study to demonstrate the contributions of our proposed losses further. Finally, we define the total loss as,
\begin{equation}
\mathcal{L}_{Total} = \mathcal{L}_{Model} + w_{v}\mathcal{L}_{VAR} + w_{b}\mathcal{L}_{BAL},
\label{loss_Total}
\end{equation}
where $w_{v}$ and $w_{b}$ are loss weights.
\subsection{Training Phases}
We repeat the following steps until our models converge:
\noindent
\textbf{1) Representation Learner:} We update our representation learner by $\mathcal{L}_{REP}$ (Eq. \ref{loss_REP}) from original data $\boldsymbol{X}$.
\noindent
\textbf{2) Discriminator:} We freeze the augmentor, and obtain augmented Mel-Spectrograms $\boldsymbol{\hat{X}}$ from original data $\boldsymbol{X}$. The discriminator $D$ is updated by $\boldsymbol{X}$ and $\boldsymbol{\hat{X}}$.
\noindent
\textbf{3) Augmentation Variance:} We freeze the representation learner, and utilize it to enhance the augmentation variance of the augmentor, by the loss $\mathcal{L}_{VAR}$ (Eq. \ref{loss_VAR}).
\noindent
\textbf{4) Emotion Preservation and Balance:} We update the augmentor. We freeze the representation learner, the discriminator $D$, then generate augmented Mel-Spectrograms $\boldsymbol{\hat{X}}$ from the augmentor. The augmentor is finally updated from the discriminator loss, $\mathcal{L}_{EMO}$ (Eq. \ref{loss_EMO}) and $\mathcal{L}_{BAL}$ (Eq. \ref{loss_BAL}).
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{tsne.jpg}
\caption{T-SNE visualization of the emotion representations from the representation learner. \textbf{LEFT}: the augmented utterances are from the augmentor trained without $\mathcal{L}_{BAL}$; \textbf{RIGHT}: the augmented utterances are from the augmentor trained with $\mathcal{L}_{BAL}$, which leads to a more clear cluster boundary.}
\label{tsne}
\vspace{-0.3cm}
\end{figure}
\subsection{Implementation Details}
In our augmentor, the encoder consists of 3 convolutional layers with LeakyReLU activation function and maps the input $\boldsymbol{X}$ into a latent 128D vector $z$. The decoder employs two deconvolutional layers with ReLU activation function to convert $z$ into a tensor $\boldsymbol{P}$ whose shape is the same as the input. The discriminator first uses four convolutional layers with the LeakyReLU activation function to process the input, then uses an LSTM and an attention layer \cite{Zhang2019AttentionaugmentedEM} to discern whether the input is original or augmented. Our representation learner employs 5 convolutional layers with LeakyReLU and an LSTM to process the Mel-Spectrogram, then the same attention layer in the discriminator is used to produce the 128 dimensional emotion representation.
For augmentation intensity $\epsilon$, we randomly sample $\epsilon$ from a uniform distribution $\text{U}(0.05, 0.3)$, because we found a random $\epsilon$ performs better than a fixed one as in \cite{Tamkin2021ViewmakerNL}. We set the $\beta$ in $\mathcal{L}_{REP}$ (Eq. \ref{loss_REP}), $\mathcal{L}_{EMO}$ (Eq. \ref{loss_EMO}) and $\mathcal{L}_{BAL}$ (Eq. \ref{loss_BAL}) as 7. For loss weights, we set $w_{g}$, $w_{r}$ and $w_{v}$ to 1, $w_{e}$ to 10 and $w_{b}$ to 8. We train all of our models with Adam optimizer for 30k iterations, and set the learning rate to 1e-6.
\section{Experiment and Results}
The proposed augmentation method is demonstrated with SER tasks. In short, the results show that: 1) Our augmentor can improve the SER with a highly imbalanced dataset. 2) Our augmentor can improve the cross-lingual SER task with very few training utterances from the target language. 3) the importance of the contribution of $\mathcal{L}_{VAR}$ (Eq. \ref{loss_VAR}, for adding augmentation variance), and $\mathcal{L}_{BAL}$ (Eq. \ref{loss_BAL}, for balancing the emotion preservation and providing variance).
\subsection{Experimental Setup}
\subsubsection{Data Preprocessing and SER Classifier}
In all our experiments, we use the same acoustic features and same emotion classifier as in \cite{Chatziagapi2019DataAU}. To extract Mel-Spectrograms from waveform files, we use a 50-millisecond window and a 50 percent overlap ratio, with 128 Mel Coefficients. A VGG19 architecture \cite{Simonyan2015VeryDC} is employed as our emotion classifier, which takes a fix-sized 128 $\times$ 128 Mel-Spectrogram (128 frames) as the input and predicts the emotion class. During the evaluation, the predicted class is determined by the majority voting of segments for each Mel-Spectrogram. To train our GAN and representation learner, we use 512 $\times$ 128 Mel-Spectrograms.
\subsubsection{Datasets}
\noindent
\textbf{Imbalanced SER} and \textbf{Ablation Study:} For the experiments of the Imbalanced SER (Sec. \ref{sec_imbalanced}) and Ablation Study (Sec. \ref{sec_ablation}) We train and test our models on IEMOCAP \cite{Busso2008IEMOCAPIE}. Like many other SER works \cite{Sahu2018OnES, Chatziagapi2019DataAU, Latif2020AugmentingGA}, we only focus on 4 classes (Angry, Sad, Neutral and Happy), resulting in 5531 speech utterances of about 7 hours total duration. Furthermore, in order to simulate the data imbalance issue, we follow \cite{Chatziagapi2019DataAU} and randomly remove 80\% of each class except Neutral. Since IEMOCAP is originally split into 5 sessions, we conduct 5 fold cross-validation, by using 4 sessions for training and 1 for testing. We train our SER classifier with original and augmented utterances. We augment each original training sample four times, resulting in a hybrid original-augmented training dataset that is roughly the same size as the original non-reduced dataset.
\noindent
\textbf{Cross-Lingual SER:} For the experiment of the Cross-Lingual SER (Sec. \ref{sec_cross}), our datasets are listed in Tab. \ref{Cross-lingual Datasdet}. These datasets span four languages, with IEMOCAP and ESD serving as source languages and the remaining datasets serving as target languages. Although ESD originally has both Chinese and English speech utterances, we only use the Chinese portion in this experiment. To create IEMOCAP\_SUB and ESD\_SUB, we randomly selected utterances from IEMOCAP and ESD, in order to make all of the target language datasets the same size. Since each dataset has different class numbers, one of the consistent ways to investigate cross-lingual SER task is by considering the binary positive/negative valence classification problem, like in many other related works \cite{Sagha2016CrossLS, Latif2018TransferLF}. We follow the same method to map emotions into valance, and split 25\% of the target language datasets for our evaluation. During the training of the SER classifier, we augment each original training sample from the target language 20 times, since the target language's data is severely lacking (only 50 utterances in some experiment cases).
\vspace{-0.3cm}
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{cross_lingual.pdf}
\caption{UAR results of cross-lingual SER. \textbf{TOP}: English (source) to 3 target languages. \textbf{BOTTOM}: Chinese (source) to 3 target languages. }
\label{cross_lingual}
\vspace{-0.3cm}
\end{figure}
\subsection{Results}
\subsubsection{SER on Imbalanced Dataset} \label{sec_imbalanced}
\noindent
\textbf{Models}: We follow \cite{Chatziagapi2019DataAU} and train our SER classifier with: \texttt{NoAUG}: the 80\% reduced dataset; \texttt{AUG}: the 80\% reduced dataset and the augmented utterances.
\noindent
\textbf{Results}: We list the Unweighted Average Recall (UAR) results reported by \cite{Chatziagapi2019DataAU} and ours in Tab. \ref{Imbalanced_Result}. One thing to note is that, although using the identical setup in \cite{Chatziagapi2019DataAU} to train our \texttt{NoAUG} model, the results (the 2nd and 4th rows) differ because the 80\% removed component is randomly chosen. The average UAR results, on the other hand, are close, implying that our comparison is fair. By comparing the two \texttt{AUG} models, we can see that our augmentation strategy is roughly 5\% higher than the augmentation approach in \cite{Chatziagapi2019DataAU}, which means a better SER performance can be achieved with our augmentation approach.
\subsubsection{Cross-Lingual SER} \label{sec_cross}
\noindent
\textbf{Models}: We train our SER classifier with: Low Target (\texttt{LT}), 100\% of the source language data but only 10\% of the target language dataset; \texttt{LT\_AUG}, same as \texttt{LT}, but the augmented data is also fed during training; Full Target (\texttt{FT}), 100\% of the source language data and 75\% of the target; \texttt{FT\_AUG}, same as \texttt{FT}, but we add the augmented data from our augmentor.
\noindent
\textbf{Results}: The UAR results on 25\% of the target language datasets are shown in Fig. \ref{cross_lingual}. By comparing with \texttt{LT} and \texttt{LT\_AUG}, our augmentation approach increases UAR performance by around 5\%, given only 10\% of the target language data. The performance can be further improved with the augmented data, when the target language data is additionally given, based on the results between \texttt{FT} and \texttt{FT\_AUG}.
\begin{table}[t]
\small
\caption{Cross-lingual datasets. IEMOCAP and ESD are used as source languages, the others serve as target languages.}
\label{Cross-lingual Datasdet}
\centering
\begin{tabular}{ c | c | c | c }
\toprule
\multirow{2}*{Dataset} & \multirow{2}*{Language} &\multicolumn{2}{c}{Valance} \\
\cline{3-4}
& & Negative & Positive \\
\midrule
IEMOCAP \cite{Busso2008IEMOCAPIE} & English & 2187 & 3344 \\
ESD \cite{Zhou2021EmotionalVC} & Chinese & 7000 & 7000 \\
\midrule
EMO-DB \cite{Burkhardt2005ADO} & German & 385 & 150 \\
EMOVO \cite{Costantini2014EMOVOCA} & Italian & 336 &252 \\
IEMOCAP\_SUB & English & 250 & 250 \\
ESD\_SUB & Chinese & 250 & 250 \\
\bottomrule
\end{tabular}
\vspace{-.3cm}
\end{table}
\begin{table}[t]
\small
\caption{UAR results on the 80\%-reduced imbalanced dataset.}
\label{Imbalanced_Result}
\centering
\begin{tabular}{ c | c | c | c | c }
\toprule
& Angry & Sad & Happy & Average \\
\midrule
NoAUG \cite{Chatziagapi2019DataAU} & 47.8 & 46.9 & 52.2 & 49.0 \\
AUG \cite{Chatziagapi2019DataAU} & 53.5 & 52.1 &55.2 &53.6 \\
\midrule
NoAUG (Ours) &54.23 &49.03 & 47.02 & 50.09 \\
\textbf{AUG (Ours)} & \textbf{62.75} &\textbf{58.11} &\textbf{53.71} &\textbf{58.19} \\
\bottomrule
\end{tabular}
\vspace{-0.2cm}
\end{table}
\begin{table}[t!]
\small
\caption{Ablation study results (UAR) of the contribution of $\mathcal{L}_{VAR}$ and $\mathcal{L}_{BAL}$. }
\label{ablation}
\centering
\begin{tabular}{ c | c | c | c | c }
\toprule
Model & Angry & Sad &Happy & Average\\
\midrule
NoAUG &54.23 &49.03 & 47.02 & 50.09 \\
\midrule
$\mathcal{L}_{Model}$ & 56.03 & 48.43 & 50.71 &51.52 \\
$\mathcal{L}_{Model}+\mathcal{L}_{VAR}$ & 36.47 & 46.61 & 45.31 &42.80 \\
$\mathcal{L}_{Total}$ & \textbf{62.75} &\textbf{58.11} &\textbf{53.71} &\textbf{58.19} \\
\bottomrule
\end{tabular}
\vspace{-0.2cm}
\end{table}
\subsubsection{Ablation Study} \label{sec_ablation}
\noindent
\textbf{Models}: For ablation study, we use the $80\%$ reduced dataset to train an SER classifier with and without augmentation. To train the augmentor, we consider three cases: 1) with only $\mathcal{L}_{Model}$; 2) with $\mathcal{L}_{Model}$ and auxiliary variance loss $\mathcal{L}_{VAR}$; 3) with the total loss that consists of both auxiliary terms $\mathcal{L}_{VAR}$ and $\mathcal{L}_{BAL}$.
\noindent
\textbf{Results}: The UAR results are listed in Tab. \ref{ablation}. We can observe that without $\mathcal{L}_{VAR}$ and $\mathcal{L}_{BAL}$, performance can only improve little because the assistance of an augmentor with low augmentation variance is limited. After we apply the loss $\mathcal{L}_{VAR}$, as we mentioned in Sec. \ref{sec_Auxiliary}, The augmentor arbitrarily performs augmentation and destroys the emotion information, resulting in a significant fall in performance. Lastly, by introducing $\mathcal{L}_{BAL}$, we can obtain the best result by balancing augmentation variance and emotion preservation.
\section{Conclusion}
This work presents a GAN-based augmentation approach to alleviate the data imbalance and scarcity issue for the SER task. Specifically, we conduct experiments and demonstrate: 1) Even with a severely imbalanced dataset, our augmentation approach can significantly increase SER performance. 2) With only about 50 training utterances of the target languages provided, our augmentation approach considerably enhances SER performance for these low-resource languages.
\bibliographystyle{IEEEtran}
|
\subsection{Industrial Control Systems}
Industrial Contol Systems (ICSs) are widely adopted in industrial sectors and critical infrastructures, such as power grids and nuclear plants. They are task-critical implementations of Cyber-Physical Systems (CPSs) to operate and automate the industrial process and consist of numerous subsystems such as SCADA systems and distributed control systems (DCS), field controllers such as Programmable Logic Controllers (PLC) and human-machine interface (HMI), physical plants such as power generators and actuators, and other intellectual electrical devices (IED) \cite{stouffer2011guide,asghar2019cybersecurity,bhamare2020cybersecurity,smidts2019next}. All physical components are organized into a hierarchical, multi-level network and are connected by the fieldbus or wireless networks to perform specific control tasks. Data transmission between levels of the network ensures real-time surveillance for security and control purposes. Fig.~\ref{fig:ics_illustration} shows a notional integrated ICS architecture in the power domain based on the commonly adopted Purdue Model \cite{williams1994purdue}.
\begin{figure}
\centering
\includegraphics[scale=0.3]{plots/ics_2.png}
\caption{An illustrative example of an ICS network of power systems. Levels 0-2 constitute the OT level for real task execution. Levels 3-5 constitute the IT level for information processing such as scheduling. All components in the ICS cooperate together to deliver electricity to support end users.
An attacker can follow the red path to attack a physical plant from an enterprise computer. AI stacks in IT and OT levels are introduced to augment the learning capabilities of IT and OT subsystems and improve the overall ICS resilience.}
\label{fig:ics_illustration}
\end{figure}
An ICS typically consists of several field sites (levels 0-2) that are functional points and can execute real-time tasks. A field site is generally equipped with local PLC controllers and HMIs for automated and human control, some IEDs such as sensors for measurements, and physical plants for specific tasks. Fieldbus connects all components to the cyber components and the controller center.
Examples of field sites include wind turbines and gas power plants.
The fieldbus, such as the CAN bus, is time-critical and fault-tolerant since any time delay and signal distortion can lead to severe faults of field sites. Besides, the fieldbus requires a large communication capacity due to the increasing scale and complexity of contemporary ICSs. For security considerations, the data transported by fieldbus should be filtered by firewalls before sending to other levels.
The control center (level 3) is a supervisory level designed for real-time system monitoring, control configurations to filed sites, and necessary human operations and interventions. At this level, HMIs provide process information and status to human operators. Control and data servers monitor and adjust control processes and store historical information. Engineering stations provide the developing and debugging environments for control operations.
The enterprise zone (level 5) is a business network performing general tasks, such as human resource management and internal communications. Data transmission between the controller center and the enterprise zone is restricted and inspected by the demilitarized zone (DMZ) (sometimes called a perimeter network). A DMZ is a subnetwork acting as an intermediary for connected security devices, which is more than a firewall. DMZ ensures that the data can only be transmitted from a high-security level to a low-security level, and the reverse is prohibited. This is because the data in the controller station is more critical and sensitive. The DMZ filters redundant data traffic from the enterprise zone to the controller center that interrupts the control process. The DMZ also adds another layer of security to the ICS. An attacker cannot bypass the DMZ unless he has direct access to the equipment within the DMZ. It prevents malicious access and attacks on the controller and physical plants from cyberspace, providing another layer of security.
The enterprise zone can generally access the Internet or other wide area networks (WANs) (level 5) through another DMZ or a firewall, depending on the security configuration. Other remote field sites can also be connected to the control center via the Internet or WANs from outside. Different security protocols are required for such remote connections.
Following the adaption of the Purdue Model, the network can be further grouped into IT and IT levels based on the functionalities. Levels 0-2 constitute the OT level and are responsible for real-time task execution such as wind turbine control. Levels 3-5 form the IT level and take charge of information processing-related work such as task scheduling and management.
Upon the classic ICS architecture, we introduce AI stacks at IT and OT levels to enable computational intelligence. The AI stack at the IT level enables real-time intelligent cyber detection and response for cyber attacks based on network data, assisting the ICS to achieve cyber resilience. The AI stack at the OT level empowers learning-based methods for controlling physical plants. Compared with traditional model-based control, the AI-enabled learning methods are particularly convenient for OT systems with complex dynamics. The AI stack also facilitates the development of real-time adaptive strategies for anomalies and disturbances. We mention that we do not specifically portray the AI components in Fig.~\ref{fig:ics_illustration} because the AI stack and computational intelligence can be algorithmically built into many digital devices such as working stations and local controllers. As its scale becomes larger and the structure more complex, the integration of AI stacks benefits the modern-day ICS to operate more adaptively and effectively against cyber attacks and unknown disturbances than traditional ones.
\subsection{Multi-Agent System Framework of Industrial Control Systems}
The classic Purdue model provides a clear visualization of a hierarchical and monolithic ICS architecture. However, it also has some limitations to achieving resilient ICS design as ICSs are becoming more integrated. On the one hand, the growing monolithic ICS architecture makes the network analysis intractable, making it challenging to design effective and resilient control strategies to cope with cyber attacks and external disturbances. On the other hand, more specialized subsystems are integrated into modern-day ICSs so that all components operate in a distributed fashion.
Therefore, we can decompose the growing ICS into a set of subsystems based on functionalities. The subsystems cooperate to accomplish the overall ICS mission. For example, in the power system, the generating plants work as individual subsystems to produce power, while the delivery grid act as another subsystem to deliver the power to end users.
Within each subsystem, we also have the cyber-physical architecture. Therefore, the IT and OT division in the Purdue model in Fig.~\ref{fig:ics_illustration} is still applicable for each subsystem. We refer to the divided systems as IT and OT systems, as shown in Fig.~\ref{fig:intro}. The IT and OT systems affect and control the subsystem in a distributed but interdependent fashion. For example, in a generating subsystem, the IT system can monitor real-time information such as generating capacity and output to the grid for power management and coordination. The OT system stabilizes the generating plants and produces power based on the requirements from the IT system.
Therefore, we propose a MAS framework to characterize modern-day ICSs using concepts in distributed control systems. The MAS framework has a two-level interpretation. At the system level, each functional subsystem is regarded as an intelligent agent with a specific objective. All agents interact and cooperate to accomplish the system-wide mission, such as power supply in the infrastructure. At the subsystem level, the IT and OT systems act as intelligent agents, and their interactions lead to achieving the subsystem-specific task objective. For example, a generating subsystem performs a smooth power generation under external requirement changes or malicious attacks on the generating plants.
The two-level MAS framework is intrinsically interdependent. The interactions at the subsystem level ensure the regular operation of the subsystem and lay the foundation for system-level cooperation. The interaction and coordination at the system level achieve the system-wide ICS mission.
Our MAS framework also enables learning for resilient ICS design and lays the foundation to achieve MAR. On the one hand, every intelligent agent (subsystem and IT or OT system in the subsystem) can utilize network data and learning to adapt to new attacks and disturbances, which provides a flexible and real-time adaption mechanism to achieve resilience. On the other hand, every agent can learn to achieve self-adaption and cooperate with others when the agent has complex dynamics and the interdependencies between agents are not known or planned for. We discuss more details on learning for resilience in Section \ref{sec:learning}.
Our multi-agent perspective aligns with the literature. For example, a hierarchical MAS framework (Hierarchical Multi-agent Dynamic System, HMADS) has been proposed in \cite{rieger2013resilient,rieger2013hierarchical} to characterize resilient control systems. The framework consists of three layers based on high-level functionality: the management layer, the coordination layer, and the execution layer. The management layer provides the high-level system objective and task scheduling. The coordination layer allocates resources to system components to align with the system objective. The execution layer takes charge of controlling and sensing field devices. A subsystem or a controller in the control system is treated as an intelligent agent and is divided into one of three layers based on its sphere of influence. Our framework absorbs the management and coordination layers into the IT, and the execution layer is subsumed under the OT.
It is clear that a divide-and-conquer approach plays an essential role in the architecture of the large-scale complex ICS. In addition to the focus on IT and OT in this chapter, human technologies (HT) are becoming increasingly prominent. They are central to the resilience of the ICSs as many attacks start with exploiting human error and vulnerabilities. Human-machine technologies have been proposed to augment human performances, e.g., securing human operators from attention failures \cite{huang2022radams}, employees from phishing \cite{huang2022advert}, and users from noncompliance \cite{huang2022zetar,casey2016compliance}. It is natural to extend our framework to include an HT layer to incorporate the emerging human solutions and their interface with IT-OT.
The hierarchical divide-and-conquer view of ICSs provides a flexible and evolving perspective that enables integrating the systems with new emerging interfaces and technologies. For example, the adoption of IoT technologies and human-machine teaming capabilities can potentially augment ICSs with new cross-layer functionalities and introduce additional dimensions to the ICS architecture.
\subsection{Security Challenges}
The interconnected IT and OT components can create potential vulnerabilities to degrade ICS performance and jeopardize the security of infrastructures. As shown by the red path in Fig.~\ref{fig:ics_illustration}, an attacker can penetrate the ICSs by compromising an enterprise computer and moving laterally within the ICS network to reach the target asset. For instance, an attacker can bring down a nuclear power plant by manipulating the sensors. We have witnessed an increasing number of such attacks, including ransomware attacks, insider threats, and APT attacks, which have raised serious security threats to ICSs \cite{makrakis2021industrial}. We summarize the critical security challenges of ICSs that remain open issues.
First, cascading failures can devastate ICSs due to the multi-level structure and the multi-agent characteristics. The cascading connections of different components provide numerous possible fault-error-failure propagation paths. A single fault of any component, such as hardware flaws or software bugs, can propagate in the network and consequently affect other components, which eventually leads to a system-wide functionality failure. These threats have been noticed and investigated in \cite{zhu2012dynamic,rieger2019industrial}. Cascading failures also make it hard to detect and diagnose the real anomaly in ICSs because of fault propagation. Fixing a single component's error is insufficient to address the general security issues of ICSs. More precise and complex fault detection and identification are required to pinpoint the initial fault and prevent further failures.
Second, the large-scale connections and heterogeneous components in contemporary ICSs lead to intrinsic complex structures, further increasing the diversity and the complexity of security vulnerabilities. Common security threats have been identified and categorized in \cite{cardenas2011attacks} and precautions can be designed to prevent the existing threats. However, many unknown factors can still cause system-wide malfunctions and jeopardize ICS security. Exploiting and identifying vulnerabilities for modern-day ICSs is still an open project. Besides, the attack surface of subnetworks in ICSs also differs from their characteristics and tasks. For example, generating plants and power grids have different network architectures and security vulnerabilities. The device compatibility is also a security concern in ICSs. It is common to observe obsolete devices in ICSs due to historical reasons. These components are generally not designed for contemporary security considerations and leave potential vulnerabilities. It is challenging to design effective and holistic strategies to make different versions of devices work together and provide enough security guarantees.
We identify cyber attacks as the third security challenge. A significant feature of contemporary ICSs compared with traditional ones is the adoption of the cyber layer. However, cyber attacks nowadays are causing a broad and profound impact on ICSs. While the cyber layer achieves efficient communications among different components and enables system automation, it also brings more security threats. The existing and recognized cyber threats include communication disruptions and malware attacks such as computer viruses and ransomware. Works such as \cite{uma2013survey,formby2017out} also summarized common cyber attacks and malware in ICSs.
Although cyber attacks are not originally designed to sabotage ICSs, several recent attacks on ICSs, including the Colonial Pipeline ransomware attack \cite{colonial2021wiki}, have shown their destructive power on infrastructures and economics. Cyber attacks also bring another level of security consideration: intelligent adversaries. Attacks can start malicious and intelligent attacks on ICSs through cyber layers, which differ from the security challenges caused by device faults. Advanced persistent threats (APTs) are an example of intelligent attacks. APT is a cyber attack that aims to penetrate the network and attack a high-value asset. It searches possible attack paths to the target assets stealthily and starts the attack when the time is perfect. APT can avoid existing security defense mechanisms strategically. So traditional methods such as routine security checks may have limited effects on APTs. Although it appears harmless for other components during the network penetration, the attack consequence is devastated because of the loss of the target asset. In short, it is critical to address the cyber threats in ICSs to maintain system resilience.
Human factors also pose security challenges in ICSs. ICSs are integrated with HMIs and workstations for human operators (or experts) to monitor the system status and debug errors. Although the contemporary ICS design keeps improving the human friendliness in operation and facilitating decision support, human operators are still indispensable. For example, human operators are required when processing emergencies such as generating plant breakdown. Besides, human operators are also capable of diagnosing complex errors such as new cyber attacks and faults of specialized devices. Therefore, human experts and their knowledge can benefit ICSs to achieve better performance and resilience. However, security threats also accompany humans and hence jeopardize ICSs.
For example, humans have limited attention. When cascading errors occur, it can be difficult for a human operator to recognize the relevant and irrelevant threats. Also, cyber attacks can launch feint attacks to distract and deceive human attention \cite{hitzel2019art,huang2021radams}. Humans are also prone to be manipulated. There have been phishing scams and espionage activities to steal critical ICS information or plant malware. Some activities can be hard to recognize. Last but not least, insider threats also significantly endanger the security of ICSs. Some of these attacks leverage human operators' access to the core system to steal data or inject malware to sabotage plants. According to empirical research \cite{homeland2016recommended}, social engineering scams and insider threats have become the primary threats to ICSs. Therefore, training and regulating human behaviors to have more efficient and user-friendly ICSs is vital and remains an open challenge.
The security challenges raise severe threats to ICSs such as power grids and nuclear plants, leading to devastating consequences. Therefore, it is critical and indispensable to improve the resilience of ICSs, so that the system can resist malicious attacks and disturbances and maintain an acceptable level of operational normalcy. Traditional threat detection and mitigation are not sufficient to deal with the complexity of modern-day ICSs and the new security challenges. The integration of AI stacks into the ICSs can augment the distributed learning capabilities of the subsystems and improve the overall ICS's resilience.
\subsection{Learning in Cyber IT System}
In the IT system, a learning agent monitors the cyber behaviors of the users and the network to detect the anomaly and adversarial behaviors. The goal of the IT operation is to mitigate the impact of the attack, which can propagate to the OT if not carefully handled. The learning process in this system depends on the sophistication of attacks and resources of the decision-maker, including his knowledge, computation, and maneuverability.
Based on the functionality, we can summarize the components in the feedback loop in Fig.~\ref{fig:ics_learning} into two modules: cyber detection and cyber mitigation. Two modules work jointly to enable cyber resilience.
\subsubsection{Learning for Cyber Detection}
Cyber detection is fundamental for cyber resilience. An AI agent can learn from cyber data such as network traffic and log traces to detect and recognize cyber threats, including internal anomalies and external attacks on cyber agents and human operators. Based on the learning pattern, the learning methods for cyber detection can be categorized as supervised learning, unsupervised learning, and semi-supervised learning \cite{mitchell2014survey}. Each type of learning provides different advantages to achieving successful detection.
\emph{Supervised learning} uses cyber data and labeled system behaviors (for example, normal or abnormal) to train a cyber detection model offline and perform detection online.
Many works have developed successful learning-based detection models based on different supervised learning techniques. For example,
in \cite{linda2009neural}, Linda et al. used neural networks (NNs) to detect cyber intrusions for critical infrastructures.
Their neural network model was trained by actual network data and showed a perfect detection rate and zero false positives on testing cyber attacks.
In \cite{wang2019detection}, Wang et al. adopted random forest (RF) to detect natural disturbances and artificial cyber attacks for power systems. The model was trained by historical network data and relevant log information. The experiments on an open-source simulated power system dataset reported a successful detection rate higher than 93\%.
Mokhtari et al. in \cite{mokhtari2021machine} used $K$-nearest neighbors (KNNs) and decision tree classifier (DTC) as a tool for cyber detection. The data collected by the SCADA system were used for detection model training, and their reported detection precision and accuracy were higher than 97\%.
Many works such as \cite{beaver2013evaluation,hink2014machine,ozay2015machine} also made a thorough comparison of different supervised learning methods for cyber detection and demonstrated the effectiveness of the learning-based methods.
Supervised learning is also used for human-related attack detection. In the work of \cite{alam2020phishing}, an RF and DTC-based model was developed to detect phishing attacks and help human operators stay safe. An NN-based detection model was also proposed in \cite{abdelaty2021daics} to reduce false alarms in ICSs to ease human operators' cognitive load.
Supervised learning enjoys a high cyber detection precision and accuracy as we have observed in the literature. However, an AI agent can hardly detect new anomalies and attacks with supervised learning because of missing labels in the dataset. Unless updating detection models constantly, an AI agent may not detect zero-day attacks and fail to provide cyber resilience for the system. Besides, acquiring labeled cyber data is often a challenge.
As a complementary approach, \emph{unsupervised learning} can handle unlabeled cyber data and provide more generic detection. The main idea behind unsupervised learning is feature-based clustering. A learning agent first identifies the features of cyber data and then searches for potential anomalies or attacks by clustering the featured data. Many recent works have focused on developing detection tools with unsupervised learning.
For example, Maglaras and Jiang in \cite{maglaras2014intrusion} adopted the One-Class Support Vector Machine (OCSVM) to detect malicious network traffic in SCADA systems. The detector extracted attributes from network traces and then was trained to classify the attacked nodes that jammed the network traffic.
Alves et al. in \cite{alves2018embedding} used K-means clustering to detect data interception, data injection attacks, and denial of service (DoS) in industrial PLC networks. The anomaly classification and learning were based on the features of real-time data streams, such as pack latency and processing information.
Kiss et al. in \cite{kiss2014data} investigated detection in critical infrastructures based on K-means and MapReduce paradigm to achieve real-time cyber attack detection.
To address the anomalous behavior detection of building energy management systems, Wijayasekara et al. in \cite{linda2012computational,wijayasekara2014mining} developed an efficient detection method based on the modified nearest neighbor clustering and fuzzy logic, which not only performed faster than traditional alarm-based detection methods, but also was able to provide linguistic interpretation of the identified anomalies.
Unsupervised learning may produce more false alarms in detection due to a lack of labels compared with supervised learning \cite{umer2022machine}. Human intervention for diagnosis is useful, but it may delay the IT operation and result in less cyber resilience in the IT system. To overcome the issue, many works consolidate supervised and unsupervised learning to create \emph{semi-supervised learning} for effective cyber detection.
In recent work \cite{kravchik2018detecting}, Kravchik and Shabtai developed a window-based cyber attack detection using convolutional neural networks (CNN) for critical water infrastructures. The CNN was trained to predict normal system behaviors, and anomalies were detected every time window based on the error between predicted and observed system behavior.
Marino et al. in \cite{marino2019cyber} adopted unsupervised learning for ICS anomaly detection classification and evaluated its performance with supervised learning. The combination of the two learning methods empowered the detection of unseen cyber anomalies and improved the detection accuracy.
We refer the readers to recent surveys \cite{umer2022machine,bhamare2020cybersecurity,anthi2021adversarial,handa2019machine} for details on machine learning algorithms for cyber detection in ICSs.
\subsubsection{Learning for Cyber Mitigation}
The cyber mitigation module serves as the decision-making and operation units in the cyber learning architecture and is the key to guaranteeing cyber resilience. However, learning for cyber mitigation is more challenging because the mitigation strategies rely on tasks and network structures. Traditional cyber mitigation strategies do not pay attention to the higher-order impact of the mitigation strategies on the system performance and the complex roles of attackers who can strategically and stealthily counteract the system strategies. To address these challenges, game-theoretic learning and reinforcement learning (RL) are used for effective and proactive cyber mitigation.
\emph{Game-theoretic learning} treats the attacker as a rational player and learns a strategic solution by observing the attacker's attack trace. It provides a suitable framework to develop attack-aware solutions to combat APTs in the IT system, where attackers aim to penetrate the IT system to attack the target asset \cite{rubio2019current}. For example, Huang and Zhu in \cite{huang2018analysis,huang2019adaptive,huang2020dynamic} proposed a Bayesian learning mechanism to defend APTs in the cyber system proactively. They used a multi-stage dynamic Bayesian game to characterize the long-term interaction between a stealthy attacker and a proactive defender. The Bayesian learning was used to update the defender's knowledge of the attacker along the interaction and assisted the defender in developing effective defense strategies for cyber resilience.
Game-theoretic learning has also been used in defensive deception to achieve effective cyber mitigation. A learning agent can proactively learn to deceive the attacker with false targets or steer the attacker to a non-hazard zone to mitigate attack consequences \cite{zhang2020game,huang2020strategic,zhu2021survey,pawlick2021game}.
A \emph{reinforcement learning} agent seeks a strategy that minimizes the attack consequence over some time from observed data. For example, in \cite{huang2019honeypot}, the network defender used RL to adaptively allocate honeypot resources to trap the attacker in the target honeypot for as long as possible.
RL has also been shown to be effective for human-related cyber attacks such as phishing and attention manipulation. For example, Huang and Zhu in \cite{huang2021advert} proposed a phishing prevention mechanism (ADVERT) using RL to generate adaptive visual aids to counteract inattention and improve the human recognition of phishing attacks. Experiments with human volunteers showed an accuracy improvement of phishing recognition from 74\% to a minimum of 86\%. Another human attention management mechanism (RADAMS) using RL was proposed in \cite{huang2021radams} to help human operators fight against informational attacks, such as fake alerts. The proposed mechanism used RL to de-emphasize alerts selectively and reduce human operators' cognitive load. Experimental results showed that the developed attention management mechanism reduced the risk by as much as 20\% compared with default strategies.
\subsection{Learning in Physical OT System}
In the OT system, the goal of a learning agent is to create a learning-based resilient control mechanism that can further mitigate the attack impact on the OT system when the attacker reaches the OT \cite{homeland2016recommended}. The OT learning mechanism also follows the feedback architecture depicted in Fig.~\ref{fig:ics_learning}. The monitoring of the OT assets, including sensors, actuators, and terminal units, creates situational awareness for the decision-makers to adaptively reconfigure the control system to threats and environmental changes.
The learning agent can either learn the uncertainties first and then make decisions or directly learn the control decisions.
Unlike the OT system, physical plants in the OT system resort to physical models to operate normally. The controller design for physical agents is also model-specific. Therefore, effective and resilient control in the OT system relies on the physical models.
However, exact physical models can be hard or impossible to obtain as contemporary ICSs become more integrated, making resilient control even more challenging. Therefore, many research resorts supervised learning to estimate the physical dynamics and then develop adaptive control strategies for possible OT system attacks. For example, Li and Zhao in \cite{li2021resilient} studied learning-based resilient and adaptive control strategies for uncertain deception attacks in CPSs. The neural network was used to approximate unknown nonlinear and switching physical system dynamics. Then a dynamic surface-based adaptive controller was designed for sensor and actuator deception attacks. The resilient controller was validated on continuously stirred tank reactor systems.
Liu et al. in \cite{liu2021adaptive} proposed observer-based adaptive neural network control for nonlinear CPSs subject to false data injection attacks. The neural network was used to adaptively approximate the unknown nonlinear functions of the CPS. Then a resilient feedback controller was designed to ensure bounded output under malicious attacks.
The model-free adaptive controller design for unknown dynamics was also studied in \cite{wang2017neural,farivar2019artificial}. Both work used supervised learning and neural network models to develop adaptive control strategies to cope with sensor faults and malicious attacks. The system resilience was guaranteed by the stabilizable controller and bounded system output.
Supervised learning has been useful for predicting physical models. However, dealing with malicious and intelligent attacks in the OT system is insufficient. To address the issue, game-theoretic learning is used to find control strategies for physical resilience because of its ability to characterize strategic interactions with adversaries. A learning agent learns to reconfigure the control system by anticipating the attacker's behavior. An example of game-theoretic learning in the OT system is moving target defense (MTD) in ICSs. The defender protects a multi-layer network system and prevents the attacker's penetration by changing the network configuration of each layer. The attacker exploits each layer's vulnerabilities and tries to penetrate the entire system. Zhu and Ba{\c{s}}ar in \cite{zhu2013game} investigated this issue by assuming two players have no knowledge of each other. They proposed an iterative game-theoretic learning algorithm for both attacker and defender to collect more information about each other and a develop better attack and defense strategy. The equilibrium solution was used as the protection plan. The work \cite{sengupta2020multi} also studied MTD strategy using game-theoretic learning in a Bayesian Stackelberg game setting.
\subsection{Cyber-Physical Co-Learning in Subsystems}
Commonly, the learning modules for IT and OT in a subsystem are designed separately. Each module monitors or senses measurements in its system and operates or controls units in the associated layer. Since the IT and OT performances are interdependent, as illustrated in Fig.~\ref{fig:cross_layer}, it is necessary to create co-learning mechanisms that allow the two learning modules to interact with each other. The principles of cross-system CPS coordination designs in the monograph \cite{zhu2020cross} provide initial attempts to develop effective co-learning mechanisms.
We discuss how the principles can be applied to co-learning designs. On the IT side, apart from learning to reconfigure IT systems to cope with cyber attacks, the IT learning also needs to consider the impact of cyber learning mechanism to the OT performance. For example, in the learning-based honeypots that aim to engage an attacker to learn his behaviors, the reinforcement learning methods used in \cite{huang2019honeypot} adaptively reconfigure the honeypots by trading off between the reward and the OT related cost. The reward is quantified by the information garnered from the the attacker's behaviors, while the cost arises from the possibility that the attacker can abscond from the honeypot and enter the production IT system which is connected with the OT, and it is relate to the OT performance. It is essential for the learning agent to consider the OT-level consequences while learning for that honeypot defense strategies. In this way, the IT system makes more effective decisions and aligns its perception of the consequences with the goal of the entire system.
\begin{figure}[h]
\centering
\includegraphics[scale=0.5]{plots/cross_layer.png}
\caption{Cyber-physical learning is a co-learning process between the IT and the OT systems. A learning agent consists of IT and OT learning modules. The attacks on the C and the P agents are denoted as $a_C$ and $a_P$.
IT learning needs to be aware of the future impact of the learning mechanism; it creates an adaptive mechanism to prevent the propagation of attack $a_C$ to the OT system with the best effort. OT learning needs to consider that IT learning can be imperfect; it creates a cyber-aware mechanism resilient to $a_P$.}
\label{fig:cross_layer}
\end{figure}
Similarly, on the OT side, the OT learning aims to learn the effective controls to stabilize the OT system and maintain the OT performance while considering possible consequences from the IT system (e.g., the probability of the attack penetrating the IT and the capability of the attacker). The knowledge sharing enabled by the learning agent provides the OT learning with appropriate countermeasures, expedites the learning process, and reduces uncertainties about the anticipated outcomes.
For example, the OT learning can learn to switch between controllers to cope with different situations and achieve OT resilience \cite{xu2015secure}. One offline but optimal controller is designed to guarantee a robust performance to anticipated uncertainties or attacks. The other online but simple controller can stabilize the system. When the OT system is attacked (e.g., sensors or actuators are compromised), switching from an optimal controller that relies on compromised sensors or actuators to a simple controller that can maintain the basic functions of the control system is key to OT resilience. Learning when to switch from one controller to the other relies on attack detection and the prior knowledge or inputs from the IT system.
We provide two baseline frameworks of cyber-physical co-designs, based on which we can further build co-learning algorithms. The work \cite{xu2018cross} proposed a resilient cyber-physical design to mitigate cyber attacks to industrial robotic control systems. A partially observed Markov decision process (POMDP) was used to model the IT behavior of the control system, and a time-delay dynamical system was adopted to characterize the OT behavior. A cyber state that could affect the OT dynamics was specified to represent the impact from the IT to the OT. A threshold-based control strategy was developed to achieve a secure and resilient mechanism by jointly considering the interdependence of the two systems.
The other example is the secure and resilient cyber-physical framework of multi-agent systems proposed in \cite{xu2015cyberphysical}. A signaling game was used to capture the communications between IT devices in adversarial environments. The OT control design was formulated as an optimal control problem. The holistic framework was established by an integrative game between IT and OT systems achieve secure and resilient control.
The cyber-physical co-learning ensures each subsystem in the ICS operates normally against cyber attacks and disturbances. The learning-based cyber detection and mitigation enable a flexible reconfiguration in the cyber system to retard attack penetration and protect OT assets. The learning-based control facilities the design of resilient control for complex OT systems. The information and knowledge sharing in the cyber-physical co-learning allows the IT and OT systems to be aware of and adapt to each other during the learning so that holistic and effective IT-OT strategies can be learned to achieve subsystem-level resilience.
\subsection{Multi-Agent Learning}
The cyber-physical co-learning enables subsystem-level resilience and is the foundation to achieving the system-level resilience and MAR. At the system level, functional subsystems in the ICS act as intelligent agents and cooperate to accomplish system-wide missions. Therefore, the overall resilience of the ICS is built upon the resilience of its subsystem. However, the ICS resilience is not a mere replica of one resilient subsystem due to the subsystems' diverse functionality and complex interdependencies. The resilience of one subsystem also affects the resilience of others. The type of coordination between the IT and OT learning of one subsystem is also needed between two subsystems. To this end, the learning structure depicted in Fig.~\ref{fig:ics_learning} can be extended to a multi-agent learning framework.
Compared with centralized architectures, the multi-agent communications and interactions in ICSs enable multiple subsystems to exchange information and improve their situational awareness and prepare for forthcoming failures. For example in \cite{chen2019control}, a feedback-based adaptive, self-configurable, and resilient framework for the unmanned aerial vehicle (UAV) overlay network has been proposed to provide coverage to underlay devices to maintain interconnectivity despite adversarial behaviors that can disrupt the network. Each agent senses the information of its connecting agents and reaches a coordinated mechanism to achieve the coverage goal. It has also been observed in \cite{chen2015resilient} that local sensing is sufficient for each agent to adapt to changes and disruptions in the power grid. It arises from the fact that the measurement of the voltages and power at local buses already contains sufficient information of the measurements from other buses through the physical laws for the learning-based control. Therefore, the multi-agent learning provides a distributed approach for subsystems to coordinate and adapt to each other, which is more resilient to agent's internal faults and malicious attacks.
Looking from the system level, many emerging sophisticated learning paradigms become applicable to develop multi-agent learning-based resilience mechanisms.
For example, \emph{federated learning} is a distributed machine learning paradigm where all learning agents collaboratively learn a shared prediction model without exchanging the local data. It has been used in the industrial Internet of Things to enable secure and privacy-preserving data access \cite{hou2021mitigating}, energy market prediction for efficient resource allocation \cite{wang2021electricity}, and vehicular networks for intrusion detection \cite{zhang2018distributed}. The application of federated learning in distributed energy systems can coordinate the sub units toward the common objective such as power management and scheduling \cite{wang2020aebis} and energy resource distribution \cite{lee2022federated} without leaking the local data. Therefore, federated learning provides a distributed learning mechanism for task-level coordination as well as safety and privacy guarantees in the learning, which improves the resilience to malicious data access and data breach attacks.
\emph{Meta-learning} provides another efficient way for a learning agent to adapt to different tasks and other agents. Meta-learning first learns a base model to fit all possible tasks, and then only uses a small amount of data to achieve fast adaptation to a specific task. Meta-learning has been used to investigate the energy dispatch mechanism in self-powered and sustainable multi-agent systems \cite{Munir2021multi} and develop power grid emergency control to maintain system reliability and security \cite{huang2022learning}.
In an ICS, such as distributed energy systems, individual subsystems can have different operational modes. For example, a power generating subsystem can operate at full load, half load, and zero loads. Different modes correspond to different OT dynamics. Therefore, other agents can use meta-learning to first maintain a general model of the power generating subsystem and then fast adapt to the specific operation mode, providing a more efficient learning mechanism. Besides, many subsystems share similar subsystem-level objectives. For example, the wind, nuclear, and coal power generating subsystems all seek to generate power. Therefore, an individual subsystem can leverage meta-learning to form a model for other subsystems and then fast adapt for cooperation. Meta-learning can also be used to estimate the attacker's model for resilience planning. When facing similar cyber attacks, the network defender can learn a general attacker's model using meta-learning and then updates the model with the new network data to develop effective defense strategies.
\emph{Stackelberg learning} is also a useful tool for security and resilience planning when facing external attacks. It involves two players, a leader and a follower, as in Stackelberg games. The leader can be a network defender who anticipates the attacker's (follower) action to make resilient defense strategies \cite{li2018false}. The leader can also be a system operator who maintains the system performance under the attacker's (follower) attack \cite{an2020stackelberg}. The associated Stackelberg game can be also extended to multiple defenders \cite{smith2014multidefender} to model multiple subsystems in ICSs. The learning agents in an ICS act as leaders and learn to protect the subsystems and the ICS from malicious attacks.
However, there are still several challenges with multi-agent learning despite its application for resilient designs. The first challenge is heterogeneity. An ICS is characterized by its heterogeneous subsystems. They need customized learning mechanisms to achieve their distinct objectives. For example, consider the wind power generating subsystem and the power delivery subsystem. Two subsystems have different OT and task objectives. The former focuses on smooth power generation while the latter cares about the power delivery on demand. Abrupt changes in the delivery side can affect the generating subsystem, while an unstable power generating subsystem can also jeopardize the stability of power grids.
The sampling efficiency comes as the next challenge. A learning agent needs sampled network data to make operational decisions and adapt to other subsystems. However, the required samples differ in tasks and subsystems. Even within the same subsystem, the IT and OT system require different sampling strategies. For example, the sampling interval for physical control in the OT system should be finer than the cyber detection in the IT system because a large sampling interval can destabilize the OT system. Time-sensitive subsystems, such as power generating subsystems, require more sampled data than time-insensitive subsystems, such as logistics subsystems. A learning agent must decide on efficient strategies to sample information from itself and other agents. Without adequate samples, the learning may not achieve sufficient resilience to deal with attacks and disturbances. Redundant sampling, however, can lead to latency and waste of memory shortage. Hence learning mechanisms need to consider the trade-off between sampling efficiency and system resilience.
The third challenge is the uncertainties in the learning. A learning agent forms a belief in other agents to anticipate their behaviors in face of uncertainties. An effective learning mechanism should help the agent form a correct belief. Incorrect ones can lead to destabilization and cascading errors in ICSs. Despite many approaches to creating the belief in learning, for example, computing empirical opponent's strategy from historical data or no-regret learning, effective belief-forming strategies should also be task-specific. There are no general rules on belief and anticipation formation. A learning agent needs a context-driven approach to learning the resilient operational strategy.
Although facing challenges, multi-agent learning for ICS resilience is still promising. It allows agents to learn on their own in a hierarchically structured way, and the entire system achieves its goal in the end. Single-point failure in learning will trigger immediate responses from other agents that will respond to the failure to prevent the system-level breakdown.
\subsection{Concept of Resilience}
The modern-day ICSs are increasingly integrated with smart devices and services enabled by the Internet of Things and smart automation technologies. This trend is accompanied by a growing attack surface that exposes ICSs to a large number of vulnerabilities across the multiple layers of the system. The traditional security mechanisms, including intrusion detection systems, firewalls, and encryption, are no longer sufficient to protect ICSs from sophisticated and unknown threats. For example, Advanced Persistent Threats (APTs) are one class of threats that can stay in the system for a long time and stealthily reach the target resource. In addition, many emerging attacks (e.g., recent Log4j attacks \cite{log4j2021mitre} and supply chain attacks \cite{kshetri2022economics}) and unknown ones have made it harder or impossible to prepare the system ahead of time and achieve perfect security. Instead, resilience is a critical aspect of cyber protection that provides a supplemental means to mitigate their impact once the traditional methods fail to thwart them. Resilience plays an even more important role in defending against unknown or unanticipated attacks as it becomes the last resort to safeguard ICSs from calamitous collapse.
An ICS is a large-scale system of systems. Its resilience is a composition of the resilience of many interacting subsystems. Each subsystem consists of its IT and OT units. Hence the resilience of the system relies on the resilience of IT and OT. For a given system, subsystem, or unit, its resilience is defined by its ability to respond to an attack, maintain its function after the attack, and recover from the attack. Illustrated in Fig.~\ref{fig:resilience_performance}, we divide the operation into three phases. The first phase is the prevention, or known as the ante impetum stage, where less sophisticated attacks are thwarted while some attacks launched at time $t_1$ succeed in bypassing the defense and starting to navigate within the system at $t_2$. The second phase is the interim impetum stage (response). The attack breaches the security defense and aims to look for and take over the control of the targeted resources.
For example, in the infamous Target data breach \cite{plachkinova2018security}, the APT attacker, after infiltrating the network by leveraging the third-party vendor’s security, moves laterally from less sensitive areas of Target’s network to areas storing consumer data. A resilient mechanism at this stage aims to respond to this breach and deter the attacker from moving forward to reach the target. Without it, the system performance will progressively deteriorate, and the system will eventually break down. In contrast, an appropriate resilient mechanism detects the breach at $t_3$ and immediately foils the attack by taking appropriate measures (e.g., disconnecting the subsystem, rebooting the system, or triggering a new set of authentication rules) to mitigate the risk. The third stage is the post impetum stage (recovery), where the defender can remove the attacker from the system and recover the compromised system to its ante impetum state or maintain an acceptable performance at the loss of the performance level $D$. Similar concepts on resilience are also discussed in \cite{mcjunkin2017electricity}.
\begin{figure}
\centering
\includegraphics[scale=0.14]{plots/ot_performance.png}
\caption{The resilience of an ICS can be measured by the OT-level performance. A resilient ICS experiences three stages after an attack is launched at $t_1$, including prevention, response, and recovery. A $(T,D)$-resilient ICS to the attack can agily respond to the attack by recovering within $T$ units of time and maintaining a loss of performance $D$. The IT-level performance is subsumed into the OT-level one. The IT-level performance determines the success rate and the timing of an attack.}
\label{fig:resilience_performance}
\end{figure}
\subsubsection{Metrics for Resilience}
For a given attack, we say that an ICS is $(T,D)$-resilient if the system can recover from the attack within $T$ units of time and maintain a maximum loss of performance $D$. Here, $T$ is a measure of how quickly the response is. From Fig.~\ref{fig:resilience_performance}, $T=t_4-t_2$ can be further viewed as the duration of the response stage. The performance loss $D$ quantifies the post impetum performance. A non-resilient system goes through an irreversible performance degradation during the operation, as depicted counterfactual-wise in Fig.~\ref{fig:resilience_performance}.
The concept of $(T,D)$-resilience provides basic metrics for ICSs. Another commonly used metric for resilience is the measure of the total performance loss after an attack. The maximum magnitude of performance loss $M$ at the interim impetum stage can also serve as an indicator of resilience. Note that the metrics discussed above build on the performance measures of the OT system or the operation at the OT level. It is important to recognize that the defense effort at the prevention stage in the IT system has a significant impact on the metrics even though it is not directly measured, and it is quantified through the OT performance. The effort to thwart the attack, such as the deployment of honeypots \cite{huang2019honeypot}, moving target defenses \cite{zhu2013game}, and zero-trust mechanisms \cite{rose2020zero}, creates more difficulty for the attacker to be successful, thus reducing the rate of successful attacks and the impact of their consequences. In Fig.~\ref{fig:resilience_performance}, this effort can be roughly measured by the time that it takes for an attack in the system to compromise the targeted OT assets. Many recent works on cyber resilience focus on resilience at this stage \cite{segovia2020cyber,mertoguno2019physics}. We can zoom into the IT system and further quantify the cyber resilience of an ICS. For example, in \cite{kerman2020implementing}, a zero-trust cyber resilient mechanism is developed for an enterprise network. The trustworthiness of each user is measured over time to disrupt unknown insiders from lateral movement. The resilience of the IT system is measured by the time it takes to deter an attack from the target asset and the distance between the attacker and the asset. Fig.~\ref{fig:resilience_performance} provides a resilience measure of ICSs through the OT performance which consolidates the impact and the performance of the IT system. We can also zoom into the prevention stage and create a similar performance measure for the IT system.
\subsubsection{Comparison with Related Concepts}
There are subtle differences between resilience and related concepts, such as fault-tolerance, robustness, and security. Fault tolerance is also a system property that aims to enable the system to maintain its core functions in the event of failures. Illustrated in Fig.~\ref{fig:resilience_fault_tolerance}, fault tolerance often prepares for anticipated failures, and those happen internally caused by fatigue, corrosion, or manufacturing flaws. Resilience, in contrast, deals with unpredictable ones and those caused by external influences, such as cyber-attacks, natural disasters, and terrorism. The event that the system aims to prepare for distinguishes the type of system properties that we need to focus on.
\begin{figure}
\centering
\includegraphics[scale=0.14]{plots/resilience_fault_tolerance.png}
\caption{The distinction between fault tolerance and resilience from the dimensions of unpredictability and the externality of the events. The two concepts overlap for some events.}
\label{fig:resilience_fault_tolerance}
\end{figure}
Robustness is a well-known system concept for control systems. The goal of robustness is to design systems to withstand a set of prescribed uncertainties or events. Robustness prepares for them and enables the system to function when they occur, which is beneficial for resilience planning. However, it does not focus on preparing the system for recovering from an unknown event that has disrupted the system. It is often either cost-prohibitive or nonviable to design systems robust to all possible events while maintaining their functions.
Robustness and resilience are complementary to each other. The low-probability events that are costly for robustness should be prepared by resilient mechanisms while the high-probability events, such as thermal noise and load variations, should be handled through robust designs. The concept of robustness is often used for OT, e.g., controller designs and operation planning.
Security is a related concept that focuses on the prevention of adversarial events to safeguard the confidentiality, integrity, and availability of IT systems. Despite the effort to create a cyber defense to detect and foil attacks from penetrating IT, there is no perfect security that can assure the system of no attacks. In contrast, cyber resilience plays a significant role that enables the IT and OT systems to function even though the attack is in the network already.
\subsection{Resilience of Distributed Systems}
ICS are large-scale systems. Modern-day power grids are composed of many distributed energy resources at the edge of the power grid. The increasing interdependencies and connectivities among subsystems have exposed the ICS to a large attack surface. As illustrated in Fig.~\ref{fig:attack_ics}, an attacker can first gain the privilege of an IT subsystem through (a) and then move laterally within the IT network to find the target asset through (b). The attacker takes down the OT asset of the targeted subsystem, and its compromise leads to cascading failures in the OT system. This type of incident has been witnessed in the Ukrainian power grid attack \cite{case2016analysis} and the oil pipeline system attack \cite{colonial2021wiki}.
They both are attributed to APTs, which carry out a kill chain consisting of a sequence of resourceful, stealthy, and strategic attacks to reach their target.
\begin{figure}[h]
\centering
\includegraphics[scale=0.23]{plots/attack_cps.png}
\caption{Some illustrative attacks on subsystems of an ICS. The connectivities among subsystems increase the attack surface. An attacker can move laterally through the IT network from (a) to (b) and locates the OT asset using (c). The attack can cause cascading failures through (d).}
\label{fig:attack_ics}
\end{figure}
As the attackers are growingly sophisticated, resilience becomes increasingly essential for ICSs. However, it is challenging for large-scale systems through centralized control and operations. One-point failure can propagate to the entire system, and a centralized operation is less flexible as the entire system needs to be reconfigured for the failure.
Distributed control, on the other hand, enables the agility of the entire system. The system can still maintain its core functions well even when a subset of subsystems fails. In addition, as the decisions are distributed, it will be more convenient and faster for each subsystem to respond to the disruption and its connected subsystems. The challenge, however, to achieve it is the need for machine intelligence to enable detection and response through fast data analytics and decision-making.
Hence, to enable distributed and collaborative resilience, we need to create an AI stack between the IT and OT systems of each subsystem. Illustrated in Fig.~\ref{fig:ai_stack}, AI agents are introduced to connect cyber IT agents (C) and physical OT agents (P)\footnote{Here we call the IT and OT systems as agents to align with the names in our MAS framework. AI agents refer to the introduced AI components such as computational devices.}.
\begin{figure}
\centering
\includegraphics[scale=0.31]{plots/ai_stack.png}
\caption{An illustration of interactions between AI, IT, and OT agents of three subsystems. The interactions include (1) interface between a C agent and an AI agent, (2) interface between a P agent and an AI agent, (3) interface between two AI agents in two subsystems.
The C and P agents from different subsystems can also be interdependent. The AI agents interact with C and P agents in the same subsystem and other subsystems to achieve distributed and collaborative resilience.
}
\label{fig:ai_stack}
\end{figure}
A more detailed and functional diagram of an AI-augmented subsystem is also demonstrated in Fig.~\ref{fig:ai_agent}. An AI agent serves as an interface between a C agent and a P agent and is equipped with necessary functional blocks, including diagnosis, prognosis, simulation, data, strategic planning, and machine learning, to create fast online responses to an unanticipated event. The AI agent can also interact with other connected AI agents. This architecture enables the AI agent to control the performance of its associated subsystem as well as coordinate with other subsystems.
When a cyber-attack occurs locally on the subsystem, the AI agent can monitor the behaviors of the C and P agents in its subsystem and make tactical control and planning decisions to respond to the event. Each subsystem can achieve its own resilience in a distributed fashion. We refer to this type of resilient mechanism for the large-scale ICS as distributed resilience. As the failures can propagate across multiple subsystems, the AI agents need to communicate and coordinate to achieve collaborative resilience. Sharing of information and intelligence among AI agents can not only improve the distributed resilience of their subsystem but also enhance the global resilience of the large-scale system. For example, in \cite{achbarou2018new}, we have seen that the sharing of information in intrusion detection systems can help the entire network defend against zero-day attacks. In \cite{sun2021data}, the collaborations among multiple machine learners have led to significant improvement in the learning as well as resilience to data poisoning attacks.
\begin{figure}
\centering
\includegraphics[scale=0.22]{plots/ai_agent.png}
\caption{An illustrative functional diagram of an AI-augmented subsystem. An AI agent of a subsystem monitors the behaviors of its C and P agents and communicates with other AI agents to create tactical control and planning decisions to respond to malicious attacks and disturbances.}
\label{fig:ai_agent}
\end{figure}
\section{Introduction} \label{sec:introduction}
\input{intro}
\section{Security Challenges of Industrial Control Systems} \label{sec:ics}
\input{ics_challenges}
\section{Resilience of Industrial Control Systems} \label{sec:resilience}
\input{ics_resilience}
\section{Learning and AI for Resilient ICSs} \label{sec:learning}
\input{ics_learning}
\section{Resilience of Distributed Energy Systems: A Conceptual Case Study} \label{sec:case}
\input{ics_case}
\section{Conclusion} \label{sec:conclusion}
\input{conclusion}
\bibliographystyle{IEEEtran}
\section{Introduction} \label{sec:introduction}
\input{intro}
\section{Security Challenges of Industrial Control Systems} \label{sec:ics}
\input{ics_challenges}
\section{Resilience of Industrial Control Systems} \label{sec:resilience}
\input{ics_resilience}
\section{Learning and AI for Resilient ICSs} \label{sec:learning}
\input{ics_learning}
\section{Resilience of Distributed Energy Systems: A Conceptual Case Study} \label{sec:case}
\input{ics_case}
\section{Conclusion} \label{sec:conclusion}
\input{conclusion}
\bibliographystyle{IEEEtran}
\subsection{Learning in Cyber IT System}
In the IT system, a learning agent monitors the cyber behaviors of the users and the network to detect the anomaly and adversarial behaviors. The goal of the IT operation is to mitigate the impact of the attack, which can propagate to the OT if not carefully handled. The learning process in this system depends on the sophistication of attacks and resources of the decision-maker, including his knowledge, computation, and maneuverability.
Based on the functionality, we can summarize the components in the feedback loop in Fig.~\ref{fig:ics_learning} into two modules: cyber detection and cyber mitigation. Two modules work jointly to enable cyber resilience.
\subsubsection{Learning for Cyber Detection}
Cyber detection is fundamental for cyber resilience. An AI agent can learn from cyber data such as network traffic and log traces to detect and recognize cyber threats, including internal anomalies and external attacks on cyber agents and human operators. Based on the learning pattern, the learning methods for cyber detection can be categorized as supervised learning, unsupervised learning, and semi-supervised learning \cite{mitchell2014survey}. Each type of learning provides different advantages to achieving successful detection.
\emph{Supervised learning} uses cyber data and labeled system behaviors (for example, normal or abnormal) to train a cyber detection model offline and perform detection online.
Many works have developed successful learning-based detection models based on different supervised learning techniques. For example,
in \cite{linda2009neural}, Linda et al. used neural networks (NNs) to detect cyber intrusions for critical infrastructures.
Their neural network model was trained by actual network data and showed a perfect detection rate and zero false positives on testing cyber attacks.
In \cite{wang2019detection}, Wang et al. adopted random forest (RF) to detect natural disturbances and artificial cyber attacks for power systems. The model was trained by historical network data and relevant log information. The experiments on an open-source simulated power system dataset reported a successful detection rate higher than 93\%.
Mokhtari et al. in \cite{mokhtari2021machine} used $K$-nearest neighbors (KNNs) and decision tree classifier (DTC) as a tool for cyber detection. The data collected by the SCADA system were used for detection model training, and their reported detection precision and accuracy were higher than 97\%.
Many works such as \cite{beaver2013evaluation,hink2014machine,ozay2015machine} also made a thorough comparison of different supervised learning methods for cyber detection and demonstrated the effectiveness of the learning-based methods.
Supervised learning is also used for human-related attack detection. In the work of \cite{alam2020phishing}, an RF and DTC-based model was developed to detect phishing attacks and help human operators stay safe. An NN-based detection model was also proposed in \cite{abdelaty2021daics} to reduce false alarms in ICSs to ease human operators' cognitive load.
Supervised learning enjoys a high cyber detection precision and accuracy as we have observed in the literature. However, an AI agent can hardly detect new anomalies and attacks with supervised learning because of missing labels in the dataset. Unless updating detection models constantly, an AI agent may not detect zero-day attacks and fail to provide cyber resilience for the system. Besides, acquiring labeled cyber data is often a challenge.
As a complementary approach, \emph{unsupervised learning} can handle unlabeled cyber data and provide more generic detection. The main idea behind unsupervised learning is feature-based clustering. A learning agent first identifies the features of cyber data and then searches for potential anomalies or attacks by clustering the featured data. Many recent works have focused on developing detection tools with unsupervised learning.
For example, Maglaras and Jiang in \cite{maglaras2014intrusion} adopted the One-Class Support Vector Machine (OCSVM) to detect malicious network traffic in SCADA systems. The detector extracted attributes from network traces and then was trained to classify the attacked nodes that jammed the network traffic.
Alves et al. in \cite{alves2018embedding} used K-means clustering to detect data interception, data injection attacks, and denial of service (DoS) in industrial PLC networks. The anomaly classification and learning were based on the features of real-time data streams, such as pack latency and processing information.
Kiss et al. in \cite{kiss2014data} investigated detection in critical infrastructures based on K-means and MapReduce paradigm to achieve real-time cyber attack detection.
To address the anomalous behavior detection of building energy management systems, Wijayasekara et al. in \cite{linda2012computational,wijayasekara2014mining} developed an efficient detection method based on the modified nearest neighbor clustering and fuzzy logic, which not only performed faster than traditional alarm-based detection methods, but also was able to provide linguistic interpretation of the identified anomalies.
Unsupervised learning may produce more false alarms in detection due to a lack of labels compared with supervised learning \cite{umer2022machine}. Human intervention for diagnosis is useful, but it may delay the IT operation and result in less cyber resilience in the IT system. To overcome the issue, many works consolidate supervised and unsupervised learning to create \emph{semi-supervised learning} for effective cyber detection.
In recent work \cite{kravchik2018detecting}, Kravchik and Shabtai developed a window-based cyber attack detection using convolutional neural networks (CNN) for critical water infrastructures. The CNN was trained to predict normal system behaviors, and anomalies were detected every time window based on the error between predicted and observed system behavior.
Marino et al. in \cite{marino2019cyber} adopted unsupervised learning for ICS anomaly detection classification and evaluated its performance with supervised learning. The combination of the two learning methods empowered the detection of unseen cyber anomalies and improved the detection accuracy.
We refer the readers to recent surveys \cite{umer2022machine,bhamare2020cybersecurity,anthi2021adversarial,handa2019machine} for details on machine learning algorithms for cyber detection in ICSs.
\subsubsection{Learning for Cyber Mitigation}
The cyber mitigation module serves as the decision-making and operation units in the cyber learning architecture and is the key to guaranteeing cyber resilience. However, learning for cyber mitigation is more challenging because the mitigation strategies rely on tasks and network structures. Traditional cyber mitigation strategies do not pay attention to the higher-order impact of the mitigation strategies on the system performance and the complex roles of attackers who can strategically and stealthily counteract the system strategies. To address these challenges, game-theoretic learning and reinforcement learning (RL) are used for effective and proactive cyber mitigation.
\emph{Game-theoretic learning} treats the attacker as a rational player and learns a strategic solution by observing the attacker's attack trace. It provides a suitable framework to develop attack-aware solutions to combat APTs in the IT system, where attackers aim to penetrate the IT system to attack the target asset \cite{rubio2019current}. For example, Huang and Zhu in \cite{huang2018analysis,huang2019adaptive,huang2020dynamic} proposed a Bayesian learning mechanism to defend APTs in the cyber system proactively. They used a multi-stage dynamic Bayesian game to characterize the long-term interaction between a stealthy attacker and a proactive defender. The Bayesian learning was used to update the defender's knowledge of the attacker along the interaction and assisted the defender in developing effective defense strategies for cyber resilience.
Game-theoretic learning has also been used in defensive deception to achieve effective cyber mitigation. A learning agent can proactively learn to deceive the attacker with false targets or steer the attacker to a non-hazard zone to mitigate attack consequences \cite{zhang2020game,huang2020strategic,zhu2021survey,pawlick2021game}.
A \emph{reinforcement learning} agent seeks a strategy that minimizes the attack consequence over some time from observed data. For example, in \cite{huang2019honeypot}, the network defender used RL to adaptively allocate honeypot resources to trap the attacker in the target honeypot for as long as possible.
RL has also been shown to be effective for human-related cyber attacks such as phishing and attention manipulation. For example, Huang and Zhu in \cite{huang2021advert} proposed a phishing prevention mechanism (ADVERT) using RL to generate adaptive visual aids to counteract inattention and improve the human recognition of phishing attacks. Experiments with human volunteers showed an accuracy improvement of phishing recognition from 74\% to a minimum of 86\%. Another human attention management mechanism (RADAMS) using RL was proposed in \cite{huang2021radams} to help human operators fight against informational attacks, such as fake alerts. The proposed mechanism used RL to de-emphasize alerts selectively and reduce human operators' cognitive load. Experimental results showed that the developed attention management mechanism reduced the risk by as much as 20\% compared with default strategies.
\subsection{Learning in Physical OT System}
In the OT system, the goal of a learning agent is to create a learning-based resilient control mechanism that can further mitigate the attack impact on the OT system when the attacker reaches the OT \cite{homeland2016recommended}. The OT learning mechanism also follows the feedback architecture depicted in Fig.~\ref{fig:ics_learning}. The monitoring of the OT assets, including sensors, actuators, and terminal units, creates situational awareness for the decision-makers to adaptively reconfigure the control system to threats and environmental changes.
The learning agent can either learn the uncertainties first and then make decisions or directly learn the control decisions.
Unlike the OT system, physical plants in the OT system resort to physical models to operate normally. The controller design for physical agents is also model-specific. Therefore, effective and resilient control in the OT system relies on the physical models.
However, exact physical models can be hard or impossible to obtain as contemporary ICSs become more integrated, making resilient control even more challenging. Therefore, many research resorts supervised learning to estimate the physical dynamics and then develop adaptive control strategies for possible OT system attacks. For example, Li and Zhao in \cite{li2021resilient} studied learning-based resilient and adaptive control strategies for uncertain deception attacks in CPSs. The neural network was used to approximate unknown nonlinear and switching physical system dynamics. Then a dynamic surface-based adaptive controller was designed for sensor and actuator deception attacks. The resilient controller was validated on continuously stirred tank reactor systems.
Liu et al. in \cite{liu2021adaptive} proposed observer-based adaptive neural network control for nonlinear CPSs subject to false data injection attacks. The neural network was used to adaptively approximate the unknown nonlinear functions of the CPS. Then a resilient feedback controller was designed to ensure bounded output under malicious attacks.
The model-free adaptive controller design for unknown dynamics was also studied in \cite{wang2017neural,farivar2019artificial}. Both work used supervised learning and neural network models to develop adaptive control strategies to cope with sensor faults and malicious attacks. The system resilience was guaranteed by the stabilizable controller and bounded system output.
Supervised learning has been useful for predicting physical models. However, dealing with malicious and intelligent attacks in the OT system is insufficient. To address the issue, game-theoretic learning is used to find control strategies for physical resilience because of its ability to characterize strategic interactions with adversaries. A learning agent learns to reconfigure the control system by anticipating the attacker's behavior. An example of game-theoretic learning in the OT system is moving target defense (MTD) in ICSs. The defender protects a multi-layer network system and prevents the attacker's penetration by changing the network configuration of each layer. The attacker exploits each layer's vulnerabilities and tries to penetrate the entire system. Zhu and Ba{\c{s}}ar in \cite{zhu2013game} investigated this issue by assuming two players have no knowledge of each other. They proposed an iterative game-theoretic learning algorithm for both attacker and defender to collect more information about each other and a develop better attack and defense strategy. The equilibrium solution was used as the protection plan. The work \cite{sengupta2020multi} also studied MTD strategy using game-theoretic learning in a Bayesian Stackelberg game setting.
\subsection{Cyber-Physical Co-Learning in Subsystems}
Commonly, the learning modules for IT and OT in a subsystem are designed separately. Each module monitors or senses measurements in its system and operates or controls units in the associated layer. Since the IT and OT performances are interdependent, as illustrated in Fig.~\ref{fig:cross_layer}, it is necessary to create co-learning mechanisms that allow the two learning modules to interact with each other. The principles of cross-system CPS coordination designs in the monograph \cite{zhu2020cross} provide initial attempts to develop effective co-learning mechanisms.
We discuss how the principles can be applied to co-learning designs. On the IT side, apart from learning to reconfigure IT systems to cope with cyber attacks, the IT learning also needs to consider the impact of cyber learning mechanism to the OT performance. For example, in the learning-based honeypots that aim to engage an attacker to learn his behaviors, the reinforcement learning methods used in \cite{huang2019honeypot} adaptively reconfigure the honeypots by trading off between the reward and the OT related cost. The reward is quantified by the information garnered from the the attacker's behaviors, while the cost arises from the possibility that the attacker can abscond from the honeypot and enter the production IT system which is connected with the OT, and it is relate to the OT performance. It is essential for the learning agent to consider the OT-level consequences while learning for that honeypot defense strategies. In this way, the IT system makes more effective decisions and aligns its perception of the consequences with the goal of the entire system.
\begin{figure}[h]
\centering
\includegraphics[scale=0.5]{plots/cross_layer.png}
\caption{Cyber-physical learning is a co-learning process between the IT and the OT systems. A learning agent consists of IT and OT learning modules. The attacks on the C and the P agents are denoted as $a_C$ and $a_P$.
IT learning needs to be aware of the future impact of the learning mechanism; it creates an adaptive mechanism to prevent the propagation of attack $a_C$ to the OT system with the best effort. OT learning needs to consider that IT learning can be imperfect; it creates a cyber-aware mechanism resilient to $a_P$.}
\label{fig:cross_layer}
\end{figure}
Similarly, on the OT side, the OT learning aims to learn the effective controls to stabilize the OT system and maintain the OT performance while considering possible consequences from the IT system (e.g., the probability of the attack penetrating the IT and the capability of the attacker). The knowledge sharing enabled by the learning agent provides the OT learning with appropriate countermeasures, expedites the learning process, and reduces uncertainties about the anticipated outcomes.
For example, the OT learning can learn to switch between controllers to cope with different situations and achieve OT resilience \cite{xu2015secure}. One offline but optimal controller is designed to guarantee a robust performance to anticipated uncertainties or attacks. The other online but simple controller can stabilize the system. When the OT system is attacked (e.g., sensors or actuators are compromised), switching from an optimal controller that relies on compromised sensors or actuators to a simple controller that can maintain the basic functions of the control system is key to OT resilience. Learning when to switch from one controller to the other relies on attack detection and the prior knowledge or inputs from the IT system.
We provide two baseline frameworks of cyber-physical co-designs, based on which we can further build co-learning algorithms. The work \cite{xu2018cross} proposed a resilient cyber-physical design to mitigate cyber attacks to industrial robotic control systems. A partially observed Markov decision process (POMDP) was used to model the IT behavior of the control system, and a time-delay dynamical system was adopted to characterize the OT behavior. A cyber state that could affect the OT dynamics was specified to represent the impact from the IT to the OT. A threshold-based control strategy was developed to achieve a secure and resilient mechanism by jointly considering the interdependence of the two systems.
The other example is the secure and resilient cyber-physical framework of multi-agent systems proposed in \cite{xu2015cyberphysical}. A signaling game was used to capture the communications between IT devices in adversarial environments. The OT control design was formulated as an optimal control problem. The holistic framework was established by an integrative game between IT and OT systems achieve secure and resilient control.
The cyber-physical co-learning ensures each subsystem in the ICS operates normally against cyber attacks and disturbances. The learning-based cyber detection and mitigation enable a flexible reconfiguration in the cyber system to retard attack penetration and protect OT assets. The learning-based control facilities the design of resilient control for complex OT systems. The information and knowledge sharing in the cyber-physical co-learning allows the IT and OT systems to be aware of and adapt to each other during the learning so that holistic and effective IT-OT strategies can be learned to achieve subsystem-level resilience.
\subsection{Multi-Agent Learning}
The cyber-physical co-learning enables subsystem-level resilience and is the foundation to achieving the system-level resilience and MAR. At the system level, functional subsystems in the ICS act as intelligent agents and cooperate to accomplish system-wide missions. Therefore, the overall resilience of the ICS is built upon the resilience of its subsystem. However, the ICS resilience is not a mere replica of one resilient subsystem due to the subsystems' diverse functionality and complex interdependencies. The resilience of one subsystem also affects the resilience of others. The type of coordination between the IT and OT learning of one subsystem is also needed between two subsystems. To this end, the learning structure depicted in Fig.~\ref{fig:ics_learning} can be extended to a multi-agent learning framework.
Compared with centralized architectures, the multi-agent communications and interactions in ICSs enable multiple subsystems to exchange information and improve their situational awareness and prepare for forthcoming failures. For example in \cite{chen2019control}, a feedback-based adaptive, self-configurable, and resilient framework for the unmanned aerial vehicle (UAV) overlay network has been proposed to provide coverage to underlay devices to maintain interconnectivity despite adversarial behaviors that can disrupt the network. Each agent senses the information of its connecting agents and reaches a coordinated mechanism to achieve the coverage goal. It has also been observed in \cite{chen2015resilient} that local sensing is sufficient for each agent to adapt to changes and disruptions in the power grid. It arises from the fact that the measurement of the voltages and power at local buses already contains sufficient information of the measurements from other buses through the physical laws for the learning-based control. Therefore, the multi-agent learning provides a distributed approach for subsystems to coordinate and adapt to each other, which is more resilient to agent's internal faults and malicious attacks.
Looking from the system level, many emerging sophisticated learning paradigms become applicable to develop multi-agent learning-based resilience mechanisms.
For example, \emph{federated learning} is a distributed machine learning paradigm where all learning agents collaboratively learn a shared prediction model without exchanging the local data. It has been used in the industrial Internet of Things to enable secure and privacy-preserving data access \cite{hou2021mitigating}, energy market prediction for efficient resource allocation \cite{wang2021electricity}, and vehicular networks for intrusion detection \cite{zhang2018distributed}. The application of federated learning in distributed energy systems can coordinate the sub units toward the common objective such as power management and scheduling \cite{wang2020aebis} and energy resource distribution \cite{lee2022federated} without leaking the local data. Therefore, federated learning provides a distributed learning mechanism for task-level coordination as well as safety and privacy guarantees in the learning, which improves the resilience to malicious data access and data breach attacks.
\emph{Meta-learning} provides another efficient way for a learning agent to adapt to different tasks and other agents. Meta-learning first learns a base model to fit all possible tasks, and then only uses a small amount of data to achieve fast adaptation to a specific task. Meta-learning has been used to investigate the energy dispatch mechanism in self-powered and sustainable multi-agent systems \cite{Munir2021multi} and develop power grid emergency control to maintain system reliability and security \cite{huang2022learning}.
In an ICS, such as distributed energy systems, individual subsystems can have different operational modes. For example, a power generating subsystem can operate at full load, half load, and zero loads. Different modes correspond to different OT dynamics. Therefore, other agents can use meta-learning to first maintain a general model of the power generating subsystem and then fast adapt to the specific operation mode, providing a more efficient learning mechanism. Besides, many subsystems share similar subsystem-level objectives. For example, the wind, nuclear, and coal power generating subsystems all seek to generate power. Therefore, an individual subsystem can leverage meta-learning to form a model for other subsystems and then fast adapt for cooperation. Meta-learning can also be used to estimate the attacker's model for resilience planning. When facing similar cyber attacks, the network defender can learn a general attacker's model using meta-learning and then updates the model with the new network data to develop effective defense strategies.
\emph{Stackelberg learning} is also a useful tool for security and resilience planning when facing external attacks. It involves two players, a leader and a follower, as in Stackelberg games. The leader can be a network defender who anticipates the attacker's (follower) action to make resilient defense strategies \cite{li2018false}. The leader can also be a system operator who maintains the system performance under the attacker's (follower) attack \cite{an2020stackelberg}. The associated Stackelberg game can be also extended to multiple defenders \cite{smith2014multidefender} to model multiple subsystems in ICSs. The learning agents in an ICS act as leaders and learn to protect the subsystems and the ICS from malicious attacks.
However, there are still several challenges with multi-agent learning despite its application for resilient designs. The first challenge is heterogeneity. An ICS is characterized by its heterogeneous subsystems. They need customized learning mechanisms to achieve their distinct objectives. For example, consider the wind power generating subsystem and the power delivery subsystem. Two subsystems have different OT and task objectives. The former focuses on smooth power generation while the latter cares about the power delivery on demand. Abrupt changes in the delivery side can affect the generating subsystem, while an unstable power generating subsystem can also jeopardize the stability of power grids.
The sampling efficiency comes as the next challenge. A learning agent needs sampled network data to make operational decisions and adapt to other subsystems. However, the required samples differ in tasks and subsystems. Even within the same subsystem, the IT and OT system require different sampling strategies. For example, the sampling interval for physical control in the OT system should be finer than the cyber detection in the IT system because a large sampling interval can destabilize the OT system. Time-sensitive subsystems, such as power generating subsystems, require more sampled data than time-insensitive subsystems, such as logistics subsystems. A learning agent must decide on efficient strategies to sample information from itself and other agents. Without adequate samples, the learning may not achieve sufficient resilience to deal with attacks and disturbances. Redundant sampling, however, can lead to latency and waste of memory shortage. Hence learning mechanisms need to consider the trade-off between sampling efficiency and system resilience.
The third challenge is the uncertainties in the learning. A learning agent forms a belief in other agents to anticipate their behaviors in face of uncertainties. An effective learning mechanism should help the agent form a correct belief. Incorrect ones can lead to destabilization and cascading errors in ICSs. Despite many approaches to creating the belief in learning, for example, computing empirical opponent's strategy from historical data or no-regret learning, effective belief-forming strategies should also be task-specific. There are no general rules on belief and anticipation formation. A learning agent needs a context-driven approach to learning the resilient operational strategy.
Although facing challenges, multi-agent learning for ICS resilience is still promising. It allows agents to learn on their own in a hierarchically structured way, and the entire system achieves its goal in the end. Single-point failure in learning will trigger immediate responses from other agents that will respond to the failure to prevent the system-level breakdown.
\subsection{Industrial Control Systems}
Industrial Contol Systems (ICSs) are widely adopted in industrial sectors and critical infrastructures, such as power grids and nuclear plants. They are task-critical implementations of Cyber-Physical Systems (CPSs) to operate and automate the industrial process and consist of numerous subsystems such as SCADA systems and distributed control systems (DCS), field controllers such as Programmable Logic Controllers (PLC) and human-machine interface (HMI), physical plants such as power generators and actuators, and other intellectual electrical devices (IED) \cite{stouffer2011guide,asghar2019cybersecurity,bhamare2020cybersecurity,smidts2019next}. All physical components are organized into a hierarchical, multi-level network and are connected by the fieldbus or wireless networks to perform specific control tasks. Data transmission between levels of the network ensures real-time surveillance for security and control purposes. Fig.~\ref{fig:ics_illustration} shows a notional integrated ICS architecture in the power domain based on the commonly adopted Purdue Model \cite{williams1994purdue}.
\begin{figure}
\centering
\includegraphics[scale=0.3]{plots/ics_2.png}
\caption{An illustrative example of an ICS network of power systems. Levels 0-2 constitute the OT level for real task execution. Levels 3-5 constitute the IT level for information processing such as scheduling. All components in the ICS cooperate together to deliver electricity to support end users.
An attacker can follow the red path to attack a physical plant from an enterprise computer. AI stacks in IT and OT levels are introduced to augment the learning capabilities of IT and OT subsystems and improve the overall ICS resilience.}
\label{fig:ics_illustration}
\end{figure}
An ICS typically consists of several field sites (levels 0-2) that are functional points and can execute real-time tasks. A field site is generally equipped with local PLC controllers and HMIs for automated and human control, some IEDs such as sensors for measurements, and physical plants for specific tasks. Fieldbus connects all components to the cyber components and the controller center.
Examples of field sites include wind turbines and gas power plants.
The fieldbus, such as the CAN bus, is time-critical and fault-tolerant since any time delay and signal distortion can lead to severe faults of field sites. Besides, the fieldbus requires a large communication capacity due to the increasing scale and complexity of contemporary ICSs. For security considerations, the data transported by fieldbus should be filtered by firewalls before sending to other levels.
The control center (level 3) is a supervisory level designed for real-time system monitoring, control configurations to filed sites, and necessary human operations and interventions. At this level, HMIs provide process information and status to human operators. Control and data servers monitor and adjust control processes and store historical information. Engineering stations provide the developing and debugging environments for control operations.
The enterprise zone (level 5) is a business network performing general tasks, such as human resource management and internal communications. Data transmission between the controller center and the enterprise zone is restricted and inspected by the demilitarized zone (DMZ) (sometimes called a perimeter network). A DMZ is a subnetwork acting as an intermediary for connected security devices, which is more than a firewall. DMZ ensures that the data can only be transmitted from a high-security level to a low-security level, and the reverse is prohibited. This is because the data in the controller station is more critical and sensitive. The DMZ filters redundant data traffic from the enterprise zone to the controller center that interrupts the control process. The DMZ also adds another layer of security to the ICS. An attacker cannot bypass the DMZ unless he has direct access to the equipment within the DMZ. It prevents malicious access and attacks on the controller and physical plants from cyberspace, providing another layer of security.
The enterprise zone can generally access the Internet or other wide area networks (WANs) (level 5) through another DMZ or a firewall, depending on the security configuration. Other remote field sites can also be connected to the control center via the Internet or WANs from outside. Different security protocols are required for such remote connections.
Following the adaption of the Purdue Model, the network can be further grouped into IT and IT levels based on the functionalities. Levels 0-2 constitute the OT level and are responsible for real-time task execution such as wind turbine control. Levels 3-5 form the IT level and take charge of information processing-related work such as task scheduling and management.
Upon the classic ICS architecture, we introduce AI stacks at IT and OT levels to enable computational intelligence. The AI stack at the IT level enables real-time intelligent cyber detection and response for cyber attacks based on network data, assisting the ICS to achieve cyber resilience. The AI stack at the OT level empowers learning-based methods for controlling physical plants. Compared with traditional model-based control, the AI-enabled learning methods are particularly convenient for OT systems with complex dynamics. The AI stack also facilitates the development of real-time adaptive strategies for anomalies and disturbances. We mention that we do not specifically portray the AI components in Fig.~\ref{fig:ics_illustration} because the AI stack and computational intelligence can be algorithmically built into many digital devices such as working stations and local controllers. As its scale becomes larger and the structure more complex, the integration of AI stacks benefits the modern-day ICS to operate more adaptively and effectively against cyber attacks and unknown disturbances than traditional ones.
\subsection{Multi-Agent System Framework of Industrial Control Systems}
The classic Purdue model provides a clear visualization of a hierarchical and monolithic ICS architecture. However, it also has some limitations to achieving resilient ICS design as ICSs are becoming more integrated. On the one hand, the growing monolithic ICS architecture makes the network analysis intractable, making it challenging to design effective and resilient control strategies to cope with cyber attacks and external disturbances. On the other hand, more specialized subsystems are integrated into modern-day ICSs so that all components operate in a distributed fashion.
Therefore, we can decompose the growing ICS into a set of subsystems based on functionalities. The subsystems cooperate to accomplish the overall ICS mission. For example, in the power system, the generating plants work as individual subsystems to produce power, while the delivery grid act as another subsystem to deliver the power to end users.
Within each subsystem, we also have the cyber-physical architecture. Therefore, the IT and OT division in the Purdue model in Fig.~\ref{fig:ics_illustration} is still applicable for each subsystem. We refer to the divided systems as IT and OT systems, as shown in Fig.~\ref{fig:intro}. The IT and OT systems affect and control the subsystem in a distributed but interdependent fashion. For example, in a generating subsystem, the IT system can monitor real-time information such as generating capacity and output to the grid for power management and coordination. The OT system stabilizes the generating plants and produces power based on the requirements from the IT system.
Therefore, we propose a MAS framework to characterize modern-day ICSs using concepts in distributed control systems. The MAS framework has a two-level interpretation. At the system level, each functional subsystem is regarded as an intelligent agent with a specific objective. All agents interact and cooperate to accomplish the system-wide mission, such as power supply in the infrastructure. At the subsystem level, the IT and OT systems act as intelligent agents, and their interactions lead to achieving the subsystem-specific task objective. For example, a generating subsystem performs a smooth power generation under external requirement changes or malicious attacks on the generating plants.
The two-level MAS framework is intrinsically interdependent. The interactions at the subsystem level ensure the regular operation of the subsystem and lay the foundation for system-level cooperation. The interaction and coordination at the system level achieve the system-wide ICS mission.
Our MAS framework also enables learning for resilient ICS design and lays the foundation to achieve MAR. On the one hand, every intelligent agent (subsystem and IT or OT system in the subsystem) can utilize network data and learning to adapt to new attacks and disturbances, which provides a flexible and real-time adaption mechanism to achieve resilience. On the other hand, every agent can learn to achieve self-adaption and cooperate with others when the agent has complex dynamics and the interdependencies between agents are not known or planned for. We discuss more details on learning for resilience in Section \ref{sec:learning}.
Our multi-agent perspective aligns with the literature. For example, a hierarchical MAS framework (Hierarchical Multi-agent Dynamic System, HMADS) has been proposed in \cite{rieger2013resilient,rieger2013hierarchical} to characterize resilient control systems. The framework consists of three layers based on high-level functionality: the management layer, the coordination layer, and the execution layer. The management layer provides the high-level system objective and task scheduling. The coordination layer allocates resources to system components to align with the system objective. The execution layer takes charge of controlling and sensing field devices. A subsystem or a controller in the control system is treated as an intelligent agent and is divided into one of three layers based on its sphere of influence. Our framework absorbs the management and coordination layers into the IT, and the execution layer is subsumed under the OT.
It is clear that a divide-and-conquer approach plays an essential role in the architecture of the large-scale complex ICS. In addition to the focus on IT and OT in this chapter, human technologies (HT) are becoming increasingly prominent. They are central to the resilience of the ICSs as many attacks start with exploiting human error and vulnerabilities. Human-machine technologies have been proposed to augment human performances, e.g., securing human operators from attention failures \cite{huang2022radams}, employees from phishing \cite{huang2022advert}, and users from noncompliance \cite{huang2022zetar,casey2016compliance}. It is natural to extend our framework to include an HT layer to incorporate the emerging human solutions and their interface with IT-OT.
The hierarchical divide-and-conquer view of ICSs provides a flexible and evolving perspective that enables integrating the systems with new emerging interfaces and technologies. For example, the adoption of IoT technologies and human-machine teaming capabilities can potentially augment ICSs with new cross-layer functionalities and introduce additional dimensions to the ICS architecture.
\subsection{Security Challenges}
The interconnected IT and OT components can create potential vulnerabilities to degrade ICS performance and jeopardize the security of infrastructures. As shown by the red path in Fig.~\ref{fig:ics_illustration}, an attacker can penetrate the ICSs by compromising an enterprise computer and moving laterally within the ICS network to reach the target asset. For instance, an attacker can bring down a nuclear power plant by manipulating the sensors. We have witnessed an increasing number of such attacks, including ransomware attacks, insider threats, and APT attacks, which have raised serious security threats to ICSs \cite{makrakis2021industrial}. We summarize the critical security challenges of ICSs that remain open issues.
First, cascading failures can devastate ICSs due to the multi-level structure and the multi-agent characteristics. The cascading connections of different components provide numerous possible fault-error-failure propagation paths. A single fault of any component, such as hardware flaws or software bugs, can propagate in the network and consequently affect other components, which eventually leads to a system-wide functionality failure. These threats have been noticed and investigated in \cite{zhu2012dynamic,rieger2019industrial}. Cascading failures also make it hard to detect and diagnose the real anomaly in ICSs because of fault propagation. Fixing a single component's error is insufficient to address the general security issues of ICSs. More precise and complex fault detection and identification are required to pinpoint the initial fault and prevent further failures.
Second, the large-scale connections and heterogeneous components in contemporary ICSs lead to intrinsic complex structures, further increasing the diversity and the complexity of security vulnerabilities. Common security threats have been identified and categorized in \cite{cardenas2011attacks} and precautions can be designed to prevent the existing threats. However, many unknown factors can still cause system-wide malfunctions and jeopardize ICS security. Exploiting and identifying vulnerabilities for modern-day ICSs is still an open project. Besides, the attack surface of subnetworks in ICSs also differs from their characteristics and tasks. For example, generating plants and power grids have different network architectures and security vulnerabilities. The device compatibility is also a security concern in ICSs. It is common to observe obsolete devices in ICSs due to historical reasons. These components are generally not designed for contemporary security considerations and leave potential vulnerabilities. It is challenging to design effective and holistic strategies to make different versions of devices work together and provide enough security guarantees.
We identify cyber attacks as the third security challenge. A significant feature of contemporary ICSs compared with traditional ones is the adoption of the cyber layer. However, cyber attacks nowadays are causing a broad and profound impact on ICSs. While the cyber layer achieves efficient communications among different components and enables system automation, it also brings more security threats. The existing and recognized cyber threats include communication disruptions and malware attacks such as computer viruses and ransomware. Works such as \cite{uma2013survey,formby2017out} also summarized common cyber attacks and malware in ICSs.
Although cyber attacks are not originally designed to sabotage ICSs, several recent attacks on ICSs, including the Colonial Pipeline ransomware attack \cite{colonial2021wiki}, have shown their destructive power on infrastructures and economics. Cyber attacks also bring another level of security consideration: intelligent adversaries. Attacks can start malicious and intelligent attacks on ICSs through cyber layers, which differ from the security challenges caused by device faults. Advanced persistent threats (APTs) are an example of intelligent attacks. APT is a cyber attack that aims to penetrate the network and attack a high-value asset. It searches possible attack paths to the target assets stealthily and starts the attack when the time is perfect. APT can avoid existing security defense mechanisms strategically. So traditional methods such as routine security checks may have limited effects on APTs. Although it appears harmless for other components during the network penetration, the attack consequence is devastated because of the loss of the target asset. In short, it is critical to address the cyber threats in ICSs to maintain system resilience.
Human factors also pose security challenges in ICSs. ICSs are integrated with HMIs and workstations for human operators (or experts) to monitor the system status and debug errors. Although the contemporary ICS design keeps improving the human friendliness in operation and facilitating decision support, human operators are still indispensable. For example, human operators are required when processing emergencies such as generating plant breakdown. Besides, human operators are also capable of diagnosing complex errors such as new cyber attacks and faults of specialized devices. Therefore, human experts and their knowledge can benefit ICSs to achieve better performance and resilience. However, security threats also accompany humans and hence jeopardize ICSs.
For example, humans have limited attention. When cascading errors occur, it can be difficult for a human operator to recognize the relevant and irrelevant threats. Also, cyber attacks can launch feint attacks to distract and deceive human attention \cite{hitzel2019art,huang2021radams}. Humans are also prone to be manipulated. There have been phishing scams and espionage activities to steal critical ICS information or plant malware. Some activities can be hard to recognize. Last but not least, insider threats also significantly endanger the security of ICSs. Some of these attacks leverage human operators' access to the core system to steal data or inject malware to sabotage plants. According to empirical research \cite{homeland2016recommended}, social engineering scams and insider threats have become the primary threats to ICSs. Therefore, training and regulating human behaviors to have more efficient and user-friendly ICSs is vital and remains an open challenge.
The security challenges raise severe threats to ICSs such as power grids and nuclear plants, leading to devastating consequences. Therefore, it is critical and indispensable to improve the resilience of ICSs, so that the system can resist malicious attacks and disturbances and maintain an acceptable level of operational normalcy. Traditional threat detection and mitigation are not sufficient to deal with the complexity of modern-day ICSs and the new security challenges. The integration of AI stacks into the ICSs can augment the distributed learning capabilities of the subsystems and improve the overall ICS's resilience.
\subsection{Concept of Resilience}
The modern-day ICSs are increasingly integrated with smart devices and services enabled by the Internet of Things and smart automation technologies. This trend is accompanied by a growing attack surface that exposes ICSs to a large number of vulnerabilities across the multiple layers of the system. The traditional security mechanisms, including intrusion detection systems, firewalls, and encryption, are no longer sufficient to protect ICSs from sophisticated and unknown threats. For example, Advanced Persistent Threats (APTs) are one class of threats that can stay in the system for a long time and stealthily reach the target resource. In addition, many emerging attacks (e.g., recent Log4j attacks \cite{log4j2021mitre} and supply chain attacks \cite{kshetri2022economics}) and unknown ones have made it harder or impossible to prepare the system ahead of time and achieve perfect security. Instead, resilience is a critical aspect of cyber protection that provides a supplemental means to mitigate their impact once the traditional methods fail to thwart them. Resilience plays an even more important role in defending against unknown or unanticipated attacks as it becomes the last resort to safeguard ICSs from calamitous collapse.
An ICS is a large-scale system of systems. Its resilience is a composition of the resilience of many interacting subsystems. Each subsystem consists of its IT and OT units. Hence the resilience of the system relies on the resilience of IT and OT. For a given system, subsystem, or unit, its resilience is defined by its ability to respond to an attack, maintain its function after the attack, and recover from the attack. Illustrated in Fig.~\ref{fig:resilience_performance}, we divide the operation into three phases. The first phase is the prevention, or known as the ante impetum stage, where less sophisticated attacks are thwarted while some attacks launched at time $t_1$ succeed in bypassing the defense and starting to navigate within the system at $t_2$. The second phase is the interim impetum stage (response). The attack breaches the security defense and aims to look for and take over the control of the targeted resources.
For example, in the infamous Target data breach \cite{plachkinova2018security}, the APT attacker, after infiltrating the network by leveraging the third-party vendor’s security, moves laterally from less sensitive areas of Target’s network to areas storing consumer data. A resilient mechanism at this stage aims to respond to this breach and deter the attacker from moving forward to reach the target. Without it, the system performance will progressively deteriorate, and the system will eventually break down. In contrast, an appropriate resilient mechanism detects the breach at $t_3$ and immediately foils the attack by taking appropriate measures (e.g., disconnecting the subsystem, rebooting the system, or triggering a new set of authentication rules) to mitigate the risk. The third stage is the post impetum stage (recovery), where the defender can remove the attacker from the system and recover the compromised system to its ante impetum state or maintain an acceptable performance at the loss of the performance level $D$. Similar concepts on resilience are also discussed in \cite{mcjunkin2017electricity}.
\begin{figure}
\centering
\includegraphics[scale=0.14]{plots/ot_performance.png}
\caption{The resilience of an ICS can be measured by the OT-level performance. A resilient ICS experiences three stages after an attack is launched at $t_1$, including prevention, response, and recovery. A $(T,D)$-resilient ICS to the attack can agily respond to the attack by recovering within $T$ units of time and maintaining a loss of performance $D$. The IT-level performance is subsumed into the OT-level one. The IT-level performance determines the success rate and the timing of an attack.}
\label{fig:resilience_performance}
\end{figure}
\subsubsection{Metrics for Resilience}
For a given attack, we say that an ICS is $(T,D)$-resilient if the system can recover from the attack within $T$ units of time and maintain a maximum loss of performance $D$. Here, $T$ is a measure of how quickly the response is. From Fig.~\ref{fig:resilience_performance}, $T=t_4-t_2$ can be further viewed as the duration of the response stage. The performance loss $D$ quantifies the post impetum performance. A non-resilient system goes through an irreversible performance degradation during the operation, as depicted counterfactual-wise in Fig.~\ref{fig:resilience_performance}.
The concept of $(T,D)$-resilience provides basic metrics for ICSs. Another commonly used metric for resilience is the measure of the total performance loss after an attack. The maximum magnitude of performance loss $M$ at the interim impetum stage can also serve as an indicator of resilience. Note that the metrics discussed above build on the performance measures of the OT system or the operation at the OT level. It is important to recognize that the defense effort at the prevention stage in the IT system has a significant impact on the metrics even though it is not directly measured, and it is quantified through the OT performance. The effort to thwart the attack, such as the deployment of honeypots \cite{huang2019honeypot}, moving target defenses \cite{zhu2013game}, and zero-trust mechanisms \cite{rose2020zero}, creates more difficulty for the attacker to be successful, thus reducing the rate of successful attacks and the impact of their consequences. In Fig.~\ref{fig:resilience_performance}, this effort can be roughly measured by the time that it takes for an attack in the system to compromise the targeted OT assets. Many recent works on cyber resilience focus on resilience at this stage \cite{segovia2020cyber,mertoguno2019physics}. We can zoom into the IT system and further quantify the cyber resilience of an ICS. For example, in \cite{kerman2020implementing}, a zero-trust cyber resilient mechanism is developed for an enterprise network. The trustworthiness of each user is measured over time to disrupt unknown insiders from lateral movement. The resilience of the IT system is measured by the time it takes to deter an attack from the target asset and the distance between the attacker and the asset. Fig.~\ref{fig:resilience_performance} provides a resilience measure of ICSs through the OT performance which consolidates the impact and the performance of the IT system. We can also zoom into the prevention stage and create a similar performance measure for the IT system.
\subsubsection{Comparison with Related Concepts}
There are subtle differences between resilience and related concepts, such as fault-tolerance, robustness, and security. Fault tolerance is also a system property that aims to enable the system to maintain its core functions in the event of failures. Illustrated in Fig.~\ref{fig:resilience_fault_tolerance}, fault tolerance often prepares for anticipated failures, and those happen internally caused by fatigue, corrosion, or manufacturing flaws. Resilience, in contrast, deals with unpredictable ones and those caused by external influences, such as cyber-attacks, natural disasters, and terrorism. The event that the system aims to prepare for distinguishes the type of system properties that we need to focus on.
\begin{figure}
\centering
\includegraphics[scale=0.14]{plots/resilience_fault_tolerance.png}
\caption{The distinction between fault tolerance and resilience from the dimensions of unpredictability and the externality of the events. The two concepts overlap for some events.}
\label{fig:resilience_fault_tolerance}
\end{figure}
Robustness is a well-known system concept for control systems. The goal of robustness is to design systems to withstand a set of prescribed uncertainties or events. Robustness prepares for them and enables the system to function when they occur, which is beneficial for resilience planning. However, it does not focus on preparing the system for recovering from an unknown event that has disrupted the system. It is often either cost-prohibitive or nonviable to design systems robust to all possible events while maintaining their functions.
Robustness and resilience are complementary to each other. The low-probability events that are costly for robustness should be prepared by resilient mechanisms while the high-probability events, such as thermal noise and load variations, should be handled through robust designs. The concept of robustness is often used for OT, e.g., controller designs and operation planning.
Security is a related concept that focuses on the prevention of adversarial events to safeguard the confidentiality, integrity, and availability of IT systems. Despite the effort to create a cyber defense to detect and foil attacks from penetrating IT, there is no perfect security that can assure the system of no attacks. In contrast, cyber resilience plays a significant role that enables the IT and OT systems to function even though the attack is in the network already.
\subsection{Resilience of Distributed Systems}
ICS are large-scale systems. Modern-day power grids are composed of many distributed energy resources at the edge of the power grid. The increasing interdependencies and connectivities among subsystems have exposed the ICS to a large attack surface. As illustrated in Fig.~\ref{fig:attack_ics}, an attacker can first gain the privilege of an IT subsystem through (a) and then move laterally within the IT network to find the target asset through (b). The attacker takes down the OT asset of the targeted subsystem, and its compromise leads to cascading failures in the OT system. This type of incident has been witnessed in the Ukrainian power grid attack \cite{case2016analysis} and the oil pipeline system attack \cite{colonial2021wiki}.
They both are attributed to APTs, which carry out a kill chain consisting of a sequence of resourceful, stealthy, and strategic attacks to reach their target.
\begin{figure}[h]
\centering
\includegraphics[scale=0.23]{plots/attack_cps.png}
\caption{Some illustrative attacks on subsystems of an ICS. The connectivities among subsystems increase the attack surface. An attacker can move laterally through the IT network from (a) to (b) and locates the OT asset using (c). The attack can cause cascading failures through (d).}
\label{fig:attack_ics}
\end{figure}
As the attackers are growingly sophisticated, resilience becomes increasingly essential for ICSs. However, it is challenging for large-scale systems through centralized control and operations. One-point failure can propagate to the entire system, and a centralized operation is less flexible as the entire system needs to be reconfigured for the failure.
Distributed control, on the other hand, enables the agility of the entire system. The system can still maintain its core functions well even when a subset of subsystems fails. In addition, as the decisions are distributed, it will be more convenient and faster for each subsystem to respond to the disruption and its connected subsystems. The challenge, however, to achieve it is the need for machine intelligence to enable detection and response through fast data analytics and decision-making.
Hence, to enable distributed and collaborative resilience, we need to create an AI stack between the IT and OT systems of each subsystem. Illustrated in Fig.~\ref{fig:ai_stack}, AI agents are introduced to connect cyber IT agents (C) and physical OT agents (P)\footnote{Here we call the IT and OT systems as agents to align with the names in our MAS framework. AI agents refer to the introduced AI components such as computational devices.}.
\begin{figure}
\centering
\includegraphics[scale=0.31]{plots/ai_stack.png}
\caption{An illustration of interactions between AI, IT, and OT agents of three subsystems. The interactions include (1) interface between a C agent and an AI agent, (2) interface between a P agent and an AI agent, (3) interface between two AI agents in two subsystems.
The C and P agents from different subsystems can also be interdependent. The AI agents interact with C and P agents in the same subsystem and other subsystems to achieve distributed and collaborative resilience.
}
\label{fig:ai_stack}
\end{figure}
A more detailed and functional diagram of an AI-augmented subsystem is also demonstrated in Fig.~\ref{fig:ai_agent}. An AI agent serves as an interface between a C agent and a P agent and is equipped with necessary functional blocks, including diagnosis, prognosis, simulation, data, strategic planning, and machine learning, to create fast online responses to an unanticipated event. The AI agent can also interact with other connected AI agents. This architecture enables the AI agent to control the performance of its associated subsystem as well as coordinate with other subsystems.
When a cyber-attack occurs locally on the subsystem, the AI agent can monitor the behaviors of the C and P agents in its subsystem and make tactical control and planning decisions to respond to the event. Each subsystem can achieve its own resilience in a distributed fashion. We refer to this type of resilient mechanism for the large-scale ICS as distributed resilience. As the failures can propagate across multiple subsystems, the AI agents need to communicate and coordinate to achieve collaborative resilience. Sharing of information and intelligence among AI agents can not only improve the distributed resilience of their subsystem but also enhance the global resilience of the large-scale system. For example, in \cite{achbarou2018new}, we have seen that the sharing of information in intrusion detection systems can help the entire network defend against zero-day attacks. In \cite{sun2021data}, the collaborations among multiple machine learners have led to significant improvement in the learning as well as resilience to data poisoning attacks.
\begin{figure}
\centering
\includegraphics[scale=0.22]{plots/ai_agent.png}
\caption{An illustrative functional diagram of an AI-augmented subsystem. An AI agent of a subsystem monitors the behaviors of its C and P agents and communicates with other AI agents to create tactical control and planning decisions to respond to malicious attacks and disturbances.}
\label{fig:ai_agent}
\end{figure}
|
\section{Introduction}
\label{sec:introduction}
Debris discs or belts have now been detected in over a hundred extrasolar systems. They are composed of a large variety of small bodies, with sizes ranging from micrometer (dust) to kilometer (planetesimals) \citep[see e.g.,][]{hughesDebrisDisksStructure2018}. The larger planetesimals (parent bodies) constantly replenish the population of short-lived dust through collisional cascades, which allows the disc to be detected. While small \textmu m-sized particles (that scatter light in the visible/near-infrared wavelength) can orbit far from the sites of their initial collisional formation \citep[e.g.,][]{wyattSpiralStructureWhen2005}, mm-sized particles (that emit in the millimetre wavelength) tend to remain near their production sites. They can thus be used to trace the orbital architecture of the parent bodies.
In the Solar System, the location, density and dynamical pattern of the asteroid and Edgeworth-Kuiper belts have long served as a useful diagnosis to constrain the formation and evolutionary history of the planets. The current structure of these debris belts has largely been shaped by their interaction with the surrounding planets, not only through the violent clearing of the orbital paths, but also through more subtle resonant and secular processes. Similarly, we expect that the structure of extrasolar debris belts can help us infer the presence of planetary perturbers and their dynamical history \citep[e.g.,][]{raymondDebrisDisksSignposts2011,pearcePlanetPopulationsInferred2022,guoPlanetesimalDynamicsPresence2022}.
In the past decades, a number of debris discs have been observed to exhibit an overall eccentric shape \citep[e.g.,][]{telescoDeep10182000,kalasPlanetarySystemOrigin2005, wyattSpiralStructureWhen2005,eiroaColdDUstNEarby2010,kristHubbleSpaceTelescope2012,macgregorALMAImagesEccentric2022}. The most common explanation for this breaking of the circular symmetry is the presence of an eccentric planet in the neighbourhood of the disc \citep{wyattHowObservationsCircumstellar1999}. This planetary perturber defines a preferred direction characterized by its longitude of periastron, and induces a so-called forced eccentricity $e_{\rm forced}$ (proportional to the planetary eccentricity) in the planetesimal disc. If the planetesimals have initially circular orbits with semi-major axis $a$, then the eccentric disc would be apsidally aligned with the planet with a width of at least $\Delta a = 2 a e_{\rm forced}$ (e.g., \citealt{kennedyUnexpectedNarrownessEccentric2020}; see Section 2.2); the width can be larger if the debris belt has a spread in semi-major axis. However, several observed eccentric debris discs have been found to challenge this simple picture. In particular, at least three systems whose millimetre emissions have been observed by ALMA contain a debris belt with width narrower than the theoretical minimum \citep[as first noted by][]{kennedyUnexpectedNarrownessEccentric2020}: (i) HD 202628 has a $22$ au-wide debris belt located at $150$ au from its G-type host star \citep{faramazScatteredlightMillimeterEmission2019}. The belt has an eccentricity $0.09$, with the corresponding theoretical minimal width of $27$ au, larger than the observed width \citep[22 au,][]{kennedyUnexpectedNarrownessEccentric2020}. (ii) Fomalhaut is a well-known system comprising a massive A star and a narrow eccentric ring at $142$ au \citep{macgregorCompleteALMAMap2017,kennedyUnexpectedNarrownessEccentric2020}. This ring has an eccentricity $0.12$ and a width of $12$ au, less than half the theoretical minimum of $34$ au. This system may harbour a planetary candidate \citep{kalasOpticalImagesExosolar2008}, although recent observations suggest it is a dispersing cloud of dust, perhaps due to a collision between unseen planetary companions \citep{gasparNewHSTData2020}. (iii) HD 53143 has the most eccentric debris disc observed to date, with an eccentricity of $0.21\pm 0.02$ \citep{macgregorALMAImagesEccentric2022}. It orbits a Solar-type star at $90$ au, and has a width of $19.7 \pm 2.5$ au, about half the theoretical minimum width of $38 \pm 4$ au. All three systems are at least several hundreds Myr old.
Several hypotheses could account for the small widths of these long-lived debris belts. First, the embedded planetesimals could have finite primordial eccentricities. To reduce the width of the belt however, the primordial eccentricity vector directions should be clustered around that of the planet---such primordial alignment is not expected to occur naturally. Second, a gas-induced damping force could act to push the planetesimal eccentricity vectors toward their equilibrium values \citep[e.g.,][]{panSelfconsistentSizeVelocity2012,linSculptingEccentricDebris2019}. Since most of the gas in the system is gone after the dissipation of the protoplanetary disc, we expect the damping force to act only in the very beginning of the system's life. This requires the planet to have acquired its mass and eccentricity at a sufficiently early time. Alternatively, the damping could be induced by collisions between planetesimals in the disc \citep{kennedyUnexpectedNarrownessEccentric2020}. This effect would vary with the sizes of the planetesimals, so that we would observe different belt widths in different wavelengths. Finally, other hypotheses have been suggested that do not involve a single planet: the disc eccentricity pattern could be due to self-stirring or to a single disrupting event (such as a collision between large planetesimals), or the disc could be shaped by several shepherding planets \citep{kennedyUnexpectedNarrownessEccentric2020}.
In this paper, we explore several new promising pathways to account for the small width of eccentric debris belts using a single planetary perturber. In Section~\ref{sec:exploration}, we lay down the theoretical framework for the debris belt eccentricity evolution assuming an evolving eccentric planet. In the following sections, we then study how this framework applies to the joint evolution of ``planet + debris disc'' systems. In Section~\ref{sec:ppdisc}, we consider the effect of the protoplanetary disc, and show that the gas-induced eccentricity excitation of the planet and the eccentricity damping of the planetesimals can both reduce the width of the final debris belt. In Section~\ref{sec:ppscattering}, we consider the scenario in which the planet eccentricity grows stochastically due to planet-planet scatterings, and show that this process has a significant probability of leading to a narrow debris belt when the scattering stops. We summarize our findings in Section~\ref{sec:conclusion}, and discuss their limitations.
\section{Test Mass perturbed by an Evolving Eccentric Planet}
\label{sec:exploration}
In this section, we summarize the key equations that govern the eccentricity evolution of a test mass under the influence of an eccentric planet, allowing the planet's mass $m_{\rm p}$ and eccentricity $e_{\rm p}$ to change in time. We assume that the orbital migration of the planet is negligible, so that the semi-major axis ratio between the planet and planetesimals remains constant.
Let $a_{\rm p}$ and $a$ be the semi-major axes of the planet and test particle. Throughout this paper, we consider debris discs exterior to the planet, thus $a > a_{\rm p}$. Let $e$ and $\varpi$ be the eccentricity and longitude of periastron of the test mass. We define the complex eccentricity
\begin{equation}
\mathcal{E}(t) \equiv e(\cos\varpi + i \sin\varpi),
\end{equation}
and similarly $\mathcal{E}_{\rm p}(t) \equiv e_{\rm p} \exp(i\varpi_{\rm p})$ for the planet.
For small $e_{\rm p}$ and $e$, the secular equation governing the evolution of the complex eccentricity of the test particle is
\begin{equation}
\dv{\mathcal{E}}{t}\,(t) = i\omega(t)\mathcal{E}(t) - i \nu(t) \mathcal{E}_{\rm p}(t), \label{eq:eqdiff}
\end{equation}
where $\omega$ and $\nu$ are frequencies given by
\begin{align}
&\omega(t) = \frac{1}{4} \mu_{\rm p}(t) \alpha b_\frac{3}{2}^{(1)}(\alpha) n = \frac{3}{4} \alpha^2 \mu_{\rm p}(t) n + O(\alpha^4),\label{eq:frequency}\\
&\nu(t) = \frac{1}{4} \mu_{\rm p}(t) \alpha b_\frac{3}{2}^{(2)}(\alpha) n = \frac{15}{16} \alpha^3 \mu_{\rm p}(t) n + O(\alpha^4).\label{eq:nu}
\end{align}
Here $\mu_{\rm p} = m_{\rm p}/M_*$ ($M_*$ is the mass of the host star), $n$ is the mean-motion of the test mass, $\alpha = a_{\rm p}/a$ and $b_i^{(j)}(\alpha)$ are the Laplace coefficients \citep{murraySolarSystemDynamics2000}. The second equalities in equations~\eqref{eq:frequency}--\eqref{eq:nu} are valid for $\alpha \ll 1$.
Equation \eqref{eq:eqdiff} has the formal solution:
\begin{equation}
\mathcal{E}(t) = \left( \int_{0}^t \left[-i\nu(t')\right] \mathcal{E}_{\rm p}(t') \exp\left[-i\mathcal{W}(t')\right] \dd{t'} + \mathcal{E}_0 \right) \exp \left[i\mathcal{W}(t)\right], \label{eq:solution}
\end{equation}
where $\mathcal{E}_0 = \mathcal{E}(t=0)$ and
\begin{equation}
\mathcal{W}(t) \equiv \int_{0}^t \omega(t') \dd{t'}. \label{eq:W}
\end{equation}
\subsection{Constant planetary mass and eccentricity}
\label{sec:classical}
Suppose the planetary mass and eccentricity grow instantly at $t=0$ and remain constant ($= \mathcal{E}_{\rm p}$) thereafter. Equation~\eqref{eq:solution} then gives
\begin{equation}
\mathcal{E}(t) = \frac{\nu}{\omega} \mathcal{E}_{\rm p} + \left(\mathcal{E}_0 - \frac{\nu}{\omega} \mathcal{E}_{\rm p} \right) \exp(i\omega t). \label{eq:classical}
\end{equation}
We see that the test mass eccentricity $\mathcal{E}(t)$ follows a circular trajectory in the complex plane. This trajectory's centre is called the ``forced eccentricity'' $e_{\rm forced}$ and its radius the ``free eccentricity'' $e_{\rm free}$:
\begin{align}
&e_{\rm forced} = \frac{\nu}{\omega} e_{\rm p} \simeq \frac{5}{4} \alpha e_{\rm p};\\ &e_{\rm free} = |\mathcal{E}_0 - \frac{\nu}{\omega} \mathcal{E}_{\rm p}|. \label{eq:eforced}
\end{align}
\subsection{Ensemble of particles}
If the debris belt were infinitely narrow initially, then it would remain an infinitely narrow belt undergoing coherent precession with an eccentricity oscillating between $|e_{\rm forced} - e_{\rm free}|$ and $(e_{\rm forced} + e_{\rm free})$. However, since a real belt inevitably has a non-zero width, the ensemble of test particles will have a range of precession frequencies $\omega$ linked to their semi-major axis distribution. The belt will lose its coherence in multiple precession periods, and rings of different eccentricities and longitudes of periastron will co-exist. The resulting structure is a belt of mean eccentricity
\begin{equation}
\langle e \rangle = e_{\rm forced}, \label{eq:emean}
\end{equation}
and width
\begin{equation}
\Delta r \simeq 2 \bar{a} e_{\rm free} + \Delta a, \label{eq:da}
\end{equation}
where $\bar{a}$ and $\Delta a$ are the mean value and spread of the semi-major axes of the test particles.
To show equations~\eqref{eq:emean} and \eqref{eq:da} explicitly, we consider an ensemble of particles with similar semi-major axes $a \simeq \bar{a}$ and write their eccentricities as
\begin{equation}
\mathcal{E}(\mathcal{W}) \equiv e(\mathcal{W}) \exp\left[i\varpi(\mathcal{W})\right] = e_{\rm free} \exp(i\mathcal{W}) + e_{\rm forced}.
\end{equation}
At a given time $t$, different particles would have different precession phases $\mathcal{W} = \omega t$ (because of the spread in $a$ and $\omega$). A given orbit characterized by $\mathcal{W}$ has a trajectory described by the polar equation
\begin{align}
r(\theta, \mathcal{W}) &{}= \frac{\bar{a} \left[1-e^2(\mathcal{W})\right]}{1 + e(\mathcal{W})\cos\left[\theta-\varpi(\mathcal{W})\right]}\nonumber\\
&{} = \bar{a} \left[ 1 - e_{\rm free} \cos\left(\mathcal{W}-\theta\right) - e_{\rm forced} \cos\theta \right]+ O(e^2),
\end{align}
where $\theta$ is the polar angle and $r$ the radius. At a given $\theta$, the minimum and maximum radii in the ensemble orbits are given by
\begin{align}
&r_{\rm min} = r(\theta, \mathcal{W} = \theta) \simeq \bar{a} \left[ 1 - e_{\rm free} - e_{\rm forced} \cos\theta \right]\\
&r_{\rm max} = r(\theta, \mathcal{W} = \theta+\upi) \simeq \bar{a} \left[ 1 + e_{\rm free} - e_{\rm forced} \cos\theta \right].
\end{align}
The average radius is then:
\begin{align}
r_{\rm av}(\theta) = \frac{1}{2}\left(r_{\rm min} + r_{\rm max} \right) = \bar{a}\left(1 - e_{\rm forced} \cos\theta\right). \label{eq:rav}
\end{align}
Thus, the apparent orbit of this ensemble has an eccentricity $e_{\rm forced}$. The width of the ensemble is
\begin{align}
\Delta r(\theta) = r_{\rm max} - r_{\rm min} = 2 \bar{a} e_{\rm free}.\label{eq:deltar}
\end{align}
These results are illustrated in Figure~\ref{fig:classical}.
\begin{figure*}
\centering
\includegraphics[width=0.45\linewidth]{ExamplePhase.pdf}
\includegraphics[width=0.45\linewidth]{NonCoherentBelt.pdf}
\caption{Schematic of the standard picture of an initially circular belt of test particles excited by an inner injected planet of eccentricity $e_{\rm p}$ (equation~\ref{eq:classical}). Left: The coloured orbits correspond to different phases of the precession. The light grey orbit represents the initial circular belt for reference. Right: Superposition of $50$ orbits with incoherent phases, but similar semi-major axes. It can be approximated by an ellipse with eccentricity $e_{\rm forced}$ and width $2 a e_{\rm free}$ (see equations~\ref{eq:rav} and \ref{eq:deltar}).} \label{fig:classical}
\end{figure*}
\subsection{Eccentricity damping}
\label{sec:eccentricitydamping}
A common hypothesis to explain the narrow eccentric debris belts in extrasolar systems (see Section~\ref{sec:introduction}) is the presence of an eccentricity-damping force (e.g. due to the friction from the surrounding gas) acting on the test particles. Equation~\eqref{eq:eqdiff} becomes
\begin{equation}
\dv{\mathcal{E}}{t} = \left(i\omega - \frac{1}{\tau_{\rm d}}\right)\mathcal{E} - i \nu \mathcal{E}_{\rm p}, \label{eq:eqdiff_damping}
\end{equation}
where $\tau_{\rm d}$ is the eccentricity damping time. For $\mathcal{E}_0 = 0$, equation~\eqref{eq:eqdiff_damping} has the solution
\begin{equation}
\mathcal{E}(t) = \frac{\nu}{\omega} \mathcal{E}_{\rm p} \frac{\tau_{\rm d}\omega}{i+\tau_{\rm d}\omega} \left[1-\exp \left(i\omega t - \frac{t}{\tau_{\rm d}} \right)\right]. \label{eq:damping}
\end{equation}
For $t \gg \tau_{\rm d}$, $\mathcal{E}(t)$ approaches a fixed value with zero free eccentricity:
\begin{equation}
\mathcal{E}(t) \xrightarrow[t\to\infty]{} \frac{\nu}{\omega}\mathcal{E}_{\rm p} \frac{\tau_{\rm d}\omega}{i+\tau_{\rm d}\omega}. \label{eq:shift}
\end{equation}
Not that for $\omega\tau_{\rm d} \gg 1$, this asymptotic eccentricity reduces to $e_{\rm forced} = (\nu/\omega) e_{\rm p}$. But if $\omega\tau_{\rm d}$ is not much larger than unity, $\mathcal{E}(t\to \infty)$ is not equal to $e_{\rm forced}$ and the eccentric disc is not apsidally aligned with the planetary perturber.
\subsection{Planetary Mass Growth}
Now consider the case where the planet has a constant eccentricity, but its mass $m_{\rm p}(t)$ is growing with time.
Both frequencies $\omega$ and $\nu$ then depend on time. However, since the ratio $\nu(t)/\omega(t)$ is constant, equation~\eqref{eq:solution} can be integrated out explicitly, which gives
\begin{equation}
\mathcal{E}(t) = \frac{\nu}{\omega} \mathcal{E}_{\rm p} + \left(\mathcal{E}_0 - \frac{\nu}{\omega} \mathcal{E}_{\rm p} \right) \exp\left[i \mathcal{W}(t)\right].\label{eq:massgrowth}
\end{equation}
Comparing to equation~\eqref{eq:classical}, we see that although the time evolution is different from the constant-$m_{\rm p}$ case, the eccentricity has exactly the same trajectory in the complex plane. This is a fundamental property of secular perturbations: their amplitude depends on the semi-major axis ratio and planet eccentricity but does not depend on the planet mass (which only impacts the timescale). Therefore, mass growth alone cannot reduce the free eccentricity and the width of the debris belt.
\subsection{Planetary Eccentricity Growth}
\label{sec:eccentricitygrowth}
Now consider the case where the planet eccentricity grows linearly in time until reaching the final value $e_{\rm p, f}$, i.e.
\begin{align}
e_{\rm p}(t) = \begin{cases}
e_{\rm p, f} \frac{t}{t_{\rm p}} &\text{for } 0 \leq t \leq t_{\rm p}\\
e_{\rm p, f} &\text{for } t > t_{\rm p}.
\end{cases}\label{eq:ecc}
\end{align}
We assume $\varpi_{\rm p} = 0$ for simplicity. Again, we can integrate equation~\eqref{eq:solution} to obtain the complex eccentricity of the test particle:
\begin{align}
&\mathcal{E}(t) = \nonumber\\&\begin{cases}
e_\mathrm{forced} \frac{1}{t_{\rm p}} \left[t +\frac{i}{\omega}(\exp i\omega t -1) \right] &\text{for } 0 \leq t \leq t_{\rm p} \\
e_\mathrm{forced} + \left[\mathcal{E}(t_{\rm p}) - e_{\rm forced}\right] \exp\left[i\omega (t-t_{\rm p})\right] &\text{for } t > t_{\rm p},
\end{cases} \label{eq:epstau}
\end{align}
where
\begin{equation}
e_\mathrm{forced} = \frac{\nu}{\omega} e_{\rm p, f} \label{eq:eforced_egrowth}
\end{equation}
is the ``final'' forced eccentricity. Clearly, the free eccentricity of the test mass at $t>t_{\rm p}$ is
\begin{align}
e_{\rm free} ={}& \left|\mathcal{E}(t_{\rm p}) - e_{\rm forced}\right|\nonumber\\
={}& e_{\rm forced} \left|\frac{2\sin\left(\frac{\omega t_{\rm p}}{2}\right)}{\omega t_{\rm p}}\right|\label{eq:efree_egrowth}
\end{align}
From equation~\eqref{eq:efree_egrowth}, we see that the free eccentricity can be lower than the forced eccentricity if $\omega t_{\rm p} \gtrsim 1$, i.e. if the planet's eccentricity growth is sufficiently slow. An example of the test particle eccentricity evolution in the complex plane in such a case is shown in Figure~\ref{fig:ecomplex}.
A similar analysis can be conducted for different eccentricity functions: quadratic $\left[e_{\rm p} \propto (t/t_{\rm p})^2\right]$ and exponential $\left[e_{\rm p} \propto \exp(t/t_{\rm p})\right]$. The results are shown in Figure~\ref{fig:efree}. We note that the final free eccentricity depends less on the growth functional form than on the timescale $t_{\rm p}$: if $t_{\rm p}$ is greater than the precession time $t_\omega = 2\upi/\omega$, then the free eccentricity is less than half of the final forced eccentricity, and so is the expected debris belt width.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{PhaseSpace_eccentricitygrowth.pdf}
\caption{Evolution of the complex eccentricity of a test particle (blue dots, equations~\ref{eq:epstau}) exterior to a planet with a linear eccentricity growth (equations~\ref{eq:ecc}), in units of the final forced eccentricity. We adopt the growth time $t_{\rm p} = 2.5 ~t_\omega$, where $t_\omega = 2\upi/\omega$. The test particle eccentricity is initially zero, its evolution is represented by $100$ dots sampled uniformly between $t = 0$ and $t_{\rm p}+t_\omega$. The corresponding forced eccentricity $\frac{\nu}{\omega}\mathcal{E}_{\rm p}(t)$ is represented by black dots linked with a grey line. The orange dot shows the final forced eccentricity (equation~\ref{eq:eforced_egrowth}), and the radius of the light blue circle is the final free eccentricity (equation~\ref{eq:efree_egrowth}). At each time-step, the complex eccentricity of the test particle (in blue) rotates anticlockwise around the current forced eccentricity (in black), which results in the half ellipse. At $x=0.4$, $e_\mathrm{forced}$ `overtakes' $\mathcal{E}$, so the half ellipse starts the cycle anew.}\label{fig:ecomplex}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{FreeEccentricity.pdf}
\caption{Final free eccentricity of an exterior test particle (in units of the final forced eccentricity, equation~\ref{eq:eforced_egrowth}) as a function of the planet eccentricity time growth $t_{\rm p}$ (in units of the precession period $t_\omega = 2\upi/\omega$, equation~\ref{eq:efree_egrowth}), for three different types of eccentricity growth functions (linear, quadratic, exponential). In all three cases, the envelope of the final free eccentricity decreases as $1/t_{\rm p}$. The lower the free eccentricity, the narrower the debris belt (equation~\ref{eq:da}).}\label{fig:efree}
\end{figure}
\section{Interaction with protoplanetary disc}
\label{sec:ppdisc}
In the previous section, we have presented a general analysis for the free eccentricity of a test particle under the influence of an evolving planet. This suggests two mechanisms of reducing the width of eccentric debris discs: eccentricity damping forces acting on the planetesimals and slow eccentricity growth of the perturbing planet. In this section, we apply these general mathematical solutions to the physical picture of ``planet+disc'' evolution, in which the gaseous protoplanetary disc damps the eccentricity of the planetesimals while exciting the eccentricity of the giant planet.
\subsection{Planet eccentricity growth in gas disc}
\label{sec:planetembedded}
Massive (gap-opening) planets interacting with protoplanetary discs can experience eccentricity growth \citep[e.g.;][]{goldreichEccentricityEvolutionPlanets2003}. Recent numerical simulations indicate that the growth time can be as long as $10^5$--$10^6$ planet orbits \citep[e.g.;][]{ragusaEccentricityEvolutionPlanetdisc2018}. On the other hand, the precession time of the planetesimals driven by the planet is (see equation~\ref{eq:frequency}):
\begin{equation}
t_\omega = \frac{2\upi}{\omega} \simeq 6\times 10^4~P_{\rm p} \left(\frac{a}{3a_{\rm p}}\right)^\frac{7}{2} \left(\frac{\mu_{\rm p}}{10^{-3}}\right)^{-1}, \label{eq:tw}
\end{equation}
where $P_{\rm p}$ is the orbital period of the planet.
We have seen in Figure~\ref{fig:efree} that as long as $t_{\rm p} \gtrsim t_\omega$, the free eccentricity of the test particle is less than $20\%$ the forced eccentricity, which would significantly reduce the physical width of the planetesimal belt. Equation~\eqref{eq:tw} indicates that to have $t_\omega \lesssim 10^5~P_{\rm p}$ would require $a/a_{\rm p} \lesssim 3.4~(\mu_{\rm p}/10^{-3})^{2/7}$, i.e. the perturbing planet must be very close to the planetesimal belt.
Since the planetary eccentricity growth time must be less than the lifetime $t_{\rm life}$ of the disc ($\sim$ Myrs), a necessary condition to reduce the width of debris belt is $t_{\omega} < t_{\rm life}$. This gives
\begin{equation}
\left(\frac{P_{\rm p}}{10~\mathrm{yr}}\right) \left(\frac{a}{3a_{\rm p}}\right)^\frac{7}{2} \lesssim \left(\frac{\mu_{\rm p}}{10^{-3}}\right).
\end{equation}
Therefore, this process requires the planet to be relatively close to both the star and the planetesimal belt.
\subsection{Planetesimal eccentricity damping in gas disc}
\label{sec:tpembedded}
Let us now consider the situation where the debris belt of planetesimals is embedded in a gaseous disc, while the planetary orbit remains constant. Such a configuration could occur in transition discs, where the planet lies inside a gas-free cavity, like has been observed in the benchmark system PDS 70 \citep{mullerOrbitalAtmosphericCharacterization2018}.
In Section~\ref{sec:eccentricitydamping}, we considered the effect of linear eccentricity damping ($\dot{e} = - e/\tau_{\rm d}$) on the free eccentricity of test particles. Such linear damping does not always apply to planetesimal-gas disc interaction. In fact, the frictional force acting on a planetesimal of radius $R$ from a gas disc of density $\rho_{\rm g}$ takes the form \citep[e.g., ][]{grishinApplicationGasDynamical2016}:
\begin{equation}
\vec{F} = -\frac{1}{2} C_{\rm d} \upi R^2 \rho_{\rm g} |\Delta v| \vec{\Delta v}, \label{eq:frictionforce}
\end{equation}
where $\vec{\Delta v}$ is the difference between the planetesimal velocity and the Keplerian velocity of the surrounding gas, and $C_{\rm d}$ is the drag coefficient ($\sim 1$ for planetesimals of $R \sim 1$--$100$ km).
The gas density $\rho_{\rm g}$ is related to the surface density $\Sigma_{\rm g}$ and disc scale height $H$ by
\begin{equation}
\rho_{\rm g} = \frac{\Sigma_{\rm g}}{2 H}.
\end{equation}
Using the disc aspect ratio $h = H/a = 0.1$ and assuming $\Sigma_{\rm g} = 2.10^{3}~\mathrm{g.cm^{-2}} (a/1~\mathrm{au})^{-3/2}$ \citep[the minimum mass solar nebula;][]{hayashiStructureSolarNebula1981}, we get
\begin{equation}
\rho_{\rm g} \simeq 10^{-9} ~\mathrm{g.cm^{-3}} \left(\frac{a}{1~\mathrm{au}}\right)^{-\frac{5}{2}}.
\end{equation}
The frictional force induces an eccentricity damping (see Appendix~\ref{sec:appendix_eccentricitydamping})
\begin{equation}
\frac{\dot{e}}{e} = -\frac{1}{\tau_{\rm d}(e)} = -\frac{e}{\tau_0}, \label{eq:nonlinearfriction}
\end{equation}
with
\begin{align}
\tau_{0} &{}\simeq \frac{0.5}{C_{\rm d}} \frac{\rho_{\rm pl}}{\rho_{\rm g}} \frac{R}{a} ~P\\
&{}\simeq 10^8~\mathrm{yr}\left(\frac{R}{10~\mathrm{km}}\right) \left(\frac{a}{100~\mathrm{au}}\right)^3 \left(\frac{M_*}{1~\mathrm{M_\odot}}\right)^{-\frac{1}{2}} \nonumber\\
&\qquad \times C_{\rm d}^{-1}\left(\frac{\rho_{\rm pl}}{2~\mathrm{g.cm^{-3}}}\right) \left(\frac{\rho_{\rm g,0}}{10^{-9}~\mathrm{g.cm^{-3} }}\right),\label{eq:tau0}
\end{align}
where $P$ is the orbital period, $\rho_{\rm pl}$ the bulk density of the planetesimal and $\rho_{\rm g, 0}$ the value of the gas density $\rho_{\rm g}$ at $1$ au. Since all quantities have a wide range of possible values, the damping constant $\tau_0$ is uncertain, and can range from $10^2$ to $10^{10}$ years (depending in particular on the location of the planetesimal belt).
\begin{figure}
\centering
\includegraphics[width=\linewidth]{ExampleDamping_nonlinear_multiple.pdf}
\caption{Eccentricity evolution of a test particle exterior to a planet companion, with the particle experiencing a non-linear friction force with the eccentricity damping timescale $\tau_{\rm d}(e) = \tau_{0}/e$ (equation~\ref{eq:nonlinearfriction}). Each line corresponds to a different ratio between $\tau_{0}$ and the precession timescale $t_\omega = 2\upi/\omega$. All cases are for $e_{\rm forced} = 0.1$. We see that the damping acts on the timescale of about $\tau_{\rm d}(e_{\rm forced}) \simeq 10 ~\tau_{0}$. }\label{fig:efree_friction}
\end{figure}
The evolution of the complex eccentricity $\mathcal{E}$ of a planetesimal driven by the perturbing planet is given by
\begin{equation}
\dv{\mathcal{E}}{t} = i\omega\mathcal{E} - \frac{|\mathcal{E}|}{\tau_{\rm 0}} \mathcal{E} - i \nu \mathcal{E}_{\rm p}.
\end{equation}
This differential equation is non-linear, and cannot be solved analytically. We use the \textsc{Python SciPy} library to solve it numerically, and plot the results of three different eccentricity evolutions in Figure~\ref{fig:efree_friction}.
We see that the test particle's eccentricity is driven toward an equilibrium value close to $e_{\rm forced} = (\nu/\omega) e_{\rm p}$ (see equation~\ref{eq:eeq} in Appendix~\ref{sec:appendix_eccentricitydamping} for the exact value) on the timescale of $\tau_0/e_{\rm forced}$. When the protoplanetary disc dissipates, the free eccentricity of the planetesimals will be roughly equal to the difference between this equilibrium and $e_{\rm forced}$: the width of the belt will thus be greatly reduced compared to the standard picture. However, this process requires the protoplanetary disc to live long enough for the damping to act, that is $\tau_{0}/e_{\rm forced} \lesssim t_{\rm life}$. From equation~\eqref{eq:tau0}, we see that this requires the debris belt to be close-in ($a \lesssim 10$ au).
Moreover, a strong damping force can shift the equilibrium eccentricity away from the forced eccentricity (as described in Section~\ref{sec:eccentricitydamping}), and subsequently hinder the reduction of the free eccentricity. Although the non-linear aspect of the damping (dependence of $\tau_{\rm d}$ on the eccentricity) reduces this effect, it becomes significant for $\tau_{0}\omega \lesssim e_{\rm forced}$ (see Appendix~\ref{sec:appendix_eccentricitydamping}). For the fiducial values we consider in this paper (equations~\ref{eq:tw} and \ref{eq:tau0} with $a = 10$--$100$ au), the damping timescale is larger than the precession timescale, so that the difference between equilibrium and forced eccentricities is negligible. Both timescale constraints are represented on Figure~\ref{fig:tau_vs_a}.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{tau_vs_a.pdf}
\caption{Parameter space ($a$, $\tau_{\rm d}$) that successfully decreases the debris belt free eccentricity (in green). The blue zone corresponds to an inefficient damping, which would require more time than the lifetime of the protoplanetary disc. The orange zone corresponds to a strong damping ($\tau_{\rm d} < t_\omega$, with $\mu_{\rm p} = 10^{-3}$ and $a_{\rm p} = 10$ au), which would shift the forced eccentricity away from its gas-free value. The black dashed line corresponds to $\tau_{\rm d} = \tau_{0}/e_{\rm forced}$ following equation~\eqref{eq:tau0} with the fiducial parameters and $e_{\rm forced} = 0.1$. For the parameters we chose, the free eccentricity can be reduced only for close-in debris belts ($a \lesssim 10$ au). In order to increase the efficiency of this mechanism to wider belts, we would need a more efficient damping (i.e. smaller planetesimals or higher gas density) and a smaller precession period (i.e. planet closer to the belt or with a higher mass).}\label{fig:tau_vs_a}
\end{figure}
\section{Planet-Planet Scattering}
\label{sec:ppscattering}
In the previous section, we showed that the interaction between the planet and debris belt and their surrounding protoplanetary disc can reduce the planet-induced free eccentricity of the planetesimals and thus increase the coherence of the belt. However, this requires either a strong and sustained frictional force or a protoplanetary disc lifetime a few times larger than the precession period of the planetesimals. These conditions may not be met for wide debris belts ($\gtrsim 100$ au) including several observed systems.
In this section, we examine another process to decrease the free eccentricity of the planetesimals, by considering the scenario in which the perturbing planet gains its eccentricity through planet-planet scattering, after the dispersal of the protoplanetary disc. Strong gravitational scatterings is a leading mechanism to produce extrasolar giant planets on eccentric orbits \citep[e.g.;][]{chatterjeeDynamicalOutcomesPlanetPlanet2008,juricDynamicalOriginExtrasolar2008,fordOriginsEccentricExtrasolar2008,andersonSituScatteringWarm2020,liGiantPlanetScatterings2021}. In this scenario, the eccentricities of the planets change in an irregular way (approximatively following random walks) until one of the planets is ejected. Our goal is to understand how the eccentricity of a planetesimal evolves during the planet-planet scattering process.
\subsection{Numerical random-walk model}
\label{sec:randomgrowth_numerical}
$N$-body simulations suggest that planet-planet scatterings can be modelled as a random process \citep{puStrongScatteringsCold2021}. Here we present a toy model where the planet eccentricity grows following a discrete random walk. We suppose that the planet has an initially circular orbit, and that we know its final eccentricity $e_{\rm p,f}$ at the end of the scattering process (e.g. $e_{\rm p, f}$ is the observed eccentricity of the perturbing planet around a debris disc).
The walk thus consists of a collection of instantaneous kicks in the 2D complex plane, beginning at $0$ and ending at $\mathcal{E}_{\rm p} = e_{\rm p,f}$ (fixed) after $N_{\rm p}$ steps. The magnitude of each kick is sampled from a Gaussian distribution of scale $\sqrt{2}\Delta e_{\rm p}$\footnote{ Note that $e_{\rm p}\cos\varpi_{\rm p}$ and $e_{\rm p}\sin\varpi_{\rm p}$ each are sampled from a Gaussian distribution of scale $\Delta e_{\rm p}$, so that $\Delta e_{\rm p}$ is the equivalent to $\sigma_{\rm p}$ in Section 4.2.} and the direction of the kick is random. The scale of the kick and the final value of the planet eccentricity give a most likely value for the number of kicks before the end of the scattering
\begin{equation}
\bar{N}_{\rm p} \equiv \frac{e_{\rm p,f}^2}{2 \Delta e_{\rm p}^2}.
\end{equation}
The actual number of kicks $N_{\rm p}$ has a distribution around $\bar{N}_{\rm p}$.
The complex eccentricity of the test particle (the planetesimal) will try to follow the evolution of the planet eccentricity. The result depends on the number of kicks that occur within one planetesimal precession period, defined as $N_\omega$ [if the typical scattering step lasts $\Delta t$, then $N_\omega \approx (\omega\Delta t)^{-1}$ ]. If $N_\omega$ is much less than the total number of steps $N_{\rm p}$, then the whole scattering process can be seen as adiabatic. On the other hand, if $N_\omega$ is much larger than the total number of steps $N_{\rm p}$, then the process can be considered instantaneous.
According to \cite{puStrongScatteringsCold2021}, the number of close encounters $N_{\rm p}$ of a scattering process between two giant planets follows a Lévy distribution peaking at $\bar{N}_{\rm p}$. In our study, we fix $\bar{N}_{\rm p} = 10^3$ \citep[the value of $\bar{N}_{\rm p}$ depends on the masses of the two planets, see Figs.~3-4 of][]{puStrongScatteringsCold2021}, and we thus produce a set of random eccentricity kicks following such a distribution (see Appendix~\ref{sec:randomwalk_method} for details).
We then use $N_\omega$ to compute the evolution of the complex eccentricity $\mathcal{E}$ of the test particle using equation~\eqref{eq:solution} (assuming $\mathcal{E}_{\rm p}$ remains constant between kicks), and use its final value $\mathcal{E}(N_{\rm p})$ to derive the free eccentricity $e_{\rm free}$ as a function of the final forced eccentricity $e_{\rm forced}$:
\begin{align}
&e_{\rm forced} = \frac{\nu}{\omega} \mathcal{E}_{\rm p} (N_{\rm p}) = \frac{\nu}{\omega} e_{\rm p,f} \label{eq:eforced_random},\\
&e_{\rm free} = |\mathcal{E}(N_{\rm p}) - e_{\rm forced}|. \label{eq:efree_random}
\end{align}
Note that the forced eccentricity does not depend on the evolutionary path taken by the planet eccentricity during the scattering process. However, the free eccentricity does, and depends in particular on $N_{\rm p}$ and $N_\omega$.
Two examples of random walks are shown in Figure~\ref{fig:ecomplex_randomgrowth}, one leading to a free eccentricity of the test particle less than the forced eccentricity, the other leading to a larger free eccentricity. Note that in this stochastic model, the free eccentricity can be either smaller or larger than $e_{\rm forced}$, contrary to the standard picture (see Section~\ref{sec:classical}).
\begin{figure}
\centering
\includegraphics[width=\linewidth]{PhaseSpace_randomgrowth_smallefree.pdf}\\
\includegraphics[width=\linewidth]{PhaseSpace_randomgrowth_largeefree.pdf}
\caption{Examples of the evolution of the complex eccentricity of the perturbing planet (in black) which undergoes planet-planet scattering, and the corresponding evolution of an outer test particle (in blue). These evolutions are computed using the numerical random-walk model presented in Section~\ref{sec:randomgrowth_numerical} with $N_\omega = \bar{N}_{\rm p} = N_{\rm p} = 1000$. The orange dotted line represents the evolution of the test particle at the end of the scattering process. The example in the upper panel leads to a particle having a small free eccentricity, the example on the bottom panel leads to a high free eccentricity.}\label{fig:ecomplex_randomgrowth}
\end{figure}
Figure~\ref{fig:efree_randomgrowthvssteps} shows the free eccentricity of the test particle in $10^4$ random walk samples as a function of their total number of steps $N_{\rm p}$. The distribution of the free eccentricities peaks at the forced eccentricity. For sufficiently large $N_{\rm p}$, the average and spread of free eccentricities appear to grow with $N_{\rm p}$. Conversely, if $N_{\rm p}$ is small, then the free eccentricities converge to the $e_{\rm forced}$---This recovers the result of the standard picture discussed in Section~\ref{sec:classical}, where the planet's eccentricity growth is instantaneous.
In the next subsection, we will demonstrate that the most relevant dependency is with $N_{\rm p}/N_\omega$, and that the root mean square of the free eccentricity has an analytical expression. We thus plot the data set of Figure \ref{fig:efree_randomgrowthvssteps} in the $(N_{\rm p}/N_\omega, e_{\rm free}/e_{\rm forced})$ space in Figure~\ref{fig:efree_randomgrowthvsstepstheory}. For $N_{\rm p} \lesssim \bar{N}_{\rm p}$, we have $\langle e_{\rm free}^2 \rangle \sim e_{\rm forced}^2$, while for $N_{\rm p} \gtrsim \bar{N}_{\rm p}$, $\langle e_{\rm free}^2 \rangle /e_{\rm forced}^2$ is proportional to $N_{\rm p}/N_\omega$. For a fixed $N_{\rm p}/N_\omega$, the free eccentricity distribution does not depend on $N_\omega$ (see Figure~\ref{fig:comparison}).
\begin{figure}
\centering
\includegraphics[width=\linewidth]{FreeEccentricity_RandomGrowth_nmax10_tw5000.pdf}
\caption{Final free eccentricity of the planetesimal as a function of the number of kicks ($N_{\rm p}$) in the planet-planet scattering process, assuming $N_\omega = 5000$ and $\bar{N}_{\rm p} = 1000$. The dots in the bottom left panel represents the individual results of 10,000 walks, the top left and bottom right panels show the histograms depicting the distributions of $N_{\rm p}$ and the free eccentricities, respectively.}\label{fig:efree_randomgrowthvssteps}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{FreeEccentricity_RandomGrowth_nmax10_theory_tw5000.pdf}
\includegraphics[width=\linewidth]{FreeEccentricity_RandomGrowth_nmax10_theory_tw5000_zoom.pdf}
\caption{Same as Figure~\ref{fig:efree_randomgrowthvssteps}, except the $x$-axis is the number of kicks normalized by $N_\omega$ and the $y$-axis represents the squared free eccentricity. The blue line and shade correspond to the moving squared average and standard deviation respectively. The theoretical expectation (equation~\ref{eq:efreerms}) is represented in orange. The bottom panel zooms in the zone of low free eccentricity in the upper panel.}\label{fig:efree_randomgrowthvsstepstheory}
\end{figure}
Since most of the walks have a length close to $\bar{N}_{\rm p}$, the overall distribution of free eccentricities depend mostly on $\bar{N}_{\rm p}/N_\omega$. Figure~\ref{fig:efree_randomgrowth} shows the cumulative distribution of free eccentricities for different $\bar{N}_{\rm p}/N_\omega$. We find that although $\langle e_{\rm free}^2 \rangle$ increases as $\bar{N}_{\rm p}/N_\omega$ increases beyond unity, a significant proportion (about $40\%$) of the random walks still lead to $e_{\rm free} \lesssim e_{\rm forced}$. This is our key result, suggesting that if the perturbing planet attains its eccentricity through random walks associated with planet-planet scatterings, the planetesimals perturbed by the planet have a significant probability of achieving small free eccentricities (compared to $e_{\rm forced}$) and the eccentric belt can therefore maintain a small width.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{FreeEccentricity_RandomGrowth_Levi_hist_cumulative.pdf}
\caption{Cumulative distribution of the final free eccentricity of the debris belt after the planet-planet scattering process (Section~\ref{sec:randomgrowth_numerical}), with $\bar{N}_{\rm p} = 1000$ and different values of $N_\omega$. Between 30\% and 50\% of the walks have a free eccentricity less than the forced eccentricity (orange vertical line), so that the resulting debris belt would be narrower than in the standard picture where the planet attains its eccentricity instantaneously. }\label{fig:efree_randomgrowth}
\end{figure
\subsection{Theoretical model}
\label{sec:randomgrowth_theory}
The numerical random-walk model (Section~\ref{sec:randomgrowth_numerical}) allowed us to get some insights on how a test particle reacts to a planet with randomly evolving eccentricity. In this subsection, we solve this problem analytically by considering a continuous random-walk model, inspired by the approach of \cite{puStrongScatteringsCold2021}.
In the continuous random-walk model, the initial and final planet eccentricities are still fixed to $0$ and $e_{\rm p, f}$, respectively. We define $t_{\rm p}$ the total time duration of a walk, and $\sigma_{\rm p}^2$ its diffusion constant (note that it has the unit of frequency). The random variable $\mathcal{E}_{\rm p}(t)$ then depends on time in the following way:
\begin{align}
&\langle \mathcal{E}_{\rm p}(t) \rangle = \frac{t}{t_{\rm p}} e_{\rm p, f},\label{eq:average}\\
&\langle \mathcal{E}_{\rm p}(t) \mathcal{E}_{\rm p}^*(s)\rangle - \langle \mathcal{E}_{\rm p}(t) \rangle \langle \mathcal{E}_{\rm p}^*(s)\rangle = 2 \min(t, s) \left[1-\frac{\max(t, s)}{t_{\rm p}}\right] \sigma_{\rm p}^2. \label{eq:covariance}
\end{align}
Assuming an initially zero eccentricity for the test particle, equation~\eqref{eq:solution} then becomes
\begin{align}
\langle \mathcal{E}(t) \rangle &{}= \int_{0}^t \langle \mathcal{E}_{\rm p}(t') \rangle \left[-i\nu \exp i \omega (t-t') \right] \dd{t'}\\
&{}= i e_\mathrm{forced} \frac{\exp(i\omega t) - i\omega t -1 }{\omega t_{\rm p}},
\end{align}
where the forced eccentricity is given by $e_{\rm forced} = (\nu/\omega) e_{\rm p,f}$ (see equation~\ref{eq:eforced_random}). Moreover, at $t = t_{\rm p}$, we have
\begin{align}
\langle |\mathcal{E}(t_{\rm p})|^2 \rangle &{}= \int_{0}^{t_{\rm p}} \int_{0}^{t_{\rm p}} \langle \mathcal{E}_{\rm p}(u) \mathcal{E}^*_{\rm p}(s)\rangle \left[-i\nu\exp i \omega (s-u) \right] \dd{s} \dd{u} \\
&{}= 2 \sigma_{\rm p}^2 \left(\frac{\nu}{\omega}\right)^2 t_{\rm p} \left(1 - \frac{2 - 2\cos\omega t_{\rm p}}{\omega^2 t_{\rm p}^2}\right) + \langle |\mathcal{E}(t_{\rm p})| \rangle^2.
\end{align}
The complex free eccentricity at the end of the random walk is given by
\begin{equation}
\mathcal{E}_{\rm free} = \mathcal{E}(t_{\rm p}) - e_{\rm forced}.
\end{equation}
The mean square of the free eccentricity is thus:
\begin{align}
\langle |\mathcal{E}_{\rm free}|^2 \rangle ={}& \langle |\mathcal{E}(t_{\rm p})|^2 \rangle - 2 \mathrm{Re}\langle e_{\rm forced}\mathcal{E}(t_{\rm p}) \rangle + e_{\rm forced}^2\nonumber \\
={}& e_\mathrm{forced}^2 \frac{2 + 2y x (x^2 - 2) - 2(1 - 2 y x)\cos x}{x^2}, \label{eq:efreerms}
\end{align}
where
\begin{align}
&x \equiv \omega t_{\rm p}, \label{eq:x}\\
&y \equiv \frac{\sigma_{\rm p}^2}{e_{\rm p, f}^2 \omega}\label{eq:y}.
\end{align}
The quantity $x$ measures the degree of adiabaticity of the planet eccentricity growth compared to the precession timescale. If the growth is fast, i.e. $x \ll 1$, then
\begin{equation}
\langle |\mathcal{E}_{\rm free}|^2 \rangle \sim e_\mathrm{forced}^2.
\end{equation}
The average $|\mathcal{E}_{\rm free}|^2$ can get lower than $e_{\rm forced}^2$ for $x\lesssim 1$ for $y^{-1} > 2x$. For $y^{-1} \gg 2x$, we recover the result of the linear planet eccentricity growth described in Section~\ref{sec:eccentricitygrowth} (equation~\ref{eq:efree_egrowth}). On the other hand, if the growth is adiabatic, i.e. $x \gg 1$, then
\begin{equation}
\langle |\mathcal{E}_{\rm free}|^2 \rangle \sim 2 e_\mathrm{forced}^2 x y \propto e_\mathrm{forced}^2~t_{\rm p}.
\end{equation}
We note that the $x$ and $y$ parametrization is analogous to the $\bar{N}_{\rm p}$ and ${N_{\rm p}}/{N_\omega}$ description introduced in Section~\ref{sec:randomgrowth_numerical}, with the correspondence
\begin{align}
&x = 2\upi \frac{N_{\rm p}}{N_\omega},\\
&y = \frac{1}{4\upi} \frac{N_\omega}{\bar{N}_{\rm p}}.
\end{align}
Note that $2xy = 1$ is equivalent to $N_{\rm p} = \bar{N}_{\rm p}$. We can thus directly compare the numerical model of Section~\ref{sec:randomgrowth_numerical} and our analytical result (equation~\ref{eq:efreerms}) in Figure~\ref{fig:efree_randomgrowthvsstepstheory}. Although we manage to capture the average free eccentricity of the test particle in the continuous random-walk model, equation~\eqref{eq:efreerms} does not give any information on the spread of free eccentricities. As shown in Section~\ref{sec:randomgrowth_numerical}, a significant proportion of the free eccentricities can be lower than the forced eccentricity even as their average is high.
\section{Summary and Discussion}
\label{sec:conclusion}
\subsection{Summary}
Planetesimals can be shaped into an eccentric debris belt by a nearby eccentric planet companion. The eccentricity of each planetesimal oscillates around an equilibrium value, the so-called forced eccentricity $e_{\rm forced}$. The period of this variation is the precession period (equation~\ref{eq:tw}), and its amplitude is the free eccentricity $e_{\rm free}$. The minimum width of the belt is $\Delta r = 2 \bar{a} e_{\rm free}$, where $\bar{a}$ is the mean semi-major axis of the belt. In the standard picture, where the planet is born eccentric, the free eccentricity of the planetesimals is equal to the forced eccentricity (Section~\ref{sec:classical}). However, observations of at least three narrow eccentric debris belts suggest that $e_{\rm free}$ can be much smaller than $e_{\rm forced}$ (see Section~\ref{sec:introduction}). In this paper, we have studied two physical processes that can reduce the free eccentricity of a debris belt: (i) interaction of the planet and/or planetesimals with the protoplanetary discs; (ii) stochastic growth of the planet eccentricity through planet-planet scatterings.
In Section~\ref{sec:exploration}, we lay down the main equations for the evolution of the planetesimals eccentricity driven by an evolving planet and demonstrate the mathematical validity of our hypotheses. We show that the free eccentricity can effectively be reduced if the planetesimals are subjected to an eccentricity damping force (Section~\ref{sec:eccentricitydamping}), or if the planet eccentricity growth is gradual (Section~\ref{sec:eccentricitygrowth}). In the first case, the force has to act for a duration comparable to the damping timescale. We also point out that if the damping is too strong, then the equilibrium eccentricity is shifted and the free eccentricity may still be large when the force stops acting. In the second case, we demonstrate that the free eccentricity is strongly reduced if the planet eccentricity reaches its final value slower than the precession timescale of the planetesimals (Figure~\ref{fig:efree}).
In Section~\ref{sec:ppdisc}, we explore these mechanisms at the beginning of the system's life, when planet and planetesimals are embedded in the protoplanetary disc. If the planet eccentricity grows due to interactions with the gas disc, a narrow eccentric debris belt could be produced. Alternatively, the gas environment can damp the free eccentricity of the planetesimals (Figure~\ref{fig:efree_friction}). In both cases however, if the separation of the belt from the host star is too large or is the planet is not very close to the belt, then the protoplanetary disc lifetime is not long enough (as compared to the precession period or the damping timescale) to induce a significant effect.
In Section~\ref{sec:ppscattering}, we study the effect of an instability phase between two planets leading to the ejection of one and the eccentricity excitation of the other. This excitation can be modelled as a random walk in the complex eccentricity plane. We combine numerical and analytical methods to study the properties of this statistical process, and find that it can lead to a wide range of free eccentricities in the neighbouring planetesimal belt. In around $40\%$ of the cases (Figure~\ref{fig:efree_randomgrowth}), the free eccentricity is smaller than the forced eccentricity, making this process a reliable candidate to account for the observed narrow widths of eccentric debris belts.
We summarize here our key points:
\begin{itemize}
\item The shape of debris belts carry information on the dynamical history of the planet companion shaping it, and on the initial conditions of the system;
\item Interaction with the protoplanetary disc can limit free eccentricities of the planetesimals (and thus the width of the belt), either through a friction force acting on the planetesimals, or by inducing a slow growth of the planet eccentricity. This requires the disc lifetime to be larger than the eccentricity damping timescale or the precession period, respectively;
\item Alternatively, the stochastic eccentricity growth associated with planet-planet scattering has a significant chance to reduce the width of a debris belt compared to the often-assumed instantaneous growth. The outcome of the process is probabilistic and could also lead to increased belt widths.
\end{itemize}
\subsection{Discussion}
We argue that the processes described in our paper could adequately account for the narrow widths of planetesimal belts. However, our demonstration uses idealized semi-analytical models, which naturally have some limitations. Our initial conditions suppose that one and only one planetary companion is directly responsible for the shape of the belt (this companion can be perturbed by a gaseous disc or an other planet, but it is the only body interacting with the planetesimals). Although it is a reasonable hypothesis when the system only harbours one giant planet, it may not be the case when several giant planets co-exist (as in our Solar System).
We are also aware of the limitations of modelling the friction effect in the protoplanetary disc with a simple non-linear force, and the large uncertainties in the gas disc properties (e.g. density and lifetime) limit our conclusion. Thus, the result of Section~\ref{sec:ppdisc} should be considered as a proof of concept that helps to identify the important quantities at play. The planet-planet scattering process described in Section~\ref{sec:ppscattering} also uses some simplifying assumptions. Treating the evolution of the complex eccentricity of the planet as a pure random walk is somewhat idealized, and the assumption that the interaction between the planet and planetesimals remain secular at all time is reasonable only if they are sufficiently separated.
In our study, we have neglected non-gravitational processes that could impact the spatial distribution of dust---we assume that they perfectly trace their parent planetesimals in the belt. Further observations, in different wavelengths, will allow us to test whether narrow eccentric debris belts are common or are the result of rare conditions of formation and evolution. The detection of the planet companions will help to confirm the relation between the debris shape and planetary architecture. The companion and debris disc characterization by direct imaging instruments, ALMA and JWST will be valuable in this regard.
\section*{Acknowledgements}
We thank the referee for their useful comments that have improved this paper. This work has been supported in part by the NSF grant AST-17152 and NASA grant 80NSSC19K0444. We made use of the \textsc{python} libraries \textsc{NumPy} \citep{harrisArrayProgrammingNumPy2020}, \textsc{SciPy} \citep{virtanenSciPyFundamentalAlgorithms2020}, and the figures were made with \textsc{Matplotlib} \citep{hunterMatplotlib2DGraphics2007}.
\section*{Data Availability}
The codes and generated data sets used for the study of stochastic planet-planet scattering evolution (Section~\ref{sec:ppscattering}) are available at \url{https://github.com/LaRodet/EccentricDebrisBelts.git}. All other figures can be directly reproduced from the equations in the paper.
\bibliographystyle{mnras}
|
\section{introduction}
The energy-momentum-tensor (EMT) form factors for hadrons in quantum chromodynamics (QCD) are physical quantities that can be measured through deeply-virtual Compton scattering~\cite{Ji:1996ek,Ji:1996nm,HERMES:2001bob,CLAS:2001wjj,ZEUS:2003pwh,H1:2005gdw} and similar processes discussed in Refs. \cite{Radyushkin:1996ru,Kharzeev:1998bz,Berger:2001xd,Frankfurt:2002ka,Guidal:2002kt,Hatta:2018ina,Guo:2021ibg,Qiu:2022bpq}. On the one hand, they are related to the perturbation of space-time induced by the hadrons; on the other hand, these form factors can also be used to characterize internal structure, such as mass, spin and momentum-current
distributions. The physical interpretation of the momentum-current (MC) ($T^{ij}(\vec{q})$) form factor $C(q)$ (or $D(q)$) in the static limit sometimes differs across the literature. They are often used in a way that assumes similarity to macroscopic fluid. For example, various components of the momentum current has been assigned the meaning of ``pressure'' and ``shear pressure" in Refs.~\cite{Polyakov:2002yz,Polyakov:2018zvc,Burkert:2018bqq,Shanahan:2018nnv}, and the ``mechanical stability" condition~\cite{Polyakov:2018zvc,Lorce:2018egm} further implies that the $D$-term form factor~\cite{Polyakov:1999gs} is negative at $q=0$. In our previous paper~\cite{Ji:2021mfb}, we suggested interpreting the form factors in terms of gravitational multipoles according to their
role in generating static gravity nearby. In particular, the gravitational tensor-monopole moment $T0$ of the momentum current is related to the $C(q^2=0)$ form factor~\cite{Ji:1996ek}. As a concrete example, we have calculated the tensor-monopole moment $\tau_H$ of hydrogen-like atom in quantum electrodynamics (QED) to leading order in the fine structure constant
$\alpha=e^2/4\pi$, and the sign is opposite of the ``mechanical stability" condition, showing the concept has little relevance in quantum mechanical systems.
As a reminder, we recall the definition of the gravitational tensor-monopole moment for static momentum-current distribution, $T^{ij}(\vec{x})$. According to our previous paper~\cite{Ji:2021mfb}, the {\it tensor monopole} $T0$ is defined as:
\begin{align}
T^{(0)} =\frac{1}{5}\int d^3\vec{x}T_{ij}(\vec{x})\left(x_ix_j-\frac{\delta_{ij}}{3}x^2\right) \ .
\end{align}
Using the conservation law or transverse condition $\partial_i T^{ij}=0$, one can show that it is related to {\it scalar momentum-current radius} ,
\begin{align}
T^{(0)} =-\frac{1}{6}\int d^3\vec{x} x^2 T_{ii}(\vec{x}) \ ,
\end{align}
where $T^{ii}(\vec{x})$ is proportional to the so-called pressure $p(r)$ in the other literature following a continuous medium~\cite{Polyakov:2002yz,Polyakov:2018zvc,Burkert:2018bqq,Shanahan:2018nnv}, although
the negative $p(r)$ demands an interpretation that deviates from the standard thermodynamics.
We choose to normalize the
tensor-monopole moment of a system as
\begin{equation}
\tau =-T^{(0)}/2 \ ,
\end{equation}
which relates to the ``$D$-term'' $D(0)$~\cite{Polyakov:1999gs} as $\tau=\frac{D(0)}{4M}$ where $M$ is the total mass of the system.
In this paper we will consider the ground state of hydrogen atom in the infinitely-heavy proton limit, where the position of the proton is fixed at origin, while the electron and photon part is treated in the background field formalism as in Ref.~\cite{Weinberg:1995mt}. In particular, the ground state $|0\rangle \equiv |0\rangle_H$ of the hydrogen atom will not be translation invariant, but describing a static spherical symmetric energy-momentum distribution around the origin. Therefore, the expectation value $\langle T^{ij}\rangle(\vec{x}) \equiv \langle 0|T^{ij}(\vec{x})|0\rangle$ of the MC density operator $T^{ij}(\vec{x})$ in the ground state $|0\rangle$ can be viewed as the MC density distribution of a classical system, from which the tensor-monopole moment can be defined as above. Due to the spherical symmetry of the ground state, in momentum space one has
\begin{align}
\langle T^{ij}\rangle (\vec{q})=(q^iq^j-\delta^{ij}q^2)\frac{C_H(q)}{m_e} \ ,
\end{align}
where $\vec{q}$ is the 3-momentum transfer taken place at the insertion of $T^{ij}$.
After a simple calculation, one can show
\begin{equation} \label{eq:deftau}
\tau = \frac{C_H(q=0)}{m_e} \ ,
\end{equation}
where $m_e$ is the electron mass.
In this paper, following up Ref.~\cite{Ji:2021mfb}, we calculate the $\tau_H$ for the hydrogen atom
to electromagnetic order ${\cal O}(\alpha)$ by including leading-order radiative corrections. The calculation is not entirely trivial since the $\tau_H$ is infrared (IR) sensitive and exhibit logarithmic IR divergences in case of a single electron~\cite{Berends:1975ah,Milton:1977je,Donoghue:2001qc}. For hydrogen-like atom, the IR divergence is regulated by the binding energy differences of order $\alpha^2 m_e$ between the ground state and excited states, but the ultraviolet contributions remain the same as the free QED. The natural scale separation $m_e\gg \alpha m_e \gg \alpha^2m_e$ and the fact that the physics of the different energy scales decouple through logarithms allows the simplification of the calculation by first working in the non-relativistic versions of QED (NRQED)~\cite{Caswell:1985ui,Labelle:1996en,Pineda:1997bj}, and then match to full-QED, in a way similar to the simplified calculation of the famous Lamb shift~\cite{Labelle:1996uc,Pineda:1997ie}.
The organization of the paper is as follows. In Section II, We first review the NRQED Lagrangian and introduce the effective EMT operator. To match to the EMT in QED, local counter-terms are required, which will be calculated in the following section. We provide a short review the calculation at leading order, emphasizing the Coulomb contributions. We then present all the relevant diagrams at one-loop level, and show by explicit power-counting that only the photonic diagram will contribute to $\tau_H$ at order $\frac{\alpha}{m_e}$. In section III, we perform the matching of EMT in NRQED to order $\frac{\alpha}{m_e}$. In section IV, we perform the calculation for the bound state. Combining the results of the two sections, we obtains the final result in Eq.~(\ref{eq:finalre}) and evaluate the sum numerically. Section V concludes the paper.
\section{Review of NRQED and overall strategy of the calculation}
In this section we introduce the EMT in NRQED and set up the overall strategy of calculating
the scalar form factor $C$ in hydrogen atom. We review the leading-order contribution to the tensor-monopole moment $\tau_H$ of the momentum current. We show that by combining the fermionic contribution, the Coulomb self-interactions and the Coulomb interference between the electron and the proton, all the Coulomb tails get removed and the resulting monopole moment $\tau$ is equal to the basic unit $\tau_0=\hbar^2/4m_e$ and {\it positive}. This example shows that the sign of the $D$-term has little to do with the ``mechanical stability''.
We consider the bound state in quantum electrodynamics (QED) between two types of fermions, the standard negative charged electron with mass $m_e$ and positive charged ``proton'' with mass $M$. At energy scale much smaller than the proton mass $M$, the proton can be approximated by an infinitely-heavy static source fixed at $\vec{x}_p=0$ ,which sources the background electric potential
\begin{align}
\nabla^2 V_p(r)=-e\delta^3(\vec{x}) \ ,
\end{align}
or $V_p=e/(4\pi r)$ ($r=|\vec{x}|$ and $e$ is the proton charge and positive). The system therefore reduces to a single electron moving in the presence of the background field $V_p$. In principle, it can be described by the dressed-Dirac theory discussed in Ref.~\cite{Weinberg:1995mt} where a complete set of solutions to the Dirac equation are being used to define the free theory, upon which radiative corrections can be added consistently. This dressed-Dirac theory has the same short distance behavior as the free QED and can be renormalized using the same renormalization constants $Z_1=Z_2$ and $Z_3$ as the free QED. This will be the underline ``first principle'' theory of this paper.
Equivalently, the first principle theory can also be treated using the Bethe-Salpeter equation approach, where the quantum nature of the proton can be preserved. See Fig.~\ref{fig:Bethe} for a depiction of the Bethe-Salpeter equation approach. Although equivalent in the infinite heavy proton limit, in this paper we will use the background field approach without mentioning otherwise.
For small fine structure constant $\alpha$, bound-states and low-energy excited states of the Dirac equation are essentially non-relativistic, and to leading order in $\alpha$ reduces to the
standard Schrodinger equation
\begin{align}\label{eq:schor}
\left(-\frac{1}{2m_e}\nabla^2-eV_p(r)\right)\psi(\vec{x})=E\psi(\vec{x}) \ .
\end{align}
The bound-state is characterized by two scales, the binding energy $\alpha^2m_e$ and the inverse Bohr radius $\alpha m_e$. One also needs the complete set of energy eigenfunctions of the above Schrodinger equation $\psi_{M}(\vec{x})$ with energy $E_M$ and the normalization condition
\begin{align}\label{eq:complet}
\sum_{M}\psi_{M}(\vec{x})\psi^{\dagger}_M(\vec{y})=\delta^{3}(\vec{x}-\vec{y}) \ .
\end{align}
It will always be understood that the set of wave functions $\psi_M$ contains both discrete and continuum spectrum. As we will show later, the contribution of the continuum spectrum is not negligible.
\begin{figure}[t]
{%
\includegraphics[height=2cm]{bethe.pdf}
}
\caption{The Bethe-Salpeter equation for the wave function $\Phi$ denoted by the oval blob. Double line represents propagator of proton field and single line represents the electron propagator. The dashed line represents the exchange of a Coulomb photon. }
\label{fig:Bethe}
\end{figure}
\subsection{Basics of NRQED}
The approach which starts from the Bethe-Salpeter equation or the fully-dressed Dirac theory in the background field as explained in \cite{Weinberg:1995mt}, is simple to understand but hard to use. This is mainly due to the complicate form of the Dirac-Coulomb propagator. For bound states with large atomic number $Z$ where relativistic effect is large, one must calculate radiative corrections numerically with Dirac-Coulomb propagator. For $Z=1$, however, the non-relativistic nature of the bound state and the emergence of the scale separation $\alpha^2m_e \ll \alpha m_e \ll m_e$ allows dramatic simplification after performing ``twist expansion'' or non-relativistic expansion in the soft scales.
The modern way to organize this expansion is through the effective field theory, more precisely, the non-relativistic reduction of QED (NRQED)~\cite{Caswell:1985ui,Labelle:1996en,Pineda:1997bj}. In NRQED, all the effective fields $\Psi$ and $A^i$ contains momentum scale comparable of smaller than $\alpha m_e$, while ultra-violet contributions are integrated out into local operators at the Lagrangian level order by order in $\frac{1}{m_e}$ , with $\alpha$-dependent {\it matching coefficients } in order to match the radiative corrections to the full-theory when expanded to the same order in $\frac{1}{m_e}$. Although not proven, it is widely believed such matching can be performed consistently to all orders in $\frac{1}{m_e}$ and $\alpha$.
For our purpose, namely, calculating the tensor-monopole moment to order $\alpha$, one only needs the Lagrangian of NRQED to order $\frac{1}{m_e}$:
\begin{align}\label{eq:LNRQED}
{\cal L}_{\rm NRQED}=&\Psi^{\dagger}(iD_0+\frac{D_iD_i}{2m_e})\Psi-c_F\frac{e}{2m_e}\Psi^{\dagger}\vec{\sigma}\cdot \vec{B}\Psi \nonumber \\
&-\frac{1}{4}F_{\mu\nu}F^{\mu\nu}+{\cal O}(\frac{1}{m_e^2}) \ ,
\end{align}
where $D^0=\partial^0-ieA^0-ieV_p$ contains the background field $V_p$ and $D^i=\partial^i-ieA^i$. The $F^{\mu\nu}=\partial^{\mu}A^{\nu}-\partial^{\nu}A^{\mu}$ is the standard field strength, $B^{i}=\frac{1}{2}\epsilon^{ijk}F^{jk}$ is the magnetic field and $c_F=1+\frac{\alpha}{2\pi}$ is the matching constants that is required to match the spin part of the vertex function in the effective theory to the full QED (to this order it is just the famous anomalous magnetic moment). Furthermore, since we are only interested in the spin-independent part of the EMT form factor to order $\frac{\alpha}{m_e}$, the spin contribution can be neglected and one needs only
\begin{align}\label{eq:NRLag}
{\cal L}_{\rm NRQED}=\Psi^{\dagger}(iD_0+\frac{D_iD_i}{2m_e})\Psi-\frac{1}{4}F_{\mu\nu}F^{\mu\nu} \ ,
\end{align}
which will be used throughout the paper.
With the complete set of energy wave functions given in Eq.~(\ref{eq:complet}) serving as the fundamental basis for the free-electron field, the above Lagrangian can be quantized in the Coulomb gauge as usual, where $A^0$
\begin{align}
A^0\equiv V_e=\frac{1}{\nabla^2}e\Psi^{\dagger}\Psi \ ,
\end{align}
are being solved in terms of the electron field explicitly, and $A^i$ contains only the transverse part, $\nabla \cdot \vec{A}=0$. In this gauge, the electric field can be separated into longitudinal and transverse (radiative) parts as
\begin{align}
\vec{E}=\vec{E}_{\parallel}+\vec{E}_{\perp} \ , \\
\vec{E}_{\parallel}=-\nabla V_e \ , \\
\vec{E}_{\perp}=-\partial^0 \vec{A} \ ,
\end{align}
and for the magnetic field there is only $\vec{B}_{\perp}$. Although not appearing in the lagrangian, the electric field of the proton reads
\begin{align}
\vec{E}_p=-\nabla V_p=\frac{e\vec{x}}{4\pi |\vec{x}|^3} \ ,
\end{align}
which will enter in the total EMT.
For the above fields in NRQED, their $\alpha$-counting rules are as follows.
For the electron field, the momentum scale is always $\alpha m$, which implies
\begin{align}\label{eq:powerphi}
\Psi^{\dagger}\Psi \sim \alpha^{3} \ .
\end{align}
For the bound states, the above Lagrangian contains contributions from both the soft photon with $|\vec{k}|={\cal O}(\alpha m_e)$ and the ultra-soft photon with $|\vec{k}|={\cal O}(\alpha^2 m_e)$. For soft radiative photon with $|\vec{k}|={\cal O}(\alpha m_e)$, one has
\begin{align}\label{eq:powerrs}
|e\vec{E}_{\perp}|_{s}, |e\vec{B}_{\perp}|_{s} \sim \alpha^4 \ ,
\end{align}
while for ultra-soft radiative photon one has
\begin{align}\label{eq:powerrus}
|e\vec{E}_{\perp}|_{us}, |e\vec{B}_{\perp}|_{us} \sim \alpha^{6} \ .
\end{align}
For Coulomb photon which is soft, one has
\begin{align}\label{eq:powercs}
|e\vec{E}_{\parallel}|_s \sim \alpha^3 \ .
\end{align}
When coupled to ultra-soft radiative photon, the Coulomb photon can also become ultra-soft. The power-counting in this case is tricky. Indeed, in coordinate state one has
\begin{align}
\langle N|\vec{E}_{\parallel}(\vec{x})|M\rangle=\int \frac{d^3\vec{k}}{(2\pi)^3 }\frac{ie\vec{k}}{|\vec{k}|^2}e^{i\vec{k}\cdot \vec{x}}(\psi^{\dagger}_M \psi_N)(\vec{k}) \ ,
\end{align}
therefore, for $|\vec{k}|={\cal O}(\alpha^2 m_e)$ the naive power-counting reads $|e\vec{E}_{\parallel}| \sim \alpha^{5}$. However, when $M \ne N$, in the matrix element one must Taylor expanding to next-leading order,
\begin{align}\label{eq:Ecspecial}
(\psi^{\dagger}_M \psi_N)(\vec{k})=-i\vec{k}\cdot \langle M|\vec{x} |N\rangle \sim \alpha \ ,
\end{align}
leading to one more power of $\alpha$, therefore in this situation the power-counting rule for ultra-soft Coulomb photon will be
\begin{align}\label{eq:powercus}
|e\vec{E}_{\parallel}|_{us} \sim \alpha^{6} \ ,
\end{align}
the same as ultra-soft radiative photon.
To regulate the UV divergences, one must specify the UV regulator, which we chose to be the standard dimensional regulator with $D=3-2\epsilon$. It has been applied in NRQED to calculate the famous Lamb~\cite{Pineda:1997ie}, ${\cal O}(\alpha^5m_e)$~\cite{Pineda:1998kn} and ${\cal O}(\alpha^6m_e)$~\cite{Czarnecki:1999mw} corrections to positronium spectrum. This means, in the intermediate steps of calculation, one must use the wave functions and matrix elements in $D=3-2\epsilon$ dimensions. For example, the Schrodinger equation becomes
\begin{align}\label{eq:schorD}
\left(-\frac{1}{2m_e}\nabla^2-eV_{p,D}(r)\right)\psi(\vec{x})=E\psi(\vec{x}) \ ,
\end{align}
with $D$-dependent potential
\begin{align}
V_{p,D}(r)=\mu^{2\epsilon}\int \frac{d^D{\vec{p}}}{(2\pi)^D|\vec{p}|^2}e^{i\vec{p}\cdot\vec{x}} \ ,
\end{align}
and the normalization condition
\begin{align}
\int d^D\vec{x} \psi^{\dagger}_M(\vec{x})\psi_N(\vec{x})=\delta_{MN} \ .
\end{align}
In terms of them, one can form the matrix elements
\begin{align}\label{eq:matrix}
\rho_{NM}(\vec{k})=\int d^D\vec{x} \psi^{\dagger}_N(\vec{x})\psi_{M}(\vec{x})e^{-i\vec{k}\cdot\vec{x}} \ , \\
\vec{v}_{NM}(\vec{k})=\int d^D\vec{x} \psi^{\dagger}_N(\vec{x})\frac{-i\vec{\nabla}}{m_e}\psi_{M}(\vec{x})e^{-i\vec{k}\cdot\vec{x}} \ ,
\end{align}
which are $D$-dependent generically, and will appear in the bound-state calculation. However, we will show that at the final stage of the calculation, by using the sum rule
\begin{align}\label{eq:sumrule}
\sum_{M}\frac{2\vec{v}_{MN}\cdot \vec{v}_{NM}}{D(E_M-E_N)} =\frac{1}{m_e} \ ,
\end{align}
coefficients of the $\frac{1}{\epsilon}$ poles are {\it $D$-independent} universal constants, while all other finite terms can be safely set to the value in $D=3$ without causing trouble.
After introducing the Lagrangian and the UV regulator, we now list the Feynman-rules of the covariant perturbation theory in Fig.~\ref{fig:rules}. The radiative photon polarization
sum is
\begin{equation}
P^{ij}(\vec{k}) = \delta^{ij} - \frac{k^ik^j}{|\vec{k}|^2} \ .
\end{equation}
All the interaction vertices are represented in plan-wave basis, in the energy eigen-basis they are replaced by the matrix-elements as in Eq.~(\ref{eq:matrix}). Fig.~\ref{fig:rulecpvertex} contains the standard instantaneous Coulomb vertices, the triple electron-photon vertices and the seagull vertex due to the $-\frac{e^2}{2m_e^2}A^iA^i\Psi^{\dagger}\Psi$ term in the Lagrangian. For non-relativistic system, sometimes it is convenient to use the old-fashioned perturbation theory as well, which can be obtained by integrating out the $k^0$ first and consists of matrix elements between free states followed by energy denominators. The matrix-elements can still be obtained from Fig.\ref{fig:rulecpvertex} by removing factor of $-i$'s for the non-instantaneous triple photon-electron vertex and the seagull vertex.
\begin{figure}[htb]
\begin{subfigure}{.2\textwidth}
\centering
\includegraphics[width=\linewidth]{rulecova.pdf}
\caption{}
\label{fig:rulecovpro}
\end{subfigure}
\begin{subfigure}{.13\textwidth}
\centering
\includegraphics[width=\linewidth]{vertex.pdf}
\caption{}
\label{fig:rulecpvertex}
\end{subfigure}
\caption{The Feynman rules of the NRQED Lagrangian Eq.~(\ref{eq:NRLag}). In Fig.~\ref{fig:rulecovpro}, we show the propagator for electron, radiative photon and Coulomb photon in covariant perturbation theory. In Fig.~\ref{fig:rulecpvertex}, we show all the interaction vertices. To obtain the rule in old-fashioned perturbation theory, simply remove an $-i$ from the non-instantaneous triple electron-photon vertex and the seagull vertex. }
\label{fig:rules}
\end{figure}
It is possible to separate the ultra-soft contribution to define a new effective theory, the potential-NRQED(pNRQED)~\cite{Pineda:1997bj,Pineda:1997ie}, where further simplifications for ultra-soft photons are performed, namely, expanding all the $e^{-i\vec{k}\cdot \vec{x}}$ in the form-factors in Eq.~(\ref{eq:matrix}). The contribution from the soft photons, however, can be calculated largely in the free-NRQED by expanding the electron propagators in the background field. Provided one uses the correct rules in the corresponding region, as we do here,
it is not necessary to introduce the pNRQED Lagrangian.
\subsection{Momentum current density of NRQED}
After introducing the NRQED in the Coulomb gauge, one now discuss its momentum current density. Clearly, the EMT of NRQED must include the non-relativistic reduction of the QED EMT, which we call the ``tree-level'' EMT. In addition, the naive EMT in theories with non-trivial UV structure can receive quantum corrections, which must be included as counter terms in the effective operator. For example, in the scalar $\phi^4$ theory in 4D, in order for off-shell matrix elements of EMT to be finite, one must add local counter-term of the form $ d(\partial^{\mu}\partial^{\nu}-g^{\mu\nu}\partial^2) \phi^2$ to the naive EMT, where $d$ is divergent order-by-order in perturbation theory. In our case,
we are only interested in corrections at one-loop that are spin-independent.
As we will show in Sec.III that to order $\frac{\alpha}{m_e}$, the EMT of NRQED has the form
\begin{align}\label{eq:Tmatch}
T^{ij}_{\rm NRQED}=T^{ij}_{\rm tree}+d_0(\partial^i\partial^j-\delta^{ij}\partial^2)\Psi^{\dagger}\Psi + {\cal O}(\alpha^2)\ ,
\end{align}
where $d_0$ depends on the UV regulator for the NRQED, but is IR insensitive.
We now consider all contributions to $T^{ij}_{\rm tree}$ and their Feynman rules. By the standard NR reduction on the energy-momentum tensor of QED, one can obtain the ``classical'' EMT for NRQED,
\begin{align}\label{eq:Tclass}
T^{ij}_{\rm tree}=T^{ij}_{e}+T_{\gamma}^{ij}+T_{\gamma p}^{ij}+T_{p}^{ij}\ .
\end{align}
where
\begin{enumerate}
\item $T^{ij}_{e}$ is the electron part of the EMT. For the non-relativistic particle, by performing the non-relativistic reduction on the full QED, one obtains:
\begin{align}
&T_{e}^{ij}=-\frac{1}{4m_e}\Psi^{\dagger}D^{i}D^j\Psi-\frac{1}{4m_e}(D^iD^j\Psi)^{\dagger}\Psi\nonumber \\
&+\frac{1}{4m_e}(D^i\Psi)^{\dagger}(D^j\Psi)+\frac{1}{4m_e}(D^j\Psi)^{\dagger}(D^i\Psi) \ .
\end{align}
Expanding all the covariant derivatives, it can be further decomposed as
\begin{align}\label{eq:Tphi}
T^{ij}_{e}=T^{ij}_{e0}+T^{ij}_{e1}+T^{ij}_{e2} \ ,
\end{align}
where
\begin{align}
&T^{ij}_{e0}=-\frac{1}{4m_e}\Psi^{\dagger}\partial^{i}\partial^j\Psi-\frac{1}{4m_e}\partial^{i}\partial^j\Psi^{\dagger}\Psi\nonumber \\& +\frac{1}{4m_e}\partial^{i}\Psi^{\dagger}\partial^{j}\Psi+\frac{1}{4m_e}\partial^j\Psi^{\dagger}\partial^i\Psi \nonumber \\
&T^{ij}_{e1}=\frac{ie}{2m_e}A^i(\Psi^{\dagger}\partial^j\Psi-\partial^j\Psi^{\dagger}\Psi)-(i\leftrightarrow j)\nonumber \\
&T^{ij}_{e2}=\frac{e^2}{m_e}A^iA^j\Psi^{\dagger}\Psi \ .
\end{align}
Notice the appearance of $T^{ij}_{e2}$, which is absent in the full QED.
\item $T^{ij}_{\gamma}$ is the standard contribution from the photon-field $A^{\mu}=(A^0,\vec{A})$,
\begin{align}\label{eq:Tphoton}
T_{\rm \gamma}^{\mu\nu}=-F^{\mu\rho}F^{\nu}_{\rho}+\frac{g^{\mu\nu}}{4}F^2 \ .
\end{align}
It is convenient to decompose the $T^{ij}_{\gamma}$ into pure-Coulomb, mixed and pure radiative parts
\begin{align}
T^{ij}_{\gamma}=T^{ij}_{\gamma\parallel}+T^{ij}_{\gamma\parallel \perp}+T^{ij}_{\gamma\perp} \ ,
\end{align}
with
\begin{align}
&T^{ij}_{\gamma\parallel}=-\partial^iV_e\partial^jV_e+\frac{1}{2}\delta^{ij}\partial^kV_e\partial^k V_e \ , \label{eq:Tcou}\\
&T^{ij}_{\gamma\parallel\perp}=\partial^{(i}V_e\partial^0A^{j)}-\delta^{ij}\partial^kV_e\partial^0A^k \ .\label{eq:Tmixed}
\end{align}
where $A^{(ij)}=\frac{1}{2}\left(A^{ij}+A^{ji}\right)$ denotes the standard symmetrization. For the pure radiative part $T^{ij}_{\gamma \perp}$, one can further decompose it into electric and magnetic part
\begin{align}\label{eq:Tradia}
T^{ij}_{\gamma\perp}=T^{ij}_{\gamma E}+T^{ij}_{\gamma B} \ ,
\end{align}
with
\begin{align}
T^{ij}_{\gamma E}=&-\partial^0A^i\partial^0A^j+\frac{\delta^{ij}}{2}\partial^0A^k\partial^0 A^k \ , \label{eq:TE} \\
T^{ij}_{\gamma B}=&(\partial^iA^k-\partial^kA^i)(\partial^jA^k-\partial^kA^j) \nonumber \\
&-\frac{\delta^{ij}}{2}(\partial^kA^l\partial^kA^l-\partial^kA^l\partial^lA^k) \ . \label{eq:TB}
\end{align}
This decomposition will be used later.
\item The $T^{ij}_{\gamma p}$ is the mixed contribution between the photon field and the proton's electric field. More precisely, it can be further decomposed as
\begin{align}
T^{ij}_{\gamma p}=T^{ij}_{\perp p}+T^{ij}_{\parallel p} \ ,
\end{align}
where $T^{ij}_{\parallel p}$ is the mixing between the electron's Coulomb field and the proton's electric field
\begin{align}
T_{\parallel p}^{ij}=\delta^{ij}\nabla V_p \cdot \nabla V_e- \partial^i V_e \partial^j V_p-\partial^iV_p\partial^jV_e \ ,
\end{align}
while $T^{ij}_{\perp p}$ is the mixing between radiative field and the proton's electric field
\begin{align}\label{eq:mixinter}
T^{ij}_{\perp p}=\partial^iV_p\partial^0A^j+\partial^jV_p\partial^0A^i-\delta^{ij}\partial^kV_p\partial^0A^k \ .
\end{align}
\item Finally, $T^{ij}_p$ is the energy momentum tensor of the proton, which can be calculated using its classical electric field $E_p^i=\frac{er^i}{4\pi r^3}$ as
\begin{align}
T^{ij}_{p}(\vec{r})=(\delta^{ij}\partial^2-\partial^i\partial^j)\frac{\alpha}{32\pi r^2}\ ,
\end{align}
which translate to momentum space as
\begin{align}\label{eq:Tp}
T^{ij}_{p}(\vec{q})=(q^iq^j-\delta^{ij}q^2)\frac{\alpha\pi}{16|q|} \ .
\end{align}
It is transverse by itself.
\end{enumerate}
One can show that sandwiched between static states with equal energies, the above $T_{\rm tree }^{ij}$ is conserved. In fact, in Appendix A we will show that $T^{ij}_e+T^{ij}_{\gamma}+T^{ij}_{\parallel p}$ and $T^{ij}_{\perp p}$ are conserved separately for the spherical symmetric ground state where $V_e(p)=V_e(|p|)$ .
For convenience of the reader, we collect all the vertices for the above momentum current density in Fig.~\ref{fig:fermionrules} and Fig.~\ref{fig:photonrules}. More precisely, for the fermion part by simply taking the matrix element of the various terms in Eq.~(\ref{eq:Tphi}) in plan wave states one has
\begin{align}
&\text{\rm Fig.~\ref{fig:Ferminkine}}=k^ik^j \ , \\
&\text{\rm Fig.~\ref{fig:Fermionphoton}}=-\frac{e}{2m_e}\delta^{il}(2p+k)^j-\frac{e}{2m_e}\delta^{jl}(2p+k)^i \ , \\
&\text{\rm Fig.~\ref{fig:Fermiontadpole}}=\frac{e^2}{m_e}(\delta^{il}\delta^{jm}+\delta^{jl}\delta^{im}) \ .
\end{align}
Similarly, for the photon part one has
\begin{align}
&\text{\rm Fig.~\ref{fig:pureCoulomb}}=-2(k-q)^{(i}(k+q)^{j)}+\delta^{ij}(\vec{k}^2-\vec{q}^2) \ , \\
&\text{\rm Fig.~\ref{fig:mixedvertex}}=2k^0(k-q)^{(i}\delta^{j)l}-\delta^{ij}k^0(k-q)^l \ ,
\end{align}
Using Eq.~(\ref{eq:TE}) and Eq.~(\ref{eq:TB}), Fig.~\ref{fig:pureradiative} can simply be obtained by taking matrix elements of $T^{ij}_{E}+T^{ij}_{B}$ in free photon states. Since the resulting expression is quite long and since in the calculation one only needs $T^{ii}_{\gamma \perp}$ and $q^iT^{ij}_{\gamma \perp}q^j$, which simplifies considerably, we will not provide the explicit formulas here.
\begin{figure}[htb]
\begin{subfigure}{.17\textwidth}
\includegraphics[width=\linewidth]{fermionoperator.pdf}
\caption{}
\label{fig:Ferminkine}
\end{subfigure}
\begin{subfigure}{.18\textwidth}
\includegraphics[width=\linewidth]{fermionphoton.pdf}
\caption{}
\label{fig:Fermionphoton}
\end{subfigure}
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\linewidth]{fermiontadpole.pdf}
\caption{}
\label{fig:Fermiontadpole}
\end{subfigure}
\caption{Vertices corresponding to different terms for $T^{ij}_{e}$ in Eq.~(\ref{eq:Tphi}). Fig.~\ref{fig:Ferminkine} corresponds to $T^{ij}_{e0}$, Fig.~\ref{fig:Fermionphoton} corresponds to$T^{ij}_{e1}$ and Fig.~\ref{fig:Fermiontadpole} corresponds to $T^{ij}_{e2}$.}
\label{fig:fermionrules}
\end{figure}
\begin{figure}[htb]
\begin{subfigure}{.2\textwidth}
\centering
\includegraphics[width=\linewidth]{purecoulomb.pdf}
\caption{}
\label{fig:pureCoulomb}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=\linewidth]{mixedveretex.pdf}
\caption{}
\label{fig:mixedvertex}
\end{subfigure}
\begin{subfigure}{.2\textwidth}
\centering
\includegraphics[width=\linewidth]{pureradiative.pdf}
\caption{}
\label{fig:pureradiative}
\end{subfigure}
\caption{Vertices corresponding to different terms for $T^{ij}_{\gamma}$ in Eq.~(\ref{eq:Tphoton}). Fig.~\ref{fig:pureCoulomb} corresponds to $T^{ij}_{\gamma \parallel}$, Fig.~\ref{fig:mixedvertex} corresponds to $T^{ij}_{\gamma \parallel \perp}$ and Fig.~\ref{fig:pureradiative} corresponds to $T^{ij}_{\gamma \perp}$. }
\label{fig:photonrules}
\end{figure}
\subsection{Leading-order momentum-current form factor}
Given the above momentum current operator, one can calculate its form factor in the ground state of hydrogen atom. The contributions to the leading order are shown in Fig.~\ref{fig:inter}. More explicitly, in Fig.~\ref{fig:electron1} one has the electron kinetic contribution, in Fig.~\ref{fig:inter1} one has the interference contribution between Coulomb photons emitted from the electron and the proton. These two terms are conserved when added together. Finally, in Fig.~\ref{fig:Coulomself1} one has the Coulomb-photon self-energy contributions from the electron and proton, respectively. Since there is no UV divergence at this order, all calculations can be performed in $D=3$ directly.
\begin{figure}[htb]
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=\linewidth]{electron1.pdf}
\caption{}
\label{fig:electron1}
\end{subfigure}
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=\linewidth]{inteference.pdf}
\caption{}
\label{fig:inter1}
\end{subfigure}
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=\linewidth]{Coulumbself.pdf}
\caption{}
\label{fig:Coulomself1}
\end{subfigure}
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=\linewidth]{Coulumpro.pdf}
\caption{}
\label{fig:Coulombpro}
\end{subfigure}
\caption{The order-${\cal O}(1)$ contributions to $T^{ij}$ for a bound state. Dashed lines represent Coulomb photons and crossed circles denote the operator insertions. Notice the infrared divergences for $C(q)$ at $q=0$ are cancelled between the interference and single electron and proton contributions.}
\label{fig:inter}
\end{figure}
More explicitly, the fermionic contribution in Fig.~\ref{fig:electron1} can be shown as
\begin{align}\label{eq:kine}
\langle T^{ij}_{e0}(\vec{x}) \rangle =-\frac{1}{4m_e}\left(\psi^{\dagger}_0\partial^{i}\partial^{j}\psi_0-\partial^{i}\psi^{\dagger}_0\partial^{j}\psi_0+{\rm c.c}\right) \ .
\end{align}
where $\langle T^{ij}\rangle \equiv \langle 0|T^{ij}|0 \rangle$ denotes the matrix element in the ground state $|0\rangle$ of the hydrogen atom with wave function $\psi_0$.
Furthermore, interference contribution, Fig.~\ref{fig:inter1} be calculated as
\begin{align}
\langle T^{ij}_{||p}(\vec{x})\rangle=\delta^{ij}\nabla V_p \cdot \nabla V_e- \partial^i V_e \partial^j V_p-\partial^iV_p\partial^jV_e \ ,
\end{align}
where the static potential $V_e$ induced by the electron reads
\begin{align}
\nabla^2 V_e(\vec{x})=e|\psi_0(\vec{x})|^2 \ ,
\end{align}
One can show that the quantum mechanical contribution
$\langle T^{ij}\rangle_{\rm e0}+\langle T^{ij}\rangle_{\rm ||p}$ is conserved by itself.
After Fourier-transformation to the momentum space,
\begin{align}\label{eq:CQM}
&\langle T^{ij}\rangle_{ e0+\parallel p}(\vec{q})=(q^iq^j-\delta^{ij}q^2)\frac{C_{ e0+\parallel p}(q)}{m_e} \ ,\\
&\frac{C_{e0+\parallel p}(q)}{m_e}=\frac{1}{2m_e(\frac{q^2}{\alpha^2m_e^2}+4)}-\frac{\alpha}{4|q|}\left(\frac{\pi}{2}-{\rm Arctan} \frac{q}{2\alpha m_e}\right) \ .
\end{align}
The resulting $C_{\rm QM}(q)$ contains a a Coulomb tail $\frac{\pi \alpha}{8|q|}$.
In order to cancel it, one must add the electron and proton Coulomb self-energy contributions, which we will show to be conserved by itself.
For the bound state, the Coulomb contribution in Fig.~\ref{fig:Coulomself1} can be calculated using the Feynman rules provided above . It turns out that the self-energy bubble for the Coulomb insertion is independent of the incoming/out going electron/proton momentum, and contributes exactly as the free-electron, which will be further dressed by the momentum dependency of the bound state wave function for the external electron and proton. The form factor from the electron therefore reads
\begin{align}
\frac{ C_{\gamma \parallel } (q)}{m_e}= \frac{\alpha \pi}{16|q|}\times \frac{16\alpha^4}{(\frac{q^2}{m_e^2}+4\alpha^2)^2} \ ,
\end{align}
where the first factor $\frac{\alpha \pi}{16|q|}$ is just the free-electron contribution, and the second factor is nothing but the dressing in the bound-state wave function
\begin{align}
\frac{16\alpha^4}{(\frac{q^2}{m_e^2}+4\alpha^2)^2}=\int \frac{d^3\vec{p}}{(2\pi)^3} \psi^{\dagger}_0(\vec{p}-\frac{\vec{q}}{2})\psi_0(\vec{p}+\frac{\vec{q}}{2}) \ ,
\end{align}
where $p\pm \frac{q}{2}$ are the external momentum of the self-energy bubble. Finally, the contribution of the proton can be obtained from Eq.~(\ref{eq:Tp}) as
\begin{align}
\frac{C_{\rm p} (p)}{m_e} = \frac{\alpha \pi}{16|q|} \ .
\end{align}
Equivalently, this can also be calculated from the $T^{ij}_{\rm Coulomb}$ for an infinitely heavy source field, which explains the representation in Fig.~\ref{fig:Coulombpro}.
In conclusion, in the region $|q|\le{\cal O} (\alpha m_e)$, the $C$ form factor of the hydrogen atom reads
\begin{align}\label{eq:hydrogenfinal}
C_H(q)=C_{e0}(q)+C_{\parallel p}(q)+C_e(q)+C_p(q) \ .
\end{align}
From these, the tensor monopole moment for the hydrogen atom is
\begin{align}
\tau_{\rm H}=\frac{C_{\rm H}(0)}{m_e}\equiv g_H\tau_0=\tau_0[1 +{\cal O}(\alpha \ln \alpha)]\ .
\end{align}
where $\tau_0 = \hbar^2/4m_e$. Therefore, $g_H=1$, except for a small correction of order $\alpha$, a result with opposite sign from a point-like boson.
\subsection{Diagrams and power-counting in ${\cal O}(\alpha)$}
To next-to-leading order in radiative corrections, relevant diagrams for the form factor
are shown in Fig.~\ref{fig:NLOf} and Fig.~\ref{fig:NLObound}. There are ten fermionic diagrams and four photonic one. We show that only the photonic contributions in Fig.~\ref{fig:NLObound} are relevant
in the soft and ultra-soft regions and contributes to $\tau_H$ to order $\alpha$.
For this purpose, it is convenient to follow the previous subsection to
introduce the leading-order momentum current independent of the external states,
\begin{align}\label{eq:Tleading}
&T_0^{ij}(\vec{q})=T^{ij}_{e0}(\vec{q})+T_{\parallel p}^{ij}(\vec{q})\nonumber \\
&+(q^iq^j-\delta^{ij}q^2)\frac{\alpha\pi}{16|q|}\Psi^{\dagger}\Psi(q)+(q^iq^j-\delta^{ij}q^2)\frac{\alpha\pi}{16|q|} \ .
\end{align}
It is easy to show by using the standard power-counting rules Eq.~(\ref{eq:powerphi}) and Eq.~(\ref{eq:powercs}) that for non-relativistic wave functions for electrons, that one has the basic power-counting rule
\begin{align}
\langle M'|T^{ij}_{\rm 0}(\vec{q})|M\rangle={\cal O}(\alpha^2)+ {\cal O}(\alpha)q+{\cal O}(1)q^2 \ ,
\end{align}
where $q$, $q^2$ denotes the expansion of $\langle M'|T_{\rm 0}^{ij}(\vec{q})|M\rangle$ to linear and quadratic orders in $q$, respectively.
\begin{figure}[htb]
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=\linewidth]{fermionver.pdf}
\caption{}
\label{fig:fermionver}
\end{subfigure}
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=\linewidth]{fermionwf.pdf}
\caption{}
\label{fig:fermionver1}
\end{subfigure}
\begin{subfigure}{.2\textwidth}
\centering
\includegraphics[width=\linewidth]{fermionop.pdf}
\caption{}
\label{fig:fermionop}
\end{subfigure}
\begin{subfigure}{.2\textwidth}
\centering
\includegraphics[width=\linewidth]{fermiontadlamb.pdf}
\caption{}
\label{fig:fermiontadlamb}
\end{subfigure}
\caption{The one-loop fermionic contributions to $T^{ij}$ for a bound state, shown in old-fashioned perturbation theory. Dashed lines represent Coulomb photons and crossed circles denote the operator insertions. Notice that Fig.~\ref{fig:fermionver1} contains two types of contributions. All of them are sub-leading in $\alpha$ and will not be calculated. }
\label{fig:NLOf}
\end{figure}
\begin{figure}[htb]
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=\linewidth]{lambmix1.pdf}
\caption{}
\label{fig:lambmix}
\end{subfigure}
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=\linewidth]{lambmixinter.pdf}
\caption{}
\label{fig:lambmixinter}
\end{subfigure}
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=\linewidth]{lambradia.pdf}
\caption{}
\label{fig:lambradia}
\end{subfigure}
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=\linewidth]{lambtad.pdf}
\caption{}
\label{fig:lambtad}
\end{subfigure}
\caption{The order-${\cal O}(\alpha)$ contributions to $T^{ij}_{\gamma}+T^{ij}_{\gamma p}$ for a bound state. Dashed lines represent Coulomb photons and crossed circles denote the operator insertions.}
\label{fig:NLObound}
\end{figure}
Given the above, it is not hard to show that all the fermionic diagrams are sub-leading in $\alpha$. For example, for the four diagrams shown in Fig.\ref{fig:fermionver}, the Feynman rule for old-fashioned perturbation theory shows that they combine to produce
\begin{align}
&\langle T^{ij}_{0}\rangle_{\rm 6a} (\vec{q})=e^2\mu^{2\epsilon}\sum_{M,M'}\int \frac{d^D\vec{k}}{(2\pi)^D}\nonumber \\
&\frac{(1-\frac{1}{D})\vec{v}_{0M'}(\vec{k})\cdot \vec{v}_{M0}(\vec{k})}{(E_0-|\vec{k}|-E_M)(E_0-|\vec{k}|-E_{M'})2|\vec{k}|}\langle M'|T_{0}^{ij}(\vec{q})|M\rangle \ ,
\end{align}
where the $T_{\rm 0}^{ij}(\vec{q})$ is defined in Eq.~(\ref{eq:Tleading}), including the proton part when $M=M'$. It is easy to see that for $D=3$, when $|\vec{k}|={\cal O}(\alpha m_e)$ or $|\vec{k}|={\cal O}(\alpha^2 m_e)$, the two energy denominators and one phase-space measure $2|\vec{k}|$ for the photon contributes to $(\alpha m_e)^{-3}$ or $(\alpha m_e)^{-6}$, which is always canceled by the integration measure $\int d^3 \vec{k}={\cal }(\alpha m_e)^3$ or $(\alpha m_e)^6$, respectively. The two form-factors for the velocity operators will contributes to $\alpha^2$, while the matrix element $\langle M'|T_{\rm 0}^{ij}(q)|M\rangle $ as shown above will contributes to ${\cal O}(\alpha^2)$ at order ${\cal O}(q^0)$ and ${\cal O}(1)$ to ${\cal O}(q^2)$. Therefore, together with the overall $e^2$, Fig.~\ref{fig:fermionver} will contributes at order ${\cal O}(\alpha^3)$ to the coefficients of $q^2$, and to ${\cal O}(\alpha^5)$ for coefficients of $q^0$, therefore not relevant for our calculation. Similar argument can be used to show that diagrams in Fig.~\ref{fig:fermionver1} and Fig.~\ref{fig:fermionop} ($T^{ij}_{e1}$) will be irrelevant to NLO as well. The last diagram Fig.~\ref{fig:fermiontadlamb}, due to the $\frac{e^2}{m_e}\Psi^{\dagger}\Psi A^iA^j$ term in $T^{ij}_{e}$, reads
\begin{align}
&\langle T^{ij}_{e2}\rangle_{6c}(q)=\int \frac{d^Dp}{(2\pi)^D} \psi^{\dagger}_0(p-\frac{q}{2})\psi_0(p+\frac{q}{2})\nonumber \\
&\times \frac{e^2}{2m_e}\mu^{2\epsilon}\int \frac{d^D\vec{k}}{(2\pi)^D|\vec{k}|}\bigg(\delta^{ij}-\frac{k^ik^j}{\vec{k}^2}\bigg) \ ,
\end{align}
which is proportional to a $q$-independent dimensionless integral and vanishes identically in dimensional regularization.
More generally, the power-counting for arbitrary diagram with interaction vertices from the Lagrangian in Eq.~(\ref{eq:NRLag}) can be performed as follows. We first consider the one-body irreducible (1PI) diagrams containing the operator insertion, but without insertion of self-energy type bubbles on external electron legs. Then it is easy to see that for arbitrary 1PI diagram one has
\begin{align}
\langle T_{0}^{ij}\rangle_{\rm s}=\bigg(q^2{\cal O}(1)+q{\cal O}(\alpha)+{\cal O}(\alpha^2)\bigg)\alpha^{\frac{3}{2}n_{V3}+2n_{V_4}} \ , \\
\langle T_{0}^{ij}\rangle_{\rm us}=\bigg(q^2{\cal O}(1)+q{\cal O}(\alpha)+{\cal O}(\alpha^2)\bigg)\alpha^{\frac{3}{2}n_{V3}+3n_{V_4}} \ ,
\end{align}
where $n_{V3}$ and $n_{V_4}$ denotes the numbers of electron-photon triple and seagull vertices, respectively, and `s', `us' denotes that all the photons in the diagram are soft or ultra-soft.
For the fermion-photon mixed operator, one has
\begin{align}
\langle T_{e1}^{ij}\rangle_{\rm s}=\bigg(q^2{\cal O}(1)+q{\cal O}(\alpha)+{\cal O}(\alpha^2)\bigg) \alpha^{\frac{3}{2}n_{V3}+2n_{V_4}+\frac{1}{2}} \ , \\
\langle T_{e1}^{ij}\rangle_{\rm us}=\bigg(q^2{\cal O}(1)+q{\cal O}(\alpha)+{\cal O}(\alpha^2)\bigg)\alpha^{\frac{3}{2}n_{V3}+3n_{V_4}+\frac{3}{2}} \ .
\end{align}
The leading contribution is shown in Fig. 6c. For the fermion-tadpole operator,
\begin{align}
\langle T_{\rm e2}^{ij}\rangle_{\rm s}=\bigg(q^2{\cal O}(1)+q{\cal O}(\alpha)+{\cal O}(\alpha^2)\bigg)\alpha^{\frac{3}{2}n_{V3}+2n_{V_4}+1} \ , \\
\langle T_{\rm e2}^{ij}\rangle_{\rm us}=\bigg(q^2{\cal O}(1)+q{\cal O}(\alpha)+{\cal O}(\alpha^2)\bigg) \alpha^{\frac{3}{2}n_{V3}+3n_{V_4}+3} \ .
\end{align}
The leading contribution is shown in Fig. 6d.
For the radiative photonic contributions which appear in Fig.\ref{fig:NLObound}, we have following counting rules,
\begin{align}
\langle T_{\gamma \perp}^{ij}\rangle_{\rm s}=\bigg(q^2{\cal O}(1)+q{\cal O}(\alpha)+{\cal O}(\alpha^2)\bigg)\alpha^{\frac{3}{2}n_{V3}+2n_{V_4}-1} \ , \\
\langle T_{\gamma \perp}^{ij}\rangle_{\rm us}=\bigg(q^2{\cal O}(1)+q{\cal O}(\alpha^2)+{\cal O}(\alpha^4)\bigg)\alpha^{\frac{3}{2}n_{V3}+3n_{V_4}-2} \ .
\end{align}
Similarly, for the mixed radiative-Coulomb operator, one has
\begin{align}
\langle T_{\gamma\parallel \perp}^{ij}\rangle_{\rm s}=\bigg(q^2{\cal O}(1)+q{\cal O}(\alpha)+{\cal O}(\alpha^2)\bigg) \alpha^{\frac{3}{2}n_{V3}+2n_{V_4}-\frac{1}{2}} \ , \\
\langle T_{\gamma\parallel \perp}^{ij}\rangle_{\rm us}=\bigg(q^2{\cal O}(1)+q{\cal O}(\alpha^2)+{\cal O}(\alpha^4)\bigg)\alpha^{\frac{3}{2}n_{V3}+3n_{V_4}-\frac{3}{2}} \ ,
\end{align}
and the same rule applies to $T^{ij}_{\perp p}$ in Eq.~(\ref{eq:mixinter}) as well.
Here we use $T^{ij}_{e0}$ and $T^{ij}_{\gamma \perp}$ as two examples to demonstrate how to derive the above power-counting rule:
\begin{itemize}
\item The operator insertion itself will contributes to $q^2{\cal O}(1)+q{\cal O}(\alpha)+{\cal O}(\alpha^2)$ in case of $T^{ij}_0$. In case of $T^{ij}_{\gamma \perp}$, one has $q^2{\cal O}(1)+q{\cal O}(\alpha)+{\cal O}(\alpha^2)$ in the soft region while $q^2{\cal O}(1)+q{\cal O}(\alpha^2)+{\cal O}(\alpha^4)$ in the ultra-soft region, which follows from Eq.~(\ref{eq:powerrs}) and Eq.~(\ref{eq:powerrus}).
\item Each of the electron-photon triple vertices contributes to one power of $\alpha$ from the velocity and half power of $\alpha$ from the interaction, leading to $\alpha^{\frac{3}{2}n_{V_3}}$.
\item Each of the seagull vertices contributes to one power of $\alpha$ from the interaction, leading to $\alpha^{n_{V_4}}$.
\item The rest of the diagram, including all the energy-denominators, phase-space measures and momentum integrals has the mass dimension $\lambda^{n_{V_4}}$ for $T^{ij}_{0}$ and $\lambda^{n_{V_4}-1}$ for $T^{ij}_{\rm \gamma \perp}$ with $\lambda=\alpha$ in soft region and $\lambda=\alpha^2$ in ultra-soft region.
\end{itemize}
Combining all the factors leads to the above results. For non-1PI diagram, it is easy to show that each self-energy-like bubble insertion on external legs will increase at least one factor of $\alpha^2$, depending on the type of insertions. Finally, diagrams with simultaneous existence of multiple scales will be more suppressed.
Given the above, it is easy to see that in order to obtain order $\alpha$ contribution at $q^2$, one needs the following:
\begin{itemize}
\item $T_{e2}^{ij}$ with $n_{V_3}=n_{V_4}=0$. This corresponds to Fig.~\ref{fig:fermiontadlamb}. However, we have shown that this diagram is $q$-independent and vanishes in DR.
\item $T^{ij}_{\gamma \parallel \perp}$ and $T^{ij}_{\rm \perp p}$ with $n_{V_3}=1,n_{V_4}=0$. They correspond to Fig.~\ref{fig:lambmix} and Fig.~\ref{fig:lambmixinter}, respectively.
\item $T^{ij}_{\gamma \perp}$ with $n_{V_3}=2, n_{V_4}=0$. This corresponds to Fig.~\ref{fig:lambradia}.
\item $T^{ij}_{\gamma \perp}$ with $n_{V_4}=1,n_{V_3}=0$. This corresponds to Fig.~\ref{fig:lambtad}.
\end{itemize}
One must notice that for $\langle T^{ij}_{\gamma \parallel \perp}\rangle$ and $\langle T^{ij}_{\perp p}\rangle$, the above power-counting using $|\vec{E}_{\parallel}|_{us}^2 \sim \alpha^9$ leads to ${\cal O}(1)$ at order $q^2$ when $n_{V_3}=1, n_{V_4}=0$ in the ultra-soft region, corresponding to Fig.~\ref{fig:lambmix} and Fig.~\ref{fig:lambmixinter}. However, we will show that in this case $\langle T^{ij}_{\perp p}\rangle=0$, and for $\langle T^{ij}_{\gamma \parallel \perp}\rangle$ one must have $M \ne N$ in Eq.~(\ref{eq:Ecspecial}). Therefore, in this case the actual power-counting should be given by Eq.~(\ref{eq:powercus}), which adds one more $\alpha$, leading to the ${\cal O}(\alpha)$ contribution as well.
In conclusion, one needs to calculate all the photonic contributions in Fig.~\ref{fig:NLObound}. More explicitly, in Fig.~\ref{fig:lambmix} one has the mixed contribution between Coulomb and radiative photon, both emitted from the electron line. In Fig.~\ref{fig:lambradia} one has a purely radiative contribution, which contributes to order $\alpha$ as well. Finally, in Fig.~\ref{fig:lambtad}, one has the tadpole contribution. The detailed results are presented in Sec. IV.
\section{Order-$\frac{\alpha}{m_e}$ Matching for $T^{ij}_{\rm NRQED}$}
To calculate the form factor $C$ to the next-to-leading order, we first consider the contribution from the quantum correction to the momentum current, and
match the $T^{ij}_{\rm NRQED}$ to that of QED at order $\alpha$. Since the matching is only sensitive to the UV contribution, it is sufficient to consider a free electron without the background field $V_p$. Furthermore, for our purpose,
we only need to consider spin-independent part.
The matching of the momentum current form factor for the free electron states starts from the full QED result after NR reduction,
\begin{align}\label{eq:QEDform}
&\langle \vec{p}+\vec{Q}|T^{ij}_{\rm QED}|\vec{p}-\vec{Q}\rangle\nonumber \\
&=\frac{p^ip^j}{m_e}A(Q^2)+(Q^iQ^j-\delta^{ij}Q^2)\tilde C_{\rm QED}(Q^2)
\end{align}
where $|\vec{p}+\vec{Q}\rangle$ denotes free-electron state with spatial-momentum equals to $\vec{p}+\vec{Q}$. We choose
the Breit frame, and for simplicity define $\vec{Q}\equiv \vec{q}/2$.The EMT form factor $A(Q^2)=1+{\cal O}(\frac{Q^2}{m_e^2})$ receives quantum corrections starting from order $\frac{\alpha Q^2}{m_e^2}$, while $\tilde C_{\rm QED}(Q^2)$ receives corrections at ${\cal O}(\frac{\alpha}{m_e})$.
$T^{ij}_{\rm NRQED}$ may receive corrections beyond the tree-level expression
\begin{align}
&T^{ij}_{\rm NRQED}=a_{\rm tree}(T^{ij}_{e}+T^{ij}_{\gamma})\nonumber \\
&+d_0(\partial^{i}\partial^j-\delta^{ij}\partial^2)\Psi^{\dagger}\Psi+{\cal O}(\frac{\alpha}{m_e^2}) \ .
\end{align}
where ${\cal O}(\frac{\alpha}{m_e^2})$ denotes high-dimensional operators such as
\begin{align}
{\cal O}_{1}^{ij}=\frac{a_1}{m_e^2} \partial^2 T^{ij}_{e0} \ , \\
{\cal O}_{2}^{ij}=\frac{a_2}{m_e^2} \partial^2(\partial^{i}\partial^j-\delta^{ij}\partial^2)\Psi^{\dagger}\Psi \ ,
\end{align}
and so on, all starting from order $\frac{\alpha}{m_e^2}$. The matching coefficients $a_{\rm tree}$, $d_0$, $a_1$, $a_2$ must be solved in order to reproduce Eq.~(\ref{eq:QEDform}) order by order in $\frac{1}{m_e}$ and $\alpha$:
\begin{align}
\langle \vec{p}+\vec{Q}|T^{ij}_{\rm QED}|\vec{p}-\vec{Q}\rangle=\langle \vec{p}+\vec{Q}|T^{ij}_{\rm NRQED}|\vec{p}-\vec{Q}\rangle \ .
\label{eq:matching}
\end{align}
For example, to match to $\frac{Q^2}{m_e^2}$ correction in $A$ one needs $a_1$ and to match to $\frac{Q^2}{m_e^2}$ correction in $\tilde C$ one needs $a_2$. Spin part of the Lagrangian will also be relevant to $a_1$ and $a_2$ as well.
However, as we have already shown in previous section, when the proper power-counting of $p^ip^j \sim {\cal O}(\alpha^2)$ are being taken into account, radiative corrections from $\langle T^{ij}_{e} \rangle_{H}$ will appear only at order $\alpha^3\frac{Q^2}{m_e}$ when averaged in the bound state. The same will apply to $a_1$ and $a_2$. Therefore the only matching constant useful in our calculation is $d_0$, which receives contribution already at order $\alpha$ and is caused purely by $T_{\gamma}^{ij}$. To obtain $d_0$, it is even simpler to work in the frame with $\vec{p}=0$, where the $A$ contribution disappears.
Thus, to compute the matching coefficients using Eq. ( \ref{eq:matching}), we first calculate
\begin{align} \label{eq:formfactor}
\langle \vec{Q}|T_{\rm tree}^{ij}|-\vec{Q}\rangle=(Q^iQ^j-\delta^{ij}Q^2)\tilde C(Q^2) \ ,
\end{align}
appearing in the right-hand side, and
obtain the explicit formula for $\tilde C(Q^2)$ at order $\frac{\alpha}{m_e}$. We show that it has the same logarithmic divergences when $Q^2\rightarrow 0$ as the full QED, but differs in UV. We perform this calculation in Coulomb gauge with the standard dimensional regularization (DR) with $D=3-2\epsilon$ for UV divergence.
\subsection{Fermionic contributions}
The relevant diagrams are shown in Fig~\ref{fig:singlefermion}. To order $\frac{1}{m}$, it is easy to show that in the Coulomb gauge, only the tadpole diagram contributes and reads
\begin{align}\label{eq:fermiontadpole}
\langle \vec{Q}|T^{ij}_{e}|-\vec{Q}\rangle=\frac{e^2}{2m_e}\mu^{2\epsilon}\int \frac{d^D\vec{k}}{(2\pi)^D|\vec{k}|}\bigg(\delta^{ij}-\frac{k^ik^j}{\vec{k}^2}\bigg) \ .
\end{align}
It is free from IR divergence, $Q$-independent therefore vanishes in DR. All other diagram are of order at least $\frac{1}{m^2}$ and will not contribute to the matching.
\begin{figure}[h!]
{%
\includegraphics[height=2cm]{fermion1l.pdf}
}
\caption{The contribution of $T^{ij}_{e}$ at one loop. The tadpole diagram, which is absent in the relativistic fermionic theory, is caused by the $\frac{e^2}{m}\Psi^{\dagger}\Psi A^iA^j$ term in $T^{ij}_{e}$. }
\label{fig:singlefermion}
\end{figure}
\subsection{Photonic contributions}
One needs to consider the diagrams in Fig.~\ref{fig:singlenonre}. To simplify notation, we write $\langle \vec{Q}|T^{ij}|-\vec{Q}\rangle$ as $\langle T^{ij}\rangle (2\vec{Q})$ and the argument $(2\vec{Q})$ are frequently omitted without causing confusion.
\begin{figure}[htb]
\begin{subfigure}{.2\textwidth}
\centering
\includegraphics[width=\linewidth]{single.pdf}
\caption{}
\label{fig:singlec}
\end{subfigure}
\begin{subfigure}{.29\textwidth}
\centering
\includegraphics[width=\linewidth]{singlemixed.pdf}
\caption{}
\label{fig:singlemixed}
\end{subfigure}
\begin{subfigure}{.2\textwidth}
\centering
\includegraphics[width=\linewidth]{singletad.pdf}
\caption{}
\label{fig:singletad}
\end{subfigure}
\caption{The contribution of $T_{\gamma}^{ij}$ in non-relativistic reduction. Notice that the tadpole diagram contributes at order $\frac{1}{m_e}$, same as the mixed one.}
\label{fig:singlenonre}
\end{figure}
We first consider the pure Coulomb contribution, shown in Fig~\ref{fig:singlec}. This term is transverse by itself and reads
\begin{align}
\langle T^{ij}_{\gamma \parallel}\rangle_{\rm 9a}(2\vec{Q}) =\frac{\alpha\pi}{8|Q|}(Q^iQ^j-\delta^{ij}Q^2) \ .
\end{align}
The mixed contribution is shown in part b) Fig~\ref{fig:singlemixed}. The two diagrams reads
\begin{align}
\langle T^{ij}_{\gamma \parallel \perp}\rangle_{\rm 9b}(2\vec{Q})=&e^2\mu^{2\epsilon}\int \frac{d^D\vec{k}}{(2\pi)^D}\frac{(k-Q)^iP^{jk}(\vec{k}+\vec{Q})Q^k+(i\leftrightarrow j)}{m_e|\vec{k}-\vec{Q}|^2|\vec{k}+\vec{Q}|} \nonumber \\
-&e^2\delta^{ij}\mu^{2\epsilon}\int \frac{d^D\vec{k}}{(2\pi)^D}\frac{(k-Q)^lP^{lk}(\vec{k}+\vec{Q})Q^k}{m_e|\vec{k}-\vec{Q}|^2|\vec{k}+\vec{Q}|} \ ,
\end{align}
where the standard triple vertex $-\frac{ie}{2m_e}(p+p')^j$ and the relation $P^{jk}(\vec{k}+\vec{Q})(Q-k)^k=2P^{jk}(\vec{k}+\vec{Q})Q^k$ have been used. Due to rotational invariance, one can reduce the above integral to two scalar integrals
\begin{align}\label{eq:7boriginal}
&\langle T^{ii}_{\gamma \parallel \perp}\rangle_{\rm 9b }(2\vec{Q}) =2e^2(D-2)\mu^{2\epsilon}\int \frac{d^D\vec{k}}{(2\pi)^D}\frac{Q^lP^{lk}(\vec{k}+\vec{Q})Q^k}{m_e|\vec{k}-\vec{Q}|^2|\vec{k}+\vec{Q}|} \ , \\
&Q^{i}\langle T^{ij}_{\gamma \parallel \perp} \rangle_{\rm 9b}(2\vec{Q}) Q^j=2e^2\mu^{2\epsilon}\int \frac{d^D\vec{k}}{(2\pi)^D}\frac{\vec{k}\cdot \vec{Q}Q^lP^{lk}(\vec{k}+\vec{Q})Q^k}{m_e|\vec{k}-\vec{Q}|^2|\vec{k}+\vec{Q}|} \ .
\end{align}
They are evaluated in Appendix.B. The result in the $\epsilon \rightarrow 0$ limit, reads
\begin{align}\label{eq: 7btranse}
\langle T^{ii}_{\gamma\parallel\perp}\rangle_{\rm 9b}=\frac{e^2}{3m_e\pi^2 \epsilon}Q^2+\frac{e^2}{3m_e\pi^2}Q^2\bigg(-\ln \frac{Q^2}{\mu^2}+c_1\bigg) \ ,
\end{align}
and
\begin{align}\label{eq: 7blongi}
Q^i\langle T^{ij}_{\gamma \parallel \perp}\rangle_{\rm 9b} Q^j=-\frac{e^2}{15m_e\pi^2 \epsilon}Q^4+\frac{e^2}{15m_e\pi^2}Q^4\bigg(\ln \frac{Q^2}{\mu^2}+c_2\bigg) \ .
\end{align}
with
\begin{align}
c_1= \gamma_E -5+\ln 4+\ln \pi +2 \psi\left(\frac{5}{2}\right) \ , \\
c_2=-\gamma_E +3-\ln 4-\ln \pi -2 \psi\left(\frac{7}{2}\right) \ .
\end{align}
Notice that the digamma function is defined as $\psi(z)=\frac{d}{dz}\ln \Gamma(z)$, $\psi(\frac{3}{2})=2-\gamma_E-\ln4$ and one has the recursive relation $\psi(z+1)=\psi(z)+\frac{1}{z}$.
From these, it is clear that the mixed contributions themselves are not transverse.
To obtain a transverse EMT, one must include the contributions from the tadpole diagrams shown in Fig~\ref{fig:singletad} as well. The detail of the calculation is present in Appendix.C, here we only present the result. First, the non-conserved part of Fig.~\ref{fig:singletad} can be calculated as
\begin{align}\label{eq:7nnoncon}
&Q^i\langle T^{ij}_{\gamma \perp}\rangle_{\rm 9c} Q^j\nonumber \\
&=e^2Q^4\mu^{2\epsilon}\frac{2^{2 \epsilon-4} (\epsilon-1) \pi ^{\epsilon-\frac{3}{2}} \left(Q^2\right)^{-\epsilon} \Gamma (3-2 \epsilon) \Gamma (\epsilon-1)}{m_e \Gamma \left(\frac{7}{2}-2 \epsilon\right)} \ .
\end{align}
Notice the appearance of $\Gamma(\epsilon-1)$ due to the quadratic divergence. Thus by expanding around $\epsilon= 0$, one has
\begin{align}\label{eq:7nnonconresult}
Q^i\langle T^{ij}_{\gamma \perp}\rangle_{\rm 9c} Q^j=\frac{e^2}{15m_e\pi^2\epsilon}Q^4+\frac{e^2}{15m_e\pi^2}Q^4\bigg(-\ln \frac{Q^2}{\mu^2}+c_2'\bigg) \ ,
\end{align}
with
\begin{align}
c_2'=\gamma_E -3+\ln 4+\ln \pi +2 \psi \left(\frac{7}{2}\right)
\end{align}
Clearly, $c_2=-c_2'$, therefore one has
\begin{align}
Q^{i}\langle T^{ij}_{\gamma \perp}\rangle_{\rm 9c} Q^j+Q^{i}\langle T^{ij}_{\gamma \perp}\rangle_{\rm 9d} Q^j==0 \ ,
\end{align}
and the $T^{ij}$ is conserved. Similarly, the trace part for Fig.~\ref{fig:singletad} is calculated in Appendix.C as
\begin{align}\label{eq:7dtrace}
\langle T^{ii}_{\gamma \perp} \rangle_{\rm 9c}(2\vec{Q})=-\frac{Q^2 e^2 c_1'}{18m_e\pi ^2 } \ ,
\end{align}
where
\begin{align}
c_1'&= 9 \gamma_E -25+3\ln(4)-3 \psi\left(\frac{3}{2}\right)+12 \psi\left(\frac{5}{2}\right)\nonumber \\
&=1-6\ln 4 \ .
\end{align}
Thus, both of the tadpole and the mixed diagram are required in order to maintain transversity. However, the divergences in the $C(Q)$ can be reads from the $T^{ii}$ for the mixed diagram Fig.\ref{fig:singlemixed} only, while the $T^{ii}$ for the tadpole diagram is logarithm-free.
\subsection{Matching to QED}
We now collect the results and match to full QED. By combining Eq.~(\ref{eq: 7btranse}) and Eq.~(\ref{eq:7dtrace}), the full contribution of $T^{ij}_{\rm tree}$ reads
\begin{align}
\langle T^{ij}_{\rm tree}\rangle (2\vec{Q})=(Q^iQ^j-\delta^{ij}Q^2)\tilde C(Q^2) \ ,
\end{align}
where
\begin{align}
\tilde C(Q^2)=\frac{\alpha\pi}{8|Q|}+\frac{e^2}{6m_e\pi^2}\bigg(-\frac{1}{\epsilon}+\ln \frac{Q^2}{\mu^2}+\gamma_E-\ln\pi-\frac{7}{6}\bigg) \ .
\end{align}
In comparison, the small-$Q$ asymptotics of $C$-form factor for relativistic electron in the full QED can be obtained from literature~\cite{Berends:1975ah,Milton:1977je} as
\begin{align}\label{eq:CQED}
\tilde C_{\rm QED}(Q^2)=\frac{\alpha\pi}{8|Q|}+\frac{e^2}{6m_e\pi^2 }\ln \frac{4Q^2}{m_e^2}-\frac{11e^2}{72m_e\pi^2 }
\end{align}
It has the same IR structure as the NRQED, but differs in UV. The required formulas are collected in Appendix.G.
It is clear now that in order to match to the full QED, one simply needs to add to the tree-level
momentum current of the NRQED the following local counter term
\begin{align}
T^{ij}_{\rm NRQED}= T^{ij}_{\rm tree}+d_0(\partial^i\partial^j-\delta^{ij}\partial^2)\Psi^{\dagger}\Psi \ ,
\end{align}
where $d_0$ contain only logarithms in $\mu$ and $m_e$,
\begin{align}
d_0=-\frac{\alpha}{6\pi m_e}\bigg(\frac{1}{\epsilon}+\ln \frac{4\mu^2}{m_e^2}+\ln \pi -\gamma_E+\frac{1}{4}\bigg) \ .
\end{align}
This concludes our construction of the momentum current $T^{ij}$ in NRQED.
\section{${\cal O}(\alpha)$ Radiative Corrections to $\tau_H$}
After obtaining the matching coefficient $d_0$, in this section we calculate the $\langle 0|T^{ij}_{\rm NRQED} |0\rangle_H$ in the bound-state and obtain the final result for $\tau_H$. More explicitly, the calculation proceeds as follows.
We will present the result for the mixed diagrams Fig.~\ref{fig:lambmix} and Fig.~\ref{fig:lambmixinter} and then for the radiative diagram Fig.~\ref{fig:lambradia}. We show explicitly that the non-conserved part $Q^{i}T^{ij}Q^j$ cancel with the tadpole diagram Fig.~\ref{fig:lambtad} that is essentially independent of the bound state and remains the same as the free NRQED calculation in previous section. The power-counting rules in the ultra-soft region are used to decouple the matrix elements and the momentum integrals, therefore in principle our calculation is only valid in this region as well. However, it is not hard to show that in the the mixed diagram, the only diagram which diverges in UV, our formulas hold in the soft region as well. Therefore, the result has the same UV structure as the free NRQED and can be matched to the full QED using the same matching coefficient $d_0$ obtained in the previous section, which leads to our final result Eq.~(\ref{eq:finalre}). Its numerical value will also be provided.
For notational simplicity, the momentum transfer will be $\vec{q}=2\vec{Q}$ and we will use the following notation frequently
\begin{align}
\langle T^{ij}\rangle (2\vec{Q})\equiv \int d^D \vec{x}e^{-i2\vec{Q}\cdot \vec{x}}\langle 0|T^{ij}(\vec{x})|0\rangle \ .
\end{align}
Without causing confusion, the argument $(2\vec{Q})$ will be omitted. Since we are only interested in the small-$Q$ behavior of the form factor, expansion to quadratic order in $Q$ will always be understood.
\subsection{The mixed diagrams Fig.~\ref{fig:lambmix} and Fig.~\ref{fig:lambmixinter}}
We first consider the mixed diagrams in Fig.~\ref{fig:lambmix} and Fig.~\ref{fig:lambmixinter}. We show that the interference diagram Fig.~\ref{fig:lambmixinter} vanishes. Indeed, using the Feynman rule, one has
\begin{align}
&\langle T^{ij}_{\perp p}\rangle_{\rm 7b} =-e^2\mu^{2\epsilon}\int\frac{d^D\vec{k}}{(2\pi)^D}\frac{(k-2Q)^i}{|\vec{k}-2\vec{Q}|^2}\frac{P^{jl}(\vec{k}) v^l_{00}(\vec{k})}{|\vec{k}|}\nonumber \\
&+(i\leftrightarrow j)+\delta^{ij}e^2\mu^{2\epsilon}\int\frac{d^D\vec{k}}{(2\pi)^D}\frac{Q^l}{|\vec{k}-2\vec{Q}|^2}\frac{P^{ll'}(\vec{k}) v^{l'}_{00}(\vec{k})}{|\vec{k}|}\ ,
\end{align}
which vanishes due to the fact that
\begin{align}
\vec{v}_{00}(\vec{k})= \int d^D\vec{x} \psi^{\dagger}_{0}(\vec{x})\frac{-i\nabla }{m_e}\psi_{0}(\vec{x})e^{-i\vec{k}\cdot \vec{x}}\propto \vec{k}
\end{align}
which contracts to zero with $P^{ij}(\vec{k})$.
Therefore, it remains to calculate Fig.~\ref{fig:lambmix}. We start with $T^{ii}$. Notice that for $k={\cal O}(\alpha^2 m_e)$, one has the standard dipole-expansion~\cite{Weinberg:1995mt,Pineda:1997ie} of the matrix elements defined in Eq.~(\ref{eq:matrix})
\begin{align}
v^{i}(\vec{k})= v^i_{MN} +{\cal O}(\alpha)\ , \\
\rho_{MN}(\vec{k}) =\delta_{MN}-i\vec{k}\cdot \vec{x}_{MN} +{\cal O}(\alpha) \ .
\end{align}
where $\vec{v}=-\frac{i\vec{\nabla}}{m_e}$. Using these one has for the trace part
\begin{align}\label{eq:5atrace}
&\langle T^{ii}_{\gamma \parallel \perp}\rangle_{\rm 7a} = -2(D-2)\mu^{2\epsilon}\frac{e^2}{D}\sum_{M}i(x^i_{0M}v^i_{M0}-v^i_{0M}x^i_{M0})\nonumber \\
&\times \int \frac{d^Dk}{(2\pi)^D}\frac{Q^lP^{lk}(\vec{k})Q^k}{|\vec{k}-2\vec{Q}|^2\bigg(|\vec{k}|+E_M-E_0\bigg)} \ .
\end{align}
Clearly, the divergent part is independent of the bound-state thanks to canonical commutation relation $[x,p]=i$ in any dimension
\begin{align}
-\frac{1}{D}\sum_{M}i(x^i_{0M}v^i_{M0}-v^i_{0M}x^i_{M0})\equiv \frac{1}{m_e}\ .
\end{align}
In particular, in the soft region where $ |\vec{k}|={\cal O}(\alpha m_e)$, the formula above is also valid at small $Q$, after neglecting the binding energies in the denominator. The integral is calculated in Appendix.D. The result reads
\begin{widetext}
\begin{align}\label{eq:5aresult1}
\langle T^{ii}_{\gamma \parallel \perp}\rangle_{\rm 7a} (2\vec{Q})= \frac{e^2}{\pi^2D}Q^2\sum_{M}i(\vec{v}_{0M}\cdot \vec{x}_{M0}-\vec{x}_{0M}\cdot \vec{v}_{M0})\bigg(\frac{1}{3\epsilon}-\frac{1}{3}\ln \frac{(E_M-E_0)^2}{\mu^2}+\frac{-3\gamma_E+3\ln \pi -1}{9} \bigg)
\end{align}
At this step, it is helpful to perform certain simplification of the matrix elements. Notice the following equalities in $D$ dimensions
\begin{align}
\langle N|[H,\vec{x}]|M\rangle=-i\langle N|\vec{v}|M\rangle=(E_N-E_M)\langle N|\vec{x}|M \rangle \ ,
\end{align}
thus for any $M \ne N$ \ ,
\begin{align}
- i (\vec{x}_{NM}\cdot \vec{v}_{MN}-\vec{v}_{NM}\cdot\vec{x}_{MN})=2\frac{\vec{v}_{NM}\cdot \vec{v}_{MN}}{E_M-E_N} \ .
\end{align}
Therefore one has
\begin{align}
\langle T^{ii}_{\gamma \parallel\perp}\rangle_{\rm 7a} (2\vec{Q})= \frac{e^2}{\pi^2}Q^2\sum_{M}\frac{2\vec{v}_{M0}\cdot \vec{v}_{0M}}{D(E_M-E_0)} \bigg(\frac{1}{3\epsilon}-\frac{1}{3}\ln \frac{(E_M-E_0)^2}{\mu^2}+\frac{-3\gamma_E+3\ln \pi -1}{9} \bigg) \ .
\end{align}
in which the matrix elements $\vec{x}_{MN}$ are eliminated.
Similarly, the non-conserved part $Q^i\langle T^{ij}_{\gamma \parallel \perp}\rangle_{\rm 7a} Q^j$ can be calculated as
\begin{align}\label{eq:5anoncon}
Q^i \langle T^{ij}_{\gamma \parallel \perp} \rangle_{\rm 7a} Q^j= \sum_{M}\frac{2\vec{v}_{M0}\cdot \vec{v}_{0M}}{D(E_M-E_0)}2e^2\mu^{2\epsilon}\int \frac{d^D\vec{k}}{(2\pi)^D}\frac{(\vec{k}\cdot \vec{Q}-\vec{Q}^2)(Q^2\vec{k}^2-(\vec{k}\cdot \vec{Q})^2)}{|\vec{k}-2\vec{Q}|^2|\vec{k}|^2(|\vec{k}|+E_M-E_0)} \ .
\end{align}
By parameterizing as usual, one obtains
\begin{align}\label{eq:5anonconresult}
Q^i \langle T^{ij}_{\gamma \parallel \perp} \rangle_{\rm 7a} Q^j= \frac{Q^4e^2}{\pi^2}\sum_{M}\frac{2\vec{v}_{M0}\cdot \vec{v}_{0M}}{D(E_M-E_0)} \bigg(-\frac{1}{15\epsilon}+\frac{1}{15}\ln \frac{(E_M-E_0)^2}{\mu^2}+\frac{-15\ln \pi+15\gamma_E-1}{225}\bigg) \ .
\end{align}
The detail of the calculation is presented in Appendix.D. It contains the same UV divergence as the mixed diagram for the single electron, but the IR divergence $\ln Q^2$ is regulated by the binding energy differences. Therefore, to cancel the $\ln Q^2$ in the tadpole diagram, there must be contributions of the form $\ln \frac{Q^2}{(E_M-E_0)^2}$. As we will see, the radiative diagram Fig.~\ref{fig:lambradia} will exactly produce this missing piece.
\subsection{Purely radiative contribution Fig.~\ref{fig:lambradia}}
The last but the most complicate diagram that remains to be calculated is the pure radiative contribution in Fig.~\ref{fig:lambradia}. We start with the $\langle T^{ii}_{\gamma \perp}\rangle_{\rm 7c}$. Using standard Feynman rule, this can be written as
\begin{align}\label{eq:radiab}
\langle T^{ii}_{\gamma \perp}\rangle_{\rm 7c}=\sum_{M}\vec{v}_{0M}\cdot \vec{v}_{M0}\frac{2ie^2}{D}\mu^{2\epsilon}\int \frac{dk^0d^{D}\vec{k}}{(2\pi)^{D+1}}\frac{A(k,Q,D){\rm tr}(P(k-Q)P(k+Q))+2(4-D)Q^TP(k-Q)P(k+Q)Q}{\bigg(k_0^2-(\vec{k}-\vec{Q})^2+i0\bigg)\bigg(k_0^2-(\vec{k}+\vec{Q})^2+i0\bigg)\bigg(E_0-k^0-E_M+i0\bigg)} \ ,
\end{align}
where $A(k,Q,D)$ is defined in Eq.~(\ref{eq:AkQD}). Notice the similarity of the integrand to the tadpole contribution in Eq.~(\ref{eq:5d}). Since the calculation is rather tedious, we present all the details in Appendix.E. The result reads
\begin{align}\label{eq:5cresult}
\langle T^{ii}_{\gamma \perp}\rangle_{\rm 7c} (2\vec{Q})=-Q^2\frac{2e^2\ln4}{3\pi^2}\sum_{M}\frac{\vec{v}_{0M}\cdot \vec{v}_{M0}}{D(E_M-E_0)} \ .
\end{align}
We then move to the non-conserved part $Q^i\langle T^{ij}_{\gamma \perp}\rangle_{\rm 7c} Q^j$. This is the most involved part of the calculation and is presented in Appendix.E. The result reads
\begin{align}\label{eq:5cnonresult}
Q^i \langle T^{ij}_{\gamma \perp}\rangle_{\rm 7c} Q^j= \frac{Q^4e^2}{\pi^2}\sum_{M}\frac{2\vec{v}_{0M}\cdot \vec{v}_{M0}}{D(E_M-E_0)}\bigg(\frac{1}{15}\ln \frac{Q^2}{(E_M-E_0)^2}+\frac{15\ln 4-46}{225}\bigg) \ .
\end{align}
As claimed in the previous subsection, it contains the missing $\ln \frac{Q^2}{(E_0-E_M)^2}$ with the correct coefficient.
\subsection{Checking conservation. }
After finishing the difficult part of the calculation, here we check the conservation of $T^{ij}$. Combining Eq.~(\ref{eq:5anonconresult}) and Eq.~(\ref{eq:5cnonresult}), the $\ln (E_M-E_0)^2$ cancels, left with
\begin{align}\label{eq:Tlongira}
Q^i\left(\langle T^{ij}_{\gamma \parallel \perp}\rangle_{\rm 7a} +\langle T^{ij}_{\gamma \perp}\rangle_{\rm 7c} \right)Q^j=\frac{Q^4e^2}{\pi^2}\sum_{M}\frac{2\vec{v}_{M0}\cdot \vec{v}_{0M}}{D(E_M-E_0)} \bigg(-\frac{1}{15\epsilon}+\frac{1}{15}\ln \frac{Q^2}{\mu^2}+\frac{15\ln4-15\ln \pi+15\gamma_E-47}{225}\bigg) \ ,
\end{align}
which can be simplified after using the sum-rule Eq.~(\ref{eq:sumrule}) as
\begin{align}\label{eq:Tlongirare}
Q^i\left(\langle T^{ij}_{\gamma \parallel \perp}\rangle_{\rm 7a} +\langle T^{ij}_{\gamma \perp}\rangle_{\rm 7c}\right)Q^j=\frac{Q^4e^2}{m_e\pi^2} \bigg(-\frac{1}{15\epsilon}+\frac{1}{15}\ln \frac{Q^2}{\mu^2}+\frac{15\ln4-15\ln \pi+15\gamma_E-47}{225}\bigg) \ .
\end{align}
Here we show that it cancels with the tadpole contribution in Fig.~\ref{fig:lambtad}. Indeed, since the tadpole contribution is essentially independent of the bound-state, one has the same expression at small $Q$ as Eq.~(\ref{eq:7nnonconresult})
\begin{align}
Q^i\langle T^{ij}_{\gamma \perp}\rangle_{\rm 7d} Q^j=\frac{e^2}{15m_e\pi^2\epsilon}Q^4+\frac{e^2}{15m_e\pi^2}Q^4\bigg(-\ln \frac{Q^2}{\mu^2}+c_2'\bigg) \ ,
\end{align}
where
\begin{align}
c_2'=\gamma_E -3+\ln 4+\ln \pi +2 \psi \left(\frac{7}{2}\right)\equiv \frac{47}{15}-\gamma_E +\ln \pi- \ln 4 \ .
\end{align}
To obtain this we used again the well-known relation for digamma function
\begin{align}
\psi(n+\frac{1}{2})=-\gamma_E-\ln 4+\sum_{k=1}^n \frac{2}{2k-1} \ .
\end{align}
Therefore, for the bound state we have shown that the momentum current at order $\frac{\alpha}{m}$ is purely transverse
\begin{align}
Q^i\left(\langle T^{ij}_{\gamma \parallel \perp}\rangle_{\rm 7a}+\langle T^{ij}_{\gamma \perp}\rangle_{\rm 7c+7d}\right)Q^j\equiv 0 \ .
\end{align}
This is the most crucial consistency check of the whole calculation.
\subsection{The total result}
After showing the conservation of $T^{ij}$, we collect all the pieces of $T^{ii}$ and obtain the final result.
First, for the mixed diagram Fig~\ref{fig:lambmix}, one has
\begin{align}
\langle T^{ii}_{\gamma \parallel\perp}\rangle_{\rm 7a}(2\vec{Q})= \frac{e^2}{\pi^2}Q^2\sum_{M}\frac{2\vec{v}_{M0}\cdot \vec{v}_{0M}}{D(E_M-E_0)} \bigg(\frac{1}{3\epsilon}-\frac{1}{3}\ln \frac{(E_M-E_0)^2}{\mu^2}+\frac{-3\gamma_E+3\ln \pi -1}{9} \bigg) \ .
\end{align}
For the radiative part Fig.~\ref{fig:lambradia}, one has
\begin{align}
\langle T^{ii}_{\gamma \perp}\rangle_{\rm 7c} (2\vec{Q})=-\frac{e^2}{\pi^2}Q^2\sum_{M}\frac{2\vec{v}_{0M}\cdot \vec{v}_{M0}}{D(E_M-E_0)}\frac{\ln4}{3} \ .
\end{align}
For the tadpole part Fig.~\ref{fig:lambtad} , one has the same small-$Q$ result as in Eq.~(\ref{eq:7dtrace})
\begin{align}
\langle T^{ii}_{\gamma \perp}\rangle_{\rm 7d} (2\vec{Q})=-\frac{e^2}{\pi^2}Q^2\sum_{M}\frac{2\vec{v}_{0M}\cdot \vec{v}_{M0}}{D(E_M-E_0)}\frac{1-6\ln 4}{18} \ .
\end{align}
Therefore, combining all them, one has
\begin{align}
\langle T_{\rm tree}^{ii}\rangle (2\vec{Q})=\frac{e^2}{\pi^2}Q^2\sum_{M}\frac{2\vec{v}_{M0}\cdot \vec{v}_{0M}}{D(E_M-E_0)} \bigg(\frac{1}{3\epsilon}-\frac{1}{3}\ln \frac{(E_M-E_0)^2}{\mu^2}+\frac{-\gamma_E+\ln \pi }{3}-\frac{1}{6} \bigg) \ ,
\end{align}
which leads to
\begin{align}
\langle T^{ij}_{\rm tree}\rangle (2\vec{Q})=(Q^iQ^j-\delta^{ij}Q^2)\tilde C_{\rm s}(Q^2) \ ,
\end{align}
where
\begin{align}
\tilde C_{\rm s}(Q^2=0)=\frac{e^2}{6\pi^2}\sum_{M}\frac{2\vec{v}_{M0}\cdot \vec{v}_{0M}}{D(E_M-E_0)} \bigg(-\frac{1}{\epsilon}+\ln \frac{(E_M-E_0)^2}{\mu^2}+\gamma_E-\ln \pi-\frac{1}{2} \bigg) \ .
\end{align}
To match it to QED, one only needs to add to the above result $-4d_0$
\begin{align}
\tilde C(Q^2=0)=\tilde C_{\rm s}(Q^2=0)-4d_0=\frac{e^2}{6\pi^2}\sum_{M}\frac{2\vec{v}_{M0}\cdot \vec{v}_{0M}}{D(E_M-E_0)} \bigg(\ln \frac{4(E_M-E_0)^2}{m_e^2}-\frac{1}{4} \bigg) \ .
\end{align}
Since our $\vec{Q}$ is twice of the momentum transfer $\vec{q}=2\vec{Q}$, one finally has
\begin{align}
\langle T^{ij}\rangle_H (\vec{q})=(q^iq^j-\delta^{ij}q^2)\frac{C_H(q)}{m_e}\ ,
\end{align}
with
\begin{align}\label{eq:finalre}
\tau_H=\frac{C_H(0)}{m_e}=\frac{1}{4m_e}+\frac{\alpha}{6\pi}\sum_{M}\frac{2\vec{v}_{M0}\cdot \vec{v}_{0M}}{D(E_M-E_0)} \bigg(\ln \frac{4(E_M-E_0)^2}{m_e^2}-\frac{1}{4} \bigg) \ .
\end{align}
This is the major result of the paper. Notice that the leading order result has been added.
\end{widetext}
To estimate how large the order $\alpha$ contribution is, one needs to calculate the sum over $M$. If $E_M-E_0$ is in the numerator, this is called the Bethe logarithm and receives large contribution from the continuum spectrum. In our case, we expect the continuum spectrum is also important. In fact, after re-scaling, the contribution can be written as
\begin{align}
\tau_H=\frac{1}{4m_e}+\frac{\alpha}{6\pi m_e} \bigg(\ln \alpha^2+\tau_{\rm d}+\tau_{\rm c}-\frac{1}{4}\bigg) \ ,
\end{align}
where $\tau_d=-0.264$ and $\tau_c=0.458$ are contributions from the discrete and continuum spectrum, which are defined and evaluated in Appendix F. Put in numbers, one has
\begin{align}\label{eq:resultnu}
\frac{\tau_H}{\tau_0}-1=\frac{4\alpha}{3\pi } \bigg(\ln \alpha-0.028\bigg)=-1.54\times 10^{-2} \ .
\end{align}
Although opposite in sign, the order $\alpha$ contribution is two orders of magnitude smaller comparing to the leading order contribution.
\section{Comment and Conclusion}
Before ending the paper, here we briefly comment on the sign of $\tau_H$. One first notice that in the result Eq.~(\ref{eq:resultnu}), $\ln \alpha^2$ dominate over the constant $-0.056$, therefore the sign at order $\alpha$ is mainly due to the logarithms, which already appears at the level of single electron. In fact, from the calculation we have learned that only the mixed diagram contributes to this logarithm, while the purely-radiative and tadpole diagrams contribute only to the constant.
Besides our calculation in NRQED with dimensional regularization, one can also perform the calculation directly from the dressed Dirac theory in a way similar to Ref.~\cite{Weinberg:1995mt}. In order to obtain the correct expansion in $\alpha$, one should separate the high-energy and low energy contributions into two parts
\begin{align}
\frac{1}{k^2+i0}\rightarrow \frac{1}{k^2-\mu^2+i0}+\bigg(\frac{1}{k^2+i0}-\frac{1}{k^2-\mu^2+i0}\bigg) \ ,
\end{align}
with the fictitious photon mass $\mu$ satisfying $\alpha^2m_e \ll \mu \ll \alpha m_e$. In the first term, the photon mass will guarantee that the ultra-soft region is non-essential, and the calculation can be performed by completely neglecting the bound-state structure for a single relativistic electron. The second term can be calculated using non-relativistic approximations for the electron as usual, with $\mu$ playing the role of the UV regulator. In fact, one may think that the first term just defines the ``matching constant'' $d_0$ in this scheme. The trouble with the photon mass regulator is that the EMT is not guaranteed to be conserved for finite $\mu$, and the power-divergences in $\mu$ requires additional attention. On contrary, the EMT in dimensional regularization is automatically conserved, and the power-divergence disappears in DR as well.
In conclusion, we have constructed the momentum current density of NRQED up to order $\frac{1}{m_e}$, from which the ${\cal O}(\alpha)$ tensor monopole moment $\tau_H$ for the ground state of hydrogen atom is calculated. Although suffering from IR divergence for a single free electron, $\tau_H$ is finite and remains positive after including the ${\cal O}(\alpha)$ correction. The IR logarithm in NRQED is naturally regulated by the binding energy differences, and the fictitious UV divergence of NRQED ``matches'' precisely with the IR divergence of the relativistic theory, guarantee the ultimate consistency of our calculation. The final
result is similar in expression to the famous Lamb shift of the energy levels.
{\it Acknowledgment.}---
This research is supported by the U.S. Department of Energy, Office of Science, Office of Nuclear Physics, under contract number DE-SC0020682, and by the Priority Research Area SciMat under the program Excellence Initiative - Research University at the Jagiellonian University in Krak\'{o}w.
|
\section{INTRODUCTION}
\label{sec:introduction}
The Cosmology Large Angular Scale Surveyor (CLASS) is a polarization-sensitive telescope array located at an altitude of 5,200 m in the Chilean Atacama Desert.
CLASS is designed to measure ``E-mode'' (even parity) and ``B-mode'' (odd parity) polarization patterns in the Cosmic Microwave Background (CMB) over large angular scales ($>1\degree$).
CLASS seeks to improve our understanding of inflation, reionization, and dark matter \cite{Tom-overview, Harrington-overview}.
CLASS is currently observing with three telescopes covering four frequency bands: one at 40~GHz (Q); one at 90~GHz (W1); and one dichroic system at 150/220~GHz (G).
The Q, W1, and G telescopes have been observing since June 2016, May 2018, and September 2019, respectively.
In-lab characterization and on-sky performance of these detector arrays has been discussed in Refs.~\citenum{Appel-Q, Dahal-W1, Dahal-HF, Dahal-multifrequency}.
In these proceedings, we discuss the redesign and preliminary in-lab characterization of new 90~GHz detectors.
Design changes were made with the goal of improving detector stability and optical efficiency.
These detectors are currently being fielded to upgrade the W1 focal plane.
We anticipate first light for the new W1 focal plane array in the austral winter of 2022.
\section{Focal Plane Design}
\label{sec:focal_plane_design}
\subsection{Detector Design}
\label{subsec:detector_design}
\begin{figure} [ht]
\begin{center}
\includegraphics[width=\textwidth, trim={2cm 0 2cm 0}]{w-band-detector-wafer_updated.jpg}
\end{center}
\caption{ \label{fig:design}
CLASS 90~GHz wafer (left), with zoom in of a detector pixel and updated transition-edge sensor (TES) architecture. For a full description of the 90~GHz wafer design, see Refs.~\citenum{Chuss-development, Denis-fabrication, Rostem-design}. Each of the 37 detector pixels consists of a symmetric planar ortho-mode transducer (OMT), which reads out two orthogonal linear polarizations over Nb microstrip transmission lines to MoAu bilayer TES bolometers.
The updated TES architecture includes a simplified absorber with a resistive meander, a direct metal connection between the TES and the Pd, and a revised choke filter circuit implementation.
These changes were made to improve stability and optical efficiency performance of the TES bolometers.}
\end{figure}
CLASS focal planes consist of arrays of highly sensitive feedhorn-coupled transition-edge sensor (TES) bolometers.
TES bolometers provide background-limited sensitivity, required to achieve CLASS' science goals.
Their design is scalable to large focal plane arrays across multiple frequencies, required for high sensitivity measurements and separation of the CMB signal from polarized Galactic foregrounds.
In Figure~\ref{fig:design}, an example CLASS 90~GHz detector wafer is shown.
The detector wafers, which integrate 37 detector pixels, are fabricated at NASA Goddard Space Flight Center.
Following the blue leak mitigation strategy presented in Ref.~\citenum{Wollack-blueleak}, the detector wafer is indium-bump bonded to a backshort wafer for signal termination and a photonic choke~\cite{Crowe-choke} wafer, which serves as an interface to the sensor array’s feedhorns.
Each of the 37 detector pixels consists of a symmetric planar ortho-mode transducer (OMT), which reads out two orthogonal linear polarizations ($+45\degree$ and $-45\degree$ from the vertical) to two TES bolometers.
Signals from opposing antenna probes are coherently combined onto a single microstrip transmission line using the difference output of a Magic Tee, which transmits a single mode.~\cite{U-Yen-magicT}
On-chip filtering and micromachined silicon packaging define the signal bandpass.~\cite{Crowe-choke}
Finally, the signal from each of the two orthogonal polarizations is passed to a MoAu bilayer TES bolometer.
During operation, the TES bolometers are voltage-biased to their superconducting transition critical temperature ($T_\mathrm{c}$) of $\sim$150~mK.
For a full description of the original 90~GHz detector design, see Refs.~\citenum{Chuss-development, Denis-fabrication, Rostem-design}.
The updated TES architecture, shown in the rightmost panel of Figure~\ref{fig:design}, includes three primary design changes from the original CLASS 90~GHz detectors:
\begin{enumerate}
\item a simplified absorber that terminates power from the sky (brought in via a Nb microstrip) onto the TES island, with a resistive PdAu meander that consists of a stepped impedance transition from Nb to PdAu;
\item the addition of a direct normal-metal connection between the TES and the heat capacity element formed by the Pd, to effectively lump the electronic heat capacity into a single element;
\item the revision of the choke filter circuit design to extend onto the membrane’s diffusive bolometer legs.
\end{enumerate}
These design changes were introduced in order to improve optical efficiency (1, 3), and improve stability of the TES transition (2), which are described in Ref.~\citenum{Dahal-multifrequency}.
In addition, the redesign introduced a more compact absorber at the Magic Tee and at the microstrip crossovers.
\subsection{Module Assembly Structure}
\label{subsec:module_assembly_structure}
\begin{figure} [ht!]
\begin{center}
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\linewidth]{mounted_wafer_labeled.jpg}
\caption{} \label{fig:mounted_wafer}
\end{subfigure}
\hfill
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\linewidth]{bonded_module_colors_adjusted_labeled.jpg}
\caption{} \label{fig:bonded_module}
\end{subfigure}
\bigskip
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\linewidth]{side_view_assembly_stack_labeled.jpg}
\caption{} \label{fig:side_view_assembly}
\end{subfigure}
\hfill
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\linewidth]{W2_in_field_labeled_small.jpg}
\caption{} \label{fig:W1_in_field}
\end{subfigure}
\end{center}
\caption[example]
{ \label{fig:assembly} The CLASS detector assemblies are integrated into focal plane array modules for ease of testing and readout; a full focal plane consists of seven modules. Panel~\ref{fig:mounted_wafer} shows a CLASS wafer mounted and aligned onto a CE7 baseplate. Panels~\ref{fig:bonded_module} and \ref{fig:side_view_assembly} show all interior layers of a fully integrated module, before it is packaged into its final configuration. For a full description of the assembly stack, see Ref.~\citenum{Dahal-W1}. Panel~\ref{fig:W1_in_field} shows the four new CLASS 90~GHz modules (outlined in blue) situated in the fielded W1 receiver. We anticipate first light for these new detectors in the austral winter of 2022.}
\end{figure}
The micromachined CLASS detector assemblies are subsequently integrated into focal plane array modules at Johns Hopkins University (JHU) to facilitate testing and readout.
A full 90~GHz focal plane consists of seven modules.
Figure~\ref{fig:assembly} highlights various stages of the assembly process, as well as the full upgraded W1 focal plane.
Smooth-walled feedhorns\cite{Feedhorns} couple light from the sky onto the TES bolometers.
The machined Cu feedhorns are individually installed onto the front of a Au-plated Si Alloy Controlled Expansion 7 (CE7) baseplate.\cite{CE7}
CE7, composed of 70\% Si and 30\% Al, is chosen due to its machinability and its low differential thermal contraction relative to Si.
The Si wafer is then mounted and aligned onto the baseplate using 1) two alignment pins and a Cu spring clip along one edge of the wafer to align the detector OMTs to the CE7 waveguide, and 2) three Be-Cu tripod spring clips to hold the Si wafer onto the baseplate.
Au wire bonds provide heat-sinking from the wafer to the baseplate, and Al bonding connects the detectors to the readout circuit (RC).
The RC consists of a printed circuit board (PCB), 8 shunt chips, 8 NIST-provided multiplexing (MUX) chips housing SQUIDs (Superconducting Quantum Interference Devices) for time-division multiplexed (TDM) readout, an Al flex circuit, Al wire bonds, and woven NbTi cable assemblies provided by Tekdata Interconnections, Ltd.\footnote{\url{https://www.tekdata-interconnect.com/}}
The RC is sandwiched between two Nb sheets, which provide magnetic shielding for the SQUIDs, and is heat-sunk with a Cu layer.
For a full description of the assembly stack, see Ref.~\citenum{Dahal-W1}.
\section{Detector Characterization}
\label{sec:detector_characterization}
In-lab characterization of the CLASS detectors is performed at JHU.
The CLASS detectors are mounted onto the 100~mK stage of a pulse tube pre-cooled dilution refrigerator \cite{Iuliano-receiver}.
The cryostat reaches an operational bath temperature ($T_{\mathrm{bath}}$) of $\sim$50~mK.
NbTi Tekdata wiring connects the feedback and bias lines from the focal plane to the SQUID Series Array (SSA) board at 4~K; the SSA board amplifies the signal to be read out at room temperature by the TDM Multi-Channel Electronics (MCE).\cite{TDM}
For estimation of electrothermal parameters and dark noise-equivalent power (NEP) was conducted in a dark configuration, with all stages of the cryostat closed with metal plates.
For bandpass measurements, the cryostat was placed in a necked-down optical configuration to prevent detector saturation, with a nylon filter on the 1~K stage, a small aperture plate with an anti-reflection coated nylon filter at the 4~K stage, and a small aperture plate with a polytetrafluoroethylene (PTFE) filter at the 60~K stage.
An ultra-high-molecular-weight polyethylene (UHMWPE) vacuum window lets in light at the front of the cryostat, while extruded polystyrene foam (XPS) filters at the 4~K stage rejects out-of-band infrared radiation.
Further in-lab characterization and on-sky performance of the new 90~GHz detectors will be discussed in a future publication.
\subsection{Electrothermal Parameters}
\label{subsec:electrothermal_parameters}
\begin{figure} [ht]
\begin{center}
\includegraphics[width=\textwidth]{IV_inset_with_Tb_legend_cbar.pdf}
\end{center}
\caption{ \label{fig:IVs}
Example I-V curves taken at multiple bath temperatures ($T_{\mathrm{bath}}$) for one detector. The inverse slope of the normal branch of the I-V yields the normal resistance ($R_\mathrm{N}$) of the TES. The saturation power ($P_{\mathrm{sat}}$) is defined to be $I_{\mathrm{TES}}\times V_{\mathrm{TES}}$ at 70\%~$R_\mathrm{N}$. The inset figure shows the measured $P_{\mathrm{sat}}$ at each $T_{\mathrm{bath}}$, from which the TES critical temperature, $T_\mathrm{c}$, and $\kappa$ can be determined.
We assume $n=4$ (ballistic phonon limit).\cite{Appel-calibration, Dahal-W1, TES-chapter}}
\end{figure}
I-V curves are used to determine the electrothermal characteristics of the TES bolometers. Over a wide range ($\sim$60-210~mK) of bath temperatures ($T_{\mathrm{bath}}$), we measure each I-V curve by ramping down the voltage bias in steps, driving the TES from normal to superconducting. We convert from the MCE digital-to-analog (DAC) feedback units and voltage bias to TES current ($I_{\mathrm{TES}}$) and TES voltage ($V_{\mathrm{TES}}$) following \S~4.1 of Ref.~\citenum{Appel-calibration}.
Figure~\ref{fig:IVs} shows an example of this process for one detector.
The inverse slope of the normal branch of the I-V yields the normal resistance ($R_\mathrm{N}$) of the TES.
At each $T_{\mathrm{bath}}$, we measure the saturation power $P_{\mathrm{sat}}$, which is the amount of power, given by $I_{\mathrm{TES}}\times V_{\mathrm{TES}}$, required to maintain the TES at its superconducting critical temperature ($T_\mathrm{c}$) with a resistance equal to 70\% of the normal resistance $R_\mathrm{N}$\cite{Appel-calibration}. We can then solve for $T_\mathrm{c}$ and $\kappa$ using the relation\cite{TES-chapter}:
\begin{equation}
P_{\mathrm{sat}} = \kappa \left(T_\mathrm{c}^n - T_{\mathrm{bath}}^n\right) \mathrm{.}
\end{equation}
\noindent The normalization prefactor $\kappa$ is related to the geometry of the stubby ballistic beam\cite{Ballistic} shown in Figure~\ref{fig:design}. We assume $n=4$ (ballistic phonon limit). The thermal conductance ($G$) of the TES, also associated with the stubby ballistic beam, is given by\cite{TES-chapter}:
\begin{equation}
G = \frac{dP}{dT}\Bigr|_{\substack{T_\mathrm{c}}} = n \kappa T_\mathrm{c}^{n-1} \mathrm{.}
\end{equation}
\begin{table}[ht]
\caption{Median (standard deviation) values for each wafer of key electrothermal parameters derived from I-V curves. We report values for approximately 45, 51, 47, and 32 bolometers for wafers 1--4, respectively.}
\label{tab:parameters}
\begin{center}
\begin{tabular}{|l|c|c|c|c|}
\hline
\rule[-1ex]{0pt}{3.5ex} & Wafer 1 & Wafer 2 & Wafer 3 & Wafer 4\\
\hline
\rule[-1ex]{0pt}{3.5ex} $G$~(pW/$K$) @ $T_\mathrm{c}$ & 269 (63) & 229 (30) & 257 (53) & 304 (56)\\
\hline
\rule[-1ex]{0pt}{3.5ex} $\kappa$~(nW/$\mathrm{K^4}$) & 16.1 (3.0) & 16.9 (1.0) & 13.9 (3.0) & 15.2 (0.9)\\
\hline
\rule[-1ex]{0pt}{3.5ex} $P_{\mathrm{sat}}$~(pW) @ 50~mK & 10.7 (2.9) & 8.4 (1.5) & 10.7 (2.6) & 12.7 (3.1)\\
\hline
\rule[-1ex]{0pt}{3.5ex} $R_\mathrm{N}$~($\mathrm{m\Omega}$) & 12.7 (0.5) & 11.0 (0.5) & 10.3 (0.4) & 10.7 (0.3) \\
\hline
\rule[-1ex]{0pt}{3.5ex} $T_\mathrm{c}$~(mK) & 163 (9) & 151 (8) & 165 (10) & 171 (10)\\
\hline
\end{tabular}
\end{center}
\end{table}
In Table~\ref{tab:parameters}, we report the median derived parameters and their standard deviations across each of the four new 90~GHz wafers. The reported values are calculated using approximately 45, 51, 47, and 32 bolometers, respectively. Using the fitted $T_\mathrm{c}$ and $\kappa$ values, we report the calculated $P_{\mathrm{sat}}$ at our operating $T_{\mathrm{bath}}$ of $\sim$~50~mK. Figure~\ref{fig:Tcwaferplot} shows the measured values of $T_\mathrm{c}$ at each detector pixel.
\begin{figure} [ht]
\begin{center}
\includegraphics[width=\textwidth,trim={1.25cm 0 0 0}]{Tc_waferplots_pm45.pdf}
\end{center}
\caption{ \label{fig:Tcwaferplot}
Measured values of $T_\mathrm{c}$ at each detector pixel. $T_\mathrm{c}$ values shown are for 175 bolometers that produced good I-V measurements during in-lab dark testing at JHU. The circles are split to show both orthogonal polarization states ($\pm45\degree$) contained within a pixel. Median values for each wafer are shown in Table~\ref{tab:parameters}.}
\end{figure}
\subsection{Bandpass Measurements}
\label{subsec:bandpass_measurements}
Bandpass measurements were performed using a Fourier transform spectrometer (FTS), which was built at JHU.\cite{FTS}
The FTS is a Martin–Puplett interferometer with a movable mirror that scans back and forth with a range of 150~mm at 0.5~mm/s.
The input signal of the FTS is a wide-band thermal source, at about $1000\degree\mathrm{C}$.
The output signal of the FTS, measured by the detectors, is modulated with a chopper at 21~Hz.
The FTS has a resolution of $\sim$1.6~GHz.
The bandpass is given by the real component of the Fourier transform of the interferogram.
The measured response plotted in Figure~\ref{fig:bandpass} is the inverse variance-weighted average over 93 bolometers that yielded high quality interferograms and bandpasses across three of the four new 90~GHz wafers.
The measured bandpass edges are in good agreement with the simulations; the discrepancies observed in-band between measured and simulated responses are likely due to optical effects related to the test setup that are not included in the simulation.
In Table~\ref{tab:bandpass}, we report the measured bandwidths and effective center frequencies associated with various types of diffuse sources following the methods described in \S~3.1 of Ref.~\citenum{Dahal-multifrequency}.
Uncertainties for the bandwidths (effective center frequencies) are given by the summation in quadrature of the standard errors on the mean and the FTS measurement resolution (half FTS measurement resolution).
\newcolumntype{P}[1]{>{\centering\arraybackslash}m{#1}}
\begin{minipage}{\linewidth}
\begin{minipage}{0.35\linewidth}
\begin{figure}[H]
\includegraphics[width=\linewidth, trim={1.25cm 0 0 0}]{JHU_FTS_log_nocorr.pdf}
\caption{\label{fig:bandpass}Simulated and inverse-variance weighted average measured spectral response (normalized to unity) for 93 bolometers across three wafers. The bandpass edges are in agreement with simulations.}
\end{figure}
\end{minipage}%
\hspace{0.4cm}
\begin{minipage}{0.59\linewidth}
\begin{table}[H]
\caption{Measured Bandwidths and Effective Center Frequencies for various types of diffuse sources.}
\label{tab:bandpass}
\begin{center}
\begin{table}[H]
\begin{tabular}{cP{1.6cm}l|c|P{2.5cm}|}
\cline{1-2} \cline{4-5}
\multicolumn{1}{|c|}{\rule[-1ex]{0pt}{3.5ex}} & \multicolumn{1}{P{1.6cm}|}{Bandpass (GHz)} & & Source & Effective Center Frequencies (GHz) \\ \cline{1-2} \cline{4-5}
\multicolumn{1}{|c|}{\rule[-1ex]{0pt}{3.5ex}FWHP} & \multicolumn{1}{P{1.6cm}|}{21.5±1.8} & & Synchrotron & 88.9$\pm$0.92 \\ \cline{1-2} \cline{4-5}
\multicolumn{1}{|c|}{\rule[-1ex]{0pt}{3.5ex}Dicke} & \multicolumn{1}{P{1.6cm}|}{32.0±1.7} & & Rayleigh–Jeans & 91.2$\pm$0.92 \\ \cline{1-2} \cline{4-5}
\rule[-1ex]{0pt}{3.5ex} & & & Dust & 92.7$\pm$0.92 \\ \cline{4-5}
\rule[-1ex]{0pt}{3.5ex} & & & CMB & 90.8$\pm$0.92 \\ \cline{4-5}
\end{tabular}
\end{table}
\end{center}
\end{table}
\end{minipage}%
\end{minipage}
\subsection{Noise-Equivalent Power}
\label{subsec:noise_performance}
We estimate the noise performance of the new detectors by measuring the NEP from dark tests for all four new 90~GHz modules.
The resulting noise spectra for 106 TES bolometers are shown in Figure~\ref{fig:NEP}.
We find a median NEP of 12.3 $\mathrm{aW\sqrt{s}}$ in the 8--12~Hz window about the CLASS modulation frequency of 10~Hz above the $1/f$ instrumental and atmospheric noise.\cite{Katie-VPM}
The observed roll-off at high frequencies is due to the MCE digital Butterworth filter that is applied to suppress noise aliasing from higher frequencies.
We anticipate the CLASS 90~GHz detectors to be photon noise limited.
The measured dark NEP is below the expected photon NEP of 32 $\mathrm{aW\sqrt{s}}$ in the field due to emission from the CMB, the atmosphere, and the instrument.
\begin{figure} [ht]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=\linewidth]{NEP.pdf}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:NEP}
Measured dark NEP for 106 TES bolometers.
Individual detectors are plotted, with line color (purple, blue, green, yellow) corresponding to each of the four wafers. The gray vertical line shows the modulated CLASS signal band.\cite{Katie-VPM} The dashed horizontal line shows the expected photon NEP at the CLASS site due to the CMB as well as instrumental and atmospheric emission.
The measured detector NEP during dark tests is below the expected NEP of 32 $\mathrm{aW\sqrt{s}}$ from photon noise in the field; therefore we expect the new 90~GHz detectors to be photon noise limited.}
\end{figure}
\section{Conclusion}
\label{sec:conclusion}
We present preliminary characterization, via electrothermal parameters, bandpass measurements, and dark noise measurements, of four new CLASS 90~GHz detector wafers.
The wafers include an updated TES architecture that aims to improve the stability and optical efficiency of the detectors.
These detectors have been installed in the field, upgrading four modules of the W1 focal plane.
The new detectors are anticipated to achieve first light in 2022.
Further in-lab and on-sky characterization will be presented in a future publication.
\acknowledgments
We acknowledge the National Science Foundation Division of Astronomical Sciences for their support of CLASS under Grant Numbers 0959349, 1429236, 1636634, 1654494, 2034400, and 2109311.
We thank Johns Hopkins University President R. Daniels and the Krieger School of Arts and Sciences Deans for their steadfast support of CLASS.
We further acknowledge the very generous support of Jim and Heather Murren (JHU A\&S '88), Matthew Polk (JHU A\&S Physics BS '71), David Nicholson, and Michael Bloomberg (JHU Engineering '64).
The CLASS project employs detector technology developed in collaboration between JHU and Goddard Space Flight Center under several previous and ongoing NASA grants. Detector development work at JHU was funded by NASA cooperative agreement 80NSSC19M0005.
Kyle Helson is supported by NASA under award number 80GSFC17M0002.
Zhilei Xu is supported by the Gordon and Betty Moore Foundation through grant GBMF5215 to the Massachusetts Institute of Technology.
We acknowledge scientific and engineering contributions from Max Abitbol, Fletcher Boone, David Carcamo, Lance Corbett, Ted Grunberg, Saianeesh Haridas, Jake Hassan, Connor Henley, Ben Keller, Lindsay Lowry, Nick Mehrle, Sasha Novak, Diva Parekh, Isu Ravi, Gary Rhodes, Daniel Swartz, Bingjie Wang, Qinan Wang, Tiffany Wei, Zi\'ang Yan, and Zhuo Zhang. We thank Miguel Angel, Jill Hanson, William Deysher, Mar\'ia Jos\'e Amaral, and Chantal Boisvert for logistical support.
We acknowledge productive collaboration with Dean Carpenter and the JHU Physical Sciences Machine Shop team.
CLASS is located in the Parque Astron\'omico Atacama in northern Chile under the auspices of the Agencia Nacional de Investigaci\'on y Desarrollo (ANID).
|
\section{Introduction}
\IEEEPARstart{T}{he} recent advancements in artificial intelligence (AI), in particular, deep learning-driven vision algorithms, and microelectronics have made possible automated surveillance on internet of things (IoT) and edge devices. Generally, these surveillance systems comprise of multiple interconnected cameras deployed in public places, such as big organizations, offices, roads, shopping malls, hospitals, and airports, to monitor humans and recognize their actions and behavior in video streams from multiple cameras. The primary objective behind the deployment of surveillance systems in the aforementioned places is to instantly detect abnormalities by recognizing the anomalous human behavior or activity in a video stream that can be harmful to the public. Human activity recognition is a process to analyze the hidden sequential pattern and predict the status of activity based on the perceptual context in input video stream. Generally, in videos, human activity is a combination of different movements of human body parts (i.e., hands, legs, or combination of both). For instance, running involves rapid movement of hands and legs, similarly, throwing object involves the backward and forward force of arm and hand. Human activity recognition has numerous potential applications, such as in smart surveillance systems \cite{huang2021abnormal}, video summarization \cite{sahu2021together}, content-based video retrieval \cite{qi2021semantics}, and human computer interaction \cite{ng2021multi}. In video, each frame contributes spatial information in sequential order which forms a sequential pattern containing human activity, that cannot be recognized in a single video frame. Considering throwing the knife and dart (that includes forward and backward force of arm and hand) have the same action pose in the starting frame; the discrimination of these two distinct activities becomes challenging while recognizing it with a single frame. Investigating the same movements of arm and hand in succeeding frames together with the information from previous frames will have the ability to effectively recognize human activities in video stream data.
The earlier developed methods in initial research for vision-based activity recognition are exclusively focused on activities performed by a single person/actor in simple and controlled environment. However, the current research focuses on more challenging and realistic human activities recorded with clutter complex background, variation in viewpoint, occlusion in background, inter- and intra-class variations, and pose variations. More categorically, the existing vision-based human activity recognition methods can be categorized into two classes namely: (i) traditional handcrafted features-based, and (ii) deep learning-based human activity recognition methods. The traditional handcrafted features-based methods \cite{asghari2020online,ehatisham2019robust,naveed2019human,franco2020multimodal,elmadany2018information} use manually designed handcrafted or hand-engineered features (requires extensive human efforts with prior knowledge of scene understanding) followed by statistical machine learning models to recognize the activity. For instance, several traditional image features have been utilized to analyze videos, such as histogram of 3D oriented gradients (HOG3D), histogram optical flow (HOF) \cite{dileep2021anomalous}, motion boundary histograms (MBH) \cite{yenduri2022fine}, and extended speeded up robust features (SURF) feature descriptors. The hand-engineered features are required to be designed specifically for each particular environment based on scene perceptual complexity. Such type of manually designed handcrafted features-based methods fail to perform when using in an environment other than that for what these methods were designed. Recently, deep learning-based methods have made incredible breakthroughs in various domain of image processing and computer vision, and have been actively used for human activity recognition problem \cite{luvizon2020multi,li2020spatio,ghose2020autofoley,lu2019gaim,liu2020multi}. These deep learning-based methods have obtained state-of-the-art performance by extracting deep progressive discriminative features using different convolutional neural network (CNN) kernels and exploiting gradient learning strategy. Unlike, traditional handcrafted features of an image, deep CNNs learn progressively strong features (containing low-level, mid-level, and high-level features) that help to keep track of all type of visual semantics in image data.
Deep learning-based methods have enhanced the activity recognition solutions in two perspectives. First, CNNs have the ability to extract more generic and semantically rich features than that of traditional handcrafted feature descriptors. Due to this generic feature extraction enabled by CNNs, CNNs have proliferated in a variety of complex computer vision tasks including 3D image reconstruction \cite{hu20213dbodynet}, image and video captioning \cite{yan2021task}, and text-to-image generation \cite{xia2021tedigan} that cannot be accomplished using traditional handcrafted features-based methods due to their limitations in terms of features and learning strategies. Secondly, deep learning offers efficient architectures called recurrent neural networks (RNNs) which have the ability to learn the representation of human activity from a bunch of frames (sequence of frames or temporal representation of human activity) rather than a single frame. Earlier traditional methods consider frame-level classification of human activity in videos, rather than understating the activity in sequence of frames that greatly limits their performance for complex and multi-person activities. To cope with this challenge, deep learning-based methods have adopted RNNs for the better understanding and recognition of complex human activities in videos. Normally, in deep learning based-methods, RNNs are placed right after CNNs, where the CNN architecture is responsible for extracting deep discriminative features from videos and the RNN is responsible for learning the hidden sequential patterns in the extracted CNN features. The performance of these deep learning methods is good compared to traditional methods; however, these methods are computationally very expensive due to their hybrid and complex CNN and RNN architectures.
The above-mentioned deep learning-based activity recognition methods have attained exceptional performance. Most of the existing AI-assisted activity recognition methods have adopted large yet effective pre-trained CNN architectures trained on a large-scale image dataset having tens of millions of trained parameters. Fusing such a computationally expensive feature descriptor backbone architecture with long short-term memory (LSTM) networks or multi-layer LSTMs (LSTMs having several layers with same settings) greatly increases the computational complexity of the overall method thereby compromising on the better tradeoff between model accuracy and complexity. Considering the demand for computationally-efficient yet effective approaches that provide a balanced tradeoff between model accuracy and complexity for deployment on resource-constrained IoT and/or edge devices, in this paper, we propose a deep learning based computationally efficient yet effective method for activity recognition problem that can be deployed even on resource-constrained edge devices in the IoT-enabled surveillance environment. Our main contributions in this work are as follows:
\begin{enumerate}
\item We propose a computationally efficient cascaded spatial-temporal learning approach for human activity recognition. The proposed system utilizes deep discriminative RGB features guided by channel-spatial attention mechanism and long-term modeling of action centric features for reliable recognition of human activities in video streams.
\item We propose a light-weight CNN architecture having a total of 8 convolutional layers where the maximum number of kernels used per layer is 64 with spatial dimension of $3 \times 3$. With these constrained settings, we have developed a compact yet efficient CNN architecture for deep discriminative feature extraction as opposed to complex deep CNNs utilized by other contemporary works in their activity recognition models using transfer learning.
\item We design a stacked dual channel-spatial attention mechanism with residual skip connection for spatial saliency extraction from video frames. The developed dual attentional module is placed after each two-consecutive convolutional layers of the developed CNN model which helps our network to extract saliency-aware deep discriminative features for localizing the action-specific regions in video frames.
\item For efficient temporal modeling of long-term action sequences, we propose a bi-directional GRU network with three bi-directional layers (having forward and backward pass) that capture the temporal patterns of human actions in both forward and backward directions, which greatly enhances the reusability of features, improves the features propagation, and alleviates the issue of gradients vanishing.
\item We demonstrate the effectiveness and suitability of the proposed encapsulated dual attention CNN and bi-directional GRU framework (DA-CNN+Bi-GRU) for resource-constrained IoT and edge devices by comparing the model accuracy and execution/inference time of the proposed framework with various baseline methods as well as contemporary human action recognition methods.
\end{enumerate}
The remainder of this paper is organized as follows. Section ~\ref{sec:SectionRelatedWork} provides the brief overview on the related works covering different type of methods introduced for human activity recognition, till date. The proposed method and its technical component are discussed in detail in Section ~\ref{sec:proposedmethod}. In Section ~\ref{sec:experimentalresults}, we present extensive experimental evaluation of our method based on different assessment strategies. Finally, we conclude the paper in Section ~\ref{sec:conclusion} with possible future research directions.
\section{Related Works on Human Activity Recognition} \label{sec:SectionRelatedWork}
In recent years, human action and activity recognition have been widely studied and have got exceptional attention of computer vision researchers due to the recent success of deep learning for image classification and object detection task. Comprehensive review on both traditional and deep learning-based methods have been presented in numerous surveys \cite{pareek2021survey,kong2022human}. The reported literature on human action and activity recognition can be summarized in terms of handcrafted features-based methods, deep learning features-based methods, long-term temporal modeling-based methods, and attention models-based methods. This section presents a brief discussion on these representative methods and a brief summary of previous related works.
\textbf{\textit{Handcrafted features-based methods}} have been used to localize the spatial and temporal variations in videos using manually hand-engineered feature descriptors. Generally, these handcrafted features-based methods can be structured as a feature extraction and encoding pipeline having three phases including key features point detection (spatial and temporal feature points), quantization of detected features, and features encoding. The first phase involves the extraction of spatial-temporal features from video frames, followed by feature quantization in the second phase that quantize local motion-centric features. Lastly, the quantized spatial-temporal features are then encoded into feature vectors (known as action feature vectors) having fixed dimensions. For instance, inspired by the features extraction mechanism of the scale-invariant feature transform (SIFT) descriptor, Scovanner et al. \cite{scovanner20073} have adopted the SIFT algorithm features extraction strategy and extend their features space from 2D to 3D for encoding hidden action patterns. As single feature representation is not capable to capture human actions, therefore numerous multi-feature representative descriptors have been proposed in the literature. Laptev et al. \cite{laptev2008learning} have proposed a multiscale spatial-temporal features-based approach by utilizing space-time extension and Harris operator. They first extract multi-scale spatial-temporal features from video frames and then characterize appearance and motion of local features using volumetric histogram of oriented gradient. The retrieved multi-scale spatial-temporal features are then fed to non-linear support vector machine (SVM) for action recognition. In \cite{ryoo2016first}, Ryoo and Matthies have inspected the behavior of local and global motion features for recognizing first-person activities in video data. Their proposed methods exclusively focus on temporal structures depicted in first person action/activity videos. These traditional handcrafted features-based methods have shown progressive improvement over the years by presenting more efficient approaches; however, these methods are time consuming (lacking end-to-end recognition strategy), labor-intensive (requiring extensive human efforts to extract generic and more discriminative features), and difficult to adopt in diverse scenarios.
\textbf{\textit{Deep learning features-based methods}} are the current mainstream methods to solve the problem of complex human action and activity recognition in videos. With the recent success in computer vision domain for high-level vision tasks including image enhancement \cite{ullah2021light}, image segmentation \cite{chen2022saliency}, and video captioning \cite{aafaq2022dense}, CNNs have been actively investigated for human action and activity recognition problem. Where numerous CNN-assisted methods have been presented \cite{karpathy2014large,simonyan2014two,wang2015towards,yue2015beyond,wu2015modeling,wang2016actions,feichtenhofer2016convolutional}, having deep CNN architectures with 2D convolution kernels applied across convolutional layers of the CNN network. These convolutional layers extract deep discriminative spatial features with translation invariance from action video frames, offering reasonable action recognition performance without using temporal modeling. For instance, Karpathy et al. \cite{karpathy2014large} have presented a single-stage CNN architecture for action recognition, where they have trained their proposed model on a large-scale sport video datasets benchmark namely Sports-1M dataset. Although, their method acheives better results than traditional handcrafted features-based methods, the presented architecture is unable to cope with temporal modeling. To overcome this issue, several two-streams CNN architectures have been introduced \cite{simonyan2014two,wang2015towards,feichtenhofer2016convolutional} to obtain both spatial and temporal modeling of human action where one architecture performs spatial modeling of spatial contextual features and the second architecture performs temporal modeling using extracted optical flow features. The addition of second network improves the performance by introducing temporal modeling to CNN-based action recognition approach; however, it equally increases the computational complexity of the overall two-stream CNN approach. To achieve spatial modeling and temporal cues within a single CNN architecture without compromising on model complexity, 3D CNNs \cite{tran2015learning,varol2017long,ji20123d} have been introduced for human action recognition task. For instance, Tran et al. \cite{tran2015learning} exploits the powerful characteristics of 3D CNN to recognize human action in sports videos, where they have trained their proposed architecture on large-scale benchmark dataset and have shown promising results. However, these 3D CNN-based approaches work well with short-term temporal modeling and lack the ability to cope with long temporal modeling.
\textbf{\textit{Temporal modeling-based methods}} have been actively presented to overcome the issue of long-term temporal modeling, where researchers have introduced a special kind of neural network called RNN, which has the ability to deal with the long-term sequences. Later, different variants of RNNs are introduced for action recognition problem including LSTM \cite{ullah2018activity}, bi-directional LSTM \cite{he2021db}, and GRU \cite{sun2022capsganet}, which are comparatively more efficient than RNNs in terms of memorizing contents for long period of time. For instance, Yue et al. \cite{yue2015beyond} have presented a two-stream CNN architecture to extract both spatial (edge, color and shape) features and temporal (optical flow) features stacked with LSTM model for temporal modeling of human activity. Similarly, Amin et al. \cite{ullah2018activity} have presented a two-stream CNN architecture followed by a multi-layer LSTM to recognize human activity in videos. They have first extracted spatial salient and optical flow features and then fed the extracted features to multi-layer LSTM for localizing human action in video sequences. Ibrahim et al. \cite{ibrahim2016hierarchical} have proposed a two-stream temporal modeling-based activity recognition framework to recognize a team or group of activities. Their proposed method consist of two LSTM networks, the first LSTM learn the representation of a single person action, whereas the second LSTM is responsible to understand collective activity by aggregating individual actions in a sequence of frames. Biswas et al. \cite{biswas2018structural} have presented a special variant of RNN named structural RNN for group activity recognition. Their proposed method consists of series of interconnected RNNs structured to analyze human actions and their mutual interactions in video sequences. To accurately learn the representation of human activity in feature-encoded video frames, Shugao et al. \cite{ullah2021efficient} have reformulated ranking loss to efficiently discriminate human activities. They have first extracted deep discriminative CNN features from video frames using VGG19, which are then fed into LSTM for analyzing hidden sequential patterns and recognition of human activities. Muhammad et al. \cite{muhammad2021ai} have presented a spatio-termporal approach for recognizing salient events in soccer videos, where they have used a pretrained ResNet50 architecture for deep features extraction and a multilayer LSTM for events recognition from the hidden sequential patterns. Although, these hybrid CNN+LSTM have shown significant performance for vision-based human action and activity recognition task, these methods are computationally complex due to intensive computation cause by CNN features extraction and human action modeling by LSTM.
\textbf{\textit{Attention mechanism-based methods}} have demonstrated great potentials for a variety of high-level vision tasks including image segmentation \cite{li2021abssnet}, video captioning \cite{deng2021syntax}, and visual questioning answering (VQA) \cite{yang2016stacked}. More recently, the attention mechanism combined with CNN and RNN networks have been widely used for human action recognition task and have achieved noticeable improvements in action recognition performance. For instance, Baradel et al. \cite{baradel2017human} have proposed a spatio-temporal attention-based approach for human action recognition, where they have exclusively focused on the tracking of human hands that helps to detect the discriminative segments of action in a video. They have used attention in recurrent style where they have embedded an attention mechanism in RNN network to efficiently model human actions. Islam et al. \cite{islam2021multi} have presented a multi-model graphical attention network for human action recognition which learn multi-model discriminative features. They have captured cross-modal relation using multi-model discriminative features extracted using message passing-based graphical attention mechanism. Long et al. \cite{long2018multimodal} have presented a method called keyless attention mechanism to effectively extract salient features, which are then fused with other extracted features to design multi-model features for human action recognition in videos. Song et al. \cite{song2018spatio} have presented a spatio-temporal attention model to examine spatial and temporal deep discriminative features for human action recognition in videos. They have used an LSTM network equipped with attention modules which is capable to exclusively focus on discriminative joints, and have applied multi-level attention on the joint-specific location. Moreover, they have also proposed a technique to model temporal action proposal for efficient action detection. Cho et al. \cite{cho2020self} have proposed a self-attention network for human action recognition which comprised of three different variants of self-attention network (SAN) named, SAN-V1, SAN-V2, and SAN-V3. Their developed models have the potential to extract high-level features by exploiting low-level correlation. Along with SAN model, they have also developed a temporal segment network (TSN) which greatly improved the overall action recognition performance of their proposed method. Although, these attention-driven methods have been widely used for human action recognition task and have obtained noticeable improvements over handcrafted features-based methods and other non-attention deep learning methods, these methods perform well only on clean red, green, and blue (RGB) video data and mostly fail while dealing with noisy color (RGB) video data.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{Figures/Updated_HAR_Framework.png}
\caption{The graphical overview of our proposed activity recognition framework. The proposed framework consists of two main modules: an attention-driven CNN followed by a bi-directional GRU network. The CNN module utilizes a dual-attention mechanism to effectively extract salient CNN features from video frames, whereas the bi-directional GRU network is used to learn the activity representation for hidden sequential patterns.}
\label{Fig:framework}
\end{figure*}
\section{Proposed Human Activity Recognition Framework} \label{sec:proposedmethod}
This section presents the detailed insights of our proposed human action recognition framework and its core components. For better understanding, the proposed approach is divided into three distinct modules, where each module is separately discussed. The first core component of our method is the newly introduced lightweight CNN architecture having a small number of trainable parameters. The second core component is a dual attention (channel and spatial attention) module, which is used to embed dual attention mechanism to the CNN module to enable our CNN model to extract salient features from video frames. The last key component of our framework is a bi-directional GRU network for learning long-term encoded patterns of human actions. The conceptual workflow of our proposed method is depicted in Figure 1.
\subsection{Overview of Proposed CNN Architecture}
Recognizing human actions in video data is indeed a challenging problem, where video data represent complex human actions over a series of frames in the form of different hidden visual contents that include temporal flow of objects in frames, varying texture, object-specific edges and colors. For better representation and modeling of human actions, these visual contents need to be analyzed effectively, which allows us to recognize the complex human actions or activity in video sequences. To effectively extract the defining visual features of these hidden action contents, CNN-based approaches are widely used to recognize human actions in videos. Although, the presented CNN-based approaches have shown remarkable performance, their computational complexity and execution/inference times are very high due to large network architectures. To avoid such high computational complexity and long runtime, we propose a tiny CNN architecture coupled with channel and spatial attention. The proposed CNN architecture contains a total of eight convolutional layers, where each two consecutive convolutional layers are followed by a max pooling layer and a dual attention block (containing both channel and spatial attention). The first two convolutional layers each apply 16 kernels on input video frames with the kernel size of $3 \times 3$, whereas the third and fourth convolutional layers each apply 32 kernels on the output of the first dual attention block with the kernel size of $3 \times 3$. Similarly, the fifth and sixth convolutional layers each apply 32 kernels on the output of the second dual attention block with the kernel size of $3 \times 3$. The last pair of the convolutional layers each apply 64 kernels on the output of the third dual attention block with the kernel size of $3 \times 3$ and then forward the estimated feature maps to the last dual attention block. The output of the last dual attention block is processed by a global average pooling layer, the output of which is then flattened by a flatten layer. The output of the flatten layer is fused with bi-directional GRU network for later long short-term sequence learning. The architectural details of our proposed CNN architecture are listed in Table \ref{tab:cnn}. It is worth noticing that we have used at most 64 convolutional kernels per layer and a fixed $3 \times 3$ kernel size that greatly help to reduce the computational complexity as low as possible with a negligible effect on model performance.
\begin{table}[t]
\centering
\caption{Architectural details of our proposed CNN architecture.}
\resizebox{\columnwidth}{!}{
\begin{tabular}{c|c|c|c|c|c|c}\hline
Layer &Input channels &Number of kernels &Kernel size &Stride & Padding & Output channels \\ \hline
Conv 1 & 3 & 16 & $3 \times 3$ & 1 &1&16\\
Conv 2 & 16 & 16 & $3 \times3$ & 1 &1&16\\
\hline \multicolumn{7}{c}{\makecell{Max pooling}}\\
\hline \multicolumn{7}{c}{\makecell{\color{green}Channel Attention}}\\
\hline \multicolumn{7}{c}{\makecell{\color{orange}Spatial Attention}}\\
\hline
Conv 3 & 32 & 32 & $3 \times 3$ & 1 &1&32\\
Conv 4 & 32 & 32 & $3 \times 3$ & 1 &1&32\\
\hline \multicolumn{7}{c}{\makecell{Max pooling}}\\
\hline \multicolumn{7}{c}{\makecell{\color{green}Channel Attention}} \\
\hline \multicolumn{7}{c}{\makecell{\color{orange}Spatial Attention}}\\
\hline
Conv 5 & 32 & 32 & 3$\times$3& 1 &1&32\\
Conv 6 & 32 & 32 & 3$\times$3& 1 &1&32\\
\hline \multicolumn{7}{c}{\makecell{Max pooling}}\\
\hline \multicolumn{7}{c}{\makecell{\color{green}Channel Attention}} \\
\hline \multicolumn{7}{c}{\makecell{\color{orange}Spatial Attention}}\\
\hline
Conv 7 & 32 & 64 & $3 \times 3$ & 1 &1&64\\
Conv 8 & 32 & 64 & $3 \times 3$ & 1 &1&64\\
\hline \multicolumn{7}{c}{\makecell{Max pooling}}\\
\hline \multicolumn{7}{c}{\makecell{\color{green}Channel Attention}} \\
\hline \multicolumn{7}{c}{\makecell{\color{orange}Spatial Attention}}\\
\hline \multicolumn{7}{c}{\makecell{\color{amethyst}Global Average Pooling}}\\
\hline \multicolumn{7}{c}{\makecell{Flatten}}\\
\hline
\end{tabular}}
\label{tab:cnn}
\end{table}
\subsection{Dual Attention Module}
To exclusively focus on most salient regions of video frames, we propose an attention-driven CNN architecture to efficiently localize the salient regions and enhance feature representation. The proposed attention mechanism is formed by fusing spatial attention module with the output of channel attention module through element-wise product operation. The detailed graphical overview of the proposed dual attention block is depicted in Figure 2. The fusion of both channel and spatial attentions not only helps to reduce the overall parameters overhead, but also enables the proposed CNN architecture to extract salient features. Therefore, we construct the formation of network layers in such way, where we place a stacked dual attention module after each two consecutive convolutional layers of our proposed network. The channel attention module estimates the weighted contribution of RGB channels by applying intermediate channel attention $\mathcal{A}_C$ on the output feature maps $F_{M}$ of the previous convolutional layer to obtain the channel attention $Att_C$. The computed output from $Att_{C}$ is then forwarded to the spatial attention module, which localizes promising object-specific regions by applying spatial attention on the computed channel attention feature maps $Att_{C}$. Finally, we obtain the refined feature maps $F_{RM}$ by fusing the spatial attention feature maps $Att_S$ with the input feature maps $F_M$ via a residual skip connection using element-wise addition operation. Mathematically, $Att_{C}$, $Att_{S}$, and $F_{RM}$ can be formulated as follows:
\begin{equation}
Att_{C}^{H\times W \times C} = \mathcal{A}_C(F_{M}^{H \times W \times C}) \otimes F_{M}^{H \times W \times C},
\label{eq:equation1}
\end{equation}
\begin{equation}
Att_{S}^{H \times W \times C} = \mathcal{A}_S(Att_{C}^{H \times W \times C}) \otimes Att_{C}^{H \times W \times C},
\label{eq:equation2}
\end{equation}
\begin{equation}
F_{RM}^{H\times W\times C} = Att_{S}^{H \times W \times C} \oplus F_{M}^{H \times W \times C}
\label{eq:equation3}
\end{equation}
\begin{figure}[t]
\includegraphics[width=90mm ]{Figures/Channel_Spatial_attention.png}
\caption{The detailed building block of dual attention block containing channel and spatial attention mechanisms.}
\label{Fig:attention_block}
\end{figure}
Here, $\mathcal{A}_C$ and $\mathcal{A}_S$ are the intermediate channel attention and the intermediate spatial attention, respectively. $F_{RM}$ is the final refined feature maps obtained by fusing spatial attention and input feature maps $F_{M}$.
\subsubsection{Channel Attention}
In pattern recognition problems, particularly in image/object recognition, each color channel contributes differently based on the appearance of color in image. During training, a CNN model construct feature maps from input image data by extracting deep discriminative features over the number of convolution layers, where a particular channel contributes more than other channels in the recognition process. Unlike the earlier attention-based approaches that used either global max pooling layer or global average pooling layer, we have used both global max pooling and global average pooling to extract more effective features. The global max pooling emphasizes on highly activated values by selecting maximum value from the receptive field, where global average pooling estimates the equally weighted feature maps for each channel.
The computed feature maps are then forwarded to a shared multilayer perceptron (MLP) containing two fully connected layers namely fc1 and fc2 having 128 and 512 nodes, respectively. The shared MLP learns the non-linearity between the two fully connected layers using ReLU activation function, and outputs two individual feature vectors namely $V_{C-max}^{1 \times 1 \times C}$ and $V_{C-avg}^{1 \times 1 \times C}$ for global max pooling and global average pooling, respectively. The computed feature vectors are then combined via an element-wise addition operation, and then forwarded to a sigmoid activation function, which normalizes the feature values to obtain intermediate channel attention features $\mathcal{A}_{C}^{1 \times 1 \times C}$. The obtained intermediate channel attention features $\mathcal{A}_{C}^{1 \times 1 \times C}$ are then fused with the input features maps $F_{M}^{H \times W \times C}$ using a residual skip connection by performing element-wise multiplication operation, which results in the ultimate channel attention feature maps $Att_{C}^{H \times W \times C}$ as depicted in Figure 2. Mathematically, the channel attention and its components can be expressed as follows:
\begin{equation}
V_{C-max}^{1 \times 1 \times C} = fc2(R_{eLU}(fc1(maxpool(F_{M}^{H \times W \times C})))),
\label{eq:equation4}
\end{equation}
\begin{equation}
V_{C-avg}^{1 \times 1 \times C} = fc2(R_{eLU}(fc1(avgpool(F_{M}^{H \times W \times C})))),
\label{eq:equation5}
\end{equation}
\begin{equation}
\mathcal{A}_C^{1 \times 1 \times C} = \sigma(V_{C-max}^{1 \times 1 \times C} \oplus V_{C-avg}^{1 \times 1 \times C}),
\label{eq:equation6}
\end{equation}
\begin{equation}
Att_{C}^{H \times W \times C} = \mathcal{A}_C^{1 \times 1 \times C} \otimes F_{M}^{H \times W \times C},
\label{eq:equation7}
\end{equation}
Here, $V_{C-max}^{1 \times 1 \times C}$ and $V_{C-avg}^{1 \times 1 \times C}$ are the obtained feature vectors from global max pooling and global average pooling operations, respectively. In the above equations, $F_{M}^{H \times W \times C}$ represents the input feature maps, $\sigma$ denotes the sigmoid activation function, whereas $Att_{C}^{H \times W \times C}$ is the final channel attention output.
\subsubsection{Spatial Attention}
The spatial attention mechanism focuses on object saliency in the given feature maps by paying more attention to important features across each color channel and localizing salient regions. To highlight the salient object-specific regions in the feature maps, we exploit inter-spatial features and their relationship among channels, which greatly help to trace the target object in the feature maps. We compute the relation of inter-spatial features among channels by applying max pooling and average pooling to the input channel attention feature maps to obtain max-pooled channel attention $Att_{C-max}^{H \times W \times 1}$ and average-pooled channel attention $Att_{C-avg}^{H \times W \times 1}$, respectively.
The max-pooled channel attention $Att_{C-max}^{H \times W \times 1}$ and average-pooled channel attention $Att_{C-avg}^{H \times W \times 1}$ are concatenated and then forwarded to a single convolutional layer $Conv^{3 \times 3}$, which applies a $3 \times 3$ convolution kernel on pooled feature maps to form single-channel convoluted feature maps. These convoluted feature maps are then processed by a sigmoid activation function, which normalizes the learned features and produces intermediate spatial attention features $\mathcal{A}_{S}^{H \times W \times 1}$. Finally, the obtained intermediate spatial attention features $\mathcal{A}_{S}^{H \times W \times 1}$ are fused with the input channel attention feature maps $Att_{C}^{H \times W \times C}$ using a residual skip connection by preforming element-wise multiplication operation, which results in final spatial attention feature maps $Att_{S}^{H \times W \times C}$, as depicted in Figure 2. Mathematically, spatial attention $Att_{S}^{W \times H \times C}$ and its component can be expressed as follows:
\begin{equation}
Att_{C-max}^{H \times W \times 1} = maxpool(Att_{C}^{H \times W \times C}),
\label{eq:equation8}
\end{equation}
\begin{equation}
Att_{C-avg}^{H \times W \times1} = avgpool(Att_{C}^{H \times W \times C}),
\label{eq:equation9}
\end{equation}
\begin{equation}
\mathcal{A}_S^{H \times W \times 1} = \sigma(Conv^{3 \times 3}(Att_{C-max}^{H \times W \times1} \biguplus Att_{C-avg}^{H \times W \times1})) ,
\label{eq:equation10}
\end{equation}
\begin{equation}
Att_{S}^{H \times W \times C} = \mathcal{A}_S^{H \times W \times 1} \otimes Att_{C}^{H \times W \times C} ,
\label{eq:equation11}
\end{equation}
where $Att_{C-max}^{H \times W \times 1}$ and $Att_{C-avg}^{H \times W \times 1}$ are the global max and average pooled features, respectively. $\sigma$ is the sigmoid activation function and $\biguplus$ represents the concatenation operation. $Att_{S}^{H \times W \times C}$ is the final obtained spatial attention. The representative saliency maps of different human actions generated by our proposed method are depicted in Figure 3.
\begin{figure}[t]
\includegraphics[width=90mm]{Figures/Attention_maps.png}
\caption{Visual representation of the computed salient object-specific regions by our dual attention mechanism.}
\label{Fig:vis_repre}
\end{figure}
\subsection{Learning Human Action Patterns via Bi-Directional GRU}
As videos are nothing but a stack of frames covering sequential flow of varying visual contents over a specific time interval. To understand the visual contents, main-stream computer vision approaches first extract deep discriminative features from the video frames using CNNs and then combine the extracted features in sequential order to maintain the semantic flow of the video. Second, the feature-encoded videos are then processed by RNNs to learn the representation of visual contents from hidden sequential patterns. Specifically, for human activity recognition problem, two special variants of RNNs are actively used by researchers that include LSTMs and GRUs. The LSTM unit comprises of different gates including input, output, forget gates and other memory components whereas the GRU unit contains an update gate, a reset gate, and an activation function. The LSTM is comparatively more complex than the GRU in terms of the number and formation of gates which leads to relatively higher computational complexity requiring more computational resources. Therefore, in this paper we propose to use GRU with bi-directional flow of learning strategy, which effectively learns from the encoded hidden sequential pattern.
The bi-directional GRU consists of two layers namely forward and backward layer, where both layers process the same sequence in different sequential order. The forward layer reads the input sequence from left to right, that is, from $X_{t-1}$ to $X_{t+n}$ where $n$ is the length of sequence. On the other hand, the backward layer reads the input sequence in reverse order from right to left, that is, from $X_{t+n}$ to $X_{t-1}$ as shown in Figure 4. Both forward and backward GRU layers consist of GRU cells, where each cell consists of two gates namely a reset $r$ and an update gate $\mu$ with two activation functions that include sigmoid and tanh. The reset gate decides whether the GRU needs to forget or retain the portion of information based on its values (between 0 and 1). When the output value of reset gate is near to 0, the reset gate forgets the information from the previous portion of the sequence, whereas if the reset gate value is near to 1, the reset gate retains the previous portion of the sequence. The update gate decides the amount of information from the previous hidden state to be retained to the current hidden state based on its values (between 0 and 1). When the value of update gate is near to 0, the updated gate simply forgets the portion of information from the previous hidden state and retains the portion of information from the previous hidden state to current hidden state when the value is close to 1. Mathematically, the operation of these gates can be expressed as follows:
\begin{figure}[t]
\includegraphics[width=90mm]{Figures/GRU.png}
\caption{The building block of bi-directional single GRU layer.}
\label{Fig:GRU_block}
\end{figure}
\begin{equation}
r_{t} = \sigma(w_{r}\cdot x_{t} + u_{r}\cdot h_{t-1}),
\label{eq:equation12}
\end{equation}
\begin{equation}
\mu_{t} = \sigma(w_{\mu}\cdot x_{t} + u_{\mu}\cdot h_{t-1}),
\label{eq:equation13}
\end{equation}
\begin{equation}
\tilde{h}_{t} = \tanh(w\cdot x_{t} + r_{t} \cdot u\cdot h_{t-1}),
\label{eq:equation14}
\end{equation}
\begin{equation}
{h}_{t} = (1-\mu_{t}) \cdot h_{t-1} + \mu_{t} \cdot \tilde{h}_{t},
\label{eq:equation15}
\end{equation}
\begin{equation}
{y}_{t} = \sigma(w_{o} \cdot h_{t}),
\label{eq:equation16}
\end{equation}
where $r_{t}$ and $\mu_{t}$ represent the reset and update gates, respectively, having values between 0 and 1. In the above equations, \emph{w} and \emph{u} are the weight variables, $x_{t}$ is the input to the GRU layer, $w_{o}$ is the weight variable between input and output layer, ${y}_{t}$ represents the output layer node at time step t. $\tilde{h}_{t}$ is the candidate hidden state of the current node, $\emph{h}_{t}$ is the current hidden state, and $h_{t-1}$ is the hidden state of the previous node.
\color{black}
\section{Experimental Results and Discussion} \label{sec:experimentalresults}
In this section, we present detailed experimental evaluation of our proposed human activity recognition framework. We evaluate the effectiveness of our proposed framework by analyzing the performance with and without the key components (channel attention, spatial attention, bi-directional GRU) of our framework. First, we describe the implementation details and performance evaluation metrics that we have used in this research. Next, we briefly discussed the datasets we have used for benchmarking experiments. We then compare our proposed framework with state-of-the-art human action recognition methods across each experimented dataset. Finally, we present the human action recognition visualization and then conduct runtime analysis of our proposed approach for real-time human activity recognition.
\subsection{Implementation Details}
The proposed framework is implemented using a well-known deep learning framework called TensorFlow version 2.0 in Python language 3 on a computing system with Intel Xeon (R) processor with processor frequency of 3.50\,GHz and having 32\,GB of dedicated main memory. The computing system is also equipped with an NVIDIA GeForce GTX 1080 graphics processing unit (GPU) having a graphics random-access memory of 8\,GB. For training and validation, we have divided the datasets into a ratio of 70\% and 30\%, where for training we have used 70\% of the data and the remaining 30\% of the data is used for validation. The training process is run for 300 epochs and the weights are initialized with a random weight initializer, whereas the batch size is set to 16. To adjust weight values during training, we have used the Adam optimizer with static learning rate of 0.0001. Our proposed network utilizes categorical cross entropy loss, which controls the weight adjustment based on network prediction during training. For sequence learning, we have used a sequence length of 16 frames without overlapping for both forward and backward pass of bi-directional GRU, where we have used three bi-directional GRU layers with 32 GRU units per layer. Moreover, we have used two different performance evaluation metrics to assess the overall performance of our proposed method. The first metric is the \textit{accuracy} metric, which is used to evaluate the activity recognition performance of our framework and other contemporary methods. The second metric is \textit{frames per second (FPS)} or alternatively \textit{seconds per frame (SFP)}, which measures the runtime of our proposed framework and other contemporary methods.
\subsection{Datasets}
To verify the effectiveness of our proposed framework, we have conducted extensive experiments on three challenging human actions datasets that include YouTube action, UCF50, and HMDB51 datasets. Each dataset consists of multiple action videos having varying duration, different view points, and frames per seconds (FPS). These datasets are discussed in detail in the following subsections.
\subsubsection{YouTube Action Dataset}
The YouTube action dataset \cite{liu2009recognizing} is a commonly used action recognition dataset containing diverse sports and other action video clips collected from YouTube. The collected videos clips are very challenging due to variation in viewpoints, camera motion, cluttered background, and varying pose and appearance of objects in the scene. The dataset contains 1640 video clips categorized into 11 action categories, where the duration of videos range between 2 to 5 seconds having a frame rate of 29 FPS and a resolution of $320 \times 240$. The collected action clips in all action categories are grouped into 25 distinct groups containing 4 or more video clips, where each video clip in the same group share common visual features, such as background, viewpoint, and the person or actor.
\subsubsection{UCF50 Dataset}
The UCF50 dataset \cite{reddy2013recognizing} is one of the challenging large-scale human activity recognition datasets, containing videos of diverse human actions captured with varying viewpoints, camera motions, object poses and appearances, and background clutter. The dataset contains a total of 6,676 video clips categorized into 50 different classes, where the duration of video clips range between 2 to 3 seconds with a frame rate of 25\;FPS and a resolution of $320 \times 240$. The video clips in all 50 categories are further grouped into 25 groups, where each group comprises of at least 4 video clips, where a video clip in a single group share common features of actions, such as the same person performing an action, the same view point, and the same background.
\subsubsection{HMDB51 Dataset}
HMDB51 \cite{kuehne2011hmdb} is one of the challenging datasets commonly used for human action recognition in videos. The videos in this dataset are collected from difference sources including movies, public databases, YouTube, and Google videos. The dataset comprises a total of 6,849 action video clips categorized into 51 classes, where each class contains at least 101 video clips having duration of 2 to 3 seconds with a frame rate of 30\;FPS and a resolution of $320 \times 240$. The collected action video clips can be generally categorized into five different types of actions that include facial actions, facial actions with object manipulation, general body movements, body movements and interaction with objects, and body movements while interacting with humans.
\begin{figure*}[t]
\centering
\begin{subfigure}[t]{0.32\linewidth}
\centering
\includegraphics[width=\textwidth]{Figures/Training_history_UCF11.png}
\caption{\centering}
\end{subfigure}
\begin{subfigure}[t]{0.32\linewidth}
\centering
\includegraphics[width=\textwidth]{Figures/Training_History_UCF50.png}
\caption{\centering}
\end{subfigure}
\begin{subfigure}[t]{0.32\linewidth}
\centering
\includegraphics[width=\textwidth]{Figures/Training_History_hmdb51.png}
\caption{\centering}
\end{subfigure}
\caption{Training history of our proposed method along with other experimented baseline methods for 300 epochs over three benchmark action datasets: (a) Training history for YouTube action dataset, (b) Training history for UCF50 dataset, And (c) Training history for HMDB51 dataset.}
\label{Fig:TrainingHistory}
\end{figure*}
\subsection{Assessment of our Framework with Baseline Methods}
This research is built up on the exploration of various possible solutions for vision-based human action recognition, where we have developed several spatial-temporal methods, assessed their performances, and developed our final proposed method. To obtain the optimal approach, we have explored different spatial-temporal solutions and successively developed four different baseline methods that include CNN+LSTM, CNN+Bi-LSTM, CNN+GRU, and CNN+Bi-GRU, and we have analyzed their performances in terms of model precision. To obtain a fair comparison, we have trained each baseline methods on three different datasets (i.e., YouTube action, UCF50, and HMDB51 datasets). These datasets are then used for training our proposed framework. The detailed network settings of these baseline methods are listed in Table \ref{tab:table2}, where it can be perceived that CNN+LSTM and CNN+GRU methods use a total of 11 spatial-temporal layers including 8 convolutional and 3 temporal layers. Similarly, CNN+Bi-LSTM and CNN+Bi-GRU methods use a total of 14 layers that include 8 convolutional and 6 temporal layers (having 3 forward and 3 backward pass layers). Finally, the proposed framework (DA-CNN+Bi-GRU) has a total of 18 layers comprising of 12 convolutional layers (8 convolutional and 4 attentional) and 6 temporal layers (having 3 forward and 3 backward pass layers).
The training performance (in terms of accuracy) of each baseline method along with our proposed method is depicted in Figure~\ref{Fig:TrainingHistory}. It can be seen from Figure~\ref{Fig:TrainingHistory} that our proposed method (DA-CNN+Bi-GRU) performs better than other baseline methods in terms of accuracy. For instance, in Figure~\ref{Fig:TrainingHistory}\,(a) for YouTube action dataset, our method achieves the best accuracy score throughout 300 epochs. In Figure~\ref{Fig:TrainingHistory}\,(b) for UCF50 dataset, our method (DA-CNN+Bi-GRU) does not perform the best in early 35 epochs, where CNN+Bi-GRU dominates; however, after 35 epochs our methods starts improving and finally trains with the best accuracy at 300th epoch. Similarly, in Figure~\ref{Fig:TrainingHistory}\,(c), our method (DA-CNN+Bi-GRU) starts as the second-best method in early training epochs where CNN+Bi-GRU dominates; however, after 20 epochs our proposed method attains the best accuracy as compared to the other baseline methods and remains the best till the end of training. The obtained performances of these baseline methods along with our proposed method across three benchmark datasets are presented in Table \ref{tab:table3}. From Table~\ref{tab:table3}, it can be noticed that the proposed framework dominates all the baseline method across each dataset. For instance, the proposed framework attains the best accuracy score of 98.0\% over YouTube action dataset among all the baseline methods, whereas CNN+Bi-GRU obtains the second-best accuracy score of 92.1\%. Similarly, on UCF50 dataset, the proposed framework obtains the highest accuracy score of 97.5\%, whereas the runner-up is CNN+Bi-GRU with an accuracy of 93.6\%. Finally, for HMDB51 dataset, it can be seen that our proposed method dominates all the baseline methods by achieving the best accuracy score of 79.3\%, whereas CNN+Bi-GRU is the runner-up method that attains the second-best accuracy score of 72.4\%. The best and the runner-up results are highlighted in bold and italic , respectively.
\begin{table}[t]
\centering
\caption{Network settings of experimented baseline methods and our proposed framework.}
\resizebox{\columnwidth}{!}{
\begin{tabular}{lcc}\hline
Method &Spatial block layers &Temporal block layers \\ \hline
CNN+LSTM & 8 convolutional & 3 LSTM \\
CNN+Bi-LSTM & 8 convolutional & 6 LSTM (3 forward and 3 backward)\\
CNN+GRU & 8 convolutional & 3 GRU \\
CNN+Bi-GRU & 8 convolutional & 6 GRU (3 forward and 3 backward)\\
DA-CNN+Bi-GRU & 12 convolutional (8 convolutional and 4 attentional) & 6 GRU (3 forward and 3 backward)\\ \hline
\end{tabular}
}
\label{tab:table2}
\end{table}
\begin{table}[t]
\centering
\caption{Quantitative comparative analysis of our proposed framework with other baseline methods.}
\resizebox{\columnwidth}{!}{
\begin{tabular}{lcc}\hline
Method &Dataset &Accuracy ($\%$)\\ \hline
CNN+LSTM & YouTube action &64.7\\
CNN+Bi-LSTM & YouTube action &84.2\\
CNN+GRU & YouTube action &88.5\\
CNN+Bi-GRU & YouTube action & \textit{92.1}\\
DA-CNN+Bi-GRU (Proposed) & YouTube action &\textbf{98.0}\\ \hline
CNN+LSTM & UCF50 &76.3\\
CNN+Bi-LSTM & UCF50 &83.3\\
CNN+GRU & UCF50 &87.6\\
CNN+Bi-GRU & UCF50 & \textit{93.6}\\
DA-CNN+Bi-GRU (Proposed) & UCF50 &\textbf{97.5}\\ \hline
CNN+LSTM & HMDB51 &56.7\\
CNN+Bi-LSTM & HMDB51 &63.2\\
CNN+GRU & HMDB51 &68.0\\
CNN+Bi-GRU & HMDB51 & \textit{72.4}\\
DA-CNN+Bi-GRU (Proposed) & HMDB51 &\textbf{79.3}\\
\hline
\end{tabular}
}
\label{tab:table3}
\end{table}
\subsection{Comparison with State-of-the-Art Methods}
To show the effectiveness of our proposed framework for the human activity recognition task, we have conducted extensive comparative analysis of our method with the state-of-the-art methods in terms of overall accuracy. The quantitative comparisons of our method with the state-of-the-art methods for YouTube action, UCF50, and HMDB51 datasets are listed in Table \ref{tab:table4}, \ref{tab:table5}, and \ref{tab:table6}, respectively. The best results in these tables are represented in bold, whereas the runner-up results are highlighted in italic text. Considering the presented results, it can be noticed that our proposed framework (DA-CNN+Bi-GRU) outperforms state-of-the-art methods on UCF50 and HMDB51 datasets, whereas it attains runner-up performance on YouTube action dataset. For YouTube action dataset, the STDN \cite{zhang2020human} has the best performance with an accuracy of 98.2\%, whereas the proposed method attains the runner-up performance by obtaining an accuracy of 98.0\%, which is within 0.2\% accuracy of the best-performing STDN \cite{zhang2020human}. Thus, for most practical purposes, our proposed framework attains comparable performance to the STDN \cite{zhang2020human}. Rest of the methods that include multi-task hierarchical clustering \cite{liu2016hierarchical}, BT-LSTM \cite{ye2018learning}, deep autoencoder \cite{ullah2019action}, two-stream attention LSTM \cite{dai2020human}, weighted entropy-variance based feature selection \cite{afza2021framework}, dilated CNN+BiLSTM+RB \cite{muhammad2021human}, DS-GRU \cite{ullah2021efficient}, and local-global features + QSVM \cite{al2021making} obtain 89.7\%, 85.3\%, 96.2\%, 96.9\%, 94.5\%, 89.0\%, 97.1\%, and 82.6\% accuracies, respectively. For the UCF50 dataset, the proposed method dominates the state-of-the-art methods by obtaining the best accuracy of 97.5\%, whereas the (LD-BF) + (LD-DF) \cite{du2022linear} obtains the second-based accuracy of 96.7\%. The local-global features + QSVM \cite{al2021making} achieves the lowest accuracy of 69.4\%, whereas the rest of the methods including multi-task hierarchical clustering \cite{liu2016hierarchical}, deep autoencoder \cite{ullah2019action}, ensemble model with sward-based optimization \cite{zhang2021intelligent}, and DS-GRU \cite{ullah2021efficient} obtain 93.2\%, 96.4\%, 92.2\%, and 95.2\% accuracies, respectively. Finally, for the HMDB51 dataset comprising of challenging action videos, our proposed method achieves the best results by obtaining an accuracy of 79.3\%, whereas the runner-up method is evidential deep learning \cite{bao2021evidential} that attains an accuracy of 77.0\%. The multi-task hierarchical clustering method \cite{liu2016hierarchical} achieves an accuracy of 51.4\%, which is the lowest among all comparative methods on HMDB51 dataset. The rest of comparative methods including STPP+LSTM \cite{wang2017two}, optical flow + multi-layer LSTM \cite{ullah2018activity}, TSN \cite{wang2018temporal}, IP-LSTM \cite{yu2019learning}, deep autoencoder \cite{ullah2019action}, TS-LSTM + temporal-inception \cite{ma2019ts}, HATNet \cite{diba2019holistic}, correlational CNN+LSTM \cite{majd2020correlational}, STDN \cite{zhang2020human}, DB-LSTM+SSPF \cite{he2021db}, DS-GRU \cite{ullah2021efficient}, TCLC \cite{zhu2021temporal}, and semi-supervised temporal gradient learning \cite{xiao2022learning} obtain accuracies of 70.5\%, 72.2\%, 70.7\%, 58.6\%, 70.3\%, 69.0\%, 74.8\%, 66.2\%, 56.5\%, 75.1\%, 72.3\%, 71.5\%, and 75.9\%, respectively. Considering the overall comparative analysis, the proposed method obtains comparable performance to the best-performing method on the YouTube action dataset, and greatly dominates the state-of-the-art comparative methods on UCF50 and HMDB51 datasets, thus demonstrating the superiority of our proposed method over the exiting action recognition methods.
\begin{table}[t]
\centering
\caption{Quantitative comparative analysis of our proposed method with the state-of-the-art action recognition methods for YouTube action dataset.}
\resizebox{\columnwidth}{!}{
\begin{tabular}{lcc}\hline
Method &Year &Accuracy ($\%$)\\ \hline
Multi-task hierarchical clustering \cite{liu2016hierarchical} &2017 &89.7 \\
BT-LSTM \cite{ye2018learning} &2018 &85.3 \\
Deep autoencoder \cite{ullah2019action} &2019 &96.2 \\
STDN \cite{zhang2020human} &2020 &\textbf{98.2} \\
Two-stream attention LSTM \cite{dai2020human} &2020 &96.9 \\
Weighted entropy-variances based\\ feature selection \cite{afza2021framework} &2021 &94.5\\
Dilated CNN+BiLSTM+RB \cite{muhammad2021human} &2021 &89.0 \\
DS-GRU \cite{ullah2021efficient} &2021 &97.1 \\
Local-global features + QSVM \cite{al2021making} &2021 &82.6 \\
DA-CNN+Bi-GRU (Proposed) &2022 & \textit{98.0} \\\hline
\end{tabular}
}
\label{tab:table4}
\end{table}
\begin{table}[t]
\centering
\caption{Quantitative comparative analysis of our proposed method with the state-of-the-art action recognition methods for UCF50 dataset.}
\resizebox{\columnwidth}{!}{
\begin{tabular}{lcc}\hline
Method &Year &Accuracy ($\%$)\\ \hline
Multi-task hierarchical clustering\cite{liu2016hierarchical} &2017 &93.2 \\
Deep autoencoder \cite{ullah2019action} &2019 &96.4 \\
Ensemble model with swarm-based\\ optimization \cite{zhang2021intelligent} &2021 &92.2 \\
DS-GRU \cite{ullah2021efficient} &2021 &95.2 \\
Local-global features + QSVM \cite{al2021making} &2021 &69.4 \\
(LD-BF) + (LD-DF) \cite{du2022linear} &2022 & \textit{97.5} \\
DA-CNN+Bi-GRU (Proposed) &2022 &\textbf{98.0} \\
\hline
\end{tabular}
}
\label{tab:table5}
\end{table}
\begin{table}[t]
\centering
\caption{Quantitative comparative analysis of our proposed method with the state-of-the-art action recognition methods for HMDB51 dataset.}
\resizebox{\columnwidth}{!}{
\begin{tabular}{lcc}\hline
Method &Year &Accuracy ($\%$)\\ \hline
Multi-task hierarchical clustering\cite{liu2016hierarchical} & 2017 & 51.4 \\
STPP+LSTM \cite{wang2017two} & 2017 & 70.5 \\
Optical flow + multi-layer LSTM \cite{ullah2018activity} & 2018 & 72.2 \\
TSN \cite{wang2018temporal} & 2018 & 70.7 \\
IP-LSTM \cite{yu2019learning} & 2019 & 58.6 \\
Deep autoencoder \cite{ullah2019action} & 2019 & 70.3 \\
TS-LSTM + temporal-inception \cite{ma2019ts} & 2019 & 69.0 \\
HATNet \cite{diba2019holistic} & 2019 & 74.8 \\
Correlational CNN + LSTM \cite{majd2020correlational} & 2020 & 66.2 \\
STDAN \cite{zhang2020human} & 2020 & 56.5 \\
DB-LSTM+SSPF \cite{he2021db} & 2021 & 75.1 \\
DS-GRU \cite{ullah2021efficient} & 2021 &72.3 \\
TCLC \cite{zhu2021temporal} & 2021 &71.5 \\
Evidential deep learning \cite{bao2021evidential} & 2021 & \textit{77.0} \\
Semi-supervised temporal gradient\\ learning \cite{xiao2022learning} & 2022 & 75.9 \\
DA-CNN+Bi-GRU (Proposed) & 2022 &\textbf{79.3} \\
\hline
\end{tabular}
}
\label{tab:table6}
\end{table}
\begin{figure*}[t]
\includegraphics[width=181mm]{Figures/Visual_recognition_results.png}
\caption{The visual recognition results of our proposed DA-CNN+Bi-GRU framework with predicted classes and their confidence scores for the test videos taken from the YouTube action, UCF50, and HMDB51 datasets.}
\label{Fig:VisualRecognition}
\end{figure*}
\subsection{Action Recognition Visualization}
To validate the recognition efficiency of our proposed framework, we have tested our framework on 15\% of test videos taken from each dataset (including YouTube action, UCF50, and HMDB51). The prepared test sets are validated for the action recognition task using our proposed framework and the visual results from the test experiments are depicted in Figure~\ref{Fig:VisualRecognition}. In Figure~\ref{Fig:VisualRecognition}, the representative frames of the predicted action clips are presented along with their ground truths, model predicted actions, and confidence scores over the probability prediction bar graphs for better understanding of readers. It can be perceived from the presented visual results that the proposed framework predicts most of the actions including brush hair, volleyball spiking, basketball, climb, fall floor, bench press, horse race, billiards, diving, baseball pitch, and hula hoop with 0.99\sout{\%} probability or 99\% confidence. Though, for some action classes, such as clap, fencing, golf swing, and high jump, the proposed framework also generates non-zero probabilities for wrong action classes; however, these probabilities for wrong action classes are still very low and thus do not affect the prediction of actual action class. Hence, the obtained qualitative visual results verify the effectiveness of our proposed framework for practical use in different vision-based human action recognition and monitoring environments.
\subsection{Runtime Analysis}
To analyze the effectiveness and feasibility of our proposed framework for practical applications in real-time environments, we have estimated the runtime of our method for action recognition tasks in terms of SPF and FPS with and without using GPU resources. The obtained runtime results are then compared with the stat-of-the-art methods. Table~\ref{tab:Runtime} presents and compares the runtime of our proposed framework with the running times of the contemporary action recognition methods. Results in Table~\ref{tab:Runtime} demonstrate that our proposed framework outperforms the state-of-the-art methods when executing on both GPU and central processing unit (CPU) platforms. Results indicate that our proposed framework attains 0.0036 SPF and 300 FPS while running on GPU, whereas it attains 0.0049 SPF and 250 FPS while running on CPU. Results further show that the second-best execution time results on GPU are achieved by \cite{wang2017two}, which are 0.0053 SPF and 186.6 FPS. In Table~\ref{tab:Runtime}, the best runtime results are highlighted with bold and runner-up results are emphasized with italic. Experimental results indicate that for the SPF metric, our proposed framework can provide an improvement of up to 18.6$\times$ when running on GPU and an improvement of 87.76$\times$ when running on CPU as compared to other contemporary activity recognition methods. Experimental results further reveal that for the FPS metric, our proposed framework can provide an improvement of up to 21.43$\times$ when running on GPU and an improvement of 166.6$\times$ when running on CPU as compared to other contemporary activity recognition methods. It is also worth mentioning here that the storage requirement of our proposed framework is just 5.4\,MB, and thus our framework can be run on resource-constrained IoT and edge devices with very limited memory including today’s smart cameras, Arduino, and Raspberry pi. These runtime and storage requirement results demonstrate that the proposed framework is a suitable candidate for deployment on resource-constrained IoT and edge devices as the proposed framework exhibits better accuracy, lower execution time, and low storage requirements as compared to contemporary activity recognition methods.
\begin{table}[t!]
\caption{Runtime analysis of our proposed framework with state-of-the-art human action recognition methods.}
\resizebox{\columnwidth}{!}{
\begin{tabular}{cccccccc}
\toprule
\multirow{2}{*}{Method} & \multicolumn{2}{c}{Seconds per Frame (SPF)} & \multirow{2}{*}{Year} & \multicolumn{2}{c}{Frames per Second (FPS)} & \\
\cmidrule{2-3} \cmidrule{5-6}
& GPU & CPU & & GPU & CPU \\ \midrule
STPP+LSTM \cite{wang2017two} & \textit{0.0053} & - & 2017 & \textit{186.6} & - \\
Optical flow + multi-layer LSTM \cite{ullah2018activity} &0.0356 & 0.18 & 2018 & 30 & 3.5 \\
Deep autoencoder \cite{ullah2019action} & 0.0430 & 0.43 & 2019 & 24 & 1.5 \\
IP-LSTM \cite{yu2019learning} & 0.0431 & - & 2019 & 23.2 & - \\
STDN \cite{zhang2020human} & 0.0075 & - & 2020 & 132 & - \\
DS-GRU \cite{ullah2021efficient} & 0.0276 & - & 2021 & 37 & - \\
(LD-BF) + (LD-DF) \cite{du2022linear} & 0.0670 & - & 2022 & 14 & - \\
DA-CNN+Bi-GRU (Proposed) & \textbf{0.0036} & \textbf{0.0049} & 2022 & \textbf{300} & \textbf{250} \\
\\ \bottomrule
\end{tabular}
}
\label{tab:Runtime}
\end{table}
\section{Conclusions and Future Research Directions} \label{sec:conclusion}
In this work, we have proposed a cascaded spatial-temporal discriminative feature learning framework for human activity recognition in video streams. The proposed method encapsulates the attentional (channel and spatial attention) CNN architecture and bi-directional GRU network as a unified framework for single instance training and efficient spatial temporal modeling of human actions. The attentional CNN architecture comprises of channel and spatial attentions, which help retrieve the prominent discriminative features from the object-specific regions, and thus generate high quality saliency-aware feature maps. The bi-directional GRU learns the temporal modeling of long-term human action sequences using two-way gradient learning (i.e., forward and backward pass), which allows our approach to utilize the learned knowledge not only from the previous frames but also from the upcoming/next frames. Such bi-directional modeling of human actions greatly helps our method to improve the learning ability while training and the prediction precision while inferencing. To evaluate the efficiency of our method, we have conducted extensive experiments on three publicly available human action benchmark datasets. The obtained experimental results are compared with the state-of-the-art methods on three benchmark human action recognition datasets, that include youtube action, UCF50, and HMDB51 datasets. Experimental results verify the effectiveness of our method in terms of both model robustness and computational efficiency. Further, we have analyzed the runtime performance of our proposed framework in terms of seconds per frame (SPF) and frames per second (FPS) for both CPU and GPU execution environments. The obtained runtime assessment results reveal that our proposed framework can attain an improvement of up to 88$\times$ for the SPF metric and up to 167$\times$ for the FPS metric as compared to other contemporary action recognition methods. Additionally, our proposed framework requires a storage of only 5.3\,MB, which makes it feasible for deployment on devices with limited memory. Thus, the overall efficiency of our framework in terms of recognition performance (accuracy), low execution time, and low storage requirements, makes our framework a strong candidate for real-time IoT and edge applications.
Currently, our proposed method only uses spatial attention (channel and spatial attention) mechanism, which is indeed very effective. However, in future we plan to use temporal attention mechanism together with spatial attention, because such hybrid attention has a great potential to improve the human activity recognition performance.
\bibliographystyle{IEEEtran}
\balance
|
\section{Introduction}
Multi-frame high dynamic range (HDR) imaging aims to generate images with a wider dynamic range and more realistic details by merging several low dynamic range (LDR) images with varying exposures, which can be well fused to an HDR image if they are aligned perfectly~\cite{ram2017deepfuse,raman2011reconstruction,ma2017robust,zhang2011gradient,mertens2007exposure,ma2019deep}. In practice, however, this ideal situation is often undermined by camera motions and foreground dynamic objects, yielding unfavorable \textit{ghosting artifacts} in the reconstructed HDR results. Various methods, commonly referred to as \textit{HDR deghosting algorithms}, have thus been proposed to acquire high-quality ghost-free HDR images.
Traditionally, several methods propose to remove ghosting artifacts by aligning the input LDR images~\cite{bogoni2000extending,hu2013hdr,kang2003high,zimmer2011freehand} or rejecting misaligned pixels~\cite{gallo2009artifact,grosch2006fast,pece2010bitmap,jacobs2008automatic,khan2006ghost} before the image fusion. However, accurate alignment is challenging, and the overall HDR effect is diminished when useful information is dropped by imprecise pixel rejection. Therefore, CNN-based learning algorithms have been introduced to solve ghosting artifact by exploring deep features in data-driven manners.
Existing CNN-based deghosting methods can be mainly classified into two categories. In the first category, LDR images are pre-aligned using homography~\cite{hartley2003multiple} or optical flow~\cite{baker2011database}, and then multi-frame fusion and HDR reconstruction are performed using a CNN~\cite{kalantari2017deep,prabhakar2019fast,prabhakar2020towards,wu2018deep}. However, homography cannot align dynamic objects in the foreground, and optical flow is unreliable in the presence of occlusions and saturations. Hence, the second category proposes end-to-end networks with implicit alignment modules~\cite{yan2019attention,liu2021adnet,chung2022high} or novel learning strategies~\cite{niu2021hdr,prabhakar2021labeled} to handle ghosting artifacts, achieving state-of-the-art performance. Nonetheless, the restraints appear when confronted with long-range object movements and heavy intensity variations. Fig.~\ref{fig:teaser} shows a representative scene where large motions and severe saturations occur, producing unexpected ghosting and distortion artifacts in the results of previous CNN-based methods. The reason lies in the intrinsic locality restriction of convolution. CNN needs to stack deep layers to obtain a large receptive field and is thus ineffective to model long-range dependency (e.g., ghosting artifacts caused by large motion)~\cite{naseer2021intriguing}. Moreover, convolutions are content-independent as the same kernels are shared within the whole image, ignoring the long-range intensity variations of different image regions~\cite{liang2021swinir}. Therefore, exploring content-dependent algorithms with long-range modeling capability is demanding for further performance improvement.
Vision Transformer (ViT)~\cite{dosovitskiy2020image} has recently received increasing research interest due to its superior long-range modeling capability. However, our experimental results indicate two major issues that hinder its applications on HDR deghosting. On the one hand, Transformers lack the inductive biases inherent to CNN and therefore do not generalize well when trained on insufficient amounts of data~\cite{dosovitskiy2020image,liang2021swinir}, despite the fact that available datasets for HDR deghosting are limited as gathering huge numbers of realistic labeled samples is prohibitively expensive. On the other hand, the neighbor pixel relationships of both intra-frame and inter-frame are critical for recovering local details across multiple frames, while the pure Transformer is ineffective for extracting such local context.
To this end, we propose a novel Context-Aware Vision Transformer (CA-ViT), which is formulated to concurrently capture both global and local dependencies with a dual-branch architecture. For the global branch, we employ a window-based multi-head Transformer encoder to capture long-range contexts. For the local branch, we design a local context extractor (LCE), which extracts the local feature maps through a convolutional block and selects the most useful features across multiple frames by channel attention mechanism. The proposed CA-ViT, therefore, makes local and global contexts work in a complementary manner. By incorporating with the CA-ViT, we propose a novel Transformer-based framework (termed as HDR-Transformer) for ghost-free HDR imaging.
Specifically, the proposed HDR-Transformer mainly consists of a feature extraction network and an HDR reconstruction network. The feature extraction network extracts shallow features and fuses them coarsely through a spatial attention module. The early convolutional layers can stabilize the training process of the vision Transformer and the spatial attention module helps to suppress undesired misalignment. The HDR reconstruction network takes the proposed CA-ViT as basic components and is constituted hierarchically. The CA-ViTs model both long-range ghosting artifacts and local pixel relationship, thus helping to reconstruct ghost-free high-quality HDR images (an example is shown in Fig.~\ref{fig:teaser}) without the need of stacking very deep convolution blocks. In summary, the main contributions of this paper can be concluded as follows:
\begin{itemize}
\item We propose a new vision Transformer, called CA-ViT, which can fully exploit both global and local image context dependencies, showing significant performance improvements over prior counterparts.
\item We present a novel HDR-Transformer that is capable of removing ghosting artifacts and reconstructing high-quality HDR images with lower computational costs. To our best knowledge, this is the first Transformer-based framework for HDR deghosting.
\item We conduct extensive experiments on three representative benchmark HDR datasets, which demonstrates the effectiveness of HDR-Transformer against existing state-of-the-art methods.
\end{itemize}
\section{Related Work}
\label{sec:related_work}
\subsection{HDR Deghosting Algorithms}
We summarize existing HDR deghosting algorithms into three categories, i.e., motion rejection methods, image registration methods, and CNN-based methods.
\mysubtitle{Motion rejection methods} Methods based on motion rejection proposed first to register the LDR images globally and then reject the pixels which are detected as misaligned. Grosch \emph{et al.} generated an error map based on the alignment color differences to reject mismatched pixels~\cite{grosch2006fast}. Pece \emph{et al.} detected motion areas using a median threshold bitmap for input LDR images~\cite{pece2010bitmap}. Jacobs \emph{et al.} identified misaligned locations using weighted intensity variance analysis~\cite{jacobs2008automatic}. Zhang \emph{et al.}~\cite{zhang2011gradient} and Khan \emph{et al.}~\cite{khan2006ghost} proposed to calculate gradient-domain weight maps and probability maps for the LDR input images, respectively. Additionally, Oh \emph{et al.} presented a rank minimization method for the purpose of detecting ghosting regions~\cite{oh2014robust}. These methods frequently produce unpleasing HDR results due to the loss of useful information while rejecting pixels.
\mysubtitle{Motion registration methods} Motion registration methods rely on aligning the non-reference LDR images to the reference one before merging them. Begoni \emph{et al} proposed using optical flow to predict motion vectors~\cite{bogoni2000extending}. Kang \emph{et al.} transferred the LDR picture intensities to the luminance domain based on the exposure time and then estimated optical flow to account for motion~\cite{kang2003high}. Zimmer \emph{et al.} reconstructed the HDR image by first registering the LDR images with optical flow~\cite{zimmer2011freehand}. Sen \emph{et al.} presented a patch-based energy minimization method that simultaneously optimizes alignment and HDR reconstruction~\cite{sen2012robust}. Hu \emph{et al.} proposed to optimize the image alignment using brightness and gradient consistencies on the transformed domain~\cite{hu2013hdr}. Motion registration methods are more robust than motion rejection methods. However, when large motions occur, this approach generates visible ghosting artifacts.
\mysubtitle{CNN-based methods} Several CNN-based methods have been recently proposed. Kalantari \emph{et al.} proposed the first CNN-based method for multi-frame HDR imaging of dynamic scenes. They employed a CNN to blend the LDR images after aligning them with optical flow~\cite{kalantari2017deep}. Wu \emph{et al.} developed the first non-flow-based framework by formulating HDR imaging as an image translation problem~\cite{wu2018deep}. Instead of using explicit alignment, Yan \emph{et al.} adopted a spatial attention module to address ghosting artifacts~\cite{yan2019attention}. Prabhakar \emph{et al.} proposed an efficient method to generate HDR images with bilateral guided upsampler~\cite{prabhakar2020towards} and further explored zero and few-shot learning for HDR Deghosting~\cite{prabhakar2021labeled}. Lately, Niu \emph{et al.} proposed the first GAN-based framework for multi-frame HDR imaging~\cite{niu2021hdr}. The approaches based on CNNs demonstrate superior capabilities and achieve state-of-the-art performance. However, ghosting artifacts can still be observed when confronted with large motion and extreme saturation.
\subsection{Vision Transformers}
Transformers have achieved huge success in the field of natural language processing~\cite{vaswani2017attention,devlin2018bert}, where the multi-head self-attention mechanism is employed to capture long-range correlations between word token embeddings. Recently, ViT~\cite{dosovitskiy2020image} has shown that a pure Transformer can be applied directly to sequences of non-overlapping image patches and performs very well on image classification tasks. Liu \emph{et al.} developed Swin Transformer, a hierarchical structure where cross-window contexts are captured through the shift-window scheme~\cite{liu2021Swin}. Chen \emph{et al.} built IPT, a pretrained Transformer model for low-level computer vision tasks~\cite{chen2021pre}. Liang \emph{et al.} extended the Swin Transformer for image restoration and proposed SwinIR, achieving state-of-the-art performance on image super-resolution and denoising~\cite{liang2021swinir}. Unlike CNN-based methods, our approach is inspired by~\cite{liu2021Swin,liang2021swinir} and built on Transformers.
\section{Method}
\label{sec:method}
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{./figures/ca_vit.pdf}
\caption{Illustration of the proposed CA-ViT. As shown in Fig.~\ref{fig:ctl} (a), the CA-ViT is designed as a dual-branch architecture where the global branch models long-range dependency among image contexts through a multi-head Transformer encoder, and the local branch explores both intra-frame local details and inner-frame feature relationship through a local context extractor. Fig.~\ref{fig:ctl} (b) depicts the key insight of our HDR deghosting approach with CA-ViT. To remove the residual ghosting artifacts caused by large motions of the hand (marked with blue), long-range contexts (marked with red), which are required to hallucinate reasonable content in the ghosting area, are modeled by the self-attention in the global branch. Meanwhile, the well-exposed non-occluded local regions (marked with green) can be effectively extracted with convolutional layers and fused by the channel attention in the local branch.}\label{fig:ctl}
\end{figure}
\subsection{CA-ViT}
\label{sec:ctl}
Unlike prior vision Transformers that adopt the pure Transformer encoder, we propose a dual-branch context-aware vision Transformer (CA-ViT), which explores both the global and local image information. As depicted in Fig.~\ref{fig:ctl} (a), the proposed CA-ViT is constructed with a global Transformer encoder branch and a local context extractor branch.
\subsubsection{Global Transformer Encoder}
For the global branch, we employ a window-based multi-head Transformer encoder~\cite{dosovitskiy2020image} to capture long-range information. The Transformer encoder consists of a multi-head self-attention (MSA) module and a multi-layer perceptron (MLP) with residual connection.
Considering the input token embeddings $E\in \mathbb{R}^{H \times W \times D}$, the global context branch can be formulated as:
\begin{equation}
\begin{aligned}\label{eq:ctl_global}
E = \textit{MSA}({\textit{LN}(E)}) + E, \\
\textit{CTX}_{global} = \textit{MLP}({\textit{LN}(E)}) + E,
\end{aligned}
\end{equation}
where $\textit{LN}$ denotes LayerNorm, and $\textit{CTX}_{global}$ denotes the global contexts captured by the Transformer encoder.
\subsubsection{Local Feature Extractor}
For the local branch, we design a local context extractor (LCE) to extract local information $\textit{CTX}_{local}$ from adjacent pixels and select cross-channel features for fusion, which is defined as:
\begin{align}\label{eq:ctl_local}
\textit{CTX}_{local} = \textit{LCE}({\textit{LN}(E)}).
\end{align}
Specifically, for the token embeddings $\textit{E}$ normalized with an LN layer, we first reshape them into $H \times W\times D$ features and use a convolution block to extract local feature maps $f_{local}$. The local features are then average pooled to a shape of $1 \times 1\times D$, and the channel-wise weights $\omega$ are calculated from two linear layers followed by a ReLU and a sigmoid activation layer, respectively. Afterward, the useful feature maps are selected through a channel-wise calibration from the original local features $f_{local}$, i.e.,
\begin{equation}
\begin{aligned}\label{eq:lce}
f_{local} &= \textit{Conv}(\textit{LN}(E)),\\
\omega &= \sigma_2(\textit{FC}(\sigma_1(\textit{FC}(f_{local})))), \\
\textit{CTX}_{local} &= \omega \odot f_{local}, \\
\end{aligned}
\end{equation}
where $\sigma_1$ and $\sigma_2$ denote the ReLU and sigmoid layer, and $\textit{FC}$ denotes the linear layer. As a result, the local context branch not only adds the locality into the Transformer encoder, but also identifies the most informative local features across multiple frames for feature fusion.
Finally, a context fusion layer is employed to combine the global and local contexts. Although other transformation functions (e.g., linear or convolution layer) can be used to implement the context fusion layer, in this paper, we simply merge the contexts by element-wise addition to reduce the influence of additional parameters.
\subsection{HDR Deghosting}
The task of deep HDR deghosting aims to reconstruct a ghost-free HDR image through deep neural networks. Following most of the previous works~\cite{kalantari2017deep,wu2018deep,yan2019attention}, we consider 3 LDR images (i.e., $I_i, i=1,2,3$) as input and refer to the middle frame $I_2$ as the reference image. To better utilize the input data, the LDR images $\{I_i\}$ are first mapped to the HDR domain using the gamma correction, generating the gamma-corrected images $\{\check{I}_i\}$:
\begin{align}\label{eq:gamma_correction}
\quad \check{I}_{i} = \frac{(I_{i})^\gamma}{t_{i}}, \quad i=1, 2, 3,
\end{align}
where $t_i$ denotes the exposure time of $I_i$, and $\gamma$ is the gamma correction parameter, which is set to 2.2 in this paper. We then concatenate the original LDR images $\{I_i\}$ and the corresponding gamma-corrected images $\{\check{I}_i\}$ into a 6-channels input $\{X_i\}$. This strategy is suggested in~\cite{kalantari2017deep} as the LDR images help to detect the noisy or saturated regions, while the gamma-corrected images are helpful for detecting misalignments. Finally, the network $\varPhi(\cdot)$ is defined as:
\begin{align}\label{eq:network}
I^{\mathrm{\hat{H}}} = \varPhi (X_i; \theta), \quad i=1, 2, 3,
\end{align}
where $I^{\mathrm{\hat{H}}}$ denotes the reconstructed HDR image, and $\theta$ is the network parameters to be optimized.
Instead of stacking very deep CNN layers to obtain a large receptive field as existing CNN-based approaches, we propose the HDR-Transformer to handle HDR deghosting. Our key insight is that, with the specifically-designed dual-branch CA-ViT, the long-range ghosting can be well modeled in the global branch, and the local branch helps to recover fine-grained details. We describe the architecture of the proposed HDR-Transformer in the next section.
\begin{figure*}[t]
\centering
\includegraphics[width=1.0\linewidth]{./figures/pipeline_eccv.pdf}
\caption{The network architecture of HDR-Transformer. The pipeline consists of two stages: (a) The feature extraction network first extracts the coarse features through a spatial attention module. (b) The extracted features are then fed into the HDR reconstruction network to recover the HDR results. The HDR reconstruction network consists of several Context-aware Transformer Blocks (CTBs), which take the proposed CA-ViT as basic components.}\label{fig:pipeline}
\end{figure*}
\subsection{Overall Architecture of HDR-Transformer}
As illustrated in Fig.~\ref{fig:pipeline}, the overall structure of our proposed HDR-Transformer mainly consists of two components, i.e., feature extraction network (Fig.~\ref{fig:pipeline} (a)) and HDR reconstruction network (Fig.~\ref{fig:pipeline} (b)). Given three input images, we first extract the spatial features through a spatial attention module. The extracted coarser features are then embedded and fed into the Transformer-based HDR reconstruction network, generating the reconstructed ghost-free HDR image.
\subsubsection{Feature Extraction Network}
The early convolution layers help to stabilize the training process of Vision Transformers~\cite{xiao2021early}. For the input images $X_{i} \in \mathbb{R}^{H \times W \times 6}, i=1, 2, 3$, we first extract the shallow features $f_{i} \in \mathbb{R}^{H \times W \times C}$ by three separate convolution layers, where $C$ is the number of channels. Then, we concatenate each non-reference feature (i.e., $f_{1}$ and $f_{3}$) with the reference feature $f_{2}$ and calculate the attention maps $m_{i}$ through a spatial attention module $\mathcal{A}$:
\begin{align}\label{eq:att_maps}
m_{i} = \mathcal{A} (f_{i}, f_{2}), \quad i=1, 3,
\end{align}
The attention features $f^{'}_{i}$ are computed by multiplying the attention maps $m_i$ by the non-reference features $f_i$, i.e.,
\begin{align}\label{eq:att_feats}
f^{'}_{i} = f_{i} \odot m_{i}, \quad i=1, 3,
\end{align}
where $\odot$ denotes the element-wise multiplication. The spatial attention module has been proved to effectively reduce undesired contents caused by foreground object movements~\cite{yan2019attention,liu2021adnet}. The convolution layers in the attention module can also increase the inductive biases for the subsequent Transformer layers.
\subsubsection{HDR Reconstruction Network}
As shown in Fig.~\ref{fig:pipeline}, the HDR reconstruction network is mainly composed of several context-aware Transformer blocks (CTBs). The input of the first CTB $f_{att} \in \mathbb{R}^{H \times W \times D}$ is obtained from $f^{'}_{1}$, $f_{2}$, and $f^{'}_{3}$ and embedded into token embeddings, where $\textit{D}$ denotes the embed dimension. The HDR result is reconstructed by $\textit{N}$ subsequent CTBs and a following convolution block. We also adopt the global skip connection to stabilize the optimization process.
\mysubtitle{Context-aware Transformer Block}
As illustrated in Fig.~\ref{fig:ctl} (b), when suffering occlusion caused by large object movements and heavy saturation, long-range context is required for removing the corresponding ghosting regions and hallucinating reasonable content, while the non-occluded areas can be fused well by the convolutional layers. To this end, we develop the context-aware Transformer block (CTB) by taking the proposed CA-ViT as the basic component.
For clarity, each CTB contains $\textit{M}$ CA-ViTs. For the $n$-th CTB with the input of $F_{n,0}$, the output of the $m$-th CA-ViT can be formulated as:
\begin{align}\label{eq:cstb_ctl}
F_{n,m} = \mathcal{C}_{n,m}(F_{n,m-1}), \quad m=1, 2,..., M,
\end{align}
where $C_{n, m}(\cdot)$ denotes the corresponding CA-ViT. Then, we feed the output of the $\textit{M}$-th CA-ViT into a dilated convolution layer. The dilated convolutional layer is employed to increase the receptive field of the context range. We also adopt the residual connection in each CTB for better convergence. Consequently, the output of the $n$-th CTB is formulated as:
\begin{align}\label{eq:cstb_out}
F_{n} = \textit{DConv}(F_{n,M}) + F_{n, 0},
\end{align}
where $DConv(\cdot)$ denotes the dilated convolutional layer, and $\textit{M}$ and $\textit{N}$ are empirically set to 6 and 3, respectively.
\subsection{Loss Function}
As HDR images are typically viewed after tonemapping, we compute the loss in the tonemapped domain using the commonly used $\mu$-law function:
\begin{align}\label{eq:mu_law}
\mathcal{T}(x)=\frac{\log (1+\mu x)}{\log(1+\mu)},
\end{align}
where $\mathcal{T}(x)$ is the tonemapped HDR image, and we set $\mu$ to 5000. Unlike previous methods~\cite{kalantari2017deep,wu2018deep,yan2019attention} that only adopt the pixel-wise loss (e.g., $l_1$ or $l_2$ error), we utilize $l_1$ loss and perceptual loss to optimize the proposed HDR-Transformer. Given the estimated HDR image $I^{\hat{H}}$ and the ground truth HDR image $I^{H}$, the $l_1$ loss term is defined as:
\begin{align}\label{eq:recon_loss}
\mathcal{L}_{r} = \parallel \mathcal{T}(I^{H}) - \mathcal{T}(I^{\hat{H}}) \parallel_{1},
\end{align}
The perceptual loss~\cite{Johnson2016Perceptual} is widely used in image inpainting~\cite{liu2018image} for better visual quality improvements. We also apply the perceptual loss to enhance the quality of the reconstructed HDR images:
\begin{align}\label{eq:percep_loss}
\mathcal{L}_{p} = \sum_{j} \parallel \varPsi_{j}(\mathcal{T}(I^{H})) - \varPsi_{j}(\mathcal{T}(I^{\hat{H}})) \parallel_{1},
\end{align}
where $\varPsi(\cdot)$ denotes the activation feature maps extracted from a pre-trained VGG-16 network~\cite{simonyan2014very}, and $j$ denotes the $j$-th layer. We analyze the effectiveness of the perceptual loss in our ablation study (Sec.~\ref{sec:ablation_loss}). Eventually, our training loss function $\mathcal{L}$ is formulated as:
\begin{align}\label{eq:loss}
\mathcal{L} = \mathcal{L}_{r} + \lambda_{p}\mathcal{L}_{p},
\end{align}
where $\lambda_{p}$ is the hyper-parameter and we set it to 0.01.
\section{Experiments}
\label{sec:experiments}
\subsection{Dataset and Implementation Details}
\mysubtitle{Datasets} Following previous methods~\cite{wu2018deep,yan2019attention,yan2020deep,niu2021hdr}, we train our network on the widely used Kalantari \emph{et al.}'s dataset~\cite{kalantari2017deep}, which consists of 74 samples for training and 15 samples for testing. Each sample from Kalantari \emph{et al.}'s dataset comprises three LDR images with exposure values of $\left\langle-2, 0, +2\right\rangle$ or $\left\langle-3, 0, +3\right\rangle$, as well as a ground truth HDR image. During the training, we first crop patches of size $128\times128$ with a stride of 64 from the training set. We then apply rotation and flipping augmentation to increase the training size. We quantitatively and qualitatively evaluate our method on Kalantari \emph{et al.}'s testing set. We also conduct evaluations on Sen \emph{et al.}~\cite{sen2012robust}'s and Tursun \emph{et al.}~\cite{tursun2016objective}'s datasets to verify the generalization ability of our method.
\mysubtitle{Evaluation Metrics} We use PSNR and SSIM as evaluation metrics. To be more precise, we calculate PSNR-$l$, PSNR-$\mu$, SSIM-$l$, and SSIM-$\mu$ scores between the reconstructed HDR images and their corresponding ground truth. The `-$l$' and `-$\mu$' denote the linear and tonemapped domain values, respectively. Given that HDR images are typically displayed on LDR displays, metrics in the tonemapped domain more accurately reflect the quality of the reconstructed HDR images. Additionally, we conduct evaluations using the HDR-VDP-2~\cite{mantiuk2011hdr}, which is developed specifically for evaluating the quality of HDR images.
\mysubtitle{Implementation Details} Our HDR-Transformer is implemented by PyTorch. We use the ADAM optimizer with an initial learning rate of 2e-4 and set $\beta_{1}$ to 0.9, $\beta_{2}$ to 0.999, and $\epsilon$ to 1e-8, respectively. We train the network from scratch with a batch size of 16 and 100 epochs enables it to converge. The whole training is conducted on four NVIDIA 2080Ti GPUs and costs about two days.
\begin{table}[t]
\centering
\caption{Quantitative comparison between previous methods and ours on Kalantari \emph{et al.}~\cite{kalantari2017deep}'s test set. We use PSNR, SSIM, and HDR-VDP-2 as evaluation metrics. The `-$\mu$' and `-$l$' refers to values calculated on the tonemapped domain and the linear domain, respectively. All values are the average over 15 testing images and higher better. The best results are highlighted and the second best are underlined.}
\label{tab:results}
\resizebox{1.0\linewidth}{!}{
\begin{tabular}{
>{\centering\arraybackslash}p{1.4cm}
>{\centering\arraybackslash}p{1.4cm}
>{\centering\arraybackslash}p{1.6cm}
>{\centering\arraybackslash}p{1.6cm}
>{\centering\arraybackslash}p{1.6cm}
>{\centering\arraybackslash}p{1.6cm}
>{\centering\arraybackslash}p{1.8cm}
>{\centering\arraybackslash}p{1.4cm}
>{\centering\arraybackslash}p{2.6cm}}
\toprule
\multirow{3}{*}{Metrics} & \multicolumn{8}{c}{Methods} \\ \cmidrule(l){2-10}
& Sen12 & Hu13 & Kalantari17 & DeepHDR & AHDRNet & NHDRRNet & HDR-GAN & SwinIR & HDR-Transformer \\
& \cite{sen2012robust} & \cite{hu2013hdr} & \cite{kalantari2017deep} & \cite{wu2018deep} & \cite{yan2019attention} & \cite{yan2020deep} & \cite{niu2021hdr} &\cite{liang2021swinir} & Ours\\ \midrule
PSNR-$\mu$ & 40.80 & 35.79 & 42.67 & 41.65 & 43.63 & 42.41 & \underline{43.92} & 43.42 & \textbf{44.32} \\
PNRR-$l$ & 38.11 & 30.76 & 41.23 & 40.88 & 41.14 & 41.43 & 41.57 & \underline{41.68} & \textbf{42.18} \\
SSIM-$\mu$ & 0.9808 & 0.9717 & 0.9888 & 0.9860 & 0.9900 & 0.9877 & \underline{0.9905} & 0.9882 & \textbf{0.9916} \\
SSIM-$l$ & 0.9721 & 0.9503 & 0.9846 & 0.9858 & 0.9702 & 0.9857 & \underline{0.9865} & 0.9861 & \textbf{0.9884} \\
HDR-VDP-2 & 59.38 & 57.05 & 65.05 & 64.90 & 64.61 & 61.21 & \underline{65.45} & 64.52 & \textbf{66.03} \\ \bottomrule
\end{tabular}
}
\end{table}
\begin{figure*}[t]
\centering
\includegraphics[width=1.0\linewidth]{./figures/result_kalantari.pdf}
\caption{More visual comparisons between the proposed method and state-of-the-art methods~\cite{sen2012robust,hu2013hdr,kalantari2017deep,wu2018deep,yan2019attention,niu2021hdr} on Kalantari \emph{et al.}~\cite{kalantari2017deep}’s dataset.}\label{fig:result_1}
\end{figure*}
\subsection{Comparison with State-of-the-art Methods}
\subsubsection{Results on Kalantari~\emph{et al.}'s Dataset}
We first compare the results of the proposed HDR-Transformer with several state-of-the-art methods, which include two patch match based methods (Sen \emph{et al.}~\cite{sen2012robust} and Hu \emph{et al.}~\cite{hu2013hdr}) and five CNN-based methods (Kalantari \emph{et al.}~\cite{kalantari2017deep}, DeepHDR~\cite{wu2018deep}, AHDRNet~\cite{yan2019attention}, NHDRRNet~\cite{yan2020deep}, and HDR-GAN~\cite{niu2021hdr}). We also compare with a tiny version of SwinIR~\cite{liang2021swinir} as the original one fails to converge on the limited dataset. Among the deep learning-based methods, Kalantari \emph{et al.}~\cite{kalantari2017deep} adopt optical flow to align the input LDR images while DeepHDR~\cite{wu2018deep} aligns the background using homography. In contrast, the left approaches and our HDR-Transformer don't require any pre-alignment. We report the quantitative and qualitative comparison results as this testing set contains ground truth HDR images.
\mysubtitle{Quantitative results} Table~\ref{tab:results} lists the quantitative results. For the sake of fairness, the results of prior works are borrowed from HDR-GAN~\cite{niu2021hdr}, and all results are averaged over 15 testing samples from Kalantari \emph{et al.}'s dataset. Several conclusions can be drawn from Table~\ref{tab:results}. Firstly, all deep learning-based algorithms have demonstrated significant performance advantages over patch match based methods. Secondly, the pure Transformer encoder adopted in SwinIR doesn't perform well for the aforementioned reasons. Thirdly, the proposed HDR-Transformer surpasses the recently published HDR-GAN~\cite{niu2021hdr} by up to 0.6dB and 0.4dB in terms of PSNR-$l$ and PSNR-$\mu$, respectively, demonstrating the effectiveness of our method.
\mysubtitle{Qualitative results} For fair comparisons, all qualitative results are obtained using the codes provided by the authors and tonemapped using the same settings in Photomatix Pro. Fig.~\ref{fig:result_1} illustrates an intractable scene that contains saturations and large motion. The first row shows the input LDR images, our tonemapped HDR result, and the corresponding zoomed LDR patches from left to right. The second row lists the compared HDR results, where the two comparison locations are highlighted in red and blue, respectively. As can be seen, the red boxed area suffers heavy intensity variation within the three input LDR images and causes long-range saturation. Previous approaches remove the ghosting artifacts induced by slight head movements but fail to hallucinate the details of the saturation regions on the face, resulting in color distortions and inconsistent details. The blue boxed patches show a large motion region caused by the hand, patch match based methods fail to discover the correct regions, and CNN-based methods fail to handle the long-range motion, leading to ghosting artifacts in the reconstructed HDR image. On the contrary, The proposed HDR-Transformer reconstructs ghost-free results while hallucinating more visually pleasing details in these areas.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{./figures/result_sen_tursun.pdf}
\caption{Comparison results on the datasets without ground truth. Scenes are obtained from the Tursun \emph{et al.}~\cite{tursun2016objective}'s and the Sen \emph{et al.}~\cite{sen2012robust}'s datasets. Our approach generates better results in the saturated boundary and hallucinates more high-frequency details when suffering heavy intensity variation.}\label{fig:result_sen_tursun}
\end{figure}
\subsubsection{Results on the Datasets w/o Ground Truth}
To validate the generalization ability of our method, we conduct evaluations on Sen \emph{et al.}~\cite{sen2012robust}'s and Tursun \emph{et al.}~\cite{tursun2016objective}'s datasets. As illustrated in Fig.~\ref{fig:result_sen_tursun}, we report the qualitative results as both datasets have no ground truth HDR images. As seen in Fig.~\ref{fig:result_sen_tursun} (a), When suffering long-range saturation, the CNN-based algorithms AHDRNet~\cite{yan2019attention} and HDR-GAN~\cite{niu2021hdr} produce undesired distortions in saturated boundaries. The Transformer-based method SwinIR~\cite{liang2021swinir} performs better but still contains noticeable distortion as the inefficiency of local context modeling. On the contrary, the proposed HDR-Transformer generates more precise boundaries (best to compare with the corresponding LDR patches), demonstrating the context-aware modeling ability of our method. Fig.~\ref{fig:result_sen_tursun} (b) shows a scene where the piano spectrum gets saturated. Previous methods lose the high-frequency details and produce blurry results, while our approach hallucinates more details than them.
\subsubsection{Analysis of Computational Budgets}
We also compare the inference times and model parameters with previous works. As shown in Table~\ref{tab:running_time}, the patch match based methods~\cite{sen2012robust,hu2013hdr} take more than 60 seconds to fuse a 1.5MP LDR sequence. Among the CNN-based methods, Kalantari \emph{et al.}~\cite{kalantari2017deep} costs more time than the left non-flow based methods because of the time-consuming optical flow preprocess. DeepHDR~\cite{wu2018deep} and NHDRRNet~\cite{yan2020deep} consume fewer inference times but need huge amounts of parameters. AHDRNet~\cite{yan2019attention} and HDR-GAN~\cite{niu2021hdr} have a better balance of performance and efficiency by taking advantage of their well-designed architectures. In contrast, HDR-Transformer outperforms the state-of-the-art method HDR-GAN~\cite{niu2021hdr} with only half computational budgets.
\begin{table}[t]
\centering
\caption{The inference times and parameters of different methods. Part of the values are from~\cite{yan2020deep}. The `-' denotes the patch match based methods have no parameters.}
\label{tab:running_time}
\resizebox{1.0\linewidth}{!}{
\begin{tabular}{
>{\centering\arraybackslash}p{1.0cm}
>{\centering\arraybackslash}p{1.0cm}
>{\centering\arraybackslash}p{1.6cm}
>{\centering\arraybackslash}p{1.6cm}
>{\centering\arraybackslash}p{1.6cm}
>{\centering\arraybackslash}p{1.6cm}
>{\centering\arraybackslash}p{1.8cm}
>{\centering\arraybackslash}p{2.8cm}}
\toprule
\multirow{2}{*}{Method} & Sen12 & Hu13 & Kalantari17 & DeepHDR & AHDRNet & NHDRRNet & HDR-GAN & HDR-Transformer \\
& \cite{sen2012robust} & \cite{hu2013hdr} & \cite{kalantari2017deep} & \cite{wu2018deep}&\cite{yan2019attention}&\cite{yan2020deep}& \cite{niu2021hdr}& Ours \\ \midrule
Environment & CPU & CPU & CPU+GPU & GPU & GPU & GPU & GPU & GPU \\
Time(s) & 61.81s & 79.77s & 29.14s & 0.24s & 0.30s & 0.31s & 0.29s & 0.15s \\
Parameters(M) & - & - & 0.3M & 20.4M & 1.24M & 38.1M & 2.56M & 1.22M \\ \bottomrule
\end{tabular}
}
\end{table}
\subsection{Ablation Study}
\label{sec:balation_study}
To analyze the effectiveness of each component, we conduct comprehensive ablation studies on Kalantari \emph{et al.}~\cite{kalantari2017deep}'s dataset. We report the PSNR and HDR-VDP-2 scores for quantitative comparison.
\subsubsection{Ablation on the network architecture}
For the network design, we compare the proposed CA-ViT, the adopted spatial attention (SA) module, and the overall HDR-Transformer with the baseline model. Specifically, we design the following variants:
\begin{itemize}
\item \textbf{Baseline}. We take a tiny version of SwinIR~\cite{liang2021swinir}, which is constituted with vanilla Transformer encoders, as our baseline model. The baseline model keeps comparable network parameters and the same training settings as our proposed HDR-Transformer.
\item +\textbf{\ CA-ViT}. This variant replaces the vanilla Transformer encoder used in the baseline model with the proposed Context-aware Vision Transformer.
\item +\textbf{\ SA}. In this variant, we add a spatial attention (SA) module to fuse the shallow features extracted from the three input LDR images.
\item +\textbf{\ CA-ViT\ }+\textbf{\ SA}. The overall network of the proposed HDR-Transformer.
\end{itemize}
\begin{table}[t]
\caption{Quantitative results of the ablation studies. BL: the baseline model, CA-ViT: the proposed Context-aware Vision Transformer, SA: the spatial attention module, $\mathcal{L}_p$: the perceptual loss term.}
\label{tab:ablation_study}
\centering
\resizebox{0.6\linewidth}{!}{
\begin{tabular}{
>{\centering\arraybackslash}p{0.85cm}
>{\centering\arraybackslash}p{1.2cm}
>{\centering\arraybackslash}p{0.85cm}
>{\centering\arraybackslash}p{0.85cm}
>{\centering\arraybackslash}p{1.6cm}
>{\centering\arraybackslash}p{1.6cm}
>{\centering\arraybackslash}p{2.0cm}}
\toprule
BL & CA-ViT & SA & $\mathcal{L}_p$ & PSNR-$\mu$ & PSNR-$l$ & HDR-VDP-2 \\ \midrule
\checkmark & & & & 43.42 & 41.68 & 64.52 \\
\checkmark & \checkmark & & & 44.03 & 41.99 & 65.94 \\
\checkmark & & \checkmark & & 43.77 & 41.78 & 65.30 \\
\checkmark & \checkmark & \checkmark & & 44.26 & 42.09 & 65.97 \\
\checkmark & \checkmark & \checkmark & \checkmark & 44.32 & 42.18 & 66.03 \\ \bottomrule
\end{tabular}
}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth]{./figures/ablation_cavit.pdf}
\caption{Qualitative results of our ablation study on the proposed CA-ViT.}\label{fig:ablation_ctl}
\end{figure}
Table~\ref{tab:ablation_study} summarizes the quantitative results of our ablation study. The first row in Table~\ref{tab:ablation_study} shows that directly applying the Transformer to HDR deghosting does not perform well. By comparing the first four rows, several conclusions can be drawn. On the one hand, the CA-ViT and SA both improve the performance, but the benefit from CA-ViT is more significant than SA. We conclude the reasons in two folds. Firstly, the inductive biases introduced by the convolution layers in the CA-ViT or SA help the Transformer be better optimized in limited data. Moreover, by incorporating the CA-ViT into each Transformer encoder, both the global and local contexts are explored, resulting in better capabilities of long-range ghosting removal and local details reconstruction. The qualitative results in Fig.~\ref{fig:ablation_ctl} also demonstrate our conclusions. On the other hand, the performance is further improved by combining all the components, which proves the effectiveness of the HDR-Transformer's pipeline design.
\subsubsection{Ablation on losses}
\label{sec:ablation_loss}
We also conduct experiments to verify the effectiveness of the perceptual loss by training the HDR-Transformer from scratch both with and without the perceptual loss term. Comparing the last two rows in Table~\ref{tab:ablation_study}, we can see that the adopted perceptual loss improves the performance of the proposed HDR-Transformer.
\section{Conclusions}
\label{sec:conclusion}
In this paper, we have proposed a dual-branch Context-aware Vision Transformer (CA-ViT), which overcomes the lack of locality in vanilla ViTs. We have extended the standard ViTs by incorporating a local feature extractor, and therefore both global and local image contexts are modeled concurrently. Furthermore, we have introduced the HDR-Transformer, a task-specific framework for ghost-free high dynamic range imaging. The HDR-Transformer incorporates the benefits of Transformers and CNNs, where the Transformer encoder and the local context extractor are used to model the long-range ghosting artifacts and short-range pixel relationship, respectively. Extensive experiments have demonstrated that the proposed method achieves state-of-the-art performance.
\noindent\textbf{Acknowledgement} This work was supported by National Natural Science Foundation of China under grants No. (61872067, 62031009 and 61720106004).
\clearpage
\bibliographystyle{splncs04}
|
\section{Introduction}\label{sec1}
Recent developments in the biomedical, social and industrial robotics have led to an increasing interest in the field of physical human-robot collaboration (pHRC).\cite{9531394,he2020admittance}
Robots that can physically collaborate with human partners enable the cognitive strengths of human and the high-precision and repeatability of robots to be combined.
Unlike fully automated industrial production lines, to effectively work with the human partner in a shared workspace involving activities such as object handovers or co-carrying, it is imperative that the cooperative robots are capable of simultaneously complying with human forces at the contact point rather than rejecting them as external disturbances.\cite{tee2010adaptive}
Therefore, impedance/admittance control based compliance control strategies have been attracting significant interest for this application.
These include the adaptive admittance control,\cite{okunev2012human} adaptive impedance control,\cite{huang2004adaptive,tsumugiwa2002variable} and robust impedance control.\cite{chan1991robust,liu1991robust}
\begin{figure}[htb]
\centering
\includegraphics[width=6.5in]{Picture3.png}
\caption{Structure of the proposed FTISMC Admittance Control Scheme.}
\label{fig_1}
\end{figure}
Consider an industrial robot that is required to work with a human partner in an uncertain environment. The controller of the robot should be capable of tracking a certain trajectory with external uncertainties and perturbations, and at the same time, be compliant to human forces without the loss of stability.
A considerable literature exists on techniques for improving the tracking performance of robot manipulators in an uncertain environment.
For many years, model-free control strategies such as PID have been extensively employed in the industry.\cite{zuo2015non} Sliding mode control (SMC) based methods have also emerged and been widely studied in the field of robotics to provide robustness against the effects of parametric uncertainties and external disturbances.\cite{van2019adaptive} The SMC scheme involves (i) the reaching phase, in which the system converges from the initial state to the sliding surface, and (ii) the sliding phase, in which the system reaches the sliding surface and tries to stay on the equilibrium point.
However, the effectiveness of SMC is limited by the reaching phase which is sensitive to disturbances. Therefore, to overcome the limitation of SMC, integral sliding mode control (ISMC) has been developed. ISMC eliminates the reaching phase by finding a suitable initial position to reject disturbances from the beginning, which is helpful for practical applications.\cite{van2019adaptive}
Furthermore, to obtain the property of fast tracking-error convergence, a further extension of ISMC, called fixed-time integral sliding mode control (FTISMC),\cite{huang2020robust,RN446,Wang_2018} which incorporates the benefits of ISMC and fixed-time sliding mode control (FxTSMC)\cite{shifix} has been proposed. Compared with traditional finite-time control strategies,\cite{yang2018adaptive,zuo2016distributed} the fixed-time controller does not depend on the initial condition of the system with the result that the convergence time can be determined a priori based on the design parameters.\cite{zuo2016distributed}
In the work of Huang and Wang,\cite{huang2020robust} a FTISMC is proposed that stabilizes a highly complex nonlinear system within a fixed time and effectively restrains the irregular and nonlinear vibrations of the system.
In the work of Li et al.\cite{RN446} a new integral high order sliding mode controller is proposed for high-order nonlinear systems with fixed-time convergence.
In addition, the singularity problem always comes up in the sense of fixed-time stability when calculating the derivative of sliding variables.
To address this issue a number of approaches have been developed to avoid or eliminate the singularity. \cite{feng2002non,yang2011nonsingular,zuo2015non,zuo2015nonsingular,van2021robust}
In the work of Feng et al.\cite{feng2002non} a global non-singular terminal sliding mode controller is presented for second-order systems to overcome the singularity problem.
In the work of Zuo,\cite{zuo2015nonsingular} a new sliding surface is proposed to circumvent the singularity problem.
In the work of Van and Ceglarek,\cite{van2021robust} a new fault-tolerant control scheme based on a nonsingular fixed-time sliding mode controller is proposed for robot manipulators, which guarantees a global fixed-time convergence.
So far there is little literature that combines admittance control and FTISMC for physical human-robot collaboration. Motivated by the above discussion this paper proposes a new ISMC control scheme for admittance control of a robot manipulator, as depicted in Fig. \ref{fig_1}.
The stability and tracking error of the closed-loop system are converged within a fixed time. Furthermore, for practical application, a nonsingular fixed-time integral sliding mode controller is designed. The proposed method is then simulated on a two-link robot manipulator and compared with other state-of-the-art control strategies. The simulation results show that the proposed controller is superior in the sense of both tracking error and convergence time. The contributions and innovations of the proposed approach can be highlighted in a comparison with other approaches as follows:
\begin{enumerate}[1.]
\item Compared to conventional admittance control,\cite{he2020admittance} the proposed approach integrates an admittance with a FTISMC to guarantee a fixed-time convergence, while compliant with external forces. Therefore, the proposed controller provides faster transient response, lower tracking errors and better disturbance rejection capacity.
\item Compared to the existing FTISMC,\cite{huang2020robust,RN446,Wang_2018} which uses the tracking errors and the velocity of the tracking errors to reconstruct the integral sliding surface, the proposed controller develops a new nonsingular fixed-time sliding surface and uses the dynamics model of the system, which is reconstructed based on the nonsingular fixed-time sliding surface, to reconstruct a new integral sliding surface. Therefore, the proposed approach is able to avail of the full advantages of both ISMC and FxTSMC.
\end{enumerate}
The paper is organized as follows. The mathematical model describing the motion of a robot manipulator and the problem formulation are presented in Section \ref{sec2}. The proposed fixed-time ISMC scheme for admittance control is developed in Section \ref{sec3}. Simulation results demonstrating the effectiveness of the proposed controller are presented in Section \ref{sec4}. Finally, section \ref{sec5} discusses the conclusions and directions for future works.
\section{Problem formulation and preliminaries}\label{sec2}
\subsection{Problem Formulation}
We consider a robot manipulator whose joint space dynamics equation can be written in the form:
\begin{equation}
\label{eq1}
M\left( q \right) \ddot{q}+C\left( q,\dot{q} \right) \dot{q}+G\left( q \right) +F\left( \dot{q} \right) =\tau _c +\tau _e
\end{equation}
where $q$, $\dot{q}$, $\ddot{q}$ are the joint positions, velocities and accelerations of the robot manipulator, respectively. $q=\left[ q_1,q_2,...,q_N \right] ^T$ and $N$ is the number of joints. $M\left( q \right)$ is the $N \times N$ mass matrix, $C\left( q,\dot{q} \right)$ is the $N \times N$ Coriolis and centrifugal forces matrix, $G\left( q \right)$ is the $N \times 1$ gravity vector, and $F\left( \dot{q} \right)$ is the $N \times 1$ vector of frictional forces. $\tau_c$ denotes the control torque acting at the joints, and $\tau_e$ is the external torque from the interaction with the human partner. $F\left( \dot{q} \right)$ and $\tau_e$ together form the time-varying lumped uncertainties of the system. Fig. \ref{fig_2} above shows the structure of a classical two-link robot manipulator interacting physically with a human.
\begin{figure}
\centering
\includegraphics[width=3.9in]{Picture1.png}
\caption{Two-Link robot manipulator.}
\label{fig_2}
\end{figure}
Since it is easier to formulate the trajectory in Cartesian space, we transfer the joint space dynamics of the robot manipulator (\ref{eq1}) into Cartesian space:
\begin{equation}
\label{eq2}
M_x\ddot{x}+C_x\dot{x}+G_x+F_x=f_c+f_e
\end{equation}
where $x=\left[ x_1,x_{2},...,x_N \right]$ is the position of the joints in Cartesian space, $M_x$, $C_x$, $G_x$ and $F_x$ are coefficient matrices in Cartesian space, $f_c$ is the control force, and $f_e$ is the external force. The joint space velocities and Cartesian space velocities are related as:
\begin{equation}
\label{eq3}
\dot{x}=J\left( q \right) \dot{q}
\end{equation}
where $J\left( q \right)$ is the Jacobian of the robot manipulator. Inserting (\ref{eq3}) into (\ref{eq1}), the coefficient matrices in Cartesian space can be written as:
$$
M_x=J^{-T}\left( q \right) M\left( q \right) J^{-1}\left( q \right)
$$
$$
C_x=J ^{-T}\left( q \right)\left( C\left( q,\dot{q} \right) -M\left( q \right) J^{-1}\left( q \right) \dot{J}\left( q \right) \right) J^{-1}\left( q \right)
$$
$$
G_x=J^{-T}\left( q \right) G\left( q \right) ,\ F_x=J^{-T}\left( q \right) F\left( \dot{q} \right)
$$
$$
f_c=J^{-T}\left( q \right) \tau _c,\ \ f_e=J^{-T}\left( q \right) \tau _e
$$
Letting $\eta _1=x$ and $\eta _2=\dot{x}$, the Cartesian space model of the robot can be represented as:
\begin{equation}
\label{eq4}
\left\{ \begin{array}{l}
\dot{\eta}_1=\eta _2\\
\dot{\eta}_2=M_x^{-1}\left( -C_x\dot{x}-G_x-F_x+f_e+f_c \right)\\
\end{array} \right.
\end{equation}
For ease of control design, let $\varXi =M_x^{-1}$, $\varGamma \left( x,\dot{x} \right) =M_x^{-1}\left( -C_x-G_x \right)$, $\varPsi \left( x,\dot{x},t \right) =M_x^{-1}\left( -F_x+f_e \right)$, and $u=f_c$, such that the Cartesian space model can be further written as:
\begin{equation}
\label{eq5}
\left\{ \begin{array}{l}
\dot{\eta}_1=\eta _2\\
\dot{\eta}_2=\varXi u+\varGamma \left( x,\dot{x} \right) +\varPsi \left( x,\dot{x},t \right)\\
\end{array} \right.
\end{equation}
\textbf{\emph{Assumption 1}}: There exists a positive constants $\rho$, such that the unknown lumped uncertainties term is bounded by:
\begin{equation}
\label{eq6}
\varPsi \left( x,\dot{x},t \right) =M_x^{-1}\left( -F_x+f_e \right)\le \rho
\end{equation}
where $\rho$ is a positive constant. $\varPsi \left( x,\dot{x},t \right)=[\varPsi_i\left( x,\dot{x},t \right), ..., \varPsi_N \left( x,\dot{x},t \right)]^T$. This assumption means that the lumped uncertainties such as joint errors, friction and external human forces are bounded, which is usually satisfied in real application scenarios.
The control objective of this paper is to develop a fixed-time ISMC scheme to make the joint position $x$ track a certain trajectory $x_r$ which is generated from admittance control. Therefore, the robot manipulator can comply with human forces rather than reject them as external disturbances. Furthermore, we require that the tracking error converges within an arbitrarily small interval of time, regardless of the initial conditions.
\subsection{Preliminaries}
Consider a time-varying differential equation:
\begin{equation}
\label{eq7}
\dot{x}=f\left( x,t \right)
\end{equation}
where $x\in R^n$, and $f:\ R^n\rightarrow R^n$ is a continuous nonlinear function. Assume that the origin of (\ref{eq7}) is the stable equilibrium point and $f\left(0\right)=0$.
\textbf{\emph{Definition 1}}\cite{zuo2015non}: The system (\ref{eq7}) is finite time stable if the origin is Lyapunov stable and any solution $x\left(t\right)$ starting from $x_0$ satisfies $\lim _{t\rightarrow T\left( x_0 \right)}x\left( t \right) \rightarrow 0$ and $x\left( t \right) =0,\forall t >T\left( x_0 \right)
$, where $T\left( x_0 \right)$ represents the settling time of the system.
\textbf{\emph{Definition 2}}\cite{zuo2015non}: The system (\ref{eq7}) is fixed-time stable if it is globally finite time stable and its settling time $T\left( x_0 \right)$ satisfies $T\left( x_0 \right) <T_{\max}$, where $T_{\max}$ is a positive number.
\textbf{\emph{Lemma 1}} \cite{zuo2016distributed}: The following inequality holds for any real number that satisfies $x_1,x_2,...,x_n>0$ and $0<k<1$:
\begin{equation}
\label{eq8}
\sum_{i=1}^n{ x_i ^{k+1}\ge \left( \sum_{i=1}^n{ x_i ^2} \right)}^{\frac{k+1}{2}}
\end{equation}
\textbf{\emph{Lemma 2}} \cite{zuo2016distributed}: The following inequality holds for any real number that satisfies $x_1,x_2,...,x_n>0$ and $k>1$:
\begin{equation}
\label{eq9}
\sum_{i=1}^n{ x_i ^k\ge n^{1-k}\left( \sum_{i=1}^n{ x_i } \right)}^k
\end{equation}
\textbf{\emph{Lemma 3}} \cite{zuo2016distributed}: Consider a scalar differential system below:
\begin{equation}
\label{eq9_1}
\dot{y}=\lambda _1\left[ y \right] ^{\alpha}+\lambda _2\left[ y \right] ^{\beta},\ y\left( 0 \right) =y_0
\end{equation}
where $\lambda _1$, $\lambda _2$, $\alpha$, $\beta$ are all positive real number satisfying $\lambda _1, \lambda _2>0$ , $0<\alpha<1$ and $\beta>1$. Then the system (\ref{eq9_1}) is fixed-time stable and the convergence time is independent with respect to the initial state of the system and upper bounded by the design parameters as:
\begin{equation}
\label{eq9_2}
T\left( x_0 \right) <T_{\max}=\frac{1}{\lambda _1\left( 1-\alpha \right)}+\frac{1}{\lambda _2\left(\beta -1 \right)}
\end{equation}
\section{Control design and stability analysis}\label{sec3}
\subsection{Admittance Trajectory Shaping}
To comply with external human forces, the system at the contact point can be modeled as a mass-spring-damper system (shown in Fig. \ref{fig_3}). The objective of modeling the system with virtual mass, spring and damper is to make sure the interaction forces are elastic and never vibrate at the contact point.
\begin{figure}[h]
\centering
\includegraphics[width=2.2in]{Picture2.png}
\caption{ The mass-spring-damper system.}
\label{fig_3}
\end{figure}
Let $\xi _i=x_{r_i}-x_{d_i}, i=1,2,...,N$, where the $x_{r_i}$ is the reference trajectory generated from admittance control and the $x_{d_i}$ is the initial desired trajectory of the robot manipulator. The dynamics for a robot manipulator rendering an impedance can be written as:
\begin{equation}
\label{eq10}
k_{m_i}\ddot{\xi}_i+k_{b_i}\dot{\xi}_i+k_{k_i}\xi _i=f_{e_i}
\end{equation}
where $k_{m_i}$, $k_{b_i}$ and $k_{k_i}$ are mass, spring and damping coefficient. $f_e$ is the external human force. By integrating the impedance equation (\ref{eq10}), we obtain $x_{r_i}$ which will be tracked by the controller.
\subsection{Design of Fixed-time ISMC}
Let the tracking error be written as:
\begin{equation}
e=x-x_r
\end{equation}
where $x_r$ is the reference trajectory generated from the admittance control, and $x$ is the real trajectory of robot. The derivative of tracking error $e$ is $\dot{e}=\dot{x}-\dot{x}_r$. To obtain a fixed-time integral sliding mode controller, the error is reconstructed based on Lemma 3:
\begin{equation}
\label{eq12}
s=\dot{e}+k_1\left[ e \right] ^{\alpha}+k_2\left[ e \right] ^{\beta}
\end{equation}
where $k _1$, $k _2$, $\alpha$, $\beta$ are all positive constants satisfying $k _1, k _2>0$ , $0<\alpha<1$ and $\beta>1$. When the sliding surface $s$ converges to zero, (\ref{eq12}) can be rewritten as:
\begin{equation}
\label{eq12_1}
\dot{e}=-k_1\left[ e \right] ^{\alpha}-k_2\left[ e \right] ^{\beta}
\end{equation}
According to Lemma 3, the convergence time of (\ref{eq12_1}) is bounded by:
\begin{equation}
\label{eq12_2}
T_{s1}\le \frac{1}{k_1\left( 1-\alpha \right)}+\frac{1}{k_2\left( \beta -1 \right)}
\end{equation}
Differentiating (\ref{eq12}) with respect to time and using (\ref{eq5}), we have:
\begin{equation}
\label{eq13}
\begin{aligned}
\dot{s}=\ddot{x}-\ddot{x}_r+k_1\alpha \left[ e \right] ^{\alpha -1}\dot{e}+k_2\beta \left[ e \right] ^{\beta -1}\dot{e}
=\varXi u+\varGamma\left(x,\dot{x}\right) +\varPsi\left(x,\dot{x},t\right) -\ddot{x}_r+k_1\alpha \left[ e \right] ^{\alpha -1}\dot{e}+k_2\beta \left[ e \right] ^{\beta -1}\dot{e}
\end{aligned}
\end{equation}
Based on (\ref{eq12}), the integral sliding surface is selected as:
\begin{equation}
\label{eq14}
\begin{aligned}
\sigma_1 \left( t \right) =s\left( t \right) -s\left( 0 \right)
-\int\limits_0^t{\left( \varXi u_0+\varGamma\left(x,\dot{x}\right) -\ddot{x}_r+k_1\alpha \left[ e \right] ^{\alpha -1}\dot{e}+k_2\beta \left[ e \right] ^{\beta -1}\dot{e} \right) dt}
\end{aligned}
\end{equation}
where term $s\left(0\right)$ represents the value of the sliding surface when $t=0$, which is added to eliminate the initial error. $u_0$ is the output of the nominal controller, which is designed for the non-disturbance system by another method (e.g., PID, CTC, etc.) and will be presented in the Section 3.4. The derivative of the integral sliding surface can be further written as:
\begin{equation}
\label{eq15}
\begin{aligned}
\dot{\sigma_1}\left( t \right) =\left( \varXi u+\varGamma\left(x,\dot{x}\right) +\varPsi\left(x,\dot{x},t\right) -\ddot{x}_r+k_1\alpha \left[ e \right] ^{\alpha -1}\dot{e}+k_2\beta \left[ e \right] ^{\beta -1}\dot{e} \right)
-\left( \varXi u_0+\varGamma\left(x,\dot{x}\right) -\ddot{x}_r+k_1\alpha \left[ e \right] ^{\alpha -1}\dot{e}+k_2\beta \left[ e \right] ^{\beta -1}\dot{e} \right)
\end{aligned}
\end{equation}
Therefore, we have:
$$\dot{\sigma_1}\left( t \right)=\varXi \left( u-u_0 \right) +\varPsi \left( x,\dot{x},t \right) $$
Letting $u=u_0+u_{s1}$, where $u_{s1}$ is designed to compensate for disturbances based on $u_0$, we obtain:
\begin{equation}
\label{eq16}
\dot{\sigma_1}\left( t \right) =\varXi u_{s1}+\varPsi \left( x,\dot{x},t \right)
\end{equation}
From (\ref{eq16}), the proposed compensating controller $u_{s1}$ is designed as:
\begin{equation}
\label{eq17}
u_{s1}=\varXi ^{-1}\left( -\left( \rho +\varepsilon \right) sign\left( \sigma_1 \right) - k_3\left[ \sigma_1 \right] ^p-k_4\left[ \sigma_1 \right] ^q \right)
\end{equation}
where the $k_3,k_4>0$, $0<p<1$ and $q>1$. Inserting the compensating controller $u_{s1}$ into (\ref{eq16}), yields:
\begin{equation}
\label{eq18}
\dot{\sigma_1}\left( t \right) =\varPsi\left( x,\dot{x},t \right) -\left( \rho +\varepsilon \right) sign\left( \sigma_1 \right) -k_3\left[ \sigma_1 \right] ^p-k_4\left[ \sigma_1 \right] ^q
\end{equation}
and
\begin{equation}
\label{eq182}
\dot{\sigma_{1 _{i}}}\left( t \right) =\varPsi_i\left( x,\dot{x},t \right) -\left( \rho +\varepsilon \right) sign\left( \sigma_{1 _{i}} \right) -k_3\left[ \sigma_{1 _{i}} \right] ^p-k_4\left[ \sigma_{1 _{i}} \right] ^q
\end{equation}
Consider a Lyapunov function candidate:
\begin{equation}
\label{eq19}
V_1=\frac{1}{2}\sum_{i=1}^N{\sigma_{1 _{i}}^{2}}
\end{equation}
Differentiating the Lyapunov function (\ref{eq19}), we have:
\begin{equation}
\label{eq20}
\begin{aligned}
\dot{V_1}&=\sum_{i=1}^N{\sigma_{1 _{i}}\dot{\sigma_{1 _{i}}}} \\
&=\sum_{i=1}^N{\sigma_{1 _{i}}\left( \varPsi_i\left( x,\dot{x},t \right) -\left( \rho +\varepsilon \right) sign\left( \sigma_{1 _{i}} \right) -k_3\left[ \sigma_{1 _{i}} \right] ^p-k_4\left[ \sigma_{1 _{i}} \right] ^q \right)} \\
&=\sum_{i=1}^N{\left( \sigma _{1 _{i}}\varPsi_i\left( x,\dot{x},t \right) -\left( \rho +\varepsilon \right) \left| \sigma_{1 _{i}} \right|-k_3\left[ \sigma_{1 _{i}} \right] ^{p+1}-k_4\left[ \sigma_{1 _{i}} \right] ^{q+1} \right)}
\end{aligned}
\end{equation}
By using Assumption 1, we have:
\begin{equation}
\label{eq20_1}
\begin{aligned}
\dot{V_1}\le \sum_{i=1}^N{\left( -k_3\left[ \sigma_{1 _{i}} \right] ^{p+1}-k_4\left[ \sigma_{1 _{i}} \right] ^{q+1} \right)}
\end{aligned}
\end{equation}
By using Lemma 1 and Lemma 2, we have:
\begin{equation}
\label{eq21}
\begin{aligned}
\dot{V_1}&\le -k_3\sum_{i=1}^N{\left| \sigma _{1 _{i}} \right|^{p+1}-}k_4\sum_{i=1}^N{\left| \sigma _{1 _{i}} \right|^{q+1}}\\
&\le -k_3\sum_{i=1}^N{\left( \left| \sigma _{1 _{i}} \right|^2 \right) ^{\frac{p+1}{2}}}-k_4\sum_{i=1}^N{\left( \left| \sigma _{1 _{i}} \right|^2 \right) ^{\frac{q+1}{2}}}\\
&\le -k_3\left( \sum_{i=1}^N{\left| \sigma _{1 _{i}} \right|^2} \right) ^{\frac{p+1}{2}}-k_4N^{\frac{1-q}{2}}\left( \sum_{i=1}^N{\left| \sigma _{1 _{i}} \right|^2} \right) ^{\frac{q+1}{2}}\\
&=-k_32^{\frac{p+1}{2}}\left( V_1 \right) ^{\frac{p+1}{2}}-k_42^{\frac{q+1}{2}}N^{\frac{1-q}{2}}\left( V_1 \right) ^{\frac{q+1}{2}}
\end{aligned}
\end{equation}
Based on the Lemma 3, the above system is fixed-time stable and the convergence time is upper bounded as:
\begin{equation}
\label{eq22}
T_{r1}\le \frac{1}{2^{\frac{p+1}{2}}k_3\left( 1-\frac{p+1}{2} \right)}+\frac{1}{2^{\frac{q+1}{2}}k_4N^{\frac{1-q}{2}}\left( \frac{q+1}{2}-1 \right)}
\end{equation}
\noindent\textbf{\emph{Theorem 1}}. The system (\ref{eq5}) with controller $u_s$ is fixed-time stable and the settling time is bounded by:
\begin{equation}
\label{eq22_1}
T\le T_{s1} + T_{r1}+T_{n}
\end{equation}
where $T_{s1}$ and $T_{r1}$ represent the convergence time of the sliding variable $s$ and the system (\ref{eq5}) with $u_{s1}$ as the controller, respectively. $T_n$ represents the convergence time of the system (\ref{eq5}) with $u_0$ as the nominal controller, which will be determined in Section 3.4.
\noindent\textbf{\emph{Proof}}. It is clear that the convergence time of the entire system is smaller than the sum of the convergence times of each component (i.e., the convergence times of the sliding surface, compensating controller and nominal controller). This completes the proof.
\begin{remark}
The value for $\rho$ of the controller in (\ref{eq17}) was selected based on \textbf{Assumption 1}. In practice, we can use experiments to determine the bounded value of the uncertainty $\rho$. Another method which can be applied to approximate the value of parameter of $\rho$ is to use an adaptive technique; for example, we can use an adaptive second-order super-twisting method \cite{van2021robust} to estimate the adaptive gain and eliminate the chattering at the same time.
\end{remark}
\begin{remark}
The use of the `sign' function in the controller (\ref{eq17}) will generate a chattering. To eliminate the chattering, a boundary layer method or second-order super-twisting method can be employed \cite{van2021robust}.
\end{remark}
\subsection{Nonsingular Problem}
From (\ref{eq13}), when $e=0$ and $\dot{e}\ne 0$ the term $k_1\alpha \left[ e \right] ^{\alpha -1}\dot{e}$ results in a singular problem. To avoid the singularity, the error can be reconstructed as:
\begin{equation}
\label{eq23}
s = e+\frac{1}{k_{5}^{m}}\left[ \dot{e}+k_6\left[ e \right] ^n \right] ^{\frac{1}{m}}
\end{equation}
where $k _5$, $k _6$, $m$, $n$ are all positive constants satisfying $k _5, k _6>0$ , $0<m<1$ and $n>1$. When the sliding surface variable converges to zero, (\ref{eq23}) is equal to (\ref{eq12}). Therefore, according to the Lemma 3, the convergence time of (\ref{eq23}) is bounded by:
\begin{equation}
\label{eq23_1}
T_{s2}\le \frac{1}{k_5\left( 1-m \right)}+\frac{1}{k_6\left( n-1 \right)}
\end{equation}
Differentiating (\ref{eq23}) with respect to time, we have:
\begin{equation}
\label{eq24}
\dot{s}=\dot{e}+\frac{1}{mk_{5}^{m}}\left[ \dot{e}+k_6\left[ e \right] ^n \right] ^{\frac{1}{m}-1}\left( \ddot{e}+k_6n\left[ e \right] ^{n-1} \dot{e}\right)
\end{equation}
The integral sliding surface is selected as:
\begin{equation}
\label{eq25}
\begin{aligned}
\sigma_2 \left( t \right) =s\left( t \right) -s\left( 0 \right)
-\int_0^t{\left(\dot{e}+\frac{1}{mk_{5}^{m}}\left[ \dot{e}+k_6\left[ e \right] ^n \right] ^{\frac{1}{m}-1}\left( \ddot{e}+k_6n\left[ e \right] ^{n-1} \right) \dot{e} \right)dt}
\end{aligned}
\end{equation}
According to the (\ref{eq5}), we have:
\begin{equation}
\label{eq25_1}
\ddot{e}=\ddot{x}-\ddot{x_r}=\varXi u+\varGamma \left( x,\dot{x} \right)-\ddot{x_r}
\end{equation}
Inserting (\ref{eq25_1}) into (\ref{eq25}) and letting:
\begin{equation}
\label{eq26}
\begin{aligned}
&T_1\left( e,\dot{e} \right) =\frac{1}{mk_{5}^{m}}\left[ \dot{e}+k_6\left[ e \right] ^n \right] ^{\frac{1}{m}-1}\\
&T_2\left( e \right) = k_6n\left[ e \right] ^{n-1} \dot{e}
\end{aligned}
\end{equation}
The derivative of integral sliding surface can be written as:
\begin{equation}
\begin{aligned}
\dot{\sigma_2}\left( t \right) =\left(\dot{e}+T_1\left( e,\dot{e} \right)\left( \varXi u+\varGamma\left( x,\dot{x} \right) +\varPsi\left( x,\dot{x},t \right)-\ddot{x}_r+T_2\left( e \right) \right)\right)
-\left(\dot{e}+T_1\left( e,\dot{e} \right)\left( \varXi u_0+\varGamma\left( x,\dot{x} \right) -\ddot{x}_r+T_2\left( e \right) \right)\right)
\end{aligned}
\end{equation}
Setting $u=u_0+u_{s2}$, we have:
\begin{equation}
\label{eq27}
\dot{\sigma_2}\left( t \right) =T_1\left( e,\dot{e} \right)\left(\varXi u_{s2}+\varPsi \left( x,\dot{x},t \right) \right)
\end{equation}
Again, the controller $u_{s2}$ can be designed as:
\begin{equation}
u_{s2}=\varXi ^{-1}\left( -\left( \rho +\varepsilon \right) sign\left( \sigma_2 \right) -k_3\left[ \sigma_2 \right] ^p-k_4\left[ \sigma_2 \right] ^q \right)
\end{equation}
Inserting the compensating controller into (\ref{eq27}), we have:
\begin{equation}
\label{eq28}
\dot{\sigma_2}\left( t \right) =T_1\left( e,\dot{e} \right)\left(\varPsi\left( x,\dot{x},t \right) -\left( \rho +\varepsilon \right) sign\left( \sigma_2 \right) -k_3\left[ \sigma_2 \right] ^p-k_4\left[ \sigma_2 \right] ^q\right)
\end{equation}
Considering a Lyapunov function candidate:
\begin{equation}
\label{eq28_1}
V_2=\frac{1}{2}\sum_{i=1}^N{\sigma_{2 _{i}}^{2}}
\end{equation}
Differentiating the Lyapunov function (\ref{eq19}), we have:
\begin{equation}
\label{eq28_2}
\begin{aligned}
\dot{V_2}&=\sum_{i=1}^N{\sigma_{2 _{i}}\dot{\sigma_{2 _{i}}}} \\
&=\sum_{i=1}^N{\sigma_{2 _{i}}T_1\left( e,\dot{e} \right)\left( \varPsi_i\left( x,\dot{x},t \right) -\left( \rho +\varepsilon \right) sign\left( \sigma_{2 _{i}} \right) -k_3\left[ \sigma_{2 _{i}} \right] ^p-k_4\left[ \sigma_{2 _{i}} \right] ^q \right)} \\
&=T_{1}\left( e,\dot{e} \right)\sum_{i=1}^N{\left( \sigma _{2 _{i}}\varPsi_i\left( x,\dot{x},t \right) -\left( \rho +\varepsilon \right) \left| \sigma_{2 _{i}} \right|-k_3\left[ \sigma_{2 _{i}} \right] ^{p+1}-k_4\left[ \sigma_{2 _{i}} \right] ^{q+1} \right)}\\
&\le T_{1}\left( e,\dot{e} \right)\sum_{i=1}^N{\left( -k_3\left[ \sigma_{2 _{i}} \right] ^{p+1}-k_4\left[ \sigma_{2 _{i}} \right] ^{q+1} \right)}\\
&\le T_{1}\left( e,\dot{e} \right)\left(-k_32^{\frac{p+1}{2}}\left( V_2 \right) ^{\frac{p+1}{2}}-k_42^{\frac{q+1}{2}}N^{\frac{1-q}{2}}\left( V_2 \right) ^{\frac{q+1}{2}}\right)
\end{aligned}
\end{equation}
\textbf{\emph{Theorem 2}}. The system (\ref{eq5}) with controller $u_{s2}$ is fixed-time stable and its settling time is bounded by:
\begin{equation}
\label{eq28_6}
T\le T_{s2} + T_{r2}+T_{n}+\epsilon \left( \tau \right)
\end{equation}
where $T_{s2}$ and $T_{r2}$ represent the convergence time of sliding variable $s$ and the system (\ref{eq5}) with $u_{s2}$ as the controller, respectively. $T_n$ represent the convergence time of the system (\ref{eq5}) with $u_0$ as the nominal controller.
\textbf{\emph{Proof}}. From (\ref{eq28}), we can see the fixed-time stability depends both on the term $\left(-k_32^{\frac{p+1}{2}}\left( V_2 \right)^{\frac{p+1}{2}}-k_42^{\frac{q+1}{2}}N^{\frac{1-q}{2}}\left( V_2 \right) ^{\frac{q+1}{2}}\right)$ and the term $T_1\left( e,\dot{e} \right)$. According to the work of Van and Ceglarek \cite{van2021robust}, for the case of $\dot{e}+k_6\left[ e \right] ^n\ne 0 $, we divide the state space into two different areas $\varOmega _1=\left\{ \left( e,\dot{e} \right) \left| T\left( e,\dot{e} \right) \ge 1 \right. \right\}$ and $\varOmega _2=\left\{ \left( e,\dot{e} \right) \left| T\left( e,\dot{e} \right) <1 \right. \right\}$.
i. When the system states are in the area of $\varOmega _1$, we have:
\begin{equation}
\label{eq28_4}
\begin{aligned}
\dot{V_2}\le-k_32^{\frac{p+1}{2}}\left( V_2 \right) ^{\frac{p+1}{2}}-k_42^{\frac{q+1}{2}}N^{\frac{1-q}{2}}\left( V_2 \right) ^{\frac{q+1}{2}}
\end{aligned}
\end{equation}
$V_2=0$ implies the sliding surface $s=0$. Therefore, according to the Lemma 3, the system states will reach the sliding surface $s=0$ within a fixed time, which is upper bounded as:
\begin{equation}
\label{eq28_5}
T_{r2}\le \frac{1}{2^{\frac{p+1}{2}}k_3\left( 1-\frac{p+1}{2} \right)}+\frac{1}{2^{\frac{q+1}{2}}k_4N^{\frac{1-q}{2}}\left( \frac{q+1}{2}-1 \right)}
\end{equation}
ii. When the system states are in the area of $\varOmega _2$, according to (\ref{eq28_2}), the sliding surface $s=0$ is still an attractor. In addition, for the case of $\dot{e}+k_6\left[ e \right] ^n = 0 $, for a given $\tau$, there exists a positive constant $\epsilon \left( \tau \right)$ such that the sliding surface $s_2=0$ can be reached from anywhere in the phase plane within a fixed time $t_r<T_{r2}+\epsilon \left( \tau \right) $. \cite{Li2017OnSC} Therefore, the total setting time $T$ is bounded as (\ref{eq28_6}). This completes the proof.
\begin{remark}
Compared to the design of the integral sliding surfaces of the existing FTISMC \cite{huang2020robust,RN446,Wang_2018}, the proposed controller uses the sliding surface in (\ref{eq25}), which takes the dynamics of the system account in the design of the integral sliding surface. This approach provides two advantages: (1) it allows us to design the nominal controller $u_0$ and the reaching controller $u_s$ separately. This facilitates the design procedure and preserves the advantages of both the nominal controller and ISMC.
\end{remark}
\subsection{Design of the Nominal Controller}
The nominal controller is designed to stabilize the system neglecting the lumped uncertainties. To stabilize the whole system within a fixed time, we choose fixed-time backstepping control \cite{Mavriplis2003} as the nominal controller. The dynamics of the robot manipulator without lumped uncertainties can be written as:
\begin{equation}
\left\{ \begin{array}{l}
\dot{\eta}_1=\eta _2\\
\dot{\eta}_2=\varXi u+\varGamma \left( x, \dot{x} \right)\\
\end{array} \right.
\end{equation}
Letting $s_1=\eta _1-x_r$, then, $\dot{s}_1=\eta_2-\dot{x_r}$, and we can design the stabilizing function as:
\begin{equation}
\alpha _s=-\left( \lambda _1s_1+\lambda _2s_{1}^{\alpha}+\lambda _3s_{1}^{\beta} \right) +\dot{x}_r
\end{equation}
where $\lambda _1$, $\lambda _2$, $\lambda _3$, $\alpha$, $\beta$ are all positive constants satisfying $0<\alpha<1$ and $\beta>1$. Thus, $\dot{s}_1$ can be written as:
\begin{equation}
\dot{s}_1=-\left( \lambda _1s_1+\lambda _2s_{1}^{\alpha}+\lambda _3s_{1}^{\beta} \right)
\end{equation}
Selecting the Lyapunov function candidate as:
\begin{equation}
V_3=\frac{1}{2}s_{1}^{T}s_1
\end{equation}
The derivative of the Lyapunov function $V_3$ is:
\begin{equation}
\begin{aligned}
\dot{V}_3&=s_{1}^{T}\dot{s}_1\\
&=-s_{1}^{T}\left( \lambda _1s_1+\lambda _2s_{1}^{\alpha}+\lambda _3s_{1}^{\beta} \right) \\
&=-\lambda _1s_{1}^{T}s_1-\lambda _2\left( s_{1}^{T}s_1 \right) ^{\frac{\alpha +1}{2}}-\lambda _3\left( s_{1}^{T}s_1 \right) ^{\frac{\beta +1}{2}}\\
&\le -\lambda _2\left( s_{1}^{T}s_1 \right) ^{\frac{\alpha +1}{2}}-\lambda _3\left( s_{1}^{T}s_1 \right) ^{\frac{\beta +1}{2}}\\
&\le -2^{\frac{\alpha +1}{2}}\lambda _2\left( \frac{1}{2}s_{1}^{T}s_1 \right) ^{\frac{\alpha +1}{2}}-2^{\frac{\beta +1}{2}}\lambda _3\left( \frac{1}{2}s_{1}^{T}s_1 \right) ^{\frac{\beta +1}{2}}\\
&=-2^{\frac{\alpha +1}{2}}\lambda _2\left( V_3 \right) ^{\frac{\alpha +1}{2}}-2^{\frac{\beta +1}{2}}\lambda _3\left( V_3 \right) ^{\frac{\beta +1}{2}}
\end{aligned}
\end{equation}
Letting $s_2=\eta _2-\alpha _s $, we have:
\begin{equation}
\label{eq29_1}
\dot{s}_2=\dot{\eta}_2-\dot{\alpha}_s=\varXi u_0-\varGamma \left( x,\dot{x} \right) -\dot{\alpha}_s
\end{equation}
Selecting the Lyapunov function candidate as:
\begin{equation}
V_4=\frac{1}{2}s_{2}^{T}s_2
\end{equation}
To obtain the property of fixed-time convergence, we design the nominal controller as:
\begin{equation}
\label{eq30}
u_0=\varXi ^{-1}\left( -\varGamma \left( x,\dot{x} \right) +\dot{\alpha}_s-\lambda _1s_2-\lambda _2s_{2}^{\alpha}-\lambda _3s_{2}^{\beta} \right)
\end{equation}
Inserting (\ref{eq30}) into (\ref{eq29_1}), we have:
\begin{equation}
\dot{s}_2=-\lambda _1s_2-\lambda _2s_{2}^{\alpha}-\lambda _3s_{2}^{\beta}
\end{equation}
The derivative of the Lyapunov function $V_2$ is:
\begin{equation}
\begin{aligned}
\dot{V}_4&=s_{2}^{T}\dot{s}_2\\
&=s_{2}^{T}\left( -\lambda _1s_2-\lambda _2s_{2}^{\alpha}-\lambda _3s_{2}^{\beta} \right) \\
&\le -2^{\frac{\alpha +1}{2}}\lambda _2\left( V_4 \right) ^{\frac{\alpha +1}{2}}-2^{\frac{\beta +1}{2}}\lambda _3\left( V_4 \right) ^{\frac{\beta +1}{2}}
\end{aligned}
\end{equation}
Choosing the Lyapunov function candidate of system as:
\begin{equation}
V_n=V_3+V_4
\end{equation}
The derivative of Lyapunov function $V$ is:
\begin{equation}
\begin{aligned}
\dot{V_n}&\le-2^{\frac{\alpha +1}{2}}\lambda _2\left( V_3 \right) ^{\frac{\alpha +1}{2}}-2^{\frac{\beta +1}{2}}\lambda _3\left( V_3 \right) ^{\frac{\beta +1}{2}}-2^{\frac{\alpha +1}{2}}\lambda _2\left( V_4 \right) ^{\frac{\alpha +1}{2}}-2^{\frac{\beta +1}{2}}\lambda _3\left( V_4 \right) ^{\frac{\beta +1}{2}}\\
&\le -2^{\frac{\alpha +1}{2}}\lambda _2\left( V_n \right) ^{\frac{\alpha +1}{2}}-2^{\frac{\beta +1}{2}}\lambda _3\left( V_n \right) ^{\frac{\beta +1}{2}}
\end{aligned}
\end{equation}
According to the Lemma 3, when applying the proposed nominal controller $u_0$, the system without lumped uncertainties is fixed-time stable and the convergence time is bounded as:
\begin{equation}
T_n\le \frac{2}{\lambda _22^{\frac{\alpha +1}{2}}\left( 1-\alpha \right)}+\frac{2}{\lambda _32^{\frac{\beta +1}{2}}\left( \beta -1 \right)}
\end{equation}
\begin{remark}
It is noted that the nominal controller can be designed using other controllers, e.g, PID, CTC and Backstepping. However, in this paper, we use the fixed-time backstepping controller to achieve global fixed-time convergence for the system.
\end{remark}
\section{Simulation and results}\label{sec4}
In the simulation, we consider a two-link robot in the horizontal plane \cite{craig1987adaptive}. The dynamics of the robot are described as:
\begin{equation}
\label{eq31}
\begin{aligned}
\tau _1&=m_2l_{2}^{2}\left( \ddot{q}_1+\ddot{q}_2 \right) +m_2l_1l_2c_2\left( 2\ddot{q}_1+\ddot{q}_2 \right) +\left( m_1+m_2 \right) l_{1}^{2}\ddot{q}_1
-m_2l_1l_2s_2\dot{q}_{2}^{2}-2m_2l_1l_2s_2\dot{q}_1\dot{q}_2+m_2l_2gc_{12}+\left( m_1+m_2 \right) l_1gc_1\\
\tau _2&=m_2l_1l_2c_2\ddot{q}_1+m_2l_1l_2s_2\dot{q}_{1}^{2}+m_2l_2gc_{12}+m_2l_{2}^{2}\left( \ddot{q}_1+\ddot{q}_2 \right)
\end{aligned}
\end{equation}
where $c_i=\cos \left( q_i \right) $, $c_{ij}=\cos \left( q_i+q_j \right) $, $s_i=\sin \left( q_i \right) $, and $s_{ij}=\sin \left( q_i+q_j \right)$, $i,j=1,2$. The value of masses are $m_1=1.5kg$, and $m_2=1.0kg$. The length of the links are $l_1=l_2=0.3m$. Based on the (\ref{eq1}), the mass matrix $M\left( q \right)$, Coriolis and centrifugal forces matrix $C\left( q,\dot{q} \right)$ and the gravity matrix $G\left( q \right)$ are given as:
\begin{equation}
\label{eq32}
M\left( q \right) =\left[ \begin{matrix}
m_2l_{2}^{2}+2m_2l_1l_2c_2+\left( m_1+m_2 \right) l_{1}^{2}& m_2l_{2}^{2}+m_2l_1l_2c_2\\
m_2l_{2}^{2}+m_2l_1l_2c_2& m_2l_{2}^{2}\\
\end{matrix} \right]
\end{equation}
\begin{equation}
\label{eq33}
C\left( q,\dot{q} \right) =\left[ \begin{matrix}
-2m_2l_1l_2s_2\dot{q}_2& -m_2l_1l_2s_2\\
m_2l_1l_2s_2\dot{q}_1& 0\\
\end{matrix} \right]
\end{equation}
\begin{equation}
\label{eq34}
G\left( q \right) =\left[ \begin{array}{c}
m_2l_2gc_{12}+\left( m_1+m_2 \right) l_1gc_1\\
m_2l_2gc_{12}\\
\end{array} \right]
\end{equation}
The bounded disturbance term is given as:
\begin{equation}
\label{eq35}
F\left( q \right) =\left[ \begin{array}{c}
2c_1s_2+5c_{1}^{2}\\
-2c_1s_2-5c_{1}^{2}\\
\end{array} \right]
\end{equation}
The Jacobian of the robot manipulator is given as:
\begin{equation}
\label{eq35_1}
J\left( q \right) =\left[ \begin{matrix}
-l_1s_{1} -l_2s_{12}& -l_2s_{12}\\
l_1c_1 +l_2c_{12}& l_2c_{12}\\
\end{matrix} \right]
\end{equation}
Assume the desired trajectory \cite{tee2010adaptive} of the two joints is:
\begin{equation}
\label{eq11}
\begin{aligned}
x_{d1}\left( t \right) &= 0.14\cos \left( 0.5t \right) \\
x_{d2}\left( t \right) &=0.14\sin \left( 0.5t \right)
\end{aligned}
\end{equation}
\begin{figure}[b]
\centering
\begin{minipage}[t]{0.45\textwidth}
\centering
\includegraphics[width=0.72\textwidth]{Desired_Trajectory.png}
\caption{Desired/reference trajectory.}
\label{fig_4}
\end{minipage}
\begin{minipage}[t]{0.45\textwidth}
\centering
\includegraphics[width=0.68\textwidth]{External_Forces.png}
\caption{External human forces.}
\label{fig_5}
\end{minipage}
\end{figure}
The external human forces, as shown in Fig. \ref{fig_5}, are given by \cite{tee2010adaptive}:
\begin{equation}
\label{eq36}
fe_i\left( t \right) \left\{ \begin{array}{l}
0\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t<10\ or\ t\ge 21\\
a_i\left( 1-\cos \pi t \right) \ \ \ \ \ \ \ \ \ 10\le t<11\\
2a_i\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ 11\le t<20\\
a_i\left( 1+\cos \pi t \right) \ \ \ \ \ \ \ \ \ 20\le t<21\\
\end{array} \right.
\end{equation}
The external human forces are applied when $t=10s$ and removed at $t=21s$. By applying admittance control, the reference trajectory of the two joints $x_{r1}$ and $x_{r2}$ can be derived by integrating equation (\ref{eq10}) twice. Fig. \ref{fig_4} shows both the desired trajectory and the reference trajectory. We can see the two trajectories are the same before the external forces are applied. When $t>10s$, the reference trajectory, which complies with thr external human forces, is different from the desired trajectory. When $t>21s$, the external human forces are removed from the contact point and the two trajectories coincide again.
\begin{figure}
\centering
\includegraphics[width=7in]{Real_Trajectory.png}
\caption{Trajectories obtained with the different controllers.}
\label{fig_6}
\end{figure}
\begin{figure}[h]
\centering
\begin{minipage}[t]{0.495\textwidth}
\centering
\includegraphics[width=1\textwidth]{Tracking_Error_1.png}
\caption{Tracking error for joint 1.}
\label{fig_7}
\end{minipage}
\begin{minipage}[t]{0.495\textwidth}
\centering
\includegraphics[width=1\textwidth]{Tracking_Error_2.png}
\caption{Tracking error for joint 2.}
\label{fig_8}
\end{minipage}
\end{figure}
\begin{figure}[h]
\centering
\begin{minipage}[t]{0.495\textwidth}
\centering
\includegraphics[width=1\textwidth]{Control_Input_1.png}
\caption{Control input for joint 1.}
\label{fig_9}
\end{minipage}
\begin{minipage}[t]{0.495\textwidth}
\centering
\includegraphics[width=1\textwidth]{Control_Input_2.png}
\caption{Control input for joint 2.}
\label{fig_10}
\end{minipage}
\end{figure}
To verify the effectiveness of the proposed controller, we compare it with PID, computed torque control (CTC), traditional ISMC with PID as the nominal controller (ISMC PID), traditional ISMC with CTC as the nominal controller (ISMC CTC), and traditional ISMC with BSP as the nominal controller (ISMC BSP). Assume the initial position of the end-effector is $q\left(0\right)=\left[0.5236,2.0944\right]^T$, and the initial value of the task variable is $x\left(0\right)=\left[0,0\right]^T$. The design parameters of the admittance control are $k_{m_i}=20$, $k_{b_i}=20$ and $k_{k_i}=100$, $i=1,2$. The PID parameters are $k_p=300$, $k_d=400$, $k_i=10$. The parameters of BSP are $\lambda_1=3$, $\lambda_2=20$, $\lambda_3=50$, $\alpha=\frac{5}{7}$, $\beta=\frac{5}{3}$. These parameters are selected based on a trial-and-error procedure and by experience. The design parameters of the proposed controller are $k_1=k_3=20$, $k_2=k_4=50$, $m=p=\frac{5}{7}$, $n=q=\frac{5}{3}$.
The trajectories of the benchmark controllers and the proposed fixed-time ISMC controller with BSP as the nominal controller are shown in Fig. \ref{fig_6}, and the corresponding Root Mean Square Error (RMSE) of tracking errors are reported in Table 1. We can see that the PID controller has the poorest tracking performance. The performances of CTC is a little better than PID. The performance of ISMC PID and ISMC CTC are better than pure PID and CTC, respectively, which means the tracking performance of the controller is improved by adding the integral sliding surface. The ISMC BSP and proposed fixed-time ISMC controller have the best tracking performance. They almost achieve perfectly tracking of the reference trajectory, which makes it difficult to compare them in Fig. \ref{fig_6}. However, as can be seen in Fig. \ref{fig_7} and Fig. \ref{fig_8}, it is clear that the proposed controller has both the smallest tracking error and shortest convergence time. Note that the proposed controller is global fixed-time convergent because both the integral sliding surface, reaching controller and the nominal controller can converge within a fixed time. In addition, from Fig. \ref{fig_9} and Fig. \ref{fig_10}, which show the control inputs of the two joints, it can be seen that the proposed solution provides the smoothest and most efficient control input.
\begin{center}
\label{table1}
\begin{table}[h]
\caption{The RMSE performance of each controller.}
\setlength{\tabcolsep}{5.3mm}{
\begin{tabular}{ccccccc}
\toprule
\makecell[c]{Joint} &\makecell[c]{PID}& \makecell[c] {CTC} &\makecell[c]{ISMC+PID} &\makecell[c]{ISMC+CTC} &\makecell[c]{ISMC+BSP} &\makecell[c]{FTISMC+BSP}\\
\midrule
\multirowcell{0.1}{1}&0.1211&0.0070&0.0446&0.0064&$2.0901\times 10^{-4}$&$5.7841\times 10^{-5}$\\
\midrule
\multirowcell{0.1}{2}&0.0845&0.0314&0.0644&0.0220&0.0016&$4.4889\times 10^{-4}
$\\
\bottomrule
\end{tabular}}
\end{table}
\end{center}
\section{Conclusions}\label{sec5}
In this paper, a fixed-time ISMC controller based on admittance control has been proposed for robot manipulators such that the robot can comply with the external human forces rather than reject them as disturbances. Furthermore, the convergence time of the system can be predefined regardless of the initial conditions. The simulation results show that the proposed controller can track the reference trajectory precisely with faster convergence of the tracking error. In future work, a fixed-time force observer of external human forces will be discussed and integrated into the system. Safety constraints of the system will also be considered. The whole system will be tested on a robotics platform such as the Baxter robot to further validate the controller's effectiveness and evaluate real world performance.
|
\section{INTRODUCTION}\label{sec1}
\subsection{Motivation}
The past few decades have seen the rapid development in physical human-robot collaboration (pHRC) which is an increasingly important area in robotics. In the past, robots and human operators have been organised in separate areas to ensure safety \cite{RN532}.
However, the physical collaboration and interaction between humans and robots are unavoidable in some specific scenarios, such as rehabilitation robots \cite{RN100} which guide a patient's arm and move with a natural fluidity with human movements.
In addition to this, the pHRC of industrial robots has been attracting a lot of interest in both academia and industry.
Such a human-in-the-loop system makes full use of the reasoning capabilities of human workers and the high precision and endurance of robots \cite{RN546}, and therefore can finish much more complicated tasks compared to fully automated systems.
To achieve a so-called \emph{compliance} in the physical interaction between human and robot, there is a vast body of work on impedance/admittance based compliance control in robotics, by which the movement of robots can be smoother, softer and more human-friendly.
However, bringing such collaborative robots into real-world use also requires safety guarantees since the robot and human share the same workspace.
It is imperative that robots are capable of enforcing safety constraints with a human-like smooth behaviour when collaborating with a human partner.
Safety should be strictly guaranteed even in some unexpected emergency situations such as human pushing the robot towards obstacles or beyond the workspace boundary by accident.
To that end, we seek to enforce the safety-critical constraints of admittance control in pHRC. In particular, we consider two position-based constraints, workspace boundary and obstacle avoidance, which are the most practical safety constraints in reality. As depicted in Fig. \ref{fig_1}, the trajectory of the robot's end-effector should always be constrained within the workspace boundary and keep a specific distance from obstacles.
The proposed system implements a two-layered control framework: the high-level trajectory planning modifies the original desired trajectory in real-time. The ECBFs-QP based framework is applied to provide optimal virtual force feedback for admittance control as a compensator of the external human forces. Then, the low-level controller tracks such trajectories to achieve safety and human-friendly behaviour for robots, as depicted in Fig. \ref{fig_2}.
\begin{figure}[t]
\centering
\includegraphics[width=3.4in]{Picture5.eps}
\caption{Motivating example scenario for a human robot force interaction with safety requirements.}
\label{fig_1}
\end{figure}
\subsection{Related Work}
The most common control methods in robotics include position control \cite{RN88015}, force control \cite{844062}, hybrid position/force control \cite{uchiyama1987hybrid} and impedance/admittance based compliance control \cite{RN104}. Pure position control, whose task is to follow a specific trajectory as accurately as possible, rejects external human forces as disturbances. Therefore, it is not suitable in pHRC.
In contrast, pure force control only tracks the given forces, and therefore cannot guarantee the position without contact with the external environment \cite{RN531}.
As a trade-off, impedance/admittance based compliance control relates both position and force. The contact point between the robot and human is assumed to be a mass-spring-damper (MSD) system whose dynamic behaviour acts like a human motion mechanism.
Subsequently, many advanced control methods have emerged based on impedance/admittance control, such as adaptive impedance/admittance control \cite{kelly1989adaptive}\cite{colbaugh1993direct}, hybrid impedance/admittance control \cite{anderson1988hybrid}\cite{liu1991robust}, robust impedance/admittance control \cite{liu1991robust}\cite{chan1991robust} and learning impedance/admittance control \cite{RN1999}\cite{RN343}, etc.
Note that the impedance and admittance are two opposite notions in the MSD system.
The system is regarded as admittance when the input of the system is force and the output is position, while it is impedance when the input of the system is position and the output is force.
In this work, we seek to get optimal interaction force feedback based on ECBFs-QP for compliance control to generate the trajectory which guarantees both safety and compliance. Therefore, we apply admittance control in our proposed framework.
In addition to compliance, another major factor of collaborative robots that hinders their use in real pHRC is safety. Previous solutions in safety-critical control in robotics mainly include: (i) path planning which aims to derive a collision-free trajectory to enforce safety, and (ii) introducing a safety filter which modifies the input of the controller to guarantee safety. Path planning mainly includes heuristic-based methods \cite{al2003efficient} \cite{li2005mobile} and potential field based methods \cite{1225434} \cite{RN449}. In the work of AI-Khatib and Saade \cite{al2003efficient}, a data-driven fuzzy approach is developed for a mobile robot to achieve path planning for moving obstacles.
In short, heuristic-based methods have distinct advantages but the major drawbacks are response time and high computational complexity \cite{RN451}.
In the work of Tang et al. \cite{1225434}, an artificial potential field (APF) based path planning was proposed for obstacle avoidance without the problem of local minimums.
However, a work of Singletary et al. \cite{RN449} compared the performance of APF and control barrier functions (CBFs), and showed that CBFs outperformed existing APF-based algorithms. Therefore, CBFs have been an increasingly popular technique in the form of a safety filter.
Using a quadratic program (QP), CBFs based safety filters can combine with an arbitrary nominal control law to enforce safety \cite{RN444} \cite{RN999}.
Importantly, the CBFs-QP based framework can mediate the extent to which different constraints are met when these objectives are in conflict \cite{RN477}.
Such a property makes CBFs ideal for pHRC since the safety of a robotics system should be a hard constraint because any violation is not acceptable, while tracking performance for the desired trajectory to finish tasks should be a soft constraint. CBFs are capable of finding a trade-off of each component to achieve an ideal behaviour. Another extension of CBFs called exponential control barrier functions (ECBFs) was developed in the work of Nguyen and Sreenath \cite{nguyen2016exponential} \cite{RN501}. which aims to handle the safety constraints with systems of higher relative degree.
So far, only a few papers have considered both compliance and safety in pHRC. Noteworthy among these is the work of Engelbrecht et al. \cite{RN588}, which proposes a novel adaptive virtual impedance algorithm for obstacle avoidance. Unlike traditional impedance control, the main difference in this work is that the impedance control is compensated by force feedback which is generated by APF. The APF provides both repulsive forces and attractive forces as adaptive feedback for impedance control to enable robots to navigate with constraints.
\begin{figure}[t]
\centering
\includegraphics[width=3.4in]{Picture3.eps}
\caption{Structure of the proposed adaptive admittance control framework.}
\label{fig_2}
\end{figure}
\subsection{Contribution}
Motivated by the above discussion, in this letter, we apply the ECBFs-QP framework to provide virtual force feedback for admittance control. Such virtual force is an adaptive compensator for the external human forces. As can been seen in Fig. \ref{fig_2}, the optimal interaction force $\hat{f_e}$ is derived from the sum of external human forces $f_e$ and the compensated forces $f_{e}^{'}$. When the system tends to approach the unsafe set, $\hat{f_e}$ modifies the desired trajectory $x_r$ to $x_f$ for the low-level controller to track and bring the system back to the safe set. The major contribution of this work can be highlighted in a comparison with other approaches as follows:
\begin{enumerate}
\item Compared with traditional path planning \cite{al2003efficient} \cite{li2005mobile}, the proposed approach integrates an adaptive force compensator with admittance control to modify the desired trajectory. Therefore, the proposed control framework has higher robustness to external human forces, as well as shorter response time and lower computational complexity.
\item Compared with existing safety filters \cite{RN444} \cite{RN999}, the proposed approach seeks to enforce safety in the \emph{force} domain. The main benefit is the improvement in the robot's human-friendly motion behaviour in the physical human robot force interaction, while simultaneously guaranteeing the safety constraints.
\end{enumerate}
\subsection{Organization}
The remainder of the letter is organised as follows. Section \ref{sec2} provides the necessary background on the robot manipulator model and ECBFs. Section \ref{sec3} presents the formulation of the proposed control framework subject to two position-based constraints: workspace constraints and obstacle avoidance. In Section \ref{sec4}, the formulation of a low-level controller is given based on our previous works. The simulation results are demonstrated on a two-link planar robotic manipulator in Section \ref{sec5}. Finally, Section \ref{sec6} discusses the conclusions and directions for future works.
\section{Preliminaries}\label{sec2}
\subsection{Model}
In this section, we begin by briefly introducing the model that captures the dynamics of the robot manipulator. The joint space dynamics of a robot manipulator can be written as:
\begin{equation}
\label{eq1}
M\left( q \right) \ddot{q}+C\left( q,\dot{q} \right) \dot{q}+G\left( q \right) +F\left( \dot{q} \right) =\tau _c +\tau _e
\end{equation}
where $q$, $\dot{q}$, $\ddot{q}$ are the joint positions, velocities and accelerations in joint space. $M\left( q \right)$, $C\left( q,\dot{q} \right)$, $G\left( q \right)$ and $F\left( \dot{q} \right)$ are mass, Coriolis force and centrifugal force, gravity and friction coefficient matrices, respectively. $\tau_c$ denotes the control torque, and $\tau_e$ is the external torque from the environment. In the design of a robotic control system, the system needs to be formulated in Cartesian space when evaluating the trajectory tracking performance. The time-varying transformation between the joint velocities and Cartesian velocities of the robot manipulator can be written as:
\begin{equation}
\label{eq2}
\dot{x}=J\left( q \right) \dot{q}
\end{equation}
where $J\left( q \right)$ is the Jacobian of the robot manipulator. By using (\ref{eq2}), we can then transfer the joint space dynamics (\ref{eq1}) into Cartesian space as:
\begin{equation}
\label{eq3}
M_x\ddot{x}+C_x\dot{x}+G_x+F_x=f_c+f_e
\end{equation}
where $x$, $\dot{x}$, $\ddot{x}$ is the position, velocity and acceleration of the robot joints in Cartesian space. $f_c=J^{-T}\left( q \right) \tau _c$, and $f_e=J^{-T}\left( q \right) \tau _e$ present the control forces and the external human forces, respectively. The coefficient matrices are given by $M_x=J^{-T}\left( q \right) M\left( q \right) J^{-1}\left( q \right) $, $C_x=J ^{-T}\left( q \right)\left( C\left( q,\dot{q} \right) -M\left( q \right) J^{-1}\left( q \right) \dot{J}\left( q \right) \right) J^{-1}\left( q \right) $, $G_x=J^{-T}\left( q \right) G\left( q \right)$ and $\ F_x=J^{-T}\left( q \right) F\left( \dot{q} \right) $. Note that another benefit of transferring joint space dynamic into Cartesian space is that we can relate the external force $f_e$ (rather than torque $\tau_e$) in (\ref{eq3}) directly to the forces in the equation of admittance control, which will be mentioned in later sections.
\subsection{Exponential Control Barrier Function}
In this section, we revisit the mathematical background of exponential control barrier functions. Consider a nonlinear control system of the form:
\begin{equation}
\left\{ \begin{array}{l}
\dot{x}=f\left( x,u \right) \\
y=\zeta \left( x \right)\\
\end{array} \right.
\label{e1}
\end{equation}
where $x\in \mathbb{R}^n$ is the system state, $u\in U\subset \mathbb{R}^m$ is the admissible control input, and $y\in \mathbb{R}^m$ is the control output. $f:\mathbb{R}^m\rightarrow \mathbb{R}^n$ is locally Lipschitz. The primary focus of control barrier functions is to force system states to remain in a given safe set $C$ which is defined as the superlevel set of a continuously differentiable function $h:\mathbb{R}^n\rightarrow \mathbb{R}$:
\begin{equation}
C=\left\{ x\in \mathbb{R}^n:h\left( x \right) \ge 0 \right\}
\label{e5}
\end{equation}
\begin{equation}
\partial C=\left\{ x\in \mathbb{R}^n:h\left( x \right) =0 \right\}
\label{e6}
\end{equation}
\begin{equation}
\mathrm{Int}\left( C \right) =\left\{ x\in \mathbb{R}^n:h\left( x \right) >0 \right\}
\label{e7}
\end{equation}
where $h\left(x\right)$ is called the constraint function. $h\left(x\right) >0$ indicates safety, while $h\left(x\right) <0$ indicates violation of the safety constraints. $C$ denotes the safe set, $\partial C$ denotes the boundary of the set $C$, and $\mathrm{Int}\left( C \right) $ denotes the interior of the set $C$. The mathematical tools that underpin application of control safety features such as CBFs are based on results from the well-known Nagumo's Theorem \cite{RN235467} which provides the necessary sufficient conditions of the invariant set:
\begin{equation}
\dot{h}\left( x \right) \ge 0, \forall x\in \partial C
\end{equation}
Using these conditions, enforcement of the safety of a control system can be converted to another question: \emph{what sufficient conditions that need to be imposed on $h\left(x\right)$ so that $\mathrm{Int}\left( C \right) $ is forward invariant?} CBFs provide a solution to this question and therefore can enforce the \emph{safety} of a control system.
\noindent\textbf{\emph{Definition 1}}. Let $u \in U$ be a control value for the system (\ref{e1}). For any initial states $x_0:=x\left(t_0\right)$, $x\left(t\right)$ is the unique solution to (\ref{e1}) in the maximum time interval $T\left( x_0 \right)$. The set $S$ is \emph{forward invariant} with respect to the control value $u$ if for every $x_0\in S$, $x\left(t \right)\in S$ for all $t\in T\left(x_0\right)$. The control system (\ref{e1}) is \emph{safe} with respect to the set $S$ if the set $S$ is forward invariant.
Essentially, there are two types of CBFs: one is \emph{reciprocal control barrier functions} which tend to infinity on the set boundary, and one is \emph{zeroing control barrier functions} which vanish on the set boundary.
1) \emph{Reciprocal control barrier functions (RCBFs)} \cite{RN444}: A RCBF $B\left(x\right)$ which tends to infinity on the set boundary $\partial C$, should satisfy the following important properties:
\begin{equation}
\inf_{x\in \mathrm{Int}\left( C \right) }B\left( x \right) \ge 0,\ \ \lim_{x\rightarrow \partial C}B\left( x \right) =\infty
\label{p1}
\end{equation}
\noindent\textbf{\emph{Definition 2}}. Consider a control system in the form of (\ref{e1}). For a given safe set $C$, admissible control input set $U$ and a continuously differentiable function $h\left(x\right)$, $B\left(x\right)$ is a \emph{reciprocal control barrier function} if there exist class $\mathcal{K}$ functions $\alpha_1$, $\alpha_2$, $\alpha_3$ such that for all $x\in \mathrm{Int}\left( C \right) $, $B\left(x\right)$ satisfies following conditions:
\begin{equation}
\frac{1}{\alpha _1\left( h\left( x \right) \right)}\le B\left( x \right) \le \frac{1}{\alpha _2\left( h\left( x \right) \right)}
\end{equation}
\begin{equation}
\inf_{u\in U}\left[ \mathscr{L}_fB\left( x ,u\right) -\alpha _3\left( h\left( x \right) \right) \right] \le 0
\end{equation}
where $\mathscr{L}_fB\left( x \right)$ is the \emph{Lie derivative} of $B\left( x \right)$ with respect to the vector field $f$. To satisfy the properties in (\ref{p1}), reciprocal control barrier function candidates are usually selected as the form of the inverse-type barrier candidate $B\left( x \right) =\frac{1}{h\left( x \right)}$, logarithmic barrier function candidate $B\left( x \right) =-\log \left( \frac{h\left( x \right)}{1+h\left( x \right)} \right)$ and so on.
\noindent\textbf{\emph{Lemma 1}} \cite{RN444}. Consider a control system in the form of (\ref{e1}). For a given safe set $C$ defined by (\ref{e5})-(\ref{e7}), admissible control input set $U$ and a RCBF $B\left(x\right)$, any locally Lipschitz continuous control input $u\left(x\right)\in U$ such that $u\left(x\right) \in K_{rcbf}\left( x \right) =\left\{ u\in U:\mathscr{L}_fB_r\left( x,u \right)-\alpha _3\left( h\left( x \right) \right) \le 0 \right\}
$ will render the set $\mathrm{Int}\left( C \right) $ forward invariant.
\noindent\textbf{\emph{Definition 3}}. Consider a continuous function $\alpha :\left[ 0,\infty \right) \rightarrow \left[ 0,\infty \right)$. It is said the $\alpha$ is a \emph{class $\mathcal{K}$ function} if it is strictly increasing and $\alpha\left(0\right)=0$. Based on class $\mathcal{K}$ functions, \emph{extended class $\mathcal{K}$ functions} are defined on the entire real line $\mathbb{R}=\left( -\infty ,\infty \right) $.
\noindent\textbf{\emph{Definition 4}}. Consider a control system in the form of (\ref{e1}). For a scalar function $B\left(x\right)$, the directional derivatives of $B\left(x\right)$ with respect to the vector field $f\left(x,u\right)$ are called the \emph{Lie derivatives} of $B\left(x\right)$ along $f\left(x,u\right)$ and are denoted by:
\begin{equation}
\mathscr{L}_fB\left( x ,u\right) =\frac{\partial B\left( x \right)}{\partial x}f\left( x,u \right)
\end{equation}
2) \emph{Zeroing control barrier functions (ZCBFs)} \cite{RN444}: Since the unbounded values of $B\left(x\right)$ may be undesirable in practical implementations, employing ZCBFs which vanish on the set boundary $\partial C$ give a solution to this problem. A ZCBF $h\left(x\right)$ satisfies the following important properties:
\begin{equation}
\inf_{x\in \mathrm{Int}\left( C \right) }h\left( x \right) \ge 0,\ \ \lim_{x\rightarrow \partial C}h\left( x \right) =0
\end{equation}
\noindent\textbf{\emph{Definition 5}}. Consider a control system in the form of (\ref{e1}). For a given safe set $C$, admissible control input set $U$ and a continuously differentiable function $h\left(x\right)$, $h\left(x\right)$ is a \emph{zeroing control barrier functions} if there exist an extended class $\mathcal{K}$ function $\alpha$ such that for all $x\in \mathrm{Int}\left( C \right) $, $h\left(x\right)$ satisfies the following conditions:
\begin{equation}
\sup_{u\in U}\left[ \mathscr{L}_fh\left( x,u \right)-\alpha \left( h\left( x \right) \right) \right] \ge 0
\label{s1}
\end{equation}
\noindent\textbf{\emph{Lemma 2}} \cite{RN444}: Consider a control system in the form of (\ref{e1}). For a given safe set $C$ defined by (\ref{e5})-(\ref{e7}), admissible control input set $U$ and a ZCBF $h\left(x\right)$, any locally Lipschitz continuous control input $u\left(x\right)\in U$ such that $u\left(x\right) \in K_{zcbf}\left( x \right) =\left\{ u\in U:\mathscr{L}_fh\left( x,u \right)-\alpha \left( h\left( x \right) \right) \ge 0 \right\} $ will render the set $\mathrm{Int}\left( C \right) $ forward invariant.
\noindent\textbf{\emph{Remark 1}}. Note that usually the $\alpha$ in (\ref{s1}) is the extended class $\mathcal{K}$ function. There is a special case in which the term $\alpha\left(h\left(x\right)\right)$ is replaced by $\lambda h\left(x\right)$, where $\lambda$ is a positive real number which is selected based on experience. In this letter, we apply this special case of ZCBFs in our proposed control framework.
The first order Lie derivatives of CBFs with higher relative degree than one do not depend explicitly on the control input offering no way to calculate what control changes need to be applied to enforce the safety condition. Therefore, an extension of the CBFs, called \emph{exponential control barrier function}, is proposed for dealing with higher relative degree constraint functions \cite{nguyen2016exponential} \cite{RN501}. In this letter, we consider two position-based constraints with relative degree two, therefore, we apply ECBFs in our proposed control framework.
\noindent\textbf{\emph{Definition 6}}. For a given control system (\ref{e1}), safe set $C$ defined by (\ref{e5})-(\ref{e7}), admissible control input set $U$ and a ZCBF $h\left(x\right)$ which has relative degree $r$, then $h\left(x\right)$ is an \emph{exponential control barrier function} if there exists a $K\in \mathbb{R}^{r}$ such that for any $x\in C$, the following condition is satisfied:
\begin{equation}
\inf_{u\in U}\left[ \mathscr{L}_{f}^{r}h\left( x,u \right) +K\xi \left( x \right) \right] \ge 0
\label{key}
\end{equation}
where $\xi \left( x \right) =\left[ h\left( x \right) ,\mathscr{L}_fh\left( x \right) ,\mathscr{L}_{f}^{2}h\left( x \right) ,...,\mathscr{L}_{f}^{r-1}h\left( x \right) \right] ^T$ and $K=\left[k^1, k^2,...k^{r-1}\right]$. Equation (\ref{key}) provides the sufficient condition for safety in our proposed framework.
\noindent\textbf{\emph{Lemma 3}} \cite{RN501}. Consider a control system in the form of (\ref{e1}). For a given safe set $C$ defined by (\ref{e5})-(\ref{e7}), admissible control input set $U$ and a ECBF $h\left(x\right)$ with relative degree $r$, any locally Lipschitz continuous control input $u\left(x\right)\in U$ such that $u\left(x\right) \in K_{ecbf}\left( x \right) =\left\{ u\in U:\mathscr{L}_{f}^{r}h\left( x,u \right) +K\xi \left( x \right) \ge 0 \right\} $ will render the set $\mathrm{Int}\left( C \right) $ forward invariant.
\section{Adaptive admittance control for safety-critical control}\label{sec3}
\subsection{Admittance control reconstruction}
To provide compliance for physical human-robot force interaction, the contact point between the human and the robot is modelled as a mass-spring-damper system to imitate human muscle mechanisms, as depicted in Fig. \ref{p2}. The virtual mass, spring, and damper ensure that the interaction forces are elastic and never vibrate at the contact point. The dynamics for a robot manipulator rendering an impedance can be written as:
\begin{small}
\begin{equation}
\label{eq10}
k_{m_i}\left( \ddot{x}_{r_i}-\ddot{x}_{d_i} \right) +k_{b_i}\left( \dot{x}_{r_i}-\dot{x}_{d_i} \right) +k_{k_i}\left( x_{r_i}-x_{d_i} \right) =f_{e_i}
\end{equation}
\end{small}
where $k_{m_i}$, $k_{b_i}$ and $k_{k_i}$ are the mass, spring and damping coefficients, respectively. $i$ is the number of degree of freedom (DOF) of the robot manipulator. $x_{d_i}$ is the desired trajectory that has been determined in advance to finish the task. $x_{r_i}$ is the reference trajectory modified by the effect of the external force by using admittance control. $f_{e_i}$ is the external human force.
\begin{figure}[h]
\centering
\includegraphics[width=1.7in]{Picture4.eps}
\caption{ The mass-spring-damper system.}
\label{p2}
\end{figure}
Let $x_{1_i}=x_{r_i}$, $x_{2_i}=\dot{x}_{r_i}$, $u_i=f_{e_i}$, from (\ref{eq10}), we have:
\begin{small}
\begin{equation}
\left\{ \begin{array}{l}
\dot{x}_{1_i}=x_{2_i}\\
\dot{x}_{2_i}=f_i\left( x_i \right) +g_i\left( x_i \right) u_i\\
\end{array} \right.
\label{s2}
\end{equation}
\end{small}
where $f_i\left( x_i \right)$ and $g_i\left( x_i \right)$ are written as $f_i\left( x_i \right) =-\frac{1}{k_{m_{\text{i}}}}\left[ k_{b_i}\left( x_{2_i}-\dot{x}_{d_i} \right) +k_{k_i}\left( x_{1_i}-x_{d_i} \right) -k_{m_i}\ddot{x}_{d_i} \right]$ and
$g_i\left( x_i \right) =\frac{1}{k_{m_i}}$.
\subsection{Workspace constraints}
In this section, we present the application of ECBFs for admittance system (\ref{s2}) with workspace constraints. We assume the end-effector of the robot is required to be constrained within a virtual box that defines its desired working area. For this case, there will be an upper boundary and a lower boundary for each sate in (\ref{s2}). These are:
\begin{equation}
x_{\min _i}\le x_{r_i}\le x_{\max _i}
\end{equation}
Then, the constraint function $h\left(x\right)$ for the upper and lower boundary of each state can be designed as:
\begin{small}
\begin{equation}
\begin{split}
h_{\max _i}\left( x \right) =\left( x_{1_i}-x_{\max _i} \right) ^2-r^2 \\
h_{\min _i}\left( x \right) =\left( x_{\min _i}-x_{1_i} \right) ^2-r^2
\end{split}
\label{h1h2}
\end{equation}
\end{small}
where $r$ denotes the safe distance between the end-effector and workspace boundary. The safe set is defined as $ C=\left\{ x\in \mathbb{R}^n:h_{\max _i}\ge 0\ \&\ h_{\min _i}\ge 0 \right\}$. The Lie derivative of the constraint function (\ref{h1h2}) with respect to $f\left(x,u\right)$ is:
\begin{small}
\begin{equation*}
\begin{split}
\mathscr{L}_fh_{\max _i}\left( x \right) =2\left( x_{1_i}-x_{\max _i} \right) \dot{x}_{1_i}=2\left( x_{1_i}-x_{\max _i} \right) x_{2_i}\ \ \\
\mathscr{L}_fh_{\min _i}\left( x \right) =-2\left( x_{\min _i}-x_{1_i} \right) \dot{x}_{1_i}=-2\left( x_{\min _i}-x_{1_i} \right) x_{2_i}
\end{split}
\end{equation*}
\end{small}
Since (\ref{h1h2}) have relative degree two, the second-order Lie derivative of the constraint functions are:
\begin{small}
\begin{equation}
\begin{aligned}
\mathscr{L}_{f}^{2}h_{\max _{\text{i}}}\left( x,u \right) &=2\left( x_{1_i}-x_{\max _i} \right) \dot{x}_{2_i}+2x_{2_i}\\
&=2\left( x_{1_i}-x_{\max _i} \right) \left( f_i\left( x \right) +g_i\left( x \right) u_i \right) +2x_{2_i}\\
\mathscr{L}_{f}^{2}h_{\min _i}\left( x,u \right) &=-2\left( x_{\min _i}-x_{1_i} \right) \dot{x}_{2_i}+2x_{2_i}\\
&=-2\left( x_{\min _i}-x_{1_i} \right) \left( f_i\left( x \right) +g_i\left( x \right) u_i \right) +2x_{2_i}
\end{aligned}
\label{c7c8}
\end{equation}
\end{small}
According to (\ref{key}), to achieve the forward invariance of the safe set $C$, the following inequalities should be satisfied:
\begin{small}
\begin{equation}
\begin{split}
\mathscr{L}_{f}^{2}h_{\max _i}\left( x,u \right) +K_{\max_i}\xi _{\max_i}\left( x \right) \ge 0\\
\mathscr{L}_{f}^{2}h_{\min _i}\left( x,u \right) +K_{\min_i}\xi _{\min_i}\left( x \right) \ge 0
\label{c1c2}
\end{split}
\end{equation}
\end{small}
where $\xi _{\max _i}\left( x \right) =\left[ h_{\max _i}\left( x \right) ,\mathscr{L}_fh_{\max _i}\left( x \right) \right] ^T$ and $\xi _{\min _i}\left( x \right) =\left[ h_{\min _i}\left( x \right) ,\mathscr{L}_fh_{\min _i}\left( x \right) \right] ^T$. The parameter $K_{\max_i}=\left[k_{\max_1},k_{\max_2}\right]$ and $K_{\min_i}=\left[k_{\min_1},k_{\min_2}\right]$ denote the positive coefficients of the ECBFs. The selection of these parameters depends on the dynamics of the robot manipulator. These parameters are selected based on a trial-and-error procedure and by experience. Then, the Quadratic Program is applied here to derive minimal modifying forces that maintain the safety conditions subject to the conditions (\ref{c1c2}):
\begin{small}
\begin{equation}
\begin{aligned}
\hat{f}_{e}^{'}&=\underset{\hat{f}_{e_i}}{arg\min}\lVert \hat{f}_e-f_e \rVert ^2\\
s.t.\ \ \ \mathscr{L}_{f}^{2}h_{\max _i}&\left( x,u \right) +K_{\max_i}\xi _{\max_i}\left( x \right) \ge 0\\
\mathscr{L}_{f}^{2}h_{\min _i}&\left( x,u \right) +K_{\min_i}\xi _{\min_i}\left( x \right) \ge 0
\label{key1}
\end{aligned}
\end{equation}
\end{small}
where $\hat{f}_{e}$ is the minimal modifying force we wish to find, $f_e$ is the external human force and $\hat{f}_{e}^{'}$ is the compensated force. By using (\ref{key1}), the minimum $\hat{f}_{e}^{'}$ is picked up based on the conditions (\ref{c1c2}). The final output from the QP $\hat{f}_{e}^{'}$ compensates the human forces as an add-on to $f_e$, resulting in the optimal interaction force $\hat{f}_{e}$ which enforces the safety of system (\ref{s2}). In addition, the natural of system (\ref{s2}) also guarantees the compliance of $\hat{f}_{e}$. We define the constraints within the QP in the form:
\begin{equation}
Au-b\le 0
\label{c5}
\end{equation}
Let $u_{nom}=f_e$, $u_{safe}=\hat{f}_{e}$. Taking maximum boundary as an example, writing (\ref{c1c2}) in the form of (\ref{c5}), we have:
\begin{small}
\begin{equation*}
\begin{split}
-\mathscr{L}_{f}^{2}h_{\max _i}\left( x,\left( u_{nom}-u_{safe} \right) \right) \le K_{\max _i}\xi _{\max _i}\left( x \right)\ \ \ \ \ \ \ \\
\mathscr{L}_{f}^{2}h_{\max _i}\left( x,u_{safe} \right) \le \mathscr{L}_{f}^{2}h_{\max _i}\left( x,u_{nom} \right) +K_{\max _i}\xi _{\max _i}\left( x \right)
\end{split}
\end{equation*}
\end{small}
According to (\ref{c7c8}), we have:
\begin{small}
\begin{equation*}
\begin{split}
2\left( x_{1_i}-x_{\max _i} \right) g_i\left( x \right)u_{safe_i} \le \ \ \ \ \ \ \ \ \ \ \\
2\left( x_{1_i}-x_{\max _i} \right) g_i\left( x \right) u_{nom_i}+K_{\max _i}\xi _{\max _i}\left( x \right)
\end{split}
\end{equation*}
\end{small}
Similarly, for minimum boundary we have:
\begin{small}
\begin{equation*}
\begin{split}
-2\left( x_{1_i}-x_{\min _i} \right) g_i\left( x \right)u_{safe_i} \le \ \ \ \ \ \ \ \ \ \ \\
-2\left( x_{1_i}-x_{\min _i} \right) g_i\left( x \right) u_{nom_i}+K_{\min _i}\xi _{\min _i}\left( x \right)
\end{split}
\end{equation*}
\end{small}
Therefore, when applying (\ref{key1}) for practical application, $A_{w_i}$ and $b_{w_i}$ for the i-th DOF of the robot manipulator can be written as:
\begin{small}
\begin{equation*}
\begin{split}
A_{w_i}=\left[ \begin{matrix}
2\left( x_{1_i}-x_{\max _i} \right) g_i\left( x \right) & 0\\
0 & -2\left( x_{\min _i}-x_{1_i} \right) g_i\left( x \right)\\
\end{matrix} \right] \\
b_{w_i}=\left[ \begin{array}{c}
2\left( x_{1_i}-x_{\max _i} \right) g_i\left( x \right) u_{nom_i}+K_{\max _i}\xi _{\max _i}\left( x \right)\\
-2\left( x_{1_i}-x_{\min _i} \right) g_i\left( x \right) u_{nom_i}+K_{\min _i}\xi _{\min _i}\left( x \right)\\
\end{array} \right]
\end{split}
\end{equation*}
\end{small}
\subsection{Obstacle avoidance}
In this section, we present the application of ECBFs for admittance system (\ref{s2}) with obstacle constraints. We assume the end-effector of the robot should always keep a safe distance from the obstacles. For this case, the constraint function is designed as:
\begin{small}
\begin{equation}
h_{obs}\left( x \right) =\sum_1^i{\left( x_{1_i}-x_{obs_i} \right) ^2-r^2}
\label{h3}
\end{equation}
\end{small}
where $x_{obs_i}$ denotes the position of obstacle in the Cartesian space and $r$ denotes the safe distance. The safe set is define as $C=\left\{ x\in \mathbb{R}^n:h_{obs}\left( x \right)\ge 0\right\}$. The first and second order Lie derivative of the constraint function (\ref{h3}) with respect to the $f\left(x,u\right)$ are:
\begin{small}
\begin{equation}
\begin{split}
\mathscr{L}_fh_{obs}\left( x \right) =\sum_1^i{\left( 2\left( x_{1_i}-x_{obs_i} \right) \dot{x}_{1_i} \right) =}\sum_1^i{\left( 2\left( x_{1_i}-x_{obs_i} \right) x_{2_i} \right)}\\
\mathscr{L}_{f}^{2}h_{obs}\left( x,u \right) =\sum_1^i{\left( 2\left( x_{1_i}-x_{obs_i} \right)
\left( f_i\left( x \right) +g_i\left( x \right) u_i \right) +2x_{2_i} \right)}
\label{c9}
\end{split}
\end{equation}
\end{small}
According to (\ref{key}), to achieve the forward invariance of the safe set $C$, the following inequalities should be satisfied:
\begin{small}
\begin{equation}
\mathscr{L}_{f}^{2}h_{obs}\left( x,u \right) +K_{obs}\xi_{obs} \left( x \right) \ge 0
\label{c6}
\end{equation}
\end{small}
where $K_{obs}=\left[k_1,k_2\right]$, $\xi_{obs} \left( x \right) =\left[ h\left( x \right) ,\mathscr{L}_fh\left( x \right) \right] ^T$. Then, the Quadratic Program is applied to derive optimal interaction forces subject to the conditions (\ref{c6}):
\begin{small}
\begin{equation}
\begin{aligned}
\hat{f}_{e}^{'}=\underset{\hat{f}_{e_i}}{arg\min}\lVert \hat{f}_e-f_e \rVert ^2\ \ \ \ \ \ \ \\
s.t. \mathscr{L}_{f}^{2}h_{obs}\left( x,u \right) +K_{obs}\xi_{obs} \left( x \right) \ge 0
\label{key2key7}
\end{aligned}
\end{equation}
\end{small}
Similarly to the workspace constraints, when applying (\ref{key2key7}) for practical application, (\ref{key2key7}) can be rewritten as:
\begin{small}
\begin{equation*}
\mathscr{L} _{f}^{2}h_{obs}\left( x,u_{safe} \right) \le \mathscr{L} _{f}^{2}h_{obs}\left( x,u_{nom} \right) +K_{obs}\xi_{obs} \left( x \right)
\end{equation*}
\end{small}
According to (\ref{c9}), we have:
\begin{small}
\begin{equation*}
\begin{aligned}
\sum_1^i{2\left( x_{1_i}-x_{obs_i} \right) g_i\left( x \right) u_{safe_i}}\le \ \ \ \ \ \ \ \ \ \ \\
\sum_1^i{2\left( x_{1_i}-x_{obs_i} \right) g_i\left( x \right) u_{nom_i}}+K_{obs}\xi_{obs} \left( x \right)
\end{aligned}
\end{equation*}
\end{small}
Therefore, when applying (\ref{key2key7}) as before within our QP constraints, $A$ and $b$ can be written as:
\begin{small}
\begin{equation*}
\begin{split}
A_{obs}=\left[ 2\left( x_{1_1}-x_{obs_1} \right) g_1\left( x \right) ,...,2\left( x_{1_i}-x_{obs_i} \right) g_i\left( x \right) \right] \\
b_{obs}=\sum_1^i{2\left( x_{1_i}-x_{obs_i} \right) g_i\left( x \right) u_{nom_i}}+K_{obs}\xi_{obs} \left( x \right)
\end{split}
\end{equation*}
\end{small}
\subsection{Workspace and obstacle constraints simultaneously}
To achieve both workspace constraints and obstacle avoidance simultaneously, we then synthesize these two kinds of constraints into QP, we have:
\begin{equation}
\begin{aligned}
\hat{f}_{e}^{'}=\underset{\hat{f}_{e_i}}{arg\min}\lVert \hat{f}_e-f_e \rVert ^2\ \ \ \ \ \ \ \\
s.t.\ \ \ \ \ \mathscr{L}_{f}^{2}h_{obs}\left( x,u \right) +K_{obs}\xi_{obs} \left( x \right) \ge 0 \\
\mathscr{L}_{f}^{2}h_{\min _i}\left( x,u \right) +K_{\min_i}\xi _{\min_i}\left( x \right) \ge 0\\
\mathscr{L}_{f}^{2}h_{\max _i}\left( x,u \right) +K_{\max_i}\xi _{\max_i}\left( x \right) \ge 0
\label{key3456}
\end{aligned}
\end{equation}
According to (\ref{c5}), the synthesized $A$ and $b$ of (\ref{key3456}) can be written as:
\begin{equation*}
A=\left[ A_{obs}, A_{w_1}, A_{w_2},...,A_{w_i} \right]^T ,b=\left[ b_{obs},b_{w_1},b_{w_2},...,b_{w_i}\right]^T
\end{equation*}
\section{Low-Level controller}\label{sec4}
As shown in Fig. 2, after the admittance control generated trajectory $x_r$ is adapted to satisfy the system safety constraints as trajectory $x_f$, the low-level controller will be designed to track the trajectory $x_f$. To guarantee the safety of the system, the low-level controller needs to guarantee that the system states will track trajectory $x_f$ precisely with very small or near zero offset. In this letter, we apply the fixed-time integral sliding mode controller (FxTISMC) \cite{RN488} as the low-level controller as it provides very high precision and tracking error convergence within a fixed time interval. The FxTISMC controller is designed as $u_c=u_0+u_{s}$, where $u_0$ is the nominal controller used to control the nominal component and $u_{s}$ is the compensating controller used to compensate for the model uncertainty and stabilise the system.
1) \emph{Nominal controller}:
Letting $\eta _1=x$ and $\eta _2=\dot{x}$, according to the dynamics of the robot manipulator in Cartesian space (\ref{eq3}), the dynamics without uncertainties can be written as:
\begin{small}
\begin{equation*}
\left\{ \begin{array}{l}
\dot{\eta}_1=\eta _2\\
\dot{\eta}_2=\varXi u+\varGamma \left( x, \dot{x} \right)\\
\end{array} \right.
\end{equation*}
\end{small}
where $\varGamma \left( x,\dot{x} \right) =M_x^{-1}\left( -C_x-G_x \right)$. Letting $s_1=\eta _1-x_r$, then, $\dot{s}_1=\eta_2-\dot{x_r}$, and the stabilizing function can designed as:
\begin{small}
\begin{equation*}
\alpha _s=-\left( \lambda _1s_1+\lambda _2s_{1}^{\alpha}+\lambda _3s_{1}^{\beta} \right) +\dot{x}_r
\end{equation*}
\end{small}
where $\lambda _1$, $\lambda _2$, $\lambda _3$, $\alpha$, $\beta$ are all positive constants satisfying $0<\alpha<1$ and $\beta>1$. Letting $s_2=\eta _2-\alpha _s $, we have:
\begin{small}
\begin{equation*}
\dot{s}_2=\dot{\eta}_2-\dot{\alpha}_s=\varXi u_0-\varGamma \left( x,\dot{x} \right) -\dot{\alpha}_s
\end{equation*}
\end{small}
The nominal controller is then designed as:
\begin{small}
\begin{equation*}
u_0=\varXi ^{-1}\left( -\varGamma \left( x,\dot{x} \right) +\dot{\alpha}_s-\lambda _1s_2-\lambda _2s_{2}^{\alpha}-\lambda _3s_{2}^{\beta} \right)
\end{equation*}
\end{small}
2) \emph{Compensating controller}:
Let the tracking error $e=x-x_r$, select the sliding variable as:
\begin{small}
\begin{equation*}
s = e+\frac{1}{\kappa_1^{m}}\left[ \dot{e}+\kappa_2\left[ e \right] ^n \right] ^{\frac{1}{m}}
\end{equation*}
\end{small}
where $\kappa_ 1$, $\kappa_2$, $\kappa_ 3$, $\kappa_4$, $m$, $n$ are all constants satisfying $\kappa_ 1, \kappa_2, \kappa_3, \kappa_4>0$ , $0<m<1$ and $n>1$. The integral sliding surface is then selected as:
\begin{small}
\begin{equation*}
\begin{split}
\sigma \left( t \right) =s\left( t \right) -s\left( 0 \right)
-\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \\
\int_0^t{\left(\dot{e}+\frac{1}{m\kappa_{1}^{m}}\left[ \dot{e}+\kappa_2\left[ e \right] ^n \right] ^{\frac{1}{m}-1}\left( \ddot{e}+\kappa_2n\left[ e \right] ^{n-1} \right) \dot{e} \right)dt}
\end{split}
\end{equation*}
\end{small}
Then, the compensating controller $u_{s}$ is designed as:
\begin{small}
\begin{equation*}
u_{s}=\varXi ^{-1}\left( -\left( \rho +\varepsilon \right) sign\left( \sigma \right) -\kappa_3\left[ \sigma \right] ^p-\kappa_4\left[ \sigma \right] ^q \right)
\end{equation*}
\end{small}
where $\rho$ and $\varepsilon$ are small positive constant. $p$ and $q$ are constants satisfying $0<p<1$ and $q>1$. After getting nominal controller $u_0$ and compensating controller $u_{s}$, the FxTISMC is derived by using $u_c=u_0+u_s$. The stability and convergence of this controller have been proven in our previous work \cite{RN488}.
\section{Numerical Example}\label{sec5}
In this section, a two-link planar robot manipulator is employed as a use case to conduct the simulation. The approach can in principle be extended to a robot with arbitrary degrees of freedom.
\subsection{Simulation details}
The dynamics of the two-link planar robot are described as \cite{RN1987}:
\begin{small}
\begin{equation*}
\begin{aligned}
\tau _1&=m_2l_{2}^{2}\left( \ddot{q}_1+\ddot{q}_2 \right) +m_2l_1l_2c_2\left( 2\ddot{q}_1+\ddot{q}_2 \right) +\left( m_1+m_2 \right) l_{1}^{2}\ddot{q}_1-\\
&m_2l_1l_2s_2\dot{q}_{2}^{2}-2m_2l_1l_2s_2\dot{q}_1\dot{q}_2+m_2l_2gc_{12}+\left( m_1+m_2 \right) l_1gc_1\\
\tau _2&=m_2l_1l_2c_2\ddot{q}_1+m_2l_1l_2s_2\dot{q}_{1}^{2}+m_2l_2gc_{12}+m_2l_{2}^{2}\left( \ddot{q}_1+\ddot{q}_2 \right)
\end{aligned}
\end{equation*}
\end{small}
where $c_i=\cos \left( q_i \right) $, $c_{ij}=\cos \left( q_i+q_j \right) $, $s_i=\sin \left( q_i \right) $, and $s_{ij}=\sin \left( q_i+q_j \right)$, $i,j=1,2$. The $M\left( q \right)$, $C\left( q,\dot{q} \right)$, $G\left( q \right)$, $F\left( q \right)$, and the Jacobian of the robot are given as:
\begin{small}
\begin{equation*}
\begin{split}
M\left( q \right) =\left[ \begin{matrix}
m_2l_{2}^{2}+2m_2l_1l_2c_2+\left( m_1+m_2 \right) l_{1}^{2}& m_2l_{2}^{2}+m_2l_1l_2c_2\\
m_2l_{2}^{2}+m_2l_1l_2c_2& m_2l_{2}^{2}\\
\end{matrix} \right] \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \\
C\left( q,\dot{q} \right) =\left[ \begin{array}{c}
-m_2l_1l_2s_2\dot{q}_{2}^{2}-2m_2l_1l_2s_2\dot{q}_1\dot{q}_2\\
m_2l_1l_2s_2\dot{q}_{1}^{2}\\
\end{array} \right] \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \\
G\left( q \right) =\left[ \begin{array}{c}
m_2l_2gc_{12}+\left( m_1+m_2 \right) l_1gc_1\\
m_2l_2gc_{12}\\
\end{array} \right]\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \\
F\left( q \right) =\left[ \begin{array}{c}
2c_1s_2+5c_{1}^{2}\\
-2c_1s_2-5c_{1}^{2}\\
\end{array} \right],
J\left( q \right) =\left[ \begin{matrix}
-l_1s_1-l_2s_{12}& -l_2s_{12}\\
l_1c_1+l_2c_{12}& l_2c_{12}\\
\end{matrix} \right] \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\end{split}
\end{equation*}
\end{small}
\begin{table}[t]
\centering
\caption{Simulation parameters.}
\begin{tabular}{cc}
\hline
Modules & Parameters\\
\hline
Initial value & \makecell{$q\left(0\right)=\left[0.5236,2.0944\right]^T$ \\ $x\left(0\right)=\left[0,0\right]^T$ } \\
\hline
Robot dynamics & \makecell{$m_1=1.5kg$, $m_2=1.0kg$ \\$l_1=l_2=0.3m$} \\
\hline
ECBFs-QP & \makecell{$K_{\max_i}=\left[500,50\right]$ \\ $K_{\min_i}=\left[500,50\right]$ \\$K=\left[700,70\right]$}\\
\hline
Admittance control & \makecell{$k_{m_i}=20$, $k_{b_i}=20$, $k_{k_i}=100$ \\ $a_1=1$, $a_2=2$} \\
\hline
Low-Level controller & \makecell{$\lambda_1=3$, $\lambda_2=20$, $\lambda_3=50$ \\ $\alpha = m=p=\frac{5}{7}$, $\beta = n=q=\frac{5}{3}$ \\$\kappa_1 = \kappa_3=20$, $\kappa_2 = \kappa_4=50$} \\
\hline
Safety constraints & \makecell{workspace boundary $\left[-0.13,0.13\right]$ \\ obstacle position $\left[-0.07,0.07\right]$ \\safe distance $r=0.04$} \\
\hline
\end{tabular}
\end{table}
Assume the desired trajectory of the end-effector is a circle with the origin as the centre and a radius of 0.14 (as can be seen in Fig. (\ref{f1b}), which can be formulated as \cite{RN100}:
\begin{small}
\begin{equation}
\label{eq11}
\begin{aligned}
x_d\left( t \right) &= 0.14\cos \left( 0.5t \right) \\
y_d\left( t \right) &=0.14\sin \left( 0.5t \right)
\end{aligned}
\end{equation}
\end{small}
The external human forces are given by \cite{RN100}:
\begin{small}
\begin{equation}
\label{eq36}
f_{e_i}\left( t \right) =\left\{ \begin{array}{l}
0\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t<5\ or\ t\ge 11\\
a_i\left( 1-\cos \pi t \right) \ \ \ \ \ \ 4\le t<5\\
2a_i\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ 5\le t<10\\
a_i\left( 1+\cos \pi t \right) \ \ \ \ \ \ 10\le t<11
\end{array} \right.
\end{equation}
\end{small}
\begin{figure}[htbp]
\begin{subfigure}{0.29\textwidth}
\centering
\includegraphics[width=.99\linewidth]{Workspace_Original_fe.eps}
\caption{Time-varying external human forces.}
\label{f1a}
\end{subfigure
\begin{subfigure}{0.19\textwidth}
\centering
\includegraphics[width=.99\linewidth]{unsafe.eps}
\caption{Unsafe trajectory.}
\label{f1b}
\end{subfigure}
\caption{Traditional admittance control.}
\label{f1}
\end{figure}
The external human forces are applied when $t=4s$ and removed at $t=11s$, as depicted in Fig. \ref{f1a}. By using the traditional admittance control, as can be seen in Fig. \ref{f1b}, the external human forces cause the deformation of the trajectory, and the resulting trajectory crosses the boundaries of the workspace (i.e., red dotted line in Fig. \ref{f1b}). In the following subsection, we integrate our proposed ECBFs-QP based adaptive admittance control into our control framework so that the safety constraints can be strictly guaranteed, while allowing the robot to comply with human interaction safely inside the task space with natural fluidity. The parameters of the simulation are given in Table. 1.
\subsection{Simulation results}
\begin{figure}[htbp]
\begin{subfigure}{0.23\textwidth}
\centering
\includegraphics[width=.99\linewidth]{Workspace.eps}
\caption{2D trajectory.}
\label{f3a}
\end{subfigure
\begin{subfigure}{0.23\textwidth}
\centering
\includegraphics[width=.99\linewidth]{Workspace_with_fe.eps}
\caption{2D trajectory.}
\label{f3b}
\end{subfigure}\\
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=.99\linewidth]{Workspace_XY.eps}
\caption{Trajectory on X and Y axis.}
\label{f3c}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=.99\linewidth]{Workspace_XY_with_fe.eps}
\caption{Trajectory on X and Y axis.}
\label{f3d}
\end{subfigure}\\
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=.99\linewidth]{Workspace_Optimal_fe.eps}
\caption{Interaction forces.}
\label{f3e}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=.99\linewidth]{Workspace_Optimal_fe_with_fe.eps}
\caption{Interaction forces.}
\label{f3f}
\end{subfigure}
\caption{Robot manipulator with workspace constraints}
\label{f3}
\end{figure}
In the case of workspace constraints, we illustrate the effectiveness of our proposed approach by applying and removing external forces. As can be seen in Fig. \ref{f3a}, the original desired trajectory $x_d$ is an ideal circle (indicated with the solid blue line). When we apply workspace constraints (i.e., rectangle indicated with the red dotted lines), the original desired trajectory is modified to the safe trajectory that is constrained within the workspace and keeps a safe distance from the boundary. When we apply the external human forces (\ref{eq36}) to both links of the robot manipulator (as shown in Fig. \ref{f3b}), the robot's trajectory changes in response to external human forces, but still remains within the restricted area. Fig. \ref{f3c} and Fig. \ref{f3d} show the Cartesian trajectory of the end-effector on the X-axis and Y-axis, separately. When $4<t<11$, compared with Fig. \ref{f3c}, the trajectory in Fig. \ref{f3d} is deformed in order to comply with the external forces. It is clear that the external loop based on the admittance control ensures the compliance of robot motion, and the internal feedback loop based on ECBFs-QP ensures the safety constraints. Fig. \ref{f3e} and Fig. \ref{f3f} show the optimal interaction forces derived by the ECBFs-QP. It is clear that our methods can provide effective compensative force feedback to admittance control such that both compliance and safety can be guaranteed.
\begin{figure}[htbp]
\begin{subfigure}{0.23\textwidth}
\centering
\includegraphics[width=.99\linewidth]{Obstacle_with_fe.eps}
\caption{2D trajectory.}
\label{f4a}
\end{subfigure
\begin{subfigure}{0.23\textwidth}
\centering
\includegraphics[width=.99\linewidth]{Obstacle_Workspace_with_fe.eps}
\caption{2D trajectory.}
\label{f4b}
\end{subfigure}\\
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=.99\linewidth]{Obstacle_XY_with_fe.eps}
\caption{Trajectory on X and Y axis.}
\label{f4c}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=.99\linewidth]{Obstacle_Workspace_XY_with_fe.eps}
\caption{Trajectory on X and Y axis.}
\label{f4d}
\end{subfigure}\\
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=.99\linewidth]{Obstacle_Optimal_fe.eps}
\caption{Interaction forces.}
\label{f4e}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=.99\linewidth]{Obstacle_Workspace_Optimal_fe.eps}
\caption{Interaction forces.}
\label{f4f}
\end{subfigure}
\caption{Robot manipulator with both two constraints}
\label{f4}
\end{figure}
In the case of obstacle avoidance, we firstly achieve obstacle avoidance only, and then synthesize the workspace constraints and obstacle constraints together in our proposed framework. As can be seen in Fig. \ref{f4a}, when we only consider obstacle avoidance, the robot moves into an unsafe proximity of the workspace boundary when trying to avoid the collision. In Fig. \ref{f4b}, after integrating the workspace constraints into the system, the robot avoids all hazards. The Cartesian trajectories of the end-effector on the X-axis and Y-axis are shown in Fig. \ref{f4c} and Fig. \ref{f4d}, and the optimal interaction forces derived from the ECBFs-QP are shown in Fig. \ref{f4e} and Fig. \ref{f4f}. It is clear that in the presence of human external forces, both compliance and safety can be ensured by our proposed method.
\section{Conclusions}\label{sec6}
In this letter, a novel control framework based on admittance control, exponential control barrier functions and the quadratic program is proposed to achieve both compliance and safety for human-robot interaction.
In particular, a virtual force feedback for admittance control is provided in real-time by using the ECBFs-QP framework as a compensator of the external human forces. Therefore, the safety of the proposed robot control system has higher robustness for external force disturbances, while simultaneously providing human-friendly dynamic behaviour. The simulation results show that the proposed approach can enforce both safety and compliance. In future work, constraints will be considered for each joint (not only the end-effector). In addition, the barrier Lyapunov function will be discussed and integrated into the system to further enforce safety.
\addtolength{\textheight}{-4cm}
\bibliographystyle{IEEEtran}
|
\section{Introduction} \label{sec:intro}
The theory of magnetospheric accretion, whereby infalling inner disk material flows along stellar magnetic field lines and forms a shock in a young star’s atmosphere, is well-established and consistent with a range of observations \citep[e.g.,][]{Koenigl1991}. X-ray emission originating from the shock front is absorbed and re-radiated as excess optical/ultraviolet Balmer continuum \citep[e.g.][]{Hartmann2016, Valenti1993, Gullbring1998, Calvet1998}, while infalling gas exhibits line emission, including the Balmer, Paschen, and Brackett series hydrogen lines.
The same accretion process has been assumed to extend to substellar masses \citep[e.g.][]{Muzerolle2005, Alcala2017}, and accretion signatures from planetary mass companions (PMCs) have been interpreted under the stellar paradigm. Recent discoveries of H$\alpha$ accretion signatures in substellar companions---both brown dwarfs (BD) \citep[e.g.,~SR12c;][]{Santamaria-Miranda2018,Santamaria-Miranda2019} and protoplanet candidates \citep[e.g., PDS 70 b and c and LkCa 15 b,][]{Haffert2019, Wagner2018, Sallum2015}---have provided incontrovertible evidence of accretion onto secondary objects in young systems. Combined with the first detections of circumplanetary disks \citep{Benisty2021}, these systems allow for direct study of planet formation processes.
Recently, \citet{Eriksson2020} discovered strong hydrogen (H$\alpha$, H$\beta$) and helium emission lines suggestive of ongoing accretion from the PMC 2MASS J01033563-5515561(AB)b, also known as Delorme 1 (AB)b. Among the first imaged circumbinary PMCs, Delorme 1 (AB)b was discovered in $L'$ band by \citet{Delorme2013} at 1\farcs77 (84 au) separation, with an estimated mass of 12--14 $M_\mathrm{Jup}$, placing it at the deuterium burning limit. Its host, Delorme 1 AB, is an M5.5 binary \citep[separation of $\sim$0\farcs25 or 12 AU;][]{Delorme2013} at 47.2 $\pm$ 3 pc \citep{Riedel2014} in the Tucana-Horologium association \citep{Gange2015}, placing its age at $\sim$30--45~Myr.
While the system shows evidence of youth, including an overluminous central binary \citep{Riedel2014}, red $JHK_s$ colors \citep[similar to other young bound nonaccreting companions;][]{Riedel2014}, and low surface gravity \citep{Liu2016},
ongoing PMC accretion at 30--45~Myr is possible, as lower mass objects have been found to have long disk dispersal timescales \citep{Luhman2022}\footnote{\citet{Luhman2022} found that in the 15-21 Myr Lower Centaurus Crux and Upper Centaurus Lupus association, disk fractions increase with decreasing mass, from 0.7\% to 9\%, indicating lower-mass stars can retain disks far longer than originally estimated ($\sim$10 Myr).}.
In this letter, we present the first detection of near infrared (NIR) emission lines from Delorme 1 (AB)b, corroborating the claim of ongoing companion accretion, and confirming the lack of accretion in the host binary system. This is the first accreting PMC with \PaB, \PaG, and Br$\gamma$ detections, and provides a critical benchmark for future NIR accretion studies of PMCs. NIR line ratios provide an important probe of the physical properties of the emitting region that can inform accretion paradigms.
\begin{figure*}[hpt!]
\centering
\includegraphics[width=\linewidth]{f1.pdf}
\caption{TripleSpec4.1 $JHK$-band spectra of Delorme 1 (AB)b (\textit{top;} green: UT 2021-11-20, magenta: UT 2022-01-24) and Delorme 1 AB (\textit{bottom;} blue: UT 2022-01-23, magenta: UT 2022-01-24)
The NIR emission lines, Pa$\gamma$, Pa$\beta$, and Br$\gamma$, are highlighted with red labels. The lines are shown in greater detail in the inset images for the companion (\textit{top}, with best fit Gaussians shown by dashed black lines) and binary host (\textit{bottom}). Other atomic and molecular features are labeled in black. Gray bands indicate regions of high atmospheric absorption.}
\label{fig1}
\end{figure*}
\section{Observations and Reductions}
Delorme 1 (AB)b was observed with the TripleSpec 4.1 Near-IR spectrograph \citep{Schlawin2014} on the Southern Astrophysical Research (SOAR) Telescope during two observing runs in 2021-2022 (ID: 2021B-0311). TripleSpec 4.1 covers 0.8-2.47 $\mu$m at moderate resolution (R$\sim$3500) with a fixed 1\farcs1 $\times$ 28\arcsec slit. Both observations were taken in good weather conditions, with seeing around 0\farcs95-1\farcs0, with the slit aligned to the parallactic angle. Delorme 1 (AB)b was observed on 2021 November 20 (epoch 1) at an airmass of 1.2. Sixteen 180~s exposures were taken in an ABBA cycle, for a total exposure time of 2880 s, yielding a final reduced spectrum with a mean SNR of $\sim90$ at $H$-band. On 2022 January 24 (epoch 2), we observed Delorme 1 (AB)b at an airmass of 1.27 with an identical observational strategy and total integration time, with the reduced spectrum achieving a mean SNR of $\sim60$ at $H$-band. We observed the binary Delorme 1 AB on 2022 January 23 (airmass 1.34), and on 2022 January 24 (airmass 1.65). We took eight 30-s exposures in an ABBA cycle, for a total of 240~s each night, yielding average final spectrum SNRs of 270 and 300 at $H$-band. As the seeing on January 23 was $\sim$1\farcs3, we were not able to sufficiently resolve the companion and did not attempt to observe it.
Spectra were reduced using a TripleSpec 4.1 version of SpeXtool \citep{Cushing2004} following the standard procedure: subtraction of A and B frames for sky removal, order identification, spectral extraction, and wavelength calibration from arc lamps. The orders were merged and areas of significant atmospheric absorption removed. A spectrophotometric standard (HIP 6364, A0V) was observed before and after Delorme 1 for both telluric correction and flux calibration, following \citet{Vacca2003} using the SpeXtool \texttt{xtellcor} software.
Due to its close distance \citep[$47.2 \pm 3.1$~pc;][]{Riedel2014}, Delorme 1 resides in the Local Bubble \citep[area of low interstellar extinction;][]{Sfeir1999}; therefore, we assume zero reddening.
\section{Results}
We detect strong \PaB, \PaG, and \BrG emission lines (Figure~\ref{fig1}) in Delorme 1 (AB)b in both epochs.
Hydrogen emission lines are not detected in the host binary (see Table~\ref{tab:results} for line flux upper limits),
providing strong confirmation they are unique to the companion.
We compute equivalent widths (EW), fluxes (\Fline) and luminosities (\Lline) for each line and epoch (Table~\ref{tab:results}).
Line fluxes are computed by integrating under a best-fit Gaussian profile after subtracting a linear fit to the local continuum. The uncertainty on the line is a function of the scatter in the continuum and the best-fit Gaussian given by
\begin{equation}
\sigma = \sqrt{N_\mathrm{pix}} \times F_\mathrm{noise} \times \Delta\lambda,
\end{equation}
where $N_\mathrm{pix}$ is the number of pixels within 3$\times$FWHM, $F_\mathrm{noise}$ is the rms of the local continuum, and $\Delta\lambda$ is the wavelength resolution per pixel at each line. EWs are obtained from the ratio of line fluxes to the average local continuum level within a 50 \AA{} window on either side of the line. We estimate EW uncertainties following \citet{Vollmann2006}.
We do not detect \BrG in epoch 2, potentially due to poorer seeing conditions. For non-detected lines, we calculate \Fline upper limits as $F_\mathrm{line}^\mathrm{upp} = 3\sigma$.
\begin{deluxetable*}{@{\extracolsep{4pt}}cccccc@{}ccc}
\tablenum{1}
\tablecaption{Delorme 1 (AB)b Line Characteristics\label{tab:results}}
\tablewidth{0pt}
\tablehead{
\colhead{} & \colhead{} & \colhead{} & \colhead{} & \multicolumn{2}{c}{stellar scaling\tablenotemark{a}} & \multicolumn{2}{c}{planetary scaling\tablenotemark{b}} & \multicolumn{1}{||c}{} \\
\cline{5-6}
\cline{7-8}
\colhead{Line} & \colhead{EW} & \colhead{\Fline} & \colhead{\Lline} & \colhead{$\log(\Lacc)$} & \colhead{$\log(\Mdot)$} & \colhead{$\log(\Lacc)$} & \colhead{$\log(\Mdot)$} & \multicolumn{1}{||c}{\multirow{-1.5}{*}{\parbox{2.5cm}{\centering Delorme 1 AB \Fline}}} \\
\colhead{} & \colhead{(\AA)} & \colhead{(10$^{-16}$ erg/s/cm$^2$)} & \colhead{($10^{-8} \Lsun$)} & \colhead{($\Lsun$)} & \colhead{($ M_\mathrm{J}$ yr$^{-1}$)}& \colhead{($\Lsun$)} & \multicolumn{1}{c}{($ M_\mathrm{J}$ yr$^{-1}$)} & \multicolumn{1}{||c}{(10$^{-15}$ erg/s/cm$^2$)}
}
\startdata
\multicolumn{8}{c}{UT 2021-11-20} & \multicolumn{1}{||c}{}\\
\hline
\PaG & -1.95$\pm$0.74 & 6.82$\pm$1.33 & 4.75$\pm$1.11 & -5.50$\pm$0.53 & -8.82$\pm$0.53 & -3.94$\pm$0.30 & -7.27$\pm$0.30 & \multicolumn{1}{||c}{$<$3.67}\\
\PaB & -2.31$\pm$0.88 & 8.05$\pm$1.49 & 5.60$\pm$1.27 & -4.92$\pm$0.62 & -8.25$\pm$0.62 & -4.02$\pm$0.30 & -7.35$\pm$0.30 & \multicolumn{1}{||c}{$<$3.71} \\
\BrG & -2.08$\pm$1.11 & 1.64$\pm$0.56 & 1.14$\pm$0.42 & -5.43$\pm$0.94 & -8.75$\pm$0.96 & -3.91$\pm$0.30 & -7.24$\pm$0.30 & \multicolumn{1}{||c}{$<$1.01}\\
\hline
\multicolumn{8}{c}{UT 2022-01-24} & \multicolumn{1}{||c}{}\\
\hline
\PaG & -1.24$\pm$0.47 & 2.94$\pm$0.77 & 2.05$\pm$0.61 & -5.95$\pm$0.55 & -9.28$\pm$0.56 & -4.25$\pm$0.30 & -7.58$\pm$0.30 & \multicolumn{1}{||c}{$<$7.91} \\
\PaB & -1.44$\pm$0.62 & 3.49$\pm$0.85 & 2.43$\pm$0.67 & -5.31$\pm$0.64 & -8.64$\pm$0.65 & -4.33$\pm$0.30 & -7.67$\pm$0.30 & \multicolumn{1}{||c}{$<$6.56} \\
\BrG & -- & $<$0.74 & $<$0.52 & $<$-5.81 & $<$-9.17 & $<$-4.20 & $<$-7.53 & \multicolumn{1}{||c}{$<$1.86} \\
\enddata
\tablenotetext{a}{\Lacc--\Lline scaling relation from \citet{Alcala2017}}
\tablenotetext{b}{\Lacc--\Lline scaling relation from \citet{Aoyama2021}}
\end{deluxetable*}
During magnetospheric accretion, the infalling column of gas is heated to $\sim10^4$ K, producing broad emission lines \citep{Hartmann2016} such as \PaB, \PaG, and \BrG. The gas travels at free-fall velocity, and heats to $10^6$ K when it shocks at the stellar photosphere, fully ionizing and preventing the formation of hydrogen line emission.
In contrast, recent simulations of accreting PMCs \citep{Aoyama2018, Aoyama2020} suggest differences in the physical conditions of the shocked region. Due to smaller masses and lower surface gravities, accreting gas travels at lower free fall velocities, leading to a non-fully ionized post-shock region. This results in shock-heated accreting gas capable of hydrogen line emission \citep{Aoyama2018}.
In other words, the detections of Paschen and Brackett-series emission from accreting objects are an unambiguous sign of accretion; however, the dominant source of line emission may be either the infalling accretion column or the post-shock region.
Given this ambiguity, we estimate accretion rates for Delorme 1 (AB)b following both families of accretion models, and discuss the differences below.
The mass accretion rate is given by:
\begin{equation}
\dot{M} = \left(1-\frac{R_\star}{R_\mathrm{in}}\right)^{-1} \frac{\Lacc R_\star}{GM_\star},
\end{equation}
where $R_\mathrm{in}$ is the inner disk radius, assumed to be 5 $R_\star$ \citep[e.g.,][]{Gullbring1998, Herczeg2008, Rigliaco2012, Alcala2017}, $R_\star$ is the radius of the accreting object, $M_\star$ is its mass, and \Lacc is the estimated accretion luminosity.
Total accretion luminosity has been found to strongly correlate with emission line luminosities in T Tauri stars \citep{Rigliaco2012, Alcala2014, Alcala2017} as
\begin{equation}
\log (L_\mathrm{acc}/L_\odot) = a\times\log (L_\mathrm{line}/L_\odot) + b,
\end{equation}
where $a$ and $b$ are the fit coefficients for each line.
These relationship can be used to estimate $\dot{M}$ from a single accretion-tracing line. However, \citet{Aoyama2020, Aoyama2021} argue that the \Lline--\Lacc relationships are not valid for planetary mass objects because of the different physical conditions of the emitting region. \citet{Aoyama2021} derived new theoretical \Lacc--\Lline relationships expected for PMCs based on the \citet{Aoyama2018} shock models. We refer to all accretion luminosities and mass accretion rates derived from \citet{Alcala2017} \Lacc--\Lline scaling relations as ``stellar" (e.g., ~$L_\mathrm{acc, ste}$/$\dot{M}_\mathrm{ste}$) and those derived from \citet{Aoyama2021} as ``planetary" (e.g., ~$L_\mathrm{acc, pla}$/$\Mdotpla$) for ease of distinguishing between the two.
\begin{figure}[tp]
\hspace{-0.6cm}
\includegraphics[width=0.5\textwidth]{f2.pdf}
\caption{Accretion rates for individual emission lines. Circles indicate data from \citet{Eriksson2020}. Stars and diamonds represent our epochs 1 and 2, respectively.
Accretion rates derived using both ``stellar" empirical scaling relations \citep[open symbols, ][]{Alcala2017} and ``planetary" accretion models \citep[filled symbols, ][]{Aoyama2021} are shown. \Mdot(H$\alpha$) is also estimated using the line luminosity model of \citet[][light gray]{Thanathibodee2019}.
\Mdot(H$\alpha$ 10\%) is estimated using a ``stellar" empirical relation \citep[dark gray, ][]{Natta2004}. }
\label{fig2}
\end{figure}
Following \citet{Eriksson2020}, we assume a companion mass of $M_p = 0.012\ M_\odot$ and radius $R_p = 0.163\ R_\odot$. We calculate $\Lacc$ following the \Lacc--\Lline scaling relations calibrated empirically for stars (\PaB: ($a,b$)=(1.06, 2.76), \PaG: ($a,b$)=(1.24, 3.58), \BrG: ($a,b$)=(1.19, 4.02)) by \citet{Alcala2017} and theoretically for PMCs (\PaB: ($a,b$)=(0.86, 2.21), \PaG: ($a,b$)=(0.85, 2.28), \BrG: ($a,b$)=(0.85, 2.84)) by \citet{Aoyama2021}. This allows us to directly compare our NIR-derived results to the accretion rates estimated by \citet{Eriksson2020}.
Our \Mdot estimates are given in Table~\ref{tab:results} for both the ``stellar" and ``planetary" relations. \Mdot estimates are relatively consistent among lines and epochs under each scaling relation; however, the \citet{Aoyama2021} models predict \Mdot's that are systematically higher by several orders of magnitude.
On average, using the stellar scaling relations of \citet{Alcala2017} we find a
$\log(\dot{M}_\mathrm{ste})$ of $-8.53 \pm 0.28 \ M_\mathrm{J}\ \mathrm{yr}^{-1}$ for epoch 1 and $-8.85\pm0.28\ M_\mathrm{J}\ \mathrm{yr}^{-1}$ for epoch 2. Using the \citet{Aoyama2021} planetary shock-model relations, we find $\log(\Mdotpla) = -7.38\pm0.23\ M_\mathrm{J}\ \mathrm{yr}^{-1}$ and $-7.19\pm0.31\ M_\mathrm{J}\ \mathrm{yr}^{-1}$ for epochs 1 and 2, respectively.
\begin{figure*}[htp!]
\centering
\includegraphics[width=\linewidth]{f3.pdf}
\caption{Predicted line ratios for local line excitation \citep[``stellar",][blue squares]{Kwan2011}, and accretion shock emission \citep[``planetary",][purple/yellow circles]{Aoyama2018} models. Delorme 1 (AB)b's NIR line ratios (panels a-c) are shown in black with markers (stars, diamonds) indicating epoch as in Figure~\ref{fig2}. Panel d shows Pa$\gamma$/H$\alpha$ line ratios, with Delorme 1 (AB)b values (unfilled markers) derived from our Pa$\gamma$ observations and the non-contemporaneous H$\alpha$ measurement of \citet{Eriksson2020}. }
\label{fig3}
\end{figure*}
\section{Discussion}
We have presented mass accretion rate estimates for Delorme 1 (AB)b derived from NIR hydrogen emission lines under two assumed scalings of \Lline to $\Lacc/\dot{M}$. Accretion rate estimates for individual NIR lines agree with one another within the ``planetary'' and ``stellar'' accretion paradigms, with the exception of the ``stellar'' Pa$\beta$ accretion rate, which is marginally inconsistent with the other ``stellar" accretion estimates.
In Figure~\ref{fig2}, we compare our NIR observations (diamonds/stars) with the marginally-resolved H$\alpha$ observations of \citet{Eriksson2020} (gray/black circles) and convert each to \Mdot using both ``stellar" (unfilled symbols, \citealt{Alcala2017}; dark gray, \citealt{Natta2004}) and ``planetary" (filled symbols, \citealt{Aoyama2019}; light gray, \citealt{Thanathibodee2019}) scaling relations. H$\alpha$ can originate from chromospheric activity, complicating its interpretation. \citet{Eriksson2020} found that the contribution to the H$\alpha$ line profile due to chromospheric activity should be minimal at this age, pointing toward Delorme 1 (AB)b experiencing ongoing accretion.
We find that our NIR \Mdot's generally agree with the \citet{Eriksson2020} estimates within uncertainties, albeit with slightly higher \Mdot values relative to the Balmer series, though our \PaB measurement is marginally inconsistent at the 1$\sigma$ level.
Given the strength of the companion's NIR EWs relative to diagnostics measured for active low-mass stars \citep[$\sim0.04-0.05$ \AA; e.g.,][]{Schofer2019}, our results are most consistent with the presence of PMC accretion.
\begin{figure*}[htp]
\centering
\includegraphics[width=\linewidth]{f4.pdf}
\caption{Mass accretion rate vs mass for all known isolated substellar accretors (gray), planetary mass companions (black squares), and a representative sample of low mass stars. Delorme 1 (AB)b's NIR derived \Mdot is highlighted (colored markers, symbol as in Fig~\ref{fig2}). The canonical $\dot{M} \propto M^{2.1}$ \citep{Muzerolle2005} relation for higher mass objects (consistent with formation via collapsing prestellar cores) is shown (solid line), and a predicted relation for substellar objects formed via disk fragmentation \citep[dashed line, $\alpha\sim0.001$, ][]{Stamatellos2015}. Delorme 1 AB(b) Pa$\gamma$ and Br$\gamma$ measurements have been offset in mass for clarity.}
\label{fig4}
\end{figure*}
We find agreement between $\Mdotpla$ and $\dot{M}$(H$\alpha$ 10\%); both are $\sim$1.5 mag higher than $\dot{M}_\mathrm{ste}$.
As $\dot{M}$(H$\alpha$ 10\%) does not rely on scaling relationships, accurate continuum subtraction, or extinction, it is considered a robust independent measure of accretion \citep{White2003, Stelzer2007}, including for the lowest mass accreting protoplanets \citep[e.g., PDS 70b;][]{Haffert2019}. As noted by \citet{Alcala2014}, the empirical relationship between H$\alpha$ 10\% width and \Mdot \citep{Natta2004} has considerable scatter, and line luminosities should also be used when possible.
However, the strong agreement between $\dot{M}$(H$\alpha$ 10\%) and $\Mdotpla$ could indicate that $\Mdotpla$ is a more accurate estimate of \Mdot for Delorme 1 (AB)b.
The marginal inconsistency in $\dot{M}_\mathrm{ste}$ could be a result of applying stellar scaling relations to an object accreting under a different paradigm; this is not seen in the $\Mdotpla$s.
To independently determine the accretion paradigm most consistent with Delorme 1 (AB)b without a reliance on scaling relations, line ratios can be used.
NIR hydrogen lines are ideal for measuring accretion line ratios \citep[see][]{Edwards2013, Bary2008} due to their small line opacities, resulting in little blue or redshifted absorption from winds or infalling gas \citep{Folha2001, Edwards2006}. By comparing observed line ratios to accretion model prediction, we can probe physical conditions of the emitting region such as number density, temperature, and infall velocity. Line ratios have discriminating power between different physical line emission sources, as different accretion models predict different line ratios.
To this end, we consider two models: the local line excitation model of \citet{Kwan2011} and the planetary shock model of \citet{Aoyama2018}. As shown in Figure~\ref{fig3}, the predicted line ratios of post-shock gas in a planetary atmosphere \citep[planetary paradigm,][purple/yellow circles]{Aoyama2018} can vary from those predicted by local line excitation models developed to describe infalling accretion columns of T Tauri stars \citep[stellar paradigm,][green/blue squares]{Kwan2011}, allowing us to infer which model better matches observations, though there is some overlap for lower densities, where we are not able to distinguish between accretion paradigms.
We calculate line ratios for each line pair and epoch (star/diamond symbols) over the whole emission range\footnote{In T Tauri stars, winds and outflow absorption can affect line ratios. As such, residual line profiles selected over regions with no opacity effects are used to calculate line ratios \citep[see][]{Edwards2013}. However, these are assumed to be negligible in PMCs.}.
In panel d, we include ratios with respect to published H$\alpha$ emission for context, noting these observations were not obtained contemporaneously with our NIR data. Line ratios may be affected by intrinsic and instrumental variability; therefore, inconsistency of the H$\alpha$ ratio with either model grid may not be indicative of variability in the physical conditions of the emitting region.
For all measurements, observed line ratios fall nearest the \citet{Aoyama2018} models and consistently diverge from the \citet{Kwan2011} models, suggesting that planetary scaling relations are likely more appropriate in this situation. Therefore, we use the \citet{Aoyama2021} models and relations for further analysis.
We extract all model physical input parameters consistent with observed line ratios within uncertainties. We find that the best-fitting models have preshock velocities of $70-170$ km/s and number densities of $10^{13-14}$ cm$^{-3}$. While the preshock velocity is consistent with measured \Mdot's and assumed mass \citep[and radius; see Figure 13 of][]{Aoyama2020}, the number density is higher than expected for the measured \Mdot assuming a pure planetary shock model. This could be explained by shock emission with a low filling factor resulting from a magnetospheric accretion flow, absorption in the post-shock region \citep{Hashimoto2020}, strong accretion column extinction \citep[][though they found that the \Mdot is too low for absorption by either gas or dust in the accretion flow]{Marleau2022}, or circumplanetary disk extinction in the line of sight \citep{Aoyama2020}. High resolution (R$\sim$10,000) spectra will help disentangle the accretion flow geometry and shed light on the nature of the accretion shock, as resolved line profiles can distinguish between geometries \citep{Aoyama2020, Marleau2022}.
In Figure~\ref{fig4}, we show the \Mdot--$M$ relation for all known accreting substellar objects, together with low mass stars (Betti et al., in prep.). The $\Mdotpla$'s for Delorme 1 (AB)b lie above the canonical $\Mdot\sim M^{2.1}$ \citep{Muzerolle2005} T Tauri star relation consistent with formation via collapsing prestellar cores. The mass accretion rates are similar to other bound planetary mass companions (black squares),
whose previous accretion rate estimations mostly come from H$\alpha$ line luminosity or H$\alpha$ 10\% width. The location of these bound PMCs in \Mdot--$M$ space is consistent with model predictions of PMC formation through disk fragmentation in disks with low viscosities \citep{Stamatellos2015}. These models predict higher accretion rates; companions that form in dynamically unstable systems have larger than expected gas mass reservoirs, allowing them to accrete more material \citep{Stamatellos2015} for longer. The high \Mdot observed for Delorme 1 (AB)b suggests that it may have formed via disk fragmentation. Its \Mdot is most consistent with \citet{Stamatellos2015} models with low disk viscosity ($\alpha \sim$0.001), and is comparable to PMCs with similar masses such as GSC~06214-00210~b, GQ~Lup~b, and DH~Tau~b, all of which have been theorized to have formed via disk fragmentation \citep{Stamatellos2015, Zhou2014}.
In summary, the strong Pa$\beta$, Pa$\gamma$, and Br$\gamma$ emission seen from Delorme 1 (AB)b indicates strong ongoing mass accretion onto the PMC.
Utilizing line ratios, we find that the NIR hydrogen emission is most consistent with models of planetary shock accretion, though the high predicted number density does not exclude magnetospheric accretion from occurring as well on the planetary surface. We conclude that higher \Mdot estimates derived from planetary scaling relations are more likely to reflect the true accretion rate, and the position of Delorme 1 (AB)b in \Mdot--$M$ space is consistent with formation via disk fragmentation. This would account for its high accretion rate, which is consistent with low disk viscosity, likely resulting in slower disk evolution and perhaps explaining why this 30--45~Myr object is still actively accreting \citep[potentially a ``Peter Pan disk";][]{Silverberg2020}. Detailed modeling of the planetary surface and disk will provide a clearer understanding of Delorme 1 (AB)b, and future observations of a wider range of line ratios will help constrain the nature of the accretion shock. Forthcoming work (Betti et al, in prep) will present detections of NIR accretion for a comprehensive sample of accreting BDs and PMCs as well as observational \Lacc--\Lline empirical relationships for the substellar regime in order to help constrain substellar formation mechanisms.
Delorme 1 (AB)b is a benchmark accreting PMC, with current observations and theoretical models suggesting the nature of its emission is in the planetary regime.
\acknowledgments
We thank the anonymous referee for their careful review.
We thank the SOAR support scientist, CTIO scientist Sean Points, for allocating some engineering time for our observations.
S.K.B. and K.B.F. acknowledge support from NSF AST-2009816.
G-DM acknowledges the support of the DFG priority program SPP 1992 ``Exploring the Diversity of Extrasolar Planets'' (MA~9185/1) and from the Swiss National Science Foundation under grant
200021\_204847 ``PlanetsInTime''.
Parts of this work have been carried out within the framework of the NCCR PlanetS supported by the Swiss National Science Foundation.
\vspace{5mm}
\facilities{SOAR(TripleSpec4.1)}
\software{astropy \citep{2013A&A...558A..33A,2018AJ....156..123A}, specutils \citep{specutils}, Spextool \citep{Vacca2003, Cushing2004}, Matplotlib \citep{Hunter2007}}
|
\section{Introduction}
Due to the rapid and advanced progress in the field of laser technology since the early 1960s, lasers have become widely used in industrial, medical, commercial, scientific and military domains. Lasers can provide us with sources having extreme properties in terms of energy, pulse width and wavelength, helping researchers to understand the fundamental concepts of radiation-matter interaction. Development of lasers with shorter wavelengths, shorter pulses and higher intensities continues unabated. The achievement of a maximum laser intensity of $10^{22}$ W/cm$^{2}$ \cite{yoon2019achieving} should lead to a better understanding of the behavior of various scattering processes \cite{ghatak2021lasers}. Therefore, lasers are currently indispensable tools for investigating physical processes, in particular, laser-matter interaction. The early studies of laser-assisted scattering in the nonrelativistic regime and at moderate field strengths are well established and documented in the literature. A comprehensive overview of this can be found in the books of Mittleman \cite{Mittleman}, Faisal \cite{Faisal}, Delone \cite{Delone}, Fedorov \cite{FEDOROV} and in some recent reviews \cite{Ehlotzky2001, Francken-Joachain,ehlotzky2009fundamental}. With the advent of very powerful laser sources, it has become important to consider laser-assisted processes in the relativistic regime. Therefore, in a laser field of relativistic intensity, many processes have been studied such as Mott scattering in an elliptically and linearly polarized laser field \cite{attaourti2004mott, Li-Berakdar2003}, laser-assisted bremsstrahlung for circular and linear polarization \cite{schnez2007laser}. Electron-proton elastic scattering has been investigated in the presence of a linearly and circularly polarized laser field in \cite{Dahiri2021, wang2019relativistic, Liu2014}. In addition, in \cite{du2018new, du2018Nonlinear}, the authors studied new phenomena in laser-assisted scattering of an electron (positron) by a muon with different polarizations. There are also some papers that studied decay processes in the presence of laser field \cite{mouslihplb,MOUSLIH2020,decay1,decay2,decay3}. In this paper, using the first Born approximation, we give complete analytical and numerical results of the scattering process $e^{-}+ \nu_{\mu} \rightarrow e^{-} + \nu_{\mu}$, assisted by an elliptically polarized laser field. These results can generate those found in previous works for a circularly polarized laser field by El Asri \textit{et al.} \cite{asri2021elastic} and for a linearly polarized laser field by Bai \textit{et al.} \cite{bai2012multiphoton}. In addition, this process has been recently studied in the framework of electroweak theory for a circularly polarized laser field by \cite{asri2022}. Overall, the goal of this work is to generalize the previous research, with a more detailed calculation in the presence of an elliptically polarized laser field. We have also compared the differential cross section (DCS) in the absence and presence of a laser field at different polarizations, and examined its dependence on the laser parameters. For the elliptical polarization of the laser field, our current calculations lead to a new form of ordinary and generalized Bessel functions \cite{korsch2006two,reiss1980effect,dattoli1992generating}. The paper is organized as follows. In Sec. \ref{THEORY}, we establish the detailed analytical calculation of the S-matrix element, in the first Born approximation, of the process $e^{-}+ \nu_{\mu} \rightarrow e^{-} + \nu_{\mu}$, as well as the expression for the DCS in the presence of an external elliptically polarized laser field. Then, in Sec. \ref{NUMERICAL_RESULTS_AND_DISCUSSION}, we present numerical results for the laser-assisted DCS and discuss its dependence on the relevant parameters. Sec. \ref{CONCLUSION} is devoted to conclusions. Throughout this work, we use the natural units ($\hbar = c =1$) and the Minkowski metric tensor $g^{\mu \nu} = \text{diag}(1,-1,-1,-1)$.
\section{Theory} \label{THEORY}
We consider the scattering process of a muon neutrino by an electron schematized as follows:
\begin{equation}\label{eq.1}
e^{-}(p_{i},s_{i}) + \nu_{\mu}(k_{i},t_{i}) \rightarrow e^{-}(p_{f},s_{f}) + \nu_{\mu}(k_{f},t_{f}),
\end{equation}
where the labels $(p_{i,f},k_{i,f})$ and $(s_{i,f},t_{i,f})$ are the associated four-momenta and spin respectively, with $i$ and $f$ stand for the initial and final states. In the framework of electroweak theory, this scattering process is mediated by the exchange of only the neutral $Z$ boson. The corresponding lowest order Feynman diagram is given by Fig.~\ref{Fig.Diagramme}.
\begin{figure}[hbtp]
\centering
\includegraphics[scale=0.3]{fig1.eps}
\caption{The tree-level Feynman diagram at the $t$-channel of the process $ e^{-} + \nu_{\mu} \rightarrow e^{-} + \nu_{\mu} $ in the framework of electroweak theory. The time flows from left to right.}\label{Fig.Diagramme}
\end{figure}\\
We treat this process in the presence of an elliptically polarized laser field described by the following classical four-potential \cite{attaourti2004mott}
\begin{equation}\label{eq.2}
\begin{split}
A^{\mu}= a_{1}^{\mu} \cos(\phi) + a_{2}^{\mu} \sin(\phi) \tan(\eta/2),
\end{split}
\end{equation}
where $ \phi= k^{\mu}x_{\mu} =(k.x)$, and $ \eta $ is the degree of ellipticity of the external electromagnetic (EM) field. The linear polarization is obtained for $ \eta= 0^{\circ} $, while the circular polarization is obtained for $ \eta= 90^{\circ} $. We choose the wave four-vector as $ k^{\mu}=(\omega,\textbf{k})=\omega(1,0,0,1) $. The polarization four-vectors $ a_{1}^{\mu}=(0,\textbf{a}_1)=|\textbf{a}|(0,1,0,0) $ and $ a_{2}^{\mu}=(0,\textbf{a}_2)=|\textbf{a}|(0,0,1,0)$ are along the $x$ and $y$-axis, respectively. $ a_{1}^{\mu} $ and $ a_{2}^{\mu} $ satisfy the normalization $ a_{1}^{2}=a_{2}^{2}=a^{2}=-|\textbf{a}|^{2}=-(\mathcal{E}_{0}/\omega)^{2} $ and the orthogonality conditions $ (a_{1}. a_{2})=0 $, with $ \mathcal{E}_{0} $ is the laser field strength and $ \omega $ is the frequency. Moreover, the four-potential $ A^{\mu} $ satisfies the Lorenz gauge condition $k_{\mu} A^{\mu}=0$, which implies that $ k_{\mu} a_{1}^{\mu}=k_{\mu} a_{2}^{\mu}=0 $, forcing the wave vector $ \textbf{k} $ to be along the $z$-axis. The incoming and outgoing muon neutrinos, which do not interact with the laser field, are treated as non-mass particles, described by Dirac wave functions, normalized to the volume $ V$ given by the following formula \cite{greiner1996gauge}
\begin{equation}\label{eq.3}
\begin{split}
\begin{matrix}
\psi^{i,f}_{\nu_{\mu}}(x) = \dfrac{1}{\sqrt{2 k^{0}_{i,f} V}} u_{\nu_{\mu}}(k_{i,f},t_{i,f})~ e^{-i k_{i,f}. x}.
\end{matrix}
\end{split}
\end{equation}
In the presence of a laser field, the electron obeys the following Dirac-Volkov equation \cite{volkov1935solution}
\begin{equation}
\begin{split}
\big\lbrace (\widehat{p_{i,f}}-e A)^{2}-m^{2}-\dfrac{1}{2} i e F_{\mu \nu} \sigma^{\mu \nu} \big\rbrace \psi^{i,f}_{e^{-}}(x)=0,
\end{split}
\end{equation}
where $ e = -|e| < 0 $ is the electric charge of electron. Dressed by an elliptically polarized laser field, the incoming and outgoing electrons can be considered as Dirac-Volkov states normalized to volume $V$ \cite{volkov1935solution}
\begin{equation}\label{eq.5}
\begin{split}
\psi^{i,f}_{e^{-}}(x) =\Big[ 1+ \dfrac{e}{2(k.p_{i,f})} \Big( \slashed k \slashed a_{1} \cos(\phi) + \slashed k \slashed a_{2} \sin(\phi)\tan(\eta/2) \Big) \Big]\dfrac{u_{e}(p_{i,f},s_{i,f})}{\sqrt{2 Q_{i,f} V}} e^{i S(q_{i,f},x)},
\end{split}
\end{equation}
where
\begin{equation}
\begin{split}
S(q_{i,f},x)=& -(q_{i,f}.x)-\dfrac{e(p_{i,f}.a_{1})}{(k.p_{i,f})}\sin(\phi) +\dfrac{e^{2} a^{2}}{8(k.p_{i,f})}(1-\tan^{2}(\eta/2)) \sin( 2 \phi) +\dfrac{e(p_{i,f}.a_{2})}{(k.p_{i,f})} \tan(\eta/2)\cos(\phi),
\end{split}
\end{equation}
where $ u_{\nu_{\mu}}(k_{i,f},t_{i,f})$ and $u_{e}(p_{i,f},s_{i,f})$ represent the Dirac bispinors satisfying $\sum_{t_{i,f}} \bar u_{\nu_{\mu}}(k_{i,f},t_{i,f}) u_{\nu_{\mu}}(k_{i,f},t_{i,f})= \slashed k_{i,f} $ and $\sum_{s_{i,f}} \bar u_{e}(p_{i,f},s_{i,f}) u_{e}(p_{i,f},s_{i,f})= \slashed p_{i,f} + m_{e}$. $ q_{i,f}=(Q_{i,f},\textbf{q}_{i,f})$ is the Volkov momentum of the electron in the presence of a laser field. That is
\begin{equation}
q_{i,f}^{\mu}=p_{i,f}^{\mu} - \dfrac{e^{2} a^{2}}{4(k.p_{i,f})}(1+\tan^{2}(\eta/2)) k^{\mu}.
\end{equation}
The square of this four-momentum shows that the mass of the dressed electron (effective mass) is proportional to the strength of the EM field as follows :
\begin{equation}
q_{i,f}^{2}= m_{*}^{2}=m_e^{2} - \dfrac{e^{2} a^{2}}{2}(1+\tan^{2}(\eta/2)),
\end{equation}
where $m_e$ is the mass of electron, and the quantity $ m_{*} $ represents the effective mass of the electron in the elliptically polarized EM field. \\
In the first Born approximation, the transition matrix element can be expressed using Feynman rules as follows:
\begin{equation}\label{eq.7}
\begin{split}
S_{fi}(e^{-} \nu_{\mu} \rightarrow e^{-} \nu_{\mu}) = & -i \int d^{4}x \int d^{4}y \Big[\bar \psi^{f}_{\nu_{\mu}}(x) \dfrac{-i g}{4 \cos(\theta_{W})} \gamma^{\mu}(1-\gamma^{5}) \psi^{i}_{\nu_{\mu}}(x) \Big] D_{\mu \nu }(x-y) \\ & \times \Big[\bar \psi^{f}_{e}(y)\dfrac{-i g}{2 \cos(\theta_{W})} \gamma^{\nu}(g_{V}-g_{A} \gamma^{5}) \psi^{i}_{e}(y) \Big],
\end{split}
\end{equation}
with $g_{V}=1/2 - 2 \sin^{2}(\theta_{W})$ and $g_{A}=-1/2$ are the coupling constants \cite{renton1990electroweak}, where $\theta_{W}$ is the Weinberg angle. Here, we choose $g_{V}=0,043$ and $g_{A}=-0,545$ \cite{greiner1996gauge}. $D_{\mu \nu }(x-y) $ is the Feynman propagator for the coupling between the $Z$-boson and the fermions given by \cite{greiner2009salam}
\begin{equation}\label{eq.8}
\begin{split}
D_{\mu \nu }(x-y) = - i \int \dfrac{d^{4}q}{(2\pi)^{4}} \dfrac{e^{- i q(x-y)}}{q^{2}-M_{Z}^{2}+i M_{Z} \Gamma_{Z}} ~\Big[ g_{\mu \nu} -\dfrac{q_{\mu} q_{\nu}}{M_{Z}^{2}} \Big],
\end{split}
\end{equation}
where $ M_{Z}=(91.1876 \pm 0.0021)~\text{GeV} $ is the rest mass of the $Z$-boson and $\Gamma_{Z}=(2.4952 \pm 0.0023)~\text{GeV}$ is its total decay rate \cite{particle2020review}. After inserting Eqs.~(\ref{eq.3}), (\ref{eq.5}) and (\ref{eq.8}) into Eq.~(\ref{eq.7}) and after some algebraic manipulations, we find
\begin{equation} \label{eq.9}
\begin{split}
S_{fi}(e^{-} \nu_{\mu} \rightarrow e^{-} \nu_{\mu})=& \dfrac{ g^{2}}{8 \cos^2(\theta_{W})~(q^{2}-M_{Z}^{2}+i M_{Z} \Gamma_{Z})}\int d^{4}xd^{4}y\dfrac{d^{4}q}{(2\pi)^{4}} \dfrac{e^{-iq(x-y)}e^{i(k_{f}-k_{i}).x}}{\sqrt{16k^{0}_{i}k^{0}_{f}Q_{i}Q_{f} V^{4}}} \\ & \times e^{i(S(q_{i},y)-S(q_{f},y))} \Big[\bar{u}_{\mu_{\mu}}(k_{f},t_{i})\gamma^{\mu}(1-\gamma_{5})u_{\nu_{\mu}}(k_{i},t_{f}) \Big] \\ & \times \Big[\bar{u}_{e}(p_{f},s_{f})\Big(1+\dfrac{e\slashed{A}\slashed{k}}{2(k.p_{f})}\Big)
\gamma_{\mu}(g_{V}-g_{A}\gamma^{5})\Big(1+\dfrac{e\slashed{k}\slashed{A}}{2(k.p_{i})}\Big)~u_{e}(p_{i},s_{i}) \Big].
\end{split}
\end{equation}
We expand the term $e^{i(S(q_{i},y)-S(q_{f},y))}$ in Eq.~(\ref{eq.9}) under the following transformation
\begin{equation}\label{eq.10}
\begin{split}
e^{i(S(q_{i},y)-S(q_{f},y))}= e^{i(q_{f}-q_{i}).y}~e^{i \big( \xi_{1} \sin(\phi) + \xi_{2} \sin(2\phi) \big)}~e^{i \xi_{3} \cos(\phi)},
\end{split}
\end{equation}
where
\begin{equation}\label{eq.11}
\xi_{1}= e \Big( \dfrac{p_{f}.a_{1}}{k.p_{f}} -\dfrac{p_{i}.a_{1}}{k.p_{i}} \Big),~~\xi_{2}= \dfrac{-e^{2} a^{2}}{8} \Big( \dfrac{1}{k.p_{f}} -\dfrac{1}{k.p_{i}} \Big) \Big( 1-\tan^{2}(\eta/2) \Big),~~\xi_{3}=-e \Big( \dfrac{p_{f}.a_{2}}{k.p_{f}} -\dfrac{p_{i}.a_{2}}{k.p_{i}} \Big) \tan(\eta/2).
\end{equation}
By introducing Eq.~(\ref{eq.10}) and using $ g^{2}/(8 \cos^2(\theta_{W}))=G_{F} M^{2}_{Z}/\sqrt{2} $, where $\textit{G}_{F}=(1.16637\pm 0.00002)\times10^{-11}~\text{MeV}^{-2}$ is the Fermi coupling constant, the expression of S-matrix becomes
\begin{equation} \label{eq.12}
\begin{split}
S_{fi}(e^{-} \nu_{\mu} \rightarrow e^{-} \nu_{\mu})=& \dfrac{ G_{F} M^{2}_{Z}}{\sqrt{2}~(q^{2}-M_{Z}^{2}+i M_{Z} \Gamma_{Z})}\int d^{4}xd^{4}y\dfrac{d^{4}q}{(2\pi)^{4}} \dfrac{e^{i(k_{f}-k_{i}-q).x}~e^{i(q_{f}-q_{i}+q).y}}{\sqrt{16k^{0}_{i}k^{0}_{f}Q_{i}Q_{f} V^{4}}} \\ & \times e^{i \big(\xi_{1} \sin(\phi) + \xi_{2} \sin(2\phi)\big)} e^{i \xi_{3} \cos(\phi)}~ \Big[\bar{u}_{\mu_{\mu}}(k_{f},t_{i})\gamma^{\mu}(1-\gamma_{5})~u_{\nu_{\mu}}(k_{i},t_{f}) \Big] \\ & \times \Big[\bar{u}_{e}(p_{f},s_{f})\Big(\chi_{0 \mu} +\chi_{1 \mu} \cos(\phi) + \chi_{2 \mu} \sin(\phi) + \chi_{3 \mu} \cos^{2}(\phi) \Big) u_{e}(p_{i},s_{i}) \Big],
\end{split}
\end{equation}
where the quantities $\chi_{0 \mu}$, $\chi_{1 \mu}$, $\chi_{2 \mu}$ and $\chi_{3 \mu}$ are expressed as follows :
\begin{equation} \label{eq.13}
\begin{split}
&\chi_{0 \mu} = \gamma_{\mu}( g_{V}-g_{A} \gamma_{5})-2 C(p_{i}) C(p_{f})a^{2} k_{\mu} \slashed k ( g_{V}-g_{A} \gamma_{5}) \tan^{2}(\eta/2) ,\\
&\chi_{1 \mu} = C(p_{f}) \slashed a_{1} \slashed k \gamma_{\mu}( g_{V}-g_{A} \gamma_{5})+C(p_{i}) \gamma_{\mu}( g_{V}-g_{A} \gamma_{5}) \slashed k \slashed a_{1},\\
&\chi_{2 \mu} = \big[C(p_{f}) \slashed a_{2} \slashed k \gamma_{\mu}( g_{V}-g_{A} \gamma_{5})+C(p_{i}) \gamma_{\mu}( g_{V}-g_{A} \gamma_{5}) \slashed k \slashed a_{2}\big]\tan(\eta/2) ,\\
&\chi_{3 \mu}= -2 C(p_{i}) C(p_{f})a^{2} k_{\mu} \slashed k ( g_{V}-g_{A} \gamma_{5}) \big(1- \tan^{2}(\eta/2) \big),
\end{split}
\end{equation}
where $ C(p_{i,f}) = e/[2(k. p_{i,f})] $. Now, we use a transformation, known as the ordinary and generalized Jacobi-Anger identity, involving ordinary $J_{s}(\xi_{3})$ and generalized $ B_{s}(\xi_{1},\xi_{2})$ Bessel functions \cite{dattoli1992generating}:
\begin{equation}
\begin{split}
e^{i(\xi_{1} \sin(\phi) + \xi_{2} \sin(2\phi))}&=\sum_{s=-\infty}^{+\infty} B_{s}(\xi_{1},\xi_{2})~e^{i s \phi} ,\\
e^{i \xi_{3} \cos(\phi)}&=\sum_{s=-\infty}^{+\infty} i^{s} j_{s}(\xi_{3})~e^{i s \phi},\\
\end{split}
\end{equation}
where $s$, the order of Bessel functions, is commonly interpreted as the number of photons exchanged between the two particles involved in our scattering process and the laser field. Explicitly, we apply the following transformation
\begin{equation}
\begin{split}
\begin{Bmatrix}
1\\
\cos(\phi)\\
\sin(\phi) \\
\cos^{2}(\phi)
\end{Bmatrix}
e^{i(\xi_{1} \sin(\phi) + \xi_{2} \sin(2\phi))}~e^{i \xi_{3} \cos(\phi)} =\sum_{s=-\infty}^{+\infty} \begin{Bmatrix}
B_{s}(\xi_{3})\\
B_{1s}(\xi_{3})\\
B_{2s}(\xi_{3})\\
B_{3s}(\xi_{3})
\end{Bmatrix}
B_{s}(\xi_{1},\xi_{2})~e^{i 2s \phi},
\end{split}
\end{equation}
where the coefficients $B_{s}(\xi_{3})$, $B_{1s}(\xi_{3})$, $B_{2s}(\xi_{3})$, $B_{3s}(\xi_{3})$ and $ B_{s}(\xi_{1},\xi_{2}) $ are expressed in terms of ordinary Bessel functions as follows \cite{lotstedt2009recursive,dattoli1992generating}:
\begin{equation}
\begin{split}
B_{s}(\xi_{3}) &= i^{s} j_{s}(\xi_{3}),\\
B_{1s}(\xi_{3}) &=\big( i^{s-1} j_{s-1}(\xi_{3})+ i^{s+1} j_{s+1}(\xi_{3})\big)/2 ,\\
B_{2s}(\xi_{3}) &=\big( i^{s-1} j_{s-1}(\xi_{3})- i^{s+1} j_{s+1}(\xi_{3})\big)/2 i ,\\
B_{3s}(\xi_{3})&=\big(2 i^{s} j_{s}(\xi_{3})+ i^{s-2} j_{s-2}(\xi_{3})+ i^{s+2} j_{s+2}(\xi_{3})\big)/4 ,\\
B_{s}(\xi_{1},\xi_{2})&= \sum_{\lambda=-\infty}^{+\infty} j_{s-2\lambda}(\xi_{1}) j_{\lambda}(\xi_{2}) .
\end{split}
\end{equation}
After integration over space-time and $d^{4}q$, and after some algebraic manipulations, we can decompose the transition matrix element into a series of terms in the form of ordinary and generalized Bessel functions
\begin{equation} \label{eq.17}
\begin{split}
S_{fi}(e^{-} \nu_{\mu} \rightarrow e^{-} \nu_{\mu})= \sum_{s=-\infty}^{+\infty} \dfrac{G_{F} M_{Z}^{2}}{\sqrt{2}(q^{2}-M_{Z}^{2}+i M_{Z} \Gamma_{Z})} \dfrac{(2\pi)^{4}~\delta^{4}(q_{f}-q_{i}+k_{f}-k_{i}+2 s k) }{\sqrt{16k^{0}_{i}k^{0}_{f}Q_{i}Q_{f} V^{4}}}
~M^{(s)}_{f i},
\end{split}
\end{equation}
where $q = q_{i}-q_{f}-2 s k$ is the relativistic four-momentum transfer in the presence of the EM field. The quantity $ M^{(s)}_{f i} $ in Eq.~(\ref{eq.17}) is defined by
\begin{equation}
\begin{split}
M^{(s)}_{f i}= \big[\bar{u}_{\nu_{\mu}}(k_{f},t_{i})\gamma^{\mu}(1-\gamma_{5})u_{\nu_{\mu}}(k_{i},t_{f}) \big] \big[\bar{u}_{e}(p_{f},s_{f}) \Gamma_{\mu}^{(s)} u_{e}(p_{i},s_{i}) \big],
\end{split}
\end{equation}
where
\begin{equation}
\begin{split}
\Gamma_{\mu}^{(s)}=\big[\chi_{0 \mu} B_{s}(\xi_{3})+ \chi_{1 \mu} B_{1s}(\xi_{3})+ \chi_{2 \mu} B_{2s}(\xi_{3}) +\chi_{3 \mu} B_{3s}(\xi_{3}) \big]B_{s}(\xi_{1},\xi_{2}).
\end{split}
\end{equation}
The DCS can be obtained by summing over the final spin states and averaging over the initial ones. Note that electrons can be in two spin states, while neutrinos are only in one negative helicity state \cite{greiner1996gauge}. Therefore, we divide the square of the matrix element $ |S_{fi}|^{2}$ by the incident particle flux $ |J_{inc}| $ and the observation time interval $ T $, and multiply it by the density of the final states. The unpolarized DCS can then be written as follows:
\begin{equation}
d\bar{\sigma}(e^{-} \nu_{\mu} \rightarrow e^{-} \nu_{\mu})=~ \int V \dfrac{d^{3}q_{f}}{(2\pi)^{3}}~ \int V \dfrac{d^{3}k_{f}}{(2\pi)^{3}}~\dfrac{1}{2} \sum_{t_{i,f},s_{i,f}}\dfrac{|S_{fi}(e^{-} \nu_{\mu} \rightarrow e^{-} \nu_{\mu})|^{2}}{ T |J_{inc}|},
\end{equation}
where $|J_{inc}|=(k_{i}.q_{i})/(k^{0}_{i} Q_{i} V )$ denotes the current of the incoming particles in the laboratory system. Applying the following relations: $d^{3}q_{f}=|\textbf{q}_{f}|^{2}d|\textbf{q}_{f}|d\Omega $ and $\delta^{4}(q_{f}-q_{i}+k_{f}-k_{i}+ 2 s k)= \delta^{0}(Q_{f}+k^{0}_{f}-Q_{i}-k^{0}_{i} + 2 s \omega)\delta^{3}(\textbf{q}_{f}+\textbf{k}_{f}-\textbf{q}_{i}-\textbf{k}_{i}+2 s \textbf{k})$ and taking simplifications, we obtain
\begin{equation}
\begin{split}
\dfrac{d\bar{\sigma}}{d\Omega}(e^{-} \nu_{\mu} \rightarrow e^{-} \nu_{\mu}) = & \sum_{s=-\infty}^{+\infty} \dfrac{\textit{G}_{F}^{2}~M_{Z}^{4}}{64(2\pi)^{2}((q^{2}-M_{Z}^{2})^{2}+ M_{Z}^{2} \Gamma_{Z}^{2})(k_{i}.q_{i})} \int\dfrac{|\textbf{q}_{f}|^{2}d|\textbf{q}_{f}|}{k^{0}_{f}Q_{f}} \delta^{0}(Q_{f}+k^{0}_{f}-Q_{i}-k^{0}_{i} + 2 s \omega)\\
& \times \sum_{s'_{i,f},s_{i,f}}|M_{fi}^{(s)}|^{2}\bigg|_{\textbf{q}_{f}+\textbf{k}_{f}-\textbf{q}_{i}-\textbf{k}_{i}+2 s \textbf{k}=0}.
\end{split}
\end{equation}
Using the following relation \cite{greiner2009salam}
\begin{equation}
\begin{split}
\int dx f(x) \delta(g(x))=\dfrac{f(x)}{|g'(x)|}\Big|_{g(x)=0},
\end{split}
\end{equation}
we can perform the remaining integral over $d|\textbf{q}_{f}|$. Therefore, the summed differential cross section (SDCS) can be decomposed into a series of discrete individual differential cross section (IDCS) for different numbers of photons exchanged. This yields
\begin{equation}\label{eq.23}
\begin{split}
\Big(\dfrac{ d \bar \sigma}{d\Omega}\Big)^{\text{with laser}}(e^{-} \nu_{\mu} \rightarrow e^{-} \nu_{\mu})= \sum_{s=-\infty}^{+\infty} \dfrac{ d\sigma^{(s)}}{d\Omega},
\end{split}
\end{equation}
where the IDCS can be written as
\begin{equation} \label{eq.24}
\begin{split}
\dfrac{ d\sigma^{(s)}}{d\Omega } = \dfrac{\textit{G}_{F}^{2}~M_{Z}^{4}}{64(2\pi)^{2}((q^{2}-M_{Z}^{2})^{2}+ M_{Z}^{2} \Gamma_{Z}^{2})(k_{i}.q_{i}) k^{0}_{f} Q_{f}} \dfrac{|\textbf{q}_{f}|^{2} }{\Big|g'(|\textbf{q}_{f}|)\Big|} \times \sum_{s_{i},s_{f}} \sum_{t_{i},t_{f}} |M_{f i}^{(s)}|^{2}\bigg|_{\textbf{q}_{f}+\textbf{k}_{f}-\textbf{q}_{i}-\textbf{k}_{i}+2 s \textbf{k}=0},
\end{split}
\end{equation}
where
\begin{equation}
\begin{split}
g'(|\textbf{q}_{f}|)= \dfrac{|\textbf{q}_{f}|}{Q_{f}} + \dfrac{|\textbf{q}_{f}|+ k^{0}_{i} \cos(\theta_{f})-|\textbf{q}_{i}| F(\theta_{i},\theta_{f},\varphi_{i},\varphi_{f}) +2 s \omega \cos(\theta_{f}) }{k^{0}_{f}},
\end{split}
\end{equation}
with
\begin{equation}
\begin{split}
F(\theta_{i},\theta_{f},\varphi_{i},\varphi_{f})= \sin(\theta_{i}) \cos(\varphi_{i}) \sin(\theta_{f}) \cos(\varphi_{f})+ \sin(\theta_{i}) \sin(\varphi_{i}) \sin(\theta_{f}) \sin(\varphi_{f})+ \cos(\theta_{i})\cos(\theta_{f}).
\end{split}
\end{equation}
The sum over spin can be converted to trace calculation as follows:
\begin{equation} \label{eq.27}
\begin{split}
\sum_{s_{i},s_{f}} \sum_{t_{i},t_{f}} |M_{f i}^{(s)}|^{2} = \text{Tr}\big[\slashed k_{f} \gamma^{\mu}(1-\gamma_{5})\slashed k_{i} \gamma^{\nu}(1-\gamma_{5}) \big]\text{Tr}\big[(\slashed p_{f}+ m_{e}) \Gamma_{\mu}^{(s)} (\slashed p_{i} + m_{e}) \bar \Gamma_{\nu}^{(s)} \big],
\end{split}
\end{equation}
where
\begin{equation}
\begin{split}
& \bar \Gamma_{\nu}^{(s)} = \gamma^{0} \Gamma_{\nu}^{(s) \dagger} \gamma^{0}, \\
& \bar \Gamma_{\nu}^{(s)}=\big[\bar \chi_{0 \nu} B^{*}_{s}(\xi_{3})+\bar \chi_{1 \nu} B^{*}_{1s}(\xi_{3})+ \bar \chi_{2 \nu} B^{*}_{2s}(\xi_{3}) +\bar \chi_{3 \nu} B^{*}_{3s}(\xi_{3}) \big]~B^{*}_{s}(\xi_{1},\xi_{2}),
\end{split}
\end{equation}
and
\begin{equation}
\begin{split}
&\bar \chi_{0 \nu} = \gamma_{\nu}( g_{V}-g_{A} \gamma_{5})-2 C(p_{i}) C(p_{f})a^{2} k_{\nu} \slashed k ( g_{V}-g_{A} \gamma_{5}) \tan^{2}(\eta/2) ,\\
&\bar \chi_{1 \nu} = C(p_{f}) \gamma_{\nu}( g_{V}-g_{A} \gamma_{5}) \slashed k \slashed a_{1} +C(p_{i}) \slashed a_{1} \slashed k \gamma_{\nu}( g_{V}-g_{A} \gamma_{5}),\\
&\bar \chi_{2 \nu} =\big[ C(p_{f}) \gamma_{\nu}( g_{V}-g_{A} \gamma_{5}) \slashed k \slashed a_{2}+ C(p_{i}) \slashed a_{2} \slashed k \gamma_{\nu}( g_{V}-g_{A} \gamma_{5})\big]\tan(\eta/2) ,\\
&\bar \chi_{3 \nu}= -2 C(p_{i}) C(p_{f})a^{2} k_{\nu} \slashed k ( g_{V}-g_{A} \gamma_{5}) \big(1- \tan^{2}(\eta/2) \big).
\end{split}
\end{equation}
The FeynCalc program \cite{feyncalc} is used to compute the traces in Eq.(\ref{eq.27}). The result obtained is written in terms of the coefficients $\Lambda_{1}$ to $\Lambda_{16}$ as follows:
\begin{equation}\label{trace-result}
\begin{split}
\sum_{s_{i},s_{f}} \sum_{t_{i},t_{f}} |M_{f i}^{(s)}|^{2}=& \big[ \Lambda_{1} |B_{s}(\xi_{3})|^{2} + \Lambda_{2} |B_{1s}(\xi_{3})|^{2} +\Lambda_{3} |B_{2s}(\xi_{3})|^{2} + \Lambda_{4}~ |B_{3s}(\xi_{3})|^{2} + \Lambda_{5} B_{s}(\xi_{3}) B^{*}_{1s}(\xi_{3})\\ & + \Lambda_{6}B_{1s}(\xi_{3}) B^{*}_{s}(\xi_{3})+ \Lambda_{7} B_{s}(\xi_{3}) B^{*}_{2s}(\xi_{3}) + \Lambda_{8}~ B_{2s}(\xi_{3}) B^{*}_{s}(\xi_{3}) + \Lambda_{9} B_{s}(\xi_{3}) B^{*}_{3s}(\xi_{3})
\\ & + \Lambda_{10}~ B_{3s}(\xi_{3}) B^{*}_{s}(\xi_{3}) + \Lambda_{11}~ B_{1s}(\xi_{3}) B^{*}_{2s}(\xi_{3}) + \Lambda_{12}~ B_{2s}(\xi_{3}) B^{*}_{1s}(\xi_{3}) + \Lambda_{13} B_{1s}(\xi_{3}) B^{*}_{3s}(\xi_{3})
\\ & + \Lambda_{14}~ B_{3s}(\xi_{3}) B^{*}_{1s}(\xi_{3}) + \Lambda_{15}~ B_{2s}(\xi_{3}) B^{*}_{3s}(\xi_{3}) + \Lambda_{16} B_{3s}(\xi_{3}) B^{*}_{2s}(\xi_{3}) \big] |B_{s}(\xi_{1},\xi_{2})|^{2} .
\end{split}
\end{equation}
We give the explicit expression of first four coefficients in the Appendix.
\section{Numericla results and discussion} \label{NUMERICAL_RESULTS_AND_DISCUSSION}
In this section, we will present the numerical results obtained and discuss their physical interpretation. We should focus on experimentally measurable quantities, in particular the behavior of the DCS assisted by an external elliptically polarized EM field. We set the incident electron momentum $\textbf{p}_{i}$ and outgoing $\textbf{p}_{f}$ in a general geometry with spherical coordinates $\theta_{i}$, $\theta_{f}$, $\varphi_{i}$ and $\varphi_{f}$. Except for Figs.~\ref{fig.DCS}, \ref{fig.DCS-4} and \ref{fig.IDCS-comparaison}, we have chosen $\theta_{i}=\varphi_{i}=\varphi_{f}=1^{\circ}$ in all the results obtained. For the incident muon neutrino, the momentum $\textbf{k}_{i}$ remains in the direction opposite to the $z$ axis. Except for Fig.~\ref{fig.DCS}, we fix the kinetic energy of the incident electron at $E_{e}^{\text{kin}}=10^{-3}$ GeV, and the initial energy of the muon neutrino is chosen as the mass of an electron at rest, $E_{\nu}^{\text{kin}}=0.5\times10^{-3}$ GeV. The direction of the field wave vector $ \textbf{k}$ is along the $z$-axis, while the polarization vectors $\textbf{a}_{1}$ and $\textbf{a}_{2}$ perpendicular to $ \textbf{k}$ are along the $x$ and $y$-axes, respectively, as illustrated in Fig.~\ref{fig.geometric}. Every $d\sigma^{(s)}/ d\Omega$ given by Eq.~(\ref{eq.24}), considering the four-momentum conservation, can be interpreted as the IDCS that describes the scattering process for each number of photons $s$ ($ s > 0 $ for absorption and $ s < 0 $ for emission). Summing over a number of exchanged photons $s$, we obtain the SDCS given by Eq.~(\ref{eq.23}).
\begin{figure}[hbtp]
\centering
\includegraphics[scale=0.3]{fig2.eps}
\caption{Geometry of the coordinate system for the scattering process. The incident electron $e^{-}(p_{i})$ moves in a spherical coordinate geometry ($\theta_{i}$, $\varphi_{i}$), while the muon neutrino $\nu_{\mu}(k_{i})$ remains in the direction opposite to the $z$-axis. The outgoing muon neutrino and final electron move in space ($xyz$).}\label{fig.geometric}
\end{figure}
We will show how the DCS, in the presence and absence of the laser field, varies as a function of the final scattering angle $\theta_{f}$, the kinetic energy of the incident electron $E^{kin}_{e}$ and various incident angles of the electron $\theta_{i}$. Then, we make comparisons of the IDCS and SDCS with other research papers. Afterward, we show how the IDCS depends on the number of photons exchanged $s$, on the geometry, on the parameters characterizing the EM field ($\mathcal{E}_{0}$, $\omega$) and on the kinetic energy of the incident electron. Before finishing, we illustrate the variation of the SDCS as a function of the final scattering angle $\theta_{f}$ and the kinetic energy of the incident electron $E^{kin}_{e}$ at different polarizations, frequencies, and electric field strengths $ \mathcal{E}_{0}$. Finally, we show how it evolves as a function of the electric field strength at different polarizations.
We start our discussion with something we are used to do in such processes occurring in an external EM field. That is we make sure that the DCS in the presence of the laser field with different polarizations is exactly equal to that in the absence of the laser field when the laser parameters tend to zero. In Fig.~\ref{fig.DCS}, we illustrate the comparison between the laser-assisted DCS of the electron-muon neutrino scattering in the framework of electroweak theory, given by Eq.(\ref{eq.24}), and the corresponding one in the absence of the laser field for a geometry $\theta_{i}=1^{\circ}$, $\varphi_{i}=\varphi_{f}=0^{\circ}$ in Fig.~\ref{fig.DCS}(a), and $\theta_{i}=\varphi_{i}=\varphi_{f}=1^{\circ}$, $\theta_{f}=2^{\circ}$ in Fig.~\ref{fig.DCS}(b).
\begin{figure}[hbtp]
\centering
\includegraphics[scale=0.45]{fig3a.eps}
\includegraphics[scale=0.45]{fig3b.eps}
\caption{The two DCSs with and without laser, illustrated as a function of the final scattering angle $ \theta_{f}$ and the kinetic energy of the incident electron for an electric field strength $ \mathcal{E}_{0}=0~\text{V/cm}$ and without any exchange of photons $(s=0)$. The remaining parameters are chosen as follows: (a) $E_{\nu}^{\text{kin}}=0.5\times10^{-3}$ GeV, $E_{e}^{\text{kin}}=10^{-3}$ GeV, $\theta_{i}=1^{\circ},~\varphi_{i}=\varphi_{f}=0^{\circ} $; (b) $E_{\nu}^{\text{kin}}=10$ GeV, $\theta_{i}=\varphi_{i}=\varphi_{f}=1^{\circ}$, $\theta_{f}=2^{\circ}$.}\label{fig.DCS}
\end{figure}
The comparison allows us to verify our results by taking the limit of the electric field strength $ \mathcal{E}_{0}=0~\text{V/cm}$ and of the number of photons exchanged $(s=0)$, where they all tend to zero. Regarding the laser field strength and frequency, these are parameters that characterize the external EM field, while the number of photons exchanged $s$ appeared due to the introduction of ordinary and generalized Bessel functions in our theoretical calculation. We note that the four graphs shown in Figs.~\ref{fig.DCS}(a) and \ref{fig.DCS}(b) are so identical as they are indistinguishable for all final scattering angles $\theta_{f}$ and kinetic energies $E^{kin}_{e}$ of the incident electron. This proves the consistency and validity of our theoretical calculations. In the next step, we will try to see the effect of the geometry in Fig.~(\ref{fig.DCS-4}), which displays the variations of the DCS in the absence of the laser field as a function of the final scattering angle $\theta_{f}$. We observe that the graphs of the DCS are distinct, which clearly shows that the geometry influences the angular distribution of the DCS. We also observe that as the incidence angle $\theta_{i}$ increases, the pic of the DCS increases with a shift towards large final scattering angles.
\begin{figure}[hbtp]
\centering
\includegraphics[scale=0.55]{fig4.eps}
\caption{The variations of the IDCS as a function of scattering angle $\theta_{f}$ for differents angles of the incident electron. The laser field strength and the number of photons are, respectively $\mathcal{E}_{0}=0$ V/cm and $s=0$. The kinetic energy of the incident muon neutrino and the incoming electron are, respectively, $E_{\nu}^{\text{kin}}=0.5\times10^{-3}$ GeV and $E_{e}^{\text{kin}}=10^{-3}$ GeV.}\label{fig.DCS-4}
\end{figure}
Another important remark is that the electron has a high probability to be scattered under a final angle $\theta_{f}$ with a value approximately approaching the incidence angle $(\theta_{f}\approx \theta_i)$. For example, if the incidence angle is $\theta_{i}=15^{\circ}$, the peak is located around the final scattering angle $\theta_{f}\approx 15^{\circ}$. In the same context, and to highlight the correctness and accuracy of our calculations in the presence of an EM field with elliptical polarization, this calculation which is general allows us to find the results previously obtained in two research papers that deal with the same scattering process in the presence of a laser field with circular \cite{asri2021elastic} or linear \cite{bai2012multiphoton} polarization. We begin with Fig.~\ref{fig.IDCS-comparaison}(a) which presents the variations of the IDCS as a function of the number of photons exchanged $s$ with a degree of ellipticity $\eta=90$ (circular polarization) for two different field strengths $\mathcal{E}_{0}$ and for a frequency $\hbar\omega=1.17~\text{eV}$. This is the same envelope obtained in a previous paper by El Asri \textit{et al.} (see Fig.~2(a) in \cite{asri2021elastic}). Thus, the theoretical formalism adopted here is general and can lead to all the results obtained in \cite{asri2021elastic}. For linear polarization, we display, in Fig.~\ref{fig.IDCS-comparaison}(b), the IDCS as a function of the number of photons exchanged $s$ with a degree of ellipticity $\eta=0$ for the laser field strength $\mathcal{E}_{0}=5.18\times10^{7}~\text{V/cm}$ and frequency $\hbar\omega=1.17~\text{eV}$.
\begin{figure}[hbtp]
\centering
\includegraphics[scale=0.5]{fig5a.eps}
\includegraphics[scale=0.55]{fig5b.eps}
\caption{The IDCS as a function of the number of photons exchanged. The free parameters are (\textbf{a}) $E_{\nu}^{\text{kin}}=0.5\times10^{-3}$ GeV, $E_{e}^{\text{kin}}=10^{-3}$ GeV, $\hbar\omega=1.17~\text{eV}$, $\theta_{i}=1^{\circ}$, $\theta_{f}=\varphi_{i}=\varphi_{f}=0^{\circ}$ (to be compared with \cite{asri2021elastic}); and (\textbf{b}) $E_{\nu}^{\text{kin}}=0.5\times10^{-3}$ GeV, $E_{e}^{\text{kin}}=10^{-3}$ GeV, $\mathcal{E}_{0}=5.18\times10^{7}~\text{V/cm}$, $\hbar\omega=1.17~\text{eV}$, $\theta_{i}=45^{\circ}$, $\varphi_{i}=\varphi_{f}=0^{\circ}$, $\theta_{f}=1^{\circ}$ (to be compared with \cite{bai2012multiphoton}).}\label{fig.IDCS-comparaison}
\end{figure}
We observe, that a large number of photons are exchanged between the laser field and the scattering process, and the cutoff number is about $s=\pm 1400$. Comparing, this figure with the one obtained by Bai \textit{et al.} (see Fig.~2(a) in \cite{bai2012multiphoton}), we get the same result.
Another comparison concerns Fig.~\ref{fig.SDCS-comparaison-6}, which displays the dependence of the SDCS given in Eq.~(\ref{eq.23}) on the final scattering angle $\theta_{f}$ for different polarizations of the EM field. In Fig.~\ref{fig.SDCS-comparaison-6}(a), we show the variation of the SDCS as a function of the final scattering angle $\theta_{f}$ for the degree of ellipticity $\eta=90$.
\begin{figure}[hbtp]
\centering
\includegraphics[scale=0.45]{fig6a.eps}
\includegraphics[scale=0.45]{fig6b.eps}
\caption{Dependence of SDCS on the scattering angle $\theta_{f}$ for various polarizations of the EM field, taking the laser parameters such as: $\mathcal{E}_{0}=10^{8}~\text{V/cm}$, $\hbar\omega=1.17 \text{eV}$. The parameters of the geometry are taken as $\theta_{i}=\varphi_{i}=\varphi_{f}=1^{\circ}$ and the remaining parameters are selected as shown in Fig.~\ref{fig.IDCS-comparaison}.}\label{fig.SDCS-comparaison-6}
\end{figure}
We obtain a symmetrical graph, which presents a peak in the vicinity of $\theta_{f}=0$. Furthermore, the SDCS remains lower than the DCS without a laser.
In Fig.~\ref{fig.SDCS-comparaison-6}(b), we represent the same graphs, but with a degree of ellipticity $\eta=0$. In this case, we see an enhancement of the SDCS compared to the DCS without laser, which is consistent with previous research in the case of linear polarization \cite{du2018new,du2018Nonlinear,schnez2007laser}.
After a detailed discussion and a comparison of the results obtained with the previous work, let us see what happens if we introduce an external EM field with elliptical polarization having a degree of ellipticity $\eta=120^{\circ} $.
\begin{figure}[hbtp]
\centering
\includegraphics[scale=0.45]{fig7a.eps}
\includegraphics[scale=0.45]{fig7b.eps} \par
\includegraphics[scale=0.45]{fig7c.eps}
\includegraphics[scale=0.5]{fig7d.eps}
\caption{The behavior of the IDCS, as a function of photon number $s$. The various parameters are (a) $\theta_{i}=1^{\circ}$, $\theta_{f}=2^{\circ}$ and $\hbar\omega=1.17$ eV, (b) $\mathcal{E}_{0}=5.18\times10^{8}$ V/cm, $\theta_{i}=1^{\circ}$ and $\theta_{f}=2^{\circ}$, (c) $\mathcal{E}_{0}=5.18\times10^{8}$ V/cm, $\hbar\omega=1.17$ eV, $\theta_{i}=1^{\circ}$ and $\theta_{f}=2^{\circ}$ and (d) $\mathcal{E}_{0}=5.18\times10^{8}$ V/cm and $\hbar\omega=1.17$ eV.}\label{fig.IDCS-elliptic}
\end{figure}
In Fig.~\ref{fig.IDCS-elliptic}, we display the variations of the IDCS (the multi-photon energy transfer phenomenon) as a function of the number of photons exchanged $s$, at different field strengths and frequencies, and at different kinetic energies and geometry of the incident electron. From Fig.~\ref{fig.IDCS-elliptic}(a), we can observe that the electron exchanges a large number of photons with the high-intensity EM field ($\mathcal{E}_{0}= 10^{9}$ V/cm), where the cutoff number is approximately $s = \pm 125$, compared to the low-intensity EM field ($\mathcal{E}_{0}= 5.18 \times 10^{8}$ V/cm), where the cutoff number is approximately equal to $s= \pm 65$. This implies that the influence of the laser field on the scattering process is more significant at higher field strengths, i.e. the electron interacts powerfully with the strong EM field. Additionally, the order of magnitude of the IDCS decreases with the increase of the field strength. In Fig.~\ref{fig.IDCS-elliptic}(b), we visualize the variations of the IDCS as a function of the number of photons exchanged $s$ at different frequencies of the laser field. We see that at low frequencies ($ \hbar \omega= 1.17$ eV), the number of exchanged photons is large and the cutoff number is about $s = \pm 65 $, compared to high frequencies ($\hbar \omega= 2$ eV) where the cutoff number is about $s = \pm 25 $. Moreover, the order of magnitude of the IDCS increases with the increase of the frequency. Consequently, the multi-photon energy transfer phenomenon between the laser and the scattering process is related to the properties of the applied laser field. Figure~\ref{fig.IDCS-elliptic}(c) describes the variations of the IDCS as a function of the number of photons exchanged $s$ at different kinetic energies of the incident electron. The multi-photon energy transfer phenomenon increases as the kinetic energy of the incident electron increases, leading to an exchange of a large number of photons, while the order of magnitude of the IDCS decreases. In Fig.~\ref{fig.IDCS-elliptic}(d), we illustrate the influence of the chosen geometry on the multi-photon process. We observe that the number of photons exchanged in the geometry ($\theta_{i}=1^{\circ},~\theta_{f}=2^{\circ}$ ) is more important than the one exchanged in the geometry ($\theta_{i}=15^{\circ},~\theta_{f}=20^{\circ}$). While the first geometry means that the incident electron is almost in the same direction as the $\textbf{k}$ field vector, the second one indicates that the electron arrives at an angle of $\theta_{i}=15^{\circ}$ with respect to the $z$-axis. Physically, this states that the electron interacts with the laser if they arrive together in the same direction more than if they are in two different directions. We also see that the oscillations of these envelopes fall sharply to the sides, which can also be best explained by the well-known behavior and properties of ordinary and generalized Bessel functions (GBF)\cite{korsch2006two}.
Let us now see what happens if we sum over all the possible number of photons exchanged. In Fig.~\ref{fig.9-SDCS}, we have plotted the variations of the SDCS as a function of the final scattering angle $\theta_{f}$, for different known frequencies and electric field strengths.
\begin{figure}[hbtp]
\centering
\includegraphics[scale=0.45]{fig8a.eps}
\includegraphics[scale=0.45]{fig8b.eps}
\caption{Variation of the laser-assisted SDCS (with the degree of ellipticity $\eta=120^{\circ}$) as a function of the scattering angle $\theta_{f}$ for various electric field strengths $\mathcal{E}_{0}$ and laser frequencies $\hbar\omega$. The kinetic energy of the incident muon neutrino and the incoming electron are, respectively, $E_{\nu}^{\text{kin}}=0.5\times10^{-3}$ GeV et $E_{e}^{\text{kin}}=10^{-3}$ GeV. The other parameters are taken as follows: (a) $\mathcal{E}_{0}=10^{8}~\text{V/cm}$, $\theta_{i}=\varphi_{i}=\varphi_{f}=1^{\circ}$, (b) $\hbar\omega=1.17~\text{eV}$, $\theta_{i}=\varphi_{i}=\varphi_{f}=1^{\circ}$.}\label{fig.9-SDCS}
\end{figure}
Figure~\ref{fig.9-SDCS}(a) shows the dependence of the SDCS assisted by elliptically polarized laser field with a degree of ellipticity $\eta=120^{\circ}$ as a function of the final scattering angle $\theta_{f}$, for different known laser frequencies $\hbar\omega$ which are the CO$_{2}$ laser ($\hbar\omega=0.117$ eV), the Nd:YAG laser ($\hbar\omega=1.17$ eV) and the He-Ne laser ($\hbar\omega=2$ eV). For $ 0^{\circ} \leqslant \theta_{f}\leqslant 180^{\circ}$, the order of magnitude of the SDCS increases with increasing laser field frequency, which confirms the result of Fig.~\ref{fig.IDCS-elliptic}(b). Since the laser field strength is also a crucial parameter, we performed the same analysis for different values of the laser field strength in Fig.~\ref{fig.9-SDCS}(b). We observe that the order of magnitude of SDCS decreases with increasing laser field strength, which is consistent with the result obtained in Fig.~\ref{fig.IDCS-elliptic}(a).
Let us now turn to the discussion of the laser field polarization effect on the SDCS. Figure~\ref{fig.SDCS} shows the variations of the SDCS as a function of the final scattering angle $\theta_{f}$ and the kinetic energy of the incident electron $E_{e}^{\text{kin}}$ at different polarizations of the strong EM field. The parameters of the laser field are chosen as follows: $\hbar \omega=1.17~\text{eV}$ and $\mathcal{E}_{0}=10^{10}~\text{V/cm}$. Our results show that the effect of the laser field polarization is clearly highlighted, since the three SDCSs with laser and without laser are now obviously distinguishable.
\begin{figure}[hbtp]
\centering
\includegraphics[scale=0.45]{fig9a.eps}
\includegraphics[scale=0.45]{fig9b.eps}
\caption{Variations of the SDCS as a function of the scattering angle $\theta_{f}$ and the kinetic energy of the incident electron for different polarizations of the strong EM field. The laser field parameters are selected as follows: $\hbar\omega=1.17~\text{eV}$ and $\mathcal{E}_{0}=10^{10}~\text{V/cm}$. The other parameters are taken as: (a) $E_{\nu}^{\text{kin}}=0.5\times10^{-3}~\text{GeV}$, $E_{e}^{\text{kin}}=10^{-3}~\text{GeV}$, $\theta_{i}=\varphi_{i}=\varphi_{f}=1^{\circ}$; (b) $E_{\nu}^{\text{kin}}=10~\text{GeV}$, $\theta_{i}=\varphi_{i}=\varphi_{f}=1^{\circ}$, $\theta_{f}=2^{\circ}$.}\label{fig.SDCS}
\end{figure}
We also see that the SDCS in the case of linear polarization ($\eta=0^{\circ}$) is enhanced compared to that without laser. Furthermore, we observe that there is a reduction of the SDCS in the case of an elliptical ($\eta=120^{\circ}$) and circular ($\eta=90^{\circ}$) polarizations. We note here that the SDCS for elliptical polarization is lower or higher than the SDCS for circular polarization depending on the value of the degree of ellipticity $\eta$. The same thing can be said about the result presented in Fig.~\ref{fig.SDCS}(b) as a function of the kinetic energy of the incident electron $E_{e}^{\text{kin}}$.\\
Finally, to understand most clearly the effect of polarization, we display the dependence of the SDCS on the laser field strength at different polarizations in Fig.~\ref{fig.SDCS-intensitie}. The laser field strength appears in the equations for determining the behavior of the SDCS through the arguments $\xi_{1}$, $\xi_{2}$ and $\xi_{3}$ (see Eq.~(\ref{eq.11})) of the ordinary and generalized Bessel functions and the coefficients $\chi_{0 \mu}$, $\chi_{1 \mu}$, $\chi_{2 \mu}$ and $\chi_{3 \mu}$ given by Eq.~(\ref{eq.13}).
\begin{figure}[hbtp]
\centering
\includegraphics[scale=0.53]{fig10a.eps}
\includegraphics[scale=0.55]{fig10b.eps} \par
\includegraphics[scale=0.55]{fig10c.eps}
\caption{Variations of SDCS as a function of the laser field strength $\mathcal{E}_{0}$ for different polarizations. The kinetic energy of the incident muon neutrino and the incoming electron are respectively $E_{\nu}^{\text{kin}}=0.5\times10^{-3}$ and $E_{e}^{\text{kin}}=10$ GeV. The frequency of the laser field is $\hbar \omega =0.117$ eV. The other parameters are chosen as follows: $\theta_{i}=\varphi_{i}=\varphi_{f}=1^{\circ}$ and $\theta_{f}=2^{\circ}$.}\label{fig.SDCS-intensitie}
\end{figure}
In Fig.~\ref{fig.SDCS-intensitie}(a), we can see that the SDCS for linear polarization ($\eta=0^{\circ}$) remains constant and equal to the DCS without laser in the field strength range between $5.18\times10^{9}$ and $6. 21\times10^{10}$ V/cm. Outside this interval, we find that the SDCS increases progressively with the laser field strengh, which is consistent with the results obtained previously in \cite{bai2012multiphoton, du2018new, wang2019relativistic, Li-Berakdar2003}. Although the ordinary and generalized Bessel functions in the SDCS vary with the laser field strength, the electron state is more perturbed, and thus the cross section is strongly changed. Consequently, the SDCS increases because the energy of the electron increases after the absorption of laser photons. On the other hand, in Figs.~\ref{fig.SDCS-intensitie}(b) and \ref{fig.SDCS-intensitie}(c), which illustrate the variations of the SDCS in the case of circular ($\eta=90^{\circ}$) and elliptical ($\eta=60^{\circ}$) polarizations, it is shown that the SDCS decreases and remains lower than the DCS without laser, which is consistent with the results found in \cite{attaourti2004mott, schnez2007laser}. Therefore, the effect of electron dressing induces very significant changes in the DCS, and a large enhancement is found in the linear polarization case.
\section{Conclusion} \label{CONCLUSION}
Laser-assisted electron-muon neutrino scattering is investigated, for elliptical polarization, in the first Born approximation and in the framework of electroweak theory. We have extended the study of this scattering process for a general polarization that leads to all particular results obtained before in linear and circular polarizations. The numerical results show that the SDCS is significantly modified by the polarization type depending on the degree of ellipticity and the laser field parameters. Moreover, the SDCS in the case of linear polarization is enhanced compared to that without laser, while it is reduced in the case of elliptical and circular polarizations. We have shown that the scattering geometry, as well as the laser field parameters and the kinetic energy of the incident electron influence the multi-photon process. We hope that the present work will serve as a stimulus for experimenters to perform such scattering experiments on electrons dressed in linear and elliptical polarization.
|
\section{\leftline {\bf{Introduction.}}}
\setcounter{equation}{0}
\subsection{Background and motivation}
As is well--known, most of the flows in nature are multi--fluid
flows. Such a terminology includes the flows of non--miscible fluids
such as air and water; gas, oil and water. For the flows of miscible
fluids, they usually form a ``new" single fluid possessing its own
rheological properties. One interesting example is the stable
emulsion between oil and water which is a non--Newtonian fluid, but
oil and water are Newtonian ones.\par
One of the classic examples of multi--fluid flows is
small amplitude waves propagating at the interface between air and
water, which is called a separated flow. In view of modeling, each
fluid obeys its own equation and couples with each other through the
free surface in this case. Here, the motion of the fluid is governed
by the pair of compressible Euler equations with free surface:
\begin{align}
\partial_{t} \rho_{i}+\nabla \cdot\left(\rho_{i} v_{i}\right) &=0, \quad i=1,2,\label{1.1} \\
\partial_{t}\left(\rho_{i} v_{i}\right)+\nabla \cdot\left(\rho_{i} v_{i} \otimes v_{i}\right)+\nabla p_i &=-g\rho_{i}
e_3\pm F_D.\label{1.2}
\end{align}
In above equations, $\rho_1$ and $v_1$ represent the density and
velocity of the upper fluid (air), and $\rho_2$ and $v_2$ denote
the density and velocity of the lower fluid (water). $p_{i}$ denotes
the pressure. $-g\rho_{i} e_3$ is the gravitational force with the
constant $g>0$ the acceleration of gravity and $e_3$ the vertical
unit vector, and $F_D$ is the drag force. As mentioned before, the
two fluids (air and water) are separated by the unknown free surface
$z=\eta(x, y, t)$, which is advected with the fluids according to
the kinematic relation:
\begin{equation}\partial_t\eta=v_{1,z}-v_{1,x}\partial_x \eta-v_{1, y}\partial_y \eta\label{1.3}\end{equation}
on two sides of the surface $z=\eta$ and the pressure is continuous
across this surface.\par When the wave's amplitude becomes large
enough, wave breaking may happen. Then, in the region around the
interface between air and water, small droplets of liquid appear in
the gas, and bubbles of gas also appear in the liquid. These
inclusions might be quite small. Due to the appearances of collapse
and fragmentation, the topologies of the free surface become quite
complicated and a wide range of length scales are involved.
Therefore, we encounter the situation where two--fluid models become
relevant if not inevitable. The classic approach to simplify the
complexity of multi--phase flows and satisfy the engineer's need of
some modeling tools is the well--known volume--averaging method (see
\cite{Ishii1, Prosperetti} for details). Thus, by performing such a
procedure, one can derive a model without surface: a two--fluid
model. More precisely, we denote $\alpha^{\pm}$ by the volume
fraction of the liquid (water) and gas (air), respectively.
Therefore, $\alpha^++\alpha^-=1$. Applying the volume--averaging
procedure to the equations \eqref{1.1} and \eqref{1.2} leads to the
following generic compressible two--fluid model:
\begin{equation}\label{1.4}
\left\{\begin{array}{l}
\partial_{t}\left(\alpha^{\pm} \rho^{\pm}\right)+\operatorname{div}\left(\alpha^{\pm} \rho^{\pm} u^{\pm}\right)=0, \\
\partial_{t}\left(\alpha^{\pm} \rho^{\pm} u^{\pm}\right)+\operatorname{div}\left(\alpha^{\pm} \rho^{\pm} u^{\pm} \otimes u^{\pm}\right)
+\alpha^{\pm} \nabla P^\pm=-g\alpha^{\pm}\rho^{\pm} e_3\pm F_D.
\end{array}\right.
\end{equation}
\par
We have already discussed the case of water waves, where a separated
flow can lead to a two--fluid model from the viewpoint of practical
modeling. As mentioned before, two--fluid flows are very common in
nature, but also in various industry applications such as nuclear
power, chemical processing, oil and gas manufacturing. According to
the context, the models used for simulation may be very different.
However, averaged models share the same structure as \eqref{1.4}. By
introducing viscosity effects and capillary pressure effects, one can generalize the above system
\eqref{1.4} to
\begin{equation}\label{1.5}
\left\{\begin{array}{l}
\partial_{t}\left(\alpha^{\pm} \rho^{\pm}\right)+\operatorname{div}\left(\alpha^{\pm} \rho^{\pm} u^{\pm}\right)=0, \\
\partial_{t}\left(\alpha^{\pm} \rho^{\pm} u^{\pm}\right)+\operatorname{div}\left(\alpha^{\pm} \rho^{\pm} u^{\pm} \otimes u^{\pm}\right)
+\alpha^{\pm} \nabla P^{\pm}\left(\rho^{\pm}\right)=\operatorname{div}\left(\alpha^{\pm} \tau^{\pm}\right), \\
P^{+}\left(\rho^{+}\right)-P^{-}\left(\rho^{-}\right)=f\left(\alpha^{-}
\rho^{-}\right),
\end{array}\right.
\end{equation} where
$\rho^{\pm}(x, t) \geqq 0, u^{\pm}(x, t)$ and
$P^{\pm}\left(\rho^{\pm}\right)=A^{\pm}\left(\rho^{\pm}\right)^{\bar{\gamma}^{\pm}}$
denote the densities, the velocities of each phase, and the two
pressure functions, respectively. $\bar{\gamma}^{\pm} \geqq 1,
A^{\pm}>0$ are positive constants. In what follows, we set
$A^{+}=A^{-}=1$ without loss of any generality. As in \cite{Evje9},
we assume that the capillary pressure $f$ belongs to $C^{3}([0, \infty))$. Moreover,
$\tau^{\pm}$ are the viscous stress tensors
\begin{equation}\label{1.6}
\tau^{\pm}:=\mu^{\pm}\left(\nabla u^{\pm}+\nabla^{t}
u^{\pm}\right)+\lambda^{\pm} \operatorname{div} u^{\pm} \mathrm{Id},
\end{equation}
where the constants $\mu^{\pm}$ and $\lambda^{\pm}$ are shear and
bulk viscosity coefficients satisfying the physical condition:
$\mu^{\pm}>0$ and $2 \mu^{\pm}+3 \lambda^{\pm} \geqq 0,$ which
implies that $\mu^{\pm}+\lambda^{\pm}>0.$ For more information
about this model, we refer to \cite{Brennen1, Bresch1, Bresch2,
Friis1, Ishii1, Prosperetti, Raja} and references therein.
However, it is well--known
that as far as mathematical analysis of two--fluid model is
concerned, there are many technical challenges. Some of them
involve, for example:
\begin{itemize}
\item The two--fluid model is a partially dissipative system.
More precisely, there is no dissipation on the mass conservation
equations, whereas the momentum equations have viscosity
dissipations;
\item The corresponding linear system of the model has
zero eigenvalue, which makes mathematical analysis (well--posedness
and stability) of the model become quite difficult and complicated;
\item Transition to single--phase regions, i.e, regions where the mass
$\alpha^{+} \rho^{+}$ or $\alpha^{-} \rho^{-}$ becomes zero, may
occur when the volume fractions $\alpha^{\pm}$ or the densities
$\rho^{\pm}$ become zero;
\item The system is non--conservative, since the non--conservative terms $\alpha^{\pm} \nabla
P^{\pm}$ are involved in the momentum equations. This brings various
mathematical difficulties for us to employ methods used
for single phase models to the two--fluid model.
\end{itemize}\par
For the case that the capillary pressure is a strictly decreasing
function near the equilibrium, namely, $f'(1)<0$, Evje--Wang--Wen \cite{Evje9} obtained global stability of the constant equilibrium state for the three--dimensional
Cauchy problem of the two--fluid model \eqref{1.5} under the assumption that the initial perturbation is small in $H^2$-norm and bounded in $L^1$-norm. It should be noted that as pointed out by Evje--Wang--Wen in \cite{Evje9}, the assumption $f'(1)<0$
played a crucial role in their analysis and appeared to have an essential stabilization effect on the
model in question.
Bretsch et al. in
the seminal work \cite{Bresch1} considered a model similar to
\eqref{1.5}. More specifically, they made the following assumptions:
\begin{itemize} \item $P^{+}=P^{-}$ (particularly, $f'(1)=0$ in this case);\\
\item inclusion of viscous terms of the form \eqref{1.2} where $\mu^{\pm}$
depends on densities $\rho^{\pm}$ and $\lambda^{\pm}=0$;\\
\item inclusion of a third order derivative of $\alpha^{\pm}
\rho^{\pm}$, which are so--called internal capillary forces
represented by the well--known Korteweg model on each phase.
\end{itemize}
They obtained the global weak solutions in the periodic domain with
$1<\overline{\gamma}^{\pm}< 6$. It is worth
mentioning that the method of \cite{Bresch1} doesn't work for the case without
the internal capillary forces. Later, Bresch--Huang--Li
\cite{Bresch2} established the global existence of weak solutions in
one space dimension without the internal capillary forces when
$\overline{\gamma}^{\pm}>1$ by taking
advantage of the one space dimension. However, the method of \cite{Bresch2} relies
crucially on the advantage of one space dimension, and particularly cannot be applied for high
dimensional problem. Recently, Wu--Yao--Zhang \cite{WYZ} showed
the global stability of the constant equilibrium state in three space dimension by
exploiting the
dissipation structure of the model (with $P^+=P^-$ and without internal capillary forces) and making full use of several key observations.
For the case of
the special density-dependent viscosities with equal viscosity coefficients and the case of general constant
viscosities, Cui--Wang--Yao--Zhu
\cite{c1} and Li--Wang--Wu--Zhang \cite {LWWZ} proved the global stability of the constant equilibrium state
for the three--dimensional
Cauchy problem with the internal capillary forces, respectively.
\par
To sum up, the works \cite{Evje9} and \cite{WYZ} rely essentially on the assumption $f'(1)<0$ and
$P^+=P^-$ (corresponding to $f'(1)=0$). Therefore, a natural and
important problem is that what will happen for the case that the capillary pressure is a strictly increasing
function near the equilibrium, namely, $f'(1)>0$. That is to say,
what about the stability of three--dimenional
Cauchy problem to the two--fluid model \eqref{1.5} with $f'(1)>0$. The main purpose of this work is to give a definite answer to this issue.
More precisely, we first employ Hodge decomposition technique and make detailed analysis of the Green's function for the corresponding linearized system to construct solutions
of the linearized problem that grow exponentially in time in the Sobolev space $H^k$, thus leading to a global
instability result for the linearized problem.
Then, based on the global
linear instability result and a local existence theorem of classical solutions to the original nonlinear system, we can
prove the instability of the nonlinear problem in the sense of Hadamard by making a delicate analysis on the properties of the semigroup.
Therefore, our result shows that for the case $f'(1)>0$, the constant equilibrium state of the two--fluid model \eqref{1.5} is linearly globally unstable and nonlinearly locally unstable
in the sense of Hadamard, which is in contrast to
the cases $f'(1)<0$ (\cite{Evje9}) and $P^+=P^-$ (corresponding to $f'(1)=0$) (\cite{WYZ}) where the constant equilibrium state of the two--fluid model \eqref{1.5} is nonlinearly globally stable.
\subsection{New formulation of system \eqref{1.5} and Main Results}
In this subsection, we devote ourselves to reformulating the system
\eqref{1.5} and stating the main results. To begin with, noting
the relation between the pressures of \eqref{1.5}$_3$, one has
\begin{equation}\label{1.7}
\mathrm{d} P^{+}-\mathrm{d} P^{-}=\mathrm{d} f\left(\alpha^{-}
\rho^{-}\right),
\end{equation}
where $P^{\pm}:=P^{\pm}\left(\rho^{\pm}\right).$ It is clear that
\[
\mathrm{d} P^{+}=s_{+}^{2} \mathrm{d} \rho^{+}, \quad \mathrm{d}
P^{-}=s_{-}^{2} \mathrm{d} \rho^{-}, \quad \text { where }
s_{\pm}^{2}:=\frac{\mathrm{d} P^{\pm}}{\mathrm{d}
\rho^{\pm}}\left(\rho^{\pm}\right)=\bar{\gamma}^{\pm}
\frac{P^{\pm}\left(\rho^{\pm}\right)}{\rho^{\pm}}.
\]
Here $s_{\pm}$ represent the sound speed of each phase respectively.
Motivated by \cite{Bresch1}, we introduce the fraction densities
\begin{equation}\label{1.8}
R^{\pm}=\alpha^{\pm} \rho^{\pm},
\end{equation}
which together with the fact that $\alpha^++\alpha^-=1$ gives
\begin{equation}\label{1.9}
\mathrm{d} \rho^{+}=\frac{1}{\alpha_{+}}\left(\mathrm{d}
R^{+}-\rho^{+} \mathrm{d} \alpha^{+}\right), \quad \mathrm{d}
\rho^{-}=\frac{1}{\alpha_{-}}\left(\mathrm{d} R^{-}+\rho^{-}
\mathrm{d} \alpha^{+}\right).
\end{equation}
By virtue of \eqref{1.7} and \eqref{1.9}, we finally get
\begin{equation}\label{1.10}
\mathrm{d} \alpha^{+}=\frac{\alpha^{-} s_{+}^{2}}{\alpha^{-}
\rho^{+} s_{+}^{2}+\alpha^{+} \rho^{-} s_{-}^{2}} \mathrm{d}
R^{+}-\frac{\alpha^{+} \alpha^{-}}{\alpha^{-} \rho^{+}
s_{+}^{2}+\alpha^{+} \rho^{-}
s_{-}^{2}}\left(\frac{s_{-}^{2}}{\alpha^{-}}+f^{\prime}\right)
\mathrm{d} R^{-}. \end{equation}
Substituting \eqref{1.10} into \eqref{1.9}, we deduce
the following expressions:
\[
\mathrm{d} \rho^{+}=\frac{\rho^{+} \rho^{-}
s_{-}^{2}}{R^{-}\left(\rho^{+}\right)^{2}
s_{+}^{2}+R^{+}\left(\rho^{-}\right)^{2} s_{-}^{2}}\left(\rho^{-}
\mathrm{d} R^{+}+\left(\rho^{+}+\rho^{+} \frac{\alpha^{-}
f^{\prime}}{s_{-}^{2}}\right) \mathrm{d} R^{-}\right),
\]
and
\[
\mathrm{d} \rho^{-}=\frac{\rho^{+} \rho^{-}
s_{+}^{2}}{R^{-}\left(\rho^{+}\right)^{2}
s_{+}^{2}+R^{+}\left(\rho^{-}\right)^{2} s_{-}^{2}}\left(\rho^{-}
\mathrm{d} R^{+}+\left(\rho^{+}-\rho^{-} \frac{\alpha^{+}
f^{\prime}}{s_{+}^{2}}\right) \mathrm{d} R^{-}\right),
\]
which together with \eqref{1.7} gives the pressure differential
$\mathrm{d} P^{\pm}$
\[
\mathrm{d} P^{+}=\mathcal{C}^{2}\left(\rho^{-} \mathrm{d}
R^{+}+\left(\rho^{+}+\rho^{+} \frac{\alpha^{-}
f^{\prime}}{s_{-}^{2}}\right) \mathrm{d} R^{-}\right) ,\] and
\[
\mathrm{d} P^{-}=\mathcal{C}^{2}\left(\rho^{-} \mathrm{d}
R^{+}+\left(\rho^{+}-\rho^{-} \frac{\alpha^{+}
f^{\prime}}{s_{+}^{2}}\right) \mathrm{d} R^{-}\right) ,\] where
\[
\mathcal{C}^{2}:=\frac{s_{-}^{2} s_{+}^{2}}{\alpha^{-} \rho^{+}
s_{+}^{2}+\alpha^{+} \rho^{-} s_{-}^{2}}.\]\par \noindent Next, by noting the
fundamental relation: $\alpha^++\alpha^-=1$, we can get the
following equality:
\begin{equation}\label{1.11}
\frac{R^+}{\rho^+}+\frac{R^-}{\rho^-}=1, ~~\hbox{and thus}~~
\rho^-=\frac{R^-\rho^+}{\rho^+-R^+}.\end{equation} Then, it holds
from the pressure relation $\eqref{1.5}_3$ that
\begin{equation}\label{1.12}
\varphi(\rho^+, R^+,
R^-):=P^+(\rho^+)-P^-{\left(\frac{R^-\rho^+}{\rho^+-R^+}\right)}-f({R^-})=0.
\end{equation}
\noindent Thus, we can employ the implicit function theorem to
define $\rho^{+}$. To see this, by differentiating the above
equation with respect to $\rho^{+}$ for given $R^{+}$ and $R^{-}$,
we get
\[
\frac{\partial\varphi}{\partial\rho^+}(\rho^+, R^+,
R^-)=s_{+}^{2}+s_{-}^{2} \frac{R^{-}
R^{+}}{\left(\rho^{+}-R^{+}\right)^{2}},
\]
which is positive for any $\rho^{+}\in(R^+, +\infty)$ and
$R^{\pm}>0.$ This together with the implicit function theorem
implies that $\rho^{+}=\rho^{+}\left(R^{+}, R^{-}\right)
\in\left(R^{+},+\infty\right)$ is the unique solution of the
equation \eqref{1.12}. By virtue of \eqref{1.8}, \eqref{1.12} and
the
fundamental fact that $\alpha^++\alpha^-=1$, $\rho^{-}$ and $\alpha^{\pm}$ can be defined by
\[
\begin{aligned}
\rho^{-}\left(R^{+}, R^{-}\right) &=\frac{R^{-} \rho^{+}\left(R^{+}, R^{-}\right)}{\rho^{+}\left(R^{+}, R^{-}\right)-R^{+}}, \\
\alpha^{+}\left(R^{+}, R^{-}\right) &=\frac{R^{+}}{\rho^{+}\left(R^{+}, R^{-}\right)}, \\
\alpha^{-}\left(R^{+}, R^{-}\right)
&=1-\frac{R^{+}}{\rho^{+}\left(R^{+},
R^{-}\right)}=\frac{R^{-}}{\rho^{-}\left(R^{+}, R^{-}\right)}.
\end{aligned}
\]
We refer the readers to [\cite{Bresch2}, P. 614] for more details.
\par
Therefore, we can rewrite system \eqref{1.5} into the following
equivalent form:
\begin{equation}\label{1.13}
\left\{\begin{array}{l}
\partial_{t} R^{\pm}+\operatorname{div}\left(R^{\pm} u^{\pm}\right)=0, \\
\partial_{t}\left(R^{+} u^{+}\right)+\operatorname{div}\left(R^{+} u^{+} \otimes u^{+}\right)+\alpha^{+} \mathcal{C}^{2}\left[\rho^{-} \nabla R^{+}+\left(\rho^{+}+\rho^{+} \frac{\alpha^{-} f^{\prime}}{s_{-}^{2}}\right) \nabla R^{-}\right] \\
\hspace{2.5cm}=\operatorname{div}\left\{\alpha^{+}\left[\mu^{+}\left(\nabla u^{+}+\nabla^{t} u^{+}\right)
+\lambda^{+} \operatorname{div} u^{+} \operatorname{Id}\right]\right\}, \\
\partial_{t}\left(R^{-} u^{-}\right)+\operatorname{div}\left(R^{-} u^{-} \otimes u^{-}\right)+\alpha^{-} \mathcal{C}^{2}\left[\rho^{-} \nabla R^{+}+\left(\rho^{+}-\rho^{-} \frac{\alpha^{+} f^{\prime}}{s_{+}^{2}}\right) \nabla R^{-}\right] \\
\hspace{2.5cm}=\operatorname{div}\left\{\alpha^{-}\left[\mu^{-}\left(\nabla
u^{-}+\nabla^{t} u^{-}\right)+\lambda^{-} \operatorname{div} u^{-}
\operatorname{Id}\right]\right\}.
\end{array}\right.
\end{equation}
In the present paper, we consider the initial value problem to
\eqref{1.13} in the whole space $\mathbb R^3$ subject to the initial condition
\begin{equation}\label{1.14} (R^{+}, u^{+}, R^{-}, u^{-})(x,
0)=(R_{0}^{+}, u_{0}^{+}, R_{0}^{-},
u_{0}^{-})(x)\rightarrow(R_{\infty}^{+}, \overrightarrow{0}, R_{\infty}^{-}, \overrightarrow{0}) \quad
\hbox{as}\quad |x|\rightarrow\infty \in \mathbb{R}^{3},
\end{equation}
where $R^{\pm}_\infty>0$ denote the background doping profile, and
for simplicity, are taken as 1 in this paper.
In this work, we investigate the instability of the constant equilibrium state for the Cauchy problem \eqref{1.13}--\eqref{1.14} in the case that $f^{\prime}(1)>0$, which should be kept in mind throughout the rest of this paper. Taking
\[ n^{\pm}=R^{\pm}-1,
\]
then we can rewrite \eqref{1.13} in terms of the varaibles $(n^+, u^+, n^-, u^-)$:
\begin{equation}\label{1.15}
\left\{\begin{array}{l}
\partial_{t} n^++\operatorname{div}u^+=F_1, \\
\partial_{t}u^{+}+\alpha_1\nabla n^++\alpha_2\nabla
n^--\nu^+_1\Delta u^+-\nu^+_2\nabla\operatorname{div} u^+=F_2, \\
\partial_{t} n^-+\operatorname{div}u^-=F_3, \\
\partial_{t}u^{-}+\alpha_3\nabla n^++\alpha_4\nabla
n^--\nu^-_1\Delta u^--\nu^-_2\nabla\operatorname{div} u^-=F_4, \\
\end{array}\right.
\end{equation}
where $\nu_{1}^{\pm}=\frac{\mu^{\pm}}{\rho^{\pm}(1,1)}$,
$\nu_{2}^{\pm}=\frac{\mu^{\pm}+\lambda^{\pm}}{\rho^{\pm}(1,1)}>0$,
$\alpha_{1}=\frac{\mathcal{C}^{2}(1,1)
\rho^{-}(1,1)}{\rho^{+}(1,1)}$,
$\alpha_{2}=\mathcal{C}^{2}(1,1)+\frac{\mathcal{C}^{2}(1,1)
\alpha^{-}(1,1) f^{\prime}(1)}{s_{-}^{2}(1,1)}$,
$\alpha_{3}=\mathcal{C}^{2}(1,1)$,
$\alpha_{4}=\frac{\mathcal{C}^{2}(1,1)
\rho^{+}(1,1)}{\rho^{-}(1,1)}-\frac{\mathcal{C}^{2}(1,1)
\alpha^{+}(1,1) f^{\prime}(1)}{s_{+}^{2}(1,1)}$,
and
the nonlinear terms are given by
\begin{align}
\label{1.16}F_{1}=&-\operatorname{div}\left(n^{+} u^{+}\right), \\
F_{2}^{i}=&-g_+\left(n^{+}, n^{-}\right) \partial_{i} n^{+}-\bar{g}_{+}\left(n^{+}, n^{-}\right) \partial_{i} n^{-}
-\left(u^{+} \cdot \nabla\right) u_{i}^{+} \nonumber\\
&+\mu^{+} h_{+}\left(n^{+}, n^{-}\right) \partial_{j}
n^{+} \partial_{j} u_{i}^{+}+\mu^{+} k_{+}\left(n^{+}, n^{-}\right) \partial_{j} n^{-} \partial_{j} u_{i}^{+} \nonumber\\
\label{1.17}&+\mu^{+} h_{+}\left(n^{+}, n^{-}\right) \partial_{j}
n^{+} \partial_{i} u_{j}^{+}+\mu^{+} k_{+}\left(n^{+}, n^{-}\right)
\partial_{j} n^{-}
\partial_{i} u_{j}^{+}\\
&+\lambda^{+} h_{+}\left(n^{+}, n^{-}\right) \partial_{i} n^{+}
\partial_{j} u_{j}^{+}+\lambda^{+} k_{+}\left(n^{+},
n^{-}\right) \partial_{i} n^{-} \partial_{j} u_{j}^{+} \nonumber\\
&+\mu^{+} l_{+}\left(n^{+}, n^{-}\right) \partial_{j}^{2} u_{i}^{+}+\left(\mu^{+}+\lambda^{+}\right) l_{+}\left(n^{+}, n^{-}\right) \partial_{i}
\partial_{j} u_{j}^{+}, \nonumber\\
\label{1.18}F_{3}=&-\operatorname{div}\left(n^{-} u^{-}\right), \\
F_{4}^{i}=&-g_-\left(n^{+}, n^{-}\right) \partial_{i} n^{-}-
\bar{g}_{-}\left(n^{+}, n^{-}\right) \partial_{i} n^{+}-\left(u^{-} \cdot \nabla\right) u_{i}^{-}\nonumber \\
&+\mu^{-} h_{-}\left(n^{+}, n^{-}\right) \partial_{j} n^{+} \partial_{j} u_{i}^{-}+\mu^{-} k_{-}\left(n^{+}, n^{-}\right)
\partial_{j} n^{-} \partial_{j} u_{i}^{-} \nonumber\\
\label{1.19}&+\mu^{-} h_{-}\left(n^{+}, n^{-}\right) \partial_{j}
n^{+}
\partial_{i} u_{j}^{-}+\mu^{-} k_{-}\left(n^{+}, n^{-}\right)
\partial_{j} n^{-} \partial_{i} u_{j}^{-} \\
&+\lambda^{-} h_{-}\left(n^{+}, n^{-}\right) \partial_{i} n^{+} \partial_{j} u_{j}^{-}+\lambda^{-} k_{-}\left(n^{+}, n^{-}\right)
\partial_{i} n^{-} \partial_{j} u_{j}^{-}\nonumber \\
&+\mu^{-} l_{-}\left(n^{+}, n^{-}\right) \partial_{j}^{2}
u_{i}^{-}+\left(\mu^{-}+\lambda^{-}\right) l_{-}\left(n^{+},
n^{-}\right) \partial_{i} \partial_{j} u_{j}^{-},\nonumber
\end{align}
where
\begin{equation}\label{1.20}
\left\{\begin{array}{l}
g_{+}\left(n^{+}, n^{-}\right)=\frac{\left(\mathcal{C}^{2} \rho^{-}\right)\left(n^{+}+1, n^{-}+1\right)}{\rho^{+}\left(n^{+}+1, n^{-}+1\right)}-\frac{\left(\mathcal{C}^{2} \rho^{-}\right)(1,1)}{\rho^{+}(1,1)}, \\
g_{-}\left(n^{+}, n^{-}\right)=\frac{\left(\mathcal{C}^{2} \rho^{+}\right)\left(n^{+}+1, n^{-}+1\right)}{\rho^{-}\left(n^{+}+1, n^{-}+1\right)}-\frac{\left(\mathcal{C}^{2} \rho^{+}\right)(1,1)}{\rho^{-}(1,1)}-\frac{f^{\prime}\left(n^{-}+1\right)\left(\mathcal{C}^{2} \alpha^{+}\right)\left(n^{+}+1, n^{-}+1\right)}{s_{+}^{2}\left(n^{+}+1, n^{-}+1\right)} \\
\hspace{2.2cm}+\frac{f^{\prime}(1)\left(\mathcal{C}^{2} \alpha^{+}\right)(1,1)}{s_{+}^{2}(1,1)}, \\
\end{array}\right.\end{equation}
\begin{equation}\label{1.21}
\left\{\begin{array}{l}
\bar{g}_{+}\left(n^{+}, n^{-}\right)=\mathcal{C}^{2}\left(n^{+}+1, n^{-}+1\right)-=\mathcal{C}^{2}\left(1, 1\right)
+\frac{f^{\prime}\left(n^{-}+1\right)\left(\mathcal{C}^{2} \alpha^{-}\right)\left(n^{+}+1, n^{-}+1\right)}{s_{-}^{2}\left(n^{+}+1, n^{-}+1\right)}\\
\hspace{2.2cm}-\frac{f^{\prime}(1)\left(\mathcal{C}^{2} \alpha^{-}\right)(1,1)}{s_{-}^{2}(1,1)},\\
\bar{g}_{-}\left(n^{+}, n^{-}\right)=\mathcal{C}^{2}\left(n^{+}+1, n^{-}+1\right)-\mathcal{C}^{2}(1,1),\\
\end{array}\right.
\end{equation}
\begin{equation}\label{1.22}
\left\{\begin{array}{l}
h_{+}\left(n^{+}, n^{-}\right)=\frac{\left(\mathcal{C}^{2}\alpha^{-}\right)\left(n^{+}+1, n^{-}+1\right)}{(n^++1)s_{-}^{2}\left(n^{+}+1, n^{-}+1\right)},\\
h_{-}\left(n^{+}, n^{-}\right)=-\frac{\left(\mathcal{C}^{2} \right)\left(n^{+}+1, n^{-}+1\right)}{(\rho^-s_{-}^{2})\left(n^{+}+1, n^{-}+1\right)},
\end{array}\right.
\end{equation}
\begin{equation}\label{1.23}
\left\{\begin{array}{l}
k_{+}\left(n^{+}, n^{-}\right)=-\left[\frac{\mathcal{C}^{2}\left(n^{+}+1, n^{-}+1\right)}{(n^++1)(s_{+}^{2}\rho^+)\left(n^{+}+1, n^{-}+1\right)}+\frac{f^{\prime}(n^-+1)\mathcal{C}^{2}\left(n^{+}+1, n^{-}+1\right)}{(\rho^+\rho^-s_{+}^{2}s_{-}^{2})\left(n^{+}+1, n^{-}+1\right)}\right],\\
k_{-}\left(n^{+}, n^{-}\right)=-\frac{\left(\alpha^+\mathcal{C}^{2}\right)\left(n^{+}+1, n^{-}+1\right)}{(n^-+1)s_{+}^{2}\left(n^{+}+1, n^{-}+1\right)}+\frac{f^{\prime}(n^-+1)\left(\alpha^+\mathcal{C}^{2}\right)\left(n^{+}+1, n^{-}+1\right)}{(\rho^-s_{+}^{2}s_{-}^{2})\left(n^{+}+1, n^{-}+1\right)},\\
\end{array}\right.
\end{equation}
\begin{equation}\label{1.24}
l_{\pm}(n^+, n^-)=\frac{1}{\rho_{\pm}\left(n^{+}+1, n^{-}+1\right)}-\frac{1}{\rho_{\pm}\left(1, 1\right)}.
\end{equation}
Taking change of variables by
\[
n^{+} \rightarrow \alpha_{1} n^{+}, \quad u^{+} \rightarrow \sqrt{\alpha_{1} u^{+}}, \quad n^{-} \rightarrow \alpha_{4} n^{-}, \quad u^{-} \rightarrow \sqrt{\alpha_{4} u^{-}},
\]
and setting
\[
\beta_{1}=\sqrt{\alpha_{1}}, \quad \beta_{2}=\frac{\alpha_{2} \sqrt{\alpha_{1}}}{\alpha_{4}},
\quad \beta_{3}=\frac{\alpha_{3} \sqrt{\alpha_{4}}}{\alpha_{1}}, \quad \beta_{4}=\sqrt{\alpha_{4}}
\]
and
\[
\beta^{+}=\sqrt{\frac{\beta_{1}}{\beta_{2}}}, \quad \beta^{-}=\sqrt{\frac{\beta_{4}}{\beta_{3}}},
\]
the Cauchy problem \eqref{1.13} and \eqref{1.14} can be reformulated as
\begin{equation}\label{1.25}
\left\{\begin{array}{l}
\partial_{t} n^{+}+\beta_{1} \operatorname{div} u^{+}=\mathcal{F}_{1}, \\
\partial_{t} u^{+}+\beta_{1} \nabla n^{+}+\beta_{2} \nabla n^{-}-v_{1}^{+} \Delta u^{+}-v_{2}^{+} \nabla \operatorname{div} u^{+}=\mathcal{F}_{2}, \\
\partial_{t} n^{-}+\beta_{4} \operatorname{div} u^{-}=\mathcal{F}_{3}, \\
\partial_{t} u^{-}+\beta_{3} \nabla n^{+}+\beta_{4} \nabla n^{-}-v_{1}^{-} \Delta u^{-}-v_{2}^{-} \nabla \operatorname{div} u^{-}=\mathcal{F}_{4},
\end{array}\right.
\end{equation}
subject to the initial condition
\begin{equation}\label{1.26}
\left(n^{+}, u^{+}, n^{-}, u^{-}\right)(x, 0)=\left(n_{0}^{+}, u_{0}^{+}, n_{0}^{-}, u_{0}^{-}\right)(x) \rightarrow(0, \overrightarrow{0}, 0, \overrightarrow{0}), \quad \text { as }|x| \rightarrow+\infty,
\end{equation}
where the nonlinear terms are given by
\[
\mathcal{F}_{1}=\alpha_{1} F_{1}\left(\frac{n^{+}}{\alpha_{1}}, \frac{u^{+}}{\sqrt{\alpha_{1}}}\right), \quad \mathcal{F}_{2}=\sqrt{\alpha_{1}} F_{2}
\left(\frac{n^{+}}{\alpha_{1}}, \frac{u^{+}}{\sqrt{\alpha_{1}}}, \frac{n^{-}}{\alpha_{4}}, \frac{u^{-}}{\sqrt{\alpha_{4}}}\right),
\]
and
\[
\mathcal{F}_{3}=\alpha_{4} F_{3}\left(\frac{n^{-}}{\alpha_{4}}, \frac{u^{-}}{\sqrt{\alpha_{4}}}\right), \quad \mathcal{F}_{4}=\sqrt{\alpha_{4}} F_{4}\left(\frac{n^{+}}{\alpha_{1}}, \frac{u^{+}}{\sqrt{\alpha_{1}}}, \frac{n^{-}}{\alpha_{4}}, \frac{u^{-}}{\sqrt{\alpha_{4}}}\right).
\]
Noticing that
\begin{equation}\label{1.27}
\beta_{1} \beta_{4}-\beta_{2} \beta_{3}=-\frac{\mathcal{C}^{2}(1,1) f^{\prime}(1)}{\sqrt{\alpha_{1} \alpha_{4}} \rho^{+}(1,1)}<0,
\end{equation}
it is clear that $\beta^+\beta^-<1$. Before stating our main results, let us state the
corresponding linearized system of \eqref{1.25} as follows:
\begin{equation}\label{1.28}
\left\{\begin{array}{l}
\partial_{t} \tilde n^{+}+\beta_{1} \operatorname{div} \tilde u^{+}=0, \\
\partial_{t}\tilde u^{+}+\beta_{1} \nabla\tilde n^{+}+\beta_{2} \nabla\tilde n^{-}-v_{1}^{+} \Delta\tilde u^{+}-v_{2}^{+} \nabla \operatorname{div}\tilde u^{+}=0, \\
\partial_{t}\tilde n^{-}+\beta_{4} \operatorname{div}\tilde u^{-}=0, \\
\partial_{t}\tilde u^{-}+\beta_{3} \nabla\tilde n^{+}+\beta_{4} \nabla\tilde n^{-}-v_{1}^{-} \Delta\tilde u^{-}-v_{2}^{-} \nabla \operatorname{div}\tilde u^{-}=0.
\end{array}\right.
\end{equation}
\bigskip
\medskip
Now, we are in a position to state our main results. The first one is concerned with
the linear instability, which is stated in the following theorem.
\smallskip
\begin{Theorem}[Linear instability]\label{2mainth} Let $\theta=\frac{\sqrt{(\nu^+\beta_4^2+\nu^-\beta_1^2)^2+4\nu^+\nu^-(\beta_1\beta_2\beta_3\beta_4-\beta_1^2\beta_4^2)}
-(\nu^+\beta_4^2+\nu^-\beta_1^2)}{2\nu^+\nu^-}$ which is positive due to \eqref{1.27}, where $\nu^\pm=\nu_1^\pm+\nu_2^\pm.$ Then for any $\vartheta>0$, the linearized system \eqref{1.28} admits an unstable solution $(\tilde n^+_\vartheta,\tilde u^+_\vartheta,\tilde n^-_\vartheta,\tilde u^-_\vartheta)$ satisfying
$$\tilde n^\pm_\vartheta \in C^0(0, \infty; H^{2}(\mathbb{R}^3))\cap C^1(0, \infty; H^{1}(\mathbb{R}^3)),\quad \text{and}\quad \tilde u^\pm_\vartheta\in C^0(0, \infty; H^{2}(\mathbb{R}^3))\cap C^1(0, \infty; L^{2}(\mathbb{R}^3)),$$
and
\begin{equation}\label{1.29}\left\|\tilde n^+_{0,\vartheta}\right\|_{L^2}\left\|\tilde u^+_{0,\vartheta}\right\|_{L^2}\left\|\tilde n^-_{0,\vartheta}\right\|_{L^2}\left\|\tilde u^-_{0,\vartheta}\right\|_{L^2}>0.
\end{equation}
Moreover, the solution satisfies the following estimate:
\begin{equation}\text{e}^{(\theta-\vartheta) t} \|\tilde{n}^\pm_{0,\vartheta}\|_{L^2}\le \|\tilde{n}^\pm_\vartheta(t)\|_{L^2}\le \text{e}^{\theta t}\|\tilde{n}^\pm_{0,\vartheta}\|_{L^2} \quad \text{and}\quad \text{e}^{(\theta-\vartheta) t}\|\tilde{u}^\pm_{0,\vartheta}\|_{L^2}\le\|\tilde{u}^\pm_\vartheta(t)\|_{L^2}\le\text{e}^{\theta t} \|\tilde{u}^\pm_{0,\vartheta}\|_{L^2}.\label{1.30}\end{equation}
\end{Theorem}
\smallskip
\begin{remark} For any $\epsilon>0$ which may be small enough, it is direct to check that $(\epsilon\tilde n^+,\epsilon\tilde u^+,\epsilon\tilde n^-,\epsilon\tilde u^-)$ is still a solution of system \eqref{1.28}. This solution is obvious unstable due to \eqref{1.29} and \eqref{1.30}.
\end{remark}
\smallskip
The second result is concerned with nonlinear instability, which is stated in the following theorem.
\smallskip
\begin{Theorem}[Nonlinear instability]\label{3mainth} The steady state $(0, \overrightarrow{0}, 0, \overrightarrow{0})$ of the system \eqref {1.25} is unstable in the Hadamard sense, that is, there exist positive constants $\theta$, $\vartheta$, $\epsilon_0$ and $\delta_0$, and functions $( \tilde n_{0,\vartheta}^+,\tilde u_{0,\vartheta}^+,\tilde n_{0,\vartheta}^-,\tilde u_{0,\vartheta}^-)\in H^4(\mathbb R^3)$, such that for any $\epsilon\in(0,\epsilon_0)$ and the initial data
\begin{equation}\label{1.32}( n_0^+, u_0^+, n_0^-,u_0^-)\triangleq\epsilon( \tilde n_{0,\vartheta}^+,\tilde u_{0,\vartheta}^+,\tilde n_{0,\vartheta}^-,\tilde u_{0,\vartheta}^-),\end{equation}
the Cauchy problem \eqref{1.25} and \eqref{1.32} admits a unique strong solution satisfying
$$ n^\pm \in C^0(0, T^{\max}; H^{4}(\mathbb{R}^3))\cap C^1(0, T^{\max}; H^{3}(\mathbb{R}^3))\quad \text{and}\quad u^\pm\in C^0(0, T^{\max}; H^{4}(\mathbb{R}^3))\cap C^1(0, T^{\max}; H^{2}(\mathbb{R}^3)),$$
and
\begin{equation}\label{1.33}\left\| (n^+,u^+,n^-,u^-)(T^\varepsilon)\right\|_{H^4}\ge \delta_0.
\end{equation}
for some escape time $T^\varepsilon\in [0,T^{\max})$, where $T^{\max}$ denotes the maximal time of existence of the solution.
\end{Theorem}
\smallskip
\begin{remark} Theorem \ref{2mainth} and Theorem \ref{3mainth} show that for the case $f'(1)>0$, the constant equilibrium state of the two--fluid model is linearly globally unstable and nonlinearly locally unstable
in the sense of Hadamard, which is in contrast to
the cases $f'(1)<0$ in Evje--Wang--Wen \cite{Evje9} and $P^+=P^-$ (corresponding to $f'(1)=0$) in Wu--Yao--Zhang \cite{WYZ} where the constant equilibrium state of the two--fluid model \eqref{1.5} is nonlinearly globally stable.
\end{remark}
Now, let us sketch the main ideas in the proofs of Theorem \ref{2mainth} and Theorem \ref{3mainth}.
For the proof of Theorem \ref{2mainth}, we need construct a solution to the linearized system \eqref{1.28} that has a growing $H^k$
norm for any $k$ and the proof can be outlined as follows.
First, we exclude the stabilizing part of the linearized system by employing the Hodge decomposition technique firstly
introduced by Danchin \cite{Dan1} to split the linearized system into three systems (see \eqref {2.1} and \eqref {2.2} for details). One is a $4\times 4$ system and
its characteristic polynomial possesses four distinct roots, the other two systems are the heat equation.
This key observation allows us to construct an unstable solution.
Second, we assume a growing mode ansatz, i.e., $$\widehat{\tilde{n}^+}=\text{e}^{\lambda(|\xi|)t}\widehat{\tilde{n}^+_0},\
\widehat{\tilde{\varphi}^+}=\text{e}^{\lambda(|\xi|)t}\widehat{\tilde{\varphi}^+_0},\
\widehat{\tilde{n}^-}=\text{e}^{\lambda(|\xi|)t}\widehat{\tilde{n}^-_0},\
\widehat{\tilde{\varphi}^-}=\text{e}^{\lambda(|\xi|)t}\widehat{\tilde{\varphi}^-_0}, ~~\hbox{for some}~\lambda,$$
and submit this ansatz into the Fourier transformation of the $4\times 4$ system to get
a time--independent system for $\lambda$.
Third, we solve the time--independent system by making careful analysis and using several key observations. Indeed, noticing that
the characteristic polynomial $F(\lambda)$ defined in \eqref{2.6} is a strictly increasing function on $(0, \infty)$, and
$F(\theta)>0$ for $\theta>0$ defined in Theorem \ref{2mainth}, we show that $0<\lambda_1<\theta$ is the unique positive root of the characteristic equation $F(\lambda)=0$, and $\theta>0$ in Theorem \ref{2mainth} is the largest possible growth rate since $Re(\lambda_i)\leq \theta$ with $1\leq i\leq 4$. Therefore, the growing mode constructed in Theorem \ref{2mainth} actually does grow in time at
the fastest possible rate.\par
For the proof of Theorem \ref{3mainth}, we deduce the nonlinear instability. Compared to \cite{Guo1,Jang,Jiang1,Jiang2,WangT} where nonlinear energy estimates and a careful bootstrap argument are employed to prove stability and instability, we need to develop new
ingredients in the proof to handle with the difficulties arising from the strong interaction of two fluids,
which requires some new thoughts. Indeed, since the strong
coupling terms are involved in the right--hand of the system \eqref{1.25}, it seems impossible to follow the energy methods of \cite{Guo1,Jang,Jiang1,Jiang2,WangT}
to get the lyapunov--type inequality: $\frac{d}{dt}\mathcal{E}(t)\leq \theta \mathcal{E}(t)$ to prove the largest possible growth rate.
Therefore, we must pursue another route by resorting to semigroup methods to capture the largest possible growth rate, but the cost is that
we need the higher regularity of the solutions. More precisely,
with the help of the global linear instability result of Theorem \ref{2mainth} and a local existence theorem of classical solutions to the original nonlinear system, we can make delicate spectral analysis for the linearized system and apply Duhamel's principle to prove the nonlinear instability
result stated in Theorem \ref{2mainth}.
\subsection{Notations and conventions.}
Throughout this paper, we denote $H^k(\mathbb R^3)$ by the usual Sobolev spaces with norm $\|\cdot\|_{H^k}$ and
denote $L^p$, $1\leq p\leq \infty$ by the usual $L^p(\mathbb R^3)$
spaces with norm $\|\cdot\|_{L^p}$. We drop the domain $\mathbb R^3$ in integrands over $\mathbb R^3$. For the sake of conciseness, we
do not precise in functional space names when they are concerned
with scalar--valued or vector--valued functions, $\|(f, g)\|_X$
denotes $\|f\|_X+\|g\|_X$. We will employ the notation $a\lesssim
b$ to mean that $a\leq Cb$ for a universal constant $C>0$ that only
depends on the parameters coming from the problem. We denote
$\nabla=\partial_x=(\partial_1,\partial_2,\partial_3)$, where
$\partial_i=\partial_{x_i}$, $\nabla_i=\partial_i$ and put
$\partial_x^\ell f=\nabla^\ell f=\nabla(\nabla^{\ell-1}f)$. Let
$\Lambda^s$ be the pseudo differential operator defined by
\begin{equation}\Lambda^sf=\mathfrak{F}^{-1}(|{\bf \xi}|^s\widehat f),~\hbox{for}~s\in \mathbb{R},\nonumber\end{equation}
where $\widehat f$ and $\mathfrak{F}(f)$ are the Fourier transform
of $f$.
\smallskip
\bigskip
\section{\leftline {\bf{Linear instability.}}}
\setcounter{equation}{0}
To construct a solution to the linearized system \eqref{1.28} that has growing $H^k$--norm for any positive integer $k$, by using a real method
as in \cite{Kowalczyk}, one need to make a detailed
analysis on the properties of the semigroup. To exclude the stabilizing part, we will employ the Hodge decomposition technique firstly
introduced by Danchin \cite{Dan1} to split the linear system into three systems. One only has four
equations and its characteristic polynomial possesses four distinct roots, the other two systems
are the heat equation. This key observation allows us to construct a unstable solution.
To see this, let
$\varphi^{\pm}=\Lambda^{-1}{\rm div}\tilde{u}^{\pm}$
be the ``compressible part" of the velocities $\tilde{u}^{\pm}$, and denote $\phi^{\pm}=\Lambda^{-1}{\rm curl}\tilde{u}^{\pm}$ (with $({\rm curl} z)_i^j
=\partial_{x_j}z^i-\partial_{x_i}z^j$) by the ``incompressible part"
of the velocities $\tilde{u}^{\pm}$. Setting $\nu^{\pm}=\nu^{\pm}_1+\nu^{\pm}_2$, the system \eqref{1.28} can be
decomposed into
the following three systems:
\begin{equation}\label{2.1}
\begin{cases}
\partial_t{\tilde{n}^+}+\beta_1\Lambda{\varphi^+}=0,\\
\partial_t{\varphi^+}-\beta_1\Lambda{\tilde{n}^+}-\beta_2\Lambda{\tilde{n}^-}+\nu^+\Lambda^2{\varphi^+}=0,\\
\partial_t{\tilde{n}^-}+\beta_4\Lambda{\varphi^-}=0,\\
\partial_t{\varphi^-}-\beta_3\Lambda{\tilde{n}^+}-\beta_4\Lambda{\tilde{n}^-}+\nu^-\Lambda^2{\varphi^-}=0,\\
\end{cases}
\end{equation}
and
\begin{equation}\label{2.2}
\begin{cases}
\partial_t\phi^++\nu^+_1\Lambda^2\phi^+=0,\\
\partial_t\phi^-+\nu^-_1\Lambda^2\phi^-=0.
\end{cases}
\end{equation}
We see that Eqs. \eqref{2.2}$_1$ and \eqref{2.2}$_2$ are the standard parabolic equations with good stability. Thus, the onset of instabilities of system \eqref{1.28} comes from \eqref{2.1}. Taking the Fourier transform to the system \eqref{2.1}, one has
\begin{equation}\label{2.3}
\begin{cases}
\partial_t\widehat{\tilde{n}^+}+\beta_1|\xi|\widehat{\varphi^+}=0,\\
\partial_t\widehat{\varphi^+}-\beta_1|\xi|\widehat{\tilde{n}^+}-\beta_2|\xi|\widehat{\tilde{n}^-}+\nu^+|\xi|^2\widehat{\varphi^+}=0,\\
\partial_t\widehat{\tilde{n}^-}+\beta_4|\xi|\widehat{\varphi^-}=0,\\
\partial_t\widehat{\varphi^-}-\beta_3|\xi|\widehat{\tilde{n}^+}-\beta_4|\xi|\widehat{\tilde{n}^-}+\nu^-|\xi|^2\widehat{\varphi^-}=0.\\
\end{cases}
\end{equation}
To construct a solution to the linearized equations \eqref{2.3} that has growing $H^k$--norm for any $k$, we shall make a growing normal mode ansatz of solutions, i.e.,
$$\widehat{\tilde{n}^+}=\text{e}^{\lambda(|\xi|)t}\widehat{\tilde{n}^+_0},\
\widehat{\tilde{\varphi}^+}=\text{e}^{\lambda(|\xi|)t}\widehat{\tilde{\varphi}^+_0},\
\widehat{\tilde{n}^-}=\text{e}^{\lambda(|\xi|)t}\widehat{\tilde{n}^-_0},\
\widehat{\tilde{\varphi}^-}=\text{e}^{\lambda(|\xi|)t}\widehat{\tilde{\varphi}^-_0}.$$
Substituting this ansazt into \eqref{2.3}, one obtains the time--independent system
\begin{equation}\label{2.4}
\begin{cases}
\lambda{\widehat{\tilde{n}^+_0}}+\beta_1|\xi|\widehat{{\varphi^+_0}}=0,\\
\lambda\widehat{{\varphi^+_0}}-\beta_1|\xi|\widehat{{\tilde{n}^+_0}}-\beta_2|\xi|\widehat{{\tilde{n}^-_0}}+\nu^+|\xi|^2\widehat{{\varphi^+_0}}=0,\\
\lambda\widehat{{\tilde{n}^-_0}}+\beta_4|\xi|\widehat{{\varphi^-_0}}=0,\\
\lambda\widehat{{\varphi^-_0}}-\beta_3|\xi|\widehat{{\tilde{n}^+_0}}-\beta_4|\xi|\widehat{{\tilde{n}^-_0}}+\nu^-|\xi|^2\widehat{{\varphi^-_0}}=0.\\
\end{cases}
\end{equation}
After a series of tedious but direct calculations, we can conclude from \eqref{2.4} that
\begin{equation}\begin{split}\label{2.5}
&[\lambda^4+(\nu^+|\xi|^2+\nu^-|\xi|^2)\lambda^3+(\beta_1^2|\xi|^2+\beta_4^2|\xi|^2+\nu^+\nu^-|\xi|^4)\lambda^2\\&+(\nu^+\beta_4^2|\xi|^4+\nu^-\beta_1^2|\xi|^4)\lambda
+\beta_1^2\beta_4^2|\xi|^4-\beta_1\beta_2\beta_3\beta_4|\xi|^4]\widehat{{\varphi^-_0}}=0.
\end{split}\end{equation}
Therefore, the system \eqref{2.4} has non--zero solutions if the characteristic equation
\begin{equation}\begin{split}\label{2.6}
F(\lambda)=&\lambda^4+(\nu^+|\xi|^2+\nu^-|\xi|^2)\lambda^3+(\beta_1^2|\xi|^2+\beta_4^2|\xi|^2+\nu^+\nu^-|\xi|^4)\lambda^2\\&+(\nu^+\beta_4^2|\xi|^4+\nu^-\beta_1^2|\xi|^4)\lambda
+\beta_1^2\beta_4^2|\xi|^4-\beta_1\beta_2\beta_3\beta_4|\xi|^4=0
\end{split}\end{equation}
has a real characteristic root.
\begin{Lemma}\label{lemma2.1}
There exists a positive constant $\eta_1\gg 1 $, such that for
$|\xi|\ge \eta_1$, the characteristic equation \eqref{2.6} admits a real positive solution satisfying the following Taylor series
expansion
\begin{equation}\lambda_1=\theta+\mathcal {O}(|\xi|^{-1}).\label{2.7}\end{equation}
Moreover, the following estimate holds
\begin{equation}\lambda_1<\theta \quad \text{for any}\quad \xi\in\mathbb R^3.\label{2.8}\end{equation}
\end{Lemma}
\begin{proof}Employing
the similar argument of Taylor series expansion as in
\cite{Mat1}, then \eqref{2.7} follows from some tedious but direct calculations. It is noticed that
$F(\lambda)$ is a strictly monotonically increasing function if $\lambda>0$. Furthermore,
\begin{equation}\nonumber\begin{split}F(\theta)>\nu^+\nu^-|\xi|^4\theta^2+(\nu^+\beta_4^2|\xi|^4+\nu^-\beta_1^2|\xi|^4)\theta
+\beta_1^2\beta_4^2|\xi|^4-\beta_1\beta_2\beta_3\beta_4|\xi|^4=0,\end{split}\end{equation}
therefore \eqref{2.8} holds and the proof of lemma is completed.
\end{proof}
\smallskip
Let $\phi\in C_0^\infty(\mathbb
R^3_{{\bf \xi}})$ be a radial function satisfying $\phi({\bf \xi})=1$ when $\frac{3}{2}\eta\le |{\bf
\xi}|\leq 3\eta$ and $\phi({\bf \xi})=0$ when $|{\bf
\xi}|\le \eta$ and $|{\bf
\xi}|\ge 4\eta$. From \eqref{2.4}, we set
$${\widehat{\tilde{n}^+_0}}=\phi({\bf \xi}),\ \widehat{{\varphi^+_0}}=-\frac{\lambda_1(|\xi|)}{\beta_1|\xi|}\phi({\bf \xi}),\ {\widehat{\tilde{n}^-_0}}=-\frac{\lambda^2_1(|\xi|)+\beta_1^2|\xi|^2+\nu^+\lambda_1(|\xi|)|\xi|^2}{\beta_1\beta_2|\xi|^2}\phi({\bf \xi})$$
and
$$\widehat{{\varphi^-_0}}=\frac{\lambda^3_1(|\xi|)+\beta_1^2\lambda_1(|\xi|)|\xi|^2+\nu^+\lambda^2_1(|\xi|)|\xi|^2}{\beta_1\beta_2\beta_4|\xi|^3}\phi({\bf \xi}).$$
Then, it is direct to check that $(\widehat{\tilde{n}^+_0},\widehat{\tilde\varphi^+_0},\widehat{\tilde{n}^-_0},\widehat{\tilde\varphi^-_0})$ is a solution of the system \eqref{2.4}. Thus, we conclude the following proposition, which implies Theorem \ref{2mainth}.
\begin{Proposition}\label{Prop2.2} Let
$${\tilde{n}^\pm}=\mathfrak{F}^{-1}\left(\text{e}^{\lambda_1t}\widehat{\tilde{n}^\pm_0}\right)\quad \text{and}\quad
{\tilde{u}^\pm}=-\Lambda^{-1}\nabla\mathfrak{F}^{-1}\left(\text{e}^{\lambda_1(|\xi|)t}\widehat{\tilde{\varphi}^\pm_0}\right).$$
Then $(\tilde{n}^+,\tilde u^+,\tilde{n}^-,\tilde u^-)$ is a solution of \eqref{1.29} and
satisfies
\begin{equation}\text{e}^{(\theta-\vartheta) t} \|\tilde{n}^\pm_0\|_{L^2}\le \|\tilde{n}^\pm(t)\|_{L^2}\le \text{e}^{\theta t}\|\tilde{n}^\pm_0\|_{L^2} \quad \text{and}\quad \text{e}^{(\theta-\vartheta) t}\|\tilde{u}^\pm_0\|_{L^2}\le\|\tilde{u}^\pm(t)\|_{L^2}\le\text{e}^{\theta t} \|\tilde{u}^\pm_0\|_{L^2},\label{2.9}\end{equation}
if $\eta_1$ large enough.
\end{Proposition}
\begin{proof} Set $\phi^\pm\equiv0$. As the definition of $\varphi^\pm$ and $\phi^\pm$, and the relation
$$\tilde u^\pm=-\Lambda^{-1}\nabla\varphi^\pm-\Lambda^{-1}\text{div}\phi^\pm,$$
it is easy to prove that $(\tilde{n}^+,\tilde u^+,\tilde{n}^-,\tilde u^-)$ is a solution of \eqref{1.29}. Moreover, in virtue of Plancherel theorem, we have
\begin{equation}\begin{split}\|\tilde u^\pm(t)\|_{L^2}^2=&\|\widehat{\tilde u^\pm}(t)\|_{L^2}^2\\
=&\int\text{e}^{2\lambda_1(|\xi|)t}|\widehat{\tilde u^\pm_0}|^2\mathrm{d}\xi\\
=&\int_{\eta\le|\xi|\le 4|\eta|}\text{e}^{2\lambda_1(|\xi|)t}|\widehat{\tilde u^\pm_0}|^2\mathrm{d}\xi\\
\ge &~\text{e}^{2(\theta-\vartheta) t}\|\tilde u^\pm_0(t)\|_{L^2}^2,
\end{split}\end{equation}
if $\eta$ is large enough. Performing the similar procedures, we can prove $\|\tilde{u}^\pm(t)\|_{L^2}\le\text{e}^{\theta t} \|\tilde{u}^\pm_0\|_{L^2}$ and $\text{e}^{(\theta-\vartheta) t} \|\tilde{n}^\pm_0\|_{L^2}\le\|\tilde{n}^\pm(t)\|_{L^2}\le\text{e}^{\theta t} \|\tilde{n}^\pm_0\|_{L^2}$. The proof of proposition is complete.
\end{proof}
\bigskip
\section{Spectral analysis and linear $L^2$--estimates}\label{1section_appendix}
In this section, we are devoted to deriving the linear $L^2$--estimates, by using a real method
as in \cite{Mat1}, one need to make a detailed analysis on the
properties of the semigroup.
\subsection{Spectral analysis for system \eqref{2.1}} We consider the Cauchy problem of \eqref{2.1} with the initial data
\begin{equation}\label{3.1}(\tilde{n}^+, \varphi^+, \tilde{n}^-, \varphi^-)\big|_{t=0}=({n}^+_0, \Lambda^{-1}{\rm div}\tilde{u}^{+}_0, {n}^-_0, \Lambda^{-1}{\rm div}\tilde{u}^{-}_0)(x)
\end{equation}
In terms of the semigroup theory, we may represent the IVP \eqref{2.1} and \eqref{3.1} for $\mathcal U=(\tilde{n}^+, \varphi^+, \tilde{n}^-, \varphi^-)^t$ as
\begin{equation}
\begin{cases}
\mathcal U_t=\mathcal B_1\mathcal U,\\
\mathcal U\big|_{t=0}=\mathcal U_0,
\end{cases} \label{3.2}
\end{equation}
where the operator $\mathcal B_1$ is defined
by
\begin{equation}\nonumber\mathcal B_1=\begin{pmatrix}
0&-\beta_1\Lambda&0&0\\
\beta_1\Lambda&-\nu^+\Lambda^2&\beta_2\Lambda&0\\
0&0&0&-\beta_4\Lambda\\
\beta_3\Lambda&0&\beta_4\Lambda&-\nu^-\Lambda^2
\end{pmatrix}.\end{equation}
Taking the Fourier transform to the system \eqref{3.2}, we obtain
\begin{equation}
\begin{cases}
\widehat {\mathcal U}_t=\mathcal A_1(\xi)\widehat {\mathcal U},\\
\widehat {\mathcal U}\big|_{t=0}=\widehat {\mathcal U}_0,
\end{cases} \label{3.3}
\end{equation}
where $\widehat {\mathcal U}(\xi,t)=\mathfrak{F}({\mathcal U}(x,t))$ and $\mathcal A_1(\xi)$ is given by
\begin{equation}\nonumber\mathcal A_1(\xi)=\begin{pmatrix}
0&-\beta|\xi|&0&0\\
\beta_1|\xi|&-\nu^+|\xi|^2&\beta_2|\xi|&0\\
0&0&0&-\beta_4|\xi|\\
\beta_3|\xi|&0&\beta_4|\xi|&-\nu^-|\xi|^2
\end{pmatrix}.\end{equation}
We compute the eigenvalues of the matrix $\mathcal A_1(\xi)$ from the determinant
\begin{equation}\begin{split}\label{3.4}&{\rm det}(\lambda{\rm I}-\mathcal A_1(\xi))\\
&=\lambda^4+(\nu^+|\xi|^2+\nu^-|\xi|^2)\lambda^3+(\beta_1^2|\xi|^2+\beta_4^2|\xi|^2+\nu^+\nu^-|\xi|^4)\lambda^2+(\nu^+\beta_4^2|\xi|^4+\nu^-\beta_1^2|\xi|^4)\lambda\\
&\quad+\beta_1^2\beta_4^2|\xi|^4-\beta_1\beta_2\beta_3\beta_4|\xi|^4\\
&=0,
\end{split}\end{equation}
which is the same as characteristic equation \eqref{2.6} and implies that the matrix $\mathcal A_1(\xi)$ possesses four different
eigenvalues:
\begin{equation}\nonumber
\lambda_1=\lambda_1(|\xi|),\quad \lambda_2=\lambda_2(|\xi|),\quad
\lambda_3=\lambda_3(|\xi|),\quad \lambda_4=\lambda_4(|\xi|).
\end{equation}
Consequently, the semigroup $e^{t\mathcal A_1}$ can be decomposed
into
\begin{equation}\label{2.21}
\text{e}^{t\mathcal A_1(\xi)}=\sum_{i=1}^4\text{e}^{\lambda_it}P_i(\xi),
\end{equation}
where the projector $P_i(\xi)$ is defined by
\begin{equation}\label{2.22}
P_i(\xi)=\prod_{j\neq i}\frac{\mathcal A_1(\xi)-\lambda_jI}{\lambda_i-\lambda_j}, \quad i,j=1,2,3,4.
\end{equation}
Thus, the solution of IVP \eqref{3.3} can be expressed as
\begin{equation}
\widehat {\mathcal U}(\xi,t)=\text{e}^{t\mathcal A_1(\xi)}\widehat {\mathcal U}_0(\xi)=\left(\sum_{i=1}^4
\text{e}^{\lambda_it}P_i(\xi)\right)\widehat {\mathcal U}_0(\xi).\label{2.23}
\end{equation}
To derive long time properties of the semigroup
$\text{e}^{t\mathcal A_1}$ in $L^2$--framework, one need to analyze
the asymptotical expansions of $\lambda_i$, $P_i$ $(i =1, 2, 3, 4)$
and $\text{e}^{t\mathcal A_1(\xi)}$. Employing the similar argument of Taylor series expansion as in
\cite{Mat1}, we have the following lemmas from tedious
calculations.
\begin{Lemma}\label{lemma3.1}
There exists a positive constant $\eta_2\ll 1 $ such that, for
$|\xi|\leq \eta_2$, the spectral has the following Taylor series
expansion:
\begin{equation}\label{3.8-1}
\left\{\begin{array}{lll}\displaystyle \lambda_1=-\left[\frac{\nu^++\nu^-}{4}-\frac{\nu^+(\beta_1^2-\beta_4^2)+\nu^-(\beta_4^2-\beta_1^2)}{8\kappa_1}\right]|\xi|^2+\sqrt{\kappa_1-\kappa_2}|\xi|+\mathcal O(|\xi|^3),\\
\displaystyle\lambda_2=-\left[\frac{\nu^++\nu^-}{4}-\frac{\nu^+(\beta_1^2-\beta_4^2)+\nu^-(\beta_4^2-\beta_1^2)}{8\kappa_1}\right]|\xi|^2-\sqrt{\kappa_1-\kappa_2}|\xi|+\mathcal O(|\xi|^3),
\\ \displaystyle \lambda_3=-\left[\frac{\nu^++\nu^-}{4}+\frac{\nu^+(\beta_1^2-\beta_4^2)+\nu^-(\beta_4^2-\beta_1^2)}{8\kappa_1}\right]|\xi|^2
+\sqrt{\kappa_2+\kappa_1}\text{i}|\xi|+\mathcal O(|\xi|^3),
\\ \displaystyle \lambda_4=-\left[\frac{\nu^++\nu^-}{4}+\frac{\nu^+(\beta_1^2-\beta_4^2)+\nu^-(\beta_4^2-\beta_1^2)}{8\kappa_1}\right]|\xi|^2-\sqrt{\kappa_2+\kappa_1}\text{i}|\xi|+\mathcal O(|\xi|^3),
\end{array}\right.
\end{equation}
where $\kappa_1=\sqrt{\frac{(\beta_1^2-\beta_4^2)^2}{4}+\beta_1\beta_2\beta_3\beta_4}$ and $\displaystyle\kappa_2=\frac{\beta_1^2+\beta_4^2}{2}$.
\end{Lemma}
For $|\xi|\leq\eta_2$, from Lemma \ref{lemma3.1}, a direct computation gives
{\small \begin{equation}\label{3.9}\begin{split}P_1(\xi)=&\begin{pmatrix}
\frac{2\kappa_1+\beta_4^2-\beta_1^2}{8\kappa_1}&
\frac{\beta_1(2\kappa_1+\beta_4^2-\beta_1^2)}{8\kappa_1\sqrt{\kappa_1-\kappa_2}}&\frac{-\beta_1\beta_2}{4\kappa_1}&\frac{-\beta_1\beta_2\beta_4}{4\kappa_1\sqrt{\kappa_1-\kappa_2}}\\
\frac{\beta_1(\beta_1^2-\beta_4^2-2\kappa_1)+2\beta_2\beta_3\beta_4}{8\kappa_1
\sqrt{\kappa_1-\kappa_2}}&\frac{2\kappa_1+\beta_4^2-\beta_1^2}{8\kappa_1}&-\frac{\beta_2\sqrt{\kappa_1-\kappa_2}}{4\kappa_1}&\frac{-\beta_2\beta_4}{4\kappa_1}\\
\frac{-\beta_3\beta_4}{4\kappa_1}&\frac{-\beta_1\beta_3\beta_4}{4\kappa_1\sqrt{\kappa_1-\kappa_2}}&\frac{2\kappa_1+\beta_1^2-\beta_4^2}{8\kappa_1}&\frac{\beta_4(2\kappa_1
+\beta_1^2-\beta_4^2)}{8\kappa_1\sqrt{\kappa_1-\kappa_2}}\\
-\frac{\beta_3\sqrt{\kappa_1-\kappa_2}}{4\kappa_1}&\frac{-\beta_1\beta_3}{4\kappa_1}&\frac{\beta_4(\beta_4^2-\beta_1^2-2\kappa_1)+2\beta_1\beta_2\beta_3}{8\kappa_1\sqrt{\kappa_1-
\kappa_2}}&\frac{2\kappa_1+\beta_1^2-\beta_4^2}{8\kappa_1}
\end{pmatrix}+\mathcal O(|\xi|),\end{split}\end{equation}
\begin{equation}\label{3.10}\begin{split}P_2(\xi)=&\begin{pmatrix}
\frac{2\kappa_1+\beta_4^2-\beta_1^2}{8\kappa_1}&\frac{-\beta_1(2\kappa_1+\beta_4^2-\beta_1^2)}{8\kappa_1\sqrt{\kappa_1-\kappa_2}}&
\frac{-\beta_1\beta_2}{4\kappa_1}&\frac{\beta_1\beta_2\beta_4}{4\kappa_1\sqrt{\kappa_1-\kappa_2}}\\
-\frac{\beta_1(\beta_1^2-\beta_4^2-2\kappa_1)+2\beta_2\beta_3\beta_4}{8\kappa_1\sqrt{\kappa_1-\kappa_2}}&
\frac{2\kappa_1+\beta_4^2-\beta_1^2}{8\kappa_1}&\frac{\beta_2\sqrt{\kappa_1-\kappa_2}}{4\kappa_1}&\frac{-\beta_2\beta_4}{4\kappa_1}\\
\frac{-\beta_3\beta_4}{4\kappa_1}&\frac{\beta_1\beta_3\beta_4}{4\kappa_1\sqrt{\kappa_1-\kappa_2}}&\frac{2\kappa_1+\beta_1^2-\beta_4^2}
{8\kappa_1}&-\frac{\beta_4(2\kappa_1+\beta_1^2-\beta_4^2)}{8\kappa_1\sqrt{\kappa_1-\kappa_2}}\\
\frac{\beta_3\sqrt{\kappa_1-\kappa_2}}{4\kappa_1}&\frac{-\beta_1\beta_3}{4\kappa_1}&-\frac{\beta_4(\beta_4^2-\beta_1^2-2\kappa_1)+2\beta_1\beta_2\beta_3}
{8\kappa_1\sqrt{\kappa_1-\kappa_2}}&\frac{2\kappa_1+\beta_1^2-\beta_4^2}{8\kappa_1}
\end{pmatrix}+\mathcal O(|\xi|),\end{split}\end{equation}
\begin{equation}\label{3.11}\begin{split}P_3(\xi)=&\begin{pmatrix}
\frac{2\kappa_1+\beta_1^2-\beta_4^2}{8\kappa_1}&\frac{\beta_1(2\kappa_1+\beta_1^2-\beta_4^2)}
{8\kappa_1\sqrt{\kappa_2+\kappa_1}}{i}&\frac{\beta_1\beta_2}{4\kappa_1}&\frac{\beta_1\beta_2\beta_4}{4\kappa_1\sqrt{\kappa_2+\kappa_1}}{i}\\
-\frac{\beta_1(\beta_1^2-\beta_4^2+2\kappa_1)+2\beta_2\beta_3\beta_4}{8\kappa_1\sqrt{\kappa_2+\kappa_1}}{i}&\frac{2\kappa_1+\beta_1^2-\beta_4^2}
{8\kappa_1}&-\frac{\beta_2\sqrt{\kappa_2+\kappa_1}}{4\kappa_1}{i}&\frac{\beta_2\beta_4}{4\kappa_1}\\
\frac{\beta_3\beta_4}{4\kappa_1}&\frac{\beta_1\beta_3\beta_4}{4\kappa_1\sqrt{\kappa_2+\kappa_1}}{i}&
\frac{2\kappa_1+\beta_4^2-\beta_1^2}{8\kappa_1}&\frac{\beta_4(2\kappa_1+\beta_4^2-\beta_1^2)}{8\kappa_1\sqrt{\kappa_2+\kappa_1}}{i}\\
-\frac{\beta_3\sqrt{\kappa_2+\kappa_1}}{4\kappa_1}{i}&\frac{\beta_1\beta_3}{4\kappa_1}&-\frac{\beta_4(\beta_4^2-\beta_1^2+2\kappa_1)
+2\beta_1\beta_2\beta_3}{8\kappa_1\sqrt{\kappa_2+\kappa_1}}{i}&\frac{2\kappa_1+\beta_4^2-\beta_1^2}{8\kappa_1}
\end{pmatrix}+\mathcal O(|\xi|),\end{split}\end{equation}
and
\begin{equation}\label{3.12}\begin{split}P_4(\xi)=&\begin{pmatrix}
\frac{2\kappa_1+\beta_1^2-\beta_4^2}{8\kappa_1}&-\frac{\beta_1(2\kappa_1+\beta_1^2-\beta_4^2)}
{8\kappa_1\sqrt{\kappa_2+\kappa_1}}{i}&\frac{\beta_1\beta_2}{4\kappa_1}&-\frac{\beta_1\beta_2\beta_4}{4\kappa_1\sqrt{\kappa_2+\kappa_1}}{i}\\
\frac{\beta_1(\beta_1^2-\beta_4^2+2\kappa_1)+2\beta_2\beta_3\beta_4}{8\kappa_1\sqrt{\kappa_2+\kappa_1}}{i}&\frac{2\kappa_1+\beta_1^2-\beta_4^2}
{8\kappa_1}&\frac{\beta_2\sqrt{\kappa_2+\kappa_1}}{4\kappa_1}{i}&\frac{\beta_2\beta_4}{4\kappa_1}\\
\frac{\beta_3\beta_4}{4\kappa_1}&-\frac{\beta_1\beta_3\beta_4}{4\kappa_1\sqrt{\kappa_2+\kappa_1}}{i}&
\frac{2\kappa_1+\beta_4^2-\beta_1^2}{8\kappa_1}&-\frac{\beta_4(2\kappa_1+\beta_4^2-\beta_1^2)}{8\kappa_1\sqrt{\kappa_2+\kappa_1}}{i}\\
\frac{\beta_3\sqrt{\kappa_2+\kappa_1}}{4\kappa_1}{i}&\frac{\beta_1\beta_3}{4\kappa_1}&\frac{\beta_4(\beta_4^2-\beta_1^2+2\kappa_1)+2\beta_1\beta_2\beta_3}
{8\kappa_1\sqrt{\kappa_2+\kappa_1}}{i}&\frac{2\kappa_1+\beta_4^2-\beta_1^2}{8\kappa_1}
\end{pmatrix}+\mathcal O(|\xi|),\end{split}\end{equation}}
\begin{Lemma}\label{lemma3.2}
For $\eta_2\le |\xi|\le \eta_1$, there exists a positive constant $C$ such that
\begin{equation}\label{3.13}\text{Re}(\lambda_i)\le \theta \quad \text{and} \quad \left | P_i\right |\le C,
\end{equation}
for $1\le i\le 4.$
\end{Lemma}
\begin{Lemma}\label{lemma3.3}
There exists a positive constants $\eta_1\gg 1 $ such that, for
$|\xi|\geq \eta_1$, the spectral has the following Taylor series
expansion:
\begin{equation}\label{3.8}
\left\{\begin{array}{lll}\displaystyle \lambda_1=\theta+\mathcal O(|\xi|^{-1}),\\
\displaystyle\lambda_2=\frac{-(\nu^+\beta_4^2+\nu^-\beta_1^2)-\kappa_3
}{2\nu^+\nu^-}+\mathcal O(|\xi|^{-1}),
\\ \displaystyle \lambda_3=-\nu^+|\xi|^2+\frac{\beta_1^2}{\nu^+}
+\mathcal O(|\xi|^{-1}),
\\ \displaystyle \lambda_4=-\nu^-|\xi|^2+\frac{\beta_4^2}{\nu^-}+\mathcal O(|\xi|^{-1}),
\end{array}\right.
\end{equation}
where $\kappa_3=\sqrt{(\nu^+\beta_4^2+\nu^-\beta_1^2)^2+4\nu^+\nu^-(\beta_1\beta_2\beta_3\beta_4-\beta_1^2\beta_4^2)}$.
\end{Lemma}
For $|\xi|\geq\eta_1$, from Lemma \ref{lemma3.3}, a direct computation gives
\begin{equation}\label{3.15}\begin{split}P_1(\xi)=&\begin{pmatrix}
\frac{\nu^+\beta_4^2-\nu^-\beta_1^2+\kappa_3}{2\kappa_3}&
0&-\frac{\beta_1\beta_2\nu^-}{\kappa_3}&0\\
0&0&0&0\\
\frac{-\beta_3\beta_4\nu^+}{\kappa_3}&0&\frac{\nu^-\beta_1^2-\nu^+\beta_4^2+\kappa_3}{2\kappa_3}&0\\
0&0&0&0
\end{pmatrix}+\mathcal O(|\xi|^{-1}),\end{split}\end{equation}
\begin{equation}\label{3.16}\begin{split}P_2(\xi)=&\begin{pmatrix}
\frac{\nu^-\beta_1^2-\nu^+\beta_4^2+\kappa_3}{2\kappa_3}&
0&\frac{\beta_1\beta_2\nu^-}{\kappa_3}&0\\
0&0&0&0\\
\frac{\beta_3\beta_4\nu^+}{\kappa_3}&0&\frac{\nu^+\beta_4^2-\nu^-\beta_1^2+\kappa_3}{2\kappa_3}&0\\
0&0&0&0
\end{pmatrix}+\mathcal O(|\xi|),\end{split}\end{equation}
\begin{equation}\label{3.17}\begin{split}P_3(\xi)=&\begin{pmatrix}
0&0&0&0\\
0&1&0&0\\
0&0&0&0\\
0&0&0&0\\
\end{pmatrix}+\mathcal O(|\xi|),\end{split}\end{equation}
and
\begin{equation}\label{3.18}\begin{split}P_4(\xi)=&\begin{pmatrix}
0&0&0&0\\
0&0&0&0\\
0&0&0&0\\
0&0&0&1\\
\end{pmatrix}+\mathcal O(|\xi|).\end{split}\end{equation}
With the help of Lemmas \ref{lemma3.1}--\ref{lemma3.3}, we can have the following proposition which is concerned with
long time properties of $L^2$--norm for the
solution.
\begin{Proposition}[$L^2$--theory]\label{Prop3.4} It holds that
\begin{equation}\|\text{e}^{t\mathcal B_1}\mathcal U(0)\|_{L^2}\lesssim \text{e}^{\theta t}\| {\mathcal U}(0)\|_{L^2},\label{2.30}\end{equation}
for any $t\geq 0$.
\end{Proposition}
\subsection{Spectral analysis for system \eqref{2.2}} We consider the Cauchy problem of \eqref{2.2} with the initial data
\begin{equation}\label{3.20}(\phi^+, \phi^-)\big|_{t=0}=(\Lambda^{-1}{\rm curl}\tilde{u}^{+}_0, \Lambda^{-1}{\rm curl}\tilde{u}^{-}_0)(x).
\end{equation}
From the classic theory of the heat equation, it is clear that the solution $\mathcal V=(\phi^+, \phi^-)^t$ to the IVP \eqref{2.2} and \eqref{3.20} satisfies the following decay
estimates.
\begin{Proposition}[$L^2$--theory]\label{Prop3.5} It holds that
\begin{equation}\|\text{e}^{-\nu^\pm t\Lambda^2}\mathcal V(0)\|_{L^2}\lesssim
\| {\mathcal V}(0)\|_{L^2},\nonumber\end{equation}
for any $t\geq 0$.
\end{Proposition}
We consider the Cauchy problem of \eqref{1.28} with the initial data
\begin{equation}\label{3.21}
\left(\tilde n^{+}, \tilde u^{+}, n^{-},\tilde u^{-}\right)(x, 0)=\left(n_{0}^{+}, \tilde u_{0}^{+}, n_{0}^{-},\tilde u_{0}^{-}\right)(x) \rightarrow(0, \overrightarrow{0}, 0, \overrightarrow{0}), \quad \text { as }|x| \rightarrow+\infty,
\end{equation}
By virtue of the definition of $\varphi^{\pm}$ and $\phi^{\pm}$, and the
fact that the relations
$$\tilde{u}^{\pm}=-\wedge^{-1}\nabla\varphi^{\pm}-\wedge^{-1}\text{div}\phi^{\pm},$$
involve pseudo--differential operators of degree zero, the estimates
in space $H^k(\mathbb R^3)$ for the original function $\tilde{u}^{\pm}$ will
be the same as for $(\varphi^{\pm}, \phi^{\pm})$. Combining Propositions
\ref{Prop3.4} and \ref{Prop3.5}, we have the
following result concerning long time properties for the solution
semigroup $\text{e}^{t\mathcal{A}}$.
\begin{Proposition}\label{Prop3.6} The global solution $\tilde{U}=(\tilde{n}^+,\tilde{u}^+,\tilde{n}^-,\tilde{u}^-)^t$ of the IVP \eqref{1.28} and \eqref{3.21} satisfies
\begin{equation}\| \text{e}^{t\mathcal{A}}\tilde{U}(0)\|_{L^2}\lesssim \text{e}^{\theta t}
\| \tilde{U}(0)\|_{L^2}.\label{3.22}\end{equation}
\end{Proposition}
\bigskip
\section{Nonlinear instability}\label{1section_appendix-2}
We mention that the local existence of strong solutions to a generic compressible
two--fluid model can be established by using the standard iteration arguments as in \cite{Wen1} whose
details are omitted. We can arrive at the following conclusion:
\begin{Proposition}\label{Prop4.1} Assume that the notations and hypotheses in Theorem \ref{3mainth} are in force. For any given initial data $\left(n_0^+,u_0^+,n_0^-,u_0^-\right)\in H^4(\mathbb R^3)$ satisfying $\inf_{x\in\mathbb R^3}\{n_0^\pm+1\}>0$, there exist a $T>0$ and a unique strong solution $(n^+,u^+,n^-,u^-)\in C^0([0,T];H^4(\mathbb R^3))$ to the Cauchy problem \eqref{1.26}--\eqref{1.27}. Moreover, the strong solution satisfies
\begin{equation}\nonumber
\begin{split}
\mathcal E(t)
\leq C(T) \mathcal E(0),
\end{split}
\end{equation}
where $\mathcal E(t)=\left\|\left(n^+,u^+,n^-,u^-\right)(t)\right\|_{H^4}$.
\end{Proposition}
\vspace{4mm}
\textbf{\textit{Proof of Theorem \ref{3mainth}.}} Now we are in a position to prove Theorem \ref{3mainth} by adopting the basic ideas in \cite{Guo1,Jang,Jiang1,Jiang2,WangT}.
In view of Theorem \ref{2mainth}, we can construct a linear solution $\left(\tilde n^+_\vartheta,\tilde u^+_\vartheta,\tilde n^-_\vartheta,\tilde u^-_\vartheta\right)\in C^0([0,\infty);H^4(\mathbb R^3))$ to the linear system \eqref{1.28}. Moreover, without loss of generality, we suppose that
\begin{equation}\mathcal E\left(\tilde n^+_{0,\vartheta},\tilde u^+_{0,\vartheta},\tilde n^-_{0,\vartheta},\tilde u^-_{0,\vartheta}\right)=\left\|\left(\tilde n^+_{0,\vartheta}, \tilde u^+_{0,\vartheta},\tilde n^-_{0,\vartheta}, \tilde u^-_{0,\vartheta}\right)\right\|_{H^4}=1.\nonumber
\end{equation}
Denote $\left(n^{+,\varepsilon}_{0,\vartheta},u^{+,\varepsilon}_{0,\vartheta},n^{-,\varepsilon}_{0,\vartheta},u^{-,\varepsilon}_{0,\vartheta}\right)\overset{\triangle}=\varepsilon \left(\tilde n^+_{0,\vartheta},\tilde u^+_{0,\vartheta},\tilde n^-_{0,\vartheta},\tilde u^-_{0,\vartheta}\right)$. Then, by virtue of Proposition \ref{Prop4.1}, there is a positive constant $\varepsilon_0$ which may be quite small such that for any $\varepsilon<\varepsilon_0$, there is a unique local strong solution $\left(n^{+,\varepsilon}_\vartheta,u^{+,\varepsilon}_\vartheta,n^{-,\varepsilon}_\vartheta,u^{-,\varepsilon}_\vartheta\right)\in C^0([0,T];H^4(\mathbb R^3))$ to the Cauchy problem \eqref{1.26}--\eqref{1.27}, emanating from the initial data $\left(n^{+,\varepsilon}_{0,\vartheta},u^{+,\varepsilon}_{0,\vartheta},n^{-,\varepsilon}_{0,\vartheta},u^{-,\varepsilon}_{0,\vartheta}\right)$ with $\mathcal E\left(n^{+,\varepsilon}_{0,\vartheta},u^{+,\varepsilon}_{0,\vartheta},n^{-,\varepsilon}_{0,\vartheta},u^{-,\varepsilon}_{0,\vartheta}\right)=\varepsilon$.
We fix $\varepsilon_0>0$ which may be small enough, then for any $\varepsilon\in(0,\varepsilon_0)$. Define
$$T^*=\sup\left\{t\in(0,T^{\max})\big|\sup\limits_{\tau\in[0,t]}\mathcal E\left(\left(n^{+,\varepsilon}_\vartheta,u^{+,\varepsilon}_\vartheta,n^{-,\varepsilon}_\vartheta,u^{-,\varepsilon}_\vartheta\right)(\tau)\right)\le \varepsilon_0\right\}$$
and
$$T^{**}=\sup\left\{t\in(0,T^{\max})\big|\sup\limits_{\tau\in[0,t]}\left\|\left(n^{+,\varepsilon}_\vartheta,u^{+,\varepsilon}_\vartheta,n^{-,\varepsilon}_\vartheta,u^{-,\varepsilon}_\vartheta\right)(\tau)\right\|_{L^2}\le \varepsilon\varepsilon_0^{-\frac{1}{3}}\text{e}^{\theta t}\right\}$$
where $T^{\max}$ denotes the maximal time of existence. Obviously, $T^*T^{**}>0$, and furthermore,
\begin{equation}\mathcal E\left(\left(n^{+,\varepsilon}_\vartheta,u^{+,\varepsilon}_\vartheta,n^{-,\varepsilon}_\vartheta,u^{-,\varepsilon}_\vartheta\right)(T^*)\right)= \varepsilon_0\quad \text{if} \quad T^*<\infty,\nonumber
\end{equation}
and
\begin{equation}\left\|\left(n^{+,\varepsilon}_\vartheta,u^{+,\varepsilon}_\vartheta,n^{-,\varepsilon}_\vartheta,u^{-,\varepsilon}_\vartheta\right)(T^{**})\right\|_{L^2}= \varepsilon\varepsilon_0^{-\frac{1}{3}}\text{e}^{\theta T^{**}}\quad \text{if} \quad T^{**}<\infty.\label{4.1}
\end{equation}
Assume $T^*=\infty$, otherwise let $T^\varepsilon=T^*$ and $\delta_0=\varepsilon_0$, we can prove Theorem \ref{3mainth} immediately.
Let
\begin{equation}\label{4.2}T^\varepsilon=\frac{1}{\theta}\ln\frac{2\varepsilon_0}{\varepsilon}\left(\text{i.e.,}\ \varepsilon\text{e}^{\theta T^\varepsilon}=2\varepsilon_0\right)\quad \text{and}\quad \vartheta=\frac{1}{T^\varepsilon}.
\end{equation}
Set $\left(n^+_{d},u^+_{d},n^-_{d},u^-_{d}\right)=\left(n^{+,\varepsilon}_{\vartheta},u^{+,\varepsilon}_{\vartheta},n^{-,\varepsilon}_{\vartheta},u^{-,\varepsilon}_{\vartheta}\right)-\varepsilon(\tilde n^+_\vartheta,\tilde u^+_\vartheta,\tilde n^-_\vartheta,\tilde u^-_\vartheta)$. Noticing that $$\left(n^+_{l},u^+_{l},n^-_{l},u^-_{l}\right)=\varepsilon(\tilde n^+_\vartheta,\tilde u^+_\vartheta,\tilde n^-_\vartheta,\tilde u^-_\vartheta)$$ is also a solution to the linear system \eqref{1.28} with the initial data $\left(n^{+,\varepsilon}_{0,\vartheta},u^{+,\varepsilon}_{0,\vartheta},n^{-,\varepsilon}_{0,\vartheta},u^{-,\varepsilon}_{0,\vartheta}\right)\in H^2(\mathbb R^3)$, it is clear that $\left(n^+_{d},u^+_{d},n^-_{d},u^-_{d}\right)$ is a solution to the system
\begin{equation}\label{4.3}
\left\{\begin{array}{l}
\partial_{t} n^{+}_d+\beta_{1} \operatorname{div} u^{+}_d=\mathbb{F}_{1}, \\
\partial_{t} u^{+}_d+\beta_{1} \nabla n^{+}_d+\beta_{2} \nabla n^{-}_d-v_{1}^{+} \Delta u^{+}_d-v_{2}^{+} \nabla \operatorname{div} u^{+}_d=\mathbb{F}_{2}, \\
\partial_{t} n^{-}_d+\beta_{4} \operatorname{div} u^{-}_d=\mathbb{F}_{3}, \\
\partial_{t} u^{-}_d+\beta_{3} \nabla n^{+}_d+\beta_{4} \nabla n^{-}_d-v_{1}^{-} \Delta u^{-}_d-v_{2}^{-} \nabla \operatorname{div} u^{-}_d=\mathbb{F}_{4},
\end{array}\right.
\end{equation}
subject to the initial condition
\begin{equation}\label{4.4}
\left(n^{+}_d, u^{+}_d, n^{-}_d, u^{-}_d\right)(x, 0)=0,
\end{equation}
where the nonlinear terms are given by
\[
\mathbb{F}_{1}=\alpha_{1} F_{1}\left(\frac{n^{+,\varepsilon}_\vartheta}{\alpha_{1}}, \frac{u^{+,\varepsilon}_\vartheta}{\sqrt{\alpha_{1}}}\right), \quad \mathbb{F}_{2}=\sqrt{\alpha_{1}} F_{2}
\left(\frac{n^{+,\varepsilon}_\vartheta}{\alpha_{1}}, \frac{u^{+,\varepsilon}_\vartheta}{\sqrt{\alpha_{1}}}, \frac{n^{+,\varepsilon}_\vartheta}{\alpha_{4}}, \frac{u^{+,\varepsilon}_\vartheta}{\sqrt{\alpha_{4}}}\right),
\]
and
\[
\mathbb{F}_{3}=\alpha_{4} F_{3}\left(\frac{n^{+,\varepsilon}_\vartheta}{\alpha_{4}}, \frac{u^{+,\varepsilon}_\vartheta}{\sqrt{\alpha_{4}}}\right), \quad \mathbb{F}_{4}=\sqrt{\alpha_{4}} F_{4}\left(\frac{n^{+,\varepsilon}_\vartheta}{\alpha_{1}}, \frac{u^{+,\varepsilon}_\vartheta}{\sqrt{\alpha_{1}}}, \frac{n^{+,\varepsilon}_\vartheta}{\alpha_{4}}, \frac{u^{+,\varepsilon}_\vartheta}{\sqrt{\alpha_{4}}}\right).
\]
\bigskip
Now, we claim that
\begin{equation}\label{4.5} T^\varepsilon=\min\{T^\varepsilon,T^{**}\},
\end{equation}
provided that $\varepsilon_0$ is small enough. Indeed, if $T^{**}=\min\{T^\varepsilon,T^{**}\}$, then $T^{**}<\infty$.
By defining $U=(n^+_d, u^+_d, n^-_d, u^-_d)^t$ and $\mathcal
F=(\mathcal{F}^1,\mathcal{F}^2,\mathcal{F}^3,\mathcal{F}^4)^t$, it holds from Duhamel's principle that
\begin{equation}\nonumber U=\int_0^t\text{e}^{(t-\tau)\mathcal{A}}\mathcal F(\tau)\mathrm{d}\tau.
\end{equation}
By virtue of Proposition \ref{Prop3.6} and \eqref{4.2}, we
have after a complicated but straightforward computation that
\begin{equation}\label{4.6}\begin{split}\|U(T^{**})\|_{L^2}\lesssim& \int_0^{T^{**}}\left\|\text{e}^{(t-\tau)\mathcal{A}}\mathcal F(\tau)\right\|_{L^2}\mathrm{d}\tau\\ \lesssim &\int_0^{T^{**}}\text{e}^{\theta(t-\tau)}\left\|\mathcal F(\tau)\right\|_{L^2}\mathrm{d}\tau
\\ \lesssim &\int_0^{T^{**}}\text{e}^{\theta(t-\tau)}\left(\left\|\left(n^{+,\varepsilon}_{\vartheta},
u^{+,\varepsilon}_{\vartheta},n^{-,\varepsilon}_{\vartheta},u^{-,\varepsilon}_{\vartheta}\right)(\tau)\right\|_{L^2}\left\|\nabla\left(n^{+,\varepsilon}_{\vartheta},
u^{+,\varepsilon}_{\vartheta},n^{-,\varepsilon}_{\vartheta},u^{-,\varepsilon}_{\vartheta}\right)(\tau)\right\|_{W^{1,\infty}}\right.
\\&+\left.\left\|\nabla\left(n^{+,\varepsilon}_{\vartheta},
u^{+,\varepsilon}_{\vartheta},n^{-,\varepsilon}_{\vartheta},u^{-,\varepsilon}_{\vartheta}\right)(\tau)\right\|_{L^4}^2\right)\mathrm{d}\tau
\\ \lesssim &~\int_0^{T^{**}}\text{e}^{\theta(t-\tau)}\varepsilon\varepsilon_0^{-\frac{1}{3}}\text{e}^{\theta \tau}\left(\left(\varepsilon\varepsilon_0^{-\frac{1}{3}}\text{e}^{\theta \tau}\right)^{\frac{1}{6}}\varepsilon_0^{\frac{5}{6}}+\left(\varepsilon\varepsilon_0^{-\frac{1}{3}}\text{e}^{\theta \tau}\right)^{\frac{1}{8}}\varepsilon_0^{\frac{7}{8}}\right)\mathrm{d}\tau
\\ \lesssim &~\varepsilon\varepsilon_0^{-\frac{1}{3}}\text{e}^{\theta T^{**}}\left(\left(\varepsilon\text{e}^{\theta T^{**}}\right)^{\frac{1}{6}}\varepsilon_0^\frac{7}{9}+\left(\varepsilon\text{e}^{\theta T^{**}}\right)^{\frac{1}{8}}\varepsilon_0^\frac{5}{6}\right)\\ \lesssim &~\varepsilon_0^\frac{17}{18}\left(\varepsilon\varepsilon_0^{-\frac{1}{3}}\text{e}^{\theta T^{**}}\right), \end{split}\end{equation}
where, by H\"older's inequality and Sobolev's inequality, we used the facts
$$\|\nabla f\|_{L^\infty}\lesssim \|f\|_{L^2}^\frac{1}{6}\|\nabla^3 f\|_{L^2}^\frac{5}{6},$$
$$\|\nabla^2 f\|_{L^\infty}\lesssim \|f\|_{L^2}^\frac{1}{8}\|\nabla^4 f\|_{L^2}^\frac{7}{8}$$
and
$$\|\nabla f\|_{L^4}\lesssim \|f\|_{L^2}^\frac{9}{16}\|\nabla^4 f\|_{L^2}^\frac{7}{16}.$$
If $\varepsilon_0$ is small enough, by Proposition \ref{Prop2.2} and \eqref{4.6}, we see that
\begin{equation}\nonumber\left\|\left(n^{+,\varepsilon}_\vartheta,u^{+,\varepsilon}_\vartheta,n^{-,\varepsilon}_\vartheta,u^{-,\varepsilon}_\vartheta\right)(T^{**})\right\|_{L^2}
\le C\left( \varepsilon\text{e}^{\theta T^{**}} +\varepsilon_0^\frac{17}{18}\left(\varepsilon\varepsilon_0^{-\frac{1}{3}}\text{e}^{\theta T^{**}}\right)\right)<\varepsilon\varepsilon_0^{-\frac{1}{3}}\text{e}^{\theta T^{**}},
\end{equation}
which contradicts with \eqref{4.1}.
Finally, performing the similar procedure as in \eqref{4.6} and using Proposition \ref{Prop2.2}, we deduce that
\begin{equation}\begin{split}&\left\|\left(n^{+,\varepsilon}_\vartheta,u^{+,\varepsilon}_\vartheta,n^{-,\varepsilon}_\vartheta,u^{-,\varepsilon}_\vartheta\right)(T^{\varepsilon})\right\|_{L^2}
\\ \ge&~\text{e}^{(\theta-\vartheta)T^\varepsilon}\varepsilon \left\|\left(\tilde n^+_{0,\vartheta},\tilde u^+_{0,\vartheta},\tilde n^-_{0,\vartheta},\tilde u^-_{0,\vartheta}\right)\right\|_{L^2}-C\varepsilon_0^\frac{17}{18}\left(\varepsilon\varepsilon_0^{-\frac{1}{3}}\text{e}^{\theta T^{\varepsilon}}\right)\\ \ge&~\frac{2\varepsilon_0m_0}{\text{e}}-C\varepsilon_0^\frac{29}{18}\\ \ge&~\frac{\varepsilon_0m_0}{\text{e}},
\end{split}\end{equation}
if $\varepsilon_0$ is small enough, where $m_0=\left\|\left(\tilde n^+_{0,\vartheta},\tilde u^+_{0,\vartheta},\tilde n^-_{0,\vartheta},\tilde u^-_{0,\vartheta}\right)\right\|_{L^2}$. This completes the proof of Theorem \ref{3mainth} by defining
$\delta_0=\min\left\{\varepsilon_0,\frac{\varepsilon_0m_0}{\text{e}}\right\}$. \hfill$\Box$
\bigskip
{\bf Statement:} No conflict of interest exists in the submission of this
manuscript, and
the datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.
\section*{Acknowledgments}
Yinghui Zhang' research is
partially supported by Guangxi Natural Science Foundation
$\#$2019JJG110003, $\#$2019AC20214 and Key Laboratory of Mathematical and Statistical Model (Guangxi Normal University), Education Department of Guangxi Zhuang Autonomous Region. Lei Yao's research is
partially supported by National Natural Science Foundation of China
$\#$12171390, $\#$11931013, $\#$11571280 and Natural Science Basic Research Plan
for Distinguished Young Scholars in Shaanxi Province of China (Grant
No. 2019JC-26).
\bigskip
|
\section{Introduction}
Systems engineering is a crucial process in the design and management of complex systems \cite{kossiakoff2011systems,friedenthal14practical,hart2015introduction,delligatti2013sysml,huang2007system}. An integral part of designing and architecting engineering systems involves producing adequate documentation to both record and support the development process (e.g. systems specifications \cite{dori2004smart}). The large amount of information to be processed naturally poses a challenge in maintaining document consistency and/or completeness. These shortcomings affect the engineering process on a higher level, particularly in terms of maintaining a `system focus' on the product and avoiding unexpected consequences, which may result in uncontrolled hazards or non-fulfilment of mission objectives during operations \cite{sawyer2005shallow,arellano2015frameworks}. Additionally, a degree of subjectivity or even arbitrariness typically associated with the human factor can also implicitly affect the design choices. In this regard, the initial design stages are especially crucial, when inadequate design choices can lead to costly product modifications after manufacturing \cite{arellano2015frameworks}. In this context, graphical modelling languages and diagrams constitute important tools for systems engineers to reduce ambiguity and maintain consistency in systems documentations \cite{friedenthal14practical,hart2015introduction,delligatti2013sysml}. However, most graphical modelling tools still require significant manual input and are thus subject to similar limitations \cite{friedenthal14practical,hart2015introduction,delligatti2013sysml}.
In engineering applications, Natural Language Processing (NLP) techniques have proven to be successful in tasks such as extracting failure labels from failure reports of engineering systems \cite{oncescu2021sensor} for developing self-supervised classification algorithms for sensor fault identification \cite{oncescu2023}, extracting requirement models from textual specifications \cite{zhao2021natural}, and detecting ambiguities in requirements \cite{ferrari2019nlp}. NLP encompasses a wide range of techniques for a variety of applications, e.g. morphological analysis tools such as lemmatisation, syntactic analysis tools such as sentence breaking, and lexical or relational semantic analysis tools such as word sense disambiguation and relationship extraction \cite{jurafsky2009speech}. Preliminary studies have also been done to demonstrate the feasibility of directly converting natural language textual descriptions to system models \cite{sawyer2005shallow}.
This paper aims to address the challenges in maintaining completeness, consistency, and objectivity in the systems design process by proposing an approach to automatically extract information from textual resources and formulate them in a formal and structured format \cite{sawyer2005shallow,arellano2015frameworks}. Specifically, it is proposed to use NLP techniques to automatically convert textual resources into Object Management Group’s Systems Modelling Language ($\textup{OMG SysML}^{\textup{TM}}$) model elements and diagrams, with a focus on structure and requirement diagrams~\cite{omg19omg}.
Textual resources are especially important in the early phases of system design, when engineers may have to collect and consult a range of unstructured textual materials, alongside existing models and specifications, in order to formulate an initial set of requirements \cite{sawyer2005shallow}. The majority of textual resources available to systems engineers are written in natural language, and include specifications, manuals, product descriptions, technical reports, user stories, customer reviews, etc. SysML is a graphical modelling language that is a subset of the Unified Modelling Language (UML) with extensions catered to systems engineering \cite{friedenthal14practical,delligatti2013sysml,huang2007system,omg19omg,friedenthal08omg,hause06thesysml}. It is able to support the specification, analysis, and design of complex systems such as hardware, software, and facilities \cite{friedenthal14practical}. SysML has also gained importance in recent years as a critical enabler of Model-based Systems Engineering \cite{friedenthal14practical,hart2015introduction}. To the best of the authors’ knowledge, this is the first study that focuses on the automatic generation of SysML diagrams from unstructured natural language text.
The main contributions of the paper are as follows:
\begin{itemize}
\item The design of a versatile automated approach to generate SysML diagrams from natural language text documents;
\item The combined use of NLP techniques and the lexical databases to facilitate and augment the generation of SysML models;
\item The mapping of textual entities (phrases and relationships) to SysML model elements using NLP techniques and heuristic rules.
\end{itemize}
The rest of this paper is structured as follows: a detailed overview of the current state-of-the-art NLP-based strategies for automating systems engineering processes is provided in Section~\ref{section:litreview}. The proposed approach is presented in Section~\ref{section:methods}, and the steps for its implementation are summarised in Section~\ref{section:implementation}. The case studies and experimental procedures are then described in Section~\ref{section:exp}. In Section~\ref{section:results}, the results are discussed.
\section{Review of state-of-the-art NLP-based techniques for automating System Engineering processes}
\label{section:litreview}
SysML diagrams embed a variety of information about the system, for example, requirements, structures, and behaviours \cite{friedenthal14practical,delligatti2013sysml,huang2007system,omg19omg,friedenthal08omg,hause06thesysml}. This study is focused on the automatic generation of requirement diagrams and structure diagrams. In this section, past studies relevant to the use of NLP techniques in systems engineering and the construction of SysML diagrams from natural language text are reviewed. Additionally, techniques from concepts and relationships extraction (ontology learning) are also surveyed~\cite{mejhedmkhinini2020combining,cimiano2005text2onto,drymonas2010unsupervised,velardi2013ontolearn,jiang2010crctol}.
\subsection{NLP in Systems Engineering}
NLP techniques have been commonly used in requirements engineering, where the majority of requirements documents and their sources, such as user reviews, are written in natural language \cite{zhao2021natural,bakar2016extracting,robeer2016automated,johann2017safe}. Past studies have used NLP for requirement elicitation \cite{abad2019supporting}, requirement enrichment \cite{korner2010semantic}, requirements tracing \cite{hayes2003improving}, requirement classification \cite{casamayor2010identification}, requirement improvement \cite{rodriguez2019efficient,ferrari2013mining}, identifying the domain of requirement \cite{thakur2016identifying}, ambiguity detection \cite{ferrari2019nlp}, generating test cases and use cases \cite{carvalho2013test,silva2016test,tiwari2019approach}, and detection of low-quality requirements \cite{ferrari2018detecting}. A more detailed review is done in \cite{zhao2021natural}, which classified the function of NLP in requirement engineering as detection, extraction, classification, modelling, tracing and relating, and search and retrieval. Of specific interest here are studies that have attempted to extract requirements and construct models from natural language text.
For example, in~\cite{loughran2006from}, the authors used NLP and aspect-oriented techniques to mine important concepts in a mix of textual requirement assets such as documentation and manuals. They then used the extracted concepts to construct models of product features~\cite{loughran2006from}. Similarly, in~\cite{ferrari2013mining}, the authors attempted to generate structure feature diagrams after extracting features from product brochures. However, these approaches relied on the user to specify relations between identified concepts and construct the models~\cite{ferrari2013mining}. Additionally, in~\cite{bakar2016extracting}, the authors extracted desired product features from online reviews using a combination of NLP techniques. They first selected the relevant reviews by calculating the term frequency-inverse document frequency (tf-idf) values and performing singular value decomposition~\cite{bakar2016extracting}. Then, features were identified by using a manually specific pattern based on Part-of-speech (PoS) tags. The features were clustered using a modification of Self Organising Map \cite{vesanto2000clustering}. Similarly, in ~\cite{johann2017safe}, the authors used an extensive list of PoS tags and sentence patterns to extract desired app features from app descriptions and reviews. However, these approaches are limited by a heavy dependence on heuristics rules in extracting desired features and do not organise the extracted features into formal models.
In contrast, other studies have attempted to construct systems engineering models from natural language text inputs. For example, in \cite{sreekumar2018extracting}, the authors extracted features from textual requirement specifications documents using a mix of NLP techniques such as noun phrase chunking and tf-idf. The relations were identified using a predefined set of association phrases \cite{sreekumar2018extracting}. The hierarchy between relations was determined by a set of heuristics rules based on features’ first occurrence in the text \cite{sreekumar2018extracting}. In addition, in \cite{al2009natural}, the authors constructed conceptual data models using PoS tags from user requirements written with specific syntax constraints. They employed a mix of manually specified sentence patterns and heuristics rules to extract entities and relationships such as superclasses/subclasses \cite{al2009natural}. Similarly, in \cite{robeer2016automated}, the authors extracted concepts from user stories with specific formats using heuristic rules based on PoS tags. They then populated a manually defined generic conceptual model with the extracted concepts \cite{robeer2016automated}. In \cite{casagrande2014nlp}, the authors used a predefined list of goal-specific keywords and syntactic patterns to extract systems goals. The taxonomy was then constructed using the Heymann algorithm based on the correlations between goals \cite{casagrande2014nlp}. Other studies have also used NLP techniques such as dependency parsing and coreference resolution to identify type dependencies. The hierarchy was then identified using predefined rules or classifiers \cite{thayasivam2011automatically,nguyen2015rule}. However, these approaches are limited by their reliance on heuristics rules \cite{thayasivam2011automatically}. Additionally, they also tended to put strict constraints on the forms of inputs, either specifying complex syntactic rules or restricting inputs to be of a certain format, for example, requirement specification documents \cite{robeer2016automated,sreekumar2018extracting,al2009natural,casagrande2014nlp}. Such constraints limit the flexibility of the approaches to adapt to the variety of textual assets available for a systems engineer and require a significant amount of prior work to produce structured requirement documents for processing.
\subsection{Modelling in SysML and UML}
Unlike requirement diagrams, which are unique to the SysML profile, structure diagrams in SysML bear a close resemblance to UML diagrams such as class diagrams and composite structure diagrams \cite{friedenthal14practical,hause06thesysml,omg19omg}. Thus, past studies aimed at generating UML diagrams from natural language text are also reviewed.
In~\cite{chen2009automatic}, the authors proposed a method to automatically generate UML class diagrams from natural language requirements by employing a Recursive Object Model (ROM) as the intermediate step. The ROM is a graphical language that treats each word and each sentence as objects and assigns semantic relations from a predefined set to these objects using lexical and syntactic rules~\cite{zeng2008recursive}. The ROM model was then traversed and converted to a class diagram using the noun objects and their relations~\cite{zeng2008recursive}. Additionally, in \cite{afreen2011generating}, the authors used PoS tagging and semantic role labelling combined with manually defined rules to generate class diagrams from requirement specifications. However, similar to past studies on generating requirement models, these studies rely on using structure requirement documents of specific formats as inputs and are thus limited in their flexibility \cite{chen2009automatic,afreen2011generating}. In contrast, in \cite{deeptimahanti2011semiautomatic}, the authors proposed a method to automatically generate UML class diagrams from natural language stakeholder requests. They used a set of predefined syntactic rules to decompose the inputs into simple sentences, then parsed the resulting sentences and identified classes and relations using a set of heuristic rules \cite{deeptimahanti2011semiautomatic}. However, the reliance on predefined rules on syntactic features to classify relations also limit the wider applicability of this approach.
\subsection{Ontology learning}
The work proposed in this paper is also inspired by a separate class of literature on generating and enriching ontologies \cite{mejhedmkhinini2020combining,cimiano2005text2onto,drymonas2010unsupervised,velardi2013ontolearn,jiang2010crctol}. Ontology learning refers to the automatic extraction of ‘concepts, relations, attributes, and hierarchies’ from text \cite{gruber1995toward,asim2018survey}. The authors suggest that the techniques in ontology generation could also be used in the automatic generation of SysML diagrams in the proposed application scenario, which assumes the availability of an informative text corpus. Additionally, both modelling approaches employ an object-centred representation of the target domain and use similar constructs such as classes/blocks and relations, allowing the techniques to be transferable from one approach to another \cite{mejhedmkhinini2020combining}.
In particular, in Text2Onto, the authors proposed a method to extract ontology models from natural language text automatically \cite{cimiano2005text2onto}. The concepts, corresponding to classes, were extracted using a mix of algorithms measuring criteria, including relative term frequency, tf-idf, entropy, and C/NC-Value \cite{cimiano2005text2onto}. The C/NC-value is a metric that combines both linguistic and statistical information~\cite{frantzi1998}. The relations were extracted using a mix of WordNet (see section \ref{sec:WordnetTfIdf} and heuristic patterns \cite{cimiano2005text2onto}. Similarly, in the approach termed OntoGain, the authors used the C/NC-Value to extract concepts \cite{drymonas2010unsupervised}. They then constructed taxonomic relations using lexical similarity-based clustering and non-taxonomic relations using a mix of association rules using the predictive apriori algorithm (based on support and confidence), and probability rules based on co-occurrence \cite{drymonas2010unsupervised}. In addition, in \cite{velardi2013ontolearn}, the authors used measures of domain consensus and relevance to extract key terms. They identified definitions and hypernyms for these key terms from both the input corpus and the web using word-class lattices \cite{velardi2013ontolearn}. This step was repeated for the extracted hypernyms and thus iteratively building up the initial noisy ontology, which was then pruned based on the optimal branching algorithm \cite{velardi2013ontolearn}. Moreover, in CRCTOL, the authors used a modified domain relevance measure to extract concepts \cite{jiang2010crctol}. They then used a mix of lexicon-syntactic patterns and term structure to extract the taxonomic relations, and used predefined patterns to extract non-taxonomic relations \cite{jiang2010crctol}. The techniques employed in automatic ontology generation were instructive to our proposed approach. However, these two diagrams serve different purposes, and the techniques in ontology generation need to be adapted to suit systems engineering \cite{mejhedmkhinini2020combining}, e.g. using input corpus as the primary source for generating an informative diagram instead of relying on the web. Another need is to classify the extracted structured information into different types of SysML diagrams \cite{omg19omg}
In this paper, a different approach to generate SysML diagrams end-to-end from unstructured natural language text based on natural language processing techniques and the semantic web is proposed. The proposed approach is different from the literature in the following aspects:
\begin{itemize}
\item Flexible and open-domain. By using techniques from open information extraction, the proposed approach does not require the input document to follow specific syntactic rules or to belong to a specific domain. Instead, it is able to extract required information for plotting from free, unstructured natural language text.
\item Automatic generation of SysML diagrams: The proposed approach takes natural language text as input and directly generates different types of SysML diagrams with no human intervention or intermediate modelling required.
\item Parameter-based specificity tuning: The proposed approach has minimal reliance on heuristic rules and predefined patterns to identify hierarchies. Instead, the approach parameterises the extraction of key entities and relationships and allows the user to adjust the degree of specificity using a set of parameters.
\end{itemize}
\section{Automatic generation of Systems Modelling Language diagrams}
\label{section:methods}
The proposed approach focuses on automatically generating a subset of structure and requirements diagrams in SysML from a corpus of natural language text documents. Specifically, this paper focuses on Block Definition Diagrams (BDD), Internal Block Diagrams (IBD), and Requirement Diagrams (REQ) \cite{omg19omg}. The basic elements of these three diagrams include blocks and their relationships. A block is an elemental modelling construct in SysML that represents both real entities, such as physical objects, and abstract entities, such as concepts~\cite{friedenthal14practical}. Relationships between blocks in SysML can be further classified into categories such as association or generalisation for BDD, and trace or containment for REQ \cite{friedenthal14practical,omg19omg,friedenthal08omg,hause06thesysml}.
To generate these diagram elements, a parallel architecture is proposed to extract two kinds of textual entities from natural language text: 1) key phrases, where a key phrase is defined as a list of one or more words that represent an important entity described in the text. 2) key relationships between the key phrases, where a key relationship is defined as one that links two key phrases. The key phrases and key relationships are used to generate blocks and the relationships between blocks, respectively. The generated blocks and their relationships are then organised into desired SysML diagrams.
The procedures for automatic diagram generation are summarised in Figure~\ref{fig:flowchart} and it consists of six steps:
\begin{enumerate}
\item The first step is the manual selection and upload of a corpus of text documents (Section~\ref{subsection:rawtextselection}).
\item The raw texts selected are then used as inputs for the key noun extraction (Section~\ref{subsection:nounextraction}).
\item The raw texts selected are used also as inputs relation extraction (Section~\ref{subsection:relationshipextraction}).
\item Steps 2 and 3 result in a collection of key nouns and a collection of relations, respectively. The two collections are then used to generate the list of key phrases and key relations (Section~\ref{subsection:keyselection}).
\item The list of key phrases and key relations are subsequently used to generate corresponding SysML model elements and augmented according to the required diagram type (Section~\ref{subsection:mappingandaugmentation}).
\item Finally, the generated SysML model elements are then organised and plotted to the corresponding SysML profile (Section~\ref{subsection:diagramgeneration}).
\end{enumerate}
An illustrative example is provided in Figure~\ref{fig:flowcharteg}. Each step is discussed in details in what follows.
\begin{figure}[ht]
\centering
\begin{subfigure}[b]{0.6\columnwidth}
\centering
\includegraphics[width=0.45\columnwidth]{figures/flowchart.pdf}
\caption{Steps for automatic SysML diagram generation
}
\label{fig:flowchart}
\end{subfigure}
\end{figure}
\begin{figure}[ht]\ContinuedFloat
\begin{subfigure}[b]{\columnwidth}
\centering
\includegraphics[width=0.6\columnwidth]{figures/flowchart_eg.pdf}
\caption{Illustration of automated diagram generation steps using an excerpt from a Windows manual as example
}
\label{fig:flowcharteg}
\end{subfigure}
\caption{Procedures and examples for automatic SysML diagram generation}
\end{figure}
\subsection{Raw texts selection}
\label{subsection:rawtextselection}
The algorithm takes as input a manually selected corpus of text documents. These documents do not need to be structured or follow a predefined writing style like requirement documents, and the proposed approach can be applied to documents typically available to systems engineers, for example manuals \cite{friedenthal14practical,delligatti2013sysml}. Additionally, the corpus should be of sufficient size. As a rule of thumb, for the effective generation of diagrams, the document count should be larger than 100 and the average word count per document should be larger than 500. A corpus can also be obtained by splitting one large document, for example one manual book, into chapters and sections, as illustrated in the case study on a Windows manual in Section~\ref{section:exp}~\cite{windows10fieldguide}.
\subsection{Key nouns extraction}
\label{subsection:nounextraction}
This step of the approach aims to select key nouns, which are nouns that represent important entities and that can serve as constituents of the key phrases. The key nouns are automatically extracted from the corpus by employing preprocessing and term frequency-inverse document frequency (tf-idf) techniques \cite{jurafsky2009speech}.
\subsubsection{Preprocessing}
\label{section:preprocessing}
Preprocessing is needed to remove grammatical features in the text to support noun extraction \cite{jurafsky2009speech}. In this study, a combination of preprocessing methods is applied to the text inputs:
\textit{Tokenisation}: Tokenisation breaks down large textual content, such as paragraphs or documents, into smaller chunks \cite{jurafsky2009speech}. In this study, the Natural Language Toolkit (NLTK) is used to tokenise each input text document into individual sentences and then tokenise each sentence into individual words. The NLTK package uses a pre-trained model for English tokenisation \cite{bird2009natural}.
\textit{Part-of-speech (PoS) tagging}: The NLTK package is used to classify each word of a sentence into different lexical categories. The PoS tagger assigns a PoS tag to each word based on its context sentence \cite{jurafsky2009speech}. Examples of PoS tags include present tense verbs (‘VBP’) and adjectives (‘JJ’). After the PoS tag for each word is assigned, the nouns are used for further processing and words of other types are removed.
\textit{Lemmatisation}: The individual nouns are further reduced to their root forms through the NLTK WordNet Lemmatiser, e.g. from ‘sensors’ to ‘sensor’. The WordNet Lemmatiser employs a combination of predefined rules and dictionary search in lemmatisation \cite{bird2009natural}.
\textit{Stop word removal}: As the last step, all stop words that are still in the text are removed, using the NLTK list of English stop words as a reference \cite{bird2009natural}.
\subsubsection{Evaluation of term frequency-inverse document frequency}
\label{section:evaluationoftfidf}
The term frequency-inverse document frequency (tf-idf) metric is employed to evaluate the domain relevance of each noun after preprocessing \cite{jurafsky2009speech,luhn1957statistical,jones1972statistical}. Note that after the preprocessing steps, the documents contain only nouns. Tf-idf treats each word as unigram, and word orders or document orders are not considered. The tf-idf weighting $w$ of a word $t$, in a document $d$ belonging to a corpus of documents $c$, is a value computed as \cite{jurafsky2009speech,luhn1957statistical,jones1972statistical}
\begin{equation}
w_{t,d} = \textup{tf}_{t,d} \times \textup{idf}_{t,c}
\end{equation}
$\textup{tf}_{t,d}$ is given by
\begin{equation}
\textup{tf}_{t,d} = \textup{log}_{10}(count_{t,d}+1)
\end{equation}
where $count_{t,d}$ is the total count of a word $t$ in a given document $d$, and the inverse document frequency $\textup{idf}_{t,c}$ is given by \cite{jurafsky2009speech,jones1972statistical}
\begin{equation}
\textup{idf}_{t,c} = \textup{log}_{10}( N_c / (1 + \textup{df}_{t,c}) ) + 1
\end{equation}
where $N_c$ is the total number of documents in the corpus $c$, and the document frequency $\textup{df}_{t,c}$ is the number of documents in the corpus $c$ that contain the word $t$ \cite{jurafsky2009speech}. In this study, the tf-idf weighting is normalised to the range $0 \leq w_{t,d} \leq 1$ by dividing by the largest tf-idf weighting in a document. The inclusion of inverse document frequency discriminates against words that are common across the corpus \cite{jurafsky2009speech,jones1972statistical}, which are assumed to be less representative of the document’s content, analogous to a set of domain-specific frequency-dependent stop words. The weighting $w_{t,d}$ obtained for each word of a document is used to identify the collection of key nouns to be retained. In particular, a tf-idf weighting threshold $\sigma_{tf-idf}$ is specified by the user (usually in terms of a user-defined percentile), and nouns with tf-idf weighting higher than the threshold are added to the collection of key nouns.
\subsection{Relationship extraction}
\label{subsection:relationshipextraction}
The input text is first tokenised into sentences using the NLTK package. After sentence tokenisation, open information extraction techniques are used to extract relationships from individual sentences \cite{jurafsky2009speech,mausam2016open}.
\subsubsection{Open Information Extraction (IE)}
In this study, the OpenIE toolbox is used for relationship extraction \cite{christensen2011analysis,pal2016demonyms,saha2017bootstrapping,saha2018open}. The toolbox assigns a confidence value to each extracted relation, and in the proposed approach, a confidence threshold $\sigma_{{relationship}}$ between $0$ and $1$ is manually defined to preliminarily filter the extracted relationships.
For each sentence, the toolbox generates a set of possible relations, where each relation is a list of textual phrases $r = (p_1,p_2,p_3,p_4,...,p_{N_r})$. In an extracted relation $r$, $p_2$ is the relation phrase, $p_1$ and $p_3$ are akin to subject and object phrases, and $p_4$ to $p_{N_r}$ are secondary argument phrases, sometimes appearing in long sentences, where $N_r$ is the number of phrases in an extracted relationship~\cite{mausam2016open}. The OpenIE toolbox is implemented as a combination of four methods \cite{christensen2011analysis,pal2016demonyms,saha2017bootstrapping,saha2018open}, which are briefly explained:
\textit{Semantic Role Labeling (SRL) based IE} SRL consists of detecting semantic arguments and their roles associated with a verb in a sentence. For example, given the sentence `I ordered a cake', SRL identifies `I' and `cake' as arguments for the verb `order', where `I' is the agent and `cake' is the patient. This component of the OpenIE tool is based on the SRL system by \cite{punyakanok2008importance} and \cite{johansson2008effect}.
\textit{Relational noun based IE} Relational noun based IE identifies relations that are mediated by nouns. For example, given the phrase `Rowing Club President James', it extracts the relation (James, be President of, Rowing Club). This is implemented by encoding a predefined set of relational nouns and nominal patterns in the OpenIE tool \cite{pal2016demonyms}.
\textit{Numerical IE} Numerical IE identifies the numerical relations in a sentence. For example, given the sentence `The company has 100,000 employees', it extracts the relation (The company; has number of employees; 100,000). This is implemented by a combination of machine-learned patterns and customisations specific to numerical relations in the OpenIE tool \cite{saha2017bootstrapping}.
\textit{Coordination analyser} This method is used to split conjunctive sentences. It first uses a dependency parser to extract candidate conjuncts, and then score their coherence based on the Berkeley Language Model \cite{pauls2011faster}. Additional linguistic constraints are also imposed to improve selection results \cite{saha2018open}.
\subsection{Key phrases and key relationships selection}
\label{subsection:keyselection}
To obtain the key phrases and key relationships useful for generating SysML model elements, a three-step selection mechanism is used. The first step is to select and refine candidate key phrases using the key nouns obtained in Section~\ref{subsection:nounextraction}. The second step is based on scoring the candidate phrases based on a combination of metrics to obtain key phrases~\cite{jurafsky2009speech,fellbaum98wordnet,WordNet}. The third step is to select key relationships based on the selected key phrases.
\subsubsection{Selecting candidate key phrases}
\label{section:selectcandidatekeyphrase}
For simplicity, only subject and object phrases ($p_1$ and $p_3$) of the extracted relationships are used for identifying the key phrase candidates. The phrases are first preprocessed using procedures described in Section~\ref{section:preprocessing}. This results in tokenised phrases where $p = (t_1,t_2,...,t_{N_p})$, and each phrase comprises only nouns in root forms; $N_p$ refers to the number of nouns in a phrase. Then, an intersection is taken between each processed phrase and the list of key nouns. Nouns in these phrases that do not map to any key nouns are removed. The user is further able to determine the specificity of the key phrases by setting a parameter $L_{phrase}$ for the maximum number of nouns in one key phrase, where $length(p) \leq L_{phrase}$. If the number of nouns in a candidate key phrase exceeds $L_{phrase}$, the nouns are ranked according to tf-idf values and the top $L_{phrase}$ key nouns are kept. This results in a collection of candidate key phrases for further processing.
\subsubsection{Key phrase selection based on tf-idf, WordNet depth, and phrase frequency}\label{sec:WordnetTfIdf}
The candidate key phrases are further selected by using a metric based on the frequency of the phrase, average tf-idf weighting, and average WordNet score~\cite{fellbaum98wordnet,WordNet}.
The count of a phrase, $count_{p,d}$, refers to the total count of a phrase $p$ in the `bag of phrases' of a document $d$, normalised with respect to the most frequent phrase in the document. This is used to account for the importance of each phrase. Only the candidate key phrases that are outputs of Section\ref{section:selectcandidatekeyphrase} are considered.
The WordNet depth $h'$ of a word $t$ refers to its semantic depth assigned by WordNet based on its synsets, and is used to account for semantic meaning when evaluating a phrase~\cite{fellbaum98wordnet,WordNet}. A general term such as `entity' would be assigned a small depth value, whereas a more specific term such as `pancreas' would be assigned a high depth value. The WordNet is a large lexical database of English, where the meanings of each word are represented as a cognitive synonym set called synsets~\cite{fellbaum98wordnet,WordNet}. The LESK algorithm is used to identify the most relevant synset for words in the input document based on the document as context \cite{lesk1986automatic}. For simplicity, it is assumed that each word in one document has only one synset. The WordNet depth of each word is normalised to range $0 \leq h' \leq 1$ by dividing by the largest depth value in a document, and the one complement of the normalised depth $h=1-h'$ is used as the WordNet score.
As both tf-idf weighting $w$ (already introduced in Section~\ref{section:evaluationoftfidf}) and the WordNet score $h$ are assigned to individual words $t$, an average value is taken for each candidate key phrase. Hence, for each candidate key phrase $p$, containing $N_p$ number of nouns, from a document $d$, a score $\lambda$ is computed by using the following formula:
\begin{equation}
\lambda_{p,d} = \sum\limits_{i}^{N_p}w_{t_i,d}/{N_p} + \sum\limits_{i}^{N_p}h_{t_i,d}/{N_p} + \textup{count}_{p,d}
\label{eq:entityscore}
\end{equation}
The first term computes the average tf-idf weighting, the second term computes the average WordNet depth, and the last term refers to the normalised count of the phrase. A score threshold $\sigma_{p}$ can then be specified by the user, where candidates with scores higher than the threshold are selected as key phrases.
\subsubsection{Key relationship selection}
Relationships are selected as key relationships only when both their subject and object phrases are key phrases. This is to prevent open-ended relationships and `floating' blocks that do not connect to any other block in SysML diagrams.
\subsection{Mapping and augmentation}
\label{subsection:mappingandaugmentation}
The selected key phrases and key relationships are \textit{textual}. For plotting SysML diagrams, they need to be first mapped to SysML model elements. The authors focus on the Block Definition Diagrams (BDD), Internal Block Diagrams (IBD), and Requirement Diagrams (REQ) in this study. The key phrases are used to generate blocks $B$, which form a fundamental unit of SysML~\cite{friedenthal14practical}. The textual relationships are first classified into different categories according to the SysML diagram type, and then used to generate the different types of relationships $R$ between blocks. The textual relationships are also used to generate requirements that blocks satisfy.
To make diagrams more complete in both structure and semantics, the textual phrases and relationships are also augmented according to the required SysML diagram type. The augmentations are used to generate additional blocks and relationships for the diagrams.
\subsubsection{Block Definition Diagrams}
\label{section:blockdefinitiondiagram}
Block Definition Diagrams are the most common diagrams used in SysML and are used to define the types of elements of other diagrams \cite{friedenthal14practical,omg19omg}. Blocks in BDD can have many features such as parts, references, and operations, whereas relations in BDD can belong to categories such as associations, generalisations, and dependencies \cite{friedenthal14practical,omg19omg}. For simplicity, the authors choose to focus on the composite association, generalisation, and reference association, and on identifying the operations features of blocks in BDD. The composite association relationship represents a structural composition where one block is part of another block \cite{friedenthal14practical,omg19omg}. On the other hand, generalisation relationship means an inheritance relationship, where one block is the generalisation of the other, more specialised, block \cite{friedenthal14practical,omg19omg}. Additionally, reference association conveys a connection between blocks where one block can access another \cite{friedenthal14practical,omg19omg}.
\textit{Relationship mapping for BDD}
The steps used to map textual relationships to inter-block relationships in BDD are:
\begin{enumerate}
\item Identification of operation: for every subject phrase $p_1$ in the textual relationships, the relation phrase $p_2$ between the subject and its object phrase is considered as the operation of the block $B_{p1}$ defined by $p_1$.
\item Classification based on relation phrase: the textual relationships are first classified according to the meaning of the relation phrase $p_2$ in the relationship. A pre-selected list of WordNet synsets that define composite associations are used, for example `include.v.01'. If the relation phrase’s synset belongs to the predefined list, then the relation is mapped to a composite association with the corresponding hierarchy.
\item Classification based on overlap: the rest of the textual relationships are then classified by the string representations of the subject and object phrases $(p_1,p_3)$. If the string representation of one phrase is contained in the other phrase in the relation (e.g. the phrase ‘prediction model’ includes the single word phrase ‘prediction’ in string representation), then the relationship is mapped to a generalisation relationship, with the block defined by the shorter phrase being the generalisation (the other being specialisation).
\item Classification based on score: the remaining textual relations are then classified according to the score of the subject and the object phrase according to Equation~\ref{eq:entityscore}. If the difference in score is above a user-defined threshold $\sigma_{rel-difference}$, the relation is mapped to a composite relationship where blocks defined by the lower-scored phrase is considered a part of the block defined by the higher-scored phrase.
\item Classification of remaining relations: the remaining textual relations which are not yet classified are then mapped to reference associations.
\end{enumerate}
\textit{Augmentation for BDD}
The steps used to augment the list of textual phrases and relationships for plotting BDDs are:
\begin{enumerate}
\item Identification of top-level phrases: The key phrases whose corresponding blocks do not form sub-blocks of another are first compiled to a set $P$, where a sub-block is defined as the block at the part end of a composite relationship, or at the specialised end of a generalisation relationship. These phrases are used for the next abstraction step.
\item Abstraction: Algorithm~\ref{algo:entityabstraction} is used to iteratively abstract higher level phrases from the top-level phrases based on a per-word score $\gamma$ that is a combination of the tf-idf weighting $w$ and WordNet score $h$ of a word $t$ in a document $d$.
\begin{algorithm}[h]
Identify set of top-level phrases $P$ \\
\SetAlgoLined
\For {\textup{phrase} p \textup{\textbf{in}} P} {
\If{\textup{length of phrase} $N_p >$ \textup{1}}{
Initialise score set $\Gamma$ \\
\For {\textup{noun} t \textup{\textbf{in}} p}{
$\gamma_{t,d} = w_{t,d} + h_{t,d}$
$\Gamma_{p,d} = \Gamma_{p,d} \cup \{\gamma_{t,d}\} $
}
sort($\Gamma_{p,d}$)
$p_{abstract} = p\setminus t$, where $\gamma_{t,d}$ is smallest in $\Gamma_{p,d}$
$P = P \cup \{p_{abstract}\}$
}}
\caption{Phrase Abstraction}
\label{algo:entityabstraction}
\end{algorithm}
Additionally, generalisation relationships are assigned to the blocks defined by the top-level phrases and their abstractions.
\item Relationship augmentation: After the abstraction step, all of the resultant set of top-level phrases are now unigrams. The WordNet is then used to identify hypernym/hyponym (corresponding to generalisation) and holonym/meronym (corresponding to composite association) relationships between these phrases and corresponding relationships are assigned between their blocks~\cite{fellbaum98wordnet,WordNet}.
\item Phrase augmentation: As the final step, the lowest common hypernyms between the phrases that remain at the top level are found using WordNet ~\cite{fellbaum98wordnet,WordNet}. The blocks defined by these remaining top-level phrases are assigned a generalisation relationship with blocks defined by their hypernyms.
\end{enumerate}
\subsubsection{Internal Block Diagrams}
Internal Block Diagrams are used to specify the internal structures of blocks, and display the parts and references of a specific block as well as the connections between its parts and references \cite{friedenthal14practical,omg19omg}. In this study, the parts of a block are defined as the block's composites, and the references of a block are defined as other blocks connected to it through the reference association. A connection means that two blocks are able to access each other in an operational system \cite{friedenthal14practical,omg19omg}. To leverage the connection between IBD diagrams and BDD diagrams, the augmented list of textual relationships and phrases from the BDD are used as inputs to the IBD algorithm.
\textit{Relationship mapping for IBD}
The steps to map textual relationships to inter-block relationships for IBD are:
\begin{enumerate}
\item Parent block generation: The user can first choose to specify a parent block to draw the IBD for. Then, all blocks that are sub-blocks or sub-sub-blocks of the user-specified block are selected. If no parent block is manually selected, IBDs will be drawn for all blocks.
\item Relationship selection: The textual relationships whose subject or object phrases correspond to a sub-block of the user-specified block are selected, except in cases where the phrase is in the lower hierarchy in the textual relationship.
\item Connection classification: For each of the selected textual relationship, if the relation phrase is not empty, then blocks defined by its subject and object phrases are assigned a port connection between each other.
\end{enumerate}
\textit{Augmentation for IBD}
The steps used to augment the list of textual phrases and relationships for plotting IBDs are:
\begin{enumerate}
\item Phrase augmentation: Pairs of key phrases that share a non-empty intersection (i.e. include one or more identical nouns) are identified and are added to the list of phrases for IBD, except when the intersection corresponds to the user-specified block.
\item Relationship augmentation: After the phrase augmentation step, a generalisation relationship is assigned to the block defined by the intersection and the intersection's parent blocks, with the intersection block at the generalised end. The new relationships are added to the list of IBD relationships.
\end{enumerate}
Importantly, the augmentation mechanism used in IBD is different to the abstraction mechanism in BDD. This is done as a redundancy measure to ensure the completeness of the plotted diagrams. For example, a phrase ‘prediction mode’ may be only abstracted to ‘prediction’ in the BDD step. However, in the IBD step, it could be additionally abstracted to ‘model’ if any of the other selected key phrases also include the word ‘model’.
\subsubsection{Requirement Diagram}
Requirement Diagrams are a unique feature of SysML diagrams that represent the requirements of the system and its components \cite{friedenthal14practical,omg19omg}. Relationships relevant to requirements in SysML include containment, trace, derive, refine, satisfy, and verify \cite{friedenthal14practical,omg19omg}. For simplicity, the authors focus on the satisfy relationships and the trace relationships between the requirements, which represents a weak dependency where a change at one end may result in the need to modify the requirement at the other end \cite{friedenthal14practical,omg19omg}.
\textit{Mapping for REQ}
The following steps are used to map textual relationships to requirements and requirement relationships for REQ diagrams.
\begin{enumerate}
\item Requirement identification: For each of the key relationships, if the relation phrase is not empty, then its corresponding raw relationship output from OpenIE is considered a requirement.
\item The requirements are assumed to be satisfied by their subject phrase blocks, and their requirement blocks are named after the subject phrases. There can be multiple requirements with the same subject phrase, and these are grouped together in the same requirement block.
\item Relationship identification: A trace relationship is assigned to requirements whose corresponding relationships share the same subject phrases or object phrases, or have subject phrases equal to object phrases of other relationships, and vice versa.
\end{enumerate}
\textit{Relationship augmentation for REQ}
The WordNet is used to identify any hypernym, hyponym, entailment, and causes relationship between the relation phrases of relationships that have been identified as a requirement ~\cite{fellbaum98wordnet,WordNet}. If a relationship is identified, the corresponding pair of requirements are considered to have a trace relationship
\subsection{Diagram Generation}
\label{subsection:diagramgeneration}
Prior to plotting the diagram, the user can choose to select a parent block for plotting the desired diagram. For IBD, this would be the parent block used to select sub-blocks and sub-sub-blocks. For BDD, all blocks at a lower hierarchy than the parent block (e.g. the parent block's sub-blocks and their sub-blocks, etc.) will be iteratively selected, and the resultant selected blocks and their relationships will be plotted instead of the full diagram comprising all generated blocks. For REQ, all requirements that contain the phrase corresponding to the selected block, and the relationships between these requirements, are extracted for plotting.
An open-source diagram generation tool (PlantUML~\cite{plantuml,plantumlguide}) is used to plot the generated blocks and their relationships. The augmented blocks and relationships are plotted with dotted lines to indicate that these do not directly map to phrases and relationships in the input text documents. The diagram generation tool uses GraphViz as its graphical engine \cite{ellson2002graphviz}.
\section{Steps for the implementation of the proposed approach}
\label{section:implementation}
The following steps are performed for the generation of SysML diagrams using the proposed approach:
\begin{enumerate}
\item Select the corpus of textual materials;
\item Select the document to draw the SysML diagram for;
\item Set threshold values for five hyperparameters
\begin{enumerate}
\item $0<\sigma_{tf-idf}<1$ for key noun selection. This sets the tf-idf threshold for a noun to be considered a key noun.
\item $0<\sigma_{{relationship}}<1$ for relationship extraction. This sets the confidence threshold for a relation extracted by OpenIE and can be used to remove an excessive number of duplicated relationships.
\item $0<\sigma_{{p}}<3$ for key phrase selection. This sets the minimum score for a phrase to be considered a key phrase via the proposed formula in Equation~\ref{eq:entityscore}.
\item $0<\sigma_{{rel-difference}}<3$ for composite relationship classification. This sets the minimum score difference for connected phrases to be mapped to composite relationships.
\item $L_{{phrase}}>0$ for length of phrases. This sets the maximum number of words in a phrase;
\end{enumerate}
\item Select the type of SysML diagram to generate (optional);
\item Select the phrase to define the parent block in the SysML diagram (optional);
\item Run the end-to-end six-steps approach as described in Section~\ref{section:methods} to automatically generate SysML diagrams, which will generate the following outputs in sequence:
\begin{enumerate}
\item List of key nouns;
\item List of extracted relationships;
\item List of key phrases and key relationships;
\item Mapping of key textual relationships to SysML relationships;
\item Required diagrams.
\end{enumerate}
\item Evaluate the quality of the results through the following steps:
\begin{enumerate}
\item Evaluation of key phrase selection;
\item Evaluation of key relation mapping;
\item Evaluation of diagram generation.
\end{enumerate}
\end{enumerate}
\subsection{Quantification of the success of phrase extraction}
To evaluate the success in key phrase extraction, manually selected lists of key phrases based on OpenIE extractions are created for each of the selected documents and used as the ground truth. The automatically extracted lists of phrases are then evaluated by precision and recall. Precision is defined as the percentage of extracted key phrases that match to a phrase in the ground truth. Recall refers to the percentage of ground truth phrases that match to a phrase in the extracted list. As the key relationships are extracted together with the key phrases, the key relationships are not evaluated separately because it is strongly coupled with the key phrases.
\subsection{Quantification of the success of relationship mapping}
To evaluate the effectiveness of the approach in mapping textual relationships to SysML relationships, the ground truth for the type of relationship between blocks defined by the subjects and objects is defined manually. This is used as a benchmark to evaluate the results from the relationship mapping algorithm. The algorithm is evaluated for its accuracy in determining the type of SysML relationship (e.g. composite association, generalisation or reference association) and the relative hierarchy between the blocks defined by the subject and the object phrase in the textual relationship (e.g. whether block defined by the subject phrase is a composite of that defined by the object, or vice versa).
\subsection{Hyperparameter selection}
The following values for the hyperparameters are suggested, and applied to the six case studies investigated:
\begin{enumerate}
\item $\sigma_{tf-idf}=0$. This means that all the nouns in the input text are treated as key nouns. This is done to illustrate the effect of selection in the subsequent steps. However, it is envisioned that the parameter can be adjusted according to the user’s needs;
\item $\sigma_{{relationship}}=0.5$. This is to filter out relationships with low confidence of being valid relationships.
\item $\sigma_{{p}}=0.6$. This implies most phrases are selected as key phrases, to ensure completeness of results.
\item $\sigma_{{rel-difference}}=0.5$. This ensures that only connected phrases with a sufficient score difference are mapped to composite relationships.
\item $L_{{phrase}} = 3$. This sets the maximum length of a phrase to be three nouns.
\end{enumerate}
\section{Experiments}
\label{section:exp}
Six case studies are investigated to evaluate the effectiveness of the proposed approach in automatically generating SysML diagrams from texts. The details of the datasets used, the experimental procedure, the evaluation procedure, and the choice of hyperparameters are discussed in this section. The results from the extraction of key phrases and classification of relationships are validated against manually constructed benchmarks. Additionally, the generated diagrams are also compared against manually designed ground truth diagrams.
It is worth emphasising here that the aim of the proposed approach is not to replace systems engineers but rather to aid them in gaining an overview of the system. Therefore, the goal is not a perfect extraction of phrases or relationships. Rather, it is to provide an overview of the selected document and system through the profiles in SysML diagrams.
\subsection{Datasets}
The applicability of the proposed approach is tested on six case studies to cover different dataset sizes, domains, and writing styles:
1) Patent descriptions from the European Patent Office (EPO)~\cite{epopatent}. Patent files in English are downloaded from the EPO public database in txt format. Each patent includes sections such as patent name, abstract, and patent description. The patent files are first parsed to remove constructs such as HTML tags and figure references. Then, only the patent description section is used as textual input to the algorithm. Each description is treated as a separate document, and 345 separate patent description documents are obtained. The patents cover areas such as data transmission devices, devices for drug admission, and bioinformatics.
2) Windows manual. A digital book named Windows 10 Field Guide in pdf format is used~\cite{windows10fieldguide}. The book is split into 28 sections according to the chapter, and converted to individual txt files with images removed. Each file is treated as a separate document. The book is a manual for the Microsoft operating system Windows 10 and describes its different functions and applications, for example Skype and Maps.
3) UK government report. A series of UK government report on the Future of Manufacturing is downloaded as pdf files ~\cite{ukgovreport}. A total of 37 reports are downloaded, and each of them is further split into different chapters, with cover pages, acknowledgements, content pages and references removed. This results in a total of 284 chapters. Each chapter is converted to txt format and treated as a separate document. The contents cover areas such as technology trends in manufacturing and investment in innovation.
4) App descriptions. App descriptions from the Apple App Store are downloaded from an online public dataset~\cite{appstore}. A total of 346 app descriptions in txt format with a file size larger than 4KB are selected from a dataset of 4075 app descriptions. This is done to ensure that each document has sufficient length. Each description is treated as a separate document. The app descriptions included apps of different genres and included contents such as user reviews and feature descriptions.
5) Research publications from IEEE. 285 papers are downloaded using IEEE Xplore ~\cite{ieeexplore}. The search criteria were to have the words ‘design’ and ‘manipulator’ in the title of the paper and the range was from 2015 to 2021. The papers were downloaded as pdfs and converted to txt files. Each paper is treated as a separate document. The papers covered areas in the design of robotic manipulators, for example mechanics and control.
6) Description of countries on Wikipedia. 193 English Wikipedia pages about countries are parsed from the parent Wikipedia page \textit{Member states of the United Nations}~\cite{wikipedia}. Each page is saved as a separate txt file and all formatting and images were removed. Each page is also treated as a separate document. The descriptions include each country's economy, geography, government, etc.
Details of each dataset are summarised in the table below. Two documents are selected from each of the datasets for demonstration\footnote{The demonstration documents are available at https://github.com/ShaohongZ/NLP-for-Systems-Engineering}. The documents are selected based on their word counts being close to the average word count of the dataset. Additionally, one block is selected for each document as the parent block, and the phrases that correspond to the parent blocks are listed in Table~\ref{table:datasets}. The use of these phrases is explained in Section~\ref{subsection:diagramgeneration}.
\begin{table}[ht!]
\centering
\begin{tabular}{M{6cm}|M{1cm} M{3.5cm} M{2,5cm}}
& Doc. no. & Word count & Selected phrase \\ \hline
\textbf{EPO patent} & & 8370 (average) & \\ \hline
A state estimator & 1 & 8418 & Prediction \\ \hline
A body fluid leakage detection\newline aqueous composition & 2 & 8343 & Starch \\ \hline
\textbf{Windows manual} & & 3027 (average) & \\ \hline
Devices & 3 & 3061 & Display \\ \hline
Maps & 4 & 2859 & Map \\ \hline
\textbf{UK government report} & & 1982 (average) & \\ \hline
Knowledge spillover - knowledge sources manufacturing Part5 & 5 & 1993 & Intangible asset \\ \hline
De-industrialisation and balance-of payments Part3 & 6 & 1958 & Manufacturing \\ \hline
\textbf{App description} & & 559 (average) & \\ \hline
Infinity Blade II & 7 & 559 & Entity \\ \hline
Ebates: Cash Back, Coupons \& Rebate Shopping App & 8 & 559 & Entity \\ \hline
\textbf{Research paper} & & 3414 (average) & \\ \hline
Design of a weight-compensated and coupled tendon-driven articulated long-reach manipulator & 9 & 3430 & Actuator \\ \hline
Multi criteria design of a spherical 3-DoF parallel manipulator for optimal dynamic performance & 10 & 3413 & Optimisation \\ \hline
\textbf{Wikipages} & & 9830 (average) & \\ \hline
Bhutan & 11 & 9867 & Industry \\ \hline
India & 12 & 9811 & Religion \\ \hline
\end{tabular}
\caption{Details of datasets, selected documents, and selected phrases}
\label{table:datasets}
\end{table}
\section{Results and Discussion}
\label{section:results}
\subsection{Key phrase extraction}
The results from the key phrase extraction are detailed in Table~\ref{table:keyentity}. It can be seen that the number of phrases extracted from a document is around the same order of magnitude as the number of unique nouns in the document. The possible reason why the number of phrases tends to be less than the number of nouns, except for one case, is that many of the phrases are in a descriptive sentence or sentences with pronouns. For example, in a descriptive sentence such as `The device is expensive', the candidate phrase `device' will not be considered as a key phrase as no relationship can be extracted from the sentence. Additionally, in sentences that involve pronouns such as `You can turn on the computer' and `It is used to predict motion', the candidate phrase `computer' or `motion' will not be selected because detecting pronoun references is out of the scope of this paper. Requiring all key phrases to be related to at least one other phrase is useful because it avoids the case of `floating blocks' where a given block is not connected to any other block in the SysML diagram. Such floating blocks are not conducive to helping the user understand the dependencies within the system or the hierarchies between the different blocks within the system, and may confuse the user.
\begin{table}[ht!]
\centering
\begin{tabular}{M{2cm}|M{2cm} M{2cm} M{2cm} M{2cm} M{2cm}}
\makecell{Document\\No.} & \makecell{Word\\Count} &\makecell{Unique\\nouns}& \makecell{No. of\\key phrases}& Precision \% & Recall \% \\ \hline
1 & 8418 & 345 & 401 & 80.8 & 76.8 \\ \hline
2 & 8343 & 430 & 351 & 91.5 & 77.3 \\ \hline
3 & 3061 & 214 & 141 & 70.2 & 57.9 \\ \hline
4 & 2859 & 193 & 119 & 66.4 & 49.7 \\ \hline
5 & 1993 & 214 & 157 & 55.4 & 81.3 \\ \hline
6 & 1958 & 223 & 140 & 70.7 & 71.2 \\ \hline
7 & 559 & 122 & 12 & 66.7 & 34.8 \\ \hline
8 & 559 & 105 & 33 & 84.8 & 63.6 \\ \hline
9 & 3430 & 216 & 131 & 61.8 & 54.0 \\ \hline
10 & 3413 & 367 & 211 & 81.0 & 82.6 \\ \hline
11 & 9867 & 1198 & 992 & 77.3 & 72.9 \\ \hline
12 & 9811 & 1224 & 854 & 58.7 & 69.3 \\ \hline
\end{tabular}
\caption{Results from key phrase extraction}
\label{table:keyentity}
\end{table}
As can be seen from Table~\ref{table:keyentity}, the precision and recall rates indicate that the algorithm has successfully extracted key phrases from the document. The algorithm is especially useful with documents that tend to be more carefully written, as can be seen from the higher success rates in extracting phrases from patent descriptions. The failure case of key phrases that are not identified includes those in descriptive sentences, those connected to pronouns, and incomplete sentences. Incomplete sentences are probably the main reason for the low recall in extracting key phrases from the Document No.7 because the descriptions contained many broken sentences such as `The best game ever'. Additionally, the algorithm is able to achieve relatively high precision, across most of the document types, validating the proposed algorithm for selecting key phrases. The precision and recall rates can be further improved by adjusting the hyperparameters. The success in extracting key phrases contributes to helping the user gain an initial understanding of the system.
\begin{table}[htp!]
\centering
\begin{tabular}{M{2cm}|M{3cm} M{3cm} M{3cm}}
\makecell{Document\\No.} & \makecell{No. of\\sentences} &\makecell{No. of\\extracted relations}& \makecell{No. of\\key relations} \\ \hline
1 & 409 & 670 & 425 \\ \hline
2 & 463 & 681 & 309 \\ \hline
3 & 186 & 257 & 102 \\ \hline
4 & 174 & 289 & 107 \\ \hline
5 & 92 & 286 & 137 \\ \hline
6 & 98 & 158 & 100 \\ \hline
7 & 20 & 21 & 7 \\ \hline
8 & 34 & 43 & 25 \\ \hline
9 & 301 & 251 & 115 \\ \hline
10 & 217 & 207 & 150 \\ \hline
11 & 508 & 1179 & 799 \\ \hline
12 & 443 & 1106 & 689 \\ \hline
\end{tabular}
\caption{Results from key relationship extraction}
\label{table:keyrelationshipextraction}
\end{table}
\subsection{Key relationship extraction}
The results from key relationship extraction are detailed in Table~\ref{table:keyrelationshipextraction}. As can be seen from the table, the OpenIE tool extracts multiple semantic relations from the same sentence even after applying a relation likelihood threshold of 0.5 on the OpenIE tool. This is likely attributed to the combination of different methods used in OpenIE \cite{christensen2011analysis,pal2016demonyms,saha2017bootstrapping,saha2018open}. This is helpful in capturing the complete semantic meaning of each sentence. It can also be seen that the key relation selection algorithm is effective in removing from 27\% to 65\% of the extracted relations. This can also be manually adjusted by tuning the parameters $sigma_{tf-idf}$, $\sigma_{p}$, and $\sigma_{{relationship}}$. The use of multiple parameters is aimed at giving the user more control over the completeness of the generated SysML diagrams.
\subsection{Relationship mapping}
The results from the mapping of key relationships for BDD diagram are detailed in Table~\ref{table:keyrelationshipclassification}. A majority of the extracted textual relationships are mapped to either composite associations or reference associations. This is as expected because a textual relationship is only mapped to generalisation if one phrase contains another, which is assumed to be a stricter criterion. At a difference threshold of 0.2, the number of composite and reference associations are around the same order of magnitude, with the number of composite associations generally larger.
\begin{table}[ht!]
\centering
\begin{tabular}{M{2.2cm}|M{2.2cm} M{2.2cm} M{2.2cm} M{2.2cm}}
\makecell{Document\\No.} & Generalisation &\makecell{Composite\\association}& \makecell{Reference\\association} & \makecell{Classification\\accuracy \%} \\ \hline
1 & 26 & 168 & 226 & 75.8 \\ \hline
2 & 11 & 92 & 204 & 70.9 \\ \hline
3 & 3 & 31 & 68 & 64.7 \\ \hline
4 & 6 & 45 & 55 & 70.1 \\ \hline
5 & 4 & 39 & 93 & 84.7 \\ \hline
6 & 1 & 26 & 72 & 71.0 \\ \hline
7 & 0 & 2 & 5 & 57.1 \\ \hline
8 & 0 & 11 & 13 & 68.0 \\ \hline
9 & 7 & 34 & 73 & 62.6 \\ \hline
10 & 1 & 33 & 115 & 73.3 \\ \hline
11 & 12 & 209 & 577 & 73.0 \\ \hline
12 & 5 & 187 & 497 & 73.6 \\ \hline
\end{tabular}
\caption{Results from key relationship classification}
\label{table:keyrelationshipclassification}
\end{table}
Having a sufficient number of composite relationships and generalisation relationships is useful because they also serve as taxonomic relationships which provide structures to the SysML diagrams. This helps the algorithm to identify the hierarchy between blocks and facilitates the graphic layout algorithm. Having such a hierarchical structure is useful in helping the user understand the overall structure of the system. Additionally, the algorithm can also be seen to have achieved high accuracy in mapping the correct relationship types. The outlier in the app description dataset is potentially due to the small number of key relations for the document. This contributes to creating a more objective starting point in SysML diagrams for users.
\subsection{Augmentation}
The results from augmentation for BDD diagrams are detailed in Table~\ref{table:augmentation}. As can be seen, the abstraction step adds a high number of augmented relationships and phrases. It helps provide a higher-level overview of the entire system and provide more structure to the SysML diagrams through taxonomic generalisation relationships. This also enables the user to quickly identify higher-level phrases within the document to plot other diagrams, for example package diagrams. Additionally, abstraction is based on the extracted phrases, ensuring that the abstracted phrases and relationships are relevant to the original corpus. This provides users with a more complete starting point to understand and design the complex system.
\begin{table}[ht!]
\centering
\begin{tabular}{M{2cm}|M{2cm} M{2cm} M{2cm} M{2cm} M{2cm}}
\makecell{Document\\No.} & \makecell{No. of\\abstraction\\relations} &\makecell{No. of\\abstracted\\phrases}&\makecell{No. of\\augmented\\top-level\\relationships}& \makecell{No. of\\lowest\\common\\hypernym\\relationships} & \makecell{No. of\\lowest\\common\\hypernyms} \\ \hline
1 & 246 & 125 & 1 & 276 & 96 \\ \hline
2 & 224 & 145 & 3 & 569 & 168 \\ \hline
3 & 98 & 61 & 1 & 159 & 58 \\ \hline
4 & 66 & 45 & 0 & 112 & 53 \\ \hline
5 & 104 & 62 & 1 & 235 & 85 \\ \hline
6 & 107 & 69 & 0 & 249 & 80 \\ \hline
7 & 10 & 10 & 0 & 14 & 9 \\ \hline
8 & 19 & 16 & 0 & 43 & 23 \\ \hline
9 & 83 & 51 & 2 & 121 & 51 \\ \hline
10 & 182 & 117 & 0 & 316 & 104 \\ \hline
11 & 736 & 462 & 26 & 1678 & 403 \\ \hline
12 & 595 & 376 & 37 & 1672 & 399 \\ \hline
\end{tabular}
\caption{Results from phrase and relationship augmentation}
\label{table:augmentation}
\end{table}
It can be seen that only a small number of relationships are identified between the top-level unigram phrases after the abstraction step, potentially due to the criteria used as only WordNet hypernyms/hyponyms and meronyms/holonyms relationship are identified~\cite{fellbaum98wordnet,WordNet}. It can also be seen that a large number of common hypernyms and hypernym relationships are augmented. The common hypernyms are able to provide another level of abstraction to the selected phrases based on semantics that are different from the abstraction based on lexical terms, helping the system engineer to gain a more complete high-level picture of the system and its different abstract categories in addition to contributing more structure to the SysML diagrams.
\subsection{Requirement identification}
The results from requirement identification are detailed in Table~\ref{table:requirement}. The number of requirements identified is the same as the number of key relationships, which is expected because the requirements in this study are derived from relationships. This contributes to a more complete set of candidate requirements for the user to consider. Additionally, it can also be seen that a large number of relationships between requirements are identified and augmented through the algorithm. This leads to most of the requirements being connected using the augmented relationships, which is useful because even though relationships between phrases may be found through using sentences explicitly stated in the text, relationships between requirements, which are derived from sentences, can be difficult to identify. The proposed algorithm is able to suggest candidate relationships between requirements, allowing the user to build on the knowledge to classify relationships between requirements further. Additionally, the relationships also provide structure to the requirement diagrams by connecting the requirements and bringing related requirements to the same cluster.
\begin{table}[ht!]
\centering
\begin{tabular}{M{2cm}|M{4cm} M{4cm}}
\makecell{Document\\No.} & \makecell{No. of\\requirements} &\makecell{No. of relationships\\between requirements} \\ \hline
1 & 425 & 636 \\ \hline
2 & 309 & 450 \\ \hline
3 & 102 & 84 \\ \hline
4 & 107 & 52 \\ \hline
5 & 137 & 85 \\ \hline
6 & 100 & 56 \\ \hline
7 & 7 & 1 \\ \hline
8 & 25 & 2 \\ \hline
9 & 115 & 149 \\ \hline
10 & 150 & 161 \\ \hline
11 & 799 & 469 \\ \hline
12 & 689 & 575 \\ \hline
\end{tabular}
\caption{Results from requirement identification and augmentation}
\label{table:requirement}
\end{table}
\subsection{BDD diagram generation}
A few examples of BDD diagrams generated in the case studies are illustrated in Figure \ref{fig:BDD_diagrams} below. It can be seen that the augmentation steps were useful in building the structure of the BDD diagrams. For example, many blocks have corresponding phrases that share the same abstracted phrase, which defines the parent block that clusters these blocks together. This can also be tuned by the choice of specificity $L_{phrase}$. This is useful because these phrases are not explicitly related in the original document, yet they are connected via the parent blocks in the generated diagrams. For example, in Figure \ref{figure:bdd_display}, the block defined by the abstracted phrase `display' clusters the sub-blocks such as `display option' and `taskbar display' into one branch. In Figure \ref{figure:bdd_prediction}, the block defined by the abstracted phrase `prediction' clusters multiple sub-blocks into one branch. Additionally, blocks corresponding to top-level unigram phrases can also be connected to the same block defined by their common hypernyms such as `entity' in Figure \ref{figure:bdd_entity}, where `entity' is a high-level common hypernym for many synsets in the WordNet ~\cite{fellbaum98wordnet,WordNet}.
\newpage
\begin{figure}[ht!]
\centering
\begin{subfigure}[b]{\columnwidth}
\centering
\includegraphics[width=0.95\columnwidth]{figures/prediction_BDD.pdf}
\caption{EPO patent dataset, selected phrase = `prediction'
}
\label{figure:bdd_prediction}
\end{subfigure}
\begin{subfigure}[b]{\columnwidth}
\centering
\includegraphics[width=0.95\columnwidth]{figures/display_BDD.pdf}
\caption{Windows manual dataset, selected phrase = `display'}
\label{figure:bdd_display}
\end{subfigure}
\end{figure}
\newpage
\begin{figure}[ht!]\ContinuedFloat
\centering
\begin{subfigure}[b]{0.6\columnwidth}
\centering
\includegraphics[width=0.6\columnwidth]{figures/entity_BDD.pdf}
\caption{App description dataset, selected phrase = `entity'}
\label{figure:bdd_entity}
\end{subfigure}
\begin{subfigure}[b]{0.25\columnwidth}
\centering
\includegraphics[width=\columnwidth]{figures/BDD_legend.pdf}
\end{subfigure}
\caption{Automatically generated BDD diagrams}
\label{fig:BDD_diagrams}
\end{figure}
These high-level phrases are augmented based on semantics, which ensures that they are relevant to the document, in addition to providing a hierarchical structure that aids user understanding. Without the augmented phrases, these extracted key phrases may become open-ended, which makes it harder for the user to understand the structure of the system from the SysML diagrams. Additionally, the two levels of abstraction capabilities, both lexical and semantic, are useful to help the systems engineer in gaining an overview of the higher-level abstract concepts in the system that are not explicitly stated in the text, and offer suggestions for candidate high-level phrases and their corresponding blocks to the systems engineers.
It can also be seen that there is a mix of extracted relationships and augmented relationships in the diagram, which implies that both steps are useful for the construction of the SysML diagrams. By combining the taxonomy derived from both generalisation and composite association relationships and both extracted and augmented relationships, the algorithm ensures a strong structure in the generated diagrams.
\subsection{IBD diagram generation}
The IBD diagram specifies the connection between different sub-blocks of a given parent block. An example of the IBD diagram is illustrated in Figure \ref{figure:ibd}. It is assumed that two blocks are connected with ports if an action phrase exists in the relationship whose subject and object phrases define these two blocks. The presence of action phrases and corresponding ports is useful to show the interaction between different blocks, in addition to their hierarchical relationship described by the BDD diagram. Additionally, it can be seen that there are more blocks in the IBD than in the BDD diagram, because the IBD diagram uses a different algorithm to identify sub-blocks. This acts as a redundancy measure for the identification of sub-blocks that are not captured by the BDD algorithm, to help generate a more complete overview of the system. Additionally, reference relationships with blocks that are outside the user-specified block are also included in the IBD to highlight potential interactions and communications with external blocks.
\subsection{REQ diagram generation}
The REQ diagram specifies the requirements satisfied by the blocks and the relationships between different requirements. An example of the REQ diagram is illustrated in Figure \ref{figure:req}. It can be seen that some of the requirements share the same name and. are satisfied by the same block, implying that the algorithm has successfully clustered them together. Additionally, most of the requirements are connected to each other via augmented trace relationships. These can serve as candidates for more specific types of relationships for the user. The relationships also help create a hierarchical structure in the requirement diagram, where the most connected requirement is assumed to be placed higher in the hierarchy. Moreover, because the requirements are derived from relationships extracted from individual sentences, they are mostly functional requirements. The extraction of non-functional requirements is out of the scope of this study. By presenting the extracted relationships as requirements to the user, the algorithm is also able to provide more sentence context to help the users understand the phrases and relationships presented in BDD and IBD diagrams.
From the generated figures, it can be seen that the algorithm is able to generate structured SysML diagrams that can potentially aid engineers in designing and architecting complex systems, alongside any existing diagrams and documentations they already possess. By defining the steps and procedures in extracting key phrases and classifying relationships, the approach provides a standardised, and arguably objective starting point for the user to understand and design different systems. Additionally, the approach also provides a degree of versatility in allowing the user to freely choose and mix the textual materials to upload, and to adjust the multiple parameters to achieve the desired specificity.
\newpage
\newgeometry{top=1cm}
\begin{figure}[p]
\centering
\begin{subfigure}[b]{\columnwidth}
\centering
\includegraphics[angle=270,width=5.8cm]{figures/actuator_IBD.pdf}
\end{subfigure}
\begin{subfigure}[b]{\columnwidth}
\centering
\includegraphics[angle=270,width=3cm]{figures/IBD_legend.pdf}
\end{subfigure}
\caption{IBD diagram for research paper dataset, selected phrase = `actuator'}
\label{figure:ibd}
\end{figure}
\restoregeometry
\newpage
\newgeometry{top=1cm}
\begin{figure}[p]
\centering
\begin{subfigure}[b]{\columnwidth}
\centering
\includegraphics[angle=270,width=4.5cm]{figures/display_RQ.pdf}
\end{subfigure}
\begin{subfigure}[b]{\columnwidth}
\centering
\includegraphics[angle=270,width=0.8cm]{figures/REQ_legend.pdf}
\end{subfigure}
\caption{REQ diagram for Windows manual dataset, selected phrase = `display'}
\label{figure:req}
\end{figure}
\restoregeometry
\section{Conclusions}
\label{section:conclusions}
In this study, an approach to automatically generate SysML diagrams end-to-end directly from unstructured natural language text was proposed. This approach consists of six steps, and it leverages open-access tools such as NLTK, OpenIE, and PlantUML. The combined use of natural language processing techniques and the WordNet (available in NLTK) with the OpenIE toolbox was effective in allowing the extraction of key textual relationships and phrases, and the mapping and augmentation of these phrases and relationships to SysML model elements. The results are successfully leveraged to plot different structured SysML diagrams, by using the PlantUML tool. This approach requires five hyperparameters to be specified by the user, and it is open-domain. The versatility of the approach was demonstrated through the use of six case studies from different domains and using different writing styles. The results obtained with the proposed approach were validated against manually extracted results, which achieved high recall and precision in key phrase extraction, high accuracy in relationship classification, and success in generation of high quality SysML diagrams.
The proposed approach would benefit systems engineers in gaining a detailed graphical overview of the system dependencies at the initial design phase, since it provides a standardised, comprehensive and automated starting point from textual resources (e.g., specifications, manuals, technical reports, maintenance reports). This will support systems engineers in understanding complex systems and in quickly assessing and designing the SysML diagrams for these systems. As far as the authors are aware, this study is the first attempt at directly generating SysML diagrams from natural language text.
Future work will focus on developing more efficient and accurate techniques for extracting phrases and relationships, and for plotting diagrams.
\section*{Acknowledgement}
The authors would like to thank Prof. Janet Pierrehumbert (University of Oxford) and Prof. Youssef Marzouk (MIT) for the helpful comments and insights at early stages of this project. Part of this work was supported by the Department of Engineering Science at the University of Oxford through the Engineering Undergraduate Research Opportunities (EUROP) in 2019 (recipient SZ, supervisors AC and AS), and by Balliol College, University of Oxford through the Career Development Fellowship in Engineering Science of AC and a research bursary to SZ.
\section{APPENDIX}
|
\section{Introduction}
We suppose that stable hot regions can be formed in the early Universe. This hypothesis was put forward on the basis of the cosmic X-ray observations and IR background \cite{kashlinsky2019electromagnetic}.
The cluster of primordial black holes (PBH) can
be responsible for such regions.
Formation of PBH clusters and their possible observational effects are now of special interest \cite{Rubin:2000dq,Rubin:2001yw,Khlopov:2004sc,Ding:2019tjk,Matsubara:2019qzv,Young:2019gfc,Kawasaki:2021zir,Inman:2019wvr,Afshordi:2003zb,Jedamzik:2020ypm,DeLuca:2020jug,Pilipenko:2022emp},
but we do not constrain possibility of such regions appearance by PBH clusters only.
PBH and their cluster formation can be the consequence of existence and breaking of some new symmetry in quantum field theory \cite{Rubin:2000dq,konoplich1998formation,deng2018cmb}.
PBH cluster can be the seed of a quasar or galaxy formation \cite{Dokuchaev:2004kr, Dokuchaev:2008hz,Khromykh:2019yyx}. Here we consider the matter trapped in this region, which can be protogalaxy or exist separately. So the prerequisites for the task in question are the regions decoupled from Hubble flow (and virialized) containing primordial plasma. Plasma must flow out to the surroundings by diffusion in the CMB field. If the region is big enough, it can survive to the present time, as it was obtained for the antimatter domain \cite{Khlopov:1998uy, Khlopov:2000as}.
The region of the size $\sim 1$ pc spreads over ambient matter after recombination ($z\approx 1000$), when the structure forms already (see, e.g., Eq.(12) from \cite{Khlopov:1998uy}). By this time, heavy chemical elements, as we show, have time to be formed in a wide range of considered model parameters, therefore areas contaminated with heavy elements can be expected to exist even if the matter had spread outside primordial region.
Moreover, cooling by conventional thermal (gamma-ray) radiation is ineffective for big regions. The escaping time of photons from the region interior (thermal time scale) at the taken parameters (given below) exceeds the modern age of the Universe. The matter inside the area can be additionally heated with respect to the surrounding one during its formation due to domain wall kinetic energy in the case of the respective mechanism of PBH cluster formation \cite{Rubin:2000dq,Rubin:2001yw,Khlopov:2004sc} including Higgs field \cite{Belotsky:2017puc,Belotsky:2017txw}.
We consider the chemical composition of such possible hot regions, whatever their origin is. The thermal evolution of such regions involves many factors.
The matter inside areas can be heated or cooled by various processes acting at the same time. These processes include the neutrino cooling \cite{belotsky2020neutrino,belotsky2020neutrino2}, inelastic reactions between elementary particles and nuclei, the radiation from star-forming hot plasma \cite{kashlinsky2019electromagnetic},
the gravitational dynamics of the system, the shock waves and diffusion of matter during the region formation \cite{belotsky2017local, axioms9020071}, energy transfer from collapsing walls mentioned above \cite{konoplich1998formation,deng2018cmb,Dokuchaev:2004kr,berezin1983thin,Davoudiasl:2021ijv}, accretion \cite{kashlinsky2019electromagnetic,axioms9020071} and the Hawking evaporation \cite{axioms9020071, dolgov1993baryon,dolgov2018massive}.
We focus here on the pure effect of inelastic reactions between elementary particles and nuclei. They may play a dominant role within a wide range of the region parameters which are specified below. We have shown earlier \cite{belotsky2020neutrino,belotsky2020neutrino2}, that neutrino emission can be decisive in the temperature evolution of such regions at the first stage. Here we extend consideration by involving reactions with the lightest element formation.
We use the results obtained in \cite{belotsky2020neutrino, belotsky2019clusters,Khromykh:2019yyx, Dokuchaev:2004kr, Dokuchaev:2008hz}, where the mass of the detached region was supposed to range $10^4-10^8 M_{\odot}$\footnote{These values have been of interest since they can provide a seed for supermassive black holes and galaxies. We do not relate the amount of PBHs with dark matter, which is strongly constrained in dependence on PBH mass value \cite{Carr:2021bzv}. Abundances and masses of PBHs inside clusters as well as of
clusters themselves are assumed to be proper ones.}. The following are the most important starting parameters: the area has a radius of $R\sim 1$ pc, a mass of $10^4 \, M_{\odot}$, and an initial temperature interval $T_0\sim 1\text{ keV}- 10$ MeV.
The goal of this work is to investigate certain reaction networks, which define the evolution of temperature and chemical composition of the regions in the early Universe.
Light element abundance ratios ($n_d/n_B$, $n_{^3He}/n_B$ and $n_{^4He}/n_B$) are finally obtained, heavier element production is discussed.
Hypothesis on existence of the regions discussed can be supported by the evidences of cosmic infrared and X-ray background correlations \cite{kashlinsky2019electromagnetic}, anomalous star existence \cite{2018AJ....156..113B, Dolgov:2019ncq}, and can be probed in direct searches for large areas with abnormal chemical composition in future. Also, such sources of high temperature radiation at the pre-recombination stage can give specific observed patterns of CMB temperature variations ($\Delta$T/T) \cite{Grachev:2010tn} because it is determined by the interaction of these fluctuations in matter density with the CMB during the Universe's expansion and cooling, which are not applicable to small scales
Section \ref{Nucl} is dedicated to the discussion on the main nuclear reactions. Subsection \ref{temp_sec} contains the information about the region temperature, subsection \ref{pn_sec} -- about proton and neutron abundances, \ref{dhe_sec} -- deuterium and helium-3, \ref{hehe_sec} -- about abundances of helium-4 and heavier elements. A closing overview of the research is provided in section \ref{concl}. We also include some useful information on reaction rates and cross-sections in appendix \ref{reac_app}.
\section{Nucleosynthesis}
\label{Nucl}
Consider the reaction between two nuclei 1 and 2. The reaction rate is proportional to the mean lifetime $\tau$ of the nuclear species in the stellar plasma. The number density change rate of nucleus 1 caused by reactions with nucleus 2 can
be expressed as
\cite{burbidge1957synthesis,clayton1983principles}
\begin{equation}\label{N122}
\Big(\frac{dn_1}{dt}\Big)_2=-(1+\delta_{12})r_{12}=-(1+\delta_{12})\frac{n_1n_2\langle \sigma v\rangle_{12}}{(1+\delta_{12})}=-n_1n_2\langle \sigma v\rangle_{12}.
\end{equation}
Here $r_{12}$ is the rate of interaction, $\delta_{12}$ is the Kronecker symbol equals one if 1 = 2 and zero if 1 $\neq$ 2, $n_1$ and $n_2$ are the number densities of nuclei of type 1 and type 2 (having the atomic numbers $Z_1$ and $Z_2$, as well as the mass numbers $A_1$ and $A_2$), and $\langle \sigma v\rangle_{12}$ represents the product of the reaction cross section and the interacting nuclei's relative velocity $v$. The case of identical initial nuclei is taken into account by the presence of the Kronecker symbol.
We will look at how the neutrons, protons, $^2$H, $^3$He and $^4$He abundances change over time due to the reactions of mostly proton-proton chain. The n + p and p + p reaction produces $^2$H, which is then destroyed by the d + p and d + $\gamma$ reactions, whereas the d + p reaction produces $^3$He, which is then destroyed by the $^3$He + $^3$He reaction producing $^4$He. We consider neutrinos to be able to leave the region freely and therefore cool it down. The essential reactions of light elements and neutrinos produced are the following:
\begin{align}
e^- + p &\longrightarrow n + \nu_e, \label{eq_ep} &&\\
e^+ + n &\longrightarrow p + \bar\nu_e, \label{eq_en} &&\\
e^+ + e^- &\longrightarrow \nu_{e,\mu,\tau}+\bar{\nu}_{e,\mu,\tau}, \label{eq_ee} &&\\
n &\longrightarrow p+e^- + \bar\nu_e, \label{n_eq}&&\\
p + p &\longrightarrow D + e^{+} +\nu_{e} + 1.44 \; \text{MeV}, &&\\
D + \gamma &\longrightarrow p + n - 2.22\; \text{MeV}, \label{dg_eq} &&\\
n+ p &\longrightarrow D + \gamma + 2.22 \; \text{MeV}, \label{npd_eq}&&\\
D + p &\longrightarrow \textsuperscript{3}He + \gamma + 5.493 \; \text{MeV}, &&\\
^{3}He + \textsuperscript{3}He &\longrightarrow \textsuperscript{4}He + 2 p + 12.861\; \text{MeV}. \label{He4_eq}
\end{align}
We neglected energy releases of less than 1 MeV.
The initial number densities are approximately described as
\begin{eqnarray}
n_p=\frac{n_B}{1+\exp\left(-\frac{\Delta m}{T_0}\right)},\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; n_n=n_p(T_0)\exp\left(-\frac{\Delta m}{T_0}\right),\label{np_eq}\\
n_{e^-}=n_e^{eq}(T_0)\exp\left(-\frac{m_e}{T_0}\right)+\Delta n_e,\;\;\;\;\;\;\;\;\;\;n_{e^+}= n_e^{eq}(T_0) \exp\left(-\frac{m_e}{T_0}\right),\label{ne}\\
n_B\equiv n_p+n_n=g_B\, \eta n_{\gamma}(T_0),\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \quad \Delta n_e\equiv n_{e^-}-n_{e^+}=n_p.
\label{nd}
\end{eqnarray}
\\
Here $\eta=n_B/n_{\gamma}\approx 0.6 \times 10^{-9}$ is the baryon to photon relation in the modern universe, $g_B\sim 1$ is the correction factors of that relation due to entropy re-distribution, $n_{\gamma}(T)=\frac{2\zeta(3)}{\pi^2}T^3$ and $n_e^{eq}(T)=\frac{3\zeta(3)}{2\pi^2}T^3$ are the equilibrium photon and electron number densities respectively, $\Delta m=m_n-m_p=1.2$ MeV. The forms of Equations \eqref{np_eq} and \eqref{ne} for number densities are chosen to fit their asymptotics in the case of thermodynamic equilibrium.
We consider all densities to be independent on space coordinates within the region. The equations \eqref{ne} are also used to calculate electron and positron current number densities with $T$ instead of $T_0$ and total electric charge instead of $n_p$ inside of $\Delta n_e$
The rates per unit volume, $\gamma_i\equiv \Gamma_i/V$, for reactions listed above are respectively
\begin{eqnarray}
\label{gamma}
\gamma_{ep}=n_{e^-}n_p\langle \sigma v\rangle_{ep},\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\gamma_{en}=n_{e^+}n_n\langle \sigma v\rangle_{en},\\
\gamma_{ee}=n_{e^-}n_{e^+}\langle \sigma v\rangle_{ee},\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\gamma_n=\frac{n_n}{\tau_n},\\
\gamma_{pp}=\frac{n_{p}^2}{2}\langle \sigma v\rangle_{pp},\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\gamma_{\gamma d}=n_{\gamma}n_{d}\langle \sigma v\rangle_{\gamma d},\\
\gamma_{np}=n_{n}n_p\langle \sigma v\rangle_{np},\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\gamma_{dp}=n_{d}n_p\langle \sigma v\rangle_{dp}, \\
\gamma_{^3He^3He}=\frac{(n_{^3He})^2}{2}\langle \sigma v\rangle_{^3He^3He}.
\end{eqnarray}
Here $n_i$ is the concentration of the respective species, $\langle \sigma v\rangle_{ij}$ is the reaction rate of interacting particles $i\, {\rm and }\,j$, $v$ is their relative velocity, for reactions \eqref{eq_ep} -- \eqref{eq_ee} $v\simeq 1$
and $\tau_n\approx 1000$ s is the neutron lifetime. The electron-electron, electron-proton and electron-neutron cross section are given by Eqs.\eqref{EE} and \eqref{SW} of Appendix.\\
The temperature balance is defined by the first law of thermodynamics
\begin{equation}\label{eq5}
\Delta Q = \delta U,
\end{equation}
where $\Delta Q$ and $\delta U$ are the heat and inner energy gains (in fact, a decrease) of the matter inside the heated area, respectively.
Expanding all the values one obtains
\begin{multline}\label{temp1}
\left[(\gamma_{pp}\cdot Q_1 - \gamma_{\gamma d}\cdot Q_2 + \gamma_{np}\cdot Q_3 + \gamma_{dp}\cdot Q_4 + \gamma_{^3He^3He}\cdot Q_5)-\right. \\
\left. ( \gamma_{ep} + \gamma_{en} + 2\gamma_{ee} + \gamma_{n} + \gamma_{pp} ) E_{\nu}\right] dt = 4b T^3 dT,
\end{multline}
where $Q_i$ is energy release of the respective reaction,
$E_{\nu}\sim T$ is the energy of outgoing neutrino, $b = \pi^2 /15$ is the radiation constant.
Using Eq. \eqref{N122} and \eqref{temp1} and reactions \eqref{eq_ee} - \eqref{He4_eq}, we can compose the following system of differential equations.
\begin{align}
\frac{d(n_n)}{dt}&= n_{e^-}n_p\langle \sigma v\rangle_{e^-p}+n_{\gamma}n_{d}\langle \sigma v\rangle_{\gamma d} -\frac{n_n}{\tau_n}-n_{n}n_p\langle \sigma v\rangle_{np}-n_{e^+}n_n\langle \sigma v\rangle_{e^+n} \label{eq_nn} &&\\
\frac{d(n_p)}{dt}&= n_{e^+}n_n\langle \sigma v\rangle_{e^+n} + \frac{n_n}{\tau_n} + n_{\gamma}n_{d}\langle \sigma v\rangle_{\gamma d} + (n_{^3He})^2\langle \sigma v\rangle_{^3He^3He} \nonumber \\& - n_{e^-}n_p\langle \sigma v\rangle_{e^-p} - n_{p}^2\langle \sigma v\rangle_{pp} - n_{d}n_p\langle \sigma v\rangle_{dp} \label{eq_pp}&&\\
\frac{d(n_d)}{dt}&= \frac{n_{p}^2}{2}\langle \sigma v\rangle_{pp} + n_{n}n_p\langle \sigma v\rangle_{np} - n_{d}n_{p}\langle \sigma v\rangle_{dp} - n_{\gamma}n_{d}\langle \sigma v\rangle_{\gamma d} \label{eq_DD} &&\\
\frac{d(n_{^3He})}{dt}&=n_{d}n_{p}\langle \sigma v\rangle_{dp}-(n_{^3He})^2\langle \sigma v\rangle_{\text{$^3$He$^3$He}} \label{eq_3He} &&\\
\frac{d(n_{^4He})}{dt}&=\frac{(n_{^3He})^2}{2}\langle \sigma v\rangle_{\text{$^3$He$^3$He}} \label{eq_4He}&&\\
\frac{d(T)}{dt}&=[(\gamma_{pp}\cdot Q_1 - \gamma_{\gamma d}\cdot Q_2 + \gamma_{np}\cdot Q_3 + \gamma_{dp}\cdot Q_4 + \gamma_{^3He^3He}\cdot Q_5) \nonumber \\& -(\gamma_{en} + \gamma_{ep} + 2\gamma_{ee} + \gamma_{n} + \gamma_{pp} ) E_{\nu}]/4b T^3 \label{eq_T}
\end{align}
The initial number densities of deuterium and helium are considered to be zero inside the region.
As can be seen from the equations, we do not consider any reactions of heavy elements production for the sake of simplicity.
Evidently, some parts of $^4$He will be transformed into heavier elements subsequently, so our estimations of its number density effectively show the number density of $^4$He together with all heavier elements.
\subsection{Temperature evolution}
\label{temp_sec}
The temperature evolution (Eq.\eqref{eq_T}) follows from the equation system above. It is dominated by the cooling due to the reaction \eqref{eq_ee}.
\begin{figure}[h]
\centering
\includegraphics[width=0.9\textwidth]{TT2.jpg}
\caption{The time behaviour of the temperature inside the heated area.}
\label{T}
\end{figure}
Figure \ref{T} shows the time dependence of the temperature for different initial temperatures $T_0$.
\subsection{Abundances of free protons and neutrons}
\label{pn_sec}
We can estimate the abundance of (free) neutrons and protons numerically using Eqs.\eqref{eq_nn} and \eqref{eq_pp}. Figure \ref{fig001} shows the evolution of the number densities, while Figure \ref{fig001ex} shows the fraction of protons (left) or neutrons (right) from the initial baryon number density
\begin{figure}[ht]
\subfigure{
\includegraphics[width=0.49\textwidth]{Np2.jpg}}
\subfigure{
\includegraphics[width=0.49\textwidth]{Nn2.jpg}}
\caption{Left: The time evolution of the protons density in the region at different initial temperatures. Right: The time evolution of the neutrons density in the region at different initial temperatures.}
\label{fig001}
\end{figure}
\begin{figure}[ht]
\subfigure{
\includegraphics[width=0.49\textwidth]{NPP.jpg}}
\subfigure{
\includegraphics[width=0.49\textwidth]{Nnn.jpg}}
\caption{Left: The time evolution of the abundance protons in the region at different initial temperatures. Right: The time evolution of the abundance neutrons in the region at different initial temperatures.}
\label{fig001ex}
\end{figure}
One can explain qualitatively these figures. There are five processes considered to affect the neutron number density. However, while the production (Eq. \eqref{npd_eq}) and destruction (Eq.\eqref{dg_eq}) of the deuterium are generally the two most active of them, their reaction rates have almost negligible difference in the most cases. Therefore, the neutron abundance is defined by neutron decays (Eq.\eqref{n_eq}) with the combined effect of electron-proton and positron-neutron reactions (Eqs.\eqref{eq_ep} and \eqref{eq_en}). At the higher initial temperatures the latter starts as dominant, slowly decreasing its effect with the fall of the temperature, until it reaches the level of the neutron decays somewhere below 1 MeV. After that, the combination of all three of these reactions causes the slow and gradual fall of neutron abundance. At low initial temperature neutron decays start as dominant process, causing the exponential drop at around $10^3$ seconds, until the decay rate matches the one of the e-p and e-n combination. After that, the neutron abundance remains stable for a long time until the temperature starts having noticeable changes, affecting the reaction rates and causing the neutron number density to have a slow and gradual fall, as in the case of high initial temperatures.
The rise in proton number density is caused by neutron decays (slowed down due to the processes described above). This effect is more visible at the high initial temperatures, as the neutrons constitute a higher part of the baryons there. The consequent fall in the proton number density is caused by the irreversible transition of the baryons to the $^3$He and $^4$He.
\subsection{Abundances of deuterium and helium-3}
\label{dhe_sec}
Figure \ref{fig002} shows the evolution of deuterium (left panel) and $^3$He fractions with time.
\begin{figure}[ht]
\subfigure{
\includegraphics[width=0.49\textwidth]{DD2.jpg}}
\subfigure{
\includegraphics[width=0.49\textwidth]{HE32.jpg}}
\caption{Left: The time evolution of the abundance ratio $(n_d/n_B)$ in the region at different initial temperatures. Right: The time evolution of the abundance ratio $(n_{^3He}/n_B)$ in the region at different initial temperatures. }
\label{fig002}
\end{figure}
The reaction rates of the deuterium production and destruction equalize themselves under the current values of the temperature and number densities of neutrons and protons in a very short amount of time, reaching the <<equilibrium>>. This equilibrium keeps adjusting to the changes in those values with time.
The abundance of helium-3 is growing most of the time, as the rate of its production is greater than the rate of its destruction into $^4$He. For the high initial temperatures, at late time, this situation reverses due to the decrease in temperature and number densities of protons and, subsequently, deuterium, and $^3$He starts falling.
The Figure
\ref{fig002} shows that
deuterium and helium-3 have very low abundances, making them very likely undetectable. Nonetheless, they play a significant role in the synthesis of heavier elements due to their high reaction rates.
\subsection{Abundance of helium-4 with heavier elements}
\label{hehe_sec}
We can estimate the helium abundance $(n_{^4He}/n_B)$ numerically from Eq. \eqref{eq_4He}. Figure \ref{fig003} shows the obtained results.
\begin{figure}[ht]
\subfigure{
\includegraphics[width=0.49\textwidth]{HE422.jpg}}
\subfigure{
\includegraphics[width=0.49\textwidth]{HE421.jpg}}
\caption{The time evolution of the density ratio $\rho_{^4He}/\rho_B = 4n_{^4He}/n_B$ in the region at different initial temperatures. Left panel: log-log scale. Right panel: log-linear scale.}
\label{fig003}
\end{figure}
As already stated above, $^4$He abundance here effectively stands for not only helium-4 itself, but also for heavier elements. While our assumptions do not allow us to estimate the metallicity of such a region, we can still make an interesting conclusion that for most of the considered initial temperatures, the dominant part of baryons will be transformed into helium-4 and subsequent elements, leaving the region with almost no hydrogen.
\section{Conclusion}
\label{concl}
We considered the possible existence of stable hot areas
formed in the early Universe. Their origin could be related to the formation of PBH clusters. There are many factors that affect the evolution of such regions, we focus here on the nuclear reactions inside them. The neutrinos produced in these processes carry away energy, what is found to play a decisive role in temperature change under our approximation (considering nuclear reactions only with the given density and reaction rate dependencies). The considered nuclear reactions tend to form heavy elements, depleting the hydrogen content. The absence of hydrogen in such areas may be a distinguishing feature for their possible search. It will be possible to relate the observed chemical composition to its initial temperature and can account for the existence of anomalous stars.
\textit{Acknowledgment}
The work was supported by the MEPhI Program Priority 2030.
\printbibliography[heading=bibintoc,title={References}]
|
\section{Introduction}\label{sec:intro}
Canonical variables such as position and momentum can be defined in terms of their commutation relations. As von Neumann showed \cite{vNunique}, this uniquely fixes the standard description of a quantum system in a Hilbert space with explicitly given operators. This approach turns out to be very useful also for studying operations on such systems. For example, a symplectic linear transformation of the underlying phase space gives just another set of operators with the same commutation relations. Hence by von Neumann's Theorem, there must be a unitary implementing it. Here we generalize this approach in two directions: On one hand, we also allow dissipative operations, which turn some pure states into mixed states, and which are given by completely positive maps. On the other, we include a classical part described by operators commuting with all others, allowing for general quantum-classical hybrid systems. This allows treating problems with a mixture of classical and quantum information, such as various measurement scenarios. The defining property of the channels which can be studied in this way is that they intertwine two actions of the phase space translations or, equivalently, take Weyl operators into Weyl operators. Such operations have been called quasifree \cite{fannes,demoen}. However, an analytic treatment in the generality needed for a practical calculus and including hybrid systems does not seem to exist. It will be provided in this paper.
While the main aim is to build an easy-to-use general calculus for quasifree hybrids, we had to go deeper into the functional analysis of such systems than we had anticipated. The reason is that, although the formal structures for states and observables for purely quantum systems and for the purely classical systems (probability) are well established, this cannot be said for the hybrid combination. Indeed, the standard approaches for the two extreme cases are not easily merged. These problems are aggravated when discussing operations (channels) between systems: Should they be described in the Schr{\"o}dinger\ or the Heisenberg picture? On which spaces should these act? Our answer is a setting, in which both pictures always make sense, for any quasifree channel. So in the practical calculus of quasifree channels, the technicalities and the conceptual issues, whose resolution forms the main body of this paper, can be taken as resolved. When questions of analytic properties of observables and channels come into play, details can be taken from our paper, but for many questions, these details can be ignored, and the calculus can be used and applied straightforwardly and rigorously to a large variety of measuring and control scenarios. It strictly includes the world of ``Gaussian Quantum Information''. But we no longer need separate definitions of Gaussian states and Gaussian measurements: Gaussianness is a property of arbitrary channel types, and it is immediate from our definition that the composition of Gaussian channels is again Gaussian. However, the quasifree setting is richer, including, for example, arbitrary channels with no input, i.e., arbitrary states.
\vskip 12pt\noindent
{\it Our paper is organized as follows:}\\
We finish the introductory Sect.~\ref{sec:intro} with three subsections placing our problem in context. Logically they can be skipped, i.e., they contain no material needed to follow the formal development in later chapters.
First (Sect.~\ref{sec:statobs}), we have some general remarks on how to choose good spaces of states and observables, how this issue has been viewed traditionally, and why hybrids pose a special challenge. We then (Sect.~\ref{sec:previousCCR}) give a very brief overview of the rich literature on the canonical commutation relations, mainly for the historical background and pointers to useful summaries. We do not take knowledge of this literature as a prerequisite, however. The body of the paper is mostly self-contained, i.e., we give arguments for the main steps, even when they could also be covered by a citation.
The literature on quantum-classical hybrids, reviewed in Sect.~\ref{sec:previous}, is more disparate, even including some approaches that fail. We provide a list of pertinent research projects with brief discussions of some aspects. We also note how these projects relate to our paper. The three background sections can be summarized by saying that we did not find any works covering all three aspects.
After introducing the basic phase space variables, Sect.~\ref{sec:states} deals with the ``good states''. Using the terminology introduced later, these are the ones with continuous characteristic functions (Thm.~\ref{thm:Bochner}), and turn out to be the state space of a C*-algebra without unit, denoted ${\mathrm C}^*\Xis{}$. Observables are treated in Sect.~\ref{sec:funcobs}. This is the most technical part of the paper. Here we get different choices of spaces of observables. Each is represented as a class of functions from the classical parameter space to the bounded operators of the quantum subsystem with varying degrees of regularity, e.g., weak measurability, strong*-continuity, and norm continuity. The aim is to show later that these properties are preserved under arbitrary quasifree channels. A direct approach would be fraught with case distinctions concerning the separation and recombination of classical and quantum parts under the channel. Therefore we go another way (Sect.~\ref{sec:semicont}), namely giving a characterization of these properties in a general operator algebraic setting, which applies to classical, quantum, and hybrid systems alike. The preservation of properties is then almost trivial to show (Prop.~\ref{prop:Heisenalg}).
Quasifree channels are defined and characterized in Sect.~\ref{sec:channels}. In order to get a feeling for the large variety of operations covered by that definition, we recommend skipping ahead to Sect.~\ref{sec:BasicOps} for examples. Meanwhile, Sect.~\ref{sec:channels} focuses on general constructions. Basic ways to combine them are described in Sect.~\ref{sec:combine}. A recurring theme is a correspondence between the quasifree channel and a state that we call its noise state. The complete positivity condition is exactly equivalent to the positivity condition for this state, which is a state on an explicitly given hybrid system. This may have a classical part, or may be entirely classical, even for channels between purely quantum systems. Thus the hybrid work needs to be done even if one wants to study only quantum channels. A key result is the factorization of any quasifree channel into preparing the noise state on some environment and then executing a noiseless quasifree operation, for which the noise state is pure and classical. Noiseless operations (Sect.~\ref{sec:noiseless}) are homomorphisms, so this is a variant of the Stinespring dilation.
Sect.~\ref{sec:BasicOps} sketches some of the possibilities of combining classical and quantum information in input and output. In particular, we parameterize covariant phase space instruments and their characteristic tradeoff between measurement accuracy and disturbance. Some aspects, like optimal cloning, the classical limit, and dynamical semigroup evolutions, are only sketched because they are treated in past or future articles of their own. In these cases, we merely indicate how these subjects fit into the framework.
\subsection{The basic problem: Good spaces for states and observables}\label{sec:statobs}
The basic statistical interpretation of quantum theory has two primitives, states and observables, which operationally stand for preparations and measurements. Mathematically they are represented in appropriate ordered Banach spaces, and the basic interpretation demands that there is a bilinear form allowing to evaluate the expectation value of any observable in any given state. In all the usual theories, however, this symmetric view is broken, and either the states or the observables are taken as primary and the dual objects as secondary, derived quantities. This section is about that chicken-and-egg situation. It is of special interest for hybrids since, in the end, we will settle for a third option.
The standard view of quantum mechanics has states described as density operators (positive trace class operators of trace $1$ on a Hilbert space ${\mathcal H}$). The maximal set of observables for which expectation value evaluations can be defined is the set of {\it all} bounded linear functionals on the trace class ${\mathcal T}({\mathcal H})$, i.e., the Banach space dual, which in this case is equal to ${\mathcal B}({\mathcal H})$, the space of bounded operators on ${\mathcal H}$, where the expectation values are expressed by $\operatorname{tr}(\rho A)$ for $\rho\in{\mathcal T}({\mathcal H})$ and $A\in{\mathcal B}({\mathcal H})$. For classical systems, two different choices are common. One can either take the algebra ${\mathcal C}(X)$ of continuous complex-valued functions on a compact space $X$ as the observables. The dual then consists of all finite Borel measures on $X$. Or else, one can take the space $L^1(X,\mu)$ of probability densities with respect to some reference measure as the space of states, of which the dual is $L^\infty(X,\mu)$. These basic options are visualized in Fig.~\ref{fig:setting1}.
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=1]
\node[left] at (0,1.6) {states};
\node[right] at (1,1.6) {observables};
\node[] at (0.5,2) {Classical};
\node [left] at (0,1) {${\mathcal C}(X)^*$};
\node [left] at (0,0) {$L^1(X,\mu)$};
\node [right] at (1,1) {$L^\infty(X,\mu)$};
\node [right] at (1,0) {${\mathcal C}(X)$};
\node [rotate=90] at (-0.6,0.5) {$\subset$};
\node [rotate=90] at (1.5,0.5) {$\subset$};
\draw[red] (0,1) --(1,0);
\draw[blue] (0,0) --(1,1);
\draw[] (0,0) --(1,0);
\node[left] at (5,1.6) {states};
\node[right] at (6,1.6) {observables};
\node[] at (6,2) {Quantum};
\node [left] at (5,0) {${\mathcal T}({\mathcal H})$};
\node [right] at (6,1) {${\mathcal B}({\mathcal H})$};
\draw[blue] (5,0) --(6,1);
\end{tikzpicture}
\captionsetup{width=0.8\textwidth}
\caption{ Dualities of spaces of states (left column of each diagram) and observables (right column). A line indicates a dual pairing, i.e., a way to compute the expectation of any observable in the right space with any state on the left. The W*-approach (blue line) starts from the states, and allows the full dual space as observables. The C*-approach (red line) makes the opposite choice. In the classical case the combination (black line) is also well defined. Traditional quantum mechanics has only the W*-approach. }
\label{fig:setting1}
\end{center}
\end{figure}
In the classical case, the structure of ${\mathcal C}(X)$ is determined by the topology of $X$, whereas the structure of $L^1$ is based on the measure theory of $(X,\mu)$. In the non-classical setting of operator algebraic quantum theory \cite{bratteli1,haag_book,emch,landsman_foundations_2017,landsman_algebraic_2009} the distinction appears as the contrast between the C*-algebraic approach, which generalizes the topological side, and the W*-algebraic approach, which generalizes the measure-theoretic side. As indicated by the letters C and W, the distinction is one between the abstract versions of norm \underline{c}losed *-subalgebras of ${\mathcal B}({\mathcal H})$ versus \underline{w}eakly closed ones (also called von Neumann algebras). Then a famous characterization theorem \cite{SakaiDuality} links this distinction to the chicken-and-egg problem mentioned in the first paragraph: The W*-algebras are precisely those C*-algebras that are dual Banach spaces, i.e., coincide with {\it all} bounded linear functionals on some state space.\newpage
As the classical case shows, neither view is in any sense more ``correct'' than the other: Topology and measure theory just capture different aspects of the system description. Curiously, one of the founding papers of the algebraic quantum field theory (AQFT) school \cite{HaagKastler} tries to make the point that the C*-algebras should be taken as primary. However, the arguments presented there can be straightforwardly dualized to give the dual conclusion, namely, that the states should be taken as primary. Luckily, the AQFT school has largely ignored the advice of \cite{HaagKastler}, and did substantial work singling out subsets of ``physically relevant states'' and separating the local part of the problem from the global aspects by defining canonical W*-algebras for local regions, while analyzing the structure at infinity (superselection sectors) in terms of the representation theory of a quasilocal C*-algebra. So it can be said that the mature version of AQFT \cite{haag_book} takes a rather refined and complex view of chicken and egg.
In axiomatic quantum mechanics, many schools have focused on the finite-dimensional case, where the problem does not arise: In that case, the spaces are reflexive, i.e., equal to their second duals.
One traditional school allowing infinite dimension from the outset is the approach by Ludwig \cite{ludwig1,ludwig2}. According to him, a complete picture of quantum mechanics should be symmetric, even though his axiomatic reconstruction just returns the asymmetric standard view shown in Fig.~\ref{fig:setting1}. The problem of choosing an appropriate space of physically realizable observables (${\mathcal D}\subset{\mathcal B}({\mathcal H})$ in Ludwig's notation) is left as an open problem. It is not one to be solved once and for all. A solution will have to depend on further specifics of the system \cite{uniformities}. This is analogous to the classical case: All $L^1$-spaces are isomorphic (if $X$ has no ``atoms'', i.e., points of positive measure, \cite[Thm.~III.1.22]{takesaki1}), so the measure theory does not even see the dimension of $X$. Similarly, ${\mathcal T}({\mathcal H})$ depends only on $\dim{\mathcal H}$, but any reasonable ${\mathcal D}$ will contain more structure.
A minimal condition on a space of observables is that there are sufficiently many observables to distinguish the states. With only this minimal condition, it is guaranteed that the weak limits of observables span the full dual of the state space. That is, all choices are weak* dense in each other. One might use this to justify an asymmetric scenario including all these limits as idealized observable elements. However, what gets lost in this asymmetric picture is a description of the physical distinguishability of states. Therefore one loses the idealized states one would similarly find in the dual of the observable space. This is, in fact, the basis of the argument in \cite{HaagKastler}, only that it works both ways (see \cite{RFWdiss} for a formal parallel development of both aspects from the Ludwig point of view). This symmetric view can be seen very clearly in the classical case (Fig.~\ref{fig:setting1}): Starting from just the bottom duality of probability densities and continuous functions, we can approximate arbitrary measures weakly by densities, including point measures. Dually we can take weak* limits in $L^\infty$ to approximate arbitrary bounded measurable functions by continuous ones. Again we gain additional extremal elements, like indicator functions, which are typically not continuous. However, there is no natural evaluation of a point probability measure on an indicator function in $L^\infty$: The elements of $L^\infty$ are classes with respect to almost everywhere equality, and so the value at a point (typically of measure zero) is not defined. Thus there is no natural pairing at the top level of the diagram of dualities (see, however, \cite{Ionescu} for the existence of {\it some} non-constructive pairing).
In the non-commutative case, the story is similar: By going to the full dual spaces, one gains idealized elements in the form of pure states and projections, whose abundance is guaranteed by the Banach--Alaoglu and Krein--Milman Theorems. These extremal elements are often the building blocks of an analysis, as in the decomposition of arbitrary states as an integral over pure states (Choquet Theory, \cite[Ch.~4]{bratteli1}) and, dually, in the spectral resolution of normal observables.
The explanation for the existence of pure states does not account for the many pure states $\kettbra\psi\in{\mathcal T}({\mathcal H})$, i.e., standard quantum mechanics, which we characterized as an instance of the W*-view. Indeed for W*-algebras of Murray--von Neumann type II and III, there are no normal pure states (see also Lem.~\ref{lem:atomic}). However, the pure states can be related to another duality: The trace class ${\mathcal T}({\mathcal H})$ is the Banach space dual of ${\mathcal K}({\mathcal H})$, the space of compact operators. The unit ball of ${\mathcal T}({\mathcal H})$, therefore, has to have many extreme points. It would be impractical, however, to use the compact operators as the basic observable algebra: It does not contain a unit, so we cannot express the normalization condition for POVM observables.
That $\idty\notin{\mathcal K}({\mathcal H})$ means that the normalization functional is not continuous in the weak* topology, so the states are not a weak*-closed subset of the unit ball and not compact. This is a crucial observation for limits in this topology. For example, consider a sequence of states averaged over an expanding range of spatial translations. Any cluster point would have to be translation invariant; but no such density operator exists, because translations have continuous spectrum. On the other hand, by the weak* compactness of the unit ball, cluster points must exist. Indeed the weak* limit is the zero functional. So it does happen that the limit of states fails to be a state. In close analogy, we can consider classical states represented by probability measures on ${\mathbb R}^n$ and their shifts or averages to infinity. These are in the dual of ${\mathcal C}_0({\mathbb R}^n)$, the space of continuous functions vanishing at $\infty$, and weak limits of shifted sequences are zero.
So in both cases, the natural state space can be located in the dual of an algebra ${\mathcal A}$ {\it without identity}, namely ${\mathcal A}={\mathcal K}({\mathcal H})$ in the quantum case, and ${\mathcal A}={\mathcal C}_0({\mathbb R}^n)$ in the classical case. But this does not mean going back to an asymmetric picture with observables taken as primary, i.e., another turn in the chicken-and-egg conundrum because ${\mathcal A}$ is not itself the algebra of observables. The full set of observables will be some set of functionals on the states in ${\mathcal A}^*$, so a subspace of ${\mathcal A}^{**}$. All such observables can be approximated weakly by elements of ${\mathcal A}$. In particular, there is an approximate unit, so $\idty$ is recovered. This scheme is outlined in Fig.~\ref{fig:setting2}. We propose to use it for hybrids, as well, and with the particular choice ${\mathcal A}={\mathcal K}({\mathcal H})\otimes{\mathcal C}_0(X)$ as the C*-tensor product of quantum and classical parts.
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=0.92]
\node[] at (0.5,2) {Classical};
\node[left] at (0,1.6) {states};
\node[right] at (1,1.6) {observables};
\node[left] at (0,0.5) {${\mathcal C}_0(X)^*$};
\node[right] at (1,1) {${\mathcal C}_0(X)^{**}$};
\node[right] at (1,0) {${\mathcal C}_0(X)$};
\draw[red] (0,0.5) --(1,0);
\draw[blue] (0,0.5) --(1,1);
\node[] at (5.5,2) {Quantum};
\node[left] at (5,1.6) {states};
\node[right] at (6,1.6) {observables};
\node[left] at (5,0.5) {${\mathcal T}({\mathcal H})$};
\node[right] at (6,1) {${\mathcal B}({\mathcal H})$};
\node[right] at (6,0) {${\mathcal K}({\mathcal H})$};
\draw[blue] (5,0.5) --(6,1);
\draw[red] (5,0.5) --(6,0);
\node[] at (10.5,2) {Hybrid};
\node[left] at (10,1.6) {states};
\node[right] at (11,1.6) {observables};
\node[left] at (10,0.5) {${\mathcal A}^*$};
\node[right] at (11,1) {${\mathcal A}^{**}$};
\node[right] at (11,0) {${\mathcal A}={\mathcal K}({\mathcal H})\otimes{\mathcal C}_0(X)$};
\draw[blue] (10,0.5) --(11,1);
\draw[red] (10,0.5) --(11,0);
\end{tikzpicture}
\captionsetup{width=0.8\textwidth}
\caption{Extended dualities suitable for a joint generalization of the classical and the quantum case to hybrids.
The states are here functionals on an underlying non-unital algebra ${\mathcal A}$, namely $\mathcal{C}_0(X)$ resp.\ ${\mathcal K}({\mathcal H})$ in the classical resp.\ quantum case. The biduals ${\mathcal C}_0(X)^{**}$ and, in general, ${\mathcal A}^{**}$ are relatively wild objects,
whose elements are not easily characterized. Selecting more managable subspaces is the subject of Sect.~\ref{sec:funcobs}. }
\label{fig:setting2}
\end{center}
\end{figure}
It turns out that this is equivalent to another way of identifying a good hybrid state space for systems of canonical variables. One takes, for each state, the expectation values of the Weyl operators (displacement operators). This is called the characteristic function of the state. Then the good states in ${\mathcal A}^*$ are precisely (Thm.~\ref{thm:Bochner} and \ref{prop:twgroup}) those with a continuous characteristic function (usually called ``regular''). Intuitively, these states live essentially in bounded regions of phase space, so the expectation of Weyl operators can be expected to depend continuously on its argument. This is in contrast to the observation that distinct Weyl operators have maximal norm distance because somewhere far out, even periodic functions with very similar periods will differ maximally. This regularity condition has an equally simple analog for quasifree channels, which is equivalent to the channel mapping regular states to regular states. We include this in our definition of quasifree channels. On the side of observables, things are more complex, however. On the maximal choice of observable algebra, the bidual ${\mathcal A}^{**}$ the Heisenberg picture of a channel is defined as the Banach space adjoint. However, the bidual is a monster, even in the classical case. We devote Sect.~\ref{sec:funcobs} to the task of finding smaller, more manageable subalgebras, which nevertheless support a Heisenberg picture for arbitrary quasifree channels. That there is some choice here is a strength of our theory rather than a weakness: It amounts to a variety of analytic conditions (weak measurability, strong*-continuity, norm continuity, and in some sense many more), which are all automatically preserved by the Heisenberg actions of quasifree channels.
Let us contrast this with an approach that naively takes inspiration from the finite case, where $\dim{\mathcal H}<\infty$, and the classical variables lie in a compact set $X$, and the appropriate observable algebra is the C*-tensor product ${\mathcal A}={\mathcal B}({\mathcal H})\otimes{\mathcal C}(X)$. When $X\cong{\mathbb R}^s$ locally compact and $\dim{\mathcal H}=\infty$, we could just replace ${\mathcal C}(X)$ by the bounded continuous functions ${{\mathcal C}_{\mathrm b}}(X)$, and consider the C*-algebra $\widetilde{\mathcal A}={\mathcal B}({\mathcal H})\otimes{{\mathcal C}_{\mathrm b}}(X)$. This algebra contains the Weyl operators as well, so we can associate with every state on $\widetilde{\mathcal A}$ a characteristic function. In fact, we started our investigation of quasifree channels with this choice, running into more and more difficulties: Since the Weyl operators do not span a norm dense subalgebra of $\widetilde{\mathcal A}$, the characteristic function will fail to characterize a state. Moreover, many states on $\widetilde{\mathcal A}$ are singular in the following sense: The classical marginal will be a state on ${{\mathcal C}_{\mathrm b}}(X)$, which is isomorphic to ${\mathcal C}(\beta X)$, where the $\beta X$ denotes the Stone--\v Cech compactification of~$X$. But if such a state has any weight on the infinite points $\beta X\setminus X$, it will not be a probability measure on $X$. Similarly, on the quantum side, a state on ${\mathcal B}({\mathcal H})$ may fail to be given by a density operator. For a usable calculus of quasifree channels, this is bad news because it would be unclear how to even define such channels on all of $\widetilde{\mathcal A}$, how to exclude that the output of any such operation is singular, and even if all that is worked out, whether the Heisenberg picture would map into the input counterpart of $\widetilde{\mathcal A}$.
\subsection{Previous works on Canonical Commutation Relations}\label{sec:previousCCR}
The description of physical systems in terms of canonical commutation relations constitutes a core part of modern mathematical physics. The tradition begins with Weyl and von Neumann and extends to the community that founded the journal {\it Communications in Mathematical Physics} in 1965. It was felt then that this structure, and more generally a C*-algebraic view of physics, was a key element of both quantum field theory in the approach of Haag, Kastler \cite{Kastler1965,HaagKastler}, and Araki \cite{araki}, and of statistical mechanics in the school of Verbeure and others (see \cite{bratteli1,bratteli2} for a textbook expounding these ideas). We are obviously building on this tradition and can hardly give due credit, not even to the major contributors.
Notable expositions are \cite{bratteli2} for a Fock space based view with statistical mechanics in mind and \cite{Derezinski} for a field theoretical one. An encyclopedic work aiming at quantum optics is \cite{Honegger}. Irreversible operations, i.e., quasifree channels as developed in Sect.~\ref{sec:channels}, came up in \cite{demoen,evansLewis}. These are, of course, also the focus of works on quantum information. In that community, the canonical systems are called systems with ``continuous variables'' as opposed to those composed of discrete quantum bits. The main interest has been in the Gaussian case because the vacuum at the empty port of a beam splitter, laser light, and cooled oscillators are all Gaussian. A collection of review articles is \cite{QICV}. Systematic expositions of the Gaussian structure are \cite[Ch.~12]{HolevoQSCI} and \cite{MCF}.
Hybrids did show up occasionally in this literature, but usually not as the main focus. For example, some core results in \cite{Honegger} are formulated without assuming the commutation form to be non-degenerate (see also \cite{manuceau}). This assumes the observable algebra on the classical side to be CCR-like, i.e., the almost periodic functions, which is not a good choice by the considerations of the previous section, but by far the most frequent choice in the literature described here.
\subsection{Previous works on hybrid systems}\label{sec:previous}
In the following list, we have collected some of the appearances of quantum-classical hybrids in the literature. The motivations are rather different, and after each brief description, we point out how the respective research project differs from the present study.
\begin{labeledlist}{l}
\item[\textit{Quantum Field Theory}] In quantum field theory, the algebra of canonical commutation relations provided a way to deal with the commutation relations of field operators in QFT without discussing tricky domain questions of these unbounded operators. In this way, quantum fields could be included in the newly forming C*-algebraic approach to quantum mechanics, which is crucially based on bounded operators. However, these works were not interested in going beyond the CCR-algebra, i.e., the C*-algebra generated by the Weyl operators. The technical difficulties coming from an infinite-dimensional test function space (phase space) and the resulting failure of von Neumann's uniqueness theorem (closely related to ``Haag's Theorem'') seemed more relevant.
It was noted only later that even for field theory, the CCR-algebra has its drawbacks. Classically it corresponds \cite{QHA} to the almost periodic functions, which by definition depend very sensitively on infinite values of the fields. This is reflected by the structure of the pure states on the algebra of almost periodic functions, which form the so-called Bohr compactification of ${\mathbb R}$. Along with the points in ${\mathbb R}$ it contains further limit points. But these new points are in no sense ``at infinity'' and themselves dense, so that the finite and the infinite are highly intertwined. Of course, this is related to the observation that almost periodic classical variables are almost impossible to measure, so the topology of the Bohr compactification is unrelated to physical distinguishability. The same criticism applies to the use of the CCR-algebra. In particular, it is difficult to implement physical dynamics as a C*-dynamical system on this algebra \cite[p.\,345]{bratteli2}. A more regular approach using resolvents rather than exponentials of the fields has therefore been proposed \cite{buchholz1} (cf. Example~\ref{Ex:resAlg}).
In contrast to the QFT literature, we consider only finite-dimensional phase spaces in this paper. In other aspects, our approach is more general, particularly by including irreversible operations. QFT focuses on reversible dynamics, so irreversible operations play no role (but see \cite{Longo}).
\item[\textit{Hybrids and canonical structure}] Both classical and quantum systems employ Hamiltonians to generate the dynamics, and it appears almost like a minor difference that one uses commutators while the other uses Poisson brackets. Consequently, there have been many attempts to fuse these structures into one common framework (see \cite{elze_quantum-classical_2013} for a review). This gains further plausibility from systems with quadratic Hamiltonians, which generate isomorphic Lie algebras in the quantum and classical worlds. On the whole, however, this basic idea has proved to be a failure \cite{peres_hybrid_2001,terno_inconsistency_2006}. Typical problems include dynamics, which do not preserve the positivity of states or allow non-local signaling due to some obscure non-linearity imported into the quantum system from the classical side. The core of these problems is actually a very familiar No-Go Theorem from quantum information theory: {\it There is no information gain without disturbance} \cite{Busch2009}. That is, whenever an interaction leads to the possibility of measuring a variable of the classical subsystem and thereby gaining information about the initial quantum state, some irreversibility must be involved. Thinking of dynamics in terms of Hamiltonians and canonical structure is, however, so tied up with reversible dynamics that any approach based on canonical structures is bound to fail, at the latest, when there is a non-trivial interaction.
For our paper, this has the consequence that we do not even assume a symplectic structure on the classical system, i.e., the classical phase space is a real vector space without further structure.
\item[\textit{Dissipation}] The No-Go Theorem strongly suggests the use of dissipative time evolutions to express the measurement interaction \cite{diosi_hybrid_2014,barchielli_1996,olkiewicz_dynamical_1999}. The quasifree case \cite{barchielli_1996} benefits especially from the clarification of the complete-positivity conditions for channels (\cite{evansLewis}, our Sect.~\ref{sec:channels}).
In \cite{Lars}, we show that, even without quasifreeness, this leads to a fusion of the classical theory of diffusion generators on one hand and Lindblad generators on the other, with a full understanding of the additional interaction terms that describe the information transfer from the quantum to the classical subsystem.
\item[\textit{Embedding the classical system into a quantum one}]
In the quantum information community, many researchers think of the observables of a classical system as the diagonal matrices embedded into a larger full matrix algebra.
Similarly, for a classical particle described by position variables in ${\mathbb R}^n$, one can get a quantum extension by including the generators of the spatial shifts, i.e., conjugate momenta, in a crossed-product \cite{takesaki1} construction.
This construction can be done at the von Neumann algebra level so that the enlarged quantum system has the full algebra of bounded operators over $L^2({\mathbb R}^n,dx)$ as observables. This is the approach to hybrids chosen, for example, in \cite{barchielli_1996}. In this setting, the distribution of the classical variables in a normal state always has an $L^1$-density, which excludes the pure states of a hybrid.
We will see later that the pure states of a modified hybrid also correspond to extremal quantum channels, so this approach excludes the optimal, e.g., minimal noise channels for some tasks.
In our approach, pure states are included from the outset, and the von Neumann algebraic crossed-product embedding is characterized as a special case for which states are norm continuous under translations (Sect.~\ref{sec:transState}).
\item[\textit{The classical system is a large quantum system}] A common point of view is that the classical world is merely emergent as a limiting case of large quantum systems. In this spirit, a hybrid would always be a large quantum system with one subsystem close to a thermodynamic limit. There is no problem then writing down Hamiltonian interactions between the almost classical and the quantum part. However, this does not resolve the No-Go Theorem for hybrid dynamics. The classical variables in such a system will generally evolve into some combination involving their conjugates, or as \cite{sherry_interaction_1978} phrases this, the classical variables lose their ``integrity''. The required physical discussion at this point would be that effectively, and to good approximation, the classical integrity is preserved. But often, the models of quantized classical systems are so simple (e.g., one degree of freedom \cite{gisin_quantum_2000}) that a physical discussion of a thermodynamic limit is not really possible. It should be noted that the classical limit is very closely related to the mean-field limit, and the latter has indeed been proposed as a model for measurement processes involving large quantum systems \cite{Hepp}. In this case, in spite of an infinite range mean field interaction, the measurement result becomes definite only in the infinite time limit.
The many-body aspect of the classical system will not come into play in our paper or even enter the formalism. Conceptually, this is because we consider that limit already being done, and we work with a much-reduced set of classical variables, a finite set of reals, such as a measurement record in a continuing observation process.
\item[\textit{Non-linearity}]
In the mean-field limit, one gets dynamical equations for quantum states, which are of canonical Hamiltonian form and allow strong non-linearities forbidden in standard quantum theory. This is not paradoxical if one realizes that the ``quantum states'' here are not states of a quantum system at all but distribution parameters for a many-body system. Nevertheless, the resulting kind of non-linear Hamiltonian evolution \cite{Bona,Duffield} has been proposed as a testable generalization of quantum mechanics \cite{Weinberg,BonaLong}. In the simplest case, i.e., a qubit, it is classical mechanics with the surface of the Bloch sphere as a phase space manifold and the surface $2$-form as symplectic form. The standard quantum evolutions are driven by Hamiltonians which are linear in the state. These are just rotations of the Bloch sphere and result from non-interacting mean-field systems. The symplectic structure on the pure states has been variously noted, but it is quite misleading to conclude that using it somehow unifies classical and quantum theory. Indeed the classical Hamiltonian structure goes against a basic impossibility claim of standard quantum mechanics, namely that all mixtures giving the same density operator (e.g., unpolarized light) are indistinguishable. To summarize: As an approach to hybrids, the theories starting from the Hamiltonian structure of quantum states resolve the tension between classical and quantum theories by turning the quantum part into a classical system.
In this paper, we stick to the ``minimal statistical interpretation'' for quantum, classical, and hybrid systems alike. The states and observables then operationally represent preparations and measurements. This interpretation implies linearity of all meaningful operations on states and observables, so linearity is not an accidental feature of the theory that can easily be dropped, and we will assume it throughout.
\item[\textit{Hybrids for gravity}]
A recent discussion of hybrids for quantum fields coupled to gravity illustrates several of the options mentioned above. In \cite{oppenheim_post-quantum_2018,oppenheim_two_2022} we find an approach making the dissipative nature of the interaction implicit. In \cite{Bose,Vedral} it is argued that gravitationally induced entanglement would serve as proof of the non-classical nature of gravity. This is contradicted by \cite{hall_two_2018}, where the authors emphasize that this will depend on the notion of hybrids and that the non-linear variant, in particular, would allow for entanglement via a classical intermediary.
Our motivation for hybrid structures is practical and comes from continuous observation and other measurement processes. Whether the resulting structures are also helpful for some fundamental theory is far beyond the scope of this paper. However, we hope that a sharper understanding of the mathematical structures will also be helpful in such projects.
\end{labeledlist}
\section{Hybrid states}\label{sec:states}
\subsection{Setup}\label{sec:setup}
In this section, we fix the basic structure of the systems we consider and the basic notations relating to phase spaces. For those who are familiar with phase space quantum mechanics, this amounts to applying a remarkably simple principle, namely just eliminating the assumption that the commutation form should be symplectic, i.e., non-degenerate.
We consider systems of $n$ quantum canonical degree of freedoms and $s$ classical ones. This means that we have a position variable $q\in{\mathbb R}^n$, and its momentum counterpart, which lies in the dual space~${\mathbb R}^n$. This only means that a scalar product $p\cdot q$ is defined, and the phase space of the system, the set of pairs $(q,p)$ carries a natural {\bf symplectic form} $\sigma((q,p),(q',p'))=p{\cdot} q' -q{\cdot} p'$. More abstractly one needs only to demand that $\sigma$ is antisymmetric and non-degenerate, i.e., the only pair $(p,q)$ such that $\sigma((q,p),(q',p'))=0$ for all $(p',q')$ is $p=q=0$. Then with a suitable choice of ``canonical coordinates'' $\sigma$ will take the given form.
We now drop the assumption of non-degeneracy, i.e., we allow non-zero null vectors for $\sigma$. In a basis this means that the $2n$ variables $p,q$ can be augmented by $0\leq s<\infty$ unpaired classical variables $x\in{\mathbb R}^s$, which can be thought of as position variables without corresponding momenta. So, as an extended {\bf phase space} $\Xi={\mathbb R}^{2n+s}$ we consider the set of triples $\xi=(q,p,x)$. The extended symplectic form will be defined
\begin{equation}\label{symp}
\sigma((q,p,x),(q',p',x'))=p{\cdot} q' -q{\cdot} p' =\sum_{ij}\xi_i\sigma_{ij}\xi'_j,
\end{equation}
which is still antisymmetric and bilinear. When we want to emphasize the generalization, we call $\Xi$ a ``hybrid phase space''. But since this is the normal case in our paper, we will often drop the adjective.
Since we later consider arbitrary linear maps on phase spaces, we usually adopt a convenient basis free view, where the (hybrid) phase space is just a real vector space $\Xi$ with antisymmetric form $\sigma$, so the type of system is given by the pair $\Xis{}$. The classical part is always singled out as the space of null vectors:
\begin{equation}\label{Xi0}
\Xi_0=\{\xi\in\Xi \mid \forall\eta:\, \sigma(\xi,\eta)=0\}.
\end{equation}
Thus we can split $\Xi=\Xi_1\oplus\Xi_0$, where $\Xi_1$ is a suitable subspace on which $\sigma$ is non-degenerate, i.e., a standard quantum system. The direct sum symbol here indicates a unique decomposition $\xi=\xi_1+\xi_0$ with $\xi_i\in\Xi_i$ for any vector $\xi$, and that the form $\sigma$ also has a block structure, as in the coordinatization \eqref{symp}. However, other than an orthogonal complement, the quantum part $\Xi_1$ is not uniquely defined, i.e., there are $\sigma$-preserving linear maps changing the decomposition. Some of our constructions depend on the decomposition $\Xi=\Xi_1\oplus\Xi_0$, but we usually do not show explicitly that this dependence is harmless. In fact, such proofs become trivial exercises once our full theory is established. The necessary isomorphisms will be noiseless quasifree in the terminology of Sect.~\ref{sec:channels}.
Going quantum means that the components of these tuples are turned into operators
\begin{equation}\label{fieldOp}
R=(R_1,\ldots,R_{2n+s})=(Q_1,\ldots,Q_n,P_1,\ldots,P_n,X_{1},\ldots,X_{s})
\end{equation}
with the commutation relations
\begin{equation}\label{CCR}
[R_j,R_k]=i\sigma_{jk}\idty.
\end{equation}
Again, we refer to the classical $X_j$ as ``operators'' out of convenience, although classical ``random variables'' might perhaps be more appropriate. These are the $R_j$ that commute with all others.
A basic symmetry of the theory are the {\bf phase space translations}, which add a constant, i.e., a multiple of the identity to each $R_j$. We denote this transformation by
\begin{equation}\label{alpha}
\alpha_\xi(R_j)=R_j+\xi_j\idty,
\end{equation}
where $\xi_j$ are the components of $\xi$. Clearly, $\alpha_\xi$ preserves the commutation relations and will always be a homomorphism (preserve operator products).
There are many subtleties in the task of finding all operators satisfying \eqref{CCR}, related to domain questions of these unbounded operators \cite{Schmue}. The main regularity condition singling out the usual case is that the operators are essentially selfadjoint on their common domain so that they generate unitary groups. These should satisfy an integrated version of \eqref{CCR}, and rather than diving into the details, we will make that our starting point. In fact, it can be argued \cite[Sect.4.1]{introGroupsQM} that the integrated version \eqref{weylrel} is historically a bit older, and due to Weyl, who proposed it to Max Born, even before the latter published \eqref{CCR}. Hence, following a strong tradition, we pass to the operators
\begin{equation}\label{weylop}
W(\xi)=\exp(i \xi{\cdot} R).
\end{equation}
We refer to the $W(\xi)$ as {\bf Weyl operators}, even though with classical arguments, they may sometimes be more like functions. In \eqref{weylop} the expression $\xi{\cdot} R:=\sum_j\xi_jR_j$ means a mixed vector/operator scalar product, by which the commutation relations become
\begin{align}
W(\xi)W(\eta)&=e^{(-i/2) \xi{\cdot}\sigma\eta} W(\xi+\eta) \label{weylrel}\\
&=e^{-i\xi{\cdot}\sigma\eta} W(\eta) W(\xi). \label{weylcom}
\end{align}
We will refer to \eqref{weylrel} as the {\bf Weyl relation}, while \eqref{weylcom} are called the {\bf canonical commutation relations (CCR)} in Weyl form. The CCR-algebra over $\Xis{}$, denoted by ${\mathrm{CCR}}\Xis{}$, is the universal C*-algebra of these generators and relations. That is, every realization of the relations by unitary operators $W(\xi)$ on a Hilbert space ${\mathcal H}$ is given by a representation of ${\mathrm{CCR}}\Xis{}$ which takes the abstract generators to the $W(\xi)$.
There are some notational choices here that we should comment on. We have not included $\sigma$ in \eqref{weylop}, simply because this would set the classical contribution to zero. In phase spaces with a proper symplectic form, this form is often used to identify the space with its dual (e.g., \cite{QHA}). Any constructions using this will not work in our context. This means that in a coordinate-free spirit, the variable $\xi$ in \eqref{weylop} does not lie in the phase space but in its dual. We will keep the notation simple by nevertheless identifying both spaces with ${\mathbb R}^{2n+s}$ and using a dot for the standard scalar product. However, a notation keeping the distinction between phase space $\Xi$ and its dual $\hXi$ can easily be restored throughout the paper using the following rules:
(a) in the above description $(p,q,r)\in\Xi$, and the argument of any shift $\alpha_\xi$ is $\xi\in\Xi$, (b) the argument of any Weyl operator $W(\xi)$ is $\xi\in\hXi$, (c) as a key parameter for quasifree channels (Sect.~\ref{sec:channels}) we use maps between phase spaces $S:\hXi_{\mathrm{out}}\to\hXi_{\mathrm{in}}$. Their transposes $S^\top:\Xi_{\mathrm{in}}\to\Xi_{\mathrm{out}}$ could be used equivalently.
\subsection{Standard Hilbert space representations and von Neumann algebras}\label{sec:stdrep}
In this section we investigate the kind of hybrid theory suggested by the {\it Hilbert space tensor product} of a classical and a quantum subsystem.
In the quantum case, there is no choice: von Neumann showed \cite{vNunique} that Schr{\"o}dinger's operators $P$ and $Q$ are the only solution of the commutation relations. As taught in every course on quantum mechanics, these live in the Hilbert space ${\mathcal H}_1=L^2({\mathbb R}^n,dq)$, with $Q_i$ acting by multiplication with the $i^{\rm th}$ coordinate, and $P_i$ by differentiating with respect to it (and a factor $i$). Equivalently, the Weyl operators are given by
\begin{equation}\label{WeylSchroe}
\bigl(W_1(q,p)\psi\bigr)(r)=e^{\frac{ip{\cdot} q}2-ip{\cdot} r}\psi(r-q), \quad\mbox{for } (q,p)\in\Xi_1,\ \psi\in{\mathcal H}_1=L^2({\mathbb R}^n,dr).
\end{equation}
In contrast, there is no such uniqueness for the classical case, basically because there are uncountably many inequivalent irreducible representations of the classical observable algebra (labelled by the points of $\Xi_0$). This non-uniqueness forces the choice of a measure $\mu$ on the classical subspace $\Xi_0$, so that the classical algebra is represented as the multiplication operators in $L^2(\Xi_0,\mu)$. A hybrid system can thus be set up in the tensor product, as defined in the following Def.~\ref{def:standard}. This kind of representation is the analog of the Schr{\"o}dinger\ representation: An explicit choice of Weyl operators satisfying the relations \eqref{weylrel}, initially without the claim that {\it all} good representations look like that. Indeed, it will be the next step to establish that claim, and hence the hybrid analog of von Neumann's result (Thm.~\ref{thm:unique} below).
\begin{defi}\label{def:standard}
Let $\Xi=\Xi_1\oplus\Xi_0$ be a hybrid phase space with antisymmetric form $\sigma=\sigma_1\oplus0$. Then a {\bf standard representation} is a representation of the Weyl relations in the Hilbert space ${\mathcal H}={\mathcal H}_1\otimes L^2(\Xi_0,\mu)$, where $\mu$ is some regular Borel measure on $\Xi_0$, and ${\mathcal H}_1$ is the Hilbert space of the Schr{\"o}dinger\ representation $W_1:\Xi_1\to{\mathcal B}({\mathcal H}_1)$ for $\Xis{_1}$. The Weyl operators are given by
\begin{equation}\label{standardrep}
W(\xi_1\oplus\xi_0)=W_1(\xi_1)\otimes W_0(\xi_0) ,
\end{equation}
where $W_0(\xi_0)$ is the multiplication operator
\begin{equation}\label{W0}
\bigl(W_0(\xi_0)\phi\bigr)(x)=e^{i\xi_0{\cdot} x} \phi(x)
\end{equation}
for $\phi\in L^2(\Xi_0,\mu)$ and $x\in\Xi_0$.
A {\it state} on the CCR-algebra is called {\bf standard} if it is given by a density operator on ${\mathcal H}$ in a standard representation.
\end{defi}
We remark that the standard representation depends on $\mu$ only up to equivalence. That is, when two measures $\mu$ and $\mu'$ have the same null sets, the Hilbert spaces $L^2(\Xi_0,\mu)$ and $L^2(\Xi_0,\mu')$ are the same by a unitary transformation that acts by multiplication (with $\sqrt{d\mu/d\mu'}$) and, in particular, intertwines the multiplication operators \eqref{W0}. We can, therefore, always choose $\mu$ to be a probability measure, typically the classical marginal of a state under consideration.
Note that the translate of a standard state is again standard, but generally not in the same representation, unless $\mu$ is quasi-invariant (equivalent to its translates). This is the case just for the Lebesgue measure and will be discussed further in Sect.~\ref{sec:transState}.
The {\bf von Neumann algebra} generated by a standard representation is
\begin{equation}\label{vNmu}
{\mathcal M}_\mu={\mathcal B}({\mathcal H}_1){\overset{{\rule{5pt}{1pt}}}\otimes} L^\infty(\Xi_0,\mu),
\end{equation}
where ${\overset{{\rule{5pt}{1pt}}}\otimes}$ denotes the tensor product of von Neumann algebras.
Indeed, since the $W_1(\xi_1)$ are irreducible on ${\mathcal H}_1$, they generate ${\mathcal B}({\mathcal H}_1)$ as a von Neumann algebra, and similarly the Weyl multiplication operators generate the maximal abelian algebra of all multiplication operators $M_f$ with $f\in L^\infty(\Xi_0,\mu)$, which is isomorphic to $L^\infty(\Xi_0,\mu)$. Putting this together, and using the commutation theorem for tensor products \cite[Thm.~IV.5.9]{takesaki1} gives \eqref{vNmu}. Note that this algebra still depends on $\mu$ because in $L^\infty(\Xi_0,\mu)$ functions, which only agree $\mu$-almost everywhere, are identified. By identifying $A\otimes f$ with the function $x\mapsto f(x)A$ we can think of the elements of ${\mathcal M}_\mu$ as measurable ${\mathcal B}({\mathcal H}_1)$-valued functions on $\Xi_0$.
We will later strive to get rid of the $\mu$-dependence in the definition of observable algebras, as is motivated at the beginning of Sect.~\ref{sec:funcobs}.
Standard states are thus normal states on some ${\mathcal M}_\mu$, hence elements of the predual
${\mathcal T}^1({\mathcal H}_1)\otimes L^1(\Xi_0,\mu)$, where ${\mathcal T}^1({\mathcal H}_1)$ denotes the trace class. They can hence be decomposed as
\begin{equation}\label{disintegrate}
\omega(A\otimes f)= \braket{\omega}{A\otimes f} = \int\mu(dx)c(x)\ f(x)\operatorname{tr}(\rho_x A),
\end{equation}
where $c\mu$ is the probability measure determining the classical marginal, i.e., the expectations of multiplication operators, and $x\mapsto\rho_x$ is a measurable family of density operators. The factor $c(x)$ is introduced to allow that $\operatorname{tr}\rho_x=1$ for all $x$. When we consider a particular state and its GNS-representation, we usually take $\mu$ directly as the classical marginal of that state, i.e., set $c(x)\equiv1$.
The required measurability conditions for the family of states $\rho_x$ are spelled out in \cite[Sect.~IV.7]{takesaki1}.
The definition of standard states brings in a dependence on $\mu$, so that it is not a priori clear that convex combinations of standard states are standard. However, the integral decomposition \eqref{disintegrate} makes clear that for a countable convex combination $\rho=\sum_j\lambda_j\rho_j$ we can take $\mu=\sum_j\lambda_jc_j\mu_j$, and, set $h_j$ to be the Radon--Nikodym\ derivative of $\lambda_jc_j\mu_j$ with respect to $\mu$. Note that $0\leq h_j(x)\leq1$, and $\sum_jh_j=1$. Then $\rho_0=1$ and
$\rho_x=\sum_jh_j\rho_{x,j}$. In particular, a normal state in a direct sum of standard representations can be rewritten as a state using just a single summand, i.e., is also standard in the sense of the above definition.
This argument also shows that the von Neumann algebra approach to hybrids can be made to work on larger and larger sets of states: If needed, one can consider any countable (and thereby any norm separable) family of states as absolutely continuous with respect to a common reference measure. However, the set of measures on $\Xi_0$ is not norm separable, so there is no single standard representation which can be used for {\it all} practical purposes. One could represent a single observable $F$ by a net of functions $F_\mu\in{\mathcal B}({\mathcal H})\otimes L^\infty(\Xi_0,\mu)$, each defined up to $\mu$-a.e. equality. Then indices are ordered by absolute continuity $\mu\ll\nu$, i.e., $\nu$ has fewer null sets than $\mu$, and in this case, $F_\nu$ is more sharply defined than $F_\mu$. There is no natural limit to such nets because we cannot include all the uncountably many point measures. However, the notion of universally measurable sets and functions (see Sect.~\ref{sec:funcobs}) does allow us to get rid of the Lebesgue completions.
Since standard states thus form a convex set, it makes sense to ask for the extreme points, i.e., the {\bf pure states}.
These are readily characterized:
\begin{lem}\label{lem:purestates}
A standard state $\omega$ on the CCR-algebra is extremal iff there is a point $x\in\Xi_0$ and a unit vector $\phi\in{\mathcal H}_1$ such that in the decomposition \eqref{disintegrate} $\mu=\delta_x$ is a point measure and $\rho_x=\kettbra\phi$.
\end{lem}
\begin{proof}
Suppose that $\omega$ is extremal. Then let $f\in L^\infty(\Xi_0)$ with $\veps<f<\idty-\veps$ for some $\veps>0$. $\omega$ is then decomposed into the sum of two positive functionals
\begin{equation}\label{notextreme}
\omega(X)=\omega(f)\,\frac{\omega (fX)}{\omega (f)}+\omega(1-f)\,\frac{\omega((1-f)X)}{\omega(1-f)}.
\end{equation}
This is a convex combination of states, so by extremality, the two states have to be proportional, i.e.,
$\omega(fX)=\lambda\omega(X)$ for all $X$. This forces $\lambda=\omega(f)$, by putting $X=\idty$, and hence we conclude that $f=\omega(f)\idty$ almost everywhere with respect to $\mu$. Hence, $\mu$ is a point measure at some point~$x$, say. The choice of $\rho_y$ for $y\neq x$ is irrelevant because the whole complement of $\{x\}$ has measure zero. The state $\rho_x$ is now given by a density operator, which clearly has to be extremal as well, so $\rho_x=\kettbra\phi$.
\end{proof}
Note that a state $\omega$ may have no extremal components, i.e., no extreme points $\omega'$ such that $\omega\geq\lambda\omega'$ with $\lambda>0$. Indeed this will be the case whenever the measure $\mu$ has no atoms (points of non-zero measure). It is therefore not a priori clear in which sense standard states can be decomposed into extreme points. This will be clarified in Sect.~\ref{sec:CstarStates}, where it will be seen that the standard states are the state space of a certain C*-algebra, so the convex combinations of extreme points are dense in a suitable weak* topology.
\subsection{Hybrid Uniqueness Theorem}
It is straightforward to check that in the standard representation, $\xi\mapsto W(\xi)$ is continuous with respect to the strong operator topology. It turns out that this characterizes standard representations. This is the main content of the following theorem, which is very close in its formulation and its proof to von Neumann's famous result \cite{vNunique}.
\begin{thm}[Hybrid Uniqueness Theorem]\label{thm:unique}
Every representation of the Weyl relations on a Hilbert space, for which the mapping $\xi\mapsto W(\xi)$ is continuous in the strong operator topology, is unitarily equivalent to a direct sum of standard representations.
\end{thm}
In the literature, it is traditional \cite{Segal} to use a weaker continuity condition, which does not demand joint continuity of $W$ in all $2n+s$ variables in $\xi\in\Xi$, but only along one-dimensional subspaces. This is the minimum required to get self-adjoint canonical operators and is usually called ``regularity'' \cite{bratteli2,Honegger}. This weaker version avoids some of the topological subtleties of infinite-dimensional $\Xi$. In the finite-dimensional case, it is still sufficient.
\begin{proof}
Consider a strongly continuous representation $W$ on a Hilbert space ${\mathcal H}$. For the most part, we will only need to use the representation $\xi_1\mapsto W(\xi_1\oplus0)$ of the subgroup $\Xi_1$. Following von Neumann, and even his notation up to a factor $2\pi$, we introduce a Gaussian function $a:\Xi_1\to{\mathbb C}$ and the operator
\begin{equation}\label{vNA}
A=\int d\xi_1\ a(\xi_1) W(\xi_1\oplus0).
\end{equation}
The integral exists as a strong integral because $W$ is continuous. Because $a$ is integrable, $A$ is clearly a bounded operator. With von Neumann's choice, it is even a projection, and in the Schr{\"o}dinger\ representation, it is just the one-dimensional projection $\kettbra\Omega$ onto the harmonic oscillator ground state vector $\Omega\in{\mathcal H}_1$. Since algebraic relations between $A$ and anything in ${\mathrm{CCR}}\Xis{_1}$ are the same in any representation, it is hardly a surprise that we have
\begin{equation}\label{cNPAP}
A W(\xi_1\oplus0)A=\brAAket\Omega{W_1(\xi_1)}\Omega\,A=:\chi(\xi_1) A,
\end{equation}
where $W_1$ is the Schr{\"o}dinger\ representation and $\chi(\xi_1)=\exp(-1/4\,\xi_1^2)$ is a Gaussian, the characteristic function of the oscillator ground state. But, of course, one can also show this (as von Neumann does) by explicit computation based on the Weyl relations.
It is a key part of von Neumann's argument that $A$ cannot vanish for any continuous representation of $\Xi_1$. Indeed, in such a representation also $W(\eta\oplus0)A W(\eta\oplus0)^*$ would vanish for all $\eta$, which is exactly of the form \eqref{vNA}, with a kernel function $a_\eta(\xi_1)=\exp(i\eta{\cdot}\sigma\xi_1)a(\xi_1)$. As a function of $\eta$, the integral of the modified \eqref{vNA} is thus the Fourier transform of an operator-valued $L^1$-function, hence vanishes only if $a$ does, which is false.
Consider now the subspace ${\mathcal H}_0:=A{\mathcal H}$ and the set $M$ of vectors of the form $W(\xi_1\oplus0)\psi_0$ for $\xi_1\in\Xi_1$ and $\psi_0=A\psi_0\in{\mathcal H}_0$. We claim that its linear span is dense. If $\Psi\in{\mathcal H}$ were orthogonal to $M$, we would have that $\braket{AW(\xi_1\oplus0)\Psi}{A\psi_0} =0$ for all $\psi_0$, so $A$ would vanish on the cyclic sub-representation space of $\Xi_1$ generated by $\Psi$, contradicting von Neumann's result $A\neq0$.
We define a function $U:M\to{\mathcal H}_1\otimes{\mathcal H}_0$ by
\begin{equation}\label{UvNu}
UW(\xi_1\oplus0)\psi_0=W_1(\xi_1)\Omega\otimes\psi_0 .
\end{equation}
Scalar products between different vectors on $M$ are preserved, so, in particular, it sends a linear combination representing the null vector again to a linear combination with vanishing norm. That is, it extends to a linear operator on the algebraic linear span of $M$. Clearly, this extension is isometric as well, so extends by continuity to ${\mathcal H}$. To summarize, \eqref{UvNu} defines an isometry $U:{\mathcal H}\to{\mathcal H}_1\otimes{\mathcal H}_0$. It is also onto, because the vectors $W_1(\xi_1)\Omega$ span ${\mathcal H}_1$. Hence $U$ is unitary and $W(\xi_1\oplus\xi_0)=U^*W_1(\xi_1)\otimes W(0\oplus\xi_0)U$.
To complete the proof, it suffices to observe that the strongly continuous representation $\xi_0\mapsto W(0\oplus\xi_0)$ of the group $\Xi_0\cong{\mathbb R}^s$ on ${\mathcal H}_0$ can be decomposed into a direct sum of cyclic ones, and the cyclic representations are of the form given in Def.~\ref{def:standard}. To see that this decomposition works together correctly with von Neumann's construction for $\Xi_1$ was the main reason to include an abridged version of his argument.
\end{proof}
\subsection{Bochner's Theorem}
A state on the CCR-algebra is completely determined by its expectations on Weyl operators, hence by the function
\begin{equation}\label{chfunc}
\chi(\xi)=\omega(W(\xi)),
\end{equation}
which we call the {\bf characteristic function} of $\omega$. Thus it is a natural question which functions exactly arise in this way. This demands unifying two well-known results: The purely classical case of this is known as Bochner's Theorem (sometimes: Bochner--Khintchine Theorem \cite{holevo_probabilistic_book}). Its quantum analog was apparently first formulated by Araki \cite{araki}, with further relevant work by \cite{Kastler1965,loupias,loupias2,parthasarathy_what,evansLewis}.
Its hybrid version (also in \cite{Honegger}) is the following.
\begin{thm}[Hybrid Bochner Theorem] \label{thm:Bochner}
Let $\Xi$ be a vector space with antisymmetric form $\sigma$. Then a function $\chi:\Xi\to{\mathbb C}$ is the characteristic function of a {standard state} on ${\mathrm{CCR}}\Xis{}$ if and only if it is
\begin{itemize}
\item[(1)] continuous,
\item[(2)] normalized, $\chi(0)=1$, and
\item[(3)] $\sigma$-twisted positive definite, which means that, for any choice $\xi_1,\ldots,\xi_N$, the $N\times N$-matrix
\begin{equation}\label{twistedPosDef}
M_{k\ell}=\chi(\xi_k-\xi_\ell)\,e^{\textstyle \frac i2 \sigma(\xi_k,\xi_\ell)}
\end{equation}
is positive semi-definite.
\end{itemize}
\end{thm}
\begin{proof}
Just conditions (2) and (3) are equivalent to $\omega$ being a state on the CCR-algebra. Indeed, the positive definiteness condition is precisely equivalent to $\omega(A^*A)\geq0$, where
$A=\sum_i c_i W(\xi_i)$, and the Weyl relations are used. By the GNS-construction, every positive linear functional comes from a Hilbert space representation, and by definition of the CCR-algebra as the universal C*-algebra of the Weyl relations, the state thus extends to the whole algebra.
Continuity of $\chi$ for a standard $\omega$ is obvious because a standard representation is strongly continuous. Conversely, suppose that $\chi$ is continuous, and let $\Omega\in{\mathcal H}_\omega$ denote the cyclic vector of the GNS-representation of $\omega$. Then
\begin{align}\label{conti}
\xi\to&\braket{\pi_\omega\bigl(W(\eta_1)\bigr)\Omega}{\pi_\omega\bigl(W(\xi)\bigr)\ \pi_\omega\bigl(W(\eta_2)\bigr)\Omega}
= \braket{\Omega}{\pi_\omega\bigl( W(-\eta_1) W(\xi) W(\eta_2) \bigr)\ \Omega}\nonumber \\
&=\exp\left(-\frac i2 \bigl(\sigma(\xi,\eta_2) +\sigma(-\eta_1,\xi+\eta_2) \bigr) \right) \braket{\Omega}{\pi_\omega\bigl( W(\xi-\eta_1+\eta_2) \bigr)\ \Omega} \nonumber\\
&= \chi(\xi-\eta_1+\eta_2) \exp\left(-\frac i2 \bigl(\sigma(\xi,\eta_2) -\sigma(\eta_1,\xi) - \sigma(\eta_1,\eta_2) \bigr) \right)
\end{align}
is continuous. Since the Weyl operators are bounded, this extends to the norm limits of linear combinations of $\pi_\omega\bigl(W(\eta_2)\bigr)\Omega$ which is, by definition, all of ${\mathcal H}_\omega$. Hence $\pi_\omega(W({\cdot}))$ is weakly continuous, but for unitary operators, this is the same as strong continuity. Hence $\omega$ is normal in a strongly continuous representation.
By Thm.~\ref{thm:unique} this is a direct sum of standard representations, and by the argument preceding it, we conclude that $\omega$ itself is standard.
\end{proof}
To see the power of the continuity condition, it may be useful to point out some rather wild states of the CCR-algebra. Indeed, this algebra is just the hybrid version of the almost periodic functions, in the precise sense of Prop.~\ref{prop:qha}. Pure states on the almost periodic functions form the Bohr compactification of $\Xi_0$ \cite[Sect.~4.7]{FollandHarmonic}, among which the points of $\Xi_0$ (i.e., their point evaluations) are just a small part, and not even an open subset. This expresses the observation that almost periodic functions cannot distinguish a point from many others that are arbitrarily far away, so the finite and the infinite are intertwined more intimately than ``observables'' would ever distinguish. An algebra whose states are better behaved may be more adequate for physics. Even for quantum field theory, where the CCR-algebra has been used extensively (with infinite-dimensional $\Xi$), this need has been felt, and a recent proposal by Buchholz to consider not the exponentials of field operators but their resolvents \cite{buchholz1,buchholz2} (see Example~\ref{Ex:resAlg}), can be seen in this light. Certainly, this eliminates the extreme sensitivity to infinite values and, for example, allows the algebra to be invariant under typical quantum mechanical time evolutions \cite{buchholz1,buchholz2}.
This suggests finding a C*-algebra whose states are just the ``good'' ones described by Bochner's Theorem. This will be done in the next section.
\subsection{The standard states as a C*-state space}\label{sec:CstarStates}
The CCR-algebra is constructed so that its representations exactly correspond to the representations of the Weyl relations. In this correspondence, the topology of $\Xi$ plays no role at all. The way to set up a similar correspondence for just the {\it continuous} unitary representations is well known from the theory of locally compact groups: One goes to the convolution algebra over the group. In fact, the term ``group algebra'' of a group is usually reserved for the C*-envelope of the convolution algebra $L^1(G)$, and not for the topology-free analog of the CCR-algebra \cite[Ch.~13]{dixmier}. Von Neumann's proof uses the same idea by introducing the operator $A$ as an integral. In this section, we follow this lead.
This will require a twisted version of the group algebra construction \cite{EdLewis}. An alternative construction would be via the group C*-algebra of a related non-abelian group, a central extension \cite[Ch.~VII]{varadarajan} of the additive group $\Xi$, called the Heisenberg group. The approach used below is a bit more direct in that it avoids the introduction of the central phase parameter, which is, in the end, integrated out anyhow. After we finished this work, we realized that the idea had already been followed through by Grundling \cite{Grundling,Grundling2}, with much the same motivation of getting a C*-description of continuous representations, and even extended to more general groups, also beyond locally compact ones. To keep this paper self-contained, we nevertheless include our version.
For $h\in L^1(\Xi,d\xi)$ and any given measurable representation $W$ of the Weyl relations we write
\begin{equation}\label{Wred}
W[h]=\int d\xi\ h(\xi)W(\xi).
\end{equation}
The bracket notation indicates that $h\mapsto W[h]$ is closely related to the representation $W$. We can then describe the algebra of such operators directly in terms of operations on $L^1(\Xi,d\xi)$. The multiplication rule and adjoints for such operators follows directly from the Weyl relations, namely $W[h]W[g]=W[h{\ast_\sigma}g]$ and $W[h]^*=W[h^*]$ with
\begin{align}\label{twistedconvol}
\bigl(h{\ast_\sigma}g\bigr)(\xi)&=\int d\eta\ h(\xi-\eta)g(\eta)e^{-\frac i2\xi{\cdot}\sigma\eta}, \\
h^*(\xi)&=\overline{h(-\xi)}.
\end{align}
These operations turn $L^1(\Xi)$ into a Banach *-algebra, which we call the {\bf $\sigma$-twisted convolution algebra} of $\Xi$. Any set of elements $h_\veps$ such that $h_\veps\geq0$, $\int d\xi\, h_\veps(\xi)=1$, and $h_\veps(\xi)=0$ for $\xi$ outside a ball of radius $\veps$ around the origin is an {\bf approximate unit}. As in the untwisted case, this follows from the strong continuity of translations on $L_1(\Xi)$.
The enveloping C*-algebra of the convolution algebra will be called the {\bf twisted group algebra} of $\Xis{}$ and denoted by ${\mathrm C}^*\Xis{}$. This is defined \cite[Ch.~2, \S7]{DixC} as the completion in the norm
\begin{equation}\label{cstnorm}
\norm h=\sup_\pi\norm{\pi(h)}= \sup_\omega\omega\bigl(h^*\ast_\sigma h\bigr)^{1/2} ,
\end{equation}
where the supremum over $\pi$ runs over all *-representations of the algebra by Hilbert space operators, and $\omega$ runs over all positive linear functionals of norm $\leq1$. By slight abuse of notation, we denote the element in ${\mathrm C}^*\Xis{}$ associated with $h\in L^1(\Xi)$ by the completion process again by $h$. This is justified by the observation that the canonical embedding $L^1\Xis{}\hookrightarrow{\mathrm C}^*\Xis{}$ is injective.
\begin{prop}\label{prop:twgroup}
Let $\Xis{}$ be a hybrid phase space. Then every state $\omega$ on ${\mathrm C}^*\Xis{}$ is given by a unique standard state $\omega'$ on ${\mathrm{CCR}}\Xis{}$ and conversely, such that
\begin{equation}\label{wwred}
\omega\bigl(h\bigr)=\int d\xi\ h(\xi)\,\omega'\bigl(W(\xi)\bigr).
\end{equation}
\end{prop}
This proposition gives us the third way of looking at Weyl elements $W(\xi)$. At first, they were defined as explicit operators in any standard representation. Secondly, they appeared as the abstract generators of a CCR-algebra. These two views are equivalent by virtue of Bochner's Theorem, which identifies standard states with linear functionals $\omega'$ on ${\mathrm{CCR}}\Xis{}$. The above proposition allows us to further introduce, for each $\xi$, the linear functional
$\omega\mapsto \omega'(W(\xi))$. This element of the bidual ${\mathrm C}^*\Xis{}^{**}$ is yet another version of the Weyl element, which we will also denote by $W(\xi)$.
It is clear that we cannot realize such an element in ${\mathrm C}^*\Xis{}$, since this algebra has no unit and hence contains no unitary elements. However, we can get close in the same sense the expression \eqref{Wred} can be close to $W(\xi)$, if $h$ is concentrated near $\xi$. The idea of the following proof is to do this limit in the GNS representation of $\omega$.
\begin{proof}
\def\Hgns{\widetilde{\mathcal H}}%
It is a general feature of the enveloping C*-algebra construction \cite[Prop.\,2.7.5]{DixC} that the states $\omega$ on ${\mathrm C}^*\Xis{}$ are in bijective correspondence to the positive linear functionals $\widetilde\omega$ on the convolution algebra with norm $1$. Here the norm is taken as a linear functional on the Banach space $L^1(\Xi)$. That is, there is a function $\chi\in L^\infty(\Xi)$ with $\norm\chi_\infty=1$ such that
\begin{equation}\label{tilw}
\widetilde\omega(h)=\int d\xi\ h(\xi)\chi(\xi)
\end{equation}
The main task of the proof is to show that the functions $\chi$ arising in this way are exactly the characteristic functions characterized by the Bochner Theorem, and in particular {\it continuous}. The uniqueness of the correspondence is clear from this equation since, on the one hand, it gives an explicit formula for $\omega$ (resp.\ $\widetilde\omega$) in terms of $\omega'$, and, on the other, two states $\omega',\omega''$ satisfying it for the same $\widetilde\omega$ would have to be equal as elements of $L^\infty(\Xi)$, hence equal almost everywhere, and hence equal by continuity.
We begin by defining a version of the Weyl operators acting on $L^1(\Xi)$, namely
\begin{equation}\label{leftWeyl}
(\widetilde W(\xi) h)(\eta):=e^{-\frac i2 \xi{\cdot}\sigma\eta}h(\eta-\xi).
\end{equation}
It is constructed so that
\begin{equation}\label{leftWeylConvolve}
h{\ast_\sigma}g=\int d\xi\ h(\xi)\, \widetilde W(\xi) g.
\end{equation}
Intuitively, we can think of $\widetilde W(\xi)$ as the operator of convolution with $\delta_\xi$, the limit of probability densities concentrated near the point $\xi$. While this is not an element of the algebra, its operation is defined analogously to the approximate unit $\delta_0$.
It is easy to check that the operators $\widetilde W(\xi)$ satisfy the Weyl multiplication rules \eqref{weylrel}. However, unitarity does not make sense since $L^1(\Xi)$ is not a Hilbert space. The crucial observation is that $\xi\mapsto \widetilde W(\xi)g$ is continuous in the norm of $L^1(\Xi)$. Indeed, it is a product of a translation and a multiplication operator, which are both strongly continuous on $L^1$.
Consider now a positive linear functional $\widetilde\omega$ on $L^1(\Xi)$. Its GNS representation space $\Hgns$ is the unique Hilbert space generated by vectors $v(h)$, $h\in L^1(\Xi)$, with the scalar product $\brAket{v(h)}{v(k)}=\widetilde\omega(h^*\ast_\sigma k)$. On these, the representation $W:L^1(\Xi)\to{\mathcal B}(\Hgns)$ acts by left multiplication in the convolution algebra, i.e., according to the formula
\begin{equation}\label{Wgns}
W[h]v(g)=v\bigl(h\ast_\sigma g\bigr).
\end{equation}
According to \cite[I.9.14]{takesaki1} the GNS space has a cyclic vector $\Omega$ and a representation $W:L^1(\Xi)\to{\mathcal B}(\Hgns)$ such that $v(h)=W[h]\Omega$, and $\braket\Omega{W[h]\Omega}=\widetilde\omega(h)$. Indeed, one has $\Omega=\lim_{\veps\to0}v(h_\veps)$, where $h_\veps$ is a bounded approximate unit.
Our next aim is to show that $W$ arises exactly as in \eqref{Wred} from the integration of a representation $W$ of the Weyl relations (recall that the two functions will be typographically distinguished by their argument brackets). The obvious candidate for the Weyl operators $W(\cdot)$ are the operators $\widetilde W(\xi)$ from \eqref{leftWeyl}, represented on $\Hgns$ in GNS style. That is, in analogy to \eqref{Wgns} we set
\begin{equation}\label{Wtgns}
W(\xi)v(g)=v\bigl(\widetilde W(\xi) g\bigr).
\end{equation}
Then it is elementary to check that $W(\xi)$ is unitary, and these operators satisfy the Weyl relations. Moreover, the $L^1$-norm continuity of $\xi\mapsto \widetilde W(\xi)g$ established earlier implies that
$\xi\mapsto W(\xi)$ is continuous in the strong operator topology. Finally, \eqref{leftWeylConvolve} implies
\begin{equation*}
W[h]v(g)=v\bigl(h\ast_\sigma g\bigr)
=\int d\xi\ h(\xi)v\bigl(\widetilde W(\xi)g\bigr)
=\int d\xi\ h(\xi) W(\xi)v\bigl(g\bigr),
\end{equation*}
i.e., the GNS-representation $W[h]$ is related to the continuous representation $W$ by \eqref{Wred}. In particular
\begin{equation*}
\widetilde\omega(h)=\int d\xi\ h(\xi)\braket\Omega{W(\xi)\Omega},
\end{equation*}
so \eqref{tilw} holds with $\chi(\xi)=\braket\Omega{W(\xi)\Omega}$, which is clearly a normalized twisted positive definite function, and continuous because $W(\xi)$ is strongly continuous.
Conversely, given a state on the CCR-algebra, we can define its characteristic function $\chi(\xi)=\omega'(W(\xi))$. In general, that might fail to be even measurable, so the formula \eqref{tilw} might make no sense. For standard states, however, $\chi$ is continuous, and the integral is well defined.
\end{proof}
We now determine the algebras ${\mathrm C}^*\Xis{}$ concretely. It turns out that this is best done by splitting into a purely classical and a purely quantum part.
\begin{prop}\label{prop10tensor}
Let $\Xis{}$ be a hybrid phase space, split as $\Xi=\Xi_1\oplus\Xi_0$ with $\sigma=\sigma_1\oplus0$. Then
\begin{equation}\label{CsXi}
{\mathrm C}^*\Xis{}={\mathrm C}^*\Xis{_1}\otimes{\mathrm C}^*(\Xi_0,0) \cong {\mathcal K}({\mathcal H}_1)\otimes{\mathcal C}_0(\Xi_0),
\end{equation}
where ${\mathcal K}({\mathcal H}_1)$ denotes the compact operators on the representation space ${\mathcal H}_1$ of the irreducible quantum system $\Xis{_1}$, and ${\mathcal C}_0(\Xi_0)$ denotes the continuous functions on $\Xi_0$ vanishing at infinity.
\end{prop}
\begin{proof}
We first observe that for the underlying $L^1$-spaces, the direct sum naturally coincides with the projective product, which is predual to the tensor product of von Neumann algebras. That is
\begin{equation}\label{L1oplustimes}
L^1(\Xi_1\oplus \Xi_0)=L^1(\Xi_1)\otimes L^1(\Xi_0).
\end{equation}
Indeed, the tensor products $f\otimes g$ on the right hand side can be identified with the product functions $fg(\xi_0\oplus\xi_1)=f(\xi_1)g(\xi_0)$, and this embedding is clearly isometric on step functions. Since the measurable structure of $\Xi_1\oplus\Xi_0$, which is $\Xi_1\times\Xi_0$ as a set, is defined as generated by rectangles, the product functions span a dense subspace of $L^1(\Xi_1\oplus\Xi_0)$. It is elementary to verify that the isomorphism \eqref{L1oplustimes} is also consistent with the definitions of adjoint operation and convolution product.
The completion in the construction of the enveloping C*-algebra also works out:
As one side of the tensor product is abelian, and the maximal and the minimal tensor product coincide, the tensor product is uniquely determined, as is the algebra in tensor product form.
In the second step, we need to show the claimed isomorphisms: Beginning with the classical case, $L^1(\Xi_0)$ is the convolution algebra of $\Xi_0\cong{\mathbb R}^s$. Its irreducible representations are described by the Gelfand isomorphism for abelian Banach algebras: In this case they are given by the point evaluations of the Fourier transform. Hence the C*-norm of the enveloping algebra, $\norm h=\sup_\pi\norm{\pi[h]}$ is equal to the supremum norm of the Fourier transform of $h$. Now by the Riemann--Lebegue Lemma, the Fourier transforms of $L^1$-functions are continuous and go to zero at infinity. On the other hand, by the Stone--Weierstra\ss{} Theorem, these Fourier transforms separate points and are hence uniformly dense. Hence ${\mathrm C}^*(\Xi_0)={\mathcal C}_0(\Xi_0)$.
For quantum systems, note that every continuous representation of the Weyl relations is isomorphic to the Schr{\"o}dinger\ representation on ${\mathcal H}_1$ by von Neumann's Uniqueness Theorem. Hence we only need to show that in that representation, the operators of the form $W[h]$ with $h\in L^1(\Xi_1)$ are compact, and these operators form a dense subalgebra of ${\mathcal K}({\mathcal H}_1)$. This follows immediately by the correspondence theory \cite[Cor.~5.1.(4)]{QHA}. An alternative approach using better known facts goes via first showing that operators $h\mapsto W[h]$ are not only continuous from $L^1$ to ${\mathcal B}({\mathcal H}_1)$ but also an isometry for the $2$-norms, i.e. $\norm{W[h]}_2=\norm h_2$, for $h\in L^1(\Xi)\cap L^2(\Xi)\equiv{\mathcal A}$, and the Schatten $2$-norm (Schmidt norm) on the operator side. Hence $W[{\mathcal A}]$ consists of Hilbert--Schmidt operators, which are compact. By taking limits in $2$-norm, we find that $W[{\mathcal A}]$ is operator norm dense in the Hilbert--Schmidt class, hence in ${\mathcal K}({\mathcal H}_1)$.
\end{proof}
We note that as a consequence of this characterization, we find that there are many extremal standard states since the state space of ${\mathrm C}^*\Xis{}$ is the weak*-closed convex hull of its extreme points, by the Banach--Alaoglu and Krein--Milman Theorems. Of course, these were already identified in Lem.~\ref{lem:purestates}.
\subsection{Restoring translation symmetry
Translations were part of our basic setup from the outset since the phase space $\Xi$ is a vector space. The notion of standard representations (Def.~\ref{def:standard}) breaks the translation symmetry. However, it is restored in the twisted convolution construction. Indeed, combining \eqref{alpha} and \eqref{weylop} we get $\alpha_\eta(W(\xi))=\exp(i\eta{\cdot}\xi)W(\xi)$. Although Weyl operators are not themselves in ${\mathrm C}^*\Xis{}$ we think of this algebra as generated by integrated Weyl operators \eqref{Wred}, and so we must define
\begin{equation}\label{redshift}
(\alpha_\eta h)(\xi)=e^{i \eta{\cdot}\xi} h(\xi).
\end{equation}
Of course, this extends to the enveloping algebra ${\mathrm C}^*\Xis{}$. In the tensor product structure of Prop.~\ref{prop10tensor} we can apply this separately to the classical parts, so $\alpha_{\eta_1\oplus\eta_0}=\alpha_{\eta_1}\alpha_{\eta_0}$. On the classical part ${\mathcal C}_0(\Xi_0)$ the action becomes the shift $(\alpha_{\eta_0}f)(\xi_0)=f(\xi_0+\eta_0)$. Similarly, we can compute the action on the quantum part, finding
\begin{equation}\label{shiftQ}
\alpha_\eta(X)=W(\sigma\eta)^*XW(\sigma\eta).
\end{equation}
In this expression, we use $\sigma$ as a matrix acting on the vector $\eta\in\Xi$, which is possible because we choose a fixed basis in $\Xi\cong{\mathbb R}^{2n+s}$ (cp.\ Sect.~\ref{sec:setup}). Since $\sigma$ vanishes on the classical part, the component $\eta_0$ of the translation argument automatically drops out, and only the quantum Weyl operators are used.
The tensor product ${\mathcal K}({\mathcal H}_1)\otimes{\mathcal C}_0(\Xi_0)$ can be considered as the algebra of norm continuous functions $F:\Xi_0\to{\mathcal K}({\mathcal H}_1)$ vanishing at infinity, by identifying $K\otimes f$ with the function $F(\xi)=f(\xi)K$. In this ``function form'', which will later extend to certain subspaces of
${\mathrm C}^*\Xis{}^{**}$ the action of translations becomes, for $\eta=\eta_1\oplus\eta_0$,
\begin{equation}\label{shiftQF}
\bigl(\alpha_{\eta}(F)\bigr)(\xi_0)=W(\sigma\eta)^*F(\xi_0+\eta_0)W(\sigma\eta).
\end{equation}
\section{Observables as functions}\label{sec:funcobs}
The aim of this section is to complement the description of hybrid states by spaces of observables. Any standard representation in the sense of Def.~\ref{def:standard} gives us a natural observable algebra, the von Neumann algebra ${\mathcal B}({\mathcal H}_1){\overset{{\rule{5pt}{1pt}}}\otimes} L^\infty(\Xi_0,\mu)$, in which observables are ${\mathcal B}({\mathcal H}_1)$-valued functions on $\Xi_0$. The catch is that this depends on $\mu$, and any choice of $\mu$ excludes some states. For example, with the Lebesgue measure for $\mu$, we exclude all pure states, and while we can add countably many point measures to $\mu$, we would still miss uncountably many others. When it comes to quasifree channels, a $\mu$-dependent description brings in additional assumptions, so it becomes much more cumbersome to formulate results that hold for all quasifree channels. In contrast, our description of states on ${\mathrm C}^*\Xis{}$ is already free of such constraints, and we will now develop a matching description of observables and, later, of channels. The advantage of the $\mu$-free point of view is best illustrated by the (otherwise not rewarding) exercise of translating our Sect.~\ref{sec:channels} to a $\mu$-dependent setting.
So we have two complementary points of view:
\begin{labeledlist}{l}
\item[\textit{The setting with fixed $\mu$}] Here, we have a natural Hilbert space ${\mathcal H}_\mu={\mathcal H}_1\otimes L^2(\Xi_0,\mu)$, where ${\mathcal H}_1$ is the Hilbert space of the Schr{\"o}dinger\ representation of the quantum part $\Xi_1$. It carries a standard representation (in the sense of Def.~\ref{def:standard}) of the Weyl operators and the basic C*-algebra ${\mathrm C}^*\Xis{}={\mathcal K}({\mathcal H}_1)\otimes{\mathcal C}_0(\Xi_0)$. The density operators on ${\mathcal H}_\mu$ precisely give those states whose distribution for $\Xi_0$ is absolutely continuous with respect to $\mu$, i.e., states in ${\mathcal T}^1({\mathcal H}_1)\otimes L^1(\Xi_0,\mu)$. Every state $\omega\in {\mathrm C}^*\Xis{}^*$ is represented in such a structure, but unless we want to go to non-separable Hilbert spaces, like the direct sum of {\it all} such ${\mathcal H}_\mu$, every standard representation misses many states.
\item[\textit{The $\mu$-free setting}] This is the point of view based on the C*-algebra ${\mathcal A}={\mathrm C}^*\Xis{}$, allowing all states of ${\mathcal A}$. The maximal space of observables, for which these states provide probability distributions, is, by definition, the bidual ${\mathcal A}^{**}$. To get the connection with the $\mu$-dependent view, consider the map
$i_\mu:{\mathcal T}^1({\mathcal H}_1)\otimes L^1(\Xi_0,\mu)\to{\mathrm C}^*\Xis{}^*$, which identifies the states in the $\mu$-dependent view with the states in the $\mu$-free setting. The adjoint of the embedding is then the restriction map, the representation
\begin{equation}\label{restrictmu}
i_\mu^*:{\mathrm C}^*\Xis{}^{**}\to {\mathcal B}({\mathcal H}_1){\overset{{\rule{5pt}{1pt}}}\otimes} L^\infty(\Xi_0,\mu)\subset {\mathcal B}\bigl({\mathcal H}_1\otimes L^2(\Xi_0,\mu)\bigr).
\end{equation}
This representation destroys all information in $A\in {\mathrm C}^*\Xis{}^{**}$, which is irrelevant for states absolutely continuous with respect to $\mu$, i.e., identifies functions coinciding $\mu$-almost everywhere.
\end{labeledlist}
\noindent
Neither of the two obvious choices for observable spaces in the $\mu$-free setting is feasible. The largest choice mentioned above is the second dual ${\mathrm C}^*\Xis{}^{**}$ which is in many ways ``too large'' so that individual elements often have no explicit description. The difficulty here arises mostly from the classical part: While the second dual of the compact operators is just the space of all bounded operators, the second dual ${\mathcal C}_0(X)^{**}$ is a rather complex object \cite{kaplan}. Elements of this space are not functions on $X$, but on a related, much larger Stonean topological space $\widehat X$. Its points are the extreme points of the normalized positive elements in the triple dual ${\mathrm C}^*\Xis{}^{***}$, another highly non-constructive object. Hence, while the elements in ${\mathrm C}^*\Xis{}^{**}$ admit a function representation, it is impossible to explicitly describe even a single point of the classical variable space on which they are supposed to be ``functions''. At the other end, for small algebras, we have ${\mathrm C}^*\Xis{}$ itself. This is in many ways ``too small''. Indeed, ${\mathrm C}^*\Xis{}$ does not have an identity, which is needed for the physical interpretation as an observable algebra. Also, it does not allow quantum operators with a continuous spectrum, barring the Weyl operators themselves. Therefore, we will have to choose some intermediate algebra ${\mathcal M}$ with ${\mathrm C}^*\Xis{}\subset {\mathcal M}\subset{\mathrm C}^*\Xis{}^{**}$. The criteria for this choice are simple:
\begin{itemize}
\item ${\mathcal M}$ should be constructed in a way that makes sense for every hybrid system.
\item Applying a quasifree channel ${\mathcal T}$ to an observable in ${\mathcal M}_{\mathrm{out}}$ for the output system should give an observable in ${\mathcal M}_{\mathrm{in}}$.
\end{itemize}
Roughly speaking, ${\mathcal M}$ will describe a degree of regularity for observables, which is preserved by all quasifree channels, leading to an automatic Heisenberg picture between the corresponding observable algebras. This could be expressed as regularity properties of operator-valued functions on the classical phase space, but such an approach introduces many case distinctions for proofs of the second of the above criteria: It depends on how inputs and outputs are split into classical and quantum parts, and how these splits are reshuffled by a quasifree channel. It turns out to be much more efficient to work with constructions that apply to classical, quantum, and hybrid systems alike. Here we follow the path of a functional analyst doing analysis, namely Gert Pedersen, whose slogan for his textbook ``Analysis NOW'' \cite{PedAnal} also stands for ``Analysis based on Norms, Operators, and Weak topologies''. The following section is based on his seminal early work and in spite of its relative abstractness, leads painlessly to just the Heisenberg picture characterizations we wanted to see.
\subsection{Semicontinuity in C*-algebras}\label{sec:semicont}
The constructions in this section are inspired by the commutative case \cite{kaplan}, but have been generalized to arbitrary C*-algebras in \cite{peder,pederC,brown} (see also \cite[III, \S6]{takesaki1} for a textbook version). The commutation relations are not needed for this so that we will consider first a general C*-algebra ${\mathcal A}$, typically without a unit, and only in the next section specialize to hybrids, i.e., ${\mathcal A}={\mathcal C}_0(X,{\mathcal K}({\mathcal H}))$. The standard states are then elements of the dual, and we are interested in well-behaved subalgebras of the bidual. A special role will be played by the pure states of ${\mathcal A}$.
For the various dualities between spaces of states and spaces of observables, we use the following notation: $\braket\omega A$ will be the expectation value of the observable $A$ in the state, where $\omega\in{\mathcal A}^*$ and $A\in{\mathcal A}$ or $A\in{\mathcal A}^{**}$, and this is extended to the whole linear spaces. We thereby identify ${\mathcal A}$ with a subspace of ${\mathcal A}^{**}$. By definition, the state space of ${\mathcal A}$ is the set of positive linear functionals of norm $1$.
Even wenn ${\mathcal A}$ has no unit, this still can be written as $\braket\omega\idty=1$, but then $\idty\in{\mathcal A}^{**}$. The weak*-topology on ${\mathcal A}^*$ is the topology making all functionals $\omega\mapsto\braket\omega A$ with $A\in{\mathcal A}$ continuous, whereas the weak-topology is similarly defined with $A\in{\mathcal A}^{**}$.
By the Banach--Alaoglu Theorem, the ``quasi-state space'' $Q=\{\omega\in{\mathcal A}^*|\omega\geq0,\norm\omega\leq1\}$ is weak*-compact, so by the Krein--Milman Theorem, $Q$ is the closed convex hull of its extreme points. Thus there are many extremal, that is ``pure'' states, and one additional extreme point $0\in Q$: When $\idty\notin{\mathcal A}$ the normalization functional is not weak*-continuous, so the state space is not weak*-compact, and there will be sequences of pure states converging to $0$.
The second dual ${\mathcal A}^{**}$ can be identified with its ``enveloping von Neumann algebra'', which is the von Neumann algebra generated by ${\mathcal A}$ in its universal representation \cite[Ch.~III.6]{takesaki1}. This is simply the direct sum of all GNS-representations of ${\mathcal A}$. Its center is the natural arena for the representation theory of ${\mathcal A}$ in the following sense: For every representation $\pi:{\mathcal A}\to{\mathcal B}({\mathcal H}_\pi)$ there is a central projection $z_\pi\in{\mathcal A}^{**}$ such that $\pi({\mathcal A})''\cong z_\pi{\mathcal A}^{**}$ as a von Neumann algebra. Representations $\pi_1,\pi_2$ which can be connected to each other by an isomorphism from $\pi_1({\mathcal A})''$ to $\pi_1({\mathcal A})''$ are called quasi-equivalent. So the central projections of ${\mathcal A}^{**}$ classify representations up to quasi-equivalence \cite[Thm.~III.2.12]{takesaki1}.
Consider now a pure state $\omega\in{\mathcal A}^*$. Its GNS-representation is irreducible, and hence the corresponding central projection $z_\pi$ is minimal, meaning that there is no projection $p$ in the center of ${\mathcal A}^{**}$, other than $0$ and $z_\pi$, such that $0\leq p\leq z_\pi$. Minimal projections are also called atoms of the projection lattice. Then the minimal projections of ${\mathcal A}^{**}$ correspond exactly to the pure states on ${\mathcal A}$. We record this simple observation as a statement for arbitrary von Neumann algebras (cp.\ \cite{Pederatomic}).
\begin{lem}\label{lem:atomic}
Let ${\mathcal M}$ be a von Neumann algebra. Then there is a one-to-one correspondence between minimal projections $p\in{\mathcal M}$ and extremal normal states $\omega\in{\mathcal M}_*$, given by
\begin{equation}\label{papMinimal}
pxp=\omega(x)p, \quad\mbox{for all\ } x\in{\mathcal M}.
\end{equation}
\end{lem}
\begin{proof}
For any projection $p$ we consider the von Neumann subalgebra $\widetilde{\mathcal M}=p{\mathcal M} p$. The crucial issue is whether this algebra is one-dimensional.
Now, $p$ is minimal iff, for any projection $q$, $0\leq q\leq p$ implies $q=0$ or $q=p$. This is equivalent to $0$ and $p$ being the only projections in $\widetilde{\mathcal M}$, i.e., to $\dim\widetilde{\mathcal M}=1$. Then also $pxp$ must be a multiple of $p$, i.e., $pxp=\omega(x)p$, and this functional is necessarily a normal state. It is also pure, because from
$\omega=\lambda\omega_1+(1-\lambda)\omega_2$ we conclude $\omega_1\leq\lambda^{-1}\omega$, hence
\begin{equation}\label{orderideal}
\abs{\braket{\omega_1}{x(1-p)}}\leq\braket{\omega_1}{(1-p)x^*x(1-p)}
\leq \lambda^{-1}\braket{\omega}{(1-p)x^*x(1-p)}=0.
\end{equation}
But then $\braket{\omega_1}{x}=\braket{\omega_1}{pxp}=\braket\omega x\braket{\omega_1}p$ and $\braket{\omega_1}p=1$ by choosing $x=\idty$ in this equation. Hence $\omega_1=\omega$, so $\omega$ is extremal.
Conversely, let $\omega$ be pure and normal, and let $p$ be its support, i.e., the smallest projection such that $\braket{\omega}p=1$. Then $\omega$ restricted to $\widetilde{\mathcal M}$ is also pure and, in addition, faithful, i.e., $x\in\widetilde{\mathcal M}$ with $\braket\omega{x^*x}=0$ implies $x=0$. Indeed, the eigenprojection $q$ of $x^*x$ for the spectral set $\{0\}$ must then satisfy $\braket\omega q=1$, and hence either $q=p$. By definition of the support projection this means $q=p=\idty_{\widetilde{\mathcal M}}$, and hence $x^*x=0$.
So consider the GNS-representation $\pi_\omega$ of a faithful normal pure state. Faithfulness implies that the representation is injective, and general representation theorems \cite[1.16.2]{sakai} imply that the image $\pi_\omega(\widetilde{\mathcal M})$ is a von Neumann algebra.
By purity, $\pi_\omega(\widetilde{\mathcal M})$ is irreducible, so $\pi_\omega(\widetilde{\mathcal M})={\mathcal B}({\mathcal H}_\omega)$. On the other hand, $\omega$ is given by a vector $\Omega\in{\mathcal H}_\omega$. So, unless $\dim{\mathcal H}_\omega=1$, there is a vector orthogonal to it and hence a non-zero element $x$ with $\pi_\omega(x)\Omega=0$, and hence $\braket\omega{x^*x}=0$, contradicting faithfulness. Hence $\dim{\mathcal H}_\omega=1$, and $\dim\widetilde{\mathcal M}=1$.
\end{proof}
For many von Neumann algebras, this is a statement about the empty set, namely when ${\mathcal M}$ is of type II or III, or $L^\infty(X,\mu)$ when there are no points with positive $\mu$-measure. However, for a second dual, there are many extreme points. Their central cover is the smallest central projection $z_{\mathfrak a}$, so that $p\leq z_{\mathfrak a}$ for all minimal projections, and $z_{\mathfrak a}{\mathcal A}^{**}$ is isomorphic to the von Neumann algebra generated by the uncountable direct sum of all normal pure state representations, called the {\it atomic} representation of ${\mathcal A}^{**}$. This is the part that will be useful for a function representation.
Indeed, consider for a moment the classical case ${\mathcal A}={\mathcal C}_0(X)$. Then there is a simple way to associate with an element $A\in{\mathcal A}^{**}$ a function $\check A$ on $X$, namely to evaluate $A$ in the pure state $\delta_x$, the point measure at $x\in X$, setting $\check A(x)=\delta_x(A)$. However, when $A$ has support in the complement of the atomic subspace (also called the diffuse subspace) we get $\delta_x(A)p=pAp=0$, where $p$ is the projection associated with $\delta_x$ via \eqref{papMinimal}. Hence $\check A=0$, so the function $\check A$ has nothing to say about $A$.
Nevertheless, for suitable subalgebras of ${\mathcal A}^{**}$ the atomic representation, and hence the function representation $\check A$ contains full information. The idea of \cite{peder} is to use monotone limits to construct useful algebras with this property. Since these constructions work in the same way in arbitrary C*-algebras, they also serve to provide a Heisenberg picture for general dual channels.
In ${\mathcal A}^{**}$ bounded, increasing nets are automatically weak*-convergent, and if this algebra is represented on a Hilbert space, the limits exist in the strong operator topology. This makes most sense in the hermitian part ${\mathcal A}^{**}_h$ of ${\mathcal A}^{**}$. For any subset $M\subset{\mathcal A}^{**}_h$ we denote by $M_\uparrow$ the set of limit points of such nets from $M$. Similarly, $M_\downarrow=-(-M_\uparrow)$ represents the limits of decreasing nets from~$M$.
\begin{defi}\label{def:multip}
Let ${\mathcal A}$ be a C*-algebra. Then
\begin{itemize}
\item The {\bf multiplier algebra} of ${\mathcal A}$, denoted by $M({\mathcal A})$, is the set of elements $m\in{\mathcal A}^{**}$ such that, for all $a\in{\mathcal A}$, $ma\in{\mathcal A}$ and $am\in{\mathcal A}$.
\item ${\mathcal A}_\uparrow:=({\mathcal A}_h+{\mathbb R}\idty)_\uparrow$ is called the {\bf lower semicontinuous} cone of ${\mathcal A}_h^{**}$. The upper semicontinuous cone is ${\mathcal A}_\downarrow=({\mathcal A}_h+{\mathbb R}\idty)_\downarrow$
\item An element $a\in{\mathcal A}^{**}_h$ is called {\bf universally measurable} if, for every state $\omega\in{\mathcal A}^*$, and every $\veps>0$, there are $x\in{\mathcal A}_\uparrow$, $y\in{\mathcal A}_\downarrow$ such that
$x\leq a\leq y$, and $\omega(y-x)<\veps$. The real vector space of universally measurable elements is denoted by ${\mathcal U}({\mathcal A})$.
\end{itemize}
\end{defi}
We remark that there are some subtle distinctions in defining the semicontinuous cone, depending on whether the unit is adjoined first (as above) and on whether a norm closure of the cone is taken. These are discussed carefully in \cite{pederC,brown}. The main observations for us are that ${\mathcal A}_\uparrow\cup{\mathcal A}_\downarrow=M({\mathcal A})_h$ \cite[Thm.~III.6.24]{takesaki1}, and that the atomic representation is isometric on ${\mathcal U}({\mathcal A})$ \cite[Thm.~III.6.37]{takesaki1}.
In the classical case, ${\mathcal A}={\mathcal C}_0(X)$ with $X$ locally compact, the lower semicontinuous cone consists just of the bounded lower semicontinuous functions $f$ in the sense of point set topology (lower level sets $\{x|f(x)\leq a\}$ are closed). The multipliers are $M({\mathcal A})={{\mathcal C}_{\mathrm b}}(X)$, all bounded continuous functions. For the universally measurable functions, note that, for a fixed measure $\mu\in{\mathcal C}_0(X)^*$, by definition, all bounded Borel measurable functions can be integrated. However, one usually completes the Borel algebra by including all $\mu$-null sets. The completion can be understood by adding all sets which can be approximated from above and below by Borel measurable sets, whose $\mu$-volume differs by arbitrarily little. The completion construction depends on $\mu$, but some sets will be added for all $\mu$, and these are called universally measurable \cite{Cohn}. The functions that are measurable for the completed $\sigma$-algebra are called $\mu$-measurable, and their classes up to $\mu$-a.e.\ equality form $L^\infty(X,\mu)$. The approximation from above and below for defining $\mu$-measurable sets has its counterpart for functions in the definition given above, with fixed $\mu=\omega$. Hence the universally measurable functions are those that are $\mu$-measurable for all $\mu$.
For these subsets of observables, there is an automatic Heisenberg picture for channels defined on states:
\begin{lem}\label{lem:autoHeisen1}
Let ${\mathcal A}$ and ${\mathcal B}$ be C*-algebras, and $T:{\mathcal A}^*\to{\mathcal B}^*$ a linear map taking states to states. Let $T^*:{\mathcal B}^{**}\to{\mathcal A}^{**}$. Then the inclusions
$T^*M({\mathcal B})\subset M({\mathcal A})$, $T^*{\mathcal B}_\uparrow \subset{\mathcal A}_\uparrow$ and $T^*{\mathcal U}({\mathcal B})\subset{\mathcal U}({\mathcal A})$ hold.
\end{lem}
\begin{proof}
Dual channels $T^*:{\mathcal B}^{**}\to{\mathcal A}^{**}$ preserve positivity and normalization. The latter condition can be written as $T^*\idty=\idty$. They map increasing nets to increasing nets and are continuous for the respective limits.
Hence $T^*{\mathcal B}_\uparrow \subset{\mathcal A}_\uparrow$. Then the characterization of multipliers as both upper and lower continuous shows $T^*M({\mathcal B})\subset M({\mathcal A})$. This is actually not so obvious just from the definition of multipliers.
For the universally measurable class we proceed directly: Fix $b\in{\mathcal U}({\mathcal B})$ and $\veps>0$. Then by definition we can find $b_i\in{\mathcal B}_i$ for $i=\uparrow,\downarrow$ such that $b_\uparrow\leq b\leq b_\downarrow$ and
$(T\omega)(b_\downarrow-b_\uparrow)\leq\veps$ dualizing $T$ in the last inequality and applying $T^*$ to the inequality for $b$ gives the required upper and lower bounds $T^*b_i$ for $T^*b$.
\end{proof}
We note that such inclusions are always equivalent to a continuity condition for $T$. If we chose subspaces $\widetilde{\mathcal A}\subset{\mathcal A}^{**}$ and $\widetilde{\mathcal B}\subset{\mathcal B}^{**}$, the inclusion $T^*(\widetilde{\mathcal B})\subset\widetilde{\mathcal A}$ is equivalent \cite[IV.2.1]{Schaefer} to the continuity with respect to the weak topologies $\sigma({\mathcal A}^*,\widetilde{\mathcal A})$ and $\sigma({\mathcal B}^*,\widetilde{\mathcal A})$, which are defined to make just those linear functionals ${\mathcal A}\to{\mathbb C}$ continuous, which are given by elements of $\widetilde{\mathcal A}$ (and similarly for~${\mathcal B}$).
\subsection{Hybrid observables as functions}
Applying the previous section to hybrids we define, for every $\Xis{}$, the spaces ${\mathcal U}\bigl({\mathrm C}^*\Xis{}\bigr)=:{\mathcal U}\Xis{}$ and $M\bigl({\mathrm C}^*\Xis{}\bigr)=:M\Xis{}$. This leads immediately to the chain of inclusions depicted on the left of Fig.~\ref{fig:setting3}.
\begin{figure}[H]
\begin{center}
\begin{tikzpicture}[scale=1]
\node at (0.5,2.35) {$\mu$-free hybrid setting};
\node[left] at (0,1.8) {states};
\node[right] at (1,1.8) {observables};
\node[right] at (1,1.25) {${\mathrm C}^*\Xis{}^{**}$};
\node[rotate=90] at (1.7,0.85) {$\subset$};
\node[] at (1.8,0.45) {${\mathcal U}\Xis{}$};
\node[rotate=90] at (1.7,0) {$\subset$};
\node[] at (1.8,-0.45) {$M\Xis{}$};
\node[rotate=90] at (1.7,-0.85) {$\subset$};
\node[left] at (0,0) {${\mathrm C}^*\Xis{}^{*}$};
\node[right] at (1,-1.25) {${\mathrm C}^*\Xis{}$};
\draw[red] (0,0) --(1,-1.25);
\draw[blue] (0,0) --(1,1.25);
\node at (6.5,2.35) {$\mu$-dependent setting};
\node[right] at (7.5,1.8) {states};
\node[right] at (4,1.8) {observables};
\node[right] at (4,1.25) {${\mathcal B}\bigl({\mathcal H}_1\otimes L^2(\Xi_0,\mu)\bigr)$};
\node[rotate=90] at (5,0.85) {$\subset$};
\node[right] at (4,0.45) {${\mathcal B}({\mathcal H}_1){\overset{{\rule{5pt}{1pt}}}\otimes} L^\infty(\Xi_0,\mu)$};
\node[rotate=90] at (5,-.45) {$\subset$};
\node[] at (1.8,-0.45) {$M\Xis{}$};
\node[rotate=90] at (1.7,-0.85) {$\subset$};
\node[right] at (8,-.45) {${\mathcal T}^1({\mathcal H}_1)\otimes L^1(\Xi_0,\mu)$};
\node[right] at (4,-1.25) {${\mathcal K}({\mathcal H})\otimes\mathcal{C}_0(\Xi_0)$};
\draw[red] (8,-.45) --(7,-1.25);
\draw[blue] (8,-.45) --(7,0.45);
\draw[->] (3,1.25)--(4,0.7);
\node at (3.5,1.2) {$i_\mu^*$};
\draw[->>] (3,.45)--(4,0.45);
\draw[->] (3,-1.25)--(4,-1.25);
\node at (3.5,-1) {$\cong$};
\end{tikzpicture}
\captionsetup{width=0.8\textwidth}
\caption{The two hybrid settings: The $\mu$-free setting is based on ${\mathrm C}^*\Xis{}$, which is represented on the
Hilbert space ${\mathcal H}_\mu={\mathcal H}_1\otimes L^2(\Xi_0,\mu)$ underlying the $\mu$-dependent approach. All $\mu$-free observable spaces are mapped to their counterparts by the restriction mapping $i_\mu^*$, the adjoint of the embedding $i_\mu$ taking the state space on the far right to the one on the left (not shown). According to Prop.~\ref{prop:multipliers}, this map is surjective from ${\mathcal U}\Xis{}$, indicated by a double arrow tip. It is injective on $M\Xis{}$ if $\mu$ has full support.}
\label{fig:setting3}
\end{center}
\end{figure}
To fully utilize the $\mu$-free structure, we need a concrete representation of multipliers and universally measurable elements of ${\mathcal A}={\mathcal C}_0(X,{\mathcal K})$, where ${\mathcal K}$ denotes the compact operators on a separable Hilbert space ${\mathcal H}_1$. In \cite[end of intro.]{brown}, this is actually suggested as the intuition-building model case for the theory we outlined in the previous section. The useful representations will be as operator-valued functions $X=\Xi_0\to{\mathcal B}({\mathcal H}_1)$. To define such a function, clearly no reference to a measure is needed. The connection to the $\mu$-dependent setting is then always made by the formula
\begin{equation}\label{omegaF}
\braket\omega F =\int\mu(dx)\ h(x) \operatorname{tr}\bigl(\rho_x F(x)\bigr),
\end{equation}
where $h\in L^1$ and the family $\rho_x$ define the state $\omega=i_\mu\bigr(\int^\oplus\!\!dx\,h(x)\rho_x\bigl)$.
Such a function representation does not exist for all $F\in{\mathrm C}^*\Xis{}^{**}$, and, as already noted above in the scalar case, the obvious candidate $\check F$, defined by $\braket{\rho\otimes\delta_x}F=\operatorname{tr}\rho \check F(x)$ may even vanish for elements such as the projection $\idty-z_{\mathfrak a}$. The notion of universal measurability for scalar functions was constructed precisely to build such a function representation. The usual ($\mu$-dependent) theory vastly extends the set of measurable functions but at the expense of ignoring any modifications on $\mu$-null sets. The largest class to which one can extend simultaneously for all measures are the universally measurable ones. In the spirit of this program, the non-commutative definitions of ${\mathcal U}({\mathcal A})$ given in the previous section were developed. Here we are looking, a bit less ambitiously, at {\it operator-valued} universally measurable functions. The statement of the proposition below is that the non-commutative version of universal measurability satisfies these needs. It would be even better to have an explicit description of the functions $F$ arising in this way, but that is beyond the scope of the present paper.
Continuing with the analogy to the scalar case, the fine points of measurability are irrelevant for continuous functions. Indeed, almost everywhere equal continuous functions are equal (provided the measure has full support, i.e., has no open null sets). Here the operator-valued case raises the interesting question of what kind of continuity must actually be demanded of an operator-valued function. The answer is given in item (1) of the following proposition. It turns out that the definition of the multiplier algebra contains two important extensions of ${\mathrm C}^*\Xis{}$: The functions in $M\Xis{}$ typically have non-compact values, and they need not decay at infinity. A key example for our project are the Weyl operators, which require both of these extensions.
\begin{prop}\label{prop:multipliers}
Let $X$ be locally compact, ${\mathcal K}$ the algebra of compact operators on a separable Hilbert space ${\mathcal H}$, and set ${\mathcal A}={\mathcal C}_0(X,{\mathcal K})$, and let $\mu$ be a measure on $X$. Then the formula \eqref{omegaF} extends to
\begin{itemize}
\item[(1)] $F\in M({\mathcal A})$, where the function $x\mapsto F(x)$ is strong*-continuous, i.e.,\\
$x\mapsto F(x)\psi$ and $x\mapsto F(x)^*\psi$ are both continuous for all $\psi\in{\mathcal H}$.
\item[(2)] $F\in{\mathcal U}({\mathcal A})$, where the function $x\mapsto F(x)$ is strong*-measurable, i.e.,\\
$x\mapsto F(x)\psi$ and $x\mapsto F(x)^*\psi$ are both $\mu$-measurable for all $\psi\in{\mathcal H}$.
\end{itemize}
In both cases, all bounded functions $F:X\to{\mathcal B}({\mathcal H})$ with the specified regularity conditions can be obtained in this way. In case (2) this coincides with the von Neumann algebra closure ${\mathcal B}({\mathcal H}){\overset{{\rule{5pt}{1pt}}}\otimes} L^\infty(X,\mu)$.
\end{prop}
\begin{proof}[Sketch of proof:]
(1) This is found in \cite[Cor~3.4]{PederMulti}. We nevertheless sketch the basic ideas of an approach not using the full-fledged theory.
According to Def.~\ref{def:multip}, multipliers can be thought of in terms of the operators
\begin{equation}
L_m,R_m:{\mathcal A} \rightarrow {\mathcal A}, \, L_m(A)=mA,\quad R_m(A)=Am .
\end{equation}
Their characteristic feature is $L_m(AB)=L_m(A)B$, and similarly for $R_m$. We note that the whole concept is redundant for C*-algebras with a unit, since then we just get multiplication with $L_m(\idty)=m\in{\mathcal A}$. This suggests that the application of $L_m(U_\lambda)$ for a bounded approximate unit $U_\lambda$ may help to characterize the action of $L_m$.
The first observation in the hybrid context is that these operators act pointwise, i.e., $\bigl(L_m(A)\bigr)(x)$ depends only on $A(x)$, or, equivalently: $A(x) = 0 \Rightarrow (L_mA)(x)=0$.
To show this, assume $A(x)=0$, and take a bounded approximate unit $U_\lambda \in {\mathcal A}$ with $\norm{U_\lambda A - A} \rightarrow 0 $ for all $A\in{\mathcal A}$. Then, because $L_m$ and $U_\lambda$ are bounded, and the product in ${\mathcal A}$ is defined pointwise:
\begin{align}
\norm{(L_m A)(x)} &\leq \norm{(L_m(A-U_\lambda A))(x)} + \norm{L_m(U_\lambda)(x) {A(x)}} \\
&\leq \norm{A-U_\lambda A} +\norm{(L_m (U_\lambda))(x)}\,\norm{A(x)}.
\end{align}
Then the second term is equal to zero, and the first goes to zero as $\lambda\to0$.
It follows that $L_m(A)(x)=L_m^x(A(x))$, where $L_m^x$ is a multiplier of ${\mathcal K}$ in the sense that it satisfies the basic relation $L_m^x(AB)=L_m^x(A)B$ for $A,B\in{\mathcal K}$. Now the multiplier algebra of ${\mathcal K}$ is known to be $M({\mathcal K})={\mathcal B}({\mathcal H})$. Indeed, for $A=\ketbra\Phi\Psi$ with a unit vector $\Psi$, we get
\begin{equation}\label{k}
L_m^x(A)=L_m^x(A\kettbra\Psi)=L_m^x(A)\kettbra\Psi=\ketbra{\widetilde\Phi}\Psi
\end{equation}
for a suitable vector $\widetilde\Phi$. Clearly, the map $\Phi\to\widetilde\Phi$ is linear, so we can set
$\widetilde\Phi=M(x)\Phi$ for an operator $M(x)$, which is easily checked to be bounded. Since the operators $\ketbra\Phi\Psi$ span a dense subspace of the compact operators, $L_m^x(A)=M(x)A$.
On the right side we get $R_m(A)(x)=\bigl(L_{m^*}(A^*)(x)\bigr)^*=\bigl(M(x)^*(A^*(x))\bigr)^*=A(x)M(x)$. It remains to check the continuity of $M$. To this end, we choose $\Phi,\Psi$ to be constant in a neighborhood of $x$. Then
$M(x)\ketbra\Phi\Psi$ has to be a norm continuous function, i.e., $M(x)\Phi$ is continuous in norm. Using the right multiplier instead, we find that $M(x)^*\Psi$ likewise has to be continuous.
We remark that in Def.~\ref{def:multip} the compacts are replaced by a more general algebra. The continuity is then in the natural topology for multipliers, the so-called strict topology, which is given by the seminorms
$\norm m_A=\norm{L_mA}+\norm{R_mA}$. For the compact operators, this coincides with the s*-topology \cite[I.8.6.3]{Blackadar}.
(2) With any element $F\in{\mathcal A}^{**}$ we can associate a function $f:X\to{\mathcal B}({\mathcal H})$ by evaluating $F$ on a pure state $\delta\otimes\rho$ for $\rho$ pure:
\begin{equation}\label{Fx}
\operatorname{tr}\rho{f(x)}=F(\delta_x\otimes\rho).
\end{equation}
Clearly, we can extend this by linearity to mixed density operators, as well. The right-hand side is then a linear functional in $\rho$, so this equation defines a unique operator $f(x)$.
However, for general $F\in{\mathcal A}^{**}$, $f$ might not be useful for learning about $F$. In particular, if $F=\idty-z_{\mathfrak a}$, the function $f$ turns out to be identically zero because a pure state will only be non-zero on the atomic representation. On the other hand, if $F\in{\mathcal U}({\mathcal A})$, this cannot happen because on this subspace, the atomic representation is isometric \cite[Thm.~III.6.37]{takesaki1}. The formula \eqref{omegaF} can be extended to larger classes of functions by standard constructions. For example, the monotone convergence theorem allows such an extension to the limits of increasing sequences of functions. When the family $x\mapsto\rho_x$ of states is continuous in a sense guaranteeing the lower semicontinuity of $\operatorname{tr}(\rho_xf(x))$ for $f\in{\mathrm C}^*\Xis{}$, then one can also invoke the monotone convergence theorem for nets \cite[Prop.~7.12]{FollandRealAnal} to extend the formula to ${\mathcal A}_\uparrow$, and hence to $M({\mathcal A})$.
The statement in the proposition bypasses all these ``universal'' considerations by going in to a standard representation depending on a measure $\mu$. Then ${\mathcal K}\otimes {\mathcal C}_0(X)$ acts as a function algebra on $\widehat{\mathcal H}={\mathcal H}\otimes L^2(X,\mu)$. According to \cite[Sect.~IV.7.]{takesaki1}, the state is given by a trace-class valued measurable function $\rho_.\in {\mathcal T}({\mathcal H})\otimes L^1(X,\mu)$, and the formula holds directly for ${\mathcal B}({\mathcal H}){\overset{{\rule{5pt}{1pt}}}\otimes} L^\infty(X,\mu)$. But also, we can follow the semicontinuity constructions directly via monotone operator limits in the separable(!) Hilbert space $\widehat{\mathcal H}$. This gives a representation of ${\mathcal U}({\mathcal A})$ in $\widehat{\mathcal H}$, which Takesaki shows \cite[Thm.~III.6.39]{takesaki1} to coincide with the von Neumann algebra closure of ${\mathcal A}$.
\end{proof}
\subsection{L\'evy continuity theorem}
This classical result relates the pointwise convergence of characteristic functions to the convergence of many more expectation values. One often finds it stated for single real random variables \cite{Lukacs}. Here is a sketch for many, possibly quantum variables.
\begin{prop}\label{prop:Levy}
Let $\omega_n$, $n\in{\mathbb N}\cup\{\infty\}$, be states in ${\mathrm C}^*\Xis{}^{*}$ with characteristic functions $\chi_n$. Then the following are equivalent:
\begin{itemize}
\item[(1)] $\lim_n\chi_n(\xi)=\chi_\infty(\xi)$ for all $\xi$,
\item[(2)] $\lim_n\braket{\omega_n}F=\braket{\omega_\infty}F$, for all $F\in{\mathrm C}^*\Xis{}$,
\item[(3)] $\lim_n\braket{\omega_n}F=\braket{\omega_\infty}F$, for all $F\in M({\mathrm C}^*\Xis{})$.
\end{itemize}
\end{prop}
\begin{proof}\
(3)$\Rightarrow$(1) is trivial because Weyl elements $W(\xi)$ are in the multiplier algebra. \\
(1)$\Rightarrow$(2) For $h\in L^1(\Xi)$, the expectation values $\braket{\omega_n}{W[h]}=\int d\xi\ h(\xi) \chi_n(\xi)$ converge by dominated convergence. This extends to ${\mathrm C}^*\Xis{}$ because elements of the form $W[h]$ are norm dense by the construction in Sect.~\ref{sec:CstarStates}. \\
(2)$\Rightarrow$(3) Let $u\in{\mathrm C}^*\Xis{}$ be an element such that $0\leq u\leq\idty$ such that $\braket{\omega_\infty} u\geq1-\veps/2$. Here $u$ serves as an ``approximate unit'' expressing the intuition that the state $\omega$ is essentially localized in a finite region of the phase space, resp. a finite-dimensional subspace of the Hilbert space. By assumption, the expectations for $u\in{\mathrm C}^*\Xis{}$ converge, so we also have
$\braket{\omega_n}u\geq1-\veps$ for sufficiently large $n$. Now for $F$ in the multiplier algebra and $n\leq\infty$,
$$\abs{\braket{\omega_n}{(1-u)F}}^2\leq \braket{\omega_n}{(1-u)FF^*(1-u)^*}\leq\norm F^2\braket{\omega_n}{(1-u)^2}\leq \norm F^2\braket{\omega_n}{(1-u)}\leq \veps\norm F^2.$$
Hence
$\abs{\braket{\omega_n-\omega_\infty}{F}}\leq \abs{\braket{\omega_n-\omega_\infty}{uF}} + 2\sqrt\veps\,\norm F$, which goes to zero, because $uF\in{\mathrm C}^*\Xis{}$ and such expectations converge by (2).
\end{proof}
It is crucial here that the pointwise limit $\chi_\infty(\xi)$ is {\it assumed} to belong to a normalized state (and, in particular, is normalized and continuous at zero, which would be sufficient). This prevents the states from wandering off to infinity, like $\alpha_{n\xi}\omega_0$ for any $\xi\neq0$. Then the limits in (2) are all zero, and those in (3) may fail to exist, so these items are no longer equivalent.
\subsection{Tensor product or function space? }\label{sec:tensprod}
In quantum mechanics and quantum information theory, the tensor product of observable algebras is usually considered the basic construction for combining two systems. In the case at hand, this would be a classical part, described by some algebra of functions on $X$ and a quantum part with observable algebra ${\mathcal A}_1={\mathcal B}({\mathcal H})$. It turns out that for hybrids, it is not so obvious how such a tensor product should be defined or whether the C*-tensor product is even a good way to describe this kind of composition. The previous section raises a doubt in this regard: While at the level of the algebras ${\mathrm C}^*\Xis{}$ the classical and quantum part are combined by a C*-tensor product (see Prop.~\ref{prop10tensor}), the same is not true for the enlarged observable algebras such as $M\Xis{}$. While all the enlarged observable algebras in Fig.~\ref{fig:setting3} consist of functions $X\to{\mathcal B}({\mathcal H}_1)$ with varying degrees of regularity, they cannot be written as a tensor product of classical and quantum parts. In this subsection, we discuss some aspects of this failure of tensor products. However, none of this is needed in the sequel, so this subsection can be skipped without much loss.
To begin with, the tensor product of C*-algebras is in general not uniquely defined \cite{takesaki1}: There is a minimal and a maximal choice of C*-norms on the linear algebra tensor product, which are in general different. However, all cross norms coincide if one of the algebras is abelian, so we will not have to worry about this ambiguity. This is the tensor product used in the basic hybrid algebra ${\mathcal A}={\mathcal C}_0(X,{\mathcal K})\cong{\mathcal K}\otimes{\mathcal C}_0(X)$.
However, we also need to include observables that do not decay at infinity and non-compact operators on the quantum side. Therefore, the question arises, whether for some algebra ${\mathcal A}_1$ (in our applications either ${\mathcal K}$ or ${\mathcal B}({\mathcal H})$) this sort of isomorphism also holds for the bounded ${\mathcal A}_1$-valued norm continuous functions ${{\mathcal C}_{\mathrm b}}(X,{\mathcal A}_1)$ with the C*-tensor product ${\mathcal A}_1\otimes{{\mathcal C}_{\mathrm b}}(X)$. The candidate for this isomorphism is
\begin{equation}\label{CXA}
\iota:{\mathcal A}_1\otimes{{\mathcal C}_{\mathrm b}}(X)\to{{\mathcal C}_{\mathrm b}}(X,{\mathcal A}_1), \qquad \iota\bigl(A\otimes f\bigr)(x)=f(x)A.
\end{equation}
This clearly can be extended to a *-homomorphism. When $X$ is compact, this is even an isomorphism \cite[Prop.\,1.22.3]{sakai}, and the subscript b can be dropped because continuous functions are automatically bounded.
However, we are interested in a locally compact space $X=\Xi_0={\mathbb R}^s$. In that case, the embedding $\iota$ is not surjective \cite{williams}. To see this, note that ${{\mathcal C}_{\mathrm b}}(X)={\mathcal C}(\beta X)$, where $\beta X$ is the Stone--\v Cech compactification, so we have $\iota({\mathcal A}_1\otimes{{\mathcal C}_{\mathrm b}}(X))={\mathcal C}(\beta X,{\mathcal A}_1)$. But the continuity of a bounded function $F:X\to{\mathcal A}_1$ does not necessarily imply the existence of a norm continuous extension to $\beta X$. In fact, if such an extension exists, the range $F(X)=\{F(x)|x\in X\}$ must have norm compact closure $F(\beta X)\subset{\mathcal A}_1$. As shown by Williams \cite{williams} this is precisely what could go wrong, i.e., we can turn this into a criterion describing the range of $\iota$.
\begin{example} An element $F\in{{\mathcal C}_{\mathrm b}}(X,{\mathcal A}_1)$ with $F\notin\iota({\mathcal A}_1\otimes{{\mathcal C}_{\mathrm b}}(X))$.\end{example}\extxt{
Take $F(x)=\sum_i P_i f_i(x)$, where $P_i$ is a family of orthogonal projections, e.g., in ${\mathcal A}_1={\mathcal K}({\mathcal H})$. The $f_i$ are chosen to have disjoint supports in the elements of some countable partition of $X$, are positive, take the value $1$ somewhere, and $\sum_if_i=\idty$. For example, with some fixed function $f_0\in{\mathcal C}_0({\mathbb R})$, such that $f_0(0)=1$ and $f_0(x)=0$ for $\abs x>1/3$, we can set $f_i(x)=f(x-i)$.
Then $\{P_i\}\subset F(X)$ does not have norm compact closure. Note that the sum defining $F$ cannot be obtained as a supremum-norm limit of finite partial sums, as would be required for $F\in{\mathcal A}_1\otimes{{\mathcal C}_{\mathrm b}}(X)$. }\\
As a consequence, it is preferable to consider the larger algebra ${{\mathcal C}_{\mathrm b}}(X,{\mathcal A}_1)$, rather than the tensor product, as a basic hybrid algebra. Note, however, that we can also change the topology of ${\mathcal A}_1$ for which we demand continuity. For example, consider ${\mathcal A}_1={\mathcal B}({\mathcal H})$, taken with the weak*-topology, and hence the space ${\mathcal C}_w(X,{\mathcal B}({\mathcal H}))$ of norm bounded, weak*- continuous functions $X\to{\mathcal B}({\mathcal H})$. Since the unit ball of ${\mathcal B}({\mathcal H})$ is now compact, the above argument of Williams no longer applies. Indeed if $F\in{\mathcal C}_w(X,{\mathcal B}({\mathcal H}))$, and $\rho\in{\mathcal B}({\mathcal H})_*$, the function $x\mapsto\operatorname{tr}\rho F(x)$ is bounded and continuous, and hence extends to $\beta X$. The value $F_\rho(\hat x)$ of this function at a point $\hat x\in\beta X$ is a bounded linear functional with respect to $\rho$, and there is an operator $F(\hat x)$ representing this functional. In other words
\begin{equation}\label{CCw}
{\mathcal C}_w(X,{\mathcal B}({\mathcal H}))={\mathcal C}_w(\beta X,{\mathcal B}({\mathcal H})).
\end{equation}
It is not obvious that this is even an algebra because the operator product is not continuous in the weak*-topology. However, it is still an order unit Banach space with well-defined positive cone and unit, so it is eligible as a space of observables (in the sense of positive operator valued measures). The same is true for ${\mathcal U}\Xis{}$.
As Prop.~\ref{prop:multipliers} shows, the multiplier algebra $M({\mathcal C}_0(X,{\mathcal K}))$ can also be considered as a set of operator-valued functions, and is clearly contained in ${\mathcal C}_w(X,{\mathcal B}({\mathcal H}))$. In this case, the algebraic product is well defined, which is readily seen both abstractly, by the definition of multipliers, and concretely, as the strong* operator topology is compatible with products. It is interesting to note that ${\mathcal C}_w(X,{\mathcal B}({\mathcal H}))$ and $M({\mathcal C}_0(X,{\mathcal K}))$ share the same ``classical part'', or center, namely the scalar functions ${\mathcal C}(\beta X)\cong{\mathcal C}_b(X)$. Here the center of an order unit space is spanned by the observables, which can be jointly measured with all others.
\subsection{Eigenvectors of translations}\label{sec:transEvecs}
We will later define quasifree channels by their covariance with respect to phase space translations. This hinges on the characterization of the joint eigenvectors of the translations, which is the topic of this section.
But only we need to introduce a notation for the translation maps themselves. In terms of canonical operators in a standard representation, a phase space shift acts as
\begin{equation}
R'_j=R_j+\eta_j\idty,
\end{equation}
where $\idty$ stands for the identity operator or the constant $1$-function. The size of the shift depends on $j$, and together these parameters form the components of a vector $\eta\in\Xi$. In terms of Weyl operators, this means
\begin{equation}\label{Weylshifted}
W'(\xi)=\exp(i\xi{\cdot} R')=\exp(i\xi{\cdot}\eta)W(\xi)=:\alpha_\eta\bigl(W(\xi)\bigr).
\end{equation}
Here we have introduced the automorphism $\alpha_\eta$, which expresses this symmetry as an automorphism on observables. If we think of $W(\xi)$ as an operator-valued function on $\Xi_0$, we have to define it on more general operator-valued functions as computed in \eqref{shiftQF}, i.e.
\begin{equation}\label{defalpha}
\alpha_\eta(F)(x)=W(\sigma\eta)^*F(x+\eta_0)W(\sigma\eta),
\end{equation}
where $\eta=\eta_1\oplus\eta_0$. This formula makes sense for any of the observable algebras that are built from ${\mathcal B}({\mathcal H}_1)$-valued bounded functions on $\Xi_0$, including the CCR-algebra in any standard representation, hence also ${\mathrm C}^*\Xis{}$. On standard states it is equivalent to
\begin{equation}\label{defalphastar}
\alpha^*_\eta(\delta_x\otimes\rho)=\delta_{x-\eta_0}\otimes W(\sigma\eta)\rho W(\sigma\eta)^*.
\end{equation}
If we define $\alpha_\eta$ as an automorphism group on ${\mathrm C}^*\Xis{}$, in \eqref{Weylshifted} we should really write $\alpha_\eta^{**}$, but we will continue to use the same symbol also for this map. Then \eqref{Weylshifted} just says that the Weyl operators $W(\xi)\in{\mathrm C}^*\Xis{}^{**}$ are joint eigenvectors of all translations. It will be crucial later on to turn this around:
\begin{lem}\label{lem:WeylEW}
Suppose that for some $F\in{\mathcal U}\Xis{}$ we have $\alpha_\eta(F)=\lambda(\eta) F$, for all $\eta$ and a suitable function $\lambda:\Xi\to{\mathbb C}$.
Then there is $\xi\in\Xi$ and $c\in{\mathbb C}$, such that $F=cW(\xi)$, and $\lambda(\eta)=\exp(i\xi{\cdot}\eta)$.
\end{lem}
\begin{proof}
Note first that we must have $\lambda(\eta+\eta')=\lambda(\eta)\lambda(\eta')$. Moreover, $\lambda$ must be a universally measurable function on $\Xi$ \cite[Prop.\,7.4.5]{PederBook}. Since the only measurable characters on ${\mathbb R}^{2s+n}$ are exponentials, we conclude that $\lambda(\eta)=\exp(i\xi{\cdot}\eta)$.
Consider $F'=FW(\xi)^*$. Then because $\alpha$ is a group of automorphisms, and $W(\xi)$ satisfies the required eigenvalue equation \eqref{Weylshifted}, we get that $\alpha_\eta(F')=F'$. It remains to prove that this implies that $F'=c\idty$, since then $F=F'W(\xi)=cW(\xi)$.
It suffices to prove this in every standard representation, where Eq.\ \eqref{defalpha} has a direct interpretation. Then, for all $\eta_0,\eta_1$, $W(\sigma_1\eta_1)^*F'(x+\eta_0)W(\sigma_1\eta_1)=F'(x)$, where the operators $W(\cdot)$ are Weyl operators in the Schr{\"o}dinger\ representation. Setting first $\eta_0=0$ we thus conclude that $F(x)=f(x)\idty$ by irreducibility of the standard quantum Weyl operators. By setting $\eta_1=0$, we get that this $f$ must be constant. Hence $F'(x)=c\idty$.
\end{proof}
\begin{example}Measurability $F\in{\mathcal U}\Xis{}$ is required.\end{example}
\extxt{
For simplicity, we will construct an example in the classical case ($\sigma=0$).
Let $\xi\to\lambda(\xi)$ be an arbitrary homomorphism ${\mathbb R}^n\to{\mathbb C}$ into the unit circle, of which we do not require any continuity or measurability. It is well known that there are many discontinuous $\lambda$, which are then necessarily non-measurable (see \cite[Ex.~3.2.4]{daoxing} or the review \cite{Rosendal}). A simple construction uses a Hamel basis of ${\mathbb R}^n$ as a vector space over ${\mathbb Q}$, i.e., a set of elements $e_j$, $j\in J$ such that every $\eta\in{\mathbb R}^n$ can be written uniquely as a finite linear combination $\eta=\sum_j\eta_je_j$. Then we just set $\lambda(\eta)=\exp i\sum_ja_i\eta_i$, for arbitrary constants $a_i$. It is easily arranged that such a function is not continuous.
Now consider the set
\begin{equation}\label{M4Markov}
{\mathcal M}=\Bigl\{F\in{\mathrm C}^*\Xis{}^{**}\Bigm| \norm F\leq1,\mbox{\ and, for all $\xi$:\ }
\delta_\xi(F)=\overline{\lambda(\xi)}\Bigr\}.
\end{equation}
As a weak*-closed subset of the unit sphere, it is compact, and it is nonempty because we can define $F$ as a functional on the linear combinations of point measures by the condition in ${\mathcal M}$ and then choose a Hahn--Banach extension. Now define the transformations $\beta_\eta=\overline{\lambda(\eta)}\alpha^{**}_\eta$. Because $\lambda$ is a character, these maps leave ${\mathcal M}$ invariant. They are also continuous and commute. Hence, by the Markov--Kakutani Fixed Point Theorem, they have a common fixed point $F$. $F$ must be non-zero because it is in ${\mathcal M}$, and as a fixed point of the $\beta_\eta$, it satisfies the equation $\alpha^{**}_\eta F=\lambda(\eta)F$. But since $\lambda$ is not continuous, it cannot be of the form given in the lemma.
}
\subsection{Continuity of state translations}\label{sec:transState}
We note that $\alpha_\eta^*$ is not strongly continuous on the Banach space of states, i.e., the function $\xi\mapsto\alpha_\eta^*\rho$ is not continuous in norm. Indeed, an arbitrarily small shift applied to a point measure moves it as far away as possible in the natural norm on states. Since translations are strongly continuous on $L^1({\mathbb R}^n,dx)$, this is different for probability measures with absolutely continuous densities. We can use this to single out one particular standard representation, namely that using the Lebesgue measure for $\mu$.
\begin{prop}\label{prop:contransl}
Let $\omega\in{\mathrm C}^*\Xis{}^*$ be a state with characteristic function $\chi$, and let $\mu$ be its marginal probability measure on $\Xi_0$. Then the following are equivalent:
\begin{itemize}
\item[(1)] $\omega$ is norm continuous under phase space translations, i.e., $\lim_{\eta\to0}\norm{\omega-\alpha_\eta^*\omega}=0$.
\item[(2)] $\mu$ is absolutely continuous with respect to the Lebesgue measure.
\item[(3)] $\omega$ is the restriction of a standard state $\widehat\omega$ on a purely quantum system, in which the classical variables in $\Xi_0$ also have conjugate momenta.
\end{itemize}
In this case $\chi\in\mathcal{C}_0(\Xi)$. As a partial converse, if $\chi\in L^p(\Xi,d\xi)$ for some $p\in[1,2]$ then the above conditions hold.
\end{prop}
\begin{proof}
(1)$\Rightarrow$(2): (2) only depends on the restriction of $\omega$ to the classical algebra. So this is a purely classical observation, which is valid for any locally compact group. Let $\alpha_x^*$, $x\in{\mathbb R}^n$ denote the action of translations on measures over ${\mathbb R}^n$. Then (1) says $\lim_{x\to0}\norm{\mu-\alpha_x^*\mu}=0$, where the norm is the dual norm of the supremum norm on ${\mathcal C}_0({\mathbb R}^n)$, i.e., the total variation norm on measures. For any $\veps>0$, there is thus a neighbourhood $U_\veps\ni0$ such that $\norm{\mu-\alpha_x^*\mu}\leq\veps$ for $x\in U_\veps$. We pick a positive measurable function $h\in L^1({\mathbb R}^n,dx)$ with integral $1$, and set
\begin{equation}\label{alfha}
\alpha^*_{h}(\mu)=\int dx\ h(x)\alpha^*_x(\mu),
\end{equation}
which is to be read as a weak* integral. By the triangle inequality $\norm{\mu-\alpha^*_h(\mu)}\leq\veps$. On the other hand, $\alpha^*_h(\mu)$ is absolutely continuous, because \eqref{alfha} is the convolution of the two measures $\mu$ and $h\,dx$. Explicitly, for an arbitrary $f\in{\mathcal C}_0({\mathbb R}^n)$, we get
\begin{equation}\label{alfhaf}
\bigl(\alpha^*_{h}\mu\bigr)(f)=\int dx\ h(x)\int \mu(dy)\ f(y-x)
=\int \mu(dy)\int dx\ h(y-x) f(x),
\end{equation}
which is to say that $\alpha^*_{h}\mu$ has density $\widetilde h(x)=\int \mu(dy)\ h(y-x)$ with respect to Lebesgue measure. Since $\alpha^*_{h}\mu$ converges in norm to $\mu$, and is thus a Cauchy sequence from $L^1({\mathbb R}^n,dx)$, its limit $\mu$ must be in $L^1$ by completeness of $L^1$.
(2)$\Rightarrow$(3):
Consider the standard representation for $\mu$ in the Hilbert space ${\mathcal H}_1\otimes L^2(\Xi_0,\mu)$. We can consider this as a subspace of ${\mathcal H}_1\otimes L^2(\Xi_0,dx)$ with the Lebesgue measure. By Def.~\ref{def:standard} $\omega$, as a state on the CCR-algebra, can be represented as a normal state on this space.
Apart from the canonical position operators in $L^2(\Xi_0,dx)$, which are already part of the hybrid setup, we can take the shift generators in this tensor factor as further canonical momentum operators. The full set of canonical operators is then clearly irreducible, so the extended system is purely quantum.
(3)$\Rightarrow$(1):
The Weyl translations in a standard representation are strongly continuous, which implies that the action $\alpha_\xi^*\rho=W(\sigma\xi)^*\rho W(\sigma\xi)$ is norm continuous for every $\rho\in{\mathcal T}({\mathcal H})$ so we can just restrict the continuity condition for the extended system to those translations $\eta$ which make sense in the original hybrid.
The final remark in the proposition is clear in one direction from the Riemann--Lebesgue Lemma and its quantum version \cite{QHA}. In the converse direction, it follows that the Fourier transform of $\chi$ is the density with respect to the Lebesgue measure, which is then even continuous and goes to zero.
\end{proof}
\subsection{\texorpdfstring{$L^p$}{Lp}-spaces}\label{sec:Lp}
If one is not interested in pure states, a good setting for hybrids is to restrict consideration to the norm continuous states characterized by Prop.~\ref{prop:contransl}. This leads to a purely von Neumann algebraic picture: Since all probability measures $\mu$ are then absolutely continuous with respect to Lebesgue measure $dx$, we can represent all states in ${\mathcal H}={\mathcal H}_1\otimes L^2(\Xi_0,dx)$, which leads to the spaces
\begin{equation}\label{L1}
L^1\Xis{}:={\mathcal T}^1({\mathcal H}_1)\otimes L^1(\Xi_0,dx)\qquad\mbox{and}\quad L^\infty\Xis{}:={\mathcal B}({\mathcal H}_1){\overset{{\rule{5pt}{1pt}}}\otimes} L^\infty(\Xi_0,dx) .
\end{equation}
Here the tensor product on the right uses the $1$-norm completion, and ${\mathcal T}^p$ denotes the Schatten classes for $1\leq p<\infty$, so ${\mathcal T}^1$ is the trace class. ${\mathcal T}^\infty$ would be ambiguous, meaning either all bounded operators ${\mathcal B}({\mathcal H})$ or just the compact operators ${\mathcal K}({\mathcal H}_1)$, so we prefer to specify explicitly. The tensor product on the right is then the von Neumann algebra version, constructed as the completion of the product operators in the weak (or similar) operator topology.
This choice was adopted, for example, in \cite{barchielli_1996,olkiewicz_dynamical_1999,bardet}, see Sect.~\ref{sec:previous}.
Note that the von Neumann algebra $L^\infty\Xis{}$ is not a subalgebra of ${\mathcal U}\Xis{}$ but a quotient, because the $(L^1,L^\infty)$-duality is defined by selecting a subspace of states (cf. Fig.~\ref{fig:setting3}). In fact, it does not make sense to evaluate an element $F\in L^\infty\Xis{}$ on a pure state because functions differing at one point, e.g., the point $x\in\Xi_0$ on which the pure state lives, are identified.
The combined Schatten/Lebesgue spaces $L^p\Xis{}$ can be obtained by a purely von Neumann algebraic construction using the semifinite trace
$\widehat\operatorname{tr}(f\otimes A)=\int dx\ f(x)\operatorname{tr} A$ on $L^\infty\Xis{}$. Then $L^p\Xis{}$ comes out as the $p$-norm completion of the elements $F\in L^\infty$ so that $\norm F_p^p:=\widehat\operatorname{tr} \abs F^p<\infty$. These spaces are also connected by interpolation \cite{ReedSimon2}. The only case of interest to us, however, is $p=2$ because of a fact, which is well--known in both the classical and the quantum case, so its generalization to hybrids is unsurprising:
The Fourier-Weyl transform\footnote{Note that the definition given here differs from that in \cite{QHA} by a symplectic matrix in the argument, which would not make sense in the classical case.}
\begin{equation}\label{FouWey}
({\mathcal F} F)(\xi)=\widehat\operatorname{tr}(F W(\xi))
\end{equation}
is a unitary isomorphism from $L^2\Xis{}$ onto $L^2(\Xi,0)$, i.e., for elements $F\in L^p\Xis{}\cap L^2\Xis{}$ and $G\in L^q\Xis{}\cap L^2\Xis{}$ with $p^{-1}+q^{-1}=1$, $1\leq p\leq\infty$ we have
\begin{equation}\label{L2scalarprod}
\widehat\operatorname{tr} F^*G=\int dx\, \operatorname{tr}(F(x)^*G(x))=(2\pi)^{-(n+s)}\int d\xi\ \overline{\operatorname{tr}(F W(\xi))}\operatorname{tr}(G W(\xi)),
\end{equation}
where $s$ and $n$ are the numbers of degrees of classical and quantum freedom, respectively (cp.\ Sect.~\ref{sec:setup}).
This extends by continuity to all of $L^2\Xis{}$, and the map is clearly onto. Its inverse is an instance of Weyl quantization, in our case, a partial one acting only on the quantum part of the system.
A similar useful formula concerns an integral over translates: For $F,G\in L^1\Xis{}\cap L^\infty\Xis{}$ we have
\begin{equation}\label{intL1}
\int d\xi\ \widehat\operatorname{tr}\bigl(F \alpha_\xi(G)\bigr)=(2\pi)^n\ \widehat\operatorname{tr}(F)\widehat\operatorname{tr}(G).
\end{equation}
The proof is immediate for the classical part, and the quantum part is essentially the square integrability of the quantum Weyl operators \cite[Lem.~3.1.]{QHA}.
\subsection{Correspondence for spaces of strongly continuous observables}\label{sec:corres}
The Weyl operators are clearly norm continuous elements of ${\mathrm C}^*\Xis{}^{**}$, i.e., as operator-valued functions on $\Xi_0$ they lie in the algebra
\begin{equation}\label{defucont}
{\mathcal C}_{\mathrm u}\Xis{}=\{F:\Xi_0\to{\mathcal B}({\mathcal H}_1)\mid \lim_{\eta\to0}\norm{F-\alpha_\eta F}=0\}.
\end{equation}
Since the norm in this definition is the supremum norm $\norm F=\sup_{\xi_0}\norm{F(\xi_0)}$, functions in this algebra are uniformly norm continuous. But this does not exhaust the condition \eqref{defucont}, since that would only refer to translations $\eta\in\Xi_0$. Thus $F\in{\mathcal C}_{\mathrm u}\Xis{}$ also imposes a condition on translations along quantum directions $\eta\in\Xi_1$.
It is suggestive to consider an alternative definition of ${\mathcal C}_{\mathrm u}$ directly in the second dual, and thus not assuming a function representation of $F$ from the outset, namely those $F\in{\mathrm C}^*\Xis{}^{**}$ for which $\alpha_\xi^{**}(F)$ is norm continuous. However, this would include some unwanted elements like $\idty -z_{\mathfrak a}$, which is even invariant under all $\alpha_\xi^{**}$ but has a vanishing function representation, as noted above. Note, however, that the condition on $F$ in \eqref{defucont} implies by Prop.~\ref{prop:multipliers}(1) that $F\in M\Xis{}$. Conversely, strong continuity of translations for elements in the multiplier algebra amounts to \eqref{defucont}, so we can write, slightly more abstractly
\begin{equation}\label{defucont1}
{\mathcal C}_{\mathrm u}\Xis{}=\{F\in M\Xis{}\mid \lim_{\eta\to0}\norm{F-\alpha^{**}_\eta F}=0\}.
\end{equation}
This algebra is still rather large, for example, not separable. In the context of Ludwig's axiomatic approach, it seemed natural to single out a norm separable subspace ${\mathcal D}\subset{\mathcal B}({\mathcal H})$ as a space of physical observables. One role of the space ${\mathcal D}$ would be to determine a more realistic assessment of the distinguishability of states compared to norm or weak topologies. It turned out \cite{QHA,uniformities} that in systems with canonical variables, the choices for ${\mathcal D}$ on the quantum side are in one-to-one correspondence with choices on the classical side, which, in turn, can often be understood in terms of compactifications of phase space. For example, the CCR-algebra corresponds to the almost periodic functions and the Bohr compactification, whereas the compact operators correspond to ${\mathcal C}_0(\Xi)$ and adjoining the identity to the one-point compactification of $X$. In this section we will show that the correspondence naturally also covers the hybrids between the fully quantum and the fully classical case. That is, the lattice of translation invariant closed subspaces of ${\mathcal C}_{\mathrm u}\Xis{}$ {\it does not depend on} $\sigma$.
This correspondence is best expressed in terms of the following notion of convolution. We denote by $\beta_-$ the automorphism of phase space inversion, satisfying $\beta_-(W(\xi))=W(-\xi)$, which is given by a coordinate change $\xi_0\mapsto-\xi_0$ on the classical part and is implemented by the parity operator on the quantum part. The sign freedom in the following definition is due to the fact that ${\mathrm C}^*(\Xi,\sigma)^*={\mathrm C}^*(\Xi,-\sigma)^*$: The twisted positive definiteness conditions \eqref{twistedPosDef} for $\sigma$ and $-\sigma$ both imply hermiticity ($\chi(-\xi)=\overline{\chi(\xi)}$), and with $\xi_k\mapsto-\xi_k$ and complex conjugation they become equivalent.
\begin{defi}\label{def:stateconvolve}
Let $\Xi$ be a real vector space with antisymmetric forms $\sigma_1$ and $\sigma_2$ and fix some signs $s_i=\pm1$ for $i=1,2$. Then, for states $\omega_i\in{\mathrm C}^*(\Xi,\sigma_i)^*$ with characteristic functions $\chi_i$, we define their {\bf convolution}, denoted by $\omega_1\ast\omega_2\in{\mathrm C}^*(\Xi,s_1\sigma_1+s_2\sigma_2)^*$ by its characteristic function $\chi(\xi)=\chi_1(\xi)\chi_2(\xi)$.\\
For $\omega\in{\mathrm C}^*(\Xi,\sigma_1)^*$ and $F\in{\mathrm C}^*(\Xi,\sigma_2)^{**}$, we define $\omega\ast F=F\ast\omega\in{\mathrm C}^*(\Xi,s_1\sigma_1+s_2\sigma_2)^{**}$ by evaluating it on an arbitrary $\omega'\in {\mathrm C}^*(\Xi,s_1\sigma_1+s_2\sigma_2)^*$:
\begin{equation}\label{Fconvolve}
\braket{\omega'}{\omega\ast F}= \braket{\omega'\ast(\beta_-\omega)}F.
\end{equation}
\end{defi}
Convolution is a bilinear operation ${\mathrm C}^*(\Xi,\sigma_1)^*\times{\mathrm C}^*(\Xi,\sigma_2)^*\to{\mathrm C}^*(\Xi,s_1\sigma_1+s_2\sigma_2)^*$, which is obviously commutative, associative, and bi-positive. It is also translation invariant in the sense that
$\alpha_\xi(\omega\ast F)=(\alpha_\xi\omega)\ast F=\omega\ast(\alpha_\xi F)$, which also shows why $\beta_-$ is needed in \eqref{Fconvolve}. The freedom of the sign in the definition is used to get a classical state or observable function as the convolution of two quantum objects.
For pointwise estimates it is useful to have a direct formula for the convolution, which bypasses the Fourier transform. When one factor is $\omega\in{\mathrm C}^*(\Xi,0)^*$, e.g.\ a classical probability measure on $\Xi$, this is the usual average over translates of the other factor:
\begin{equation}
\omega\ast F=\int\omega(d\xi)\ \alpha_\xi(F),
\end{equation}
where $F$ could be an observable or another state, and the symplectic form is the same for $F$ and $\omega\ast F$.
For a state $\omega\in{\mathrm C}^*(\Xi,\sigma)^*$ and an observable
$F\in{\mathcal C}_{\mathrm u}\Xis{}$ on the same hybrid, we get a uniformly continuous function $\omega\ast F\in{\mathcal C}_{\mathrm u}(\Xi,0)$:
\begin{equation}\label{convolvetofct}
\bigl(\omega\ast F\bigr)(\xi)=\omega\bigl(\alpha_\xi\beta_-(F)\bigr).
\end{equation}
The hybrid generalization of correspondence theory \cite{QHA,Fulsche} is given in the following proposition. $\omega\in{\mathrm C}^*(\Xi,\sigma_1-\sigma_2)^*$ is called {\bf regular} if it is norm continuous under translations (cf. Prop.~\ref{prop:contransl}) and its characteristic function vanishes nowhere.
\begin{prop}\label{prop:qha}
Let $\Xi$ be a vector space. Then the lattice of $\alpha$-invariant closed subspaces of ${\mathcal C}_{\mathrm u}\Xis{}$ does not depend on $\sigma$.
More precisely, let ${\mathcal D}_i\subset{\mathcal C}_{\mathrm u}(\Xi,\sigma_i)$ be $\alpha$-invariant closed subspaces, and $\omega_0\in{\mathrm C}^*(\Xi,\sigma_1-\sigma_2)^*$ regular.
Then the following are equivalent:
\begin{itemize}
\item[(1)] $\omega\ast{\mathcal D}_1\subset{\mathcal D}_2$ and $\omega\ast{\mathcal D}_2\subset{\mathcal D}_1$ for all $\omega\in{\mathrm C}^*(\Xi,\sigma_1-\sigma_2)^*$.
\item[(2)] The inclusions (1) hold for $\omega=\omega_0$.
\item[(3)] ${\mathcal D}_2$ is the closure of $\omega_0\ast{\mathcal D}_1$.
\item[(4)] ${\mathcal D}_2=\{A\in{\mathcal C}_{\mathrm u}(\Xi,\sigma_1)\mid \omega_0\ast A\in{\mathcal D}_1\}$.
\end{itemize}
\end{prop}
Note that because (1) does not depend on $\omega_0$, each of the following items holds for all regular $\omega_0$ if it holds for any one, and by the same token, is also equivalent to the same condition with ${\mathcal D}_1$ and ${\mathcal D}_2$ exchanged.
\begin{proof}
(See \cite[Thm.~4.1]{QHA} for more details.) The crucial fact here is Wiener's approximation theorem, which states that the translates of $\rho\in L^1(\Xi,d\xi)$ span a norm dense subspace iff the Fourier transform vanishes nowhere. These are precisely the regular elements of ${\mathrm C}^*(\Xi,0)^*$. The proof uses the following arguments:
\begin{lem} Let ${\mathcal D}\subset{\mathcal C}_{\mathrm u}\Xis{}$ be an $\alpha$-invariant closed subspace, and $\rho\in {\mathrm C}^*(\Xi,0)^*$. Then
\begin{itemize}
\item[(1)] $\rho\ast{\mathcal D}\subset{\mathcal D}$,
\item[(2)] when $\rho$ is regular, this inclusion is norm dense.
\end{itemize}
\end{lem}
\begin{proof}
(1) Now, $\rho$ is a classical standard state, i.e., a probability measure on $\Xi$. The convolution integral $\rho\ast A=\int \rho(d\xi)\alpha_\xi\beta_-(A)$ can be approximated for strongly continuous $A$ by partitioning the integration domain into regions, over which either $\alpha_\xi(A)$ changes little, or which have small total weight with respect to $\rho$. We may then replace $\alpha_\xi(A)$ by a constant in each region, thus approximating the convolution uniformly by a linear combination of translates $\alpha_\xi(A)$.
(2) For $A\in{\mathcal C}_{\mathrm u}\Xis{}$ we can find $\rho'\in L^1$ with sufficiently small support around the origin so that $\norm{\rho'\ast A-A}$ is small. Approximating $\rho'$ by a linear combination of translates $\alpha_\xi\rho$, we find that $A$ itself lies in the closure of the translation-invariant subspace generated by $\rho\ast A$.
\end{proof}
Coming back to the proof of the proposition, note that (1)$\Rightarrow$(2) is trivial. Given (2) we get
$\omega_0\ast\omega_0\ast{\mathcal D}_1\subset\omega_0\ast{\mathcal D}_2\subset{\mathcal D}_1$. But since $\omega_0\ast\omega_0$ is regular this inclusion is dense, which proves (3).
Next, we verify (3)$\Leftrightarrow$(4) by showing that the spaces defined by these conditions, that we temporarily call ${\mathcal D}_2^{(3)}$ and ${\mathcal D}_2^{(4)}$, are equivalent.
Suppose that $A\in{\mathcal D}_2^{(4)}$. Then because $A\in{\mathcal C}_{\mathrm u}(\Xi,\sigma_2)$, $A$ lies in the closed translation invariant subspace generated by $\omega_0\ast\omega_0\ast A\in\omega_0\ast{\mathcal D}_1$, which is ${\mathcal D}_2^{(3)}$.
Conversely, if $A\in{\mathcal D}_2^{(3)}$, it can be approximated by elements of the form $\omega_0\ast A_1$, so $\omega_0\ast A\approx \omega_0\ast\omega_0\ast A_1\in{\mathcal D}_1$, which means that $A\in{\mathcal D}_2^{(4)}$.
It remains to show that (3)$\Rightarrow$(1). Indeed $\omega\ast{\mathcal D}_2\subset\overline{\omega\ast\omega_0\ast{\mathcal D}_1}\subset{\mathcal D}_1$. On the other hand, since $\omega_0\ast\omega_0\ast{\mathcal D}_1\subset{\mathcal D}_1$ is dense, we find for arbitrary $\omega$: $\omega\ast{\mathcal D}_1\subset\overline{\omega\ast\omega_0\ast\omega_0\ast{\mathcal D}_1}\subset \overline{\omega\ast\omega_0\ast{\mathcal D}_2}\subset{\mathcal D}_2$.
\end{proof}
\section{Quasifree channels: Definition and constructions}\label{sec:channels}
The notion of quasifree states and operations arose in field theory and statistical mechanics \cite{demoen,fannes,loupias}. In statistical mechanics, a free time evolution is the non-interacting time evolution of a many-particle system. Indeed, in the absence of interaction, the time evolution on the one-particle Hilbert space should be automatically lifted to an evolution for the full system. Similarly, in field theory, one may get an evolution of the quantum field from a transformation of the test function space. In the setting of this paper, we are much less ambitious since our phase spaces, the analog of the one-particle spaces or the test function spaces, are finite-dimensional. We keep as the hallmark of quasifree evolutions that they can be characterized completely by linear operators at the phase space level. In contrast to the typical applications to unitary dynamics, we moreover include irreversible (completely positive) operations and general hybrids (see \cite{evansLewis,DaviesDiff} for some early extensions in the irreversible direction). We do not, however, consider quasifree maps on the CAR-algebra \cite{ArakiCAR,evansLewis} since the commutation of classical variables forms a much less happy combination with the anticommutation of fermionic degrees of freedom.
The linearity at the phase space level can be expressed as a {\it covariance condition} with respect to phase space translations. For general covariant channels, one has to fix representations of the symmetry group under consideration in the input system as well as in the output system, with the desired operations intertwining these two representations. In the case at hand, these will be two representations of the group of phase translations, and the difference between the representations is parametrized by a linear operator $S:\Xi_{\mathrm{out}}\to\Xi_{\mathrm{in}}$. Our first step will be to characterize all channels satisfying such a covariance condition plus a regularity condition, which ensures that standard states in the sense of the previous sections are mapped to standard states. The action of these channels on states, i.e., the Schr{\"o}dinger\ picture, will then be obvious. This was, in fact, the starting point of the present study. However, the corresponding Heisenberg pictures seemed initially rather unclear. Having clarified the necessary spaces in the previous section, we can now go on to apply these ideas and get Heisenberg picture channels for all quasifree channels without the need for any extra assumptions.
\subsection{Definition}\label{sec:defqf}
In the Sch\"odinger picture, a channel is a completely positive, normalization preserving, linear map
${\mathcal T}:{\mathrm C}^*\Xis_{\mathrm{in}}^*\to{\mathrm C}^*\Xis_{\mathrm{out}}^*$. It thus takes the input states to a device to the output states. Such channels include {\it measurements} when $\Xis_{\mathrm{out}}$ is classical (i.e., $\sigma_{\mathrm{out}}=0$), {\it preparations} ($\Xi_{\mathrm{in}}=\{0\}$), and all kinds of combinations in which, in addition to an operation on the quantum subsystem, classical information is used as an input, or is read out in the process (cf.\ Sect.~\ref{sec:BasicOps}). The Heisenberg picture is always denoted by ${\mathcal T}^*$, and ${\mathcal T}^*(A)$ for an observable $A$ of the output system is interpreted as that observable on the input system, which is obtained by first operating with the quantum device and then measuring $A$. The two pictures are thus related as two ways of viewing the same experiment. Since all observables have expectations in the standard hybrid state, they can be considered as elements of the dual, i.e., ${\mathrm C}^*\Xis{}^{**}$, and from this interpretation, it is clear that ${\mathcal T}^*$ must indeed be the Banach space adjoint of ${\mathcal T}$. In the definition we use the notation $S^\top:\Xi_{\mathrm{in}}\to\Xi_{\mathrm{out}}$ for the linear algebra transpose (or adjoint) of a linear map $S:\Xi_{\mathrm{out}}\to\Xi_{\mathrm{in}}$.
\begin{defi}\label{def:channel}
Let $\Xis_{\mathrm{in}}$ and $\Xis_{\mathrm{out}}$ be hybrid phase spaces, and $S:\Xi_{\mathrm{out}}\to\Xi_{\mathrm{in}}$ a linear map.
Then an {\bf $S$-covariant} channel is a completely positive linear operator ${\mathcal T}:{\mathrm C}^*\Xis_{\mathrm{in}}^*\to{\mathrm C}^*\Xis_{\mathrm{out}}^*$ such that, for all $\xi\in\Xi_{\mathrm{in}}$,
\begin{equation}\label{trcov}
{\mathcal T}(\alpha^{\mathrm{in}}_\xi)^*=(\alpha^{\mathrm{out}}_{S^\top\xi})^*{\mathcal T}.
\end{equation}
A {\bf quasifree channel} is a channel, which is $S$-covariant for some $S$.
\end{defi}
There is an alternative characterization in terms of ${\mathcal T}^*$, which also clarifies the data needed to specify an $S$-covariant channel.
\begin{prop}\label{prop:ChanOnW}
Let ${\mathcal T}$ be an $S$-covariant channel. Then there is a unique continuous and normalized function $f:\Xi_{\mathrm{out}}\to{\mathbb C}$, which is twisted positive definite with respect to the antisymmetric form
\begin{equation}\label{delsig}
\Delta\sigma=\sigma_{\mathrm{out}}-S^\top\sigma_{\mathrm{in}} S,
\end{equation}
such that, for all $\xi\in\Xi_{\mathrm{out}}$,
\begin{equation}\label{Scovf}
{\mathcal T}^*(W_{\mathrm{out}}(\xi))=f(\xi) W_{\mathrm{in}}(S\xi).
\end{equation}
Conversely, every function $f$ with this property defines an $S$-covariant channel.
\\
\noindent{\bf Terminology:\ } We will refer to $f$ as the {\bf noise function} of the channel ${\mathcal T}$, and to the hybrid state on $(\Xi_{\mathrm{out}},\Delta\sigma)$ with characteristic function $f$ as its
{\bf noise state}, and denote it typically by $\tau$.
\end{prop}
Before going into the proof, let us explain why the form \eqref{Scovf} determines a unique state. Prima facie it defines the action of the channel only on the Weyl operators, hence by norm limits on the CCR-algebra, but no further. The point is that the formula really defines a transformation ${\mathcal T}:{\mathrm C}^*\Xis_{\mathrm{in}}^*\to{\mathrm C}^*\Xis_{\mathrm{out}}^*$ on {\it states}: By taking expectations of \eqref{Scovf} with $\omega_{\mathrm{in}}$ we get the characteristic function of $\omega_{\mathrm{out}}={\mathcal T}\omega_{\mathrm{in}}$ as
\begin{equation}\label{Tchi}
\chi_{\mathrm{out}}(\xi)=f(\xi) \chi_{\mathrm{in}}(S\xi).
\end{equation}
This shows that the channel is indeed specified completely by $S$ and $f$. Another way to put this is to note that since the expectations of Weyl operators specify the state, the linear hull of these operators is weak*- dense in the bidual ${\mathrm C}^*\Xis{}^{**}$, and correspondingly in all the observable spaces. Therefore, \eqref{Scovf} suffices to define the Heisenberg picture channel by first a linear extension and then an extension by weak*-continuity.
\begin{proof}[Proof of Prop.~\ref{prop:ChanOnW}]
Applying $\alpha^{\mathrm{in}}_\xi$ to \eqref{Scovf} and using the eigenvalue equation \eqref{Weylshifted} for $W_{\mathrm{out}}(\xi)$ we find that
\begin{equation}
\alpha^{\mathrm{in}}_\xi{\mathcal T}^*\bigl(W(\eta)\bigr)
= {\mathcal T}^*\bigl( \alpha^{\mathrm{out}}_{S^\top\xi} W(\eta)\bigr)
= {\mathcal T}^*\bigl(e^{i S^\top\xi {\cdot} \eta } W(\eta)\bigr)
= e^{i \xi {\cdot} S\eta } {\mathcal T}^*\bigl( W(\eta)\bigr).
\end{equation}
That is, ${\mathcal T}^*\bigl(W(\eta)\bigr)$ is a joint eigenvector of the translations and must be proportional to
$W(S\eta)$. We denote the proportionality factor by $f(\eta)$, see Sect.~\ref{sec:transEvecs}. This immediately implies \eqref{Scovf}. Now we can choose a state $\omega_{\mathrm{in}}$ such that $\chi_{\mathrm{in}}$ vanishes nowhere, for example, a Gaussian. Since $\chi_{\mathrm{in}}$ is continuous by Bochner's Theorem, and the channel maps standard states to standard states, so $\chi_{\mathrm{out}}$ is also continuous, we conclude that $f$ is continuous.
We now have to analyze the condition for complete positivity. Here one should remember that the channel ${\mathcal T}$ is primarily defined on ${\mathrm C}^*(\Xi_{\mathrm{in}}\sigma_{\mathrm{in}})^*$ and complete positivity just means that ${\mathcal T}\otimes{\rm id}_n$ preserves positivity (i.e., positive semidefiniteness) for all $n$, where ${\rm id}_n$ is the identity on the $n\times n$-matrices ${\mathcal M}_n$ (viewed as density matrices). In order to give an equivalent formulation in the Heisenberg picture, one can check complete positivity on any subalgebra ${\mathcal A}\subset{\mathrm C}^*\Xis_{\mathrm{out}}$ so that the positivity of any element of $\omega_n\in{\mathrm C}^*\Xis_{\mathrm{out}}^*\otimes {\mathcal M}_n$ can be expressed as the positivity of expectation values of positive elements in ${\mathcal A}\otimes {\mathcal M}_n$, i.e., the positive cones are dual to each other. For this, any weak*-dense subalgebra ${\mathcal A}$ will do, and we take here the linear span of the Weyl operators for ${\mathcal A}$.
We now assume that ${\mathcal T}^*$ is completely positive and aim at deriving the stated twisted definiteness condition for $f$. To this end, we use that, for a completely positive operator ${\mathcal T}^*$, and any choice of finitely many $a_j,b_j$ we have $X=\sum_{jk}a_j{\mathcal T}^*(b_j^*b_k)a_k^*\geq0$. Here we choose $a_j=c_jW(S\xi_j)$ and $b_j=W(\xi_j)$ for an arbitrary choice of $\xi_1,\ldots,\xi_n\in \Xi_{\mathrm{out}}$, and $c_j\in{\mathbb C}$. The idea is that then $X$ becomes a multiple of the identity, namely \def\grr{}%
\begin{align}\label{MijWeyl}
0\leq X&=\sum_{jk}\overline{c_j}W(S\xi_j)\ {\mathcal T}^*\bigl(W(\xi_j)^*W(\xi_k)\bigr) c_kW(S\xi_k)^* \nonumber\\
&= \sum_{jk}\overline{c_j} c_k e^{i \xi_j{\cdot}\sigma_{\mathrm{out}}\xi_k/2}\
W(S\xi_j){\mathcal T}^*(W(-\xi_j+\xi_k))W(S\xi_k)^* \nonumber\\
&=\sum_{jk}\overline{c_j} c_k e^{i \xi_j{\cdot}\sigma_{\mathrm{out}}\xi_k/2}f(-\xi_j+\xi_k)\
W(S\xi_j)W(-S\xi_j+S\xi_k)W(S\xi_k)^* \nonumber\\
&= \sum_{jk}\overline{c_j} c_k
e^{i \xi_j{\cdot}\sigma_{\mathrm{out}}\xi_k/2}f(-\xi_j+\xi_k)\,e^{i(S\xi_j){\cdot}\sigma_{\mathrm{in}}(S\xi_k)}\ W(S\xi_j)W(S\xi_j)^*W(S\xi_k)W(S\xi_k)^*\nonumber\\
&=\sum_{jk}\overline{c_j} f(-\xi_j+\xi_k)\,e^{i \xi_j{\cdot}\Delta\sigma\xi_k/2}c_k\ \idty. \nonumber
\end{align}
The positivity of this expression for arbitrary $c_j$ and $\xi_j$ is exactly the stated twisted definiteness condition.
Conversely, when $f$ satisfies the conditions, we can define ${\mathcal T}$ acting on ${\mathrm C}^*\Xis_{\mathrm{in}}$ by Eq.\ \eqref{Tchi}, using Bochner's Theorem, and Prop.~\ref{prop:twgroup}. Continuity and normalization of the output characteristic function are then guaranteed by the corresponding properties of $f$. Positivity will be addressed together with complete positivity.
We have to extend Bochner's Theorem to a version involving an additional tensor factor ${\mathcal M}_n$. So let $\omega^{\mathrm{in}}\in{\mathrm C}^*\Xis_{\mathrm{in}}\otimes{\mathcal M}_n$ be positive. The matrix elements $\omega^{\mathrm{in}}_{\alpha\beta}$ then have characteristic functions $\chi^{\mathrm{in}}_{\alpha\beta}(\eta)=\omega^{\mathrm{in}}_{\alpha\beta}(W(\eta))$, and the positivity condition for $\omega^{\mathrm{in}}$ is the positivity of the matrix
\begin{equation}\label{omPos}
\chi^{\mathrm{in}}_{\alpha\beta}(-\eta_j+\eta_k) e^{i\eta_j{\cdot}\sigma_{\mathrm{in}}\eta_k} ,
\end{equation}
for arbitrary $\eta_1,\ldots,\eta_N$, where the indices of this matrix are considered to be the pairs $(\alpha,j)$ and $(\beta,k)$. Applying the channel ${\mathcal T}\otimes{\rm id}_n$ to $\omega^{\mathrm{in}}$ means the application of \eqref{Tchi} to each matrix element, resulting in a similar matrix for $\omega^{\mathrm{out}}=({\mathcal T}\otimes{\rm id}_n)\omega^{\mathrm{in}}$, namely
\begin{equation}\label{TchiMatricial}
\chi^{\mathrm{out}}_{\alpha\beta}(\xi_i-\xi_j)e^{\textstyle \frac i2 \xi_i{\cdot}\sigma_{\mathrm{out}}\xi_j}
=\Bigl(f(\xi_i-\xi_j)e^{\textstyle \frac i2 \xi_i{\cdot}\Delta\sigma\xi_j}\Bigr)\
\Bigl(\chi^{\mathrm{in}}_{\alpha\beta}(S\xi_i-S\xi_j)e^{\textstyle \frac i2 S\xi_i{\cdot}\sigma_{\mathrm{in}} S\xi_j}\Bigr).
\end{equation}
Here we used the definition of $\Delta\sigma$. By assumption, the matrix in the first factor is positive definite.
The second factor is positive because the input state \eqref{omPos} is positive with the substitution $\eta_j=S\xi_j$. Hence the left-hand side is also positive definite as the Hadamard product of two positive definite matrices.
\end{proof}
\subsection{State-channel correspondence}
In this section, we will describe in more detail the geometry of the correspondence between an $S$-covariant channel ${\mathcal T}$ and its noise state $\tau$, which was set up in Prop.~\ref{prop:ChanOnW}. The operator $S$ will be fixed, and this is necessary if we want to consider a correspondence of convex sets: The convex combination of quasifree channels with different $S$ is simply not quasifree. However, the design possibilities for channels by engineering $\tau$ are not exhausted by convex combinations. Since arbitrary states are allowed, superpositions work just as well (see, e.g., \cite{Volkoff}).
We begin with some general remarks on state-channel correspondence and cones in quantum theory.
State-channel correspondence has been a very useful tool in quantum information theory. It originated in Choi's thesis \cite{choi}, which is often cited together with Jamio{\l}kowski\ \cite{jamiol}\footnote{ However, Jamio{\l}kowski's work appeared before the importance of complete positivity was generally recognized and gets the right isomorphism only up to an additional partial transpose operation.}. If we restrict for the moment to finite-dimensional systems, we can summarize this by saying in quantum theory, there is only one isomorphism type of positive cone for the basic objects: For {\it observables} it is the elements of the form $A^*A$, for {\it states} it is the dual of the observable cone, and for {\it channels} it is the completely positive cone. The inclusion of direct sums of positive semidefinite cones extends this statement to quantum-classical hybrid systems. As an immediate consequence, we find that there is only one kind of order interval, which in an ordered vector space is a set of the kind $[x_1,x_2]=\{ x |x_1\leq x\leq x_2\}$, which is obviously determined by just the order relations. In particular, the possible decompositions $\rho=\rho_1+\rho_2$ of a fixed state $\rho$ into a sum of positive $\rho_i$ is isomorphic to the corresponding interval $[0,\idty]$ in which decompositions are just two-valued observables, and decompositions ${\mathcal T}={\mathcal T}_1+{\mathcal T}_2$ of a channel into completely positive terms, i.e., an instrument with overall state change ${\mathcal T}$. This correspondence of order intervals is, in a sense, more robust than the correspondence of cones: It persists in infinite-dimensional systems while the isomorphism of cones breaks down. For example, ${\mathcal B}({\mathcal H})$ has an order unit (an element $u$ so that $a\leq\lambda u$ for all $a\geq0$ and suitable $\lambda$, here: $u=\idty$), whereas the trace class has none.
From the finite-dimensional case, it is clear that the difference between the spaces of states, observables, and channels lies in the respective normalization conditions. This is also reflected in the different structures of the convex sets of normalized elements: The extreme points are the projective Hilbert space for states, the projection lattice for observables, and something more complicated for channels.
Moreover, we get different natural norms: The trace norm, the operator norm, and the ``norm of complete boundedness'' \cite{paulsen}, denoted by $\cbnorm\cdot$, which is often also (sometimes only in the Schr{\"o}dinger picture) referred to as the diamond norm \cite{kitaev}. The cb-norm has a reputation of being not easy to compute \cite{paulsen,watrouscb}.
One surprising fact about the isomorphism ${\mathcal T}\leftrightarrow\tau$ is that it connects the normalized subsets of different categories: states on the one hand and quasifree channels on the other. In the light of the above explanations this readily traced to the normalization conditions: For a channel the normalization condition is ${\mathcal T}^*\idty=\idty$, and for general completely positive maps $\cbnorm{\mathcal T}=\norm{{\mathcal T}^*\idty}$. This is not a linear function of ${\mathcal T}$. However, for a general bounded covariant map ${\mathcal T}$, we have shown (see \eqref{Scovf} with $\xi=0$) that ${\mathcal T}^*(\idty)=f(0)\idty$, so $\cbnorm{\mathcal T}=f(0)=\norm\tau$. So for positive elements, both norms depend only on one number, and this dependence is linear, i.e., the norm is additive on the positive cone. This is exactly what makes a full channel-state correspondence possible here. This feature is shared by other classes of covariant channels, i.e., channels that intertwine automorphic actions of a group $G$, i.e., $(\alpha_g^{\mathrm{out}})^*{\mathcal T}={\mathcal T}(\alpha_g^{\mathrm{in}})^*$ for $g\in G$. When the $\alpha_g^{\mathrm{in}}$ are implemented by an irreducible set of unitaries, a projective representation of $G$, then, once again, ${\mathcal T}^*\idty$ is a multiple of the identity, and the class of covariant channels is affinely isomorphic to a state space of a quantum system that can be computed from the representations involved \cite{Reeb}. We see here that the irreducibility of the implementing unitaries is not the key condition since, on the classical subsystem, no such unitaries exist. Instead, the decisive condition is that the representation on the input side has only the multiples of $\idty$ as invariant elements, in the hybrid case a special case of Lem.~\ref{lem:WeylEW}.
The following corollary summarizes the above discussion and lists some transfers-of-properties for the correspondence.
\begin{cor}\label{cor:statechannel}
Fix hybrid systems with phase spaces $\Xis_{\mathrm{in}}$ and $\Xis_{\mathrm{out}}$, and a linear map $S:\Xi_{\mathrm{in}}\to\Xi_{\mathrm{out}}$. Then there is a bijective correspondence between $S$-covariant channels ${\mathcal T}$ in the sense of Def.~\ref{def:channel}, and noise states $\tau$ on the hybrid system $(\Xi_{\mathrm{out}},\Delta\sigma)$ as stated in Prop.~\ref{prop:ChanOnW}. Then if ${\mathcal T},{\mathcal T}_1,{\mathcal T}_2$ correspond to $\tau,\tau_1,\tau_2$, respectively, and $\lambda\in{\mathbb R}$, then
\begin{itemize}
\item[(1)] ${\mathcal T}=\lambda{\mathcal T}_1+(1-\lambda){\mathcal T}_2$ iff $\tau=\lambda\tau_1+(1-\lambda)\tau_2$,
\item[(2)] $\lambda{\mathcal T}_2-{\mathcal T}_1$ is completely positive iff $\lambda\tau_2-\tau_1\geq0$,
\item[(3)] $\cbnorm{{\mathcal T}_1-{\mathcal T}_2}\leq\norm{\tau_1-\tau_2}$,
\item[(4)] for $\xi\in\Xi_{\mathrm{out}}$, ${\mathcal T}_1=\alpha_\xi^*{\mathcal T}$ iff $\tau_1=\alpha_\xi^*\tau$,
\item[(5)] $\tau$ is extremal (= pure) iff ${\mathcal T}$ is noiseless in the sense of Sect.~\ref{sec:noiseless},
\item[(6)] $\tau$ is norm continuous under translations iff ${\mathcal T}$ is smoothing in the sense of Sect.~\ref{sec:smooth}.
\end{itemize}
\end{cor}
\begin{proof}
The bijective correspondence is directly from Prop.~\ref{prop:ChanOnW}.
(1) and (2) are obvious, and (5),(6) will be shown in the sections mentioned.
(4) follows by noting that under the translations by $\xi$ stated in that item, the noise function $f(\eta)=\braket \tau{W(\eta)}$ changes by a factor $\exp(i\xi{\cdot}\eta)$.
(3) Both norms are additive on the positive cone and coincide there. There is then a largest norm on the real linear span of the positive elements with this property, called the base norm \cite{Nagel}. The norm on states is of this type, which implies the inequality. A bit more explicitly, the base norm has the smallest unit ball of all the norms with the given restriction, just the convex hull of the positive and the negative elements of norm one.
\begin{eqnarray}\label{basenorm}
\cbnorm{{\mathcal T}_1-{\mathcal T}_2}&\leq& \inf\bigl\lbrace p_++p_-\bigm| {\mathcal T}_\pm\mbox{\ channels},\ ({\mathcal T}_1-{\mathcal T}_2)= p_+{\mathcal T}_+-p_-{\mathcal T}_-\bigr\rbrace \nonumber \\
&=& \inf\bigl\lbrace p_++p_-\bigm| \tau_\pm\mbox{\ states},\ (\tau_1-\tau_2)= p_+\tau_+-p_-\tau_-\bigr\rbrace \nonumber \\
&=&\norm{\tau_1-\tau_2}.
\end{eqnarray}
\end{proof}
\subsection{Heisenberg pictures for \texorpdfstring{$S$}{S}-covariant channels}\label{Heisenqf}
The Heisenberg picture ${\mathcal T}^*$ of a quasifree channel is initially defined on the bidual ${\mathrm C}^*\Xis_{\mathrm{out}}^{**}$. However, it also maps better-behaved algebras into each other, so one can settle for one of these algebras as the basic observables in some context. Since the definitions given in Sect.~\ref{sec:funcobs} work for arbitrary C*-algebras, commutative, quantum, or hybrid, the analytic properties defining these more special algebras are automatically preserved for all quasifree channels. As remarked already after Lem.~\ref{lem:autoHeisen1} each inclusion ${\mathcal T}^*{\mathcal A}\subset{\mathcal B}$ can also be read as a continuity condition for ${\mathcal T}$.
\begin{prop}\label{prop:Heisenalg}
For $\Xis{}=\Xis_{\mathrm{in}}$ or $\Xis{}=\Xis_{\mathrm{out}}$ consider the algebras
\begin{equation}\label{Heisenbergs}
{\mathrm{CCR}}\Xis{}\subset{\mathcal C}_{\mathrm u}\Xis{}\subset M\Xis{}\subset {\mathcal U}\Xis{}\subset{\mathrm C}^*\Xis{}^{**}.
\end{equation}
Let ${\mathcal T}^*$ be the Banach space adjoint of a quasifree channel ${\mathcal T}:{\mathrm C}^*(\Xi_{\mathrm{in}}\sigma_{\mathrm{in}})^*\to{\mathrm C}^*(\Xi_{\mathrm{out}}\sigma_{\mathrm{out}})^*$. Then ${\mathcal T}^*$ maps the ``out'' version of an algebra in this inclusion chain to the corresponding ``in'' version.
\end{prop}
\begin{proof}
${\mathcal T}^*$ is initially defined on the bidual, i.e., the largest element in the chain, so for this one, there is nothing to prove. We have shown this for arbitrary dual channels for $M\Xis{}$ and ${\mathcal U}\Xis{}$ in Sect.~\ref{sec:funcobs}. The other statements use the quasifree structure. This is obvious from Prop.~\ref{prop:ChanOnW} for ${\mathrm{CCR}}\Xis{}$. For ${\mathcal C}_{\mathrm u}\Xis{}$ it follows from the observation that
if $\xi\mapsto \alpha^{\mathrm{out}}_\xi(F)$ is norm continuous for some $F\in{\mathcal U}\Xis_{\mathrm{out}}$, then
\begin{equation}\label{ucontmapsto}
\xi\mapsto S\xi\mapsto \alpha^{\mathrm{out}}_{S\xi}(F)= {\mathcal T}^*\alpha^{\mathrm{out}}_{S\xi}(F)
\mapsto\alpha^{\mathrm{in}}_\xi\ {\mathcal T}^*(F)
\end{equation}
is also continuous.
\end{proof}
The algebra ${\mathrm C}^*\Xis{}$ is conspicuously absent from the proposition's list of algebras with an automatic Heisenberg picture. Indeed, it does not belong on that list. A simple counterexample is a depolarizing channel, for which $S=0$, and $f=\chi_0$ is the characteristic function of some output state $\omega_0$. Then, after \eqref{Tchi}, $\chi_{\mathrm{out}}=\chi_0$ for all input states. This translates to the Heisenberg picture as ${\mathcal T} (A)=\omega_0(A)\idty$. So even if $A\in{\mathrm C}^*\Xis_{\mathrm{out}}$ its image under the Heisenberg picture channel map is a multiple of the identity $\notin{\mathrm C}^*\Xis_{\mathrm{in}}$. Nevertheless, there is an easily checkable condition that will ensure the Heisenberg picture also in this case:
\begin{lem}Let ${\mathcal T}$ be an $S$-covariant channel. Then either
\begin{itemize}
\item[(1)] $S\Xi_{\mathrm{out}}=\Xi_{\mathrm{in}}$ and ${\mathcal T}^*\,{\mathrm C}^*\Xis_{\mathrm{out}}\subset{\mathrm C}^*\Xis_{\mathrm{in}}$, or
\item[(2)] $S\Xi_{\mathrm{out}}\neq\Xi_{\mathrm{in}}$ and $\bigl({\mathcal T}^*\,{\mathrm C}^*\Xis_{\mathrm{out}}\bigr)\cap{\mathrm C}^*\Xis_{\mathrm{in}}=\{0\}$.
\end{itemize}
\end{lem}
\begin{proof}
(1) Take an element $W[h]=\int d\xi\ h(\xi)W(\xi)\in{\mathrm C}^*\Xis_{\mathrm{out}}$. By definition of the algebra ${\mathrm C}^*\Xis_{\mathrm{out}}$ as the C*-envelope of the twisted convolution algebra, such elements are dense. It, therefore, suffices to show that the image under the channel is again given by such an integral. Applying the channel gives
\begin{equation}\label{actOnComp}
{\mathcal T}^* W[h]=\int d\xi\ h(\xi)f(\xi) W(S\xi).
\end{equation}
We can split the integration variables into $\xi=(\xi_\perp,\xi_\Vert)$ with $S\xi_\perp=0$, and a variable $\xi_\Vert$ in a suitable linear complement of the kernel. Then $\xi_\Vert$ uniquely specifies a point $S\xi_\Vert=\eta\in\Xi_{\mathrm{in}}$. Carrying out the integral over $\xi_\perp$ leaves ${\mathcal T}^* W[h]=W[h']$ with a function $h'(\eta)=\int d\xi_\perp \ h(\xi_\perp,\xi_\Vert)f(\xi_\perp,\xi_\Vert)$ which clearly lies in $L^1(\Xi_{\mathrm{in}},d\eta)$.
(2) When $S$ is not surjective, there is a non-zero vector $\eta$ orthogonal to $S\Xi_{\mathrm{out}}$. Then we have
$\alpha_\eta (W(S\xi))=\exp(i\eta{\cdot} S\xi)W(S\xi)=W(S\xi)$ for all $\eta$. Integrating with an arbitrary $h\in L^1(\Xi_{\mathrm{out}})$, it follows that
\begin{equation}\label{f}
\alpha_\eta{\mathcal T}^* W[h]= \int d\xi\ h(\xi)f(\xi) \alpha_\eta (W(S\xi))={\mathcal T}^* W[h].
\end{equation}
This transfers to ${\mathrm C}^*\Xis_{\mathrm{out}}$ by continuity. The image therefore consists of $F\in{\mathcal C}_{\mathrm u}\Xis_{\mathrm{in}}$ satisfying $\alpha_\eta F=F$. We will show that together with $F\in{\mathrm C}^*\Xis_{\mathrm{in}}$ this implies $F=0$. With \eqref{defalpha}, the action of translations on functions $F\in{\mathcal C}_{\mathrm u}\Xis_{\mathrm{in}}$ is given by
\begin{equation}\label{Fshifted}
\bigl(\alpha_\eta F\bigr)(\xi_0)= W(\sigma\eta)^*F(\xi_0+\eta_0)W(\sigma\eta) =F(\xi_0),
\end{equation}
where the last equality expresses our first conclusion. We take the norm on both sides so that the non-zero vector $\eta$ enters only through its classical part $\eta_0$. We claim that this classical part must vanish. Indeed, the sequence $n\mapsto \xi_0+n\eta_0$ goes to infinity for all $\xi_0$, and since $F\in{\mathrm C}^*\Xis_{\mathrm{in}}$, we have
$\norm{F(\xi_0+n\eta_0)}=0$. But then $F(\xi_0)=0$ for all $\xi_0$, and $F=0$. Hence $\eta_0=0$.
Now, for any fixed $\xi_0$ \eqref{Fshifted} says that the (supposedly) compact operator $F(\xi_0)$ commutes with a one-parameter subgroup of Weyl operators. In particular, the finite-dimensional eigenspaces of $F(\xi_0)+F(\xi_0)^*$ would have to be invariant under such a group. But since the generators in the Schr{\"o}dinger\ representation have a continuous spectrum, this is impossible. So the eigenspaces for non-zero eigenvalues have to be empty, which implies $F(\xi_0)+F(\xi_0)^*=0$. Repeating this argument for $i(F(\xi_0)-F(\xi_0)^*)$ we get $F(\xi_0)=0$ for all $\xi_0$, hence $F=0$.
\end{proof}
It may also be advantageous to single out a translation-invariant subspace of ${\mathcal C}_{\mathrm u}\Xis{}$. The selection may even be done uniformly for all $\Xi$ so that another instance of an ``automatic Heisenberg picture'' results. Rather than expanding this theory, let us give an example.
\begin{example}\label{Ex:resAlg} The resolvent algebra.\end{example}\extxt{
Let us consider the resolvent algebras defined in \cite{buchholz1,buchholz2} from the point of view of the present paper, particularly the correspondence theory of Prop.~\ref{prop:qha}. In contrast to the cited works we thus restrict to finite-dimensional $\Xi$. By definition, the resolvent algebra ${\mathcal R}\Xis{}$ is the C*-algebra generated by the resolvent elements $(u\idty-\sum_i\xi_i R_i)^{-1}$, where the $R_i$ are the field operators from \eqref{fieldOp}, and $u\in{\mathbb C}$ with $\Im m\,u\neq0$. Let us first consider the C*-algebra generated by the resolvents with one fixed $\xi$. Now the functions $t\mapsto1/(u+t)$ generate the C*-algebra ${\mathcal C}_0({\mathbb R})$ by the Stone-Weierstra\ss\ Theorem, but, by the resolvent equation, it actually suffices to take the linear span of these functions. Moreover, this space is $\alpha$-translation invariant, since
\begin{equation}\label{resolventshift}
\alpha_\eta\Bigl(\bigl(u\idty-\sum_i\xi R_i\bigr))^{-1}\Bigr)=\bigl((u-\xi{\cdot}\eta)\idty-\sum_i\xi R_i\bigr)^{-1}.
\end{equation}
Hence the resolvents are constant in the direction of the subspace $M=\xi^\perp:=\{\eta|\xi{\cdot}\eta=0\}$ and go to zero transversally to this subspace. More generally, we define, for any linear subspace $M\subset\Xi$:
\begin{align}\label{CC0XiM}
{\mathcal C}_0(\Xi/M,\sigma)=\{A\in{\mathcal C}_{\mathrm u}\Xis{}\mid\ & \alpha_\xi(A)=A\ \mbox{for}\ \xi\in M, \nonumber \\
&\text{and}\ w^*{-}\!\lim_\xi\alpha_\xi(A)=0\ \text{for}\ \xi+M\to\infty\}.
\end{align}
Here the limit condition just means that for fixed $\omega\in{\mathrm C}^*\Xis{}^*$, $|\braket\omega{\alpha_x(A)}|$ becomes arbitrarily small as soon as $\xi$ is outside a cylinder, which is infinite in the $M$-directions and compact transverse to it. One easily checks that these spaces are corresponding in the sense of Prop.~\ref{prop:qha}. Moreover, it is clear that with decreasing $M$ these algebras interpolate between ${\mathcal C}_0(\Xi/\Xi,\sigma)={\mathbb C}\idty$ and ${\mathcal C}_0\Xis{}$, and that products are evaluated according to ${\mathcal C}_0(\Xi/M_1,\sigma){\mathcal C}_0(\Xi/M_2,\sigma)={\mathcal C}_0(\Xi/(M_1\cap M_2),\sigma)$. Hence
\begin{equation}\label{resalg}
{\mathcal R}\Xis{}=\sum_M {\mathcal C}_0(\Xi/M,\sigma).
\end{equation}
In the specific sense of Prop.~\ref{prop:qha}, ${\mathcal R}\Xis{}$ is independent of $\sigma$. It hence suffices to check in the classical case ($\sigma=0$) that it is closed under noiseless channels and tensor product expansions (see Sect.~\ref{sec:noisefac}), and hence under arbitrary quasifree channels.
It would be interesting to get the exact relationship between the above analysis, which is manifestly independent of $\sigma$, and the one in \cite{buchholz1}, which focuses particularly on the non-degenerate subspaces $M\subset\Xi$, i.e., the subspaces on which $\sigma$ is symplectic.
}
\subsection{Smoothing channels}\label{sec:smooth}
The $\mu$-dependent setting, with the special choice $\mu=dx$ as the Lebesgue measure, has been singled out by the norm continuity of states under translations in Sect.~\ref{sec:transState}. The resulting structure also supports other $L^p$ spaces and corresponding Schatten classes (Sect.~\ref{sec:Lp}). A natural question is then whether a given quasifree channel preserves the continuity of states and therefore can be seen as a normal map between the corresponding hybrid von Neumann algebras $L^\infty\Xis{}$ as defined in \eqref{L1}. The identity channel obviously has this property, but, for example, a depolarizing channel with a pure output state does not. The following lemma gives a positive answer for general non-singular $S$ and arbitrary noise function. Because all $S_t$ in a matrix semigroup are non-singular, we conclude that the von Neumann algebra $L^\infty\Xis{}$, as used in \cite{barchielli_1996}, is a sufficient arena for quasifree semigroups.
\begin{lem}\label{lem:ChanL1}
Let ${\mathcal T}$ be the quasifree channel given by $S:\Xi_{\mathrm{out}}\to\Xi_{\mathrm{in}}$ and $f:\Xi_{\mathrm{out}}\to{\mathbb C}$. Suppose that $S$ is injective.
Then ${\mathcal T}$ maps norm continuous states to norm continuous states.
\end{lem}
\begin{proof}
When $S$ is injective, $S^\top\Xi_{\mathrm{in}}\to\Xi_{\mathrm{out}}$ is surjective. So let $\xi\in\Xi_{\mathrm{out}}$, which we can consequently write as $\xi=S^\top\eta$. Suppose that $\rho_{\mathrm{in}}$ is norm continuous under translations, and consider $\rho_{\mathrm{out}}={\mathcal T}\rho_{\mathrm{in}}$. Then by \eqref{trcov} the function $t\mapsto\alpha^*_{t\xi}\rho_{\mathrm{out}}=\alpha^*_{tS^\top\eta}{\mathcal T}\rho_{\mathrm{in}}={\mathcal T}\bigl(\alpha^*_{t\eta}\rho_{\mathrm{in}}\bigr)$ is continuous in norm. Since this holds for all $\xi\in\Xi_{\mathrm{out}}$ and the translations commute, $\xi\mapsto \alpha_\xi^*\rho_{\mathrm{out}}$ is also norm continuous.
\end{proof}
For other channels {\it all} output states, not just those from continuous input states, are continuous. For that we need sufficient noise, and the following proposition collects some basic observations.
\begin{prop}\label{prop:smoothing}
Let ${\mathcal T}$ be a quasifree channel with noise state $\tau$ and noise function $f$. Consider the following statements:
\begin{itemize}
\item[(1)] $f\in L^p(\Xi,d\xi)$ for some $p\in[1,2]$.
\item[(2)] $\tau$ is norm continuous under translations.
\item[(3)] $\lim_{\xi\to0}\cbnorm{\alpha_\xi^*{\mathcal T}-{\mathcal T}}=0$.
\item[(4)] For all $\omega\in{\mathrm C}^*\Xis_{\mathrm{in}}^*$, ${\mathcal T}\omega$ is norm continuous under translations.
\item[(5)] For all $A\in{\mathcal U}\Xis_{\mathrm{out}}$, ${\mathcal T}^*A\in{\mathcal C}_{\mathrm u}\Xis_{\mathrm{in}}$.
\end{itemize}
Then (1)$\Rightarrow$(2)$\Rightarrow$(3)$\Rightarrow$(4) and (5).
A channel with the property (3) will be called {\bf smoothing}.
\end{prop}
\begin{proof}
(1)$\Rightarrow$(2) is a part of Prop.~\ref{prop:contransl} applied to the noise state. (1)$\Rightarrow$(2) is an immediate consequence of items (3) and (4) of Cor.~\ref{cor:statechannel}, which also proves item (5) of that corollary. The remaining items follow from (3) by applying ${\mathcal T}$ or ${\mathcal T}^*$ to the respective arguments. For a converse at this point one would need a uniformity condition on the modulus of continuity: For example, demanding the existence of an $\omega$-independent function $\veps(\xi)$, which goes to zero as $\xi\to0$, and satisfies $\norm{\omega-(\alpha^*_\xi{\mathcal T})\otimes{\rm id}_n\omega}\leq\veps(\xi)$ for all states on ${\mathrm C}^*\Xis_{\mathrm{in}}\otimes{\mathcal M}_n$ is equivalent to (3).
\end{proof}
The term smoothing also suggests higher orders, and indeed the idea can be applied immediately to differentiability conditions. If $f$ is a Schwartz function, or even a Gaussian, as in most applications,
$\alpha_\xi^*\tau$ has Taylor approximations of all orders, where the coefficients are in ${\mathrm C}^*(\Xi_{\mathrm{out}},\Delta\sigma)$, and the error terms are likewise bounded in norm. This then transfers directly to $\alpha_\xi^*{\mathcal T}$ as well as all channel outputs.
\subsection{Squeezing and gauge invariance}
Squeezing is an important operation in quantum optics. It did not show up so far because we treated systems with the same number of quantum and classical degrees of freedom as equivalent. In quantum optics, however, there is an additional structure, the free time evolution/Hamiltonian. Closely related is the vacuum state, the unique ground state of the Hamiltonian.
A compact way to encode it is to specify a {\bf complex structure} on $\Xi$, i.e., a symplectic real-linear operator $I:\Xi\to\Xi$ so that $I^2=-\idty$, and
$\braket\xi\eta=\xi{\cdot}\sigma I\eta+ i\xi{\cdot}\sigma\eta$ turns $\Xi$ into an $n$-dimensional complex Hilbert space, in which multiplication by the scalar $i$ is interpreted as the operator $I$. The characteristic function of the vacuum is then $\chi(\xi)=\exp\bigl(-\braket\xi\xi/2\bigr)$. Very often this structure is assumed from the outset, i.e., the phase space $\Xi$ is taken to be a Hilbert space, and the symplectic form is the imaginary part of the scalar product.
Let us denote by $\gamma_t$ the quasifree channel with $S_t=\exp(It)=(\cos t) \idty+(\sin t)I$ and $f(\xi)\equiv1$, which is also called the group of {\it gauge transformations}. In the hybrid case, we extend it by $S_t=\idty$ on the classical variables. In the Schr{\"o}dinger\ representation of the quantum part, this is a unitarily implemented group generated by the number operator. This gives the Schr{\"o}dinger\ representation a Fock space structure with the Hilbert space $\Xi$ as the one particle space and a decomposition of the canonical operators $R_k$ into a creation and an annihilation part.
Our main reason for reiterating these well-known facts here is to avoid potential confusion about the term ``linear''. In the language of (quantum) optics, linear elements are beam splitters, phase shifters, and other passive dielectric components of an optical setup. They are indeed quasifree operations, i.e., described in terms of symplectic linear operations on phase space, but in addition, they commute with the free time evolution, i.e., they intertwine the respective complex structures $I_{\mathrm{in}}$ and $I_{\mathrm{out}}$: They are {\it complex} linear. Usually, this is applied to unitary channels only, but it is natural to demand that the noise function be gauge invariant as well. So we define a {\it gauge invariant channel} between phase spaces with complex structure as one with $\gamma_t^{\mathrm{out}}{\mathcal T}={\mathcal T}\gamma^{\mathrm{in}}_t$. This applies to states, as well, whereby gauge-invariant states are just those commuting with the number operator. For a channel it is somewhat weaker than particle number conservation: A damping channel ($S=e^{-\lambda t}\idty$, $f(\xi)=\exp(-(1-e^{-\lambda t})\norm\xi^2/2)$) is gauge invariant, but reduces particle number. Some of the interesting hybrid channels, e.g., position observables, are not gauge invariant.
Squeezing components are just those that are not gauge invariant, such as the preparation of a (perhaps Gaussian) state with different variances for $P$ and $Q$, or in quantum optics for the field quadratures. In order to achieve a certain task, for example, the preparation of an entangled state of two modes from a laser, it may be necessary to use squeezing elements. The framework for quantifying just how much squeezing is required (not counting any gauge invariant intermediate steps) is called the resource theory of squeezing \cite{Braunstein,Wolfquetsch}. It would carry us too far to include a systematic presentation here. A key element is the singular value decomposition, here also called the Bloch--Messiah decomposition \cite{Braunstein}, $S=S_1S_2S_3$ with $S_1,S_3$ orthogonal symplectic (non-squeezing) and $S_2$ purely squeezing, i.e., multiplication with a positive factor in some directions and multiplication with the inverse in the conjugate ones.
\subsection{Composition, concatenation, convolution}\label{sec:combine}
In this section, we briefly consider three ways of combining channels or states.
They correspond roughly to the parallel and serial execution of operations and to the addition of phase space variables. For none of these, a case distinction for different configurations of classical and quantum variables is needed. \\
\subsubsection*{Composition of subsystems}
Composition is usually rendered as the tensor product of Hilbert spaces or observable algebras. In our context, this indeed corresponds to C*-tensor products at the level of the non-unital algebras ${\mathrm C}^*\Xis{}$.
Given hybrid phase spaces $(\Xi_{j},\sigma_{j})$ with indices $j=1,2$, their composition is the hybrid phase space $(\Xi_{12},\sigma_{12})=(\Xi_{1}\oplus\Xi_{2},\sigma_{1}\oplus\sigma_{2})$. The C*-tensor product ${\mathrm C}^*(\Xi_{12},\sigma_{12})={\mathrm C}^*(\Xi_{1},\sigma_{1})\otimes{\mathrm C}^*(\Xi_{2},\sigma_{2})$ is uniquely defined, because the algebras involved are nuclear, so maximal and minimal C*-norm \cite{takesaki1} on the algebraic tensor product coincide. This entails that states in ${\mathrm C}^*(\Xi_{12},\sigma_{12})^*$ can be weakly approximated by product elements, but the resulting ``tensor product of state spaces'' requires more than norm limits of product elements. For observable algebras a simple approach using norm limits also fails (cf. Sect.~\ref{sec:tensprod}). It is clear that the compactification of a product is usually not the product of the compactifications. Even for the one-point compactification, corresponding to the observable algebras ${\mathcal A}_i={\mathrm C}^*(\Xi_i,\sigma_i)\oplus{\mathbb C}\idty$, we get additional components, like $\idty\otimes{\mathrm C}^*\Xis{_2}\not\subset{\mathrm C}^*(\Xi_{12},\sigma_{12})\oplus{\mathbb C}\idty$.
In spite of these subtleties, quasifree channels allow a straightforward composition operation ${\mathcal T}_1\otimes{\mathcal T}_2$. When $S_i:\Xi_{i,\mathrm{out}}\to\Xi_{i,\mathrm{in}}$ and $f_i:\Xi_{i,\mathrm{out}}\to{\mathbb C}$ are the data defining ${\mathcal T}_i$, the tensor product has
\begin{equation}\label{tensorcompose}
\begin{array}{rl}
S(\xi_1\oplus\xi_2)&=(S_1\xi_1)\oplus(S_2\xi_2), \\ f(\xi_1\oplus\xi_2)&=f_1(\xi_1)f_2(\xi_2).\end{array}
\end{equation}
Of course, with the composition of quantum systems comes {\bf entanglement}. It is the observation that, while general states on a composite system can be approximated by product elements, these product elements cannot be taken to be positive. Indeed, non-entangled or ``separable'' states are nowadays {\it defined} by the existence of a positive product approximation \cite{Popescu,HolevoInfdimEnt}.
Entanglement in Gaussian states is well understood \cite{MCF,eisert_gaussian_channels_2005,wolf_extremality_gaussian_states_2006,weedbrook_gaussian_2012,WolfWernerBoundEnt}, but the hybrid scenario creates no new interesting possibilities: The classical part of a composite hybrid is just the product of the classical parts. The pure classical states are point measures on a cartesian product, and hence product states. This is just saying that classical systems cannot be entangled. In the integral decomposition \eqref{disintegrate} of an arbitrary hybrid state, all entanglement is therefore in the states $\rho_x$, where $x=(x_1,x_2)$ is a point in the cartesian product.
This is true in spite of the recent proposal \cite{hall_two_2018} to use entanglement generation via a classical intermediary in a classical theory of gravity, see Sect.~\ref{sec:previous}.\\
\subsubsection*{Concatenation}
Executing one operation after the other is called, depending on community or context, concatenation, composition or multiplication. We use the first term, which derives from Latin for chaining because the second is too unspecific (see previous paragraph), and the third is too overloaded.
Clearly, when ${\mathcal T}_1,{\mathcal T}_2$ are quasifree channels, so is concatenation ${\mathcal T}={\mathcal T}_1{\mathcal T}_2$. When we take $S_1,S_2,S$ and $f_1,f_2,f$ as the defining parameters of these channels then
\begin{equation}\label{compose}
\begin{array}{rl}
S&= S_2S_1, \\ f(\xi)&= f_1(\xi)f_2(S_1\xi).\end{array}
\end{equation}
We thus get a {\bf category} whose objects are the hybrid systems and whose morphisms are the quasifree channels. Objects in a category are ``the same'' if they are connected by a morphism and its inverse morphism. Isomorphism classes in our setting are labelled by the pairs $(n,s)\in{\mathbb N}\times{\mathbb N}$, where $n$ is the number of quantum degrees of freedom, so that $\Xi_1={\mathbb R}^{2n}$ as a vector space on which $\sigma$ is non-degenerate, and $s$ is the number of classical dimensions, i.e., $\Xi_0=\ker\sigma={\mathbb R}^s$. Note that, in particular, our theory depends only on $\Xis{}$ and not on a particular splitting $\Xi=\Xi_1\oplus\Xi_0$. We have used such splittings above, although only $\Xi_0$, the null space of $\sigma$ is intrinsically defined by the structure $\Xis{}$, and different complements $\Xi_1$ could be chosen. Changing this splitting is an isomorphism leaving $\Xi_0$ fixed. It acts by an $\xi_0$-dependent phase space translation of the quantum part, which is clearly quasifree and invertible as such.
Other categorical features (monomorphism, epimorphisms, etc.) can be worked out.
An important result of this kind is a characterization of channels with a one-sided inverse, see Prop.~\ref{prop:retracts}.
A trivial but frequently used concatenation is the formation of {\bf marginals} of a channel, i.e., considering only one of the outputs and discarding the other (see Sect.~\ref{sec:instru} below). The discarding operation is itself a channel, the noiseless one with $S:\Xi_1\to\Xi_1\oplus\Xi_2$, $S\xi_1=\xi_1\oplus0$. Equivalently, it is the tensor product of the identity on $\Xi_1$ with the {\bf destructive channel} defined by $\Xi_{\mathrm{out}}=\{0\}$, and consequently $S:0\mapsto0\in\Xi_2$.
\subsubsection*{Convolution}
We have already met the convolution of states in Def.~\ref{def:stateconvolve}. As in all group representation theory, one should think of convolutions as a contravariant encoding of the group multiplication. So it is here: Let us consider two systems with the same set $\Xi=\Xi_1=\Xi_2$, so the addition of phase space elements makes sense. For the moment, we do not care whether they are classical or quantum. Can we {\bf add} signals of these types? The model for this is the addition of random variables. It corresponds to setting the Fourier arguments dual to the random variables $x_1$ and $x_2$ equal: The characteristic function for a sum is the expectation of $\exp(ik(x_1+x_2))$, which we obtain from that of the joint distribution. So convolution in general corresponds to the linear map $S\xi=\xi\oplus\xi\in\Xi_1\oplus\Xi_2$. So this would suggest a channel acting as ${\mathcal T}(\rho_1\otimes\rho_2)=\rho_1\ast\rho_2$. This works as a noiseless channel when one of the factors is classical. In the quantum case, however, although the convolution of arbitrary states is well defined, the map ${\mathcal T}$ with this property would not extend as a channel to entangled states. Thus one could either add noise or modify the definition by inverting the symplectic form in one factor, i.e., setting ${\mathcal T}(\rho_1\otimes\rho_2^\top)=\rho_1\ast\rho_2$, where $\rho^\top$ denotes transposition (= inversion of momenta) or the application of any other antiunitary quasifree symmetry.
This idea will be used in our analysis of teleportation (Sect.~\ref{sec:teleport}).
\subsection{Noiseless operations}\label{sec:noiseless}
Every quasifree channel can be modified by multiplying $f$ with an arbitrary (untwisted) positive definite function $g$. This corresponds to adding classical noise or averaging the output over translations $\alpha_\xi$ with a noise probability measure whose characteristic function is $g$. Since $\abs{g(\xi)}\leq1$ this always decreases $\abs f$. In fact, unless the noise measure is concentrated on a single point, and we thus have a simple translation, we have $\abs{g(\xi)}<1$ for some $\xi$ and the decrease of $\abs{f(\xi)}$ is strict, at least for some $\xi$.
A channel can thus be called a {\it minimal noise channel} if it cannot be constructed in this way with $\abs g\neq1$. Those channels will be characterized below as the extremal $S$-covariant channels. In the same spirit, we call a channel {\it noiseless}, if $\abs{f(\xi)}=1$, for all $\xi$, i.e., it is as large as consistent with any kind of twisted positive definiteness. These are characterized by the following proposition.\newpage
\begin{prop}\label{prop:noiseless} For a quasifree channel ${\mathcal T}$, specified by $S:\Xi_{\mathrm{out}}\to\Xi_{\mathrm{in}}$ and $f:\Xi_{\mathrm{out}}\to{\mathbb C}$, the following conditions are equivalent:
\begin{itemize}
\item[(1)] $\abs{f(\xi)}=1$ for all $\xi$, i.e., ${\mathcal T}$ is noiseless.
\item[(2)] $\Delta\sigma=0$, and there is some $\eta$ such that $f(\xi)=\exp{(i \xi{\cdot}\eta)}$.
\item[(3)] ${\mathcal T}^*:{\mathrm{CCR}}\Xis_{\mathrm{out}}\to{\mathrm{CCR}}\Xis_{\mathrm{in}}$ is a homomorphism.
\item[(4)] ${\mathcal T}^*:{\mathrm C}^*\Xis_{\mathrm{out}}^{**}\to{\mathrm C}^*\Xis_{\mathrm{in}}^{**}$ is a homomorphism.
\end{itemize}
\end{prop}
\begin{proof}
Let us begin by establishing an equivalent condition for (3) in terms of $S$ and $f$. Clearly, using the norm continuity of ${\mathcal T}^*$ it suffices to establish that ${\mathcal T}^*\bigl(W(\xi)W(\eta)\bigr)={\mathcal T}^*\bigl(W(\xi)\bigr){\mathcal T}^*\bigl(W(\eta)\bigr)$. Writing this out using
\eqref{Scovf} we get
\begin{equation}\label{homf}
f(\xi+\eta)=e^{i \xi{\cdot}(\Delta\sigma)\eta}\ f(\xi)f(\eta).
\end{equation}
Clearly, this is satisfied when (2) holds, proving (2)$\Rightarrow$(3). Moreover, \eqref{homf} implies that $\xi\mapsto \abs{f(\xi)}$ is a homomorphism. Since $f$ is twisted positive definite, we must have $\abs{f(\xi)}\leq1$, and by the homomorphism property $1=\abs{f(\xi)}\abs{f(-\xi)}$ so $\abs{f(\xi)}\geq1$, i.e.,
$\abs{f(\xi)}=1$. This shows that (3)$\Rightarrow$(1). The direction (1)$\Rightarrow$(2) follows immediately from Lem.~\ref{lem:constchi} below (see also the remark following the proof).
It remains to verify the equivalence (3)$\Leftrightarrow$(4). Here the direction (4)$\Rightarrow$(3) is trivial because
${\mathrm{CCR}}\Xis{}\subset{\mathrm C}^*\Xis{}^{**}$, and Weyl operators go to Weyl operators (see also the discussion in Sect.~\ref{Heisenqf}). For the converse direction, note that in a von Neumann algebra $x\to xy$ is weak*-continuous. So the relation ${\mathcal T}^*(AB)={\mathcal T}^*(A){\mathcal T}^*(B)$, which is assumed to hold for $A,B\in{\mathrm{CCR}}$ transfers to arbitrary $B\in{\mathrm C}^*\Xis_{\mathrm{out}}^{**}$ by weak*-continuity, and because ${\mathrm{CCR}}$ is weak*-dense. Repeating this argument for the first factor extends the relation to all $A,B$.
\end{proof}
The following lemma is needed in the proof, and we separated it because it is of independent interest.
\begin{lem}\label{lem:constchi}
Let $\Xis{}$ be a vector space with antisymmetric form, and suppose that $\chi$ is a normalized $\sigma$-twisted positive definite function on $\Xi$. Suppose that $\abs{\chi(\eta)}=1$ for some $\eta\neq0$.
Then $\sigma\eta=0$ and, for all $\xi\in\Xi$,
\begin{equation}\label{chiis1}
\chi(\xi+\eta)=\chi(\xi)\chi(\eta).
\end{equation}
\end{lem}
\begin{proof}
Consider a $3\times3$-matrix $M$ of the form \eqref{twistedPosDef}. Abbreviating the matrix entries as
$M_{k\ell}=\chi(\xi_k-\xi_\ell)\,\exp({\textstyle \frac i2 \sigma(\xi_k,\xi_\ell)})$, it is of the form
\begin{equation}\label{mm3}
M=\begin{pmatrix}1& M_{12}&\overline{M_{31}}\\ \overline{M_{12}}&1&M_{23}\\ M_{31}&\overline{M_{13}}&1
\end{pmatrix}.
\end{equation}
Its determinant is \begin{equation}
0\leq \det M=1+M_{12}M_{23}M_{31}+ \overline{M_{12}M_{23}M_{31}}- \abs{M_{12}}^2- \abs{M_{23}}^2- \abs{M_{31}}^2.
\end{equation}
Now take the triple of vectors as $(-\eta,0,\xi)$, where $\xi\in\Xi$ is arbitrary. Then
$M_{12}=\chi(-\eta)=\overline{\chi(\eta)}$, $M_{23}=\overline{\chi(\xi)}$, $M_{31}=\chi(\xi+\eta)\exp(-\frac{i}2\xi{\cdot}\sigma\eta)$.
In particular, $\abs{M_{12}}=\abs{\chi(\eta)}=1$, so this expression simplifies to
\begin{equation}
0\leq\det M= -\bigl\vert M_{31}-\overline{M_{12}}\,\overline{M_{23}}\bigr\vert^2.
\end{equation}
This can only be positive if the absolute value vanishes, which means that
\begin{equation}
\chi(\xi+\eta)=\chi(\xi)\chi(\eta)e^{\frac{i}2\xi{\cdot}\sigma\eta}.
\end{equation}
Changing $\xi\mapsto-\xi$, and $\eta\mapsto-\eta$, which also satisfies the assumption of the lemma, every characteristic function in the last expressions changes to its complex conjugate, while the exponent does not. Hence the exponential factor has to be $1$. Since $\xi\mapsto\lambda\xi$ is also allowed, the exponent has to be zero.
\end{proof}
This shows that the maximal absolute value of $\chi$ can only be reached on the classical subsystem. We have not assumed that $\abs{\chi(\lambda\eta)}=1$ also holds for all scalar multiples $\lambda\eta$ as well. If that is the case, and $\chi$ is continuous, then $\chi(\eta)=\exp(i\mu{\cdot}\eta)$ for some $\mu\in\Xi$. We remark that this assumption may fail, and so, even in $1$ dimension, we cannot conclude from the assumptions of the lemma that $\chi$ is the characteristic function of a point measure. For example, the classical characteristic function of a measure supported by the integers is $2\pi$-periodic, so $\chi(2\pi)=1$, but except for a point measure we have $\abs{\chi(\eta)}<1$ for $0<\eta<2\pi$.
Let us recapitulate which of the basic operations are noiseless.
\begin{itemize}
\item The {\it states} with the homomorphism property, i.e. ($\omega(AB)=\omega(A)\omega(B)$), are only the pure states of classical systems, corresponding to point measures on $\Xi=\Xi_0$. Noiseless quantum states do not exist, which also excludes such states on hybrids with non-vanishing $\sigma$.
\item Noiseless {\it observables} are the projection valued ones: The homomorphism property implies $F(M)^2={\mathcal T}^*(\chi_M)^2={\mathcal T}^*(\chi_M^2)={\mathcal T}^*(\chi_M)$. When the observable is considered as acting on a function algebra ${{\mathcal C}_{\mathrm b}}(\Xi_{\mathrm{out}})$ this is the property of having a von Neumann-style functional calculus,
${\mathcal T}^*(\Phi(A))=\Phi({\mathcal T}^*(A))$ for $\Phi:{\mathbb R}\to{\mathbb R}$. That is, postprocessing of outcomes with a function $\Phi$ is the same as applying this function to the operator in the functional calculus.
\item{}
Noiseless channels from an irreducible quantum system to itself act by unitary transformation, where the unitary operator belongs to the {\it metaplectic representation} \cite{Gosson} of the affine symplectic group.
\end{itemize}
\noindent The following proposition characterizes a further class of noiseless channels, namely those with a right inverse in the Schr{\"o}dinger\ picture.
\begin{prop}\label{prop:retracts}
Let ${\mathcal T}_1:{\mathrm C}^*\Xis{_1}^*\to{\mathrm C}^*\Xis{_2}^*$ and ${\mathcal T}_2:{\mathrm C}^*\Xis{_2}^*\to{\mathrm C}^*\Xis{_1}^*$ be quasifree channels such that ${\mathcal T}_1{\mathcal T}_2={\rm id}$. Then
\begin{itemize}
\item[(1)] ${\mathcal T}_1$ is noiseless, and $S_1:\Xi_2\to\Xi_1$ is injective,
\item[(2)] ${\mathcal T}_2$ is an expansion, i.e., there is a system $(\Xi_{\mathrm{e}},\sigma_{\mathrm{e}})$ such that there is an isomorphism
$$\Xis{_1}\cong(\Xi_2\oplus\Xi_{\mathrm{e}},\sigma_2\oplus\sigma_{\mathrm{e}}),$$
and ${\mathcal T}_2={\rm id}_2\otimes {\mathcal P_{\mathrm{e}}}$, where ${\mathcal P_{\mathrm{e}}}$ is a preparation of a $(\Xi_{\mathrm{e}},\sigma_{\mathrm{e}})$-system,
\item[(3)] under the isomorphism from (2), $S_1:\Xi_2\to\Xi_2\oplus\Xi_{\mathrm{e}}$ is the embedding into the first summand.
\end{itemize}
Moreover, for $i=1,2$ if a channel ${\mathcal T}_i$ satisfies the condition (i), then there is a channel ${\mathcal T}_{3-i}$, so that ${\mathcal T}_1$ and ${\mathcal T}_2$ satisfy all the above conditions.
\end{prop}
\begin{proof}
The composition relation \eqref{compose} gives that $\idty=S_2S_1$ and $f_1(\xi)f_2(S_1\xi)=1$. Since $\abs{f_i(\xi)}\leq1$ by positive definiteness, we must have $\abs{f_1(\xi)}=\abs{f_2(S_1\xi)}=1$ for all $\xi$.
In particular, ${\mathcal T}_1$ must be noiseless, and since $S_1$ has a left inverse, it is injective. This shows (1).
Now consider (2). $f_2$ is twisted positive definite for $\Delta\sigma=\sigma_1-S_2^\top\sigma_2S_2$. Moreover, on the subspace $S_1\Xi_1$ this function has the maximal modulus, so by Lem.~\ref{lem:constchi} the range of $S_1$ is in the null space of $\Delta\sigma$. This is equivalent to the matrix equation $(\Delta\sigma)S_1=0$. Using $S_2S_1=\idty$ gives
\begin{equation}\label{fkha}
\sigma_1S_1=S_2^\top\sigma_2.
\end{equation}
Since $S_1S_2$ is an idempotent operator, every $\xi\in\Xi_1$ is naturally split as $\xi=S_1S_2\xi+(\idty-S_1S_2)\xi$, where the first summand is obviously in the range $S_1\Xi_2$ and the second satisfies $S_2(\idty-S_1S_2)\xi=(S_2-S_2)\xi=0$. Therefore, by \eqref{fkha}$^\top$, these parts are $\sigma_1$-orthogonal:
\begin{equation}\label{hyposympdirsum}
S_1\xi\cdot\sigma_1(\idty-S_1S_2)\eta=\xi\cdot\sigma_2S_2(\idty-S_1S_2)\eta=0.
\end{equation}
Moreover, $S_1:\Xi_2\to\Xi_1$ is an isomorphism onto its range, changing $\sigma_2$ to the restriction of $\sigma_1$. This proves the decomposition with $\Xi_{\mathrm{e}}=(\idty-S_1S_2)\Xi_1$ and $\sigma_{\mathrm{e}}$ the restriction of $\sigma_1$ to this subspace. The action of $S_2$ is very simple in these terms: It acts separately on the two summands, which makes the corresponding channel a tensor product. On the first summand, $S_1\Xi_2$, it just inverts the isomorphism~$S_1$. Hence, after identifying the $2$ subsystem of $\Xi_1$ with $\Xi_2$, it acts like the identity channel on this part. The second summand $\Xi_{\mathrm{e}}$ is annihilated by $S_2$, which is the hallmark of a preparation (see above). The state prepared lives on $(\Xi_{\mathrm{e}},\sigma_{\mathrm{e}})$ and has characteristic function $\chi_{\mathrm{e}}(\xi_{\mathrm{e}})=f_2(\xi_{\mathrm{e}})$.
\end{proof}
\subsection{Noise factorization and dilations}\label{sec:noisefac}
The Stinespring dilation is one of the most powerful tools in quantum information theory. In the standard setting, it is a structure theorem for completely positive maps ${\mathcal T}^*:{\mathcal A}_{\mathrm{out}}\to{\mathcal B}({\mathcal H}_{\mathrm{in}})$, where we have added the star on ${\mathcal T}$ and the labels ``in'' and ``out'' to be consistent with the above notation. It provides an additional Hilbert space ${\mathcal K}$, an isometry $V:{\mathcal H}_{\mathrm{in}}\to{\mathcal K}$, and a representation $\pi:{\mathcal A}_{\mathrm{out}}\to{\mathcal B}({\mathcal K})$ such that
\begin{equation}\label{stinespring}
{\mathcal T}^*(A)=V^*\pi(A)V.
\end{equation}
For a quasifree channel, ${\mathcal T}^*$ will map some subalgebra ${\mathcal A}_{\mathrm{out}}\subset {\mathrm C}^*\Xis_{\mathrm{out}}$ into a representation of a subalgebra ${\mathcal A}_{\mathrm{in}}\subset {\mathrm C}^*\Xis_{\mathrm{in}}$, so that there are many choices to be made, and consequently many variations on the dilation theme. All these variations have the structure of factorizations through an intermediate system, ${\mathcal B}({\mathcal K})$. The first step (in the direction from input to output) is the embedding of the input states in ${\mathcal H}_{\mathrm{in}}$ into this ``larger system'', an expansion. The second step, done here by the representation $\pi$, is a noiseless operation in the sense of the previous paragraph. These features can be phrased entirely in the category of quasifree maps. What is more, the factorization can be done for arbitrary quasifree channels. This is the content of Thm.~\ref{thm:noisedec}. We will discuss later how it relates to Stinespring-like results.
Note that the channels are written here in the Schr{\"o}dinger\ picture, so in the factorization ${\mathcal T}={\mathcal T}_N{\mathcal T}_E$, the expansion ${\mathcal T}_E$ is applied to the physical system first, and ${\mathcal T}_N$ acts on the expanded system. If we write the expansion channel as tensoring with a fixed state $\omega_E$, the factorization is written as
\begin{equation}\label{factorChan}
{\mathcal T}\omega={\mathcal T}_N(\omega\otimes\omega_E).
\end{equation}
\begin{thm}\label{thm:noisedec}
Every quasifree channel can be decomposed into ${\mathcal T}={\mathcal T}_N{\mathcal T}_E$, where ${\mathcal T}_E$ is an expansion and ${\mathcal T}_N$ is a noiseless channel. The phase space of the extension system
is $\Xi_\Delta=\Xi_{\mathrm{out}}$ as a vector space, but with antisymmetric form $\Delta\sigma=\sigma_{\mathrm{out}}-S^\top\sigma_{\mathrm{in}} S$. The salient linear maps and noise functions are
\begin{equation}\label{gaussdilate}
\begin{array}{lll}
S_N:\Xi_{\mathrm{out}}\to\Xi_{\mathrm{in}}\oplus\Xi_\Delta \qquad &S_N\xi=S\xi\oplus\xi & f_N(\xi)=1\\
S_E:\Xi_{\mathrm{in}}\oplus\Xi_\Delta\to\Xi_{\mathrm{in}} &S_E(\xi_1\oplus\xi_2)=\xi_1 \qquad&f_E(\xi_1\oplus\xi_2)=f(\xi_2).
\end{array}
\end{equation}
\end{thm}
\begin{proof}
Let us first verify that the given data for ${\mathcal T}_N$ and ${\mathcal T}_E$ satisfy the positivity condition for channels. The respective difference forms are
\begin{eqnarray}
\xi{\cdot}(\Delta\sigma)_N\eta&=& \xi{\cdot}\sigma_{\mathrm{out}}\eta- (S_N\xi){\cdot}(\sigma_{\mathrm{in}}\oplus\sigma_\Delta)S_N\eta\nonumber\\
&=& \xi{\cdot}\sigma_{\mathrm{out}}\eta- (S\xi){\cdot}\sigma_{\mathrm{in}} S\eta-\xi{\cdot}\sigma_\Delta\eta=0 \label{delNdilate} \\[7pt]
(\xi_1\oplus\xi_2){\cdot}(\Delta\sigma)_E(\eta_1\oplus\eta_2)&=&(\xi_1\oplus\xi_2){\cdot}(\sigma_{\mathrm{in}}\oplus\sigma_\Delta)(\eta_1\oplus\eta_2)
- S_E(\xi_1\oplus\xi_2){\cdot}\sigma_{\mathrm{in}} S_E(\eta_1\oplus\eta_2) \nonumber\\
&=& \xi_1{\cdot}\sigma_{\mathrm{in}}\eta_1+ \xi_2{\cdot}\sigma_\Delta\eta_2- \xi_1{\cdot}\sigma_{\mathrm{in}}\eta_1=\xi_2{\cdot}\sigma_\Delta\eta_2. \label{delEdilate}
\end{eqnarray}
Hence $(\Delta\sigma)_N=0$, as required of a noiseless channel, which makes $f=1$ a legitimate choice. Moreover, $(\Delta\sigma)_E=(0\oplus\sigma_\Delta)$, which is exactly the noise function for which $f_E$ has to be twisted positive definite. Hence ${\mathcal T}_N$ and ${\mathcal T}_E$ are well defined.
It remains to verify the concatenation relation. Of course, the product of two channels in our class is again in the class, and there is a simple general formula for the data $(S',f')$ of the product. By \eqref{compose}, this gives
\begin{equation}\label{prodchan}
S'=S_ES_N=S \quad\mbox{and}\quad f'(\xi)=f_E(S_N\xi)f_N(\xi)=f(\xi).
\end{equation}
Hence we have ${\mathcal T}_N{\mathcal T}_E={\mathcal T}$, as claimed.
\end{proof}
When only one fixed channel ${\mathcal T}$ is under consideration, the above representation may be very wasteful. For example, when ${\mathcal T}$ is itself noiseless, one can clearly choose ${\mathcal T}_E$ to be the identity, and there is no need to adjoin an additional system $\Xi_\Delta$, i.e., \eqref{gaussdilate} is not a ``minimal'' factorization. In order to get to a move in that direction, for a general quasifree channel, consider the noise function $f:\Xi_{\mathrm{out}}\to{\mathbb C}$. Let $N\subset\Xi_\Delta$ denote the largest subspace on which $\abs{f(\xi)}=1$. Then by Lem.~\ref{lem:constchi}, $f$ is a character on $N$, and hence of the form $f_1(\xi)=\exp(i\lambda{\cdot}\xi)$ for some $\lambda\in\Xi_{\mathrm{out}}$. Here $\lambda$ is not uniquely determined, but that is also not needed. The remainder $f'(\xi)=f(\xi)/f_1(\xi)$ is then a legitimate noise function with $f'(\eta+\xi)=f'(\eta)$ for $\xi\in N$. We may therefore consider $f'$ as a function $f_{\mathrm{mid}}$ on the quotient $\Xi_{\mathrm{mid}}=\Xi_{\mathrm{out}}/N$. Denoting the quotient map by $S_{\mathrm{mid}}:\Xi_{\mathrm{out}}\to\Xi_{\mathrm{mid}}$, this amounts to $f(\xi)=f_1(\xi)f_{\mathrm{mid}}(S_{\mathrm{mid}}\xi)$. By Lem.~\ref{lem:constchi}, $N$ is also contained in the null space of $\Delta\sigma$, so this form also passes to the quotient as $\sigma_{\mathrm{mid}}$.
This gives an alternative noise factorization ${\mathcal T}={\mathcal T}_N{\mathcal T}_E$, closely related to \eqref{gaussdilate}, but with the intermediate system $(\Xi_\Delta,\Delta\sigma)$ replaced by $(\Xi_{\mathrm{mid}},\sigma_{\mathrm{mid}})$,
\begin{equation}\label{gaussdilmin}
\begin{array}{lll}
S_N:\Xi_{\mathrm{out}}\to\Xi_{\mathrm{in}}\oplus\Xi_{\mathrm{mid}} \qquad &S_N\xi=S\xi\oplus S_{\mathrm{mid}}\xi & f_1(\xi)=\exp(i\lambda{\cdot}\xi)\\
S_E:\Xi_{\mathrm{in}}\oplus\Xi_{\mathrm{mid}}\to\Xi_{\mathrm{in}} &S_E(\xi_1\oplus\xi_2)=\xi_1 \qquad&f_E(\xi_1\oplus\xi_2)=f_{\mathrm{mid}}(\xi_2).
\end{array}
\end{equation}
The map $S_N$ in this construction is connected to the previous one by another noiseless channel based on the quotient map $S_{\mathrm{mid}}$. That is, the modification just described moves in the direction of including as much of the channel into the noiseless part as possible. This follows a categorical approach described in \cite{Westerbaan} as the {\it Paschke dilation}. This generalizes the Stinespring construction to the category of W*-algebras with normal completely positive maps, when the range in the Heisenberg picture (corresponding to the input in physical terms) is no longer of the form ${\mathcal B}({\mathcal H})$. Fig.~\ref{fig:Paschke} summarizes the defining minimality condition.
\begin{figure}[ht]
\begin{center}
\begin{tikzpicture}[scale=0.9]
\node[scale=1.1] at (0,0) {${\mathcal A}$};
\node[scale=1.1] at (0.9,-1.25) {${\mathcal D}$};
\node[scale=1.1] at (3.1,1.25) {${\mathcal E}$};
\node[scale=1.1] at (4,0) {${\mathcal B}$};
\draw[very thick,->] (0.3,0.25)--(2.7,1.20);
\draw[very thick,->] (0.3,-0.3)--(0.7,-0.95);
\draw[->] (0.3,0)--(3.7,0);
\draw[very thick,->] (1.15,-0.95)--(2.85,0.95);
\draw[dashed,->] (1.3,-1.2)--(3.7,-0.25);
\draw[dashed,->] (3.3,0.95)--(3.7,0.3);
\end{tikzpicture}
\captionsetup{width=0.8\textwidth}
\caption{Comparison of factorizations of a morphism ${\mathcal A}\to{\mathcal B}$ in the category of von Neumann algebras with normal completely positive unital maps.
The thick arrows represent *-homomorphisms, corresponding to noisefree channels. The upper factorization has a ``larger'' noisefree factor.
The Paschke dilation is defined as the one with the largest noisefree factor. }
\label{fig:Paschke}
\end{center}
\end{figure}
A Paschke dilation, or a factorization ${\mathcal T}={\mathcal T}_N{\mathcal T}_E$ can be turned into a Stinespring dilation in the usual sense by taking the input algebra as faithfully represented on a Hilbert space ${\mathcal H}$, so in the Heisenberg picture the channel maps into ${\mathcal B}({\mathcal H})$, and hence we realize the standard setting for the Stinespring construction. This step hardly depends on the hybrid structure: We will only use that ${\mathcal T}_E$ is an expansion, and ${\mathcal T}_N$ is a *-homomorphism. As seen from the $\mu$-free setting, this requires the choice of some measure $\mu$ on the classical part of the input phase space. Then ${\mathcal H}={\mathcal H}_\mu={\mathcal H}_1\otimes L^2(\Xi_{\mathrm{in},0},\mu)$ is the space of a standard representation (see Sect.~\ref{sec:stdrep}), and the input von Neumann algebra is ${\mathcal A}_{\mathrm{in}}={\mathcal B}({\mathcal H}_1){\overset{{\rule{5pt}{1pt}}}\otimes} L^\infty(\Xi_{\mathrm{in},0},\mu)\subset{\mathcal B}({\mathcal H}_\mu)$. On the output side, we define the von Neumann algebra ${\mathcal A}_{\mathrm{out}}$ as the weak closure of the GNS representation of an output state $\omega_{\mathrm{out}}={\mathcal T}\omega_{\mathrm{in}}$, when $\omega_{\mathrm{in}}$ is a suitable faithful input state. The expansion state is some normal state $\omega_E$ on a von Neumann algebra ${\mathcal A}_E$. Thus ${\mathcal T}_E\rho=\rho\otimes\omega_E$, a state on ${\mathcal A}_{\mathrm{in}}\otimes{\mathcal A}_E$, and the factorization uses some normal *-homomorphism ${\mathcal T}_N^*:{\mathcal A}_{\mathrm{out}}\to{\mathcal A}_{\mathrm{in}}{\overset{{\rule{5pt}{1pt}}}\otimes}{\mathcal A}_E$.
Now, let $({\mathcal H}_E,\pi_E,\Omega_E)$ denote the GNS-representation of the expansion state. Then we define
\begin{equation}\label{Vexpand}
\begin{array}{lll}
V:{\mathcal H}\to{\mathcal H}\otimes{\mathcal H}_E, &\quad\text{with}\ V\phi=\phi\otimes\Omega_E, \\
\pi:{\mathcal A}_{\mathrm{out}}\to{\mathcal B}({\mathcal H}\otimes{\mathcal H}_E), &\quad\text{with}\ \pi=({\rm id}\otimes\pi_E){\mathcal T}^*.
\end{array}
\end{equation}
Then $V^*({\rm id}\otimes\pi_E)(A\otimes B)V=\omega_E(B) A={\mathcal T}_E(A\otimes B)$, and composing that with ${\mathcal T}_N^*$ we get ${\mathcal T}^*(X)={\mathcal T}^*_E{\mathcal T}^*_N(X)=V^*({\rm id}\otimes\pi_E)({\mathcal T}^*_N(X))V=V^*\pi(X)V$.
Hence, \eqref{Vexpand} defines a Stinespring dilation, which is, however, usually not the unique minimal one. Of course, the minimal representation is contained in this by choosing an appropriate subspace of ${\mathcal H}_\mu\otimes{\mathcal H}_E$ and restricting $\pi$ accordingly. Since that depends on further details of the channels involved, we will not pursue this here.
\section{Basic physical operations}\label{sec:BasicOps}
In the unified picture given here, every operation, including preparations and measurements, is given by a quasifree channel. The purpose of this section is to advertise this unification by showing how basic quantum operations fit into the framework. We will assume only the basic definitions (Sect.~\ref{sec:defqf}) and the parametrization of channels by a linear map $S:\Xi_{\mathrm{out}}\to\Xi_{\mathrm{in}}$ and the noise function $f$, respectively the noise state $\tau$. Typically, $S$ specifies the kind of operation one is considering, the number of classical/quantum inputs/outputs, and how they are basically related. It will typically be fixed at the beginning of each of the subsections below. This fixes a hybrid system $(\Xi_{\mathrm{out}},\Delta\sigma)$, and hence the possible noise states $\tau$, respectively noise functions $f$. While the knowledge of the definitions suffices to verify how the respective examples fit in the general framework, we do sometimes draw on the general results above or illustrate them in the particular case.
\subsection{States}
States are the mathematical description of a system preparation. The input system is, therefore, a trivial one, $\Xi_{\mathrm{in}}=\{0\}$. Hence $S=0$, and $\Delta\sigma=\sigma_{\mathrm{out}}$. The positivity condition for $f$ thus demands that $f$ is a characteristic function of a standard state for the out-system. There is no further condition, i.e., {\it all states} are quasifree channels in this sense. We caution the reader that this is in contrast to another well-established use of the term, by which only Gaussian states are called ``quasifree'' \cite{qfstates,fannes}.
In the theory of channel capacity, e.g., for the Holevo bound \cite{Holevo73}, \cite[12.3]{Wilde}, one needs {\bf state ensembles} (or ``assemblages''), usually written as a collection of states with probability weights. When hybrids are considered as systems in their own right, this is just the same as a state on a hybrid. This view of state ensembles naturally extends also to continuous ensembles, in which the convex weights are replaced by a non-discrete measure.
\subsection{Disturbance}
The word disturbance always refers to a situation deviating from an ideal, in our case, a deviation from the identity channel. That is, we look at how much the output states differ from the inputs. This requires
input and output to be of the same type, i.e., $\Xis_{\mathrm{in}}=\Xis_{\mathrm{out}}=\Xis{}$, and since the ideal channel, or `no disturbance', should be a special case, we choose $S=\idty$.
Then $\Delta\sigma=0$, so the condition on the noise function $f$ is the classical Bochner condition. Hence $f$ is the Fourier transform of a probability measure $\tau$, appropriately called the noise measure. The channel acts as
\begin{equation}\label{noisechan}
{\mathcal T}\omega=\int\tau(d\xi)\ \alpha_\xi(\omega)=\omega\ast\tau,
\end{equation}
where the convolution is taken in the sense of Def.~\ref{def:stateconvolve}.
The size of the noise can be ascertained in different ways. A norm bound (cf.~Cor.~\ref{cor:statechannel}) is $\cbnorm{{\rm id}-{\mathcal T}}=\sup_\rho\norm{\rho-{\mathcal T}\rho}_1=\norm{\delta_0-\nu}_1$, where $\delta_0$ is the point measure at $0$, and the last norm is the norm for classical states, also known as the total variation. If we decompose $\nu=(1-\lambda)\delta_0+\lambda\nu'$ for some probability measure with $\nu'(\{0\})=0$ we get
$\norm{\delta_0-\nu}_1=2\lambda$. That is, this norm measure of noise is only small if we have a large convex component of ${\mathcal T}$ which is equal to ${\rm id}$. In particular, a channel that introduces a small shift ($\nu=\delta_\xi$, $\xi\approx0$) is always at a maximal distance. Better measures of the noise for many purposes are variances or, more generally, transport distances \cite{Villani}. In many cases, it is not necessary to condense the size of the noise into a single number, and the most accurate description is the noise measure itself.
\subsection{Observables}
An {observable} is a channel with classical output, i.e., $\sigma_{\mathrm{out}}=0$, and $\Xi_{\mathrm{out}}$ is the space of measurement outputs. In the quasifree setting, the observable automatically gets a covariance property with respect to shifts of the outputs.
The theory laid out in Sect.~\ref{sec:funcobs} shows that the two ways of looking at an observable, namely as a positive operator-valued measure (POVM) on the one hand and as an operator on continuous functions on the other, are equivalent Heisenberg pictures of such a channel. For the POVM view, we have to identify, for every measurable set $M\subset\Xi_{\mathrm{out}}$, an effect operator $F(M)$ on the input system. Thus we need a Heisenberg picture map ${\mathcal T}^*$ which is well defined on the indicator function $1_M\in{\mathcal U}(\Xi_{\mathrm{out}},0)$. The appropriate Heisenberg picture is thus ${\mathcal T}^*:{\mathcal U}(\Xi_{\mathrm{out}},0)\to{\mathcal U}\Xis_{\mathrm{in}}$, and the positive operator-valued measure describing the observable is $F(M)={\mathcal T}^*(1_M)$. Equivalently, we can consider the observable as a map on bounded continuous functions $\phi\in M(\Xi_{\mathrm{out}},0)$, and ${\mathcal T}^*\phi=\int F(d\xi) \phi(\xi)$.
The further characterization of the class of covariant observables so described depends on the range of $S:\Xi_{\mathrm{out}}\to\Xi_{\mathrm{in}}$, and especially on the restriction of $\sigma_{\mathrm{in}}$ to the range $S\Xi_{\mathrm{out}}$.
This is basically the question of whether the quantities measured are subject to a quantum uncertainty constraint or not. We will consider the two extreme cases, a position observable and a phase space or position and momentum observable, separately below. In either case, we take $S$ to be injective because otherwise, we would have directions in the output space that have distributions not depending on the quantum input.
By virtue of \eqref{trcov}, quasifree observables fit into the framework of observables covariant with respect to a projective unitary representation of a group $G$. In this traditional subject, \cite{Davies,holevo_probabilistic_book}, the basic construction of all covariant observables uses a covariant version \cite{scutaru} of the Stinespring dilation (called Naimark's dilation for the case of classical output) to reduce the construction to the noiseless, i.e., projection valued case, which is then solved by Mackey's theory of induced representations (see \cite{cattaneo}, and \cite{screen} for a worked example). What has apparently not been considered in detail was the nature of the noise. In our framework, there is a clear distinction of the position vs.\ the phase space case, requiring classical vs.\ quantum noise. We, therefore, treat these cases separately below.
A traditional subject in the general theory is the existence of a direct formula for the output probability density at a point in the outcome set. If such a formula exists, it will be given by the expectation value of a positive possibly unbounded operator, which is called the operator-valued Radon--Nikodym\ density the observable (see e.g., \cite[Sect.~IV.2.]{holevo_probabilistic_book}, \cite[Sect.~I.5.G]{Schroeck}, and \cite[Thm.~4.5.2]{Davies} for the compact group case). That is, we are looking for a family of positive, possibly unbounded operators $\dot F(x)$ such that the observable is expressed as
\begin{equation}\label{RNposition}
{\mathcal T}^*(g)=\int dx\ \dot F(x)\ g(x).
\end{equation}
Recall that, following \eqref{alpha}, our convention for the action of translations is $(\alpha_xg)(y)=g(x+y)$. So the covariance condition \eqref{trcov} translates to $\alpha_{\xi}\bigl(\dot F(x)\bigr)=\dot F(x-S^\top\xi)$. Now since $S$ is injective, $S^\top$ is onto, so this equation determines the function $\dot F(x)$ from one of the values, say $\dot F(0)=:\dot F$:
\begin{equation}\label{FdotCov}
\dot F(x)=\alpha_{-\xi}(\dot F),\quad\text{for any }\xi\in\Xi_{\mathrm{in}}\ \text{ such that}\ S^\top\xi=x.
\end{equation}
Since $S^\top$ might have a kernel, this also implies the invariance of $\dot F$ under $\alpha_\xi$ with $S^\top\xi=0$.
Of course, there is also an expression for $\dot F$ in terms of the noise function $f$, since both quantities determine the observable. For that we put $g(x)=\exp(ik{\cdot} x)$, i.e., $g=W_{\mathrm{out}}(k)$, in the above equation, and solve for $\dot F$ by inverse Fourier transform. With $n={\dim\Xi_{\mathrm{out}}}$ we get:
\begin{equation}\label{FdotInt}
\dot F= \frac1{(2\pi)^n}\int dk f(k) W_{\mathrm{in}}(Sk).
\end{equation}
In general, e.g., for the canonical position observable, neither $\dot F$ nor this integral makes sense. However, with sufficient noise, seen by the decay of $f$ at infinity, both do.
\subsubsection*{Position observables}
The canonical {position observable} of a purely quantum system belongs to the selfadjoint operators $Q_j$ from Sect.~\ref{sec:setup}. The characteristic function of the output probability distribution is hence the expectation of $\exp(ik\cdot Q)=W(k,0)$. So this is quasifree with $\Xi_{\mathrm{out}}=\{k\}={\mathbb R}^n$ and $Sk=(k,0)$, when the variables are arranged as in Sect.~\ref{sec:setup}. Of course, one could also include some classical hybrid variables. Since the noise function vanishes, the observable is projection valued, which can be said in two equivalent ways, namely that $F(M)$ is always a projection or that ${\mathcal T}^*$ is a homomorphism (also compare Prop.~\ref{prop:noiseless}). For any input density operator $\rho$, we write ${\mathcal T}\rho=\rho^Q$, and call it the position distribution of $\rho$. Similarly, we define $\rho^P$ as the momentum distribution.
The beauty of the quasifree formalism is here that it automatically includes noisy versions. These are characterized by choosing the same $S$, but allowing $f$ to be more general. This defines the class of generalized position observables, which share the covariance condition with the canonical one. The structure theory is then immediate: Since $\Delta\sigma=0$ the noise is necessarily classical, so the most general position observable has the output distribution $\nu\ast\rho^Q$, where $\nu$ is some fixed noise measure on position space which is independent of $\rho$, and $\rho^Q$ is the output distribution of the standard position observable. Thus we can always think of such a measurement as executing the standard one and then adding, from a statistically independent source, noise with distribution~$\nu$.
When the noise distribution has a Radon--Nikodym\ density $\dot\nu$ with respect to the Lebesgue measure, we have $\dot F=\dot\nu(Q)$ in the functional calculus of the commuting selfadjoint operators $Q_k$. In contrast, for the canonical observable itself, the expectation of $\dot F$ in the state vector $\Psi$ should be $\abs{\Psi(x)}^2$, which might be given a meaning as a sesquilinear form on Schwartz space. But there is no closable operator $\dot F$ corresponding to this. This is also seen in the difficulty of making sense of \eqref{FdotInt}.
\subsubsection*{Phase space observables}
Here we demand a joint measurement of all positions and momenta. So we have $\Xi_{\mathrm{out}}=\Xi_{\mathrm{in}}$ and $S=\idty$, but the symplectic forms are different, namely the standard quantum one on $\Xi_{\mathrm{in}}$ and $0$ on $\Xi_{\mathrm{out}}$. Hence, $\Delta\sigma=\sigma_{\mathrm{in}}$, and the admissible noise functions are exactly the characteristic functions of quantum states $\tau$.
Hence the relation \eqref{Tchi} is exactly that for a {\bf convolution} of quantum states in the sense of \cite{QHA} and Def.~\ref{def:stateconvolve}. When $\tau$ is the quantum state defining the observable, and $\rho$ is the input state, the output distribution is thus $\tau\ast\rho$. Comparing the expression \eqref{convolvetofct} with \eqref{RNposition} we find the Radon--Nikodym\ density of the POVM to be
\begin{equation}\label{densityPS}
\dot F=\beta_-(\tau).
\end{equation}
This is a density operator in two different meanings of the word: A Radon--Nikodym\ density, and also a positive operator with trace $1$ (choosing the correct, cf. \cite{QHA}) normalization of phase space Lebesgue measure). This characterization of covariant phase space observables is well-known \cite{Davies,QHA,holevo_probabilistic_book}. The Gaussian special case is known in quantum optics as the Husimi distribution or Q-function of $\rho$. But as the quasifree formalism clearly indicates, any $\tau$, pure or mixed, will work analogously.
Of course, such a joint position/momentum measurement necessarily includes errors, which is the subject of {\bf measurement uncertainty} relations \cite{BLW}. By this, we mean any relation expressing that one can either get a fairly good position measurement with large errors for momenta or conversely. For uncertainty relations, the covariance condition is an unwanted restriction, but the proof of the general case \cite{BLW} works via showing that among the optimal solutions, there is always a covariant one. This makes the tradeoffs extremely easy to describe. Indeed, the position marginal of the output distribution is $(\tau\ast\rho)^Q=\tau^Q\ast\rho^Q$, a relation which is shown by setting one set of variables equal to zero in the product of characteristic functions of $\tau$ and $\rho$. In other words, the position marginal of phase space observable is a noisy position observable. That statement is obvious from the covariance conditions, but here we also learn that the noise measure is itself the position distribution $\tau^Q$ of a quantum state $\tau$. The same holds for momentum, and, crucially, it is {\it the same} quantum state $\tau$ that enters. In other words, the tradeoff between the noises in the marginals of a phase space observable is the same as the tradeoff between the concentration of the position distribution $\tau^Q$ and the momentum distribution $\tau^P$ of a quantum state. This tradeoff is known as {\bf preparation uncertainty}. The equality of measurement uncertainty and preparation uncertainty is false for most other observable pairs but persists \cite{PSuncert} for more general observable pairs, which are related by the Fourier transformation of some locally compact abelian group. This includes angle and number, or qubit strings looked at in different Pauli bases.
\subsection{Dynamics}\label{sec:dynamics} For time evolutions, the input and output systems are the same. Let us first consider {\bf reversible} evolutions, for which the time parameter $t$ in ${\mathcal T}_t$ is allowed to be positive or negative, i.e., the ${\mathcal T}_t$ form a one-parameter group rather than just a semigroup. Then $\Delta\sigma$ has to vanish, and each ${\mathcal T}_t$ must be a noiseless operation (cf.\ Sect.~\ref{sec:noiseless}), and ${\mathcal T}_t^*$ must be a homomorphism. Actually, this conclusion is valid even without the quasifree form, just using that equality in the Schwarz inequality for completely positive maps (${\mathcal T}(x^*x)\geq{\mathcal T}(x)^*{\mathcal T}(x)$) implies the homomorphism property. Hence for a reversible evolution, the center of the algebra, i.e., the classical part, must be invariant as a set, and there is a well-defined restriction of ${\mathcal T}_t$ to the classical subsystem. That is, by observing the classical subsystem, we can never find out anything about the initial state of the quantum subsystem. This {\bf no-interaction theorem} blocks any understanding of the quantum measurement process by reversible, e.g., Hamiltonian couplings. It is quite expected on general grounds: Any information gained about a quantum system requires a disturbance, and this is not compatible with reversibility.
This No-Go theorem is lifted as soon as we allow irreversible evolutions. Indeed, one can develop a joint generalization of the theory of diffusions on the classical side and Lindblad Master equations on the quantum side, in which the salient information-disturbance tradeoffs have a natural and rigorous formulation.
A traditional subject in classical probability are processes with {\bf independent increments}. Since the increments are supposed to have the same distribution for any current state, this implies translation invariance, and since successive increments are assumed independent, we get a convolution semigroup ($S_t\equiv\idty$). The classic result is the L\'evy--Khintchine Theorem (see, e.g., \cite{AppleLevy}), characterizing the generators as a combination of a Gaussian part and a jump part. If we likewise stick to the choice of trivial $S_t$, this result applies verbatim to arbitrary hybrids. Even without quasifreeness assumption, it is treated in \cite{barchielli_1996}.
For the general case of an arbitrary semigroup $S_t$, the precise and general characterization of generators is lacking so far. It is easy to see that the L\'evy--Khintchine formula is still valid, but there are uncertainty-type constraints needed to ensure complete positivity. These are readily solved in the purely Gaussian case: The logarithmic derivative of the noise function at $t=0$ has to be an admissible quantum covariance matrix for the ``symplectic form'' computed as the derivative of $\Delta\sigma$. In general, the same constraint hold for the combination of the Gaussian part and the second order moments of the jump measure. Moreover, one is always free to add classical L\'evy--Khintchine noise. Many interesting features of the classical theory are connected to jump measures that do not have finite moments. Covering the quantum counterparts and getting the relevant Lindblad (or more general \cite{Inken,ArvesonBook}) form for the resulting semigroup generators is an interesting project.
Many applications use the quasifree structure. Especially when time-dependent generators are involved, as in the case of {\bf feedback and control}, it is vastly easier to put the process together in phase space than to multiply cp maps on the infinite-dimensional observable algebra. {\bf Continual observation} is likewise a hybrid scenario, in which the classical part can be observed completely and at all times without incurring disturbance costs. Doing justice to this field would require a book of its own, and we do not even try to review the literature. The hybrid aspects are typically neglected, as are the demands of building usable observable algebras.
\subsection{Classical limit}
The classical limit, $\hbar\to0$, characterizes the behavior of states and observables which do not change appreciably over phase space regions whose size is measured by $\hbar$. We have suppressed this parameter, which implicitly means that we used units for quantum position and quantum momentum, which make $\hbar=1$. For the discussion of the classical limit, it is better to make this parameter explicit as a factor to the commutation form \eqref{CCR}, just as physics textbooks have it. The identity map $S$ between universes with different $\hbar$ is then not symplectic, but one can build a (necessarily noisy) quasifree channel between such universes, allowing the comparison of observables. Equivalently, one can scale all phase space variables by $\sqrt\hbar$. The connection maps are then used to formulate a notion of {\it convergent sequences} by a Cauchy-like condition. This approach to the classical limit \cite{hbar20} is as close to a limit of the entire theory (not just isolated aspects such as WKB wave functions or partition functions) as one can get. The limit is a classical canonical system, with quantum Hamiltonian dynamics going to its classical counterpart. For our context, it should be noted that it can be taken for parts of the system (like the heavy particles in a Born--Oppenheimer approximation) and, due to the complete positivity of the connection maps, composes well with further degrees of freedom, i.e., can be applied to hybrids.
\subsection{Cloning}
Cloning, also known as copying or broadcasting, is a process that generates copies of a quantum system \cite{LindbladClone}. Of course, the well-known No-cloning Theorem says that this cannot be done without error. Quasifree maps are ideally suited as a simple testbed for this basic operation and the unavoidable errors. Let us consider a fixed system type $\Xis{}$, which also serves as the input. At the output, we have $N$ such systems in parallel, so $\Xi_{\mathrm{out}}=\bigoplus_j^N\Xi_j$ where $\Xi_j$ is just an isomorphic copy of the underlying $\Xi=\Xi_{\mathrm{in}}$. The marginals of interest forget all but one output and are thus described by a disturbance channel with $S=\idty$ (see above). This fixes $S$ on each of the subspaces in $\Xi_{\mathrm{out}}$, and hence by linearity, the overall map $S$:
\begin{equation}\label{cloneS}
S\Bigl(\bigoplus_j\xi_j\Bigr)=\sum_j\xi_j.
\end{equation}
In other words, this map is exactly what one would write down for an ideal copier if one had never heard of the No-cloning Theorem. The quasifree formalism then generates all possible error tradeoffs consistent with this overall behavior.
The optimal solution of this problem depends on how the quality of the clones is assessed, and in particular, whether one uses the average fidelity of the clones or the closeness of the overall output to a product state, i.e., whether one also demands the output systems to be nearly uncorrelated. The optimization problem should be stated without assuming quasifreeness, but one can {\it prove} that the optimal cloners will be quasifree with the above $S$. It turns out that for the criterion of overall product state fidelity, the optimal cloner is Gaussian, whereas for the average single state fidelity criterion, it is not, although the best Gaussian cloner performs only a few percent below optimum~\cite{Gcloners}. One can also look at asymmetric scenarios, in which the various copies satisfy different quality requirements, i.e., the output state is not permutation symmetric.
\subsection{Instruments}\label{sec:instru}
An instrument, according to a now-standard terminology by Davies and Lewis \cite{DavLew,Davies} is a channel with both a classical and a quantum output, i.e., a hybrid output. This is the setting in which one can discuss the tradeoff between information gain on the classical part of the output and disturbance on the quantum output (see Fig.~\ref{fig:example_covinst}).
\begin{figure}[ht]
\begin{center}
\begin{tikzpicture}[scale=1]
\coordinate (A) at (-1.5,1);
\coordinate (B) at (1.5,1);
\coordinate (C) at (1.5,-1);
\coordinate (D) at (-1.5,-1);
\draw[] (A) --(B) --(C) --(D) --(A);
\draw[thick,->] (-2.5,0) -- (-1.5,0);
\draw[thick,->] (1.5,0.5) -- (2.5,0.5);
\draw[double,->] (1.5,-0.5) -- (2.5,-0.5);
\node [left] at (-2.5,0) {$ \Xis_{\mathrm{in}}\cong\Xis{}$};
\node [right] at (2.5,0.5) {$\Xis{}$};
\node [right] at (2.75,0) {$\oplus$};
\node [right] at (2.5,-0.5) {$(\Xi_c,0)$};
\node [right] at (3.75,0) {$\cong \Xis_{\mathrm{out}}$};
\draw[decoration={brace,raise=5pt},decorate] (3.5,0.75) --(3.5,-0.75);
\node [] at (0,0) {${\mathcal T}$};
\end{tikzpicture}
\captionsetup{width=0.8\textwidth}
\caption{A covariant instrument: A quantum system with the phase space $\Xis{}$ is measured by the instrument ${\mathcal T}$. The output is a hybrid system with a quantum part on the same space $\Xis{}$ joined by a classical system, the measurement result, with some classical system $(\Xi_c,0)$. }
\label{fig:example_covinst}
\end{center}
\end{figure}
Concretely, let $\Xi_{\mathrm{out}}=\Xi_{\mathrm{in}}\oplus\Xi_c$, where $\Xi_c$ is the classical output. As in the case of a cloner, linearity of $S$ implies that we just have to fix our demands for the marginals, i.e., the actions on the summands $\Xi_{\mathrm{in}}$ and $\Xi_c$, to get the overall map $S$. On the first summands, we just take the identity, in keeping with our intention to discuss the disturbance inflicted by the instrument. The case of ``no disturbance'' should be included, so we should take $S=\idty$ on the summand $\Xi_{\mathrm{in}}$. For the second summand, $\Xi_c$, we just have to say which variable or combination of variables we wish to measure, i.e., $S$ is chosen exactly as the corresponding map $S$ from the above description of observables. To distinguish it from the overall $S$, we denote this by $S_c$.
Putting these parts together, we get
\begin{equation}\label{Sinstrument}
S(\xi\oplus\eta)=\xi+S_c\eta
\end{equation}
or, equivalently, $S^\top\xi=\xi\oplus S_c^\top\xi$. The noise functions consistent with this choice then parametrize the class of covariant phase space instruments.
Their analysis is a nice illustration of our theory. The main interest is again in the marginals, which reflect the tradeoffs between disturbance and information gain. We treat them in analogy to the corresponding observables.
Just as for observables, the theory of quasifree instruments fits into the theory of covariant instruments for more general groups \cite{DaviesCovInst,HolevoRadNy,Carmeli,Erkka}. We begin by outlining a heuristic argument suggesting a form for general covariant instruments. We will verify later how this form comes out of our approach. As in the case of observables, we assume an operator density for the outputs as a function of the measured parameter: Its interpretation is the quantum channel conditioned on the classical output $x$. This captures a typical use of instruments, where the quantum state is updated based on the classical result. We are thus looking for a family of cp maps ${\mathcal T}_x$ such that the following analog of \eqref{RNposition} holds:
\begin{equation}\label{instCond1}
{\mathcal T}^*(A\otimes g)=\int dx\ {\mathcal T}^*_x(A) g(x).
\end{equation}
Putting $A=\idty$, it is clear that ${\mathcal T}_x$ is not a channel, as it is not normalized to the identity. Instead ${\mathcal T}_x(\idty)=\dot F(x)$ is the Radon--Nikodym\ density of the classical marginal observable. Thus, if the classical marginal has no density, then ${\mathcal T}_x$ cannot be defined either. On the other hand, if $\dot F(x)$ exists, we can look for a bona fide channel $\widetilde{\mathcal T}_x$ such that, with the abbreviation $D(x)=\dot F(x)^{1/2}$, we have
$D(x)\widetilde{\mathcal T}_x(A)D(x)={\mathcal T}_x(A)$. A feature shared with the observable case and the general group case is that ${\mathcal T}_x(A)$ needs only be known at one point because this can be transferred to all $x$ by covariance. Indeed, the covariance of the instrument is equivalent to ${\mathcal T}^*_{x+S_c^\top\xi}=\alpha_{-\xi}{\mathcal T}^*_{x} \alpha_{\xi}$. Thus if $K_j$ is a set of Kraus operators for the channel $\widetilde{\mathcal T}_0$, and $\dot F=D^2$ is the density at the origin, we can write
\begin{equation}\label{instCondi}
{\mathcal T}^*_x(A)= \sum_j \alpha_{-\xi}\bigl(K_jD\bigr)^*A\ \alpha_{-\xi}\bigl(K_jD\bigr),
\quad \text{where}\ S^\top\xi=x.
\end{equation}
In other words, now the Kraus operators are subject to the group translation. They are only constrained by the overall normalization $\sum_jK_j^*K_j=\idty$ and the invariance condition arising from the possibility that $S_c^\top\xi=0$ might have non-zero solutions $\xi$. In that case, we must demand that the $K_j$ and the $\alpha_\xi(K_j)$ describe the same channel. In particular, for extremal instruments, when there is only one Kraus operator, it has to be invariant up to a phase.
\subsubsection*{Position instruments}
We take $S_c$ from the position observable as above, i.e., $S(p,q,k)=(p+k,q)$. All these quantities can be vectors $p,q,k\in{\mathbb R}^n$. Then
\begin{equation}\label{posIDelta}
(p,q,k)\cdot\Delta\sigma(p',q',k')=p{\cdot} q'-q{\cdot} p' -(p+k){\cdot} q'+q{\cdot}(p'+k')=q{\cdot} k'-k{\cdot} q'.
\end{equation}
We will illustrate our formalism by executing the task of finding all position instruments twice: Once directly via the characteristic functions and Prop.~\ref{prop:ChanOnW}, and once in the way inspired by general covariance theory, i.e., via \eqref{instCondi}. For simplicity, we look only at the pure case, i.e., we are happy to find the simplest solutions from which all others arise by mixture.
For the approach of this paper, note that \eqref{posIDelta} is the commutation form of a hybrid phase space with quantum coordinates $(q,k)$ and a classical direction $p$. A pure state on this hybrid fixes the classical part (cf. Lem.~\ref{lem:purestates}) to a point $a$, say, and is given on the quantum part by a vector $\psi$ on the Hilbert space of $n$ degrees of freedom, defining the noise state $\tau$. This gives the noise function
\begin{equation}\label{fPosInst}
f(p,q,k)=e^{i a{\cdot} p}\,\chi_\tau(q,k)\ = e^{i a{\cdot} p}\, \braket\psi{W(q,k)\psi}.
\end{equation}
Together with $S$, this is a complete description of the instrument.
In the approach via \eqref{instCondi}, purity brings us to the case of a single Kraus operator $K$, which is determined by the operation up to a phase. The kernel of $S_c^\top$ consist of the $\xi=(q,0)$, so $W(q,0)$ is a multiplication operator in the position representation. $K$ must be an eigenvector: $W(q,0)^*KW(q,0)=u(q)K$, for some phases $u(q)$. This must be a character, $u(q)=\exp(i a{\cdot} q)$. Hence $KW(0,a)$ is an operator commuting with all multiplication operators in position representation, hence is itself such an operator, i.e., $\psi(Q)$ for some function. To summarize,
\begin{equation}\label{KrausPos}
\bigl(KD\phi\bigr)(x)=\psi(x)\phi(x+a).
\end{equation}
The required assumptions on the function $\psi$ are not immediately obvious. One readily checks that multiplication
by $\abs{\psi(x-a)}^2$ is the density for the position observable. This must have integral $1$ with respect to $x$, so $\psi\in L^2$ seems like a natural condition, even if it makes the Radon--Nikodym\ density $\dot F(x)$ a potentially unbounded operator. In this approach, it is maybe not obvious that the instrument is well-defined for arbitrary unbounded functions $\psi\in L^2$. However, computing the characteristic function of the overall channel (maybe with some additional regularity assumptions) gives exactly \eqref{fPosInst} with the same $\psi$, $a$. So the two approaches give the same result, only with less analytical pain in our quasifree theory.
We are interested in the tradeoffs for the marginals, namely the quantum output, which is necessarily of the type discussed above under ``disturbance'', and the measurement output, which is of the type discussed under position observables. Both can be read off directly from the $\chi_{\mathrm{out}}(p,q,k)=f(p,q,k)\chi_{\mathrm{in}}(p+k,q)$, by setting suitable variables to zero:
\begin{equation}\label{margePosInsT}
\begin{array}{rll}
\mbox{classical marginal:}&p=q=0,& \qquad\mbox{noise measure}= \tau^Q, \\
\mbox{quantum marginal:}&k=0,& \qquad\mbox{noise measure}=\delta_a\times \tau^P.
\end{array}
\end{equation}
This is a very concise formulation of a well-known intuition: $\tau^Q$ is the distribution of the noise added to the measurement outcomes, i.e., the ``error'' of the measurement. $\tau^P$, on the other hand, is the disturbance of the momentum variable. So these are reciprocal in exactly the way known for quantum states. We remark that noise could also occur in the quantum position direction, here given by a deterministic shift $a$. Non-pure instruments will have the distribution for that as well, and $\tau$ in the above description generally depends on $a$, allowing all the complex correlations in a hybrid noise state.
\subsubsection*{Phase space instruments}
In this case, $S(\xi\oplus\eta)=\xi+\eta$, and $\Delta\sigma$ is non-degenerate, so the noise state is a quantum state of twice the number of degrees of freedom. In the pure case, it is given by a vector $\psi\in L^2({\mathbb R}^n\times{\mathbb R}^n,dx_1,dx_2)$. Such a vector can be identified with a Hilbert--Schmidt operator over the system Hilbert space ${\mathcal H}=L^2({\mathbb R}^n,dx)$, and we will see that this is precisely the required form of the local Kraus operator $KD$. This general form for phase space instruments was also obtained independently in \cite{Erkka}. In the following proposition, which is a straightforward application of our formalism, we also describe the resulting tradeoff between disturbance (noise in the quantum marginal) and precision (noise in the classical marginal). They are precisely related by Fourier transformation almost exactly as in the case of joint measurements of position and momentum. Only the Fourier transform is not between position and momentum but between the operator side and the function side of quantum harmonic analysis.
\begin{prop}\label{prop:phspInstr}
\begin{itemize}
\item[(1)] Every extremal quasifree phase space instrument is characterized by a Hilbert--Schmidt operator $\hat\Psi$ with $\operatorname{tr}(\hat\Psi^*\hat\Psi)=1$ such that
\begin{equation}\label{phspInstr}
{\mathcal T}^*(A\otimes g)=\int d\xi\ \alpha_{-\xi}(\hat\Psi)^*\, A\, \alpha_{-\xi}(\hat\Psi)\ g(\xi).
\end{equation}
\item[(2)] Conversely, any such operator $\hat\Psi$ determines an instrument and is determined by it up to a phase.
\item[(3)] The classical marginal is a covariant phase space observable with density $\dot F=\hat\Psi^*\hat\Psi$.
\item[(4)] The quantum marginal is addition of translation noise: $\rho\mapsto\int d\xi\ m(\xi)\alpha_\xi(\rho)$ with $m\in L^1(\Xi)$
\begin{equation}\label{phspIMarge}
m(\xi)=\abs{({\mathcal F}\hat\Psi)(-\sigma\xi)}^2.
\end{equation}
\end{itemize}
\end{prop}
Note that since ${\mathcal F}$ is unitary from the Hilbert--Schmidt class onto $L^2(\Xi)$, not only all operator densities $\dot F$ but also all $L^1$-densities $m$ can occur. The prototype of this tradeoff is the case of a single degree of freedom with additional covariance under harmonic oscillator rotations. In particular, we can look at the Gaussians $\hat\Psi=c\,\exp(-\beta H)$ with $H=(P^2+Q^2)/2$. Then the Fourier transform is also Gaussian, and proportional to $\exp(-\coth(\beta/2))\xi^2/4$, where $\xi^2=(p^2+q^2)/2$. Now for $\beta\to0$, $\hat\Psi$ is a small multiple of the identity, so it can approximately be interchanged with $A$ in \eqref{phspInstr}. This even works in trace norm for the action on a trace class operator for the dual channel. This means that the disturbance goes to zero, and this is borne out by the computation of $m$, which for small $\beta$ is Gaussian with variance $\propto1/\beta$. On the other hand, the phase space density of the classical marginal becomes very broad, and the measurement outputs reveal very little about the state. In the other direction, $\beta\to\infty$, $\hat\Psi$ becomes a coherent state projection, and the output distribution becomes the Husimi function. The quantum noise $m$ is still Gaussian, with a variance on the order of standard quantum uncertainties.
\begin{proof}[Proof of Prop.~\ref{prop:phspInstr}]
The difference symplectic form is now
\begin{equation}\label{phspDelta}
(\xi,\eta)\cdot\Delta\sigma(\xi',\eta')=\xi\cdot\sigma\xi'-(\xi+\eta)\cdot\sigma(\xi'+\eta').
\end{equation}
Rather than expanding this, we just choose a twisted definite function, evaluated for the independent variables $\xi$ and $\xi+\eta$. That is, for the extremal case, we choose a pure state on a doubled system, given by a vector $\Psi\in{\mathcal H}\otimes{\mathcal H}$ such that
\begin{equation}\label{phspDeltaPsi}
f(\xi\oplus\eta)=\brAAket\Psi{W(\xi)\otimes \overline{W(\xi+\eta)}}\Psi.
\end{equation}
Here the bar indicates complex conjugation $ \overline{W(\xi)}=\theta^* W(\xi)\theta$ with respect to an arbitrary antilinear involution $\theta$, which has the effect of reversing the symplectic form and hence takes care of the minus sign in \eqref{phspDelta}. This completes the parametrization of the family of instruments. What is left is rewriting this in the stated form and computing the marginals.
To this end, we introduce the isomorphism $\Psi\mapsto\hat\Psi$ form ${\mathcal H}\otimes{\mathcal H}$ to Hilbert--Schmidt operators on ${\mathcal H}$ given by $\psi_1\otimes\psi_2\mapsto\ketbra{\psi_1}{\theta\psi_2}$. Note that the involution $\theta$ is needed here so that both sides of the identification are linear in $\psi_2$. We next express the action of the Weyl operators in \eqref{phspDeltaPsi} in terms of the Hilbert--Schmidt operators. For $\Psi=\psi_1\otimes\psi_2$, we get
\begin{align}
{W(\xi)\otimes \overline{W(\xi+\eta)}}\Psi
&=(W(\xi)\psi_1) \otimes (\theta^*W(\xi+\eta)\theta\psi_2)\nonumber\\
&\mapsto \ketbra{W(\xi)\psi_1}{W(\xi+\eta)\theta\psi_2}
= W(\xi)\,\hat\Psi\ W(\xi+\eta)^*.
\end{align}
Inserting this into \eqref{phspDeltaPsi} gives the equivalent expression
\begin{equation}\label{phspDeltaPsihat}
f(\xi\oplus\eta)=\operatorname{tr}\bigl(\hat\Psi^* W(\xi)\hat\Psi W(\xi+\eta)^*\bigr).
\end{equation}
Denoting the Weyl elements on the classical output by $W_0$, and using the identity $\int d\zeta\ \alpha_\zeta(A)=\operatorname{tr}(A)\idty$, we find
\begin{align}\label{phspDeltaPs}
{\mathcal T}^*\bigl(W(\xi)\otimes W_0(\eta)\bigr)
&=\operatorname{tr}\bigl(\hat\Psi^* W(\xi)\hat\Psi W(\xi+\eta)^*\bigr)\ W(\xi+\eta) \nonumber\\
&=\int d\zeta\ \alpha_\zeta\bigl(\hat\Psi^* W(\xi)\hat\Psi W(\xi+\eta)^*\bigr)\ W(\xi+\eta)\nonumber\\
&=\int d\zeta\ \alpha_\zeta\bigl(\hat\Psi\bigr)^*\ e^{i\zeta\cdot\xi}W(\xi)\ \alpha_\zeta\bigl(\hat\Psi\bigr)
\ e^{-i\zeta\cdot(\xi+\eta)}\ W(\xi+\eta)^*\ W(\xi+\eta)\nonumber\\
&=\int d\zeta\ \alpha_\zeta\bigl(\hat\Psi\bigr)^*\ W(\xi)\ \alpha_\zeta\bigl(\hat\Psi\bigr)
\ e^{-i\zeta\cdot\eta}\nonumber\\
&=\int d\zeta\ \alpha_{-\zeta}\bigl(\hat\Psi\bigr)^*\ W(\xi)\ \alpha_{-\zeta}\bigl(\hat\Psi\bigr)
\ W_0(\eta)(\zeta).
\end{align}
This coincides with \eqref{instCond1} and \eqref{instCondi} with $g=W_0(\eta)$, $A=W(\xi)$ and $KD=\hat\Psi$.
The form of the classical marginal is obvious from \eqref{phspInstr} by putting $A=\idty$ (resp.~$\xi=0$ in \eqref{phspDeltaPs}). For the classical marginal, putting $g=1$ leads to a form from which it is not even clear that it is just convolution with noise. For that, it is better to go back to the characteristic functions. Indeed, the function $m$ in \eqref{phspIMarge} is just the inverse Fourier transform of $f(\xi\oplus0)$, i.e.,
\begin{align}\label{phspDeltaPs2}
m(\eta)&=(2\pi)^{-2n}\int d\xi\ e^{i\eta\cdot\xi}\ f(\xi\oplus0)=(2\pi)^{-2n}\int d\xi\ e^{i\eta\cdot\xi}\ \operatorname{tr}\bigl(\hat\Psi^* W(\xi)\hat\Psi W(\xi)^*\bigr) \nonumber\\
&=(2\pi)^{-2n}\int d\xi\ e^{i\eta\cdot\xi}\ \operatorname{tr}\bigl(\hat\Psi^* \alpha_{\sigma\xi}(\hat\Psi)\bigr)
=(2\pi)^{-2n}\int d\xi\ \operatorname{tr}\bigl(\hat\Psi^* \alpha_{\sigma\xi}(\hat\Psi W(-\sigma\eta))W(\sigma\eta)\bigr)\nonumber\\
&= \operatorname{tr}\bigl(\hat\Psi W(-\sigma\eta)\bigr)\operatorname{tr}\bigl(\hat\Psi^*W(\sigma\eta)\bigr) =\bigl|({\mathcal F}\hat\Psi)(-\sigma\eta)\bigr|^2.
\end{align}
In the second line, we used the eigenvalue equation \eqref{Weylshifted} to absorb the exponential factor and \eqref{intL1} in the last line to evaluate the integral.
\end{proof}
\subsection{Teleportation and dense coding}\label{sec:teleport}
The quasifree setting also provides a special angle on the well-known protocols of dense coding, and teleportation \cite{Bennetele,Bennedense,Wertele}.
This is traditionally treated in finite-dimensional settings. Our setting can largely be generalized to cover finite dimensions and, in fact, arbitrary phase spaces built as the Cartesian product of a locally compact abelian group for position and its dual group for momentum. With a finite group the Hilbert spaces become finite-dimensional, and in the simplest case, this is the one-bit (=two-element) group with the Pauli matrices and identity as the Weyl operators. We will now take the qubit case as a guide and obtain a painless quasifree approach to ``continuous variable teleportation'', generalizing the usual Gaussian schemes \cite{CVteleport} to arbitrary non-Gaussian entangled resource states.
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=.8]
{\def\bx(#1,#2);{\draw[] (#1,#2) -- +(1.5,0) -- +(1.5,1) -- +(0,1) -- +(0,0);}
\node at (2.6,-1) {Teleportation}; \node at (12,-1) {Dense coding };
\bx(0,0); \bx(0,2.5); \bx(4,2.5);
\draw[thick,->] (.75,1) -- (.75,2.5); \draw[thick,->] (1.5,.5) -- (4.75,.5) -- (4.75,2.5);
\draw[thick,->] (1.5,3.05) -- (4,3.05); \draw[thick,->] (1.5,2.95) -- (4,2.95); \draw[thick,->] (-1,3)--(0,3); \draw[thick,->] (5.5,3)--(6.5,3);
\node at (0,1.75) {$(\Xi,-\sigma)$}; \node at (2.6,0.8) {$(\Xi,\sigma)$}; \node at (-.8,3.4) {$(\Xi,\sigma)$}; \node at (2.6,3.4) {$(\Xi,0)$}; \node at (6.2,3.4) {$(\Xi,\sigma)$};
\bx(9,0); \bx(9,2.5); \bx(13,2.5);
\draw[thick,->] (9.75,1) -- (9.75,2.5); \draw[thick,->] (10.5,.5) -- (13.75,.5) -- (13.75,2.5);
\draw[thick,->] (10.5,3.05) -- (13,3.05); \draw[thick,->] (8,2.95) -- (9,2.95); \draw[thick,->] (8,3.05)--(9,3.05);
\draw[thick,->] (14.5,3.05)--(15.5,3.05);\draw[thick,->] (14.5,2.95)--(15.5,2.95);
\node at (9.1,1.75) {$(\Xi,\sigma)$}; \node at (11.6,0.8) {$(\Xi,-\sigma)$}; \node at (8.2,3.4) {$(\Xi,0)$}; \node at (11.6,3.4) {$(\Xi,\sigma)$}; \node at (15.2,3.4) {$(\Xi,0)$};
}
\end{tikzpicture}
\captionsetup{width=0.8\textwidth}
\caption{The protocols for teleportation and dense coding. Classical information is indicated by a double arrow.
All operations in the top row are noiseless with the map $S\xi=\xi\oplus\xi$. The two protocols are related by swapping the equipment for sending and receiving sides. The noise arises from the entangled resource state and can be chosen to be zero in the finite cases. }
\label{fig:teleport}
\end{center}
\end{figure}
This will give some quasifree teleportation schemes, but not all have this property (cf. \cite{Wertele}). In any case, the quasifree angle suggests a natural interpretation of why the classical signals require 2 bit in the 1 qubit version: This is just the {\it phase space} associated to the qubit system. So we will take all systems involved as systems with the same phase space $\Xi$ but different symplectic forms. Sender (Alice) and receiver (Bob) have quasifree devices with the same $S$, namely $S\xi=\xi\oplus\xi$. Only the symplectic forms need to be chosen so that the devices can be chosen to be noiseless (see Fig.~\ref{fig:teleport}). The verification of the protocol is then trivial and identical for teleportation and dense coding:
The overall map $S$ takes $\xi\mapsto\xi\oplus\xi\mapsto\xi\oplus\xi\oplus\xi\mapsto\xi$, where the last step is evaluation with the entangled state provided, say with characteristic function $\chi(\xi\oplus\eta)$. Noise occurs only in this last step, so the overall noise is just a translation average with an $L^1$ probability density, whose characteristic function is $\chi(\xi\oplus\xi)$. By choice of symplectic forms, the diagonal subspace of $\Xi\oplus\Xi$ lies in the null space of the symplectic form, so it can be considered as ``position'' in a doubled system and therefore allows an arbitrarily sharply concentrated distribution. Actually, this argument has been used since 1935 \cite{EPR}, when it was realized that in a two-particle system $(Q_1-Q_2)$ and $(P_1+P_2)$ commute.
Note, however, that we do {\it not} get the ``original EPR state'', for which the distribution would be a $\delta$-function, and which would therefore make teleportation work ideally. This is not the position distribution of a standard quantum state (cf. Sect.~\ref{sec:states} and Prop.~\ref{prop:contransl}). Rather it is a singular state \cite{KSW03} which mostly lives at infinity (on some compactification of phase space).
\newpage
\section{Summary and Outlook}
We have developed a framework for canonical hybrid systems in which quasifree channels can be discussed with remarkable ease and full generality. In several ways, this theory is simpler than more specialized versions. This is an instance of the inventor's paradox (``The more general problem may have the simpler solution''). For example, if one is not interested in measurement, and classical inputs and outputs, one could have expected a simpler theory by dropping all the classical variables and restricting to purely quantum systems. However, channels in that context would still satisfy a positivity condition belonging to a hybrid state, and the noise factorization (Thm.~\ref{thm:noisedec}) would provide an analysis of the noise in the channel as partly classical and partly quantum. This would suggest allowing hybrids from the outset, and indeed we saw that this does not make the theory any harder.
A second case of the inventor's paradox in this paper is the lack of a Gaussian assumption. Gaussian quasifree channels are those for which $f$ has a Gaussian form and is hence given by a covariance matrix. We actually started out by looking at Heisenberg picture questions for this class, e.g., ``Is phase space continuity (as in ${\mathcal C}_{\mathrm u}\Xis{}$) automatically preserved by Gaussian channels?''. It turned out that the Gaussian simplification did not help at all for this, and more and more such issues were resolved in the general quasifree setting of the current paper.
Another simplification lies in the $\mu$-free approach, whose distinction from a $\mu$-dependent one is sketched at the beginning of Sect.~\ref{sec:funcobs}. The gain is to include pure states and, taken together with the previous paragraph, extremal channels. Here we had to go to considerable functional analytic lengths, but the result is simple and easy to apply: A variety of choices for hybrid observable algebras that can be used systematically with automatic Heisenberg picture description for the full class of quasifree channels.
Several directions for further work present themselves. Some have already been mentioned above:
\begin{itemize}
\item Specialize to the {\it Gaussian} case, i.e., the case where all noise functions have Gaussian form. This class is well known \cite{Holevo11,MCF}, and practically important \cite{Lammers}, and allows a complete reduction to the finite-dimensional analysis of covariance matrices together with the $S$-operators between phase spaces. One gets a simple toolbox in which the tradeoffs of information gain and disturbance can all be described in finite-dimensional matrix terms.
\item Generalize to hybrids with general, i.e., {\it not quasifree} channels. The key element in Sect.~\ref{sec:funcobs} is the local compactness of the classical parameter space, but to get good channels, we also used the continuity of characteristic functions, and hence the group structure of phase space. Can one get a good class of channels without that?
\item Replace the phase space by an arbitrary locally compact abelian group, and the Weyl operators by a {\it projective representation}. A lot of the theory described here will carry over, but it is a matter of careful screening to identify the limits of this generalization.
\item Consider the {\it Fermionic} and mixed CAR/CCR case.
\item Allow {\it infinite dimensional} $\Xi$. The aim would be applications in quantum field theory. So far, mostly the case of symplectic maps as been considered under the heading of Bogolyubov transformations. However, in order to bring some operational elements to the theory, noisy operations like counting processes and other interventions are very interesting, and the quasifree category is an ideal testing ground. These aspects are sorely underdeveloped in all schools of QFT, but a better understanding seems to be emerging \cite{VerchFewster,Jubb}.
\item Analyze {\it dynamical semigroups}.
This was described in more detail in Sect.~\ref{sec:dynamics}. A missing piece seems to be a precise description of the constraints on the data in the L\'evy--Khintchine formula to make a completely positive semigroup with non-symplectic $S_t$.
\item The intersection of the previous two items gives quasifree hybrid semigroups on infinite dimensional spaces \cite{DaviesDiff,hellmich,Blanchard}. Thorough work exists in the case of classical noise, e.g., when a unitary group is controlled by a driving Markov process \cite{Blanchard}. One interesting issue is the possibility and structure of quantum dynamical semigroup generators, which are {\it not} of {\it Lindblad} (or Arveson type I \cite{ArvesonBook}) form \cite{Inken}.
\item One of the beautiful results in Gaussian Quantum Information is the growing evidence \cite{VittHolGauss,newHolGauss} that the variational problems in the capacity theory of Gaussian channels have {\it Gaussian maximizers}. This involves the discussion of relative entropies and Gibbs states for quadratic Hamiltonians, which surely have hybrid versions, possibly even with some relevance to the Gaussian maximizer conjecture.
\item Explore the {\it resolvent algebra}, and potential applications of Example~\ref{Ex:resAlg} to quantum field theory. In particular, analyze the sum decomposition \eqref{resalg}, and the ideals of ${\mathcal R}\Xis{}$ in the light of correspondence theory, and study the continuity of dynamical evolutions.
\item Further explore the understanding of the {\it Paschke dilation} \cite{Westerbaan,Paschke} as the basic dilation statement in the category of W*-algebras with completely positive normal maps. This reduces to the Stinespring dilation when the input system is a quantum system with observable algebra ${\mathcal B}({\mathcal H})$. A good start has been made in \cite{Westerbaan}, but many issues that have been treated traditionally by the Stinespring construction should allow a treatment in this more general setting.
\end{itemize}
\section*{Acknowledgements}
We thank Alexander Stottmeister and Lauritz van Luijk for critical reading and prompting several clarifications and improvements.
Support by the Graduiertenkolleg 1991 of the DFG, the CRC DQ-mat, and the Network QlinkX of the BMBF is gratefully acknowledged.
\newpage
\bibliographystyle{abbrv}
|
\section{Introduction}\label{sec: Introduction}
Einstein’s field equations in general relativity predict that the curvature is produced not only by the distribution of mass-energy but also by its motion \cite{Hartle2002}. Candidate metric geometries that can intrinsically describe the motion are the Finsler and Finsler-like geometries which constitute metrical generalizations of Riemannian geometry and depend on position and velocity/momentum/scalar coordinates. These are dynamic geometries that can describe locally anisotropic phenomena and Lorentz violations
\cite{Asanov:1991rmp, Stavrinos:1999rmp, Kostelecky:2011qz, AlanKostelecky:2012yjr, Stavrinos:2012kv, Minguzzi:2015xka, Foster:2015yta, Antonelli:2018fbv, Edwards:2018lsn, Ikeda:2019ckp, Relancio:2020mpa}
as well as with field equations, FRW and Raychaudhuri equations, geodesics, dark matter and dark energy effects \cite{Konitopoulos:2021eav, Stavrinos:2021ygh, Kouretsis:2008ha, Mavromatos:2010jt, Stavrinos:2020llm, Hama:2022vob}.
By considering this approach, the gravitational field is interpreted as the metric of a generalized spacetime and constitutes a force-field which contains the motion. This possibility reveals the Finslerian geometrical character of spacetime.
In the framework of applications of Finsler geometry, many works in different directions of geometrical and physical structures have contributed to the extension of research for theoretical and observational approaches during the last years. We cite some works from the literature of the applications of Finsler geometry
\cite{Gibbons:2007iu, Skakala:2010hw, Kostelecky:2011qz, AlanKostelecky:2012yjr, GallegoTorrome:2012qra, Stavrinos:2013neo, Fuster:2015tua, Voicu:2015uta, Hohmann:2018rpp, Edwards:2018lsn, Colladay:2019lig, Caponio:2020ofw, Hama:2021frk, Hama:2022vob}.
In the first period of development of applications of Finsler geometry to Physics, especially to General Relativity, remarkable works were
published by G.~Randers~\cite{Randers1941}, J.~I.~Horvath~\cite{Horvath1950} and A.~Moor~\cite{HorvathMoor}.
Later, Einstein’s field equations were formulated in the Finslerian framework by the works of J.~I.~Horvath~\cite{Horvath1950,HorvathMoor},
Y.~Takano~\cite{Takano1968} and S.~Ikeda~\cite{Ikeda1981}. In these studies, the field equations had been considered without calculus of variations.
G.~S.~Asanov~\cite{Asanov1983} explored the Finslerian gravitational field by using Riemannian osculating methods and derived Einstein field equations using the variational principle.
A class of Finsler spaces (FR standing
for Finsler-Randers) originated by
G.~Randers~\cite{Randers1941} who studied the physical properties of spacetime with an asymmetrical metric which provides the uni-direction of time-like intervals. This consideration gives a particular interest in a generalized metric structure of the Riemannian spacetime.
Based on this form of spacetime, it is possible to investigate the gravitational field with more degrees of freedom in the framework of a tangent/vector/scalar bundle \cite{Triantafyllopoulos:2018bli, Triantafyllopoulos:2020ogl, Hama:2021frk}.
The FR cosmological model was first introduced in \cite{stavrinos2005, Stavrinos:2006rf}. It is of special interest since the Friedmann equations include an extra geometrical term that acts as a dark energy-fluid.
The Finsler-Randers-type spacetime can be considered as a direction-dependent motion of the Riemannian/FRW model.\\
The local anisotropic structure of spacetime affects the gravitational field and leads to modified cosmological considerations. Based on Finsler or Finsler-like cosmologies, the Friedmann equations include extra terms which influence the cosmological evolution \cite{Kouretsis:2008ha, Triantafyllopoulos:2018bli,Hama:2021frk, Konitopoulos:2021eav}.
When Lorentz symmetry holds, the spacetime is isotropic in the sense that all directions and uniform motions are equivalent.
The introduction of a vector field in the structure of spacetime causes relativity violations and local anisotropy which arise from breaking the Lorentz symmetry and which affect the metric, curvature, geodesics and null cone \cite{Girelli:2006fw, Minguzzi:2014aua, Javaloyes:2018lex, Silva:2013xba, Kostelecky:2008be, Vacaru:2010fi, Stavrinos:2012ty, Hohmann:2016pyt}.
An FR space has a metric function of the form
\begin{equation}\label{lagrangian}
F(x,y) = (-a_{\mu\nu}(x)y^{\mu}y^{\nu})^{1/2} + u_{\alpha}y^{\alpha}
\end{equation}
where $u_{\alpha}$ is a covector with $||u_{\alpha}||\ll 1$, $y^{\alpha}=\frac{dx^{\alpha}}{d\tau}$ and $a_{\mu\nu}(x)$ is a Riemannian metric for which the Lorentzian signature $(-,+,+,+)$ has been assumed and the indices $\mu, \nu, \alpha$ take the values $0,1,2,3$. The geodesics of this space can be produced by \eqref{lagrangian} and the Euler-Lagrange equations. If $u_{\alpha}$ denotes a force field $f_{\alpha}$ and $y^{\alpha}$ is substituted with $d x^{\alpha}$ then $f_{\alpha}dx^{\alpha}$ represents the spacetime effective energy produced by the anisotropic force field $f_{\alpha}$, therefore equation \eqref{lagrangian} is written as
\begin{equation}\label{lagrangian2}
F(x,dx) = \left(-a_{\mu\nu}(x)dx^{\mu}dx^{\nu}\right)^{1/2} + f_{\alpha}dx^{\alpha}
\end{equation}
This form of metric provides a dynamical effective structure of spacetime. A small differentiation is presented between GR and the FR gravitation model. This is because of the work provided by the one-form $A_{\gamma}$ which gives an external motion to the Riemannian spacetime. This motion is an internal concept for the FR spacetime.\\
A cosmological model can be considered by
Eq.~(\ref{lagrangian2}) if we assume the FRW cosmological metric instead of the general type of the Riemannian one \cite{stavrinos2005, Stavrinos:2006rf}. In this case, we get a Friedmann-Finsler-Randers cosmological model in the following form
\begin{equation}
a_{\mu\nu}(x) = \mathrm{diag}\left[-1,\frac{a^2}{1-\kappa r^{2}},a^{2}r^{2}, a^{2}r^{2}\sin^{2}\theta\right]
\end{equation}
This model was also further studied later in \cite{Stavrinos:2002rc, Chang:2007vq, Stavrinos:2012kv, Basilakos:2013hua, Basilakos:2013ij, Brody:2015zra, Silva:2015ptj, Stavrinos:2016xyg, Papagiannopoulos:2017whb, Chaubey:2018wph, Chanda:2019guf, Chanda:2019mro, Heefer:2020hra, Raushan:2020mkh, Papagiannopoulos:2020mmm, Silva:2020tqr, Lou:2021gwk, Hama:2022vob, Angit:2022lfu}.\\[7pt]
In the present paper, we continue the investigation of the Schwarzschild-Finsler-Randers spacetime (SFR) which has been studied in previous works
by a subset of the present authors \cite{Triantafyllopoulos:2020vkx, Kapsabelis:2021dpb}. The structure of the model is given in Section \ref{sec: Basic structure}. In this framework, the geodesics are studied and a dynamical analysis is presented in Section \ref{sec: Geodesics}. We also compare our results with GR and discuss the corresponding similarities
and differences.
A dynamical analysis for the effective potential of this spacetime is provided in the Section \ref{sec: Effective Potential},
where upon suitable assumptions, the phase portraits
of both models (SFR and GR) are presented.
The deflection angle of the SFR spacetime is investigated in Section \ref{sec: Deflection}. Finally, the conclusions of our study
and some possible directions for future exploration
are presented in Section \ref{sec: Conclusions}.
\section{Basic structure of the model}\label{sec: Basic structure}
In this section, we briefly present the underlying geometry of the SFR gravitational model, as well as the field equations for the SFR metric. The solution of these equations for this metric is presented at the end of the section. An extended study of this model can be found in \cite{Triantafyllopoulos:2020vkx,Triantafyllopoulos:2020ogl}.
The Lorentz tangent bundle $TM$ is a fibered 8-dimensional manifold with local coordinates $\{x^\mu,y^\alpha\}$ where the indices of the $x$ variables are $\kappa,\lambda,\mu,\nu,\ldots = 0,\ldots,3$ and the indices of the $y$ variables are $\alpha,\beta,\ldots,\theta = 4,\ldots,7$.
The tangent space at a point of $TM$ is spanned by the so-called adapted basis
$\{E_A\} = \,\{\delta_\mu,\dot\partial_\alpha\} $ with
\begin{equation}
\delta_\mu = \dfrac{\delta}{\delta x^\mu}= \pder{}{x^\mu} - N^\alpha_\mu(x,y)\pder{}{y^\alpha} \label{delta x}
\end{equation}
and
\begin{equation}
\dot \partial_\alpha = \pder{}{y^\alpha}
\end{equation}
where $N^\alpha_\mu$ are the components of a nonlinear connection $N=N^{\alpha}_{\mu}(x,y)\,\mathrm{d} x^{\mu}\otimes \pdot{\alpha}$.
The nonlinear connection induces a split of the total space $TTM$ into a horizontal distribution $T_HTM$ and a vertical distribution $T_VTM$. The above-mentioned split is expressed with the Whitney sum:
\begin{equation}
TTM = T_HTM \oplus T_VTM
\end{equation}
The anholonomy coefficients of the nonlinear connection are defined as
\begin{equation}\label{Omega}
\Omega^\alpha_{\nu\kappa} = \dder{N^\alpha_\nu}{x^\kappa} - \dder{N^\alpha_\kappa}{x^\nu}
\end{equation}
A Sasaki-type metric
\cite{Miron:1994nvt, Vacaru:2005ht}
$\mathcal{G}$ on $TM$ is:
\begin{equation}
\mathcal{G} = g_{\mu\nu}(x,y)\,\mathrm{d}x^\mu \otimes \mathrm{d}x^\nu + v_{\alpha\beta}(x,y)\,\delta y^\alpha \otimes \delta y^\beta \label{bundle metric}
\end{equation}
where we have defined the metrics $g_{\mu\nu}$ and $v_{\alpha\beta}$ to be pseudo-Finslerian.
A pseudo-Finslerian metric $ f_{\alpha\beta}(x,y) $ is defined as one that has a Lorentzian signature of $(-,+,+,+)$ and that also obeys the following form:
\begin{align}
f_{\alpha\beta}(x,y) = \pm\frac{1}{2}\pdder{F^2}{y^\alpha}{y^\beta} \label{Fg}
\end{align}
where the function $F$ satisfies the following conditions \cite{Miron:1994nvt}:
\begin{enumerate}
\item $F$ is continuous on $TM$ and smooth on $ \widetilde{TM}\equiv TM\setminus \{0\} $, i.e., the tangent bundle minus the null set $ \{(x,y)\in TM | F(x,y)=0\}$ . \label{finsler field of definition}
\item $ F $ is positively homogeneous of first degree on its second argument:
\begin{equation}
F(x^\mu,ky^\alpha) = kF(x^\mu,y^\alpha), \qquad k>0 \label{finsler homogeneity}
\end{equation}
\item The form
\begin{equation}
f_{\alpha\beta}(x,y) = \dfrac{1}{2}\pdder{F^2}{y^\alpha}{y^\beta} \label{finsler metric}
\end{equation}
defines a non-degenerate matrix: \label{finsler nondegeneracy}
\begin{equation}
\det\left[f_{\alpha\beta}\right] \neq 0 \label{finsler nondegenerate}
\end{equation}
\end{enumerate}
where the plus-minus sign in \eqref{Fg} is chosen so that the metric has the correct signature.
In this work, we will follow the model presented in \cite{Triantafyllopoulos:2020vkx}. The metric $g_{\mu\nu}$ is the classic Schwarzschild one:
\begin{align}\label{Schwarzchild}
g_{\mu\nu}\mathrm{d} x^\mu \mathrm{d} x^\nu
= -fdt^2 + \frac{dr^2}{f} + r^2 d\theta^2 + r^2 \sin^{2}\theta\, d\phi^2
\end{align}
with $f=1-\frac{R_s}{r}$ and $R_s=2GM$ the Schwarzschild radius (we assume units where the speed of light $c=1$).
Hereafter, we consider an $\alpha$-Randers type metric as the one in rel.\eqref{lagrangian} which is distinguished from the $\beta$-Randers type metric that is investigated in the Standard Model Extension (SME) \cite{Kostelecky:2011qz,AlanKostelecky:2012yjr,Silva:2013xba,Foster:2015yta}.
The metric $v_{\alpha\beta}$ is derived from a metric function $F_v$ of the $\alpha$-Randers type:
\begin{equation}\label{RandersL}
F_v = \sqrt{-g_{\alpha\beta}(x)y^\alpha y^\beta} + A_\gamma(x) y^\gamma
\end{equation}
where $g_{\alpha\beta}=g_{\mu\nu}\tilde\delta^{\mu}_{\alpha}\tilde\delta^{\nu}_{\beta}$ is the Schwarzschild metric from Eq.~\eqref{Schwarzchild} and $A_{\gamma}(x)$ is a covector which expresses a deviation from general relativity, with $|A_\gamma(x)|\ll 1$, i.e.,
we assume that that deviation is small. We choose a non-linear connection with the following form:
\begin{equation}\label{Nconnection}
N^\alpha_\mu = \frac{1}{2}y^\beta g^{\alpha\gamma}\partial_\mu g_{\beta\gamma}
\end{equation}
The metric tensor $v_{\alpha\beta}$ of \eqref{RandersL} is derived from \eqref{Fg} after omitting higher order terms $O(A^2)$:
\begin{equation}\label{vab}
v_{\alpha\beta}(x,y) = g_{\alpha\beta}(x) + h_{\alpha\beta}(x,y),
\end{equation}
where
\begin{align}\label{hab}
h_{\alpha\beta} = \frac{1}{\tilde{a}}(A_{\beta}g_{\alpha\gamma}y^\gamma + A_{\gamma}g_{\alpha\beta}y^\gamma + A_{\alpha}g_{\beta\gamma}y^\gamma) + \frac{1}{\tilde{a}^3}A_{\gamma}g_{\alpha\epsilon}g_{\beta\delta}y^\gamma y^\delta y^\epsilon
\end{align}
with $\tilde{a} = \sqrt{-g_{\alpha\beta}y^{\alpha}y^{\beta}}$.
The total metric defined in the previous steps is called the \textit{Schwarzschild-Finsler-Randers} (SFR) metric. As we can see, the term $h_{\alpha\beta}(x,y)$ can be considered as a perturbation of the Schwarzschild metric since $|A_\gamma(x)|\ll 1$.
The nonzero coefficients of a canonical and distinguished $d-$connection $\mathcal D$ on $TM$ are:
\begin{align}
L^\mu_{\nu\kappa} & = \frac{1}{2}g^{\mu\rho}\left(\delta_kg_{\rho\nu} + \delta_\nu g_{\rho\kappa} - \delta_\rho g_{\nu\kappa}\right) \label{metric d-connection 1} \\
L^\alpha_{\beta\kappa} & = \dot{\partial}_\beta N^\alpha_\kappa + \frac{1}{2}v^{\alpha\gamma}\left(\delta_\kappa v_{\beta\gamma} - v_{\delta\gamma}\,\dot{\partial}_\beta N^\delta_\kappa - v_{\beta\delta}\,\dot{\partial}_\gamma N^\delta_\kappa\right) \label{metric d-connection 2} \\
C^\mu_{\nu\gamma} & = \frac{1}{2}g^{\mu\rho}\dot{\partial}_\gamma g_{\rho\nu} \label{metric d-connection 3} \\
C^\alpha_{\beta\gamma} & = \frac{1}{2}v^{\alpha\delta}\left(\dot{\partial}_\gamma v_{\delta\beta} + \dot{\partial}_\beta v_{\delta\gamma} - \dot{\partial}_\delta v_{\beta\gamma}\right). \label{metric d-connection 4}
\end{align}
See Appendix \ref{sec:d-connection} for more details.
The field equations for our model have been derived in previous works and can be found in Appendix \ref{sec:field_eqs}. Solving the field equations \eqref{feq1}, \eqref{feq2} and \eqref{feq3} to first order in $A_\gamma(x)$ in vacuum ($T_{\mu\nu} = Y_{\alpha\beta} = \mathcal Z^\kappa_\alpha = 0$), we get \cite{Triantafyllopoulos:2020vkx}:
\begin{equation}\label{Asolution}
A_\gamma(x) = \left[\tilde A_0 \left(1-\frac{R_S}{r}\right) ^{1/2}, 0, 0, 0 \right]=\left[\tilde A_0 f^{1/2}, 0, 0, 0 \right]
\end{equation}
with $\tilde A_0$ a constant.
While this is an approximate solution, it will be sufficient
for our purposes given the assumption $|A_\gamma(x)|\ll 1$.
\section{Geodesics}\label{sec: Geodesics}
\label{geodesics}
{In this section, we will study the geodesics of the SFR and perform a dynamical analysis. We compare our results with the corresponding ones of GR.}
From the definition of the metric function \eqref{RandersL} we have:
\begin{equation}
\label{metric function}
F(x,dx) = \left(-g_{\mu\nu}(x)dx^{\mu}dx^{\nu}\right)^{1/2} + A_{\gamma}(x)dx^{\gamma}
\end{equation}
where $g_{\mu\nu}(x)$ is the Schwarzschild metric and $A_{\gamma}(x)$ is a one-form vector field with $|A_\gamma(x)|\ll 1$.\\[7pt]
By using Eq.~\eqref{Schwarzchild} and Eq.~\eqref{metric function} is written as:
\begin{equation}
F(x,dx) = \Big[fdt^2 - \frac{dr^2}{f} - r^2 d\theta^2 - r^2 \sin^{2}\theta\, d\phi^2\Big]^{1/2} + A_{\gamma}(x)dx^{\gamma}.
\label{metrical function}
\end{equation}
We define the Lagrangian
\begin{equation}\label{Lagrangian}
L(x,\dot x)=F(x,\dot x)=\Big[f\dot{t^{2}} - \frac{\dot{r^{2}}}{f}-r^{2}\dot\theta^{2}-r^{2}sin^{2}\theta\dot{\phi^2}\Big]^{1/2}+\tilde{A_{0}}f^{1/2}\dot{t},
\end{equation}
where we have used Eqs.~\eqref{metrical function} and~\eqref{Asolution}.
From the Euler-Lagrange equations
\begin{equation}
\frac{d}{d\tau}\frac{\partial L}{\partial \dot{x}^{\mu}}=\frac{\partial L}{\partial x^{\mu}}
\end{equation}
we find the equations for the geodesics:
\begin{equation}
\ddot{x}^{\lambda}+\Gamma^{\lambda}_{\mu\nu}\dot{x}^{\mu}\dot{x}^{\nu}+g^{\kappa\lambda}\Phi_{\kappa\mu}\dot{x}^{\mu}=0,
\end{equation}
where $\Gamma^{\lambda}_{\mu\nu}$ are the Christoffel symbols of Riemann geometry, $\Phi_{\kappa\mu}=\partial_{\kappa}A_{\mu}-\partial_{\mu}A_{\kappa}$ and $A_{\mu}$ is the solution from Eq.~\eqref{Asolution}. We notice that from the definition of $\Phi_{\kappa\mu}$ we get a rotation form of geodesics. If $A_{\mu}$ is a gradient of a scalar field, $A_{\mu}=\pder{\Phi}{x^{\mu}}$ then $\Phi_{\kappa\mu}=0$ and the geodesics of our model are identified with the Riemmanian ones.
\\[7pt]
The geodesics of our model can then be explicitly written in the form:
\begin{align}
&\ddot t + \frac{1-f}{rf}\dot r \dot t =-\tilde{A}_{0}\dot{r}\frac{f^{-3/2}(1-f)}{2r}
\label{geodesics0}\\
&\ddot r + \frac{f(1-f)}{2r} \dot t^2 - \frac{1-f}{2rf} \dot r^2 - rf \big( \dot \theta^2 + \sin^2 \theta \dot \phi^2 \big)=-\tilde{A}_{0}\dot{t}\frac{f^{1/2}(1-f)}{2r}
\label{geodesics1}\\
&\ddot \theta + \frac{2}{r} \dot \theta \dot r - \frac{1}{2}\sin 2\theta \, \dot \phi^2=0
\label{geodesics2}\\
&\ddot \phi + \frac{2}{r} \dot \phi \dot r + 2\cot \theta \, \dot \theta \dot \phi=0
\label{geodesics3}
\end{align}
{From the relations \eqref{geodesics0}-\eqref{geodesics3} we notice that the first two dynamical equations involve a contribution of extra terms
particular to the SFR spacetime while the last two relations are the same as in GR.} \\
We now make a key assumption regarding the angular
dependence of the model. Namely,
by using $\theta=\frac{\pi}{2}$ we notice that Eq.~\eqref{geodesics2} is satisfied and equations \eqref{geodesics0}, \eqref{geodesics1} and \eqref{geodesics3} can be written as:
\begin{align}
&\ddot t + \frac{1-f}{rf}\dot r \dot t =-\tilde{A}_{0}\dot{r}\frac{f^{-3/2}(1-f)}{2r}
\label{geodesics0v2}\\
&\ddot r + \frac{f(1-f)}{2r} \dot t^2 - \frac{1-f}{2rf} \dot r^2 - rf\dot\phi^2 =-\tilde{A}_{0}\dot{t}\frac{f^{1/2}(1-f)}{2r}
\label{geodesics1v2}\\
&\ddot \phi + \frac{2}{r} \dot \phi \dot r=0
\label{geodesics3v2}
\end{align}
From Eq.\eqref{geodesics3v2} we find:
\begin{equation}
r^{2}\dot{\phi}=J=const.
\label{angular momentum}
\end{equation}
where $J$ is the angular momentum and the relevant equation
represents its conservation law.
If we use the relation $f'=\frac{1-f}{r}$ where $f=1-\frac{2GM}{r}$ and the Leibniz chain-rule $\frac{d}{d\tau}=\frac{dr}{d\tau}\frac{d}{dr}=\dot r \frac{d}{dr}$, then Eq.~\eqref{geodesics0v2} can be written as:
\begin{equation}
f\ddot t + \frac{df}{d\tau}\dot t = -\tilde{A}_{0}\frac{df^{1/2}}{d\tau}
\end{equation}
which, in turn, gives us
\begin{equation}\label{energy0}
f\dot{t}+\tilde{A}_{0}f^{1/2} = \mathcal{E}_R=const.
\end{equation}
where $\mathcal{E}_R$ is the energy of the particle moving along the geodesic.
We notice
that the first term constitutes the energy for a particle moving along the geodesics in general relativity, $E_{GR}=f\dot{t}$ and we can rewrite the
relevant expression as:
\begin{equation}
E_{GR}+\tilde{A}_{0}f^{1/2}=\mathcal{E}_{R}.
\end{equation}
By using Eq.~\eqref{geodesics1v2} with \eqref{angular momentum},
and \eqref{energy0},
we arrive at the (effectively one-degree-of-freedom) radial equation:
\begin{equation}
\ddot r + \frac{1-f}{2rf}(\mathcal{E}_{R}^{2}-\dot{r}^{2}) -\frac{fJ^{2}}{r^{3}}=\tilde{A}_{0}\mathcal{E}_{R}\frac{f^{-1/2}(1-f)}{2r}
\label{radial1}
\end{equation}
where we omitted $O(\tilde A_{0}^{2})$ terms.
As before, we use the relation $f'=\frac{1-f}{r}$ in \eqref{radial1} to
bring it to the equivalent form:
\begin{equation}
\label{radial2}
\ddot r + \frac{f'}{2f}(\mathcal{E}_{R}^{2}-\dot r^{2})-\frac{fJ^{2}}{r^{3}}=\frac{\tilde A_{0}\mathcal{E}_{R}}{2}f^{-1/2}f'
\end{equation}
We can further simplify the Eq.~\eqref{radial2} by using the Leibniz chain-rule $\frac{d}{d\tau}=\frac{dr}{d\tau}\frac{d}{dr}=\dot r \frac{d}{dr}$ and upon deriving the first integral of the motion, we obtain:
\begin{equation}
\dot{r}^{2} + f\left(\frac{J^{2}}{r^{2}}+\epsilon\right)+2\tilde{A}_{0}\mathcal{E}_{R}f^{1/2}=\mathcal{E}_{R}^{2}
\label{radial3}
\end{equation}
where $\epsilon$ is a constant and for $\epsilon = 0$ we have null geodesics.
The first two terms from \eqref{radial3} constitute the total energy in general relativity (GR), $\mathcal{E}^{2}_{GR}=\dot{r}^{2} + f(\frac{J^{2}}{r^{2}}+\epsilon)$ and the third term emerges from the structure of SFR spacetime and its energetic contribution.
Therefore Eq.~\eqref{radial3} can be written as:
\begin{equation}
\label{energy}
\mathcal{E}^{2}_{GR}+2\tilde{A}_{0}\mathcal{E}_{R}f^{1/2}=\mathcal{E}_{R}^{2}
\end{equation}
Eq.~\eqref{energy} shows that the term $A_{\gamma}(x)$ from \eqref{Asolution} provides an additional energy contribution to the system of GR.
Below, we give the Figures \ref{graph-sfr-r-s}, \ref{graph-sfr-phi-s} and Fig. \ref{graph-sfr-x-y} for the geodesics of GR and SFR we have obtained by solving the equations \eqref{geodesics0}-\eqref{geodesics3}.
The relevant ordinary differential equations are solved via
a standard solver within Mathematica and $(r,\phi)$ are presented
as a function of $\tau$, while Fig.~\ref{graph-sfr-x-y} presents
the evolution in the original $(x,y)$ plane. In our case, we assume $R_{s}=2$ and initial radial distance $r_{0}=3$, so the photons are found on the photonsphere with $r_{ph}=\frac{3}{2}R_{s}=3$ in the GR case. The deviation between
the trajectories of the SFR and those of the GR is clearly discernible
in both figures.
\begin{figure}[H]
\begin{subfigure}{0.5\textwidth}
\includegraphics[scale=0.4]{SFR-GR-r-s.PNG}
\caption{This is an ($\tau$,r) graph for the geodesics of photons\\ for angular momentum $J=4$ and initial radial\\ distance $r_{0}=3$. The red line shows the SFR \\geodesics and the blue line the GR geodesics.}
\label{graph-sfr-r-s}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\includegraphics[scale=0.4]{SFR-GR-phi-s.PNG}
\caption{This is an $(\tau,\phi)$ graph for the geodesics of photons for angular momentum $J=4$ and initial radial distance $r_{0}=3$. The red line shows the SFR geodesics and the blue line the GR geodesics.}
\label{graph-sfr-phi-s}
\end{subfigure}
\end{figure}
From Fig.~\ref{graph-sfr-r-s}, we can see that the radial component in the SFR model takes lower values compared to the GR one which remains constant. This difference between the r-components of SFR and GR can be interpreted as the
increase of the radius of the photonsphere due to the one-form $A_{\gamma}$ as we have shown in \cite{Kapsabelis:2021dpb}.
This leads the orbit of the photon to fall inside the event horizon because the initial distance $r_{0}=3$ and energy are not sufficient to allow circular orbits of the photonsphere.
That means for an
orbit with $r$ constant in the SFR model, the particle needs more energy compared to the GR case.
In Fig.~\ref{graph-sfr-x-y}, the geodesics of GR and SFR are depicted. In the case of GR, the photons move in circular orbits around the black hole. In the SFR model, the photons follow a spiral orbit and fall inside the event horizon.
It is important to remind the reader here that underlying these results
is the key assumption of $\theta=\frac{\pi}{2}$ which allows the
reduction of the model to an \textit{effective single degree-of-freedom
system}. It is important in future work to consider how deviations
from this equilibrium value (and the corresponding incorporation
of the full dynamical system) may affect the conclusions presented
above. However, as the latter is outside the scope of the present
study, we now focus on the further analysis of the effective
potential of the SFR model and its implications for the
phase portrait of the relevant system.
\begin{figure}
\includegraphics[scale=0.5]{SFR-GR-x-y.PNG}
\caption{This is an x-y graph for the geodesics of photons for angular momentum $J=4$ and initial radial distance $r_{0}=3$. The red line shows the SFR geodesics and the blue line the GR geodesics.}
\label{graph-sfr-x-y}
\end{figure}
\section{Effective Potential of SFR model}\label{sec: Effective Potential}
In this section, we will study the effective potential of the SFR model and compare it with the effective potential of GR.
The equation of the energy in GR reads:
\begin{equation}
\label{energy eq GR}
\dot{r}^{2} + f\left(\frac{J^{2}}{r^{2}}+\epsilon\right)=\mathcal{E}_{GR}^{2} \end{equation}
We see from \eqref{energy eq GR} that the effective potential
energy landscape is given by:
\begin{equation}
V_{eff,GR}=\frac{1}{2}f \left(\frac{J^{2}}{r^{2}}+\epsilon \right)
\end{equation}\\
In Fig.\ref{VeffGR}, we show the graph for the effective potential in GR for angular momentum $J=3,J=4$ and $J=5$ to examine its variation for
different values of the angular momentum.
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{Veff-GR.PNG}
\caption{Graph for the $V_{eff}(r)$ in the GR model for angular momentum $J=3$, $J=4$ and $J=5$.}
\label{VeffGR}
\end{figure}
We now recall the key difference (and associated additional contribution)
to the energetics of the SFR model. In particular,
the energy equation for the latter, derived from Eq.~\eqref{radial2}, is given as:
\begin{equation}
\dot{r}^{2} + f\left(\frac{J^{2}}{r^{2}}+\epsilon \right) + 2\tilde{A}_{0}\mathcal{E}_{R}f^{1/2}= \mathcal{E}_{R}^{2}
\label{energy eq SFR}
\end{equation}
In \eqref{energy eq SFR} the effective potential can be written in the form
\begin{equation}
\label{VeffSFReq}
V_{eff,SFR}=\frac{1}{2}f \left(\frac{J^{2}}{r^{2}}+\epsilon \right)+ \tilde{A}_{0}\mathcal{E}_{R}f^{1/2}
\end{equation}\\
The graph for the effective potential in the SFR model $(V_{eff},r)$ is depicted in Fig.~\ref{VeffSFR}, in this case
for different values of angular momentum.
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{Veff-SFR.PNG}
\caption{Graph for the $V_{eff}(r)$ in the SFR model for angular momentum $J=3$, $J=4$ and $J=5$.}
\label{VeffSFR}
\end{figure}
In Fig. \ref{VeffSFRGR1} and \ref{VeffSFRGR5} we show the effective potentials of the SFR and GR models comparing the two for $J=1$ and $J=5$.
As we can see in Fig.~\ref{VeffSFRGR1}, the difference between GR and SFR is bigger than that of Fig.~\ref{VeffSFRGR5}. Notably, when the contribution
of the angular momentum is weaker, the difference between the two models
is more substantial/clearly discernible. When the angular momentum becomes
large, the relevant difference is rather weak and the $V_{eff}$
of the two models become proximal.
\begin{figure}[H]
\begin{subfigure}{0.5\textwidth}
\includegraphics[scale=0.4]{Veff-SFR-GR-J=1.PNG}
\caption{Graph for the $V_{eff}(r)$ in GR (blue line)\\ and SFR (red line) for angular momentum $J=1$.}
\label{VeffSFRGR1}
\end{subfigure}
\begin{subfigure}{0.4\textwidth}
\includegraphics[scale=0.4]{Veff-SFR-GR-J=5.PNG}
\caption{This is a graph for the $V_{eff}(r)$ in GR (blue line) and SFR (red line) for angular momentum $J=5$.}
\label{VeffSFRGR5}
\end{subfigure}
\end{figure}
In Figs.~\ref{phase-sfr}-\ref{phase-gr} and also~\ref{phase-sfr-gr},
we observe the phase portraits associated with the
effective potentials depicted above. These phase portraits
reflect the existence of an energy barrier whose precise
height depends on the value of the angular momentum.
Energies below this barrier height result in reflection from
the outside and trapping from the inside. On the other hand,
energies higher than those of the barrier result in
reaching the Schwarzschild radius (if the particle is
coming from the outside) or reaching infinity (if the
particle is moving outward from the inside). The
latter figure demonstrates the differences between the two
phase portraits which are quantitative but not qualitative.
\begin{figure}[H]
\begin{subfigure}{0.55\textwidth}
\includegraphics[scale=0.5]{SFR_phase_plot.jpg}
\caption{This is a phase plot for the radial geodesics of \\SFR,
representing the trajectories in $(r, \dot{r})$ space.}
\label{phase-sfr}
\end{subfigure}
\begin{subfigure}{0.55\textwidth}
\includegraphics[scale=0.5]{GR_phase_plot.jpg}
\caption{This is a phase plot for the radial geodesics of GR
representing the trajectories in $(r, \dot{r})$ space.}
\label{phase-gr}
\end{subfigure}
\end{figure}
\begin{figure}[h]
\includegraphics[scale=0.5]{GR-SFR_phase_plot.jpg}
\caption{This is a comparison between the radial phase portraits of the GR (blue) and SFR (yellow) models.}
\label{phase-sfr-gr}
\end{figure}
\section{Deflection Angle}\label{sec: Deflection}
In this section, we will deal with the deflection angle of the SFR model and we will compare our findings with the corresponding ones of the GR model. In this consideration, we take into account photons that pass close to a central mass M.
From Eq.~\eqref{radial3} for photons, we put $\epsilon=0$ and we get:
\begin{equation}
\label{b1eq}
\frac{\dot{r}^{2}}{J^2}+\frac{f}{r^{2}}+
\frac{2\tilde{A_{0}}f^{1/2}}{Jb}=\frac{1}{b^2}
\end{equation}
where $b=J/\mathcal{E}_{R}$.\\
By using the Leibniz chain-rule $\dot{\phi}=\frac{d\phi}{d\tau}=\frac{d\phi}{dr}\frac{dr}{d\tau}=\frac{d\phi}{dr}\dot{r}$ with the relations \eqref{angular momentum} and \eqref{b1eq} we have:
\begin{equation}
\frac{\dot{r}^{2}}{\dot{\phi}^{2}}=r^{4}\left(\frac{1}{b^2}-\frac{f}{r^{2}}-
\frac{2\tilde{A_{0}}f^{1/2}}{Jb}\right)
\end{equation}
After some rearrangements we find:
\begin{equation}
\label{phideflect}
\frac{d\phi}{dr}=\frac{1}{r^2}\left(\frac{1}{b^2}-\frac{f}{r^{2}}-
\frac{2\tilde{A_{0}}f^{1/2}}{Jb}\right)^{-1/2}
\end{equation}
The deflection angle is calculated by the integration of \eqref{phideflect}:
\begin{equation}
\label{int1}
\Delta\phi_{SFR}=2\int_{r_1}^{\infty}\frac{dr}{r^2}\left[\frac{1}{b^2}-\frac{1}{r^{2}}\left(1-\frac{2GM}{r}\right)-
\frac{2\tilde{A_{0}}}{Jb}\left(1-\frac{2GM}{r}\right)^{1/2}\right]^{-1/2}
\end{equation}
where we have used $f=1-\frac{2GM}{r}$.\\
We perform a change of variables in the integral of Eq.~\eqref{int1}:
\begin{equation}
\Delta\phi_{SFR}=2\int^{w_{1}}_{0}dw\left[1-w^{2}\left(1-\frac{2GM}{b}w\right)-2a\left(1-\frac{2GM}{b}w\right)^{1/2}\right]^{-1/2}\label{intwdeflection}
\end{equation}
where we have set $w=\frac{b}{r}$ and $a=\frac{\tilde{A_{0}}b}{J}$.\\
If we expand the integral in powers of $\frac{2GM}{b}$ and $a$ we find:
\begin{equation}
\Delta\phi_{SFR}\approx 2\int_{0}^{w_1}dw\frac{1+\frac{GM}{b}w}{\left[\left(1-2a\right)+\frac{2GM}{b}w-w^{2}\right]^{1/2}}
\label{intdeflection}
\end{equation}
where we have omitted second order terms.\\
By evaluating the integral in Eq.~\eqref{intdeflection} we get (see Appendix 3):
\begin{equation}
\Delta\phi_{SFR}=\pi+\frac{4GM}{b}\frac{1-a}{\sqrt{1-2a}}
\end{equation}
The deflection angle $\delta\phi_{SFR}$ can be found as:
\begin{align}
&\delta\phi_{SFR}=\Delta\phi_{SFR}-\pi\Rightarrow\nonumber\\[7pt]
&\delta\phi_{SFR}=\frac{4GM}{b}\frac{1-a}{\sqrt{1-2a}} \label{SFRangle}
\end{align}
If we expand Eq.~\eqref{SFRangle} in powers of $a=\frac{\tilde{A}_{0}b}{J}$ the deflection angle can be written as:
\begin{equation} \delta\phi_{SFR}\approx \left(1+\frac{a^{2}}{2} \right)\frac{4GM}{b}
\end{equation}
The deflection angle $\delta\phi$ of GR \cite{Hartle2002, Carroll}
is given by:
\begin{equation}
\label{angleGR}
\delta\phi_{GR}=\frac{4GM}{b}
\end{equation}
Therefore, we notice that the deflection angle of SFR includes a small additional Randers contribution term $a$ which shows a small deviation from GR because $|\tilde{A_0}|\ll 1$.
We can see from Eq.~\eqref{SFRangle} that:
\begin{equation}
\lim_{\tilde{A_0}\rightarrow 0}\delta\phi_{SFR}=\delta\phi_{GR}
\end{equation}
The small difference of the deflection angle of the SFR model from the GR one can plausibly be attributed to the Lorentz violations \cite{Kostelecky:2011qz} or on the small amount of energy which is added to the gravitational potential of SFR. \\[7pt]
\textbf{\textit{Remark}}: By considering the following relation, we can connect the geometrical concept of the curvature $\kappa_{\phi}=\frac{d\phi}{d\tau}$ of a path with the deflection angle $\delta\phi$ in the following way:
\begin{align}
\dot{\phi}=\frac{d\phi}{d\tau}&=\frac{d\phi}{dr}\frac{dr}{d\tau}=\frac{d\phi}{dr}\dot{r}\Rightarrow\\
\label{chainrule}
&\kappa_{\phi}=\frac{d\phi}{dr}\dot{r}\Rightarrow\\
&\delta\phi=\int\frac{\kappa_{\phi}}{\dot r} dr
\end{align}
This form of curvature can be called \textit{deflection curvature}.
\section{Conclusions \& Future Challenges}\label{sec: Conclusions}
In this article, we investigated the analytic form of the geodesics of the model SFR which was introduced in previous works \cite{Triantafyllopoulos:2020vkx,Kapsabelis:2021dpb}. A dynamical analysis was presented based on the energy and angular momentum of a particle along of geodesics (null or timelike) of
the SFR spacetime. Comparisons between the SFR and GR
were provided. We found that there is a small deviation from
the GR model which is due to the dynamical term $A_{\gamma }(x)$.
We also formulated and studied an effective potential of our model and we compared the one of the SFR case once again
with the effective potential of GR attributing
the small but discernible differences to the specific
structure of (and perturbation incorporated within)
the SFR spacetime. The relevant differences in the
trajectories were illustrated both in the evolution
over the time-variable $\tau$ and in the $(x,y)$ plane.
In addition, we calculated the deflection angle
for the SFR spacetime and we compared with the corresponding one of GR. The result is a small difference of the
SFR model from GR, it is possibly caused by Lorentz violations or by the small amount of energy which is added to the gravitational potential of SFR spacetime.
It is important to note that this work opens a number of
interesting directions of further study for the future.
On the one hand, the traditional assumption of $\theta=\pi/2$
made over here is clearly a restrictive one that
simplifies the equations of motion automatically satisfying
the dynamics for the angular variable $\theta$ with the
latter being at steady state. However, more generally,
one can straightforwardly envision scenarios
where this condition is no longer satisfied.
It is then of interest to explore
if one
starts in the vicinity of $\pi/2$ whether one stays
in that neighborhood or perhaps if one deviates
away from this steady state and how the associated
dynamics of the full 4-degree-of-freedom space is accordingly
explored.
Another aspect that is also worth further exploring
is that of the small amplitude covector deviation
from the General Relativity standard model. Here,
we have limited our considerations to the realm
of associated small amplitude perturbations (where
leading order expansions of the field would suffice).
However, it would also be of interest to explore the situation
when one gradually deviates from the realm of this
approximation as well.
{In addition, applications of geodesics of the SFR model can be pursued for
more concrete cosmological studies such as, e.g., for the case of the S2 stars orbiting the black hole in Sagittarius A* in which the geodesics of the star are perturbed from the classical Keplerian orbits because of the distribution of stellar remnants.
} Indeed, our hope is that this work may
pave the way towards testing the Schwarzschild-Finsler-Randers gravitational model which incorporates features going beyond the standard Riemannian geometry of spacetime.
In this vein, some of the above topics are presently
under consideration and associated results will
be presented in future publications.
\section{Acknowledgements}
The authors would like to thank Gary Gibbons for his valuable suggestions.
|
\section{Introduction}\label{sec:intro}
The motion of a compressible viscous, heat-conductive, and Newtonian polytropic fluid occupying a spatial domain $\Omega \subset \mathbb{R}^n$ for $n\ge 2$
is governed by the following full compressible Navier-Stokes system (\textbf{CNS}) in the Eulerian coordinates:
\begin{equation}\label{eqs:FNS}
\begin{cases}
\partial_t \rho + {\mathrm {div}}(\rho \u)=0,\\
\partial_t (\rho \u ) + {\mathrm {div}} ( \rho \u \otimes \u ) +\nabla P = {\mathrm {div}}\,\mathbb{S},\\
\partial_t(\rho E)+{\mathrm {div}} ( \rho E\u + P\u ) = {\mathrm {div}} (\u\mathbb{S}) +\text{div}(\kappa\nabla e).
\end{cases}
\end{equation}
Here and throughout the paper, $\rho\geq 0$ denotes the mass density, $\u=(U^1, \dotsc, U^n )^\top$ the fluid velocity,
$P$ the pressure of the fluid, $E=e+\frac{1}{2}\snorm{\u}^2$ the specific total energy,
$e$ the specific internal energy, ${\boldsymbol{x}}=(x^1,\dotsc, x^n)^\top\in \Omega$ the Eulerian spatial coordinates,
and $t\geq 0$ the time coordinate. The equations of state for polytropic fluids are
\begin{equation}\label{2}
P=R\rho\theta=(\gamma-1)\rho e,\hspace{7mm} e=c_v\theta,\hspace{7mm}c_v=\frac{R}{\gamma-1},
\end{equation}
where $R$ is the gas constant, $\theta$ the absolute temperature, $c_v$ the specific heat at constant volume,
and $\gamma>1$ the adiabatic exponent. The viscosity stress tensor $\mathbb{S}$ is given by
\begin{equation}\label{3}
\mathbb{S}=2\mu D(\u)+\lambda\,{\mathrm {div}}\u\,\mathbb{I}_n,
\end{equation}
where $D(\u)=\frac{\nabla \u+(\nabla \u)^\top}{2}$ is the deformation tensor, $\mathbb{I}_n$ the $n\times n$ identity matrix,
$\mu$ the shear viscosity coefficient, and $\lambda+\frac{2}{n}\mu$ the bulk viscosity coefficient.
The constant, $\kappa$, in the energy equation is defined by $\kappa\vcentcolon= \frac{\kappa_{Q}}{c_{V}}$,
where $\kappa_{Q}\ge0$ is the coefficient of thermal conductivity. Furthermore, $(\mu,\lambda,\kappa)$ satisfies the following physical condition:
\begin{equation}\label{eqs:viscoef}
\mu > 0, \qquad \lambda+\dfrac{2}{n}\mu\ge 0, \qquad \kappa>0.
\end{equation}
In this paper,
we are concerned with global spherically symmetric
solutions of the form:
\begin{equation*}
(\rho,\u,e)({\boldsymbol{x}},t)
=(\rho(\snorm{{\boldsymbol{x}}},t),u(\snorm{{\boldsymbol{x}}},t)\dfrac{{\boldsymbol{x}}}{\snorm{{\boldsymbol{x}}}},
e(\snorm{{\boldsymbol{x}}},t))
\end{equation*}
of system \eqref{eqs:FNS}--\eqref{eqs:viscoef} in domain $\Omega={\mathbb R}^n$ with Cauchy data:
\begin{equation}\label{eqs:CauchyInit}
(\rho,\u,e)({\boldsymbol{x}},0)=(\rho_0,\u_0,e_0)({\boldsymbol{x}})= ( \rho_0(\snorm{{\boldsymbol{x}}}),u_0(\norm{{\boldsymbol{x}}})\dfrac{{\boldsymbol{x}}}{\snorm{{\boldsymbol{x}}}},e_0(\snorm{{\boldsymbol{x}}})) \qquad\, \text{for ${\boldsymbol{x}}\in {\mathbb R}^n$.}
\end{equation}
Therefore, for the spherically symmetric case with the radial coordinate variable $r=|{\boldsymbol{x}}|$, problem \eqref{eqs:FNS}--\eqref{eqs:CauchyInit} can be reformulated as:
\begin{equation}\label{eqs:SFNS}
\begin{cases}
\partial_t\rho + \partial_r(\rho u) + m\dfrac{\rho u}{r} =0,\\[2pt]
\rho \partial_t u +\rho u \partial_r u + \partial_r P(\rho,e) = (2\mu+\lambda) \partial_r\big(\partial_ru+m\dfrac{u}{r}\big),\\[2pt]
\rho \partial_t e +\rho u \partial_r e +P(\rho,e)\big(\partial_r u+m\dfrac{u}{r}\big)
= 2\mu \big(\snorm{\partial_ru}^2 + m \dfrac{u^2}{r^2}\big) + \lambda \big(\partial_ru+m\dfrac{u}{r}\big)^2
+ \kappa \big(\partial_r^2e+ m\dfrac{\partial_re}{r}\big)
\end{cases}
\end{equation}
in $\Omega_T \vcentcolon=[0,\infty)\times[0,T]$ with $m\equiv n-1$, and
\begin{align}
& u(0,t)=0, \qquad \partial_r e(0,t)=0 && \text{for} \ \ t\in[0,T], \label{BC}\\
&(\rho,u,e)(r,0)= (\rho_0,u_0,e_0)(r) && \text{for} \ \ r\in [0,\infty). \label{IC}
\end{align}
The boundary conditions \eqref{BC} is derived from the continuity of ${\boldsymbol{x}}\mapsto (\u,\nabla e)({\boldsymbol{x}},t)$ at $\snorm{{\boldsymbol{x}}}=0$.
There is a large literature regarding the full \textbf{CNS}. For the general three-dimensional (3-D) flow, in the absence
of vacuum ({\it i.e.}, $\inf_{{\boldsymbol{x}}}\rho_0({\boldsymbol{x}})>0$), the local well-posedness of classical solutions of the Cauchy problem
for \eqref{eqs:FNS} follows from the standard symmetric hyperbolic-parabolic structure satisfying the well-known Kawashima's
condition ({\it cf}. \cites{I1, I2, KA, Nash, serrin,Tani}), which was extended to the global one
by Matsumura-Nishida \cite{MN} for the initial data close to a non-vacuum equilibrium in some Sobolev space $H^s(\mathbb{R}^3)$ for $s>\frac{5}{2}$.
For both isentropic and non-isentropic flow,
we also refer the reader to
Danchin \cites{danchin,danchin2} for strong solutions
with small data in Besov spaces of ${\mathbb R}^n$ for $n\ge 2$,
and Hoff \cites{H4,H3} for weak solutions with small and discontinuous data in some Sobolev spaces of ${\mathbb R}^n$ for $n=2$ or $3$.
Considering the global well-posedness of classical or weak solutions with arbitrarily large initial data,
a subtle issue is whether, in the absence of smallness conditions, cavitation may occur in solutions of \eqref{eqs:FNS}
for barotropic and/or full \text{CNS},
which has no clear answer yet so far, except the one-dimensional (1-D) case.
Specifically, for the 1-D flow,
the unique global classical solution was obtained by Kazhikhov-Shelukhin
\cite{KS} for large initial data in a bounded interval with $\inf_{{\boldsymbol{x}}} \rho_0({\boldsymbol{x}})>0$, and
this theory was then extended to the unbounded domains in Kawashima-Nishida in \cite{KN}.
The key estimate developed in \cites{KS,KN} is the pointwise upper and lower bounds of the density,
which have led to many notable results for the problems concerning 1-D flow ({\it e.g.}, \cites{H1,CHT}).
The global existence and stability of 1-D weak solutions with large and discontinuous initial data can be found in
Chen-Hoff-Trivisa \cite{CHT}, Jiang-Zlotnik \cite{JZlotnik}, Zlotnik-Amosov \cites{ZA1,ZA2}, and the references cited therein.
Furthermore, it was proved in Hoff-Smoller \cite{HS} that the weak solutions of the 1-D \textbf{CNS} do not develop
vacuum regions, as long as no
vacuum states are present initially.
While the multidimensional (M-D) case
is much more complicated: up to now, very few solid progress is known on the problems
of well-posedness or vacuum formation with large data.
It is worth pointing out that the approaches used in the references mentioned above do not work directly
when initial vacuum appears ({\it i.e.}, $\inf_{\boldsymbol{x}} {\rho_0({\boldsymbol{x}})}=0$), which occurs when some physical requirements are imposed,
such as finite total initial mass and energy in the whole space.
One of the main issues in the presence of vacuum is the degeneracy of the time evolution operator,
which makes it difficult to understand the behavior of the velocity field near the vacuum.
In the terms of structures, system \eqref{eqs:FNS} is a hyperbolic-parabolic coupled system in the fluids region,
but degenerates to a hyperbolic-elliptic one near the vacuum region.
By imposing initially some compatibility conditions, Cho-Kim in \cite{CK} firstly established the local well-posedness
of 3-D strong solutions with vacuum in terms of $(\rho, u,\theta)$ in some homogeneous Sobolev spaces,
which has been extended recently to be global ones with small energy but large oscillations by Huang-Li \cite{HL}, Wen-Zhu \cite{WZ},
and the references cited therein.
The first global existence of M-D weak solutions of finite energy for the isentropic \textbf{CNS}
with generic data and vacuum was established in Lions \cite{L2},
especially for $\gamma \ge\frac{9}{5}$ when $n=3$;
see also Feireisl-Novotn\'{y}-Petzeltov\'{a} \cite{fu1} for $\gamma>\frac{3}{2}$ when $n=3$.
Moreover, this theory has been partially extended to the non-isentropic compressible flow with
$(\mu,\,\lambda,\,\kappa)$ depending on the temperature by Feireisl in
\cites{fu2,fu3}, where the thermal energy equation holds as an inequality in the sense of distributions.
However, the uniqueness problem of these weak solutions is widely open due to their fairly low regularity.
Also see
\cites{L1,LX1,BJ,VY,BVY} and the references cited therein.
For the M-D spherically symmetric flow, some favourable regularity properties may be expected, since the equations exhibit largely 1-D behaviour when the flow is away from both the origin and the far-field. However, compared with the 1-D case, in order to establish the existences of M-D spherically symmetric classical or weak solutions with arbitrarily large initial data of \eqref{eqs:SFNS}, some essential difficulties are encountered including the coordinate singularity at the origin, indicated by the singular factor $\frac{1}{r}$ in system \eqref{eqs:SFNS} and infinite initial total mass/energy from the far field;
also see \cites{JZ,JZ1,JZ2,H2,HJ,MRS}.
To circumvent these issues, it is customary to consider instead the initial-boundary value problem posed in some annular domains
\begin{equation*}
\{ ({\boldsymbol{x}},t)\in {\mathbb R}^n \times [0,\infty)\,\vcentcolon\,a \le \snorm{{\boldsymbol{x}}} \le b \}
\end{equation*}
with proper boundary conditions, where $a>0$ and $b>0$ are both finite constants. The global well-posedenss of 3-D
strong solutions with cylindrical symmetry and strictly positive initial density to the full \textbf{CNS}
has been given by Frid-Shelukhin \cite{FS} in an infinite cylinder of finite radius parallel to and centered
along the $x^3$--axis in $\mathbb{R}^3$. In the domain between a static solid core and a free boundary
connected to a surrounding vacuum state, Yashima-Benabidallah in
\cite{YB} first established the global existence of spherically symmetric weak solutions with large and discontinuous initial data
to the full \textbf{CNS} with boundary data prescribed so that the uniform boundedness of the total energy of solutions is direct to achieve,
and later Chen-Kratka in \cite{CGQK} constructed the global large solutions in a more regular space with free normal stress on
the free boundary which is a natural physical situation for the fluids with the surrounding vacuum state.
Under the assumption that $0<\inf_{\boldsymbol{x}}\rho_0({\boldsymbol{x}})< \sup_{\boldsymbol{x}}\rho_0({\boldsymbol{x}}) < \infty$, based on an elaborate truncation procedure on the initial data,
Jiang in
\cite{J} proved the global existence of spherically symmetric smooth solutions for (large) initial
data in the domain exterior to a ball
\begin{equation*}
\{ ({\boldsymbol{x}},t)\in {\mathbb R}^n \times [0,\infty)\,\vcentcolon\, r=\snorm{{\boldsymbol{x}}}\ge a \} \qquad\,\,\,\mbox{in $\mathbb{R}^n$ for $n=2$ or $3$}
\end{equation*}
with the following boundary condition and the far-field behavior:
\begin{equation*}
\begin{cases}
u(a,t)=0, \ \ \partial_r e(a,t)=0 \qquad\,\,\,\, \text{for $t\in[0,T]$},\\[2pt]
\lim\limits_{r\to\infty} ( \rho, u , e , \partial_r e )(r,t) = (1,0,1,0).
\end{cases}
\end{equation*}
A further natural question is whether the uniform estimates in $a>0$ could be obtained so that
the region containing the origin can be included as $a\searrow 0$. In the absence of the vacuum,
Hoff \cite{H2} first constructed the global spherically symmetric weak solutions of the Cauchy problem
for the isothermal flow with large data and strictly positive initial density via the limit
process as $a\searrow 0$ in $\mathbb{R}^n$ for $n \geq 2$;
see also Haspot \cite{Has} for the case of isentropic flow with degenerate viscosity.
For the initial data with finite total initial mass and energy in the whole space,
Jiang-Zhang \cites{JZ,JZ1,JZ2} proved the global existence of spherically symmetric weak solutions of the Cauchy problem
for the isentropic flow in $\mathbb{R}^n$ for $n = 2, 3$ when $\gamma>1$.
Also see Xin-Yuan \cite{XY} for some analysis on the vacuum formation problems of these weak solutions.
Later, when $\inf_{\boldsymbol{x}} \rho_0({\boldsymbol{x}})>0$,
Hoff-Jenssen in \cite{HJ} proved the global existence of spherically symmetric weak solutions of the initial-boundary
value problems of the non-isentropic flow in a finite ball in ${\mathbb R}^3$, which also works for the cylindrically symmetric case.
However, the carefully designed energy estimates and pointwise estimates applied in \cite{HJ} depends strictly on the boundedness
of the domain under the consideration, which seems difficult to adapt the arguments to the corresponding Cauchy problem.
In fact, this Cauchy problem is highly nontrivial due to the strong coupling between the velocity and temperature,
the coordinate singularity at the origin, and infinite initial total mass and energy from the far field.
To the best of our knowledge, in the existing literature,
there have been no such results in this direction for the non-isentropic compressible flow with infinite energy.
The main aim of this paper is to establish the global-in-time existence of weak solutions of the Cauchy problem
for the full Navier-Stokes equations for M-D compressible heat-conducting flow
in $\mathbb{R}^n$ for $n\ge 2$,
with initial data that are large, discontinuous, spherically symmetric, and away from the vacuum, and to analyze the underlying properties of these weak solutions.
To achieve these, our main strategy is to regard the Cauchy problem as the limit of a series of initial-boundary value problems (IBVPs) that are formulated in finite annular regions and, as the key argument, to establish a set of {\it a-priori} estimates that are uniform with respect to both the inner and outer radii of the annuli considered in the spherically symmetric Lagrangian coordinates. One of the first obstacles in deriving such estimates is the possible degeneracy or singularity of the temporal derivative operators $\rho \partial_t u$ and $\rho \partial_t e$, which are influenced by the behavior of the density. This observation motivates the estimate of pointwise upper and lower bounds of the density; if it can be achieved, then the dissipation structure of the momentum and energy equations can be utilized to gain further information on the solutions.
Such
density bounds
are derived by
exploiting the underlying structure of
the continuity and momentum equations, and
the regional parameters in the derivation must be given careful consideration in order to obtain the density bounds
uniformly with respect to the size of domain.
For this, we face two major difficulties: The first is to ensure that such bounds are independent of
the inner radius $a\in(0,1)$ of the annular region.
To deal with this, we use the crucial observation that the density bounds
can actually be expressed explicitly in terms of the coordinate variables
so that, using this explicit formula, the dependence on $a\in(0,1)$ can be circumvented
by restricting to the spatial regions bounded away from the inner boundary.
The second difficulty is to show that the density bounds
are also uniform over the
outer radius of the annular region, which is resolved by taking the advantages of the specific form
of the entropy inequality for the solutions in the Lagrangian coordinates.
This is
the main motivation for our proof to start by constructing approximate solutions under the Lagrangian formulation of system $\eqref{eqs:SFNS}$,
rather than their Eulerian form. Once the uniform pointwise bounds of the density are established,
it can be used to derive further estimates stemming from the parabolic structure of the momentum and energy equations.
However, suitable spatial cut-off functions have to be introduced in their derivations,
due to the fact that the density bounds described above are restricted on the domain bounded away from the inner boundary.
This leads to certain problematic integrals near the boundary that cannot be estimated with the standard parabolic argument;
we resolve this issue by incorporating the dissipation terms in the entropy inequality.
By overcoming all these difficulties, we obtain the desired domain-independent {\it a-priori} estimates,
which then allow us to construct a global weak solution of the Cauchy problem via the compactness argument.
We now outline the organization of the rest of this paper.
In \S \ref{sec:main}, we first introduce the definition of weak solutions of the Cauchy problem \eqref{eqs:FNS}--\eqref{eqs:CauchyInit}
and then state our main results.
In \S \ref{sec:reform}, the main strategy and the overall scheme of the proof are described.
The Cauchy problem considered can be regarded as the limit of a series of initial-boundary value problems
that are formulated in finite annular regions.
For such approximation problems, in \S \ref{sec:rhobd}, we derive some {\it a-priori} estimates that are independent of
both the inner and outer radii of the annuli considered in the spherically symmetric Lagrangian coordinates.
In \S \ref{sec:WSCEx}, we establish the existence of weak solutions of some exterior problems in the Eulerian coordinates.
Finally, in \S \ref{sec:ato0}, the weak solutions of the original problem are attained via careful compactness arguments
applied to this set of approximate solutions in the Eulerian coordinates.
\section{Global Solutions and Main Theorem}\label{sec:main}
In this section, we present the main theorem of this paper.
For this purpose, we first introduce the following definition
of global weak solutions of the Cauchy problem \eqref{eqs:FNS}--\eqref{eqs:CauchyInit}.
\begin{definition}[Weak Solutions]\label{def:WFAV}
The vector function $(\rho,\u,e)({\boldsymbol{x}},t)$ is called to be a weak solution of
the Cauchy problem \eqref{eqs:FNS}--\eqref{eqs:CauchyInit} in ${\mathbb R}^n \times [0,T]$ if it satisfies the following{\rm :}
\begin{enumerate}[label=(\roman*),ref=(\roman*),font={\normalfont\rmfamily}]
\item There exists a upper semi-continuous function $\underline{r}(t) \vcentcolon [0,T] \to [0,\infty)$ and a constant $C_0=C_0(\rho_0,\u_0,e_0,\gamma,\mu,\lambda,\kappa,n)>0$ such that
$$
\lim\limits_{t\searrow 0}\underline{r}(t)=0, \qquad 0 \le \underline{r}(t)\le C_0 \ \ \text{for all $\, t\in[0,T]$}.
$$
With this, define
$\mathcal{F}\vcentcolon
=\{({\boldsymbol{x}},t)\in {\mathbb R}^n\times[0,T]\,\vcentcolon\, \snorm{{\boldsymbol{x}}}>\underline{r}(t)\,\, \mbox{for $t\in [0,T]$} \}$ as the fluid region.
\smallskip
\item $\rho\in L_{\textnormal{loc}}^{\infty}(\mathcal{F})$, $\rho({\boldsymbol{x}},t)=0$
for $({\boldsymbol{x}},t)\in ({\mathbb R}^n \times [0,T])\backslash \mathcal{F}$ almost everywhere,
$(\u,e)({\boldsymbol{x}},t)$ is locally H\"older continuous in $\mathcal{F}$, and $\mathring{\mathcal{F}}\vcentcolon=\mathcal{F}\cap\{0<t<T\}$ is an open set.
\smallskip
\item For each $\Phi \in C^1\big([0,T]; C_{\rm c}^1 ({\mathbb R}^n) \big)$,
\begin{align*}
\int_{{\mathbb R}^n} \rho({\boldsymbol{x}},t) \Phi({\boldsymbol{x}},t)\, {\mathrm d} {\boldsymbol{x}} - \int_{{\mathbb R}^n}\rho_0({\boldsymbol{x}}) \Phi({\boldsymbol{x}},0)\,{\mathrm d} {\boldsymbol{x}}
=\int_{0}^{t}\int_{{\mathbb R}^n} (\rho\partial_t\Phi+\rho\u\cdot\nabla\Phi)
\, {\mathrm d}{\boldsymbol{x}} {\mathrm d} s.
\end{align*}
\item $\nabla \u\in L^2_{\textnormal{loc}}(\mathcal{F})$.
Moreover, for any $\Psi\in C^{1}\big( [0,T] ; C^1_{\rm c}({\mathbb R}^n) \big)$ with
${\mathrm {supp}}(\Psi) \Subset \mathcal{F}$ $($i.e., ${\mathrm {supp}}(\Phi)$ is a compact subset of $\mathcal{F}${\rm )},
\begin{align*}
&\int_{{\mathbb R}^n} (\rho U^i) ({\boldsymbol{x}},t) \Psi({\boldsymbol{x}},t)\,{\mathrm d} {\boldsymbol{x}}
- \int_{{\mathbb R}^n} (\rho_0 U_0^i) ({\boldsymbol{x}}) \Psi({\boldsymbol{x}},0)\,{\mathrm d}{\boldsymbol{x}} \\
&\begin{aligned}
=& \int_{0}^{t}\int_{{\mathbb R}^n} \big( \rho U^{i} \partial_t \Psi + \rho U^{i} (\u \cdot \nabla) \Psi
+ (\gamma-1)\rho e \partial_{x^i} \Psi \big)
\,{\mathrm d}{\boldsymbol{x}} {\mathrm d} s \\
&- \int_{0}^{t}\int_{{\mathbb R}^n}\big( \mu \nabla \Psi \cdot \nabla U^{i}
+ (\mu+\lambda) \partial_{x^i} \Psi\,{\mathrm {div}}\,\u \big)
\,{\mathrm d} {\boldsymbol{x}} {\mathrm d} s \qquad \, \text{ for each $i=1,\dotsc,n$.}
\end{aligned}
\end{align*}
\item $\nabla e\in L^2_{\textnormal{loc}}(\mathcal{F})$.
Moreover, for any $\Phi\in C^{1}\big( [0,T] ; C^1_{\rm c}({\mathbb R}^n) \big)$ with
${\mathrm {supp}}(\Phi) \Subset\mathcal{F}$,
\begin{align*}
&\int_{{\mathbb R}^n} (\rho E)({\boldsymbol{x}},t) \Phi ({\boldsymbol{x}},t)\,{\mathrm d}{\boldsymbol{x}} - \int_{{\mathbb R}^n} (\rho_0 E_0)({\boldsymbol{x}}) \Phi({\boldsymbol{x}},0)\,{\mathrm d} {\boldsymbol{x}} \\
&\begin{aligned}
=& \int_{0}^{t}\int_{{\mathbb R}^n} \big(\rho E \partial_t \Phi
+(\rho E + P ) (\u \cdot \nabla) \Phi \big)
\,{\mathrm d} {\boldsymbol{x}} {\mathrm d} s \\
&-\int_{0}^{t} \int_{{\mathbb R}^n} \Big( \kappa \nabla e + \dfrac{\mu}{2} \nabla \snorm{\u}^2
+ \lambda \u {\mathrm {div}} \u + \mu (\nabla \u )\cdot \u \Big) \cdot \nabla \Phi
\,{\mathrm d}{\boldsymbol{x}} {\mathrm d} s.
\end{aligned}
\end{align*}
\end{enumerate}
\end{definition}
The main theorem provided below states the global-in-time existence and properties of weak solutions
in $\mathbb{R}^n$ for $n\ge 2$.
\begin{theorem}[Main Theorem]\label{thm:WSAV}
Let $(\rho_0,\u_0,e_0)({\boldsymbol{x}})=(\rho_0(r),u_0(r)\frac{{\boldsymbol{x}}}{r},e_0(r)), r\vcentcolon=|{\boldsymbol{x}}|$,
be spherically symmetric initial data defined in $\boldsymbol{x}\in \mathbb{R}^n$ satisfying
\begin{equation}\label{eqs:init}
\begin{split}
&e_0(r)\ge C_*^{-1}, \quad C_*^{-1} \le \rho_0(r) \le C_* \qquad\, \text{ for all $r\in[0,\infty)$,}\\
&\int_{0}^{\infty}\Big(\rho_0\big(\dfrac{1}{2}\snorm{u_0}^2+ \psi(e_0)\big) + (\gamma-1)G(\rho_0)
+\snorm{(\rho_0-1, u_0^2, e_0-1)}^2\Big)(r)\,r^m {\mathrm d} r \le C_*,
\end{split}
\end{equation}
for some given constant $C_*>0$, where $G(\zeta)=1-\zeta+\zeta\log \zeta$ and $\psi(\zeta) = \zeta-1-\log \zeta$.
Then, for each $T>0$, there exists a global spherically symmetric weak solution $(\rho,\u,e)({\boldsymbol{x}},t)$ of the
Cauchy problem \eqref{eqs:FNS}--\eqref{eqs:CauchyInit} in the sense of {\rm Definition \ref{def:WFAV}}
such that the following statements hold\textnormal{:}
\begin{enumerate}[label=(\roman*),ref=(\roman*),font={\normalfont\rmfamily}]
\item\label{item:WSAV1} There exists a continuous increasing function $g\vcentcolon [0,\infty) \to [0,\infty)$ with $g(y)\searrow 0$
as $y\searrow 0$ such that, for each bounded measurable set $E \subset {\mathbb R}^n$,
\begin{align*}
\qquad \esssup\limits_{t\in[0,T]}\int_{{\mathbb R}^n}\big(\frac{1}{2}\rho\snorm{\u}^2 + G(\rho)
\big)({\boldsymbol{x}},t)\,{\mathrm d} {\boldsymbol{x}} \le C(T),
\quad \esssup\limits_{t\in[0,T]}\int_{E}(\rho e)({\boldsymbol{x}},t)\,{\mathrm d} {\boldsymbol{x}}
\le C(T)\big(1 + g(|E|)\big),
\end{align*}
where $|E|$ denotes the Lebesgue measure of $E$ and $C(T)=C(T,C_*,\gamma,\mu,\lambda,\kappa,n)>1$ is a constant depending on $T>0$.
Moreover, for $\eta\in (0,1)$,
\begin{align*}
&\int_{0}^{T} \sup_{\snorm{{\boldsymbol{x}}}\ge\underline{r}(t)+\eta}\dfrac{\snorm{\u}}{\sqrt{e}}({\boldsymbol{x}},t)\,{\mathrm d} t
\le C(T)\big(\eta^{\frac{2-n}{2}} + \eta^{2-n} \big) && \text{if } \ n=2,\, 3,\\
&\int_{0}^{T} \sup_{\snorm{{\boldsymbol{x}}}\ge\underline{r}(t)+\eta} \log \big( \max\big\{ 1, e({\boldsymbol{x}},t)\big\} \big)\, {\mathrm d} t
\le C(T)\big(1+\sqrt{|\log \eta|}\big) && \text{if } \ n=2,\\
&\int_{0}^{T} \sup_{\snorm{{\boldsymbol{x}}}\ge\underline{r}(t)+\eta} \log \big( \max \big\{ 1, e^{\pm 1}({\boldsymbol{x}},t) \big\} \big)\,{\mathrm d} t
\le C(T) \eta^{2-n} && \text{if } \ n=3.
\end{align*}
\smallskip
\item\label{item:WSAV2} There exists a continuous function $(y,t)\mapsto \widetilde{r}(y,t) \vcentcolon [0,\infty)\times[0,T] \to [0,\infty)$
such that $y\mapsto \widetilde{r}(y,t)$ is strictly increasing, and $\underline{r}(t) = \lim_{y\searrow 0} \widetilde{r}(y,t)$ for $t\in[0,T]$. Moreover,
there exists a constant $C_0=C_0(C_*,\gamma,\mu,\lambda,\kappa,n)>0$ such that
\begin{equation*}
\begin{dcases*}
\int_{\{\underline{r}(t) < \snorm{{\boldsymbol{x}}} < \widetilde{r}(y,t) \}} \rho({\boldsymbol{x}},t)\,{\mathrm d} {\boldsymbol{x}} = y & for a.e. $(y,t)\in[0,\infty)\times[0,T]$,\\
\widetilde{r}(y,t) = \widetilde{r}_0(y) + \int_{0}^{t} u(\widetilde{r}(y,s),s)\,{\mathrm d} s & for a.e. $(y,t)\in (0,\infty)\times[0,T]$,\\
n y \psi^{-1}_{-}( \dfrac{C_0}{y}) \le \widetilde{r}^{\,n}(y,t) \le C_0(1 +y) & for all $(y,t)\in(0,\infty)\times[0,T]$,
\end{dcases*}
\end{equation*}
where $u(\snorm{{\boldsymbol{x}}},t)\vcentcolon= \u({\boldsymbol{x}},t)\cdot\frac{{\boldsymbol{x}}}{\snorm{{\boldsymbol{x}}}}$, $\psi^{-1}_{-}$ is
the left inverse of $\psi(\zeta)=\zeta-1-\log \zeta$, and the function $y\mapsto\widetilde{r}_0(y) \vcentcolon [0,\infty)\to [0,\infty)$ is implicitly defined as
\begin{equation*}
y= \int_{0}^{\widetilde{r}_0(y)} \rho_0(r)\, r^m {\mathrm d} r \qquad\text{for each $y\in[0,\infty)$.}
\end{equation*}
\item\label{item:WSAV3}
For $\varepsilon\in(0,1]$, there exists a constant $C(\varepsilon)=C(\varepsilon,T,C_*,\gamma,\mu,\lambda,\kappa,n)>0$ such that
\begin{equation*}
\begin{cases*}
\snorm{\widetilde{r}(y_1,t)-\widetilde{r}(y_2,t)} \le C(\varepsilon) \snorm{y_1-y_2} &\quad for all $(y_1,y_2,t)\in [\varepsilon,\infty)^2\times[0,T]$, \\
\norm{\widetilde{r}(y,t_1)-\widetilde{r}(y,t_2)} \le C(\varepsilon) \snorm{t_1^{\frac{3}{4}}-t_2^{\frac{3}{4}}} &\quad for all $(y,t_1,t_2)\in [\varepsilon,\infty)\times[0,T]^2$.
\end{cases*}
\end{equation*}
Moreover, let $\sigma(t)\vcentcolon=\min\{1,t\}$
and $B^{\rm c}(r)\vcentcolon=\{{\boldsymbol{x}}\in{\mathbb R}^n\vcentcolon \snorm{{\boldsymbol{x}}} \ge r\}$. Then
\begin{equation*}
C^{-1}(\varepsilon)
\le \rho({\boldsymbol{x}},t) \le C(\varepsilon), \,\,\, \snorm{\u({\boldsymbol{x}},t)} \le C(\varepsilon)\sigma^{-\frac{1}{4}}(t),
\,\,\, 0<e({\boldsymbol{x}},t) \le C(\varepsilon)\sigma^{-\frac{1}{2}}(t)
\end{equation*}
for a.e. $({\boldsymbol{x}},t)$ such that $t\in[0,T]$ and ${\boldsymbol{x}} \in B^{\rm c}(\widetilde{r}(\varepsilon,t))$, and
\begin{align*}
&\sup_{ 0\le t \le T} \int_{B^c(\widetilde{r}(\varepsilon,t))} \norm{(\rho-1,\snorm{\u}^2,e-1,\sqrt{\sigma}\nabla \u,\sigma \nabla e)}^2({\boldsymbol{x}},t)\, {\mathrm d} {\boldsymbol{x}}\\
&\,\, +\int_{0}^{T}\int_{B^c(\widetilde{r}(\varepsilon,t))} \norm{(\nabla \u, (\u \cdot \nabla) \u, \nabla e, \sqrt{\sigma}\partial_t \u, \sigma \partial_t e)}^2({\boldsymbol{x}},t)\,{\mathrm d} {\boldsymbol{x}} {\mathrm d} t \le C(\varepsilon).
\end{align*}
In addition, for all $t\in(0,T]$ and ${\boldsymbol{x}},{\boldsymbol{y}} \in B^{\rm c}(\widetilde{r}(\varepsilon,t))$,
\begin{equation*}
\sigma^{\frac{1}{2}}(t)\snorm{\u({\boldsymbol{x}},t)-\u({\boldsymbol{y}},t)} + \sigma(t)\snorm{e({\boldsymbol{x}},t)-e({\boldsymbol{y}},t)} \le C(\varepsilon) \snorm{{\boldsymbol{x}}-{\boldsymbol{y}}}^{\frac{1}{2}},
\end{equation*}
and, for all $0<t_1<t_2\le T$ and $\snorm{{\boldsymbol{x}}} \ge \sup_{t_1\le t \le t_2}\widetilde{r}(\varepsilon,t)$,
\begin{equation*}
\sigma^{\frac{1}{2}}(t_1)\snorm{\u({\boldsymbol{x}},t_1)-\u({\boldsymbol{x}},t_2)} + \sigma(t_1)\snorm{e({\boldsymbol{x}},t_1)-e({\boldsymbol{x}},t_2)} \le C(\varepsilon) \snorm{t_2-t_1}^{\frac{1}{4}}.
\end{equation*}
Furthermore, defining
$\rho^{(\varepsilon)} ({\boldsymbol{x}},t) \vcentcolon = \rho({\boldsymbol{x}},t) \chi_\varepsilon({\boldsymbol{x}},t)$ with
\begin{equation*}
\chi_\varepsilon({\boldsymbol{x}},t) \vcentcolon = \begin{dcases*}
1 & if $t\in[0,T]$ and ${\boldsymbol{x}} \in B^{c}(\widetilde{r}(\varepsilon,t))$,\\
0 & otherwise,
\end{dcases*}
\end{equation*}
then, for all $L\in\mathbb{N}$, $\rho^{(\varepsilon)}({\boldsymbol{x}},t)\in C^0\big([0,T];H^{-1}(B_L)\big)$ and
\begin{equation*}
\big\|\rho^{(\varepsilon)}(\cdot, t_1) - \rho^{(\varepsilon)}(\cdot,t_2)\big\|_{ H^{-1}(B_L)}
\le C(\varepsilon)\norm{t_1-t_2} \qquad \text{ for all $0\le t_1 \le t_2 \le T$,}
\end{equation*}
where $C(\varepsilon)$ is independent of $L\in\mathbb{N}$ as before
and $B_L\vcentcolon=\{ {\boldsymbol{x}}\in{\mathbb R}^n \vcentcolon \snorm{{\boldsymbol{x}}}<L \}$.
\end{enumerate}
\end{theorem}
\begin{remark}\label{rem:moment}
The domain for which the momentum equations hold weakly can be extended to the entire space-time domain.
However, the viscosity term may be present in the weak form as a non-standard limit distribution
of approximate solutions.
More specifically, let $(\rho,\u,e)({\boldsymbol{x}},t)$ be a weak solution with initial data $(\rho_0,\u_0,e_0)({\boldsymbol{x}})$
obtained in {\rm Theorem} \textnormal{\ref{thm:WSAV}}.
Let $\xi \vcentcolon [0,\infty)\to [0,1]$ be a fixed increasing $C^1$--function with
$\xi \equiv0$ on $[0,1]$ and $\xi\equiv1$ on $[2,\infty)$,
and set $\xi^R(r)\vcentcolon= \xi(\frac{r}{R})$.
Then there exists a sequence of vector valued functions
$\{\u_{\alpha}({\boldsymbol{x}},t)\}_{\alpha\in\mathbb{N}}$ such that
$\u_{\alpha}^{i}\in L^2_{\textnormal{loc}}({\mathbb R}^n\times[0,T])$
for each $\alpha\in\mathbb{N}$ and $i=1,\dotsc,n$.
Moreover, for each $i=1,\dotsc, n$ and each $\Psi\in C^2\big([0,T];C^2_{\rm c}({\mathbb R}^n)\big)$
with $\Psi\vcentcolon {\mathbb R}^n\times[0,T] \to \mathbb{R}$, defining
the distribution $\mathbb{M}(\cdot,t) \in \big(C^2\big([0,T];C^2_{\rm c}({\mathbb R}^n)\big)\big)^{\ast}$ by
\begin{align*}
\mathbb{M}^i(\Psi,t) := \lim\limits_{R\searrow 0} \lim\limits_{\alpha\to\infty} \int_{0}^{t}
\int_{\{\snorm{{\boldsymbol{x}}}\ge \alpha^{-1}\}} \big\{(\mu + \lambda) (\u_{\alpha} \cdot \nabla)
\partial_i \Psi^R + \mu \boldsymbol{U}_{\alpha}^i\Delta \Psi^R \big\}\,{\mathrm d} {\boldsymbol{x}} {\mathrm d} s,
\end{align*}
with $\Psi^R({\boldsymbol{x}},t)\vcentcolon= \xi^{R}(\snorm{{\boldsymbol{x}}})\Psi({\boldsymbol{x}},t)$,
then
\begin{equation*}
\begin{aligned}
&\int_{{\mathbb R}^n} (\rho U^i)({\boldsymbol{x}},t) \Psi({\boldsymbol{x}},t)\,{\mathrm d} {\boldsymbol{x}} - \int_{{\mathbb R}^n} (\rho_0 U_0^i)({\boldsymbol{x}}) \Psi({\boldsymbol{x}},0)\,{\mathrm d} {\boldsymbol{x}} \\
&= \int_{0}^{t}\int_{{\mathbb R}^n} \big(\rho U^{i} \partial_t \Psi
+ \rho U^{i} (\u\cdot \nabla) \Psi + (\gamma-1)\rho e \partial_{{\boldsymbol{x}}^i} \Psi\big)({\boldsymbol{x}},s)\,{\mathrm d} {\boldsymbol{x}} {\mathrm d} s
+ \mathbb{M}^i(\Psi,t) \quad\mbox{for each $i=1,\dotsc,n$}.
\end{aligned}
\end{equation*}
In fact, $\u_{\alpha}({\boldsymbol{x}},t)$ is the velocity of solution to the exterior sphere problem, which satisfies the auxiliary boundary condition\textnormal{:}
\begin{equation*}
\u_{\alpha}({\boldsymbol{x}},t) =0, \ \ \nabla e_{\alpha}({\boldsymbol{x}},t) =0
\qquad \text{for $\snorm{{\boldsymbol{x}}}=\alpha^{-1}$ and $t\in [0,T]$.}
\end{equation*}
The existence of such solution $(\rho_{\alpha},\u_{\alpha},e_{\alpha})$ is stated
in {\rm Theorem} \textnormal{\ref{thm:WSCEx}} in {\rm \S}\textnormal{\ref{subsec:eSFNS}}.
The derivation of the above result is obtained by using the uniform entropy estimates on
the approximate solutions $(\rho_\alpha,\u_\alpha,e_{\alpha})_{\alpha\in\mathbb{N}}$
\textnormal{(}\textit{cf}. {\rm Theorem} \textnormal{\ref{thm:WSCEx}\ref{item:WSCEx1};} also see \cite{HJ}\textnormal{)}.
An important followup
is
whether the following equality hold{\rm :}
\begin{equation*}
\mathbb{M}^i(\Psi,t) \overset{?}{=} \int_{0}^{t}
\int_{{\mathbb R}^n} \big\{(\mu + \lambda) {\mathrm {div}} \u
\partial_i \Psi + \mu \nabla \u^i \nabla \Psi \big\}\,{\mathrm d} {\boldsymbol{x}} {\mathrm d} s,
\end{equation*}
which is beyond the scope of this paper and will further be addressed in a separate
paper.
\end{remark}
\section{Main Strategy and Reformulation}\label{sec:reform}
Throughout this paper, we use the following notation for five types of generic constants
that are all larger than $1$ and may be different at different occurrence (without loss of generality)
to indicate their specific
dependence on various parameters:
\begin{equation*}
\begin{aligned}
& C\equiv C(\gamma,\mu,\lambda,\kappa,n), \quad C_0\equiv C_0(C,C_*),
\quad C(T)\equiv C(T,C_0),\\
& C(a)\equiv C(a,T,C_0), \quad C(\varepsilon)\equiv C(\varepsilon,T,C_0) \,
\qquad \text{ for each $a\in(0,1)$ and $\varepsilon\in(0,1]$},
\end{aligned}
\end{equation*}
where $C_*=C_*(\rho_0,u_0,e_0)>1$ depends only on the initial data as
in \eqref{eqs:init} of Theorem \ref{thm:WSAV}.
We also denote $\beta\vcentcolon=2\mu + \lambda$.
\subsection{Main strategy}\label{subsec:MS}
Our approach for proving Theorem \ref{thm:WSAV} is to split the original problem \eqref{eqs:SFNS} into the following two parts:
\smallskip
\begin{enumerate}[label=(\Roman*), ref=(\Roman*)]
\item\label{item:1} For any fixed $a\in(0,1)$ and $T>0$,
we prove the global-in-time existence of a weak solution $(\rho_a,u_a,e_a)(r,t)$
of the initial-boundary value problem (IBVP) of system \eqref{eqs:SFNS}
in the exterior domain $(r,t)\in[a,\infty)\times[0,T]$ with the following initial-boundary conditions:
\begin{equation*}
\begin{cases*}
(\rho_a,u_a,e_a)(r,0) = (\rho_a^0,u_a^0,e_a^0)(r) &\,\, for $r\in[a,\infty)$,\\
u(a,t)=\partial_r e(a,t)=0 &\,\, for $t\in[0,T]$,
\end{cases*}
\end{equation*}
where $(\rho_a^0,u_a^0,e_a^0)$ is the approximate initial data constructed from
the initial data $(\rho_0,u_0,e_0)$ given in Theorem \ref{thm:WSAV}.
The detailed construction can be found in \S \ref{subsec:mollify}, and $(\rho_a^0,u_a^0,e_a^0)$ is shown to satisfy the
initial condition \eqref{eqs:init} with a constant $C_0>0$ independent of $a\in(0,1)$ in Propositions \ref{prop:mollify}.
Notice that
the corresponding solutions $\{(\rho_a,u_a,e_a)(r,t)\}_{a\in(0,1)}$ satisfy certain high-order
estimates uniformly in $a\in(0,1)$ (see Theorem \ref{thm:WSCEx} in \S \ref{subsec:eSFNS}).
\smallskip
\item We employ the uniform estimates, independent of $a$ obtained in \ref{item:1},
and make careful compactness arguments to prove that there is a subsequence $a_j \searrow 0$ as $j\to\infty$ such that
the corresponding limit triple $(\rho,u,e)$ is a global weak solution described
in Theorem \ref{thm:WSAV}. \label{item:2}
\end{enumerate}
\smallskip
Part \ref{item:1} of our approach will be achieved via the following six steps:
\begin{enumerate}[label=(I.\arabic*),ref=(I.\arabic*),font={\normalfont\rmfamily}]
\item\label{item:i} We modify the original initial data $(\rho_0,u_{0},e_0)$ into $(\rho_a^0,u_a^{0},e_{a}^0)(r)$ in the Eulerian coordinates, as constructed in \S \ref{subsec:mollify}.
\item\label{item:ii} The solution, $(\rho_a,u_a,e_a)$, in \ref{item:1} is obtained as the limit of solutions of the approximate problems
posed in bounded annular domains. For this purpose, we apply a truncation procedure on the mollified initial data $(\rho_a^0,u_a^0,e_a^0)$
in the Eulerian coordinates near the far field region ({\it i.e.}, $r\to \infty$), which is parameterized by integer $k\in\mathbb{N}$.
\item\label{item:iii} We transform the IBVP in the exterior domain $(r,t)\in[a,\infty)\times[0,T]$
stated in \ref{item:1} into a set of IBVPs shown in \eqref{eqs:LFNS-k}--\eqref{eqs:LFNS-kb} in the Lagrangian coordinates in the spatial interval $[0,k]$. Indeed, the main point of
the above reduction and modification of the initial data is to ensure that, for each $T>0$, there exists a unique global-in-time smooth solution of the approximate IBVP \eqref{eqs:LFNS-k}--\eqref{eqs:LFNS-kb} in
$(x,t)\in[0,k]\times[0,T]$.
\item\label{item:iv} We derive the entropy inequalities on these solutions. Consequently, we obtain the upper and lower bounds of the density.
These density bounds are independent of the approximation parameter $k\in \mathbb{N}$;
in addition, they are independent of $a\in(0,1)$ on the regions away from the origin in the Lagrangian coordinates.
\item\label{item:v} Based on the bounds of the density obtained in \ref{item:iv}, we seek for a set of high-order estimates
on the regions away from the origin, $x=0$, in the Lagrangian coordinates,
which is independent of the approximate parameter $a\in(0,1)$.
In fact, these estimates can be obtained by introducing a set of $C^1$ cut-off functions $g_\varepsilon(x)$,
for each $\varepsilon\in(0,1)$, away from the origin (see \eqref{eqs:g}--\eqref{eqs:grbd}).
However, the inclusion of these cut-off functions leads to the difficulties in estimating certain
integral terms near the boundary: $x=\varepsilon$, which is described in the proof of Lemma \ref{lemma:eL2}.
In order to overcome these difficulties, we make full use of the dissipation terms in the entropy estimates from Step \ref{item:iv}, which can be found in Lemma \ref{lemma:euLinf}.
\item\label{item:vi} Finally, we denote $(\widetilde{v}_{a,k},\widetilde{u}_{a,k},\widetilde{e}_{a,k})(x,s)$
(with the specific volume $\widetilde{v}_{a,k} := (\widetilde{\rho}_{a,k})^{-1}$) as the approximate solution of problem \eqref{eqs:LFNS-k}--\eqref{eqs:LFNS-kb}
in the Lagrangian domain $(x,s)\in[0,k]\times[0,T]$, and $$
\widetilde{r}_{a,k}(x,s)=\Big(a^n + n \int_{0}^{x}\widetilde{v}_{a,k}(y,s)\,{\mathrm d} y\Big)^{\frac{1}{n}}
$$
as the corresponding particle path function.
For each $(a,k)\in(0,1)\times\mathbb{N}$, we formulate the coordinate transformation $(r,t)=\mathcal{T}_{a,k}(x,s)=(\mathcal{T}_{a,k}^{(1)},\mathcal{T}_{a,k}^{(2)})(x,s)$ as
\begin{equation*}
\mathcal{T}_{a,k}^{(1)}(x,s)=\widetilde{r}_{a,k}(x,s), \quad
\mathcal{T}_{a,k}^{(2)}(x,s)=s \, \qquad \text{ for $(x,s)\in [0,k]\times[0,T]$.}
\end{equation*}
Owing to the density bounds from Step \ref{item:iv}, the inverse function $\mathcal{T}_{a,k}^{-1}$ exists.
By these coordinate transformations, we define
\begin{equation*}
(\overline{v}_{a,k},\overline{u}_{a,k},\overline{e}_{a,k})(r,t) \vcentcolon=( \widetilde{v}_{a,k},\widetilde{u}_{a,k},\widetilde{e}_{a,k}) ( \mathcal{T}_{a,k}^{-1}(r,t)) \
\qquad \text{for $(r,t)\in R_{a,k}$},
\end{equation*}
where $R_{a,k} \vcentcolon= \mathcal{T}_{a,k}\big([0,k]\times [0,T] \big) = \big\{ (r,t)\,\vcentcolon\, t\in[0,T], r\in [a,\widetilde{r}_{a,k}(k,t)] \big\}$.
We then extend these functions into the entire exterior Eulerian domain $(r,t)\in[a,\infty)\times[0,T]$ by introducing the following test functions:
Let $\xi \vcentcolon {\mathbb R} \to [0,1]$ be a function with $\xi\in C^{\infty}$ such that $\xi(\zeta)=1$ if $\zeta\le 0$, $\xi(\zeta)=0$
if $\zeta\ge 1$, and $\snorm{\xi^{\prime}(\zeta)} \le 2$ for all $\zeta\in{\mathbb R}$.
With this, we define the following cut-off functions:
\begin{equation*}
\varphi_{a,k}(r,t) = \xi( \dfrac{2r-\widetilde{r}_{a,k}(k,t)}{\widetilde{r}_{a,k}(k,t)}).
\end{equation*}
It is shown in \S \ref{subsec:Eulext} that ${\mathrm {supp}}(\varphi_{a,k})\subseteq R_{a,k}$
so that the approximate solutions in the Eulerian coordinates can be extended
to domain $[a,\infty)\times [0,T]$:
\begin{equation}\label{eqs:extfunc}
\qquad\,\, (\rho_{a,k},u_{a,k},e_{a,k})(r,t)=(1,0,1)+\begin{dcases}
\varphi_{a,k}({\overline{v}}^{-1}_{a,k} - 1,
\overline{u}_{a,k},
\overline{e}_{a,k}- 1)(r,t)\quad\, \text{for $(r,t)\in R_{a,k}$,} \\
(0,0,0)\qquad\qquad\,\,\quad\text{for $(r,t)\in [a,\infty)\times[0,T]\backslash R_{a,k}$,}
\end{dcases}
\end{equation}
With
the uniform {\it a-priori} estimates derived in Steps \ref{item:iv}--\ref{item:v},
there exits
a subsequence $(\rho_{a,k_j},u_{a,k_j},e_{a,k_j})_{j\in\mathbb{N}}$ such that
its limit function $(\rho_a,u_a,e_a)$ is a weak solution of the exterior problem
stated in \ref{item:1}. Furthermore, it can also be shown that such solutions continue to enjoy
the uniform regularity estimates in $a\in(0,1)$ attained in Steps \ref{item:iv}--\ref{item:v}.
The detailed analysis of this procedure can be found in \S \ref{sec:WSCEx}.
\end{enumerate}
\medskip
Part \ref{item:2} of our approach can be achieved via the following four steps:
\begin{enumerate}[label=(II.\arabic*),ref=(II.\arabic*),font={\normalfont\rmfamily}]
\item\label{item:two1} For given approximate initial data $(\rho_a^0,u_a^0,e_a^0)$ in Part \ref{item:1}, denote
\begin{equation*}
\big\{ (\rho_a, u_a, e_a)(r,t)\vcentcolon [a,\infty) \times [0,T] \to [0,\infty)\times{\mathbb R}\times[0,\infty) \big\}_{a\in(0,1)}
\end{equation*}
as the global weak solutions of the exterior problem
obtained in Part \ref{item:1}. Based on the entropy estimate, we prove the uniform integrability of the approximate
solutions $(\rho_a,u_a,e_a)_{a\in(0,1)}$ with respect to the approximate parameter $a\in(0,1)$ in \S \ref{subsec:meas},
which plays key roles in proving that the limit vector function of these approximate solutions
are indeed a weak solution of the original problem (see Lemmas \ref{lemma:alimRhoWF}--\ref{lemma:eWF}).
\item\label{item:two2} The particle path functions $\widetilde{r}_a(x,t)\vcentcolon[0,\infty]\times[0,T]\to [a,\infty)$
associated with the approximate density $\rho_a$ for each $a\in(0,1)$ are defined to be the functions satisfying
\begin{equation*}
x = \int_{a}^{\widetilde{r}_{a}(x,t)} \rho_{a}(r,t)\,r^m {\mathrm d} r \qquad\,\, \text{for {\it a.e.} $(x,t)\in[0,\infty)\times[0,T]$,}
\end{equation*}
which can be found in Lemma \ref{lemma:rkcvg} in \S \ref{subsec:klimpath}.
Then, by the estimates obtained in Step \ref{item:two1} and compactness arguments,
we show in \S \ref{subsec:alimpath} that there exists a function $\widetilde{r}(x,t)\vcentcolon [0,\infty)\times[0,T]\to [0,\infty)$ that
is the limit function of sequence $\{\widetilde{r}_a(x,t)\}_{a\in(0,1)}$ when $a \searrow 0$.
Moreover, we show that, for $t\in[0,T]$, $\lim_{x\searrow0}\widetilde{r}(x,t)$ does not necessarily equal to zero,
but a function of $t$ instead in general: $\underline{r}(t)\vcentcolon [0,T]\to [0,\infty)$.
In fact, $\widetilde{r}(x,t)$ is the particle path of our original problem \eqref{eqs:SFNS},
and $\underline{r}(t)$ is the vacuum interface in \eqref{eqs:SFNS} that separates the fluid region
and the possible vacuum, which can be verified later.
Furthermore, it is shown in Lemma \ref{lemma:alimpath}\ref{item:apath3} that,
for each $x>0$, there exists $\delta(x)>0$ depending only on $x>0$
such that
\begin{equation}\label{eqs:rlower}
\inf_{t\in[0,T]}\widetilde{r}(x,t)\ge \delta(x)>0.
\end{equation}
This
is crucial for the next step where the limit, $a\searrow0$, is taken in $\{(\rho_a,u_a,e_a)\}_{a\in(0,1)}$.
\smallskip
\item\label{item:two3} Next, for each $\varepsilon\in(0,1]$, we fix a domain away from the origin by
\begin{equation}\label{eqs:Fve}
\mathcal{F}_\varepsilon\vcentcolon=\big\{(r,t)\in[0,\infty)\times[0,T]\,\vcentcolon\,r\in[\widetilde{r}(\varepsilon,t),\infty)\big\},
\end{equation}
where $\widetilde{r}(x,t)$ is obtained in Step \ref{item:two2}.
Moreover, we set
$$
\mathcal{F}\vcentcolon=\{(r,t)\in[0,\infty)\times[0,T]\,\vcentcolon\, r> \underline{r}(t)\}.
$$
By the high-order estimates derived in Step \ref{item:v}, a limit function $(\rho,u,e)(r,t)\vcentcolon \mathcal{F}\to [0,\infty)\times{\mathbb R} \times[0,\infty)$ is constructed in \S \ref{subsec:alim}
as follows: By an application of the Arzel\'a-Ascoli theorem, we first obtain $(\rho_\varepsilon,u_\varepsilon,e_\varepsilon)(r,t)$ defined in domain $\mathcal{F}_\varepsilon$ for each $\varepsilon\in(0,1]$,
then construct the vector function $(\rho,u,e)$ by extending $(\rho_\varepsilon,u_\varepsilon,e_\varepsilon)$ with
\begin{equation*}
(\rho_\varepsilon,u_\varepsilon,e_\varepsilon)(r,t) \vcentcolon
= \begin{dcases*}
(0,0,0) & if $(r,t)\in \mathcal{F}\backslash \mathcal{F}_\varepsilon$,\\
(\rho_\varepsilon,u_\varepsilon,e_\varepsilon)(r,t) & if $(r,t)\in\mathcal{F}_\varepsilon$,
\end{dcases*}
\end{equation*}
and finally glue $\big\{(\rho_\varepsilon,u_\varepsilon,e_\varepsilon)(r,t)\big\}_{\varepsilon\in(0,1]}$ together
over parameter $\varepsilon\in(0,1]$ by the telescoping sum:
\begin{equation*}
f = f_{\varepsilon_1} + \sum_{i=1}^{\infty} (f_{\varepsilon_{i+1}} - f_{\varepsilon_i}),
\end{equation*}
for $f= \rho, u, e$, respectively, and $\varepsilon_i\equiv \frac{1}{i}$.
The detail of this procedure is given in Propositions \ref{prop:aaExt}--\ref{prop:scHm1} in Appendix \ref{subsec:AAV}.
In addition, it is also shown in Lemma \ref{lemma:alimrho} that
\begin{equation*}
x = \int_{\underline{r}(t)}^{\widetilde{r}(x,t)} \rho(r,t)\,r^m {\mathrm d} r
\qquad \text{for {\it a.e.} $(x,t)\in[0,\infty)\times[0,T]$,}
\end{equation*}
which justifies that $\underline{r}(t)$ obtained in Step \ref{item:two2} is the vacuum interface curve of our original problem \eqref{eqs:SFNS}.
\item\label{item:two4} Finally, by the estimates obtained in Step \ref{item:two1} and the convergences in Steps \ref{item:two2}--\ref{item:two3},
we prove the weak forms of solutions listed in Theorem \ref{thm:WSAV}.
The weak forms of solutions in the Eulerian coordinates $({\boldsymbol{x}},t)\in {\mathbb R}^n\times[0,T]$ are shown in \S \ref{subsec:alimWF}.
\end{enumerate}
\begin{remark}
Before proceeding, we first give the following remarks{\rm :}
\begin{enumerate}[label=(\alph*), font=\normalfont]
\item In Steps {\normalfont \ref{item:i}--\ref{item:ii}}, it is important to show that the
modified initial data sequence satisfies the entropy estimates uniformly with respect
to {\normalfont $(a,k)\in(0,1)\times\mathbb{N}$} and strongly
converges to the original initial data {\normalfont$(\rho_0,u_{0},e_0)(r)$}
in proper functional spaces. These results will be needed later in the compactness
arguments in Step {\normalfont \ref{item:vi}}, which is detailed in {\rm \S \ref{sec:WSCEx}}.
\item The main purpose of the reformulation from the Eulerian to the Lagrangian coordinates
in Step {\normalfont\ref{item:iii}} is to obtain the crucial upper and lower bounds of
the density $($see statement {\normalfont\ref{item:prio1}} of {\rm Theorem \ref{thm:priori}}
in {\rm \S \ref{sec:rhobd})}, which are independent of $k\in\mathbb{N}$.
Note that these bounds of the density are an improvement on the result of Jiang \cite{J}.
\item As mentioned in Step {\normalfont \ref{item:v}}, the introduction of
{\normalfont$C^1$} cut-off functions away from the origin leads to several hurdles in the estimates,
and these issues have also been encountered in Hoff-Jenssen \cite{HJ}.
In this paper, we resolve these issues by making use of the dissipation terms that are present
in the entropy estimate. This is detailed in {\rm Lemma \ref{lemma:euLinf}}
in {\rm \S \ref{subsec:L2total}}.
\item The purpose of the coordinate transformation back into the Eulerian domain,
introduced in Step {\normalfont\ref{item:vi}} is due to the following consideration{\rm :}
As {\normalfont $k\to\infty$}, a difficulty arises due to the nonlinear second-order spatial
differential operators in the Lagrangian forms of the momentum and internal energy equations.
Since these operators are linear in the Eulerian coordinates, this difficulty is resolved by
converting the approximate solutions, originally obtained in the Lagrangian coordinates,
into the functions
in the Eulerian domain. This allows us to take limit {\normalfont $k\to\infty$} and
obtain the desired weak forms in the Eulerian coordinates.
\item In Step {\normalfont\ref{item:vi}}, after taking limit {\normalfont $k\to\infty$},
it requires to show that the limit solution still satisfies the entropy inequality,
since it will be required in Step {\normalfont\ref{item:two1}}.
This is achieved by an application of Mazur's lemma $(${\rm Lemma \ref{lemma:Ent}}$)$
in {\rm \S \ref{subsec:klimEnt}} and {\rm Proposition \ref{prop:mazur}}.
\item In order to apply the compactness arguments from the high-order estimates obtained in {\normalfont\ref{item:v}},
it is crucial that $\mathcal{F}_{\varepsilon}$ defined in \eqref{eqs:Fve} is a set strictly bounded away from the
origin, $r=0$, for each $\varepsilon>0$.
We obtain this by assertion \eqref{eqs:rlower} in Step {\normalfont\ref{item:two2}}.
\end{enumerate}
\end{remark}
The rest of this section is organized as follows: In \S \ref{subsec:eSFNS}, we show our main existence
results for
weak solutions with large data of the exterior problem mentioned in Part \ref{item:1}. In \S \ref{subsec:mollify}-- \S \ref{subsec:ffaprox},
we describe the detailed procedures for the modification and truncation of the initial data listed in Steps \ref{item:ii}--\ref{item:iii} above.
Finally, in \S \ref{subsec:lageqs}, we reformulate the exterior problem in the Eulerian coordinates into the approximate Lagrangian problems
in domain $(x,t)\in[0,k]\times[0,T]$ as described in Step \ref{item:iv},
and establish the corresponding global-in-time well-posedness of the unique strong solution with regular initial data.
\subsection{Mollification of the initial data in the Eulerian domain}\label{subsec:mollify}
Let $(\rho_0,u_0,e_0)(r)$, $r\in[0,\infty)$, be the initial data in Theorem \ref{thm:WSAV}.
First, we extend the initial data to
$\mathbb{R}$
by
\begin{equation}
(\hat{\rho}_0,\hat{u}_0,\hat{e}_0)(\zeta) \vcentcolon=\begin{dcases*}
(\rho_0(\zeta),u_0(\zeta),e_0(\zeta)) & if $\zeta\in[0,\infty)$,\\[1mm]
(\rho_0(-\zeta),-u_0(-\zeta),e_0(-\zeta)) & if $\zeta\in(-\infty,0)$.
\end{dcases*}
\end{equation}
Then, for each $a\in(0,1)$, let $j_{a}\in C_{\rm c}^{\infty}([-1,1])$ be
the standard 1-D mollifiers such that ${\mathrm {supp}}(j_a)\subseteq [-\frac{a}{2},\frac{a}{2}]$. We define
$$
(\hat{\rho}_{a}^0,\hat{u}_{a}^0,\hat{e}_{a}^0) \vcentcolon= ( \hat{\rho}_0 \ast j_{a}, \hat{u}_0\ast j_{a}, \hat{e}_0\ast j_{a}).
$$
It follows that, for some $C_*>0$, independent of $a\in (0,1)$,
\begin{equation}\label{prftemp0}
\begin{split}
&(\hat{\rho}_{a}^0,\hat{u}_{a}^0,\hat{e}_{a}^0 )\in C^{\infty}({\mathbb R}),\\
&\hat{e}_{a}^0(\zeta)\ge C_*^{-1}, \,\,\,\,
C_*^{-1}\le \hat{\rho}_{a}^0(\zeta) \le C_* \qquad\
\text{for {\it a.e.} $\zeta\in{\mathbb R}$,}\\
&(\hat{\rho}_{a}^0,\hat{u}_{a}^0,\hat{e}_{a}^0)(\zeta) \to (\rho_0, u_0, e_0)(r)
\qquad\quad\ \text{as $a\searrow 0$ for {\it a.e.} $\zeta\in(0,\infty)$.}
\end{split}
\end{equation}
Next, let $\chi\vcentcolon[0,\infty)\to [0,1]$ be such that
$\chi\in C^{\infty}$, $\chi(\zeta)=0$ if $\zeta\le1$, and $\chi(\zeta)=1$ if $\zeta\ge 2$.
Set the cut-off functions:
$$
\chi_a(r)=\chi(\frac{r}{a}) \qquad\mbox{for each $a\in(0,1)$}.
$$
It follows that $\chi_a(r)=0$ for $r\le a$ and $\chi_a(r)=1$ for $r\ge 2a$.
With this, we define the approximate exterior initial data:
\begin{equation}\label{eqs:amolint}
(\rho_a^0, u_a^0, e_a^0)(r)
\vcentcolon= \big( (\hat{\rho}_a^0-1)\chi_a +1,\, \hat{u}_a^0\chi_a,\,
(\hat{e}_a^0-1)\chi_a+1 \big)(r) \qquad \text{for all $r\in (0,\infty)$.}
\end{equation}
\begin{proposition}\label{prop:mollify}
The approximate exterior initial data $(\rho_a^0, u_a^0, e_a^0)(r)$ satisfy
\begin{enumerate}[label=(\roman*), ref=(\roman*), font={\textnormal}]
\item\label{item:mollify1} $(\rho_{a}^0, u_{a}^0, e_{a}^0)\in C^{\infty}([0,\infty))$
and $u_{a}^0(a)= \partial_r e_{a}^0(a)=0$.
\item\label{item:mollify2} As $a\searrow 0$,
\begin{equation}\label{temp:mollify1}
\begin{split}
&\sbnorm{(\rho_{a}^0-1,\snorm{ u_{a}^0}^2, e_{a}^0-1) - (\rho_0-1,\snorm{u_0}^2, e_0-1) }_{L^2([0,\infty),r^m {\mathrm d} r)} \to 0,\\
&(\rho_{a}^0, u_{a}^0, e_{a}^0)(r) \to ( \rho_0, u_0, e_0 )(r) \qquad\ \text{for a.e. $r\in[0,\infty)$.}
\end{split}
\end{equation}
\item\label{item:mollify3} There exists some constant $C_0>0$ independent of $a\in (0,1)$ such that
\begin{equation}\label{temp:mollify2}
\begin{split}
&\,\, e_a^0(r)\ge C_0^{-1}, \ \ \, C_0^{-1}\le \rho_{a}^0(r) \le C_0 \qquad\
\text{ for all $(a,r)\in(0,1)\times [0,\infty)$,}\\
&\sup\limits_{a\in(0,1)}\int_{a}^{\infty}\Big(\rho_{a}^0\big(\frac{1}{2}\snorm{u_{a}^0}^2 + \psi(e_{a}^0)\big)
+G(\rho_a^0) + \snorm{(\rho_a^0-1, |u_a^0|^2, e_a^0-1)}^2 \Big)(r)\,r^m {\mathrm d} r \le C_0.
\end{split}
\end{equation}
\end{enumerate}
\end{proposition}
\begin{proof}
Statements \ref{item:mollify1} and
$C_*^{-1}\le \rho_{a}^0(r) \le C_*$
for all $r\in[0,\infty)$ follow immediately from construction \eqref{eqs:amolint},
while the almost everywhere convergence $\eqref{temp:mollify1}_2$ follows from \eqref{prftemp0}.
Next, by the Minkowski
inequality and \eqref{eqs:amolint}, we have
\begin{align*}
\sbnorm{u_a^0}_{L^4([0,\infty),r^m {\mathrm d} r)}
\le \int_{-\frac{a}{2}}^{\frac{a}{2}} \Big( \int_{0}^{\infty} \snorm{ \chi_a(r) \hat{u}_0(r-\zeta) j_a(\zeta) }^4
\,r^m {\mathrm d} r \Big)^{\frac{1}{4}} {\mathrm d} \zeta
\le \int_{-\frac{a}{2}}^{\frac{a}{2}} j_a(\zeta) \Big(\int_{a}^{\infty}
\snorm{u_0(r-\zeta)}^4\,r^m {\mathrm d} r\Big)^{\frac{1}{4}} {\mathrm d} \zeta.
\end{align*}
Let $\xi\equiv r- \zeta$.
Then $0 \le 1 + \frac{\zeta}{\xi} \le 2$ for all $(\zeta,\xi)\in [-\frac{a}{2},\frac{a}{2}]\times [\frac{a}{2},\infty)$.
It follows that
\begin{align*}
\sbnorm{u_a^0}_{L^4([0,\infty),r^m {\mathrm d} r)}
\le \int_{-\frac{a}{2}}^{\frac{a}{2}} j_a(\zeta)
\Big( \int_{a-\zeta}^{\infty} \snorm{u_0(\xi)}^4 \xi^m ( 1+ \frac{\zeta}{\xi})^m {\mathrm d} \xi \Big)^{\frac{1}{4}}
{\mathrm d} \zeta
\le 2^{\frac{m}{4}}C_* \int_{-\frac{a}{2}}^{\frac{a}{2}} j_a(\zeta) {\mathrm d} \zeta = 2^{\frac{m}{4}}C_*.
\end{align*}
Applying the same argument to $(\rho_a^0-1,e_a^0-1)$ yields the third term of $\eqref{temp:mollify2}_2$.
Next, by the H\"older inequality,
\begin{align*}
\sbnorm{\snorm{u_a^0}^2-\snorm{u_0}^2}_{L^2([0,\infty),r^m{\mathrm d} r)}\le (1 + 2^{\frac{m}{4}}) C_*
\sbnorm{u_a^0-u_0}_{L^4([0,\infty),r^m {\mathrm d} r)}.
\end{align*}
From this, we now show $\sbnorm{u_a^0-u_0}_{L^4([0,\infty),r^m {\mathrm d} r)}\to 0$ as $a\searrow 0$.
By construction \eqref{eqs:amolint},
\begin{align*}
\sbnorm{u_a^0-u_0}_{L^4([0,\infty),r^m {\mathrm d} r)}
&\le \sbnorm{\chi_a(\hat{u}_a^0-u_0)}_{L^4([0,\infty),r^m {\mathrm d} r)}
+ \Big( \int_{0}^{2a} (1-\chi_a)^4 \snorm{u_0}^4(r)\,r^m {\mathrm d} r \Big)^{\frac{1}{4}}.
\end{align*}
Using the fact that $\sbnorm{u_0}_{L^4([0,\infty),r^m {\mathrm d} r)}\le C_*$, $\chi_a \to 1$ as $a\searrow 0$ almost everywhere,
$0\le \chi_a\le 1$, and the dominated convergence theorem, we have
\begin{align*}
\lim\limits_{a\searrow 0}\int_{0}^{2a}
(1-\chi_a)^4 \snorm{u_0}^4(r)\,r^m {\mathrm d} r =0.
\end{align*}
Thus,
we conclude that
\begin{align*}
\lim\limits_{a\searrow 0}\sbnorm{\snorm{u_a^0}^2-\snorm{u_0}^2}_{L^2([0,\infty),r^m{\mathrm d} r)}
\le ( 1 + 2^{\frac{m}{4}} ) C_*
\lim\limits_{a\searrow 0}\sbnorm{\chi_a(\hat{u}_a^0-u_0)}_{L^4([0,\infty),r^m {\mathrm d} r)}.
\end{align*}
Therefore, it is left to prove the limit:
$\sbnorm{\chi_a(\hat{u}_a^0-u_0)}_{L^4([0,\infty),r^m {\mathrm d} r)} \to 0$ as $a\searrow 0$.
To show this, it follows from the Minkowski
inequality that
\begin{align*}
\sbnorm{\chi_a(\hat{u}_a^0-u_0)}_{L^4([0,\infty),r^m {\mathrm d} r)}
&\le \int_{-\frac{a}{2}}^{\frac{a}{2}} j_a(\zeta)
\Big( \int_{a}^{\infty} \snorm{u_0(r-\zeta)-u_0(r)}^4\, r^m {\mathrm d} r \Big)^{\frac{1}{4}} {\mathrm d} \zeta\\
&\le 2^{\frac{m}{4}}\int_{-\frac{a}{2}}^{\frac{a}{2}} j_a(\zeta)
\Big( \int_{a-\zeta}^{\infty} \snorm{u_0(\xi)-u_0(\xi+\zeta)}^4\,\xi^m {\mathrm d} \xi \Big)^{\frac{1}{4}}{\mathrm d}\zeta,
\end{align*}
where we have used that $0 \le 1 + \frac{\zeta}{\xi}\le 2$
if $\zeta\in[-\frac{a}{2},\frac{a}{2}]$ and $\xi\ge a-\zeta$.
Now, by the separability of
space
$L^4( [0,\infty),r^m {\mathrm d} r )$, for each $\delta>0$, there exists $h_\delta\in C_{\rm c}^{\infty}([0,\infty))$ such that
$\sbnorm{h_\delta-\hat{u}_0}_{L^4([0,\infty),r^m {\mathrm d} r)}\le 3^{-\frac{m}{4}} \frac{\delta}{4}$.
By the triangle inequality,
it follows that
\begin{align*}
2^{-\frac{m}{4}}\sbnorm{\chi_a(\hat{u}_a^0-u_0)}_{L^4([0,\infty),r^m {\mathrm d} r)}
&\le \int_{-\frac{a}{2}}^{\frac{a}{2}}j_a(\zeta) \Big( \int_{\frac{a}{2}}^{\infty} \snorm{u_0(\xi)-h_\delta(\xi)}^4
\,\xi^m {\mathrm d} \xi \Big)^{\frac{1}{4}} {\mathrm d} \zeta\\
&\quad + \int_{-\frac{a}{2}}^{\frac{a}{2}}j_a(\zeta) \Big( \int_{\frac{a}{2}}^{\infty}
\snorm{h_\delta(\xi)-h_\delta(\xi+\zeta)}^4\,\xi^m {\mathrm d} \xi \Big)^{\frac{1}{4}} {\mathrm d} \zeta \\
&\quad + \int_{-\frac{a}{2}}^{\frac{a}{2}} j_a(\zeta) \Big( \int_{a-\zeta}^{\infty} \snorm{h_\delta(\xi+\zeta)-u_0(\xi+\zeta)}^4\,\xi^m {\mathrm d} \xi \Big)^{\frac{1}{4}} {\mathrm d} \zeta\\
&=\vcentcolon \sum_{i=1}^3 I_i.
\end{align*}
First, $I_1$ and $I_3$ are estimated as:
\begin{equation*}
\begin{aligned}
I_1\vcentcolon
=& \int_{-\frac{a}{2}}^{\frac{a}{2}} j_a(\zeta) \Big( \int_{\frac{a}{2}}^{\infty}
\snorm{u_0(\xi)-h_\delta(\xi)}^4\,\xi^m {\mathrm d} \xi \Big)^{\frac{1}{4}} {\mathrm d} \zeta
\le \dfrac{\delta}{4} \int_{-\frac{a}{2}}^{\frac{a}{2}} j_a(\zeta)\,{\mathrm d} \zeta = \dfrac{\delta}{4},\\
I_3
\le&\, 3^{\frac{m}{4}} \int_{-\frac{a}{2}}^{\frac{a}{2}} j_a(\zeta)
\Big( \int_{a}^{\infty} \snorm{h_\delta(r)-u_0(r)}^4\,r^m {\mathrm d} r \Big)^{\frac{1}{4}} {\mathrm d} \zeta
\le \dfrac{\delta}{4} \int_{-\frac{a}{2}}^{\frac{a}{2}} j_a(\zeta)\,{\mathrm d} \zeta = \dfrac{\delta}{4},
\end{aligned}
\end{equation*}
where we have used the fact that $\frac{1}{2} \le 1 - \frac{\zeta}{r} \le \frac{3}{2}$
for all $(\zeta,r)\in [-\frac{a}{2},\frac{a}{2}]\times [a,\infty)$.
Since $h_\delta$ is compactly supported for each $\delta>0$,
there exists $a_{\delta}\in(0,1)$ such that
\begin{align*}
I_2 \vcentcolon= \int_{-\frac{a}{2}}^{\frac{a}{2}} j_a(\zeta)
\Big( \int_{\frac{a}{2}}^{\infty} \snorm{h_\delta(\xi)-h_\delta(\xi+\zeta)}^4\,\xi^m {\mathrm d} \xi \Big)^{\frac{1}{4}} {\mathrm d} \zeta
\le \dfrac{\delta}{2}\int_{-\frac{a}{2}}^{\frac{a}{2}} j_a\,{\mathrm d} \zeta
= \dfrac{\delta}{2}
\qquad \text{ if $a\in(0,a_{\delta})$.}
\end{align*}
Thus, for given $\delta>0$, there exists $a_\delta\in(0,1)$ such that $2^{-\frac{m}{4}}\sbnorm{\chi_a(\hat{u}_a^0-u_0)}_{L^4([0,\infty),r^m {\mathrm d} r)} \le\delta$ for all $a\in(0,a_\delta)$.
This implies that
$\lim_{a\searrow 0}\sbnorm{\snorm{u_a^0}^2-\snorm{u_0}^2}_{L^2([0,\infty),r^m{\mathrm d} r)}=0$.
The proof for $\rho_a^0$ and $e_a^0$ follows from the same argument.
Next, by the Jensen inequality and the fact
that $G(\zeta), \psi(\zeta)$, and $\snorm{\zeta}^2$ are convex functions of $\zeta$,
it follows that
$G(\rho_{a}^0)\le G(\hat{\rho}_0) \ast j_{a}$, $\psi(e_{a}^0)\le \psi (\hat{e}_0) \ast j_{a}$,
and $\snorm{u_{a}^0} \le \snorm{\hat{u}_0}^2\ast j_{a}$ for $r\in(0,\infty)$.
Using Fubini-Tonelli's theorem, it follows that
\begin{align*}
\int_{a}^{\infty} G(\rho_a^0)(r)\,r^m {\mathrm d} r
= \int_{-\frac{a}{2}}^{\frac{a}{2}} \Big(\int_{a-\zeta}^{\infty} G(\rho_0)(\xi) (\zeta+\xi)^m {\mathrm d} \xi\Big)
j_{a}(\zeta)\, {\mathrm d} \zeta
\le 2^m \int_{-\frac{a}{2}}^{\frac{a}{2}} C_* j_{a}(\zeta)\, {\mathrm d} \zeta = 2^m C_*,
\end{align*}
where we have used the initial condition $\eqref{eqs:init}_2$. In a similar way, we have
\begin{align*}
\int_{a}^{\infty} \rho_{a}^0 \psi(e_{a}^0) (r) r^m {\mathrm d} r
\le 2^m C_* \int_{-\frac{a}{2}}^{\frac{a}{2}}
\Big(\int_{0}^{\infty} \rho_0\psi(e_0) (\xi) \xi^m {\mathrm d} \xi\Big) j_a(\zeta) \,{\mathrm d} \zeta \le 2^m C_*^3.
\end{align*}
Similarly, we also obtain
$$
\sbnorm{\rho_a^0 \snorm{u_a^0}^2}_{L^1([0,\infty),r^m{\mathrm d} r)}
\le 2^m C_*^3.
$$
Setting $C_0= 2^m C_*^3$,
we complete the proof.
\end{proof}
\subsection{The approximation problem in the exterior Eulerian domain}\label{subsec:eSFNS}
To resolve the coordinate singularity at the origin, which poses a significant difficulty
for the existence of solutions, we first consider the approximation problem
in the exterior domain: $\Omega_{a,T}\vcentcolon=[a,\infty)\times[0,T]$ for fixed $a\in(0,1)$,
and find a solution $(\rho_a,u_a,e_a)$ of the exterior problem for system \eqref{eqs:SFNS} over
domain $\Omega_{a,T}$ with the following initial-boundary conditions:
\begin{equation}\label{eqs:eSFNS}
\begin{cases}
u_a(a,t)=0, \quad \partial_r e_a(a,t)=0 \qquad & \text{for $t\in[0,T]$},\\[1mm]
(\rho_a,u_a,e_a)(r,0)=(\rho_a^0,u_a^0,e_a^0)(r) \qquad & \text{for $r\in[a,\infty)$}.
\end{cases}
\end{equation}
Now the boundary conditions in \eqref{eqs:eSFNS}
can be interpreted physically as the presence of an insulating ball at the origin,
with no slip-boundary condition for the velocity field and gradient field of the internal energy. Moreover,
the initial data
are given by $(\rho_a^0,u_a^{0},e_{a}^0)(r)$ constructed in \S \ref{subsec:mollify} above.
To prove Theorem \ref{thm:WSAV}, the existence of weak solutions of
the exterior problem \eqref{eqs:SFNS} and \eqref{eqs:eSFNS} for each $a\in(0,1)$ is first obtained
so that some compactness arguments
can be applied when $a\searrow 0$.
\begin{definition}[Spherically Symmetric Weak Solutions in the Exterior Domain]\label{def:SWeak}
Let
\begin{align*}
\mathcal{D}^a\vcentcolon
=&\big\{ \phi\in C^{\infty}\big([a,\infty)\times[0,T]\big)\,\vcentcolon\, \exists N>a\, \ \text{such that} \ \phi(r,t)=0 \
\text{for $r\ge N$}\big\},\\[1mm]
\mathcal{D}_0^a \vcentcolon
=&\big\{ \phi\in \mathcal{D}^a\,\vcentcolon\, \phi(a,t)=0 \ \text{ for all $t\in[0,T]$} \big\}.
\end{align*}
Then $(\rho_a,u_a,e_a)(r,t)$ is a weak solution of problem \eqref{eqs:SFNS} and \eqref{eqs:eSFNS},
provided that
\begin{enumerate}[label=(\roman*),ref=(\roman*),font={\normalfont\rmfamily}]
\item\label{item:SWeak1} For each $\phi\in \mathcal{D}^a$,
\begin{align*}
&\int_{a}^{\infty} \rho_a(r,t) \phi(r,t)\,r^m {\mathrm d} r
- \int_{a}^{\infty} \rho_a^0(r) \phi(r,0)\,r^m {\mathrm d} r
=\int_{0}^{t}\int_{a}^{\infty} \big(\rho_a \partial_t \phi + \rho_a u_a \partial_r \phi\big)\,r^m {\mathrm d} r {\mathrm d} s.
\end{align*}
\item\label{item:SWeak2} For each $\phi\in \mathcal{D}^a_0$,
\begin{align*}
&\int_{a}^{\infty} (\rho_a u_a)\phi(r,t)\,r^m {\mathrm d} r
- \int_{a}^{\infty} (\rho_a^0 u_a^0)(r) \phi(r,0)\,r^m {\mathrm d} r
- \int_{0}^{t}\int_{a}^{\infty} \rho_a u_a\big(\partial_t \phi + u_a \partial_r \phi\big)\,r^m {\mathrm d} r {\mathrm d} s \\
&= \int_{0}^{t}\int_{a}^{\infty} \Big( P_a - \beta \big( \partial_r u_a + m \dfrac{u_a}{r}\big)\Big) \big( \partial_r \phi + m\dfrac{\phi}{r}\big)\,r^m {\mathrm d} r {\mathrm d} s,
\end{align*}
where $\beta\vcentcolon= 2\mu + \lambda$.
\smallskip
\item\label{item:SWeak3} For each $\phi\in \mathcal{D}^a$,
\begin{align*}
&\int_{a}^{\infty} (\rho_a E_a)(r,t) \phi (r,t)\,r^m {\mathrm d} r
- \int_{a}^{\infty} (\rho_a^0 E_a^0)(r) \phi(r,0)\,r^m {\mathrm d} r\\
&= \int_{0}^{t}\int_{a}^{\infty} \big\{ \rho_a E_a \partial_t \phi +(\rho_a E_a + P_a ) u_a \partial_r \phi \big\}
\,r^m {\mathrm d} r {\mathrm d} s\\
&\quad -\int_{0}^{t} \int_{a}^{\infty} \Big\{ 2\mu u_a \partial_r u_a +\lambda u_a \big( \partial_r u_a + m \dfrac{u_a}{r} \big)
+ \kappa \partial_r e_a \Big\} \partial_r\phi
\,r^m {\mathrm d} r {\mathrm d} s,
\end{align*}
where $P_a\vcentcolon= (\gamma-1) \rho_a e_a$, $E_a\vcentcolon=\frac{1}{2}\snorm{u_a}^2 + e_a$, and $E_a^0\vcentcolon= \frac{1}{2}\snorm{u_a^0}^2 + e_a^0$.
\end{enumerate}
\end{definition}
From now on, we denote $\sigma=\sigma(t)\vcentcolon= \min\{1,t\}$.
Moreover, for any function $y(t)\vcentcolon [0,T] \to [0,\infty)$, we denote
\begin{equation*}
\begin{split}
\mathcal{E}[\rho,u,e;y](T)
\vcentcolon=&\sup_{ t\in [0,T]} \int_{y(t)}^{\infty}
\snorm{( \rho-1, u^2 , e-1, \sqrt{\sigma} \partial_r u,\sigma \partial_r e)}^2(r,t)\,r^m {\mathrm d} r\\
&+ \int_{0}^{T}\int_{y(t)}^{\infty}
\snorm{(\partial_r u, \partial_r e, u \partial_r u,\sqrt{\sigma} \partial_t u, \sigma\partial_t e)}^2
\,r^m{\mathrm d} r {\mathrm d} t.
\end{split}
\end{equation*}
\begin{theorem}\label{thm:WSCEx}
For fixed $a\in(0,1)$, let $(\rho_a^0,u_a^0,e_a^0)(r)$ be the initial data constructed in \textnormal{Proposition \ref{prop:mollify}}. Then, for each $T>0$, there exists a weak solution
$(\rho_a,u_a,e_a)(r,t)$ of the exterior problem \eqref{eqs:SFNS} and \eqref{eqs:eSFNS}.
Moreover, the weak solution $(\rho_a,u_a,e_a)(r,t)$ satisfies the following properties{\rm :}
\begin{enumerate}[label=(\roman*),ref=(\roman*),font={\normalfont\rmfamily}]
\item\label{item:WSCEx1} The entropy inequality holds\textnormal{:}
\begin{equation*}
\begin{aligned}
\quad \ &\esssup\limits_{t\in[0,T]}\int_{a}^{\infty}
\Big(\rho_a\big(\frac{1}{2}\snorm{u_a}^2+ \psi(e_a)\big)+ (\gamma-1)G(\rho_a) \Big)(r,t)\,r^m {\mathrm d} r
+ \kappa\int_{0}^{T}\int_{a}^{\infty}\dfrac{\snorm{\partial_r e_a}^2}{ e_a^2}\,r^m {\mathrm d} r{\mathrm d} t \\
&\,\,+ \int_{0}^{T}\int_{a}^{\infty} \Big\{\big(\dfrac{2\mu}{n}+\lambda \big)\dfrac{\snorm{\partial_r u_a+m \frac{u_a}{r}}^2}{e_a} + \dfrac{2m\mu}{n} \dfrac{\snorm{\partial_r u_a-\frac{u_a}{r}}^2}{e_a}\Big\}\,r^m{\mathrm d} r{\mathrm d} t \le C(T).
\end{aligned}
\end{equation*}
\smallskip
\item\label{item:WSCEx2}
$\mathcal{E}[\rho_a,u_a,e_a;a](T)\le C(a)$ and, for a.e. $(r,t)\in [a,\infty)\times[0,T]$,
\begin{equation*}
\qquad\quad \ C^{-1}(a) \le \rho_a(r,t) \le C(a),
\ \snorm{u_a(r,t)} \le C(a)\sigma^{-\frac{1}{4}}(t), \ C^{-1}(a) \le e_a(r,t) \le C(a)\sigma^{-\frac{1}{2}}(t).
\end{equation*}
\item\label{item:WSCEx3} There exists a continuous function $(x,t)\mapsto \widetilde{r}_a(x,t) \vcentcolon [0,\infty)\times[0,T] \to [a,\infty)$
such that $x\mapsto \widetilde{r}_a(x,t)$ is strictly increasing for each $t\in[0,T]$ and
\begin{equation*}
\begin{dcases*}
\int_{a}^{\widetilde{r}_a(x,t)} \rho_a(r,t)\,r^m {\mathrm d} r =x &\quad for a.e. $(x,t)\in[0,\infty)\times[0,T]$,\\
\widetilde{r}_a(x,t) = \widetilde{r}_a^0(x) + \int_{0}^{t} u_a(\widetilde{r}_a(x,s),s)\,{\mathrm d} s &\quad for all $(x,t)\in[0,\infty)\times[0,T]$,\\
n x \psi^{-1}_{-}(\dfrac{C_0}{x})
\le ({\widetilde{r}}_a)^{n}(x,t) \le C_0( 1 +x ) &\quad for all $(x,t)\in[0,\infty)\times[0,T]$,
\end{dcases*}
\end{equation*}
where the function $x\mapsto\widetilde{r}_a^0(x) \vcentcolon [0,\infty)\to [a,\infty)$ is implicitly defined as
\begin{equation*}
x= \int_{a}^{\widetilde{r}_a^0(x)} \rho_a^0(r)\,r^m {\mathrm d} r \qquad
\text{for each $x\in[0,\infty)$.}
\end{equation*}
\smallskip
\item\label{item:WSCEx4} For each $\varepsilon\in(0,1]$,
$\mathcal{E}[\rho_a,u_a,e_a;\widetilde{r}_a(\varepsilon,\cdot)](T) \le C(\varepsilon)$ and
\begin{equation*}
\begin{dcases}
\snorm{\widetilde{r}_a(x_1,t)-\widetilde{r}_a(x_2,t)} \le C(\varepsilon) \snorm{x_1-x_2} &\quad \text{for $(x_1,x_2,t)\in [\varepsilon,\infty)^2\times[0,T]$}, \\
\snorm{\widetilde{r}_a(x,t_1)-\widetilde{r}_a(x,t_2)} \le C(\varepsilon) \snorm{t_1^{\frac{3}{4}}-t_2^{\frac{3}{4}}}\,\, &\quad \text{for $(x,t_1,t_2)\in [\varepsilon,\infty)\times[0,T]^2$}.
\end{dcases}
\end{equation*}
Furthermore, for a.e. $(r,t)$ such that $t\in[0,T]$ and $r\in [\widetilde{r}_a(\varepsilon,t),\infty)${\rm :}
\begin{equation*}
C^{-1}(\varepsilon) \le \rho_a(r,t) \le C(\varepsilon),\quad \snorm{u_a(r,t)} \le C(\varepsilon)\sigma^{-\frac{1}{4}}(t),
\quad e_a(r,t) \le C(\varepsilon)\sigma^{-\frac{1}{2}}(t).
\end{equation*}
In addition,
for all $r_1,\,r_2\in [\widetilde{r}_a(\varepsilon,t),\infty)$ and $t\in[0,T]$,
\begin{equation*}
\sigma^{\frac{1}{2}}(t)\snorm{u_a(r_1,t)-u_a(r_2,t)} + \sigma(t)\snorm{e_a(r_1,t)-e_a(r_2,t)} \le C(\varepsilon)\snorm{r_1-r_2}^{\frac{1}{2}},
\end{equation*}
and, for all $(r,t_1,t_2)$ such that $0<t_1<t_2\le T$
and $r\ge \sup_{t_1\le t \le t_2}\widetilde{r}_a(\varepsilon,t)$,
\begin{equation*}
\sigma^{\frac{1}{2}}(t_1)\snorm{u_a(r,t_1)-u_a(r,t_2)} + \sigma(t_1) \snorm{e_a(r,t_1)-e_a(r,t_2)}\le C(\varepsilon)\snorm{t_2-t_1}^{\frac{1}{4}}.
\end{equation*}
\item\label{item:WSCEx5} For each connected interval $I\subseteq [0,\infty)$, denote $\widetilde{H}_0^1(I,r^m{\mathrm d} r)$ as the closure of
\begin{align*}
\quad\,\, \, \mathcal{D}_0(I)\vcentcolon =\big\{ \phi\in C^{\infty}(I)\vcentcolon\,\exists N>0
\ \text{such that $[0,N]\subset I$ and $\phi(r)=0$ for $r\in I\cap[N,\infty)$}\big\}
\end{align*}
via the $H^1(I,r^m{\mathrm d} r)$--norm, and denote its dual space as $\widetilde{H}^{-1}( I, r^m {\mathrm d} r)$.
For each $\varepsilon>0$, define $\rho_a^{(\varepsilon)}(r,t) \vcentcolon = \rho_a(r,t) \chi_{\varepsilon}^a(r,t)$
with the indicator function $\chi_{\varepsilon}^a(r,t)${\rm :}
\begin{equation*}
\chi_{\varepsilon}^a(r,t) \vcentcolon = \begin{dcases*}
1 & if $t\in[0,T]$ and $r\in [\widetilde{r}_a(\varepsilon,t),\infty)$,\\
0 & otherwise.
\end{dcases*}
\end{equation*}
Then, for all
$L\in\mathbb{N}$, $\rho_{a}^{(\varepsilon)}\in C^{0}\big([0,T];\widetilde{H}^{-1}([0,L],r^m {\mathrm d} r) \big)$ and
\begin{equation*}
\sbnorm{\rho_a^{(\varepsilon)}(\cdot,t_1)-\rho_a^{(\varepsilon)}(\cdot,t_2)}_{\widetilde{H}^{-1}([0,L],r^m{\mathrm d} r)} \le C(\varepsilon)\snorm{t_1-t_2} \qquad\, \text{for all $t_1, t_2 \in [0,T]$.}
\end{equation*}
\smallskip
\item\label{item:WSCEx6} For each $\eta\in (a,1)$, the following estimates hold\textnormal{:}
\begin{equation*}
\qquad \ \begin{dcases*}
\int_{0}^{T} \sup_{r\ge\eta}\dfrac{\snorm{u_a}}{\sqrt{e_a}}(r,t)\,{\mathrm d} t
\le C(T)\big(\eta^{\frac{2-n}{2}}+ \eta^{2-n}\big) & if $n=2$, $3$,\\
\int_{0}^{T} \sup_{r\ge \eta} \log \big( \max\big\{ 1, e_a(r,t)\big\} \big)\, {\mathrm d} t
\le C(T)\big( 1 + \sqrt{|\log\eta|}
\big) & if $n=2$,\\
\int_{0}^{T} \sup_{r\ge\eta} \log \big( \max \big\{ 1, e_a^{\pm 1}(r,t) \big\} \big)\,{\mathrm d} t
\le C(T) \eta^{2-n} & if $n=3$.
\end{dcases*}
\end{equation*}
\end{enumerate}
\end{theorem}
\begin{remark}
For the smooth initial data $(\rho_a^0,u_a^0,e_a^0)$ constructed
in {\rm \S \ref{subsec:mollify}}, it can be shown that there exists a unique global-in-time classical
solution of problem \eqref{eqs:SFNS} and \eqref{eqs:eSFNS} \textnormal{(}see \cite{J}\textnormal{)}. In this paper, since the higher regularity estimates for establishing classical solutions will not be needed
for the proof of the main theorem, we omit them for simplicity.
\end{remark}
\subsection{Truncation of the initial data in the exterior Eulerian domain}\label{subsec:ffaprox}
Since $(\rho_a,u_a,e_a)$ in Theorem \ref{thm:WSCEx} is obtained as the limit of solutions
of the approximate problems posed in bounded annular domains, we need to construct the corresponding approximate initial data, which is the main purpose of this subsection.
In particular, we construct $(\rho_{a,k}^0,u_{a,k}^0,e_{a,k}^0)$ for $k\in\mathbb{N}$ by truncating $(\rho_a^{0},u_a^0,e_a^0)$ with suitable cut-off functions $\varphi_{a,k}^0$.
We first consider the function:
$r \mapsto\, \int_{a}^{r} \rho_a^0(\zeta)\,\zeta^m {\mathrm d} \zeta$,
which is a strictly increasing continuous function since $C_0^{-1} \le \rho_a^0(r) \le C_0$.
Hence, the inverse function $\widetilde{r}_a^0(x)\vcentcolon [0,\infty)\to [a,\infty)$ exists so that
\begin{equation}\label{eqs:intPath}
x = \int_{a}^{\widetilde{r}_a^0(x)} \rho_a^0(r)\,r^m {\mathrm d} r\, \qquad \text{for each $x\in[0,\infty)$.}
\end{equation}
Moreover, due to the regularity of $\rho_a^0$, it follows from the inverse function theorem
that $\widetilde{r}_a^0(x) \vcentcolon [0,\infty)\to [a,\infty)$ is a smooth, bijective, strictly increasing map.
From now on, we use ${\mathrm D}_x$ to represent the derivative with respect to $x\in[0,\infty)$.
\begin{proposition}\label{prop:r0diff}
For all $x\in [0,\infty)$,
\begin{equation}\label{3,2a}
\begin{aligned}
&{\mathrm D}_x \widetilde{r}_a^0(x) = \dfrac{1}{(\widetilde{r}_a^0)^m(x)\rho_a^0(\widetilde{r}_a^0)(x)},\qquad
\ \widetilde{r}_a^0(x) = \Big(a^n + n \int_{0}^{x} \dfrac{1}{\rho_a^0(\widetilde{r}_a^0)(y)} {\mathrm d} y\Big)^{\frac{1}{n}}.
\end{aligned}
\end{equation}
\end{proposition}
\begin{proof}
Since $\rho_a^0(r)\in C^{\infty}$ and $C_0^{-1}\le \rho_a^0(r)\le C_0$ by construction \eqref{eqs:amolint}, it follows that
\begin{align*}
\lim\limits_{r\to \widetilde{r}_a^0(x)}\dfrac{1}{r-\widetilde{r}_a^0(x)}\int_{\widetilde{r}_a^0(x)}^{r}
\rho_a^0(\zeta)\,\zeta^m {\mathrm d} \zeta
=(\widetilde{r}_a^0)^m(x)\rho_a^0(\widetilde{r}_a^0(x))
\qquad \text{for each $x\in[0,\infty)$.}
\end{align*}
Since
$y\mapsto \widetilde{r}_a^0(y)$ is continuous, then, for
all $x\in[0,\infty)$,
\begin{align*}
\lim\limits_{y\to x}\dfrac{1}{\widetilde{r}_a^0(y)-\widetilde{r}_a^0(x)}\int_{\widetilde{r}_a^0(x)}^{\widetilde{r}_a^0(y)} \rho_a^0(\zeta)\,\zeta^m {\mathrm d} \zeta
=(\widetilde{r}_a^0)^m(x) \rho_a^0(\widetilde{r}_a^0)(x).
\end{align*}
Notice that, by \eqref{eqs:intPath},
\begin{align*}
\dfrac{\widetilde{r}_a^0(y)-\widetilde{r}_a^0(x)}{y-x}
\Big(\dfrac{1}{\widetilde{r}_a^0(y)-\widetilde{r}_a^0(x)}
\int_{\widetilde{r}_a^0(x)}^{\widetilde{r}_a^0(y)}\rho_a^0(r)\,r^m {\mathrm d} r\Big)
= \dfrac{1}{y-x} \int_{\widetilde{r}_a^0(x)}^{\widetilde{r}_a^0(y)}\rho_a^0(r)\,r^m {\mathrm d} r = 1.
\end{align*}
Then, using the fact that
$0 < C_0^{-1} a^m \le \rho_0(\widetilde{r}_0)(y)\widetilde{r}_0^{\,m}(y)
\le C_0 \widetilde{r}_0^{\,m}(y) < \infty$ for all $y\in[0,\infty)$, it follows that, for all $x\in[0,\infty)$,
\begin{align*}
\lim\limits_{y\to\ x} \dfrac{\widetilde{r}_a^0(y)-\widetilde{r}_a^0(x)}{y-x}
= \lim\limits_{y\to x} \Big( \dfrac{1}{\widetilde{r}_a^0(y)
-\widetilde{r}_a^0(x)} \int_{\widetilde{r}_a^0(x)}^{\widetilde{r}_a^0(y)}\rho_a^0(r)r^m {\mathrm d} r\Big)^{-1}
= \dfrac{1}{(\widetilde{r}_a^0)^m(x)\rho_a^0(\widetilde{r}_a^0)(x)}.
\end{align*}
Thus, the derivative of $\widetilde{r}_a^0(x)$ exists and $\eqref{3,2a}_1$ holds.
Next, it follows by the chain rule that
\begin{align*}
\dfrac{1}{n} {\mathrm D}_x ( \widetilde{r}_a^0)^n(x)
= (\widetilde{r}_a^0)^m(x) {\mathrm D}_x \widetilde{r}_a^0(x) = \frac{1}{\rho_a^0(\widetilde{r}_a^0)(x)} \qquad \text{for $x\in[0,\infty)$.}
\end{align*}
Integrating in $y\in[0,x]$ and using $\widetilde{r}_a^0(0)=a$, we obtain $\eqref{3,2a}_2$.
The proof is complete.
\end{proof}
Now, let $\chi \vcentcolon {\mathbb R} \to [0,1]$ be such that $\chi\in C^{\infty}$, $\chi(\zeta)=1$ if $\zeta\le 0$, $\chi(\zeta)=0$ if $\zeta\ge 1$, and $\snorm{\chi^{\prime}(\zeta)}\le 2$ for all $\zeta\in{\mathbb R}$.
With this, the cut-off functions $\varphi_{a,k}^0$ are defined as
\begin{equation*}
\varphi_{a,k}^0(r) = \chi\big(\dfrac{2r-\widetilde{r}_a^{0}(k)}{\widetilde{r}_a^{0}(k)}\big) \qquad\,\, \text{for each $(a,k)\in(0,1)\times\mathbb{N}$.}
\end{equation*}
It follows that $\varphi_{a,k}^0(r)\in C^{\infty}([a,\infty))$ and
\begin{equation*}
\varphi_{a,k}^0(r)=1 \ \ \text{for $r\in[a,\frac{\widetilde{r}_a^0(k)}{2}]$,} \qquad\,\,\, \varphi_{a,k}^0(r)=0 \ \ \text{for $r\in[\widetilde{r}_a^0(k),\infty)$.}
\end{equation*}
Using this, we define the truncated initial data:
\begin{equation}\label{def:appInitr1}
(\rho_{a,k}^0,u_{a,k}^0,e_{a,k}^0)(r)
\vcentcolon = ((\rho_a^0-1)\varphi_{a,k}^0+1,\, u_{a}^0\varphi_{a,k}^0, \,(e_a^0-1)\varphi_{a,k}^0+1)(r)
\quad \text{ for $r\in [a,\infty)$.}
\end{equation}
With above construction, it can be verified that
\begin{equation}\label{eqs:akInitProp}
\begin{split}
&( \rho_{a,k}^0, u_{a,k}^0, e_{a,k}^0 ) \in C^{\infty}([a,\infty)),
\qquad\ u_{a,k}^0(r) = \partial_r e_{a,k}^0(r) = 0 \,\,\,\, \text{for $r\in [a, \widetilde{r}_a^0(k)]$},\\[1mm]
&(\rho_{a,k}^0, u_{a,k}^0, e_{a,k}^0)(r) =
\begin{dcases*}
(\rho_a^0, u_a^0, e_a^0)(r) \quad & for all $r \in [a,\frac{\widetilde{r}_a^0(k)}{2})$,\\
(1,0,1) & for all $r \in [\widetilde{r}_a^0(k),\infty)$.
\end{dcases*}
\end{split}
\end{equation}
\begin{proposition}\label{prop:kapprox}
For $G(z)=1-z+z\log z$ and $\psi(z)=z-1-\log z$,
then
following hold:
\begin{equation*}
\begin{split}
& C_0^{-1} \le \rho_{a,k}^0(r) \le C_0, \quad e_{a,k}^0(r)\ge C_0^{-1}\qquad \ \text{ for all $(r,k)\in[a,\infty)\times\mathbb{N}$ and $a\in(0,1)$,}\\
&\sup_{\substack{a\in(0,1)\\ k\in\mathbb{N}}} \int_{a}^{\infty}\!\!
\Big\{\rho_{a,k}^0\big(\frac{1}{2}\snorm{u_{a,k}^0}^2 +\psi (e_{a,k}^0)\big) + G(\rho_{a,k}^0) + \big|(\rho_{a,k}^0-1, |u_{a,k}^0|^2, e_{a,k}^0-1)\big|^2\Big\}\,r^m {\mathrm d} r \le C_0.
\end{split}
\end{equation*}
\end{proposition}
\begin{proof}
By Proposition \ref{prop:mollify}, we see that
\begin{align*}
\int_{a}^{\infty} \Big\{ \snorm{\rho_{a,k}^0-1}^2 + \snorm{u_{a,k}^0}^4
+\snorm{e_{a,k}^0-1}^2 \Big\}\,r^m {\mathrm d} r
\le \int_{a}^{\infty} \Big\{ \snorm{\rho_a^0-1}^2 + \snorm{u_a^0}^4 +\snorm{e_a^0-1}^2 \Big\}\,r^m{\mathrm d} r\le C_0,
\end{align*}
and
$C_0^{-1}\le \rho_{a,k}^0(r) = \rho_a^0(r) \varphi_{a,k}^0(r) + 1-\varphi_{a,k}^0(r) \le C_0$
for all $r\in[a,\infty)$.
By the same argument, we can show that $e_{a,k}^0(r)\ge C_0^{-1}$ for all $r\in[a,\infty)$.
Since $G$ and $\psi$ are convex, $G(1)=\psi(1)=0$, and $0\le \varphi_{a,k}^0\le 1$,
it follows that
\begin{align*}
G(\rho_{a,k}^0) =&\, G(\rho_a^0 \varphi_{a,k}^0 + 1-\varphi_{a,k}^0)\le\varphi_{a,k}^0 G(\rho_a^0)+(1-\varphi_{a,k}^0)G(1)=\varphi_{a,k}^0 G(\rho_a^0),\\[1mm]
\psi(e_{a,k}^0) =&\, \psi( e_a^0\varphi_{a,k}^0 + 1-\varphi_{a,k}^0 ) \le \varphi_{a,k}^0 \psi(e_a^0) + (1-\varphi_{a,k}^0) \psi(1) = \varphi_{a,k}^0 \psi(e_a^0).
\end{align*}
Then, using Proposition \ref{prop:mollify}, we have
\begin{equation*}
\int_{a}^{\infty} \Big\{ \rho_{a,k}^0\big(\frac{1}{2}\snorm{u_{a,k}^0}^2+\psi(e_{a,k}^0)\big)+ G(\rho_{a,k}^0)\Big\}\,r^m {\mathrm d} r
\le C_0\int_{a}^{\infty}\Big\{\rho_a^0\big(\dfrac{1}{2}\snorm{u_a^0}^2+ \psi(e_a^0)\big) +G(\rho_a^0)\Big\}\,r^m{\mathrm d} r \le C_0.
\end{equation*}
This completes the proof.
\end{proof}
\subsection{Reformulation in the bounded Lagrangian domain}\label{subsec:lageqs}
For each $(a,k)\in(0,1)\times\mathbb{N}$, we transform $(\rho_{a,k}^0,u_{a,k}^0,e_{a,k}^0)(r)$ into the Lagrangian domain as
\begin{equation}\label{def:appInit}
(\widetilde{v}_{a,k}^0,\widetilde{u}_{a,k}^0,\widetilde{e}_{a,k}^0)(x)\vcentcolon= \big((\rho_{a,k}^0)^{-1},u_{a,k}^0,e_{a,k}^0\big)(\widetilde{r}_a^0(x))\qquad\, \text{for $x\in [0,\infty)$.}
\end{equation}
Since $\widetilde{r}_a^0(x)\in C^{\infty}([0,\infty))$ due to $\rho_{a}^0 \in C^{\infty}([a,\infty))$, Proposition \ref{prop:r0diff} and \eqref{eqs:akInitProp} imply that
\begin{equation*}
(\widetilde{v}_{a,k}^0,\widetilde{u}_{a,k}^0,\widetilde{e}_{a,k}^0)\in C^{\infty}([0,\infty)),
\, \,\, \widetilde{u}_{a,k}^0 (x)\big\vert_{x=0,\,k} = {\mathrm D}_x \widetilde{e}_{a,k}^0 (x)\big\vert_{x=0,\,k} = 0
\qquad \text{ for $(a,k) \in (0,1)\times\mathbb{N}$.}
\end{equation*}
\begin{proposition}\label{prop:kapproxL}
The following uniform estimate holds{\rm :}
\begin{equation*}
\sup_{a\in(0,1)}\sup_{k\in\mathbb{N}}\int_{0}^{k} \Big(\dfrac{1}{2}\snorm{\widetilde{u}_{a,k}^0}^2
+ \psi(\widetilde{e}_{a,k}^0) + \psi(\widetilde{v}_{a,k}^0)
+ \snorm{(\widetilde{v}_{a,k}^0-1, |\widetilde{u}_{a,k}^0|^2, \widetilde{e}_{a,k}^0-1)}^2\Big)(x)\,{\mathrm d} x \le C_0.
\end{equation*}
\end{proposition}
\begin{proof}
First, it follows from Proposition \ref{prop:r0diff} that the map: $x\mapsto \widetilde{r}_a^0(x)$ satisfies
\begin{equation*}
0<C_0^{-1}\, (\widetilde{r}_a^0)^{-m}(x)\le \snorm{{\mathrm D}_x \widetilde{r}_a^0(x)}
= (\widetilde{r}_a^0)^{-m}(x)\, (\rho_a^0)^{-1}(\widetilde{r}_a^0(x)) \le C_0 a^{-m}<\infty \qquad
\text{for $x\in[0,\infty)$,}
\end{equation*}
so that the Jacobian of transformation $r=\widetilde{r}_a^0(x)$ is bounded.
Then, for any function $f(r)\vcentcolon [a,\infty)\to \mathbb{R}$ such that $f\in L^1_{\text{loc}}$, the following coordinate transformation is satisfied:
\begin{align*}
\int_{a}^{\widetilde{r}_a^0(k)} f(r) \rho_a^0(r)\,r^m {\mathrm d} r
= \int_{0}^{k} f(\widetilde{r}_a^0)(x)\,{\mathrm d} x.
\end{align*}
Using this, \eqref{eqs:akInitProp}, and Propositions \ref{prop:mollify}--\ref{prop:kapprox},
it follows that, for each $(a,k)\in(0,1)\times\mathbb{N}$,
\begin{align*}
\int_{0}^{k} \snorm{\widetilde{v}_{a,k}^0-1}^2 (x)\,{\mathrm d} x
= \int_{0}^{k} \norm{1-\rho_{a,k}^0\big(\widetilde{r}_a^0(x)\big)}^2 \norm{\rho_{a,k}^0\big(\widetilde{r}_a^0(x)\big)}^{-2}{\mathrm d} x
\le C_0\int_{a}^{\infty}\norm{\rho_a^0(r)-1}^2 r^m {\mathrm d} r \le C_0.
\end{align*}
By the same argument,
$\sbnorm{(\snorm{\widetilde{u}_{a,k}^0}^2,\widetilde{e}_{a,k}^0-1)}_{L^2([a,\infty),r^m{\mathrm d} r)}\le C_0$.
Next, by Propositions \ref{prop:mollify}--\ref{prop:kapprox}
and the identity: $z\psi(z^{-1})= 1-z+z\log z = G(z)$, we have
\begin{align*}
&\int_{0}^{k} \Big(\frac{1}{2}\snorm{\widetilde{u}_{a,k}^0}^2+\psi(\widetilde{e}_{a,k}^0)+ (\gamma-1)\psi(\widetilde{v}_{a,k}^0)\Big)(x)\,{\mathrm d} x\\
&= \int_{a}^{\infty} \dfrac{\rho_a^0}{\rho_{a,k}^0}
\Big( \rho_{a,k}^0\big(\frac{1}{2}\snorm{u_{a,k}^0}^2+ \psi (e_{a,k}^0)\big)
+ (\gamma-1) G(\rho_{a,k}^0) \Big)(r)\,r^m {\mathrm d} r \le C_0
\end{align*}
for all $(a,k)\in(0,1)\times\mathbb{N}$. This completes the proof.
\end{proof}
The vector function $(\widetilde{v}_{a,k},\widetilde{u}_{a,k},\widetilde{e}_{a,k})$ is called a solution of the $(a,k)$-approximate IBVP in the Lagrangian coordinates,
provided that $(\widetilde{v}_{a,k},\widetilde{u}_{a,k},\widetilde{e}_{a,k})$ satisfies the equations:
\begin{equation}\label{eqs:LFNS-k}
\begin{dcases}
{\mathrm D}_t \widetilde{v}_{a,k} - {\mathrm D}_x(\widetilde{r}_{a,k}^m \widetilde{u}_{a,k}) =0,\\
{\mathrm D}_t \widetilde{u}_{a,k} + \widetilde{r}_{a,k}^m {\mathrm D}_x p(\widetilde{v}_{a,k},\widetilde{e}_{a,k}) = \beta \widetilde{r}_{a,k}^m {\mathrm D}_x \Big(\dfrac{{\mathrm D}_x (\widetilde{r}_{a,k}^m \widetilde{u}_{a,k})}{\widetilde{v}_{a,k}} \Big),\\
{\mathrm D}_t \widetilde{e}_{a,k} +p(\widetilde{v}_{a,k},\widetilde{e}_{a,k}){\mathrm D}_x(\widetilde{r}_{a,k}^m \widetilde{u}_{a,k}) = \mathcal{G}_{a,k}, \\
\widetilde{r}_{a,k}(x,t) = \Big(a^n + n\displaystyle\int_{0}^x \widetilde{v}_{a,k}(y,t){\mathrm d} y\Big)^{\frac{1}{n}}
\end{dcases}
\end{equation}
in $[0,k]\times[0,T]$ with $p(v,e)\vcentcolon=\frac{(\gamma-1)e}{v}$, $\beta\vcentcolon=2\mu + \lambda$, and
\begin{equation*}
\mathcal{G}_{a,k} \vcentcolon= \beta \dfrac{\snorm{{\mathrm D}_x(\widetilde{r}_{a,k}^m \widetilde{u}_{a,k})}^2}{\widetilde{v}_{a,k}} - 2m\mu {\mathrm D}_x (\widetilde{r}_{a,k}^{m-1}\widetilde{u}_{a,k}^2) + \kappa {\mathrm D}_x\Big( \dfrac{r_{a,k}^{2m} {\mathrm D}_x \widetilde{e}_{a,k}}{\widetilde{v}_{a,k}} \Big),
\end{equation*}
and the initial-boundary conditions:
\begin{equation}\label{eqs:LFNS-kb}
\begin{dcases}
\widetilde{u}_{a,k}(x,t)\big\vert_{x=0,\,k}={\mathrm D}_x \widetilde{e}_{a,k}(x,t)\big\vert_{x=0,\,k}=0 & \text{for $t\in[0,T]$},\\
(\widetilde{v}_{a,k},\widetilde{u}_{a,k},\widetilde{e}_{a,k})(x,0)
= (\widetilde{v}_{a,k}^0,\widetilde{u}_{a,k}^0,\widetilde{e}_{a,k}^0)(x)\qquad
&\text{for $x\in [0,k]$},
\end{dcases}
\end{equation}
\begin{remark}
The equations in \eqref{eqs:LFNS-k} are formally derived
from system \eqref{eqs:SFNS} and \eqref{eqs:eSFNS} as follows{\rm :} for any smooth solution $(\rho,u,e)$ of \eqref{eqs:SFNS} and \eqref{eqs:eSFNS}, the transformation
from the Eulerian coordinates $(r,t)$ to the Lagrangian coordinates $(x,t)$ is defined as
\begin{equation}
x(r,t)\vcentcolon = \int_{a}^{r} \rho(s,t)\,s^m {\mathrm d} s, \quad t(r,t)\vcentcolon = t
\qquad \text{ for $(r,t)\in [a,\infty)\times[0,T]$}.
\end{equation}
The continuity equation $\eqref{eqs:SFNS}_1$ implies that
$\partial_t x(r,t) = -\rho u r^m$.
In addition, by the implicit function theorem,
on a neighbourhood where $\rho>0$, there is a function $r(x,t)$ such that
\begin{equation}\label{eqs:impl}
\int_a^{r(x,t)}\rho(\zeta,t)\,\zeta^m {\mathrm d} \zeta = x.
\end{equation}
Denote $(v,u,e)(x,t):=(\rho^{-1}(r(x,t),t),u(r(x,t),t),e(r(x,t),t))$
and $({\mathrm D}_t, {\mathrm D}_x)$ as the derivative with respect to the Lagrangian coordinates.
Then it follows by taking derivative ${\mathrm D}_t$ on $\eqref{eqs:impl}$ and the Leibniz theorem
that ${\mathrm D}_t r(x,t) = u(x,t)$. These relations can be summarized as
\begin{equation}\label{eqs:jacob}
\dfrac{\partial (x,t)}{\partial (r,t)} =
\begin{pmatrix}
\rho r^m & -\rho u r^m \\[1mm]
0 & 1
\end{pmatrix}, \qquad
\dfrac{\partial (r,t)}{\partial (x,t)} =
\begin{pmatrix}
v r^{-m} & u \\[1mm]
0 & 1
\end{pmatrix}.
\end{equation}
Furthermore, from the continuity equation $\eqref{eqs:SFNS}_1$,
\begin{equation}\label{eqs:kinetic}
r(x,t) = r_0(x) + \int_{0}^{t} u(x,\tau)\,{\mathrm d}\tau,
\qquad
r^n(x,t) = a^n + n\int_{0}^x v(y,t)\,{\mathrm d} y,
\end{equation}
where $\displaystyle r_0(x) \vcentcolon=\big( a^n + n\int_{0}^{x} \rho_0^{-1}(r(y))\,{\mathrm d} y\big)^{\frac{1}{n}}$. Equations $\eqref{eqs:LFNS-k}_1$--$\eqref{eqs:LFNS-k}_3$ are then obtained by translating \eqref{eqs:SFNS} using the coordinate transformation relations \eqref{eqs:jacob}.
\end{remark}
By virtue of the well-known global existence theorem for problem \eqref{eqs:LFNS-k}--\eqref{eqs:LFNS-kb}, the initial data constructed in \eqref{def:appInit} implies the following theorem (see \cites{J,YB,YB2}):
\begin{theorem}\label{thm:localt}
Given initial data $(\widetilde{v}_{a,k}^0,\widetilde{u}_{a,k}^0,\widetilde{e}_{a,k}^0)(x)$ for $x\in[0,k]$,
then there exists a unique classical solution
$(v_{a,k},u_{a,k},e_{a,k})(x,t)$ of problem \eqref{eqs:LFNS-k}--\eqref{eqs:LFNS-kb} in domain $(x,t)\in[0,k]\times[0,\infty)$
satisfying
\begin{equation*}
v_{a,k}(x,t)\in C^{1+\alpha,1+\frac{\alpha}{2}}\big([0,k]\times[0,\infty)\big),
\quad (u_{a,k},e_{a,k})(x,t)\in C^{2+\alpha,1+\frac{\alpha}{2}}\big([0,k]\times[0,\infty)\big)
\end{equation*}
for some $\alpha\in(0,1)$.
\end{theorem}
By the above theorem, there exists a unique global classical solution of \eqref{eqs:LFNS-k}--\eqref{eqs:LFNS-kb} for each fixed $(a,k)\in(0,1)\times\mathbb{N}$. This allows us to derive the {\it a-priori} estimates in \S 4 below.
\section{The A-Priori Estimates}\label{sec:rhobd}
This section is devoted to the derivation of the {\it a-priori} estimates for solutions
of the approximation problem \eqref{eqs:LFNS-k}--\eqref{eqs:LFNS-kb}.
For simplicity, we suppress the approximation parameters $(a,k)\in(0,1)\times\mathbb{N}$
and denote
$(v,u,e,r)\equiv (\widetilde{v}_{a,k},\widetilde{u}_{a,k},\widetilde{e}_{a,k},\widetilde{r}_{a,k})$ as the solution obtained from Theorem \ref{thm:localt} in $[0,k]\times[0,T]$ for each $T>0$.
First, we state the main theorem of this section.
\begin{theorem}\label{thm:priori}
The following estimates are satisfied{\rm :}
\begin{enumerate}[label=(\roman*), ref=(\roman*),font={\normalfont\rmfamily}]
\item\label{item:prio0}
The entropy estimate holds with $\psi(\zeta):=\zeta-1-\log\zeta${\rm :}
\begin{align}\label{eqs:entropy}
&\sup\limits_{t\in[0,T]}\int_{0}^{k}
\Big(\dfrac{1}{2}\snorm{u}^2 + \psi(e)+(\gamma-1)\psi(v)\Big)(x,t)\,{\mathrm d} x
+ \kappa\int_{0}^{T}\int_{0}^{k} \dfrac{r^{2m}\snorm{{\mathrm D}_x e}^2}{v e^2}\,{\mathrm d} x{\mathrm d} t\nonumber\\
&+ \int_{0}^{T}\int_{0}^{k} \Big\{\Big(\lambda + \dfrac{2\mu}{n}\Big)\dfrac{\snorm{{\mathrm D}_x (r^m u)}^2}{ve}+2m\mu \dfrac{v}{e}\Big(\dfrac{{\mathrm D}_x\big(r^m u\big)}{v\sqrt{n}}-\sqrt{n}\dfrac{u}{r}\Big)^2\Big\}\,{\mathrm d} x {\mathrm d} t
\le C_0.
\end{align}
\item\label{item:prio1}
For each given constant $\varepsilon\ge0$,
\begin{equation}\label{eqs:vulb}
\underline{v}(\varepsilon,t,a) \le v(x,t) \le \overline{v}(\varepsilon,t,a) \qquad
\text{for all $(x,t) \in [\varepsilon,\infty)\times[0,T)$},
\end{equation}
with $\overline{v}=\overline{v}(z,t,a)$ and $\underline{v}=\underline{v}(z,t,a)$ explicitly given by
\begin{equation}\label{vbarname}
\begin{dcases*}
\overline{v}=C_0 (1+t)^2 f(z,a) \exp\big\{ C_0 t f(z,a) + C_0 (1+t) h(z,a) f(z,a) \exp\{C_0 t f(z,a)\} \big\},\\
\underline{v}=C_0 (1+t)^{-1}f^{-1}(z,a)\exp\{ -t C_0 f^2(z,a) \}\Gamma(z,t,a),
\end{dcases*}
\end{equation}
where $h(z,a)$, $f(z,a)$, and $\Gamma(z,t,a)$ are defined by
\begin{equation}\label{fhF}
\begin{aligned}
& h(z,a)\vcentcolon=\big(a^n+nz\psi_{-}^{-1}(\frac{C_0}{z})\big)^{-\frac{2m}{n}},
\quad f(z,a)\vcentcolon= \exp\big\{C_0h(z,a)\big\}, \\
& \Gamma(z,t,a)\vcentcolon = \exp\big\{-\dfrac{mC_0t}{\beta} h^{\frac{n}{2m}}(z,a)\big\},
\end{aligned}
\end{equation}
with $\psi_{-}^{-1}(y)\vcentcolon [0,\infty)\to (0,1]$ as
the left
inverse of
$\psi$.
In particular, for each $\varepsilon>0$,
\begin{equation}\label{4.34.3}
\begin{dcases*}
C^{-1}(T) \le v(x,t) \le C(T) \quad & for all $(x,t) \in [1,k]\times[0,T)$,\\
C^{-1}(\varepsilon) \le v(x,t) \le C(\varepsilon) & for all $(x,t) \in [\varepsilon,k]\times[0,T]$,\\
C^{-1}(a) \le v(x,t) \le C(a) & for all $(x,t) \in [0,k]\times[0,T]$.
\end{dcases*}
\end{equation}
\item\label{item:prio2}
For each $a\in (0, 1),\, \mathcal{L}_0 [v,u,e](T)\le C(a)$ and, for $(x,t)\in [0,k]\times[0,T]$,
\begin{equation*}
\qquad \ \ C^{-1}(a) \le v(x,t) \le C(a),\, \,\, \snorm{u(x,t)}\le C(a)\sigma^{-\frac{1}{4}}(t),
\,\,\, C^{-1}(a) \le e(x,t) \le C(a)\sigma^{-\frac{1}{2}}(t);
\end{equation*}
and, for each $\varepsilon\in(0,1]$, $\mathcal{L}_y [v,u,e](T)\le C(\varepsilon)$ and
\begin{equation*}
\qquad C^{-1}(\varepsilon) \le v(x,t) \le C(\varepsilon), \,\,\,
\sigma(t)^{\frac{1}{4}}\snorm{u(x,t)}+\sigma^{\frac{1}{2}}(t) e(x,t) \le C(\varepsilon)
\quad\mbox{for $(x,t)\in[\varepsilon,k]\times[0,T]$},
\end{equation*}
where $\sigma(t)\vcentcolon= \min\{1,t\}$ and, for $y\in[0,1)$,
$\mathcal{L}_{y}[v,u,e](T)$ is defined by
\begin{equation*}
\begin{split}
\mathcal{L}_{y} [v,u,e](T)\vcentcolon
=&\sup_{ t\in [0,T]} \int_{y}^{k} \norm{\big( v-1, u^2 , e-1, \sqrt{\sigma}r^m {\mathrm D}_x u,\sigma r^m {\mathrm D}_x e\big)}^2(x,t)\,{\mathrm d} x\\
&+ \int_{0}^{T}\int_{y}^{k} \norm{\big(r^m{\mathrm D}_x u, r^m {\mathrm D}_x e, r^m u {\mathrm D}_x u,\sqrt{\sigma} {\mathrm D}_t u, \sigma {\mathrm D}_t e\big)}^2
\,{\mathrm d} x {\mathrm d} t.
\end{split}
\end{equation*}
\end{enumerate}
\end{theorem}
\smallskip
\subsection{Entropy estimate}\label{subsec:entEst}
We now start the proof of Theorem \ref{thm:priori} with the derivation of entropy estimate stated in \ref{item:prio0}, which is motivated from the second law of thermodynamics. It encapsulates the dissipative effect of viscosity and thermal diffusion.
\begin{proof}[Proof of {\rm \ref{item:prio0}} in {\rm Theorem \ref{thm:priori}}]
It follows from $\eqref{eqs:LFNS-k}_1 $--$ \eqref{eqs:LFNS-k}_3$ and direct calculations that
\begin{align}\label{temp2}
&\dfrac{{\mathrm d}}{{\mathrm d} t}\int_{0}^{k}
\Big(\frac{1}{2}\snorm{u}^2+ \psi(e)+(\gamma-1)\psi(v)\Big)\,{\mathrm d} x\nonumber\\
&= - \beta \int_{0}^{k} \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{ev}{\mathrm d} x
+ 2m\mu\int_{0}^{k} \dfrac{1}{e} {\mathrm D}_x(r^{m-1}u^2)\,{\mathrm d} x
- \kappa \int_{0}^{k} \dfrac{r^{2m}\snorm{{\mathrm D}_x e}^2}{v e^2}{\mathrm d} x.
\end{align}
According to $\eqref{eqs:LFNS-k}_1$ and $\eqref{eqs:LFNS-k}_4$,
we can obtain
\begin{equation*}
2m\mu \dfrac{1}{e} {\mathrm D}_x(r^{m-1}u^2) - \beta\dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{ev}
= -\Big(\dfrac{2\mu}{n}+\lambda\Big)\dfrac{\snorm{{\mathrm D}_x (r^m u)}^2}{ve} - 2m\mu \dfrac{v}{e}\Big(\dfrac{{\mathrm D}_x(r^m u)}{v\sqrt{n}}-\dfrac{u\sqrt{n}}{r}\Big)^2,
\end{equation*}
which, along with \eqref{temp2}, yields \eqref{eqs:entropy}.
\end{proof}
\smallskip
\subsection{Upper and lower bounds of the density}\label{subsec:rhobd}
Now we aim to prove \eqref{eqs:vulb}--\eqref{4.34.3}.
By \eqref{eqs:entropy}, we first obtain some upper and lower bounds of the particle path function $r(x,t)$.
\begin{lemma}
\label{lemma:rbd} Let $\psi_{-}^{-1}(\cdot)$
be the left branch inverse of $\psi$. Then
\begin{equation}\label{eqs:rbd}
a^n + nx\psi^{-1}_{-}(\dfrac{C_0}{x}) \le r^n(x,t) \le C_0(1+x)
\qquad \text{ for $(x,t)\in [0,k]\times[0,\infty)$}.
\end{equation}
\end{lemma}
\begin{proof}
Since $\psi(\zeta)\vcentcolon= \zeta-\log \zeta-1$ is strictly convex in $(0,\infty)$ and $\psi(1)=0$,
then there exists the left inverse function $\psi_{-}^{-1}\vcentcolon [0,\infty)\to (0,1]$
and the right inverse function $\psi_{+}^{-1}\vcentcolon [0,\infty)\to [1,\infty)$.
In particular, it is direct to see that $y\mapsto \psi_{-}^{-1}(y)$ is strictly decreasing
($\psi_{-}^{-1}(0)=1$ and $\psi_{-}^{-1}(y)\to 0$ as $y\to \infty$),
while $y\mapsto \psi_{+}^{-1}(y)$ is strictly increasing ($\psi_{+}^{-1}(0)=1$ and $\psi_{+}^{-1}(y)\to \infty$ as $y\to \infty$).
Now, from relation \eqref{eqs:kinetic}, Theorem \ref{thm:priori}\ref{item:prio0}, and the Jensen inequality,
it follows that, for all $(x,t)\in [0,k]\times[0,\infty)$,
\begin{equation}\label{temp3}
\psi(\dfrac{r^n(x,t)-a^n}{nx})
= \psi(\dfrac{1}{x}\int_{0}^x v(y,t)\,{\mathrm d} y)
\le \dfrac{1}{x}\int_{0}^x \psi(v)(y,t)\,{\mathrm d} y \le \dfrac{C_0}{x}.
\end{equation}
Next, the proof is divided into two cases: $0<\frac{r^n-a^n}{nx}\le 1$ and $\frac{r^n-a^n}{nx}\ge 1$.
\smallskip
Case 1: $0<\frac{r^n-a^n}{nx}\le 1$. Then $\frac{r^n-a^n}{nx}=\psi_{-}^{-1}(\psi(\frac{r^n-a^n}{nx}))$.
Since $\psi_{-}^{-1}(\cdot)$ is monotone decreasing, it follows from \eqref{temp3} that
\begin{equation*}
\dfrac{r^n-a^n}{nx}=\psi_{-}^{-1}(\psi(\dfrac{r^n-a^n}{nx}))\ge \psi_{-}^{-1}(\dfrac{C_0}{x}).
\end{equation*}
On the other hand, since $\frac{r^n-a^n}{nx}\le 1$, it follows from $a<1$ that
\begin{equation*}
r^n(x,t)\le a^n + nx
\le n (1+x).
\end{equation*}
\smallskip
Case 2: $\frac{r^n-a^n}{nx}\ge 1$. Then it follows from $\psi_{-}^{-1}(y)\le 1$ for all $y\in[0,\infty)$ that
\begin{equation*}
r^n(x,t)\ge a^n+nx\ge a^n+nx\psi_{-}^{-1}(\dfrac{C_0}{x}).
\end{equation*}
Moreover, $\frac{r^n-a^n}{nx}=\psi_{+}^{-1}(\psi(\frac{r^n-a^n}{nx}))$. Since $\psi_{+}^{-1}(\cdot)$ is
monotone increasing, it follows from \eqref{temp3} that
\begin{equation*}
\dfrac{r^n-a^n}{nx}=\psi_{+}^{-1}(\psi(\dfrac{r^n-a^n}{nx}))\le\psi_{+}^{-1}(\dfrac{C_0}{x}).
\end{equation*}
It is verified in Proposition \ref{prop:omega}
that $x\mapsto x \psi_{+}^{-1}(\frac{C_0}{x})$ is monotone increasing.
If $x\le 1$, then $r^n(x,t)\le a^n+n x \psi_{+}^{-1}(\frac{C_0}{x}) \le 1 + n \psi_{+}^{-1}(C_0) \le C_0 + C_0 x$.
If $x> 1$, since $y\mapsto \psi_{+}^{-1}(y)$ is monotone increasing,
then $r^n(x,t)\le a^n+n x \psi_{+}^{-1}(\frac{C_0}{x}) \le 1 + n x \psi_{+}^{-1}(C_0) \le C_0 + C_0 x$.
This concludes the proof.
\end{proof}
\begin{lemma}\label{lemma:select} Fix $t\in[0,T]$. For each $i=1, \dotsc, k$,
there exist
$A_i(t)$, $B_i(t) \in(i-1,i)$ so that
\begin{align*}
C_0^{-1}\le v(A_i(t),t),\, e(B_i(t),t)\le C_0.
\end{align*}
\end{lemma}
\begin{proof}
We give our proof only for $A_i(t)$, since the proof for $B_i(t)$ is the same.
Fix an integer $i=1,\dotsc,k$. Since $\psi(\cdot)$
is convex, it follows from Jensen's inequality and Theorem \ref{thm:priori}\ref{item:prio0} that
\begin{align*}
\psi( \int_{i-1}^{i} v(x,t)\,{\mathrm d} x)\le \int_{i-1}^{i}\psi(v(x,t))\,{\mathrm d} x \le\dfrac{C_0}{\gamma-1}.
\end{align*}
As in Lemma \ref{lemma:rbd}, denote $z\mapsto\psi_{\pm}^{-1}(z)$ to be the left and
right branch of the convex function $\psi(\zeta)$. Since $\psi(1)=0$, we have
\begin{align*}
0<\psi_{-}^{-1}(\dfrac{C_0}{\gamma-1})
\le \int_{i-1}^{i} v(x,t)\,{\mathrm d} x\le\psi_{+}^{-1}( \dfrac{C_0}{\gamma-1})<\infty.
\end{align*}
By the mean value theorem, there exists a point $A_i(t)\in(i-1,i)$ such that
\begin{align*}
\psi_{-}^{-1}(\dfrac{C_0}{\gamma-1})
\le v(A_i(t), t) \le \psi_{+}^{-1}(\dfrac{C_0}{\gamma-1}).
\end{align*}
Since $e(x,t)$ satisfies the same estimate that differs only by constant $\gamma-1$ in Theorem \ref{thm:priori}\ref{item:prio0},
by the same argument, there also exists $B_i(t)\in(i-1,i)$ such that $C_0^{-1}\le e(B_i(t),t)\le C_0$.
\end{proof}
Using Lemmas \ref{lemma:rbd}--\ref{lemma:select} and Theorem \ref{thm:priori}\ref{item:prio0},
we are ready to prove Theorem \ref{thm:priori}\ref{item:prio1}.
\begin{proof}[Proof of Theorem \textnormal{\ref{thm:priori}\ref{item:prio1}}]
Fix
$i=1,\dotsc,k-2$. Since we focus on obtaining an explicit bound for $v(x,t)$ near the origin $x=0$,
the case for $i\ge k-1$ is not considered in the proof (but can be handled by repeating the same argument).
We divide the proof into six steps.
\medskip
1. Applying Lemma \ref{lemma:select}, it follows that, for each $(x,t)\in[i-1,i+1]\times[0,T)$,
there exist $A_i(t)\in (x, x+3)$ and $ B_i(t)\in (i,i+1)$ such that
\begin{equation}\label{eqs:crb}
C_0^{-1}\le v(A_i(t),t),\, e(B_i(t),t) \le C_0.
\end{equation}
Now, from $\eqref{eqs:LFNS-k}_1$--$\eqref{eqs:LFNS-k}_2$,
\begin{equation}\label{temp4}
{\mathrm D}_t u + r^m {\mathrm D}_x p = \beta r^m {\mathrm D}_x {\mathrm D}_t \log v.
\end{equation}
Dividing both sides of \eqref{temp4} by $\beta r^{m}$ and integrating over $[x,A_i(t)]\times[0,t]$ for some $(x,t)\in(i-1,i+1]\times[0,T)$, we have
{
\begin{equation*}
\log \dfrac{v(A_i(t),t)v_0(x)}{v_0(A_i(t))v(x,t)}
=\dfrac{1}{\beta}\int_{x}^{A_i(t)}\dfrac{u}{r^m}\bigg\vert_0^t{\mathrm d} y
+\dfrac{m}{\beta}\int_{x}^{A_i(t)}\int_{0}^{t}\dfrac{u^2}{r^n}\,{\mathrm d} s{\mathrm d} y+\dfrac{1}{\beta}\int_{0}^{t}\big(p(A_i(t),s)-p(x,s)\big)\,{\mathrm d} s,
\end{equation*}
}
where $p(x,s)\vcentcolon = p(v(x,s),e(x,s))$. Taking the exponential on both sides yields
\begin{equation}\label{temp5}
\dfrac{E(x,t)}{D(x,t)}Y(t)\dfrac{v_0(A_i(t))v(x,t)}{v(A_i(t),t)v_0(x)}
= \exp\Big\{\dfrac{1}{\beta}\int_{0}^{t}p(x,s)\,{\mathrm d} s\Big\},
\end{equation}
where
{\small
\begin{equation*}
\begin{split}
&E(x,t)\vcentcolon = \exp\Big\{\dfrac{m}{\beta}\int_{0}^{t}\int_{x}^{A_i(t)}\dfrac{u^2}{r^n}\,{\mathrm d} y {\mathrm d} s\Big\},
\quad Y(t)\vcentcolon = \exp\Big\{\dfrac{1}{\beta}\int_{0}^{t}p(A_i(t),s)\,{\mathrm d} s\Big\}, \\
&D(x,t)\vcentcolon = \exp\Big\{\dfrac{1}{\beta}\int_{x}^{A_i(t)} \Big(\dfrac{u_0(y)}{r_0^m(y)}
-\dfrac{u(y,t)}{r^m(y,t)}\Big)\,{\mathrm d} y \Big\}.
\end{split}
\end{equation*}
}
Multiplying $\eqref{temp5}$ with $\beta^{-1}p(x,t)$ and integrating in time, we have
\begin{equation*}
1 + \dfrac{1}{\beta}\int_{0}^{t} \dfrac{E(x,s)}{D(x,s)}Y(s)p(x,s)
\dfrac{v_0(A(x,s))v(x,s)}{v(A(x,s),s)v_0(x)}\,{\mathrm d} s
=\exp\Big\{\dfrac{1}{\beta}\int_{0}^{t}p(x,s)\,{\mathrm d} s\Big\}.
\end{equation*}
Substituting this back into \eqref{temp5}, it follows that
\begin{equation}\label{eqs:repre}
\dfrac{E(x,t)}{D(x,t)}Y(t)\dfrac{v_0(A_i(t))v(x,t)}{v(A_i(t),t)v_0(x)}
=1 + \dfrac{1}{\beta}\int_{0}^{t} \dfrac{E(x,s)}{D(x,s)}Y(s)p(x,s)
\dfrac{v_0(A(x,s))v(x,s)}{v(A(x,s),s)v_0(x)}\,{\mathrm d} s.
\end{equation}
\smallskip
2. First, for $D(x,t)$, it follows from Lemma \ref{lemma:rbd}--\ref{lemma:select}
and Theorem \ref{thm:priori}\ref{item:prio0} that
\begin{align*}
\Bignorm{\int_{x}^{A_i(t)}\dfrac{u}{r^m}(y,t)\,{\mathrm d} y }
&\le \Big(\int_{x}^{A_i(t)}r^{-2m}\,{\mathrm d} y\Big)^{\frac{1}{2}}
\Big(\int_{x}^{A_i(t)}\snorm{u}^2\,{\mathrm d} y\Big)^{\frac{1}{2}}\\
&\le \Big(\int_{x}^{A_i(t)}\big(a^n+ny\psi_{-}^{-1}(\frac{C_0}{y})\big)^{-\frac{2m}{n}} {\mathrm d} y\Big)^{\frac{1}{2}}C_0^{\frac{1}{2}}\\
&\le C_0 + C_0 \big(a^n + n x \psi_{-}^{-1}(\frac{C_0}{x})\big)^{-\frac{2m}{n}}.
\end{align*}
Then, using the definition of $f(x,a)$ in \eqref{fhF} and Lemma \ref{lemma:select}, we have
\begin{equation}\label{eqs:D}
(e^{C_0} f(x,a))^{-1} \le D(x,t) \le e^{C_0} f(x,a).
\end{equation}
\smallskip
3. Now we estimate the terms: $Y(t)$ and $E(x,t)$.
In view of the definition of $\Gamma(z,t,a)$ in \eqref{fhF}, Theorem \ref{thm:priori}\ref{item:prio0},
and Lemma \ref{lemma:rbd}, we see that, for $t\ge s \ge 0$,
\begin{equation}\label{eqs:E}
1 \le E(x,t) \le \Gamma^{-1}(x,t,a),\,\,\,\,\,
\Gamma(x,t-s,a)\le\dfrac{E(x,s)}{E(x,t)}
=\exp\Big\{\dfrac{m}{\beta}\int_{s}^{t}\int_{i-1}^{i}\dfrac{u^2}{r^n}\,{\mathrm d} y {\mathrm d} \tau\Big\} \le 1.
\end{equation}
Combining \eqref{eqs:crb} and \eqref{eqs:D}--\eqref{eqs:E} with \eqref{temp5} yields that, for all $(x,t)\in(i-1,i+1]\times[0,T)$,
\begin{equation}\label{eqs:gron1}
v(x,t) Y(t)
\le C_0 f^2(x,a)\Big\{1 + \int_{0}^{t} e(x,s)Y(s)\,{\mathrm d} s \Big\}.
\end{equation}
Since the above argument is true for any $(x,t)\in(i-1,i+1]\times[0,T)$, for a fixed $x\in(i-1,i]$,
we can integrate both sides of the above equation in $y\in[x,x+1]\subset(i-1,i+1]$. It then follows from the entropy inequality \eqref{eqs:entropy} and Jensen's inequality that
\begin{equation*}
Y(t) \le C_0 f^2(x,a) \Big\{ 1 + \int_{0}^{t} Y(s)\,{\mathrm d} s \Big\} \qquad \text{for all $t\in[0,T)$,}
\end{equation*}
where we have used that $x\mapsto f(x,a)$ is monotone decreasing. Thus, by Gr\"onwall's inequality,
\begin{equation}\label{eqs:Y}
1\le Y(t)
= C_0(1+t)f^3(x,a)\exp\big\{C_0 t f^2(x,a) \big\} \qquad \text{for each $t\in(0,T]$.}
\end{equation}
4. Since $B_i(t)\in(i,i+1)$ from \eqref{eqs:crb}, then $x<i<B_i(t)$ for all $t\in[0,T)$.
Using the Cauchy-Schwartz inequality, Lemma \ref{lemma:rbd}, estimate \eqref{eqs:entropy}, and Jensen's inequality, it follows that, for each $y\in[x,i+1]\subseteq (i-1,i+1]$ and $s\in[0,T)$,
\begin{align}\label{temp:esqrt}
&\Bignorm{\sqrt{e(B_i(s),s)}- \sqrt{e(y,s)}}
= \dfrac{1}{2}\Bignorm{\int_{y}^{B_i(s)} \dfrac{{\mathrm D}_x e}{\sqrt{e}}(z,s)\,{\mathrm d} z }\nonumber\\
&\le \dfrac{1}{2} \Big(\int_{0}^{k} \dfrac{r^{2m}\snorm{{\mathrm D}_x e}^2}{v e^2}(z,s)\,{\mathrm d} z\Big)^{\frac{1}{2}}\Big(\sup\limits_{z\in[x,i+1]}\frac{v}{r^{2m}}(z,s) \int_{i-1}^{i+1}e(z,s)\,{\mathrm d} z \Big)^{\frac{1}{2}}
\nonumber\\
&\le C_0\Big(a^{n}+nx\psi_{-}^{-1}(\dfrac{C_0}{x}) \Big)^{-\frac{m}{n}} \big(Q(s)V(x,s)\big)^{\frac{1}{2}},
\end{align}
where
\begin{equation}\label{temp:VJ}
V(x,s)\vcentcolon= \sup\limits_{z\in[x,i+1]}v(z,s),
\qquad Q(s) \vcentcolon= \int_{0}^{k}\dfrac{r^{2m}\snorm{{\mathrm D}_x e}^2}{v e^2}(z,s)\,{\mathrm d} z,
\end{equation}
and $Q(t)\in L^1(0,T)$ due to the entropy estimate \eqref{eqs:entropy}. According to the definition of $h(z,a)$ in \eqref{fhF}, the construction in \eqref{eqs:crb} implies that, for all $y\in[x,i+1]\subseteq(i-1,i+1]$,
\begin{equation}\label{eqs:eineq}
e(y,t) \le C_0 + C_0 h(x,a) Q(s)V(x,s)\qquad \text{for all $(y,s)\in[x,i+1]\times[0,T)$}.
\end{equation}
Substituting \eqref{eqs:eineq} and \eqref{eqs:Y} into \eqref{eqs:gron1}, it implies that, for all $y\in [x,i+1]\subseteq (i-1,i+1]$,
\begin{align*}
v(y,t)&\le v(y,t)Y(t) \\
&\le C_0(1+t)f^5(x,a)\exp\big\{C_0 t f^2(x,a)\big\}\Big\{1+t+\int_{0}^{t}h(x,a)Q(s)V(x,s){\mathrm d} s\Big\},
\end{align*}
where we used that $x\mapsto f(x,a)$ is decreasing.
Taking the supremum over $y\in[x , i+1]$, we have
\begin{align}\label{eqs:gronv}
V(x,t) \le C_0(1+t)f^5(x,a)\exp\big\{C_0 t f^2(x,a)\big\}\Big\{1+t+\int_{0}^{t}h(x,a)Q(s)V(x,s)\,{\mathrm d} s\Big\}.
\end{align}
Applying Gr\"onwall's inequality to \eqref{eqs:gronv}, we obtain that, for all $(x,t)\in[i-1,i]\times[0,T)$,
\begin{align*}
v(x,t) \le V(x,t) \le C_0(1+t)^2 f^{5}(x,a)\exp\Big\{C_0t f^2(x,a) + C_0(1+t)(f^5h)(x,a)\exp\{C_0 t f^2(x,a)\} \Big\}.
\end{align*}
In addition, for the lower bound of $v(x,t)$, it follows from \eqref{temp5} that
\begin{equation*}
v(x,t) = \dfrac{D(x,t)}{Y(t)E(x,t)} \dfrac{v_0(x)v(A_i(t),t)}{v_0(A_i(t))} \exp\Big\{\int_{0}^{t} p(x,s)\,{\mathrm d} s \Big\}
\ge C_0^{-1}\dfrac{D(x,t)}{Y(t)E(x,t)} .
\end{equation*}
By \eqref{eqs:crb}, \eqref{eqs:D}--\eqref{eqs:E}, and estimates \eqref{eqs:Y}, we find that, for all $(x,t)\in [i-1,i]\times [0,T]$,
\begin{align*}
v(x,t)\ge C_0 f^{-1} Y^{-1} E^{-1} \ge C_0 (1+t)^{-1} f^{-4}(x,a)\exp\{-C_0 t f^2(x,a)\} \Gamma(x,t,a).
\end{align*}
Since this is true for each $i=1,\dotsc, k-2$, it follows that, for all $(x,t)\in[0,k-2]\times[0,T]$,
\begin{equation*}
\begin{dcases}
v(x,t)\le C_0 (1+t)^2 f^{5}(x,a) \exp\Big\{C_0 t f^2(x,a) + C_0 (1+t) (f^5 h)(x,a) \exp\{C_0 t f^2(x,a)\} \Big\} ,\\
v(x,t) \ge C_0 f^{-4}\exp\{-C_0t f^2\} \Gamma(x,t,a).
\end{dcases}
\end{equation*}
\smallskip
5. From definition \eqref{fhF}, it follows that, for all $x\in[1,k-2]$,
\begin{align*}
f(x,a)\le\exp\big\{C_0(n\psi_{-}^{-1}(C_0))^{-\frac{2m}{n}}\big\},
\,\,\, h(x,a) \le ( n\psi_{-}^{-1}(C_0))^{-\frac{2m}{n}},
\,\,\, \Gamma(x,t,a)\ge\exp\big\{-\dfrac{mC_0t}{n\beta \psi_{-}^{-1}(C_0)} \big\},
\end{align*}
so that $C^{-1}(T)\le v(x,t)\le C(T)$ for all $(x,t)\in[1,k-2]\times[0,T]$. Since $x\mapsto x \psi_{-}^{-1}(\frac{C_0}{x})$ is positive and monotone increasing, then, for all $x\in[\varepsilon,k-2]$,
\begin{align*}
f(x,a)\le \exp\big\{C_0(n\varepsilon\psi_{-}^{-1}(\frac{C_0}{\varepsilon}))^{-\frac{m}{n}}\big\},\,\,\,
h(x,a)\le (n\varepsilon\psi_{-}^{-1}(\frac{C_0}{\varepsilon}))^{-\frac{m}{n}},\,\,\,
\Gamma(x,t,a) \ge \exp\big\{-\dfrac{mC_0t}{n\beta\varepsilon\psi_{-}^{-1}(\frac{C_0}{\varepsilon})} \big\},
\end{align*}
which yield that, for each $\varepsilon\in(0,1]$, $C^{-1}(\varepsilon)\le v(x,t) \le C(\varepsilon)$ for all $(x,t)\in[\varepsilon,k-2]\times[0,T]$.
\smallskip
6. Finally, we remark that, for the case when $x\in(k-2,k]$, we can choose $k\ge 3$ and $A_k(t), B_k(t)\in(k-1,k)$ by Lemma \ref{lemma:select}, and
then repeat the same argument presented above. From this, we conclude that $C^{-1}(T)\le v(x,t)\le C(T)$ for all $(x,t)\in(k-2,k]\times[0,T)$.
\end{proof}
\subsection{Exterior \texorpdfstring{$L^{\infty}\big(0,T;L^2\big)$}{LTL2}--estimates}\label{subsec:L2total}
To obtain the uniform estimates in $(a,k)\in(0,1)\times\mathbb{N}$, the derivations are restricted to domain $[\varepsilon,k]\times[0,T]$.
Thus, given fixed $\varepsilon>0$, we introduce the spatial cut-off function $g_\varepsilon(x)\in C^1$:
{\small
\begin{equation}\label{eqs:g}
g_{\varepsilon}(x)\vcentcolon=
\begin{dcases*}
0 & if $x\le \dfrac{\varepsilon}{2}$,\\
\dfrac{8}{\varepsilon^2}\big(x-\frac{\varepsilon}{2}\big)^2 & if $\dfrac{\varepsilon}{2} \le x\le \dfrac{3\varepsilon}{4}$,\\
1-\dfrac{8}{\varepsilon^2}(x-\varepsilon)^2 & if $\dfrac{3\varepsilon}{4} \le x\le \varepsilon$,\\
1 & if $x\ge \varepsilon$.
\end{dcases*}
\end{equation}
}
It can directly be verified that $g_\varepsilon(x)$ satisfies the following properties:
\begin{equation}\label{eqs:gprop}
{\mathrm {supp}} (g_{\varepsilon}^{\prime}) = [\frac{\varepsilon}{2},\varepsilon], \quad\,
0\le g_\varepsilon(x) \le 1, \quad g_\varepsilon^{\prime}(x)\ge 0, \quad
\snorm{g_{\varepsilon}^{\prime}(x)}^2\le \dfrac{32}{\varepsilon^2} g_{\varepsilon}(x) \qquad\,\mbox{for all $x\in[0,\infty)$}.
\end{equation}
Moreover, it is crucial that the estimates are independent of the approximation parameters $(a,k)\in(0,1)\times\mathbb{N}$. In order to achieve this,
the upper and lower bounds of $r(x,t)$ must be given careful consideration. Recall that, from Lemma \ref{lemma:rbd}, $r(x,t)$ satisfies
\begin{equation*}
n x\psi_{-}^{-1}(\frac{C_0}{x}) \le r^n(x,t) \le C_0(1 + x)
\qquad \text{for $(x,t)\in[0,k]\times[0,T]$}.
\end{equation*}
It follows that, for each $\varepsilon\in(0,1]$,
\begin{equation}\label{eqs:grbd}
\sup_{x\in[\frac{\varepsilon}{2},\varepsilon]}r(x,t)\le C(\varepsilon), \qquad \,\, \sup_{x\in[\varepsilon,k]}r^{-1}(x,t) \le C(\varepsilon).
\end{equation}
Using $g_\varepsilon(x)$ and Theorem \ref{thm:priori}\ref{item:prio0}--\ref{item:prio1},
our aim is to derive the $L^{\infty}(0,T;L^2)$--estimates listed in Theorem \ref{thm:priori}\ref{item:prio2}.
In order to do this, we first prove the following lemma, which is necessary for resolving the problematic boundary integrals for $x\in[\frac{\varepsilon}{2},\varepsilon]$.
\begin{lemma}[Exterior $L^{1}(0,T;L^{\infty})$--Estimates of $e$ and $u^2$]\label{lemma:euLinf}
For any fixed $\varepsilon>0$,
\begin{equation}\label{temp:euLinf}
\int_{0}^{T} \sup_{y\in [\varepsilon,k]} \big(u^2+e\big)(y,s)\, {\mathrm d} s \le C(\varepsilon).
\end{equation}
\end{lemma}
\begin{proof}
By Lemma \ref{lemma:select}, for each $(y,s)\in[\varepsilon,k]\times[0,T]$, there exists $B_i(s)\in[0,k]$
such that $\norm{y-B_i(s)}\le 1$ and $C_0^{-1} \le e(B_i(s),s) \le C_0$. By the same calculation in \eqref{temp:esqrt}, we have
\begin{align}
\snorm{\sqrt{e(y,s)} - \sqrt{e(B_i(s),s)}}^{2}
\le&\, \frac{1}{4}\sup_{z\in[\varepsilon,k]} \dfrac{v}{r^{2m}} (z,s) \Bignorm{\int_{B_i(s)}^{y} e(x,s)\,{\mathrm d} x }
\Bignorm{ \int_{B_i(s)}^{y} \dfrac{r^{2m}\,\snorm{{\mathrm D}_x e}^2}{ve^2}(x,s)\,{\mathrm d} x }\nonumber\\
\le &\, C(\varepsilon)\int_{0}^{k} \dfrac{r^{2m}\,\snorm{{\mathrm D}_x e}^2}{v e^2}(x,s)\,{\mathrm d} x, \label{4.25a}
\end{align}
where
we have used Theorem \ref{thm:priori}\ref{item:prio1}, \eqref{eqs:g}--\eqref{eqs:grbd},
and $\sbnorm{\psi(e)}_{L^{1}(0,k)}\le C_0$ from Theorem \ref{thm:priori}\ref{item:prio0} and Jensen's inequality.
Taking the supremum over $y\in[\varepsilon,k]$ on the left-hand side of \eqref{4.25a} and integrating in $t\in[0,T]$,
it follows from
Lemma \ref{lemma:select} and Theorem \ref{thm:priori}\ref{item:prio0} that
\begin{equation}\label{temp:infE}
\int_{0}^{T} \sup_{y\in [\varepsilon,k]} e(y,s)\,{\mathrm d} s
\le C_0 T + C(\varepsilon)\int_{0}^{T}\int_{0}^{k}\dfrac{r^{2m}\,\snorm{{\mathrm D}_x e}^2}{ve^2}
\,{\mathrm d} x {\mathrm d} s
\le C(\varepsilon).
\end{equation}
Next, notice that
\begin{align*}
\snorm{u{\mathrm D}_x u}=\snorm{r^{-m}u {\mathrm D}_x(r^m u)-m r^{-n}v u^2}
\le 2r^{-2m} \snorm{{\mathrm D}_x(r^m u)}^2 +2 m^2 r^{-2n} v^2 \snorm{u}^2.
\end{align*}
Then, by the Sobolev embedding theorem: $W^{1,1}(\varepsilon,k)\xhookrightarrow[]{}C^0(\varepsilon,k)$, and Theorem \ref{thm:priori}\ref{item:prio0}--\ref{item:prio1},
\begin{align}\label{eqs:uSob}
\sup_{y\in[\varepsilon,k]}u^2(y,t)
&\le C\int_{\varepsilon}^{k} u^2\,{\mathrm d} x + C\int_{\varepsilon}^{k}\snorm{u{\mathrm D}_x u}\,{\mathrm d} x
\nonumber\\
&\le C_0 + C \sup_{y\in[\varepsilon,k]}\Big(\dfrac{v}{r^{2m}}e
+\dfrac{v}{r^n}\Big)(y,s)\int_{\varepsilon}^{k} u^2(x,t)\,{\mathrm d} x
+ C\int_{\varepsilon}^{k} \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{ve}\,{\mathrm d} x\nonumber\\
&\le C(\varepsilon) + C(\varepsilon) \sup_{y\in[\varepsilon,k]}e(y,t) + C\int_{0}^{k} \dfrac{\snorm{{\mathrm D}_x (r^mu)}^2}{ve}\,{\mathrm d} x.
\end{align}
$\eqref{temp:euLinf}_2$ follows from integrating the above in $[0,T]$ and then using \eqref{temp:infE} and Theorem \ref{thm:priori}\ref{item:prio0}.
\end{proof}
\begin{lemma}[Exterior $L^{\infty}(0,T;L^2)$--Estimate of the Total Energy]\label{lemma:eL2}
For any fixed $\varepsilon>0$,
\begin{align}\label{eqs:eL2}
\sup\limits_{t\in[0,T]}\int_{\varepsilon}^{k}\big(\snorm{e-1}^2 + u^4\big)(x,t)\,{\mathrm d} x
+ \int_{0}^{T}\sup\limits_{x\in[\varepsilon,k]}e^2(x,s)\,{\mathrm d} s
+\int_{0}^{T}\int_{\varepsilon}^{k}\dfrac{r^{2m}}{v}
\big(\snorm{{\mathrm D}_x e}^2 + \snorm{u{\mathrm D}_x u}^2\big)\,{\mathrm d} x {\mathrm d} s \le C(\varepsilon).
\end{align}
\end{lemma}
\begin{proof} We divide the proof into four steps.
\smallskip
1. Denote the total energy density function $w\vcentcolon= \frac{1}{2}u^2+e-1$.
Then multiplying $\eqref{eqs:LFNS-k}_2$ with $u$ and adding it to $\eqref{eqs:LFNS-k}_3$ yield
\begin{equation}\label{TEtemp12}
{\mathrm D}_t w = {\mathrm D}_x\big(r^m u \widetilde{F}\big) -2m\mu {\mathrm D}_x(r^{m-1}u^2)
+ \kappa {\mathrm D}_x\big(\dfrac{r^{2m}}{v}{\mathrm D}_x e\big),
\end{equation}
where $\widetilde{F}\vcentcolon=\beta v^{-1}{\mathrm D}_x( r^m u)-p(v,e)$.
Multiplying \eqref{TEtemp12} with $g_\varepsilon w$ and then integrating in $x$, we have
\begin{align*}
\dfrac{1}{2}\dfrac{{\mathrm d}}{{\mathrm d} t}\int_{0}^k g_\varepsilon w^2\,{\mathrm d} x
&= -\int_{0}^{k} g_\varepsilon r^m u\widetilde{F} {\mathrm D}_x w\,{\mathrm d} x
+ 2m\mu \int_{0}^{k} g_\varepsilon r^{m-1}u^2 {\mathrm D}_x w\,{\mathrm d} x
- \kappa \int_{0}^{k} g_\varepsilon \dfrac{r^{2m}}{v} {\mathrm D}_x e {\mathrm D}_x w\,{\mathrm d} x \\
&\quad\, +\int_{0}^{k} g_\varepsilon^{\prime} \Big\{-r^m u w \widetilde{F}
+2m\mu r^{m-1}u^2 w
-\kappa \dfrac{r^{2m}}{v} w {\mathrm D}_x e\Big\}\,{\mathrm d} x \\
&=\vcentcolon \sum_{i=1}^{4} I_i.
\end{align*}
2. The terms $I_j, j=1,2,3$, can be estimated as follows:
\begin{align*}
I_1
=& -\beta \int_{0}^{k} g_\varepsilon \big\{\dfrac{r^{2m}}{v} u{\mathrm D}_x u {\mathrm D}_x e + m r^{m-1} u^3 {\mathrm D}_x u
+ m r^{m-1} u^2 {\mathrm D}_x e \big\}\,{\mathrm d} x \\
&+(\gamma-1)\int_{0}^{k}g_\varepsilon\dfrac{e}{v}r^m\big\{u^2{\mathrm D}_x u+u{\mathrm D}_xe\big\}\,{\mathrm d} x
-\beta\int_{0}^{k} g_\varepsilon\dfrac{r^{2m}}{v}\snorm{u {\mathrm D}_x u}^2\,{\mathrm d} x \\
\le &\, \big(\dfrac{5\beta^2}{2\kappa}-\dfrac{2\beta}{3}\big) \int_{0}^{k}g_\varepsilon\dfrac{\snorm{r^m u{\mathrm D}_x u}^2}{v}\,{\mathrm d} x +\dfrac{3\kappa}{10} \int_{0}^{k} g_\varepsilon \dfrac{\snorm{r^m{\mathrm D}_x e}^2}{v}\,{\mathrm d} x
+C\int_{0}^k g_\varepsilon \Big( \dfrac{v u^4}{r^2} + \dfrac{e^2u^2}{v} \Big)\,{\mathrm d} x,\\
I_2
=&\, 2m\mu \int_{0}^{k} g_\varepsilon r^{m-1} (u^3 {\mathrm D}_x u + u^2 {\mathrm D}_x e)\,{\mathrm d} x \\
\le& \,\dfrac{\beta}{6} \int_{0}^{k} g_\varepsilon\dfrac{r^{2m}}{v}\snorm{u {\mathrm D}_x u}^2\,{\mathrm d} x
+\dfrac{\kappa}{10} \int_{0}^{k} g_\varepsilon\dfrac{r^{2m}}{v}\snorm{{\mathrm D}_x e}^2\,{\mathrm d} x
+ C \int_{0}^{k} g_\varepsilon \dfrac{v}{r^2} u^4\,{\mathrm d} x,\\
I_3
=& - \kappa \int_{0}^{k} g_\varepsilon \dfrac{r^{2m}}{v} {\mathrm D}_x e\{{\mathrm D}_x e + u {\mathrm D}_x u\}\,{\mathrm d} x\\
\le & -\dfrac{9}{10}\kappa\int_{0}^{k}g_\varepsilon\dfrac{r^{2m}}{v}\snorm{{\mathrm D}_x e}^2\,{\mathrm d} x
+\dfrac{5\kappa}{2}\int_{0}^{k} g_\varepsilon\dfrac{r^{2m}}{v}\snorm{u {\mathrm D}_x u}^2\,{\mathrm d} x.
\end{align*}
Using \eqref{eqs:g}--\eqref{eqs:grbd} and Theorem \ref{thm:priori}\ref{item:prio1},
we have
\begin{align*}
I_4
=& \int_{\frac{\varepsilon}{2}}^{\varepsilon} g_\varepsilon^{\prime} \Big\{-r^m u w \Big(\beta\dfrac{ r^m}{v} {\mathrm D}_x u + \dfrac{m\beta}{r} u - (\gamma-1)\dfrac{e}{v}\Big) + 2m\mu r^{m-1}u^2 w -\kappa \dfrac{r^{2m}}{v} w {\mathrm D}_x e \Big\}\,{\mathrm d} x\\
\le& \int_{\frac{\varepsilon}{2}}^{\varepsilon} \dfrac{\varepsilon^2\snorm{g_{\varepsilon}^{\prime}}^2}{32}\dfrac{r^{2m}}{v}
\Big\{ \dfrac{\beta}{2}\snorm{u {\mathrm D}_xu}^2+\dfrac{\kappa}{4}\snorm{{\mathrm D}_x e}^2 \Big\}\,{\mathrm d} x
+ C\int_{\frac{\varepsilon}{2}}^{\varepsilon}\Big\{ \dfrac{\varepsilon^2\snorm{g_{\varepsilon}^{\prime}}^2}{32}\big(\dfrac{vu^4}{r^2}+\dfrac{e^2 u^2}{v}\big)
+ \dfrac{32}{\varepsilon^2}\dfrac{r^{2m}}{v}w^2 \Big\}\,{\mathrm d} x\\
\le& \int_{0}^{k}g_\varepsilon \dfrac{r^{2m}}{v}
\Big\{ \dfrac{\beta}{2}\norm{u {\mathrm D}_xu}^2 + \dfrac{\kappa}{4}\norm{{\mathrm D}_x e}^2 \Big\}\,{\mathrm d} x
+ C\int_{0}^{k}g_\varepsilon\Big\{\dfrac{v}{r^2}u^4 +\dfrac{e^2 u^2}{v}\Big\}\,{\mathrm d} x
+ C(\varepsilon)\int_{\frac{\varepsilon}{2}}^{\varepsilon} w^2\,{\mathrm d} x.
\end{align*}
Combining all the above estimates, it follows that
\begin{equation}\label{TEtemp13}
\begin{aligned}
\dfrac{1}{2}\dfrac{{\mathrm d}}{{\mathrm d} t}\int_{0}^k g_\varepsilon w^2\,{\mathrm d} x
\le& -\dfrac{\kappa}{4}\int_{0}^{k} g_\varepsilon \dfrac{r^{2m}}{v} \norm{{\mathrm D}_x e}^2 {\mathrm d} x + \dfrac{5(\beta^2+\kappa^2)}{2\kappa}\int_{0}^{k} g_\varepsilon \dfrac{r^{2m}}{v} \norm{u {\mathrm D}_x u}^2 {\mathrm d} x \\
&+ C \int_{0}^{k} g_\varepsilon \dfrac{v}{r^2} u^4\,{\mathrm d} x
+ C \int_{0}^{k} g_\varepsilon \dfrac{e^2 u^2}{v}\,{\mathrm d} x
+ C(\varepsilon)\mathcal{R}_{\varepsilon}(t),
\end{aligned}
\end{equation}
where $\mathcal{R}_{\varepsilon}(t)$ is the problematic boundary integral near the origin, given by
\begin{equation}\label{TEtemp20}
\mathcal{R}_\varepsilon(t)
\vcentcolon = \int_{\frac{\varepsilon}{2}}^{\varepsilon}\big((e-1)^2 + u^4\big)(x,t)\,{\mathrm d} x.
\end{equation}
3. Multiplying $g_\varepsilon u^3$ to the momentum equation $\eqref{eqs:LFNS-k}_2$ and integrating in $x\in[0,k]$, we obtain the following $L^4$-estimate for $u$:
\begin{align*}
\dfrac{1}{4}\dfrac{{\mathrm d}}{{\mathrm d} t}\int_{0}^{k}g_\varepsilon \norm{u}^4 {\mathrm d} x
&= -\beta\int_{0}^{k} g_\varepsilon\dfrac{1}{v} {\mathrm D}_x(r^m u){\mathrm D}_x(r^m u^3)\,{\mathrm d} x
+(\gamma-1) \int_{0}^{k} g_\varepsilon \dfrac{e}{v} {\mathrm D}_x(r^m u^3)\,{\mathrm d} x \\
&\quad + \int_{0}^{k} g_\varepsilon^{\prime} \Big\{\dfrac{(\gamma-1)e}{v} r^m u^3
-\dfrac{\beta}{v} {\mathrm D}_x(r^m u) r^m u^3\Big\}\,{\mathrm d} x\\
&=\vcentcolon \,\sum_{i=1}^3 J_i.
\end{align*}
The terms $J_1$ and $J_2$ can be estimated as follows:
\begin{align*}
J_1
=& -\beta \int_{0}^{k} g_\varepsilon \Big( 3\dfrac{\snorm{r^{m}u{\mathrm D}_x u}^2}{v}
+ 4m r^{m-1}u^3 {\mathrm D}_x u + m^2 \dfrac{u^4v}{r^2} \Big)\,{\mathrm d} x\\
\le&\, -\beta \int_{0}^{k} g_\varepsilon \dfrac{r^{2m}}{v}\snorm{u {\mathrm D}_x u}^2\,{\mathrm d} x
+ m^2 \beta \int_{0}^{k} g_\varepsilon \dfrac{v}{r^2} u^4 \,{\mathrm d} x,\\
J_2
\le&\, \dfrac{\beta}{4}\int_{0}^{k}g_\varepsilon \dfrac{\snorm{r^mu{\mathrm D}_x u}^2}{v}\,{\mathrm d} x + C \int_{0}^{k} g_\varepsilon \Big( \dfrac{e^2u^2}{v} + \dfrac{vu^4}{r^2} \Big)\,{\mathrm d} x.
\end{align*}
Using \eqref{eqs:g}--\eqref{eqs:grbd}
and Theorem \ref{thm:priori}\ref{item:prio1}, we have
\begin{align*}
J_3
\le &\, \dfrac{\gamma-1}{2} \int_{\frac{\varepsilon}{2}}^{\varepsilon} \dfrac{\varepsilon^2}{32}\snorm{g_\varepsilon^{\prime}}^2 \dfrac{e^2 u^2}{v}\,{\mathrm d} x
+\dfrac{\gamma-1}{2}\int_{\frac{\varepsilon}{2}}^{\varepsilon}\dfrac{32}{\varepsilon^2}\dfrac{r^{2m}}{v}u^4\,{\mathrm d} x
+ C \int_{\frac{\varepsilon}{2}}^{\varepsilon} \dfrac{1}{\varepsilon} r^{m-1} u^4\,{\mathrm d} x\\
&\, + \dfrac{\beta}{4} \int_{\frac{\varepsilon}{2}}^{\varepsilon} \dfrac{\varepsilon^2}{32}
\snorm{g_\varepsilon^{\prime}}^2 \dfrac{r^{2m}}{v} \snorm{u {\mathrm D}_x u}^2\,{\mathrm d} x
+\beta\int_{\frac{\varepsilon}{2}}^{\varepsilon}\dfrac{32}{\varepsilon^2} \dfrac{r^{2m}}{v} u^4\,{\mathrm d} x \\
\le &\, \dfrac{\beta}{4}\int_{0}^{k}g_\varepsilon\dfrac{r^{2m}}{v}\snorm{u {\mathrm D}_x u}^2\,{\mathrm d} x
+\dfrac{\gamma-1}{2}\int_{0}^{k} g_\varepsilon \dfrac{e^2 u^2 }{v}\,{\mathrm d} x
+ C(\varepsilon) \int_{\frac{\varepsilon}{2}}^{\varepsilon} u^4\,{\mathrm d} x.
\end{align*}
Combining all the above estimates, it follows that
\begin{equation}\label{TEtemp14}
\begin{aligned}
\dfrac{{\mathrm d}}{{\mathrm d} t}\int_{0}^{k}g_\varepsilon \dfrac{\snorm{u}^4}{4} {\mathrm d} x
\le &\, C\int_{0}^kg_\varepsilon\Big(\dfrac{e^2 u^2}{v}+\dfrac{vu^4}{r^2}\Big)\,{\mathrm d} x
-\dfrac{\beta}{2}\int_{0}^{k} g_\varepsilon \dfrac{\snorm{r^{m}u{\mathrm D}_x u}^2}{v}\,{\mathrm d} x
+ C(\varepsilon) \mathcal{R}_\varepsilon(t).
\end{aligned}
\end{equation}
4. Multiplying \eqref{TEtemp14} by $4\widetilde{C}\equiv \frac{10(\beta^2+\kappa^2)}{\kappa\beta}$ and adding it to \eqref{TEtemp13},
it follows from \eqref{eqs:entropy} that
\begin{align}\label{TEtemp15}
&\widetilde{C}\dfrac{{\mathrm d}}{{\mathrm d} t}\int_{0}^{k} g_\varepsilon u^4\,{\mathrm d} x
+ \dfrac{1}{2}\dfrac{{\mathrm d}}{{\mathrm d} t} \int_{0}^k g_\varepsilon w^2\,{\mathrm d} x
+ \int_{0}^{k} g_\varepsilon \dfrac{r^{2m}}{v} \Big\{ \dfrac{\kappa}{4}\snorm{{\mathrm D}_x e}^2 + \beta\widetilde{C} \snorm{u {\mathrm D}_x u}^2 \Big\}\,{\mathrm d} x
\nonumber\\
&\le C(\varepsilon) \mathcal{R}_\varepsilon(t) + C \int_{0}^{k} g_\varepsilon \Big(\dfrac{v}{r^2}u^4 +\dfrac{\norm{ue}^2}{v}\Big)\,{\mathrm d} x.
\end{align}
Using Theorem \ref{thm:priori}\ref{item:prio0}--\ref{item:prio1} and \eqref{eqs:g}--\eqref{eqs:grbd},
we have
\begin{align}\label{TEtemp16}
\int_{0}^{k} g_\varepsilon \Big(\dfrac{v}{r^2}u^4 +\dfrac{\norm{ue}^2}{v}\Big)\,{\mathrm d} x
\le C(\varepsilon) \int_{0}^{k} g_\varepsilon u^4\,{\mathrm d} x + C(\varepsilon) \sup_{y\in[\frac{\varepsilon}{2},k]}(g_\varepsilon\snorm{e}^2)(y,t).
\end{align}
By Lemma \ref{lemma:select}, for each $(x,t)\in[0,k]\times[0,T]$,
there exist both an integer $i\in\{0,\dotsc, k-1\}$ and a point $B_i(t)\in[0,k]$
such that $x, B_i(t) \in [i,i+1]$ and $C_0^{-1}\le e(B_i(t),t) \le C_0$.
Using Theorem \ref{thm:priori}\ref{item:prio0}, $\snorm{x-B_i(t)}\le 1$, and \eqref{eqs:g}--\eqref{eqs:grbd},
we have
\begin{align*}
&\snorm{(g_\varepsilon e)(x,t) - (g_\varepsilon e)(B_i(t),t)}^2
=\Big(\int_{B_i(t)}^{x} (g_\varepsilon {\mathrm D}_x e + g_\varepsilon^{\prime} e)(y,t)\,{\mathrm d} y \Big)^2 \\
&\le 2\Big(\int_{B_i(t)}^{x}g_\varepsilon\dfrac{r^{2m}}{v}\dfrac{\snorm{{\mathrm D}_x e}^2}{e^2}
\,{\mathrm d} y \Big) \Big( \int_{B_i(t)}^{x}g_\varepsilon\dfrac{v}{r^{2m}}e^2\,{\mathrm d} y \Big) + C(\varepsilon)\Big(\int_{B_i(t)}^{x}e(y,t)\,{\mathrm d} y\Big)^2\\
&\le 4\Big(\int_{0}^{k} \dfrac{r^{2m}}{v}\dfrac{\norm{{\mathrm D}_x e}^2}{e^2}\,{\mathrm d} y\Big) \Big\{ \Bignorm{\int_{B_i(t)}^{x}g_\varepsilon\dfrac{v}{r^{2m}}\snorm{e-1}^2\,{\mathrm d} y} + \Bignorm{\int_{B_i(t)}^{x}\dfrac{g_\varepsilon v}{r^{2m}}\,{\mathrm d} y}\Big\} + C(\varepsilon)\Big(\int_{i}^{i+1}e\,{\mathrm d} y\Big)^2\\
&= C(\varepsilon)\Big\{1+ Q(t)\Big(1+\Bignorm{\int_{B_i(t)}^{x}g_\varepsilon\snorm{e-1}^2\,{\mathrm d} y}\Big)\Big\},
\end{align*}
where $Q(t)$ is defined in \eqref{temp:VJ}.
Taking the supremum over $x\in[\frac{\varepsilon}{2},k]$, it follows from Theorem \ref{thm:priori}\ref{item:prio1} and \eqref{eqs:g}--\eqref{eqs:grbd}
that
\begin{equation}\label{TEtemp17}
\sup_{x\in[\frac{\varepsilon}{2},k]} (g_\varepsilon\snorm{e}^2)(x,t)
\le C(\varepsilon)\Big\{1 + Q(t) \Big(1 + \int_{0}^{k} g_\varepsilon\snorm{e-1}^2\,{\mathrm d} x\Big)\Big\}.
\end{equation}
Substituting \eqref{TEtemp17} into \eqref{TEtemp16} yields
\begin{equation}\label{TEtemp21}
\begin{aligned}
\int_{0}^{k} g_\varepsilon \Big(\dfrac{v}{r^2}u^4 +\dfrac{\snorm{ue}^2}{v}\Big)\,{\mathrm d} x
\le C(\varepsilon) \big( 1+ Q(t)\big) \Big( 1+ \int_{0}^{k} g_\varepsilon ( u^4 + w^2 )\,{\mathrm d} x \Big),
\end{aligned}
\end{equation}
where we have used:
$\snorm{u}^4 + w^2
\ge \dfrac{3}{4}\snorm{u}^4 + \dfrac{1}{2}(e-1)^2$.
Substituting \eqref{TEtemp21} into \eqref{TEtemp15}, it follows that
\begin{align*}
&\dfrac{{\mathrm d}}{{\mathrm d} t}\int_{0}^{k} g_\varepsilon (u^4 + w^2)\,{\mathrm d} x
+ \int_{0}^{k} g_\varepsilon \dfrac{r^{2m}}{v}\big(\snorm{{\mathrm D}_x e}^2
\snorm{u {\mathrm D}_x u}^2\big)\,{\mathrm d} x \\
&\le C(\varepsilon) \big( 1+ Q(t)\big) \Big( 1+ \int_{0}^{k} g_\varepsilon ( u^4 + w^2)\,{\mathrm d} x \Big) + C(\varepsilon) \mathcal{R}_{\varepsilon}(t) .
\end{align*}
Therefore, applying Gr\"onwall's inequality and using $Q(t)\in L^1(0,T)$,
we have
\begin{equation}\label{TEtemp18}
\begin{aligned}
\int_{0}^{k} \big(\snorm{u}^4 + \snorm{e-1}^2\big)(x,t)\,{\mathrm d} x
\le C(\varepsilon)\Big(1+ \int_{0}^{k} (u_0^4 + w_0^2)\,{\mathrm d} x
+ \int_{0}^{t} \mathcal{R}_\varepsilon\,{\mathrm d} s\Big).
\end{aligned}
\end{equation}
5. Using Lemma \ref{lemma:euLinf}, we claim that $\mathcal{R}_\varepsilon(t)$ defined in \eqref{TEtemp20} satisfies:
\begin{equation}\label{TEtemp19}
\int_{0}^t \mathcal{R}_\varepsilon(s)\,{\mathrm d} s
= \int_{0}^t \int_{\frac{\varepsilon}{2}}^{\varepsilon}\big(\snorm{e-1}^2 + \snorm{u}^4\big)\,{\mathrm d} x {\mathrm d} s \le C(\varepsilon).
\end{equation}
To show this claim, we first use Lemma \ref{lemma:euLinf} and Theorem \ref{thm:priori}\ref{item:prio0} to obtain
\begin{align*}
\int_{0}^{t} \mathcal{R}_\varepsilon(s)\,{\mathrm d} s
&\le C\int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},\varepsilon]}\snorm{e-1}(y,s)
\big(\int_{\frac{\varepsilon}{2}}^{\varepsilon}(e+1)(x,s)\,{\mathrm d} x\big)\, {\mathrm d} s
+C\int_{0}^t\sup_{y\in[\frac{\varepsilon}{2},\varepsilon]}u^2(y,s)\big(\int_{\frac{\varepsilon}{2}}^{\varepsilon}u^2(x,s)\,{\mathrm d} x\big)\,{\mathrm d} s\\
&\le C(\varepsilon)\Big(1+
\int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},\varepsilon]} e(y,s) \big(\int_{\frac{\varepsilon}{2}}^{\varepsilon} e(x,s)\,{\mathrm d} x\big)\,{\mathrm d} s\Big).
\end{align*}
It follows by Jensen's inequality and Theorem \ref{thm:priori}\ref{item:prio0} that
\begin{align*}
\psi( \dfrac{2}{\varepsilon} \int_{\frac{\varepsilon}{2}}^{\varepsilon} e(x,s)\,{\mathrm d} x )
\le \dfrac{2}{\varepsilon} \int_{\frac{\varepsilon}{2}}^{\varepsilon} \psi(e)(x,s)\,{\mathrm d} x
\le \dfrac{2C_0}{\varepsilon},
\end{align*}
so that
$$
\int_{\frac{\varepsilon}{2}}^{\varepsilon} e(x,s)\, {\mathrm d} x \le \dfrac{\varepsilon}{2} \psi^{-1}_{+} (\dfrac{2C_0}{\varepsilon}) \le C(\varepsilon).
$$
Substituting this back into the estimate of $\mathcal{R}_\varepsilon$
and using Lemma \ref{lemma:euLinf}, it follows that
\begin{align*}
\int_{0}^{t} \mathcal{R}_\varepsilon(s)\,{\mathrm d} s
\le C(\varepsilon) \Big(1+ \int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},\varepsilon]} e(y,s)\,{\mathrm d} s\Big)
\le C(\varepsilon).
\end{align*}
This prove the claim \eqref{TEtemp19}.
\medskip
6. Using Gr\"onwall's inequality and \eqref{TEtemp18}--\eqref{TEtemp19},
we have
\begin{align*}
&\int_{0}^{k} g_\varepsilon(x)(u^4 + w^2)(x,t)\,{\mathrm d} x
+ \int_{0}^{t}\int_{0}^{k} g_\varepsilon \dfrac{r^{2m}}{v} \snorm{{\mathrm D}_x e}^2
\snorm{u {\mathrm D}_x u}^2\,{\mathrm d} x{\mathrm d} s
\le \int_{0}^{k} ( u_0^4 + w_0^2 )\, {\mathrm d} x + C(\varepsilon).
\end{align*}
By the inequality:
$\snorm{u}^4 + w^2
\ge \frac{3}{4}\snorm{u}^4 + \frac{1}{2}(e-1)^2$,
we see that $\sbnorm{(e-1,u^2)}_{L^{\infty}(0,T;L^2(\varepsilon,k))}\le C(\varepsilon)$.
Finally, substituting the above estimate into \eqref{TEtemp17}, we obtain
\begin{align*}
\int_{0}^{T}\sup\limits_{x\in[\varepsilon,k]} e^2(x,t)\,{\mathrm d} t
&\le \int_{0}^{T}\sup_{x\in[\frac{\varepsilon}{2},k]} (g_\varepsilon e^2)(x,t)\,{\mathrm d} t
\le C(\varepsilon).
\end{align*}
This completes the proof.
\end{proof}
\begin{lemma}[Exterior $L^{\infty}(0,T;L^2)$--Estimate of the Specific Volume]\label{lemma:vL2}
For each $\varepsilon>0$,
\begin{align*}
\sup\limits_{t\in[0,T]}\int_{\varepsilon}^{k} \snorm{v-1}^2(x,t)\,{\mathrm d} x \le C(\varepsilon).
\end{align*}
\end{lemma}
\begin{proof}
Multiplying $\eqref{eqs:LFNS-k}_1$ with $g_\varepsilon(v-1)$ and integrating in $[0,k]$, then, by Theorem \ref{thm:priori}\ref{item:prio1},
\begin{align*}
\dfrac{1}{2} {\mathrm D}_t \int_{0}^k g_\varepsilon (v-1)^2\,{\mathrm d} x
= \int_{0}^k g_\varepsilon (v-1) {\mathrm D}_x ( r^m u )\,{\mathrm d} x
\le C(\varepsilon) \int_{0}^{k} g_\varepsilon (v-1)^2\,{\mathrm d} x
+ \dfrac{1}{2} \int_{0}^k g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}\,{\mathrm d} x.
\end{align*}
Applying Lemma \ref{lemma:eL2} and Gr\"onwall's inequality, the lemma is proved.
\end{proof}
The following lemma is necessary for the high-order estimates of $(u,e)$
in \S \ref{subsec:HRue}
below.
\begin{lemma}[Exterior $L^{2}$--estimate of ${\mathrm D}_x u$] \label{lemma:L2Dxu}
For each $\varepsilon>0$,
\begin{equation}\label{eqs:dissip}
\int_{0}^{T}\int_{\varepsilon}^{k} \Big(\dfrac{\snorm{{\mathrm D}_x (r^m u)}^2}{v} +\dfrac{r^{2m}\snorm{{\mathrm D}_x u}^2}{v} \Big)\,{\mathrm d} x {\mathrm d} s \le C(\varepsilon).
\end{equation}
\end{lemma}
\begin{proof}
Multiplying $\eqref{eqs:LFNS-k}_1$ by $g_\varepsilon$ and integrating over $[0,k]$,
we have
\begin{equation}\label{temp17}
\int_{0}^{k} g_\varepsilon(x) {\mathrm D}_t (v(x,t)-1)\,{\mathrm d} x
= - \int_{0}^{k} g_{\varepsilon}^{\prime}(x) (r^m u)(x,t)\,{\mathrm d} x.
\end{equation}
Multiplying $\eqref{eqs:LFNS-k}_2$ by $g_\varepsilon u$ and integrating over $x\in[0,k]$,
using Theorem \ref{thm:priori}\ref{item:prio0}--\ref{item:prio1}, Lemma \ref{lemma:eL2},
and \eqref{temp17}, it follows that
\begin{align}\label{461-di1}
&\frac{1}{2}\dfrac{{\mathrm d} }{{\mathrm d} t}\int_{0}^{k} g_{\varepsilon}\snorm{u}^2\,{\mathrm d} x
+\beta\int_{0}^{k} g_{\varepsilon}\dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x \nonumber\\
&= (\gamma-1)\int_{0}^{k} g_{\varepsilon} \dfrac{ {\mathrm D}_x(r^m u)}{v} (e-1)\,{\mathrm d} x + (\gamma-1)\int_{0}^{k} g_{\varepsilon} \dfrac{ {\mathrm D}_t v}{v} {\mathrm d} x
+ \int_{0}^{k} g_{\varepsilon}^{\prime}(x) \dfrac{r^mu}{v}\big((\gamma-1)e
-\beta {\mathrm D}_x(r^m u)\big)\,{\mathrm d} x \nonumber\\
&=\vcentcolon \sum_{i=1}^3 I_i.
\end{align}
The term $I_1$ is estimated by using Theorem \ref{thm:priori}\ref{item:prio1}
and Lemma \ref{lemma:eL2}:
\begin{equation}\label{461-di2}
\begin{split}
&I_1 \vcentcolon
= (\gamma-1)\int_{0}^{k} g_{\varepsilon} \dfrac{ {\mathrm D}_x( r^m u)}{v} (e-1)\,{\mathrm d} x
\le \dfrac{\beta}{4} \int_{0}^k g_\varepsilon \dfrac{\snorm{{\mathrm D}_x( r^m u)}^2}{v} {\mathrm d} x+ C(\varepsilon).
\end{split}
\end{equation}
$I_2$ is estimated by using $\eqref{eqs:LFNS-k}_1$, Theorem \ref{thm:priori}\ref{item:prio0},
and \eqref{eqs:g}--\eqref{eqs:grbd}:
\begin{align}\label{461-di3}
\dfrac{I_2}{\gamma-1} &\vcentcolon = \int_{0}^{k} g_{\varepsilon} \dfrac{ {\mathrm D}_t v}{v} {\mathrm d} x
= \int_{0}^{k} g_{\varepsilon} {\mathrm D}_t (\log v-v+1)\,{\mathrm d} x
- \int_{0}^{k} g_{\varepsilon}^{\prime} r^m u\,{\mathrm d} x\nonumber\\
&\le \int_{0}^{k} g_{\varepsilon} {\mathrm D}_t (\log v -v +1)\,{\mathrm d} x
+ \Big(\int_{\frac{\varepsilon}{2}}^{\varepsilon} r^{2m} {\mathrm d} x\Big)^{\frac{1}{2}}
\Big( \int_{\frac{\varepsilon}{2}}^{\varepsilon} u^2 {\mathrm d} x \Big)^{\frac{1}{2}}\nonumber\\
&\le \int_{0}^{k} g_{\varepsilon} {\mathrm D}_t ( \log v -v +1 )\,{\mathrm d} x + C(\varepsilon).
\end{align}
$I_3$ is estimated by using Lemma \ref{lemma:eL2} and \eqref{eqs:g}--\eqref{eqs:grbd}:
\begin{align}\label{461-di4}
I_3
\le &\, \dfrac{(\gamma-1)^2}{2}\sup\limits_{y\in[\frac{\varepsilon}{2},\varepsilon]} (v^{-2}r^{2m})(y,t)
\int_{\frac{\varepsilon}{2}}^{\varepsilon} u^2 {\mathrm d} x
+ (\gamma-1) \Big( \int_{\frac{\varepsilon}{2}}^{\varepsilon} v^{-2} r^{2m} {\mathrm d} x \Big)^{\frac{1}{2}}
\Big( \int_{\frac{\varepsilon}{2}}^{\varepsilon} u^2 {\mathrm d} x \Big)^{\frac{1}{2}}
\nonumber\\
& + \dfrac{1}{2}\int_{\frac{\varepsilon}{2}}^{\varepsilon} e^2 {\mathrm d} x
+ \dfrac{\beta}{4} \int_{\frac{\varepsilon}{2}}^{\varepsilon}\dfrac{\varepsilon^2}{32}\snorm{g_\varepsilon^{\prime}(x)}^2
\dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x + \dfrac{32\beta}{\varepsilon^2}\sup\limits_{y\in[\frac{\varepsilon}{2},\varepsilon]}\dfrac{r^{2m}}{v}(y,t)
\int_{\frac{\varepsilon}{2}}^{\varepsilon} u^2\,{\mathrm d} x\nonumber\\
\le&\, \dfrac{\beta}{4} \int_{0}^{k}g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x + C(\varepsilon).
\end{align}
According to \eqref{461-di1}--\eqref{461-di4}, we have
\begin{align*}
\frac{1}{2}&\int_{0}^{k}g_\varepsilon(x)u^2(x,t)\,{\mathrm d} x
+\dfrac{\beta}{2}\int_{0}^{t}\int_{0}^{k}g_{\varepsilon}\dfrac{\snorm{{\mathrm D}_x (r^m u)}^2}{v} {\mathrm d} x {\mathrm d} s\\
&\le C(\varepsilon) + (\gamma-1) \int_{0}^{k} g_\varepsilon (\log v - v +1)\,{\mathrm d} x
+ (\gamma-1) \int_{0}^{k} g_\varepsilon (v_0 - \log v_0 - 1)\,{\mathrm d} x.
\end{align*}
Since $\log v - v +1 \le 0$ and $C_0 \le v_0(x) \le C_0^{-1}$,
it implies from Taylor's theorem that
\begin{equation}\label{eqs:disp}
\dfrac{\beta}{2}\int_{0}^{t}\int_{0}^{k}g_{\varepsilon}\dfrac{\snorm{{\mathrm D}_x (r^m u)}^2}{v} {\mathrm d} x {\mathrm d} s \le C(\varepsilon) + (\gamma-1) C_0 \int_{0}^{k} (v_0-1)^2 {\mathrm d} x \le C(\varepsilon).
\end{equation}
The other term in \eqref{eqs:dissip} is estimated by rewriting \eqref{eqs:disp}
and using Theorem \ref{thm:priori}\ref{item:prio0}--\ref{item:prio1}.
\end{proof}
\subsection{Exterior \texorpdfstring{$L^{\infty}\big(0,T;H^1\big)$}{LTH1}--estimates of the velocity and internal energy}\label{subsec:HRue}
We now derive the high-regularity estimates for $(u,e)$ in Theorem \ref{thm:priori}\ref{item:prio2}. For simplicity, we define
\begin{equation*}
\Lambda(v,e) \equiv (v-1)^2 + (e-1)^2, \quad \,
\sigma(t) \vcentcolon = \min\{1,t\}.
\end{equation*}
It follows from Lemma \ref{lemma:eL2}-\ref{lemma:vL2} that $\bnorm{\Lambda(v,e)}_{L^{\infty}(0,T;L^1(\varepsilon,k))}\le C(\varepsilon)$. Furthermore, we introduce the effective viscosity flux $F$, which is defined by
\begin{align*}
F\vcentcolon= \beta\dfrac{{\mathrm D}_x(r^m u)}{v} - p(v,e) +(\gamma-1) = \beta\dfrac{{\mathrm D}_x(r^m u)}{v} - (\gamma-1)\dfrac{e-v}{v} .
\end{align*}
Using
Lemmas \ref{lemma:eL2}--\ref{lemma:L2Dxu},
we have the following observation on $F$:
\begin{lemma}[Estimates on Effective Viscosity Flux]\label{lemma:estF}
\begin{equation}\label{eqs:SL2F}
\begin{split}
&\int_{0}^{k} g_\varepsilon(x) F^2(x,t)\,{\mathrm d} x
\le C(\varepsilon)\Big(1+\int_{0}^{k} g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}{\mathrm d} x\Big),\\
&\int_{0}^{T}\int_{0}^{k} g_\varepsilon(x) F^2(x,t)\,{\mathrm d} x {\mathrm d} t \le C(\varepsilon),\\
&\int_{0}^T \int_{0}^{k} g_\varepsilon(x) (r^{2m}\snorm{{\mathrm D}_x F}^2)(x,t)\, {\mathrm d} x{\mathrm d} t
\le \int_{0}^{T}\int_{0}^{k} g_\varepsilon(x) \snorm{{\mathrm D}_t u}^2 (x,t)\,{\mathrm d} x {\mathrm d} t.
\end{split}
\end{equation}
\end{lemma}
\begin{proof}
It follows from the definition of $F$ and the momentum equation $\eqref{eqs:LFNS-k}_2$ that
\begin{align*}
&F
= \beta \dfrac{{\mathrm D}_x(r^m u)}{v} - (\gamma-1)\dfrac{e-1}{v} + (\gamma-1) \dfrac{v-1}{v},\\
&{\mathrm D}_t u = \beta r^m {\mathrm D}_x\Big(\dfrac{{\mathrm D}_x(r^m u)}{v}\Big) - r^m {\mathrm D}_x( p - (\gamma-1) ) = r^m {\mathrm D}_x F,
\end{align*}
which, along with Theorem \ref{thm:priori}\ref{item:prio1} and Lemmas \ref{lemma:eL2}--\ref{lemma:L2Dxu}, yields that
\begin{align*}
\int_{0}^{k}g_\varepsilon(x) F^2 (x,t)\,{\mathrm d} x
\le& \, 2\beta^2 \int_{0}^{k}g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v^2} {\mathrm d} x + 4(\gamma-1)^2 \int_{0}^{k}g_\varepsilon \dfrac{\Lambda(v,e)}{v^2} {\mathrm d} x \\
\le& \,C(\varepsilon) \int_{0}^{k}g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x + C(\varepsilon),
\end{align*}
and $\eqref{eqs:SL2F}_2$--$\eqref{eqs:SL2F}_3$ obviously.
\end{proof}
\begin{lemma}[Intermediate Step]\label{lemma:uH1int}
\begin{equation}\label{yinlis.9}
\int_{0}^{T}\int_{0}^{k} \sigma g_\varepsilon \snorm{{\mathrm D}_t u}^2\,{\mathrm d} x {\mathrm d} s
+\sup_{0\le s\le T}\int_{0}^k\Big(\sigma g_\varepsilon
\dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}\Big)(x,s)\,{\mathrm d} x
\le C(\varepsilon)\Big\{1
+ \Big(\int_{0}^T\int_{0}^k \sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\,{\mathrm d} x {\mathrm d} s\Big)^{\frac{1}{2}}\Big\}.
\end{equation}
\end{lemma}
\begin{proof} We divide the proof into six steps.
\smallskip
1. Multiplying equations $\eqref{eqs:LFNS-k}_2$ with $g_\varepsilon \sigma {\mathrm D}_t u$ and
then integrating yield
\begin{align*}
&\int_{0}^{t}\int_{0}^{k} \sigma g_\varepsilon \snorm{{\mathrm D}_t u}^2\,{\mathrm d} x {\mathrm d} s
+ \dfrac{\beta}{2}\int_{0}^{k} \Big(\sigma g_\varepsilon\dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}\Big)(x,t)\,{\mathrm d} x \\
&= \dfrac{\beta}{2}\int_{0}^{t}\int_{0}^k \sigma^{\prime} g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x {\mathrm d} s + m\beta \int_{0}^{t} \int_{0}^{k} \sigma g_\varepsilon {\mathrm D}_x ( r^{m-1} u^2 ) \dfrac{{\mathrm D}_x(r^m u)}{v} {\mathrm d} x {\mathrm d} s \\
&\quad -\int_{0}^t \int_{0}^{k} \sigma g_\varepsilon r^m {\mathrm D}_t u {\mathrm D}_x p\,{\mathrm d} x {\mathrm d} s
+ \dfrac{\beta}{2} \int_{0}^{t} \int_{0}^{k} \sigma g_\varepsilon {\mathrm D}_t (v^{-1}) \snorm{{\mathrm D}_x (r^m u)}^2\,{\mathrm d} x {\mathrm d} s\\
&\quad-\beta \int_{0}^t \int_{0}^{k} \sigma g_\varepsilon^{\prime} r^m {\mathrm D}_t u \dfrac{{\mathrm D}_x(r^m u)}{v} {\mathrm d} x {\mathrm d} s \\
&=\vcentcolon \sum_{i=1}^5 I_i.
\end{align*}
$I_1$ is estimated by using $\sigma(t)=\min\{ 1,t \}$
and Lemma \ref{lemma:L2Dxu} as
\begin{align*}
I_1 \vcentcolon = \dfrac{\beta}{2}\int_{0}^{t}\int_{0}^k \sigma^{\prime} g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x {\mathrm d} s \le \dfrac{\beta}{2}\int_{0}^{1}\int_{0}^k g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x {\mathrm d} s \le C(\varepsilon).
\end{align*}
$I_2$ is estimated by using Lemmas \ref{lemma:eL2}--\ref{lemma:L2Dxu}, Theorem \ref{thm:priori}\ref{item:prio1}, and \eqref{eqs:g}--\eqref{eqs:grbd} as
\begin{align*}
I_2
&= 2m\beta \int_{0}^{t} \int_{0}^{k} \sigma g_\varepsilon r^{m-1} u {\mathrm D}_x u \dfrac{{\mathrm D}_x(r^m u)}{v} {\mathrm d} x {\mathrm d} s + m(m-1)\beta\int_{0}^{t}\int_{0}^{k} \sigma g_\varepsilon \dfrac{v}{r^2} u^2 \dfrac{{\mathrm D}_x(r^m u)}{v} {\mathrm d} x {\mathrm d} s \\
&\le m\beta \int_{0}^{t} \sup_{y\in[\frac{\varepsilon}{2},k]} r^{-2}(y,s)
\big(\int_{0}^{k} g_\varepsilon \dfrac{r^{2m}}{v}\snorm{ u {\mathrm D}_x u}^2\,{\mathrm d} x\big)\,{\mathrm d} s
+ m\beta \int_{0}^{t} \int_{0}^{k}g_\varepsilon\dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x {\mathrm d} s \\
&\quad + \dfrac{m(m-1)\beta}{2}\Big\{
\int_{0}^{t} \sup_{y\in[\frac{\varepsilon}{2},k]}\dfrac{v}{r^4}(y,t)\big(\int_{0}^{k} g_\varepsilon u^4\,{\mathrm d} x\big)\,{\mathrm d} s
+ \int_{0}^{t}\int_{0}^{k} g_\varepsilon \dfrac{\snorm{{\mathrm D}_x (r^m u)}^2}{v} {\mathrm d} x {\mathrm d} s
\Big\}\le C(\varepsilon).
\end{align*}
2. Using $p-\gamma+1 = \frac{(\gamma-1)(e-v)}{v}$, the term $I_3$ is separated into six parts:
\begin{align*}
\dfrac{I_3}{\gamma-1}
&= \int_{0}^{k} \Big(\sigma g_\varepsilon {\mathrm D}_x(r^m u) \dfrac{e-v}{v}\Big)(x,t)\,{\mathrm d} x
- \int_{0}^t \int_{0}^{k} \sigma^{\prime} g_\varepsilon {\mathrm D}_x(r^m u) \dfrac{e-v}{v}
{\mathrm d} x {\mathrm d} s \\
&\quad - \int_{0}^t\int_{0}^{k} \sigma g_\varepsilon {\mathrm D}_x(r^m u) {\mathrm D}_t \Big(\dfrac{e-v}{v}\Big)\,{\mathrm d} x {\mathrm d} s -\int_{0}^t \int_{0}^{k} \sigma g_\varepsilon {\mathrm D}_x (r^{m-1} u^2) \dfrac{e-v}{v}{\mathrm d} x {\mathrm d} s\\
&\quad + \int_{0}^t \int_{0}^{k} \sigma g_\varepsilon^{\prime} {\mathrm D}_t (r^m u) \dfrac{e-v}{v}{\mathrm d} x {\mathrm d} s
- \int_{0}^t \int_{0}^{k} \sigma g_\varepsilon^{\prime} r^{m-1} u^2 \dfrac{e-v}{v}{\mathrm d} x {\mathrm d} s\\
&= \vcentcolon \sum_{i=1}^6 I_3^{(i)}.
\end{align*}
$I_3^{(1)}$ is estimated by using Theorem \ref{thm:priori}\ref{item:prio1} and Lemma \ref{lemma:eL2}--\ref{lemma:L2Dxu}: \begin{align*}
I_3^{(1)}&\vcentcolon = \int_{0}^{k} \sigma g_\varepsilon {\mathrm D}_x(r^m u) \dfrac{e-v}{v} {\mathrm d} x
= \int_{0}^{k} \sigma g_\varepsilon {\mathrm D}_x(r^m u) \big(\dfrac{e-1}{v} + \dfrac{1-v}{v}\big) \,{\mathrm d} x\\
&\le \dfrac{1}{\beta} \int_{0}^{k} \sigma g_\varepsilon \dfrac{\Lambda(v,e)}{v} {\mathrm d} x + \dfrac{\beta}{4}\int_{0}^{k} \sigma g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x \\
&\le C(\varepsilon) + \dfrac{\beta}{4}\int_{0}^{k} \Big(\sigma g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}\Big)(x,t)\,{\mathrm d} x.
\end{align*}
$I_3^{(2)}$ is estimated by using the definition of $\sigma$, Theorem \ref{thm:priori}\ref{item:prio1}, and Lemma \ref{lemma:eL2}--\ref{lemma:L2Dxu}:
\begin{align*}
I_3^{(2)}
&\vcentcolon= -\int_{0}^t \int_{0}^{k} \sigma^{\prime} g_\varepsilon {\mathrm D}_x(r^m u) \dfrac{e-v}{v} {\mathrm d} x {\mathrm d} s
= -\int_{0}^t\int_{0}^{k}\sigma^{\prime} g_\varepsilon {\mathrm D}_x(r^m u)
\Big( \dfrac{e-1}{v} + \dfrac{1-v}{v} \Big) {\mathrm d} x {\mathrm d} s\\
&\le \int_{0}^{1} \int_{0}^k g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x {\mathrm d} s + \dfrac{1}{4}\int_{0}^{1}\int_{0}^{k} g_\varepsilon \dfrac{\Lambda(v,e)}{v} {\mathrm d} x {\mathrm d} s \le C(\varepsilon).
\end{align*}
$I_3^{(3)}$ is estimated by using Theorem \ref{thm:priori}\ref{item:prio1} and Lemma \ref{lemma:L2Dxu}:
\begin{align*}
I_3^{(3)}
&= -\int_{0}^t\int_{0}^{k} \sigma g_\varepsilon {\mathrm D}_x(r^m u) \Big\{\dfrac{{\mathrm D}_t e}{v} - \dfrac{e}{v^2}{\mathrm D}_x(r^m u)\Big\}\,{\mathrm d} x {\mathrm d} s\\
&\le C(\varepsilon)\Big(\int_{0}^t \int_{\frac{\varepsilon}{2}}^k \dfrac{\snorm{{\mathrm D}_x( r^m u)}^2}{v}{\mathrm d} x {\mathrm d} s\Big)^{\frac{1}{2}}\Big(\int_{0}^{t}\int_{0}^k \sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\,{\mathrm d} x {\mathrm d} s\Big)^{\frac{1}{2}}\\
&\quad + C(\varepsilon)\int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},k]} e(y,s)
\big(\int_{0}^{k}\sigma g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x\big)\,{\mathrm d} s \\
&\le C(\varepsilon)\Big(\int_{0}^{t}\int_{0}^k \sigma^2 g_\varepsilon^2 \norm{{\mathrm D}_t e}^2 {\mathrm d} x {\mathrm d} s\Big)^{\frac{1}{2}}
+ C(\varepsilon)\int_{0}^t\Big(\sup_{y\in[\frac{\varepsilon}{2},k]} e(y,s)
\int_{0}^{k}\sigma g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x\Big){\mathrm d} s.
\end{align*}
$I_3^{(4)}$ is estimated by using Theorem \ref{thm:priori}\ref{item:prio1} and Lemmas \ref{lemma:eL2}--\ref{lemma:vL2}:
\begin{align*}
I_3^{(4)}
&=-\int_{0}^t\int_{0}^{k}\sigma g_\varepsilon \Big\{ 2 r^{m-1} u {\mathrm D}_x u + (n-2) \dfrac{v}{r^2} u^2\Big\}\Big(\dfrac{e-1}{v}+\dfrac{1-v}{v}\Big)\,{\mathrm d} x {\mathrm d} s\\
&\le \int_{0}^{t}\int_{0}^{k} \sigma g_\varepsilon \dfrac{r^{2m}}{v}\snorm{u{\mathrm D}_x u}^2\,{\mathrm d} x {\mathrm d} s
+ C(\varepsilon)\int_{0}^{t}\int_{0}^{k} g_\varepsilon
\big\{ u^4 + \Lambda(v,e) \big\}\,{\mathrm d} x {\mathrm d} s \le C(\varepsilon).
\end{align*}
For $I_3^{(5)}$,
we use Theorem \ref{thm:priori}\ref{item:prio1},
Lemmas \ref{lemma:eL2}--\ref{lemma:vL2},
and \eqref{eqs:g}--\eqref{eqs:grbd}
to obtain
\begin{align*}
I_3^{(5)}
&=\int_{0}^t \int_{0}^{k} \sigma g_\varepsilon^{\prime}\big(m r^{m-1} u^2 + r^m {\mathrm D}_t u\big)
\Big(\dfrac{e-1}{v}+\dfrac{1-v}{v} \Big)\,{\mathrm d} x {\mathrm d} s\\
&\le C\Big( \int_{0}^t\int_{\frac{\varepsilon}{2}}^{\varepsilon} \sigma \dfrac{\varepsilon^2}{32}\snorm{g_\varepsilon^{\prime}}^2 \snorm{{\mathrm D}_t u}^2
\,{\mathrm d} x {\mathrm d} s \Big)^{\frac{1}{2}}
\Big( \int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},\varepsilon]} \dfrac{r^{2m}}{v^2}(y,s)
\big(\int_{\frac{\varepsilon}{2}}^{\varepsilon} \dfrac{32}{\varepsilon^2}
\Lambda(v,e)\,{\mathrm d} x\big)\,{\mathrm d} s\Big)^{\frac{1}{2}}\\
&\quad + C\Big( \int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},\varepsilon]} r^{-2}(y,s)
\big(\int_{\frac{\varepsilon}{2}}^{\varepsilon} u^4 {\mathrm d} x\big)\,{\mathrm d} s \Big)^{\frac{1}{2}}\Big(\int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},\varepsilon]}\dfrac{r^{2m}}{v^2}(y,s)\big(\int_{\frac{\varepsilon}{2}}^{\varepsilon}\Lambda(v,e)
\,{\mathrm d} x\big)\,{\mathrm d} s \Big)^{\frac{1}{2}}\\
&\le \dfrac{1}{8}\int_{0}^t\int_{0}^k\sigma g_\varepsilon\snorm{{\mathrm D}_t u}^2\,{\mathrm d} x{\mathrm d} s +C(\varepsilon),
\end{align*}
and $I_3^{(5)}$
is estimated by using Theorem \ref{thm:priori}\ref{item:prio1}, Lemma \ref{lemma:eL2}--\ref{lemma:vL2}, and \eqref{eqs:g}--\eqref{eqs:grbd}.
\begin{align*}
I_3^{(5)}
&\le \dfrac{1}{2}\int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},\varepsilon]} \dfrac{r^{2(m-1)}}{v^2}(y,s)
\big(\int_{\frac{\varepsilon}{2}}^{\varepsilon} \snorm{g_\varepsilon^{\prime}} u^4\,{\mathrm d} x\big)\,{\mathrm d} s
+ \dfrac{1}{2} \int_{0}^t \int_{\frac{\varepsilon}{2}}^{\varepsilon} \snorm{g_\varepsilon^{\prime}} \Lambda(v,e)
\,{\mathrm d} x {\mathrm d} s \le C(\varepsilon).
\end{align*}
Combining the estimates of $I_3^{(1)}$--$I_3^{(6)}$,
it follows that
\begin{align*}
I_3
\le &\, \dfrac{\beta}{4} \int_{0}^k\Big(\sigma g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}\Big)(x,t)\,{\mathrm d} x
+\dfrac{1}{8}\int_{0}^{t}\int_{0}^{k}\sigma g_\varepsilon\snorm{{\mathrm D}_t u}^2 {\mathrm d} x {\mathrm d} s
+ C(\varepsilon)\\
&+ C(\varepsilon)\Big(\int_{0}^{t}\int_{0}^k \sigma g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2
{\mathrm d} x {\mathrm d} s\Big)^{\frac{1}{2}}
+ C(\varepsilon)\int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},k]} e(y,s)
\big(\int_{0}^{k}\Big(\sigma g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}\Big)(x,s)\, {\mathrm d} x\big) {\mathrm d} s.
\end{align*}
4. $I_4$ is rewritten into three parts by using the continuity equation $\eqref{eqs:LFNS-k}_1$ and the definition of $F$:
\begin{align*}
I_4
= & -\dfrac{1}{2\beta} \int_{0}^t \int_{0}^k \sigma g_\varepsilon F^2 {\mathrm D}_x(r^m u) \,{\mathrm d} x {\mathrm d} s
-\dfrac{\gamma-1}{\beta} \int_{0}^t \int_{0}^k \sigma g_\varepsilon F \dfrac{e-v}{v} {\mathrm D}_x(r^m u)\,{\mathrm d} x {\mathrm d} s \\
& -\dfrac{(\gamma-1)^2}{2\beta} \int_{0}^t \int_{0}^k \sigma g_\varepsilon
\Big(\dfrac{e-v}{v}\Big)^2 {\mathrm D}_x(r^m u)\,{\mathrm d} x {\mathrm d} s\\
\vcentcolon=&\, \sum_{i=1}^3 I_4^{(i)}.
\end{align*}
$I_4^{(1)}$ is estimated by using integration by parts, Lemma \ref{lemma:estF},
and \eqref{eqs:g}--\eqref{eqs:grbd}.
\begin{align*}
\beta I_4^{(1)}
&= \int_{0}^t \int_{0}^k \sigma g_\varepsilon u F r^m {\mathrm D}_x F\,{\mathrm d} x {\mathrm d} s
+\dfrac{1}{2}\int_{0}^t\int_{0}^k\sigma g_\varepsilon^{\prime} F^2 r^m u\,{\mathrm d} x{\mathrm d} s\\
&\le 2 \int_{0}^{t}\int_{0}^{k} \sigma g_\varepsilon u^2 F^2\,{\mathrm d} x {\mathrm d} s
+ \dfrac{1}{8} \int_{0}^{t}\int_{0}^{k} \sigma g_\varepsilon r^{2m} \snorm{{\mathrm D}_x F}^2\,{\mathrm d} x {\mathrm d} s\\
&\quad +\dfrac{1}{4}\int_{0}^t \int_{\frac{\varepsilon}{2}}^{\varepsilon}\sigma \dfrac{\varepsilon^2}{32}\snorm{g_\varepsilon^{\prime}}^2
\snorm{u}^2 F^2\,{\mathrm d} x {\mathrm d} s
+ \dfrac{1}{4}\int_{0}^t \int_{\frac{\varepsilon}{2}}^{\varepsilon} \dfrac{32}{\varepsilon^2} r^{2m} F^2\,{\mathrm d} x {\mathrm d} s \\
&\le \dfrac{1}{8} \int_{0}^{t}\int_{0}^{k} \sigma g_\varepsilon \snorm{{\mathrm D}_t u}^2\,{\mathrm d} x {\mathrm d} s
+C(\varepsilon) + C(\varepsilon)\int_{0}^{t} \sup_{y\in[\frac{\varepsilon}{2},k]}\snorm{u}^2(y,s)
\big(\int_{0}^k \sigma g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x\big){\mathrm d} s.
\end{align*}
$I_4^{(2)}$ is estimated by using Theorem \ref{thm:priori}\ref{item:prio1}
and
Lemmas \ref{lemma:eL2}--\ref{lemma:estF}:
\begin{align*}
\dfrac{\beta}{\gamma-1}I_4^{(2)}
&\le \int_{0}^t \int_{0}^k \sigma g_\varepsilon F^2\,{\mathrm d} x {\mathrm d} s
+ \int_{0}^{t} \sup_{y\in[\frac{\varepsilon}{2},k]} \dfrac{(e-v)^2}{v}
\big(\int_{0}^{k} \sigma g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x\big)\,{\mathrm d} s\\
&\le C(\varepsilon) + C(\varepsilon)\int_{0}^{t}\sup_{y\in[\frac{\varepsilon}{2},k]} \snorm{e(y,s)}^2 \big(\int_{0}^{k} \sigma g_\varepsilon
\dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x\big)\,{\mathrm d} s.
\end{align*}
$I_4^{(3)}$ is estimated by using Theorem \ref{thm:priori}\ref{item:prio1}
and Lemma \ref{lemma:eL2}--\ref{lemma:L2Dxu}:
\begin{align*}
\dfrac{2\beta}{(\gamma-1)^2}I_4^{(3)}
&\le 2 \int_{0}^t \int_{0}^k \sigma g_\varepsilon \dfrac{\Lambda(v,e)}{v^2} \norm{{\mathrm D}_x(r^m u)}\, {\mathrm d} x {\mathrm d} s \\
&\le \dfrac{(\gamma-1)^2}{2\beta} \int_{0}^t \int_{0}^k \sigma g_\varepsilon \dfrac{\Lambda(v,e)^2}{v^3} {\mathrm d} x {\mathrm d} s + \dfrac{(\gamma-1)^2}{2\beta} \int_{0}^t\int_{0}^k \sigma g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x {\mathrm d} s \\
&\le C(\varepsilon) + C(\varepsilon) \int_{0}^{t}
\big(\sup_{y\in[\frac{\varepsilon}{2},k]} \snorm{e-1}^2(y,s)+1\big)
\big(\int_{0}^{k} \sigma g_\varepsilon \snorm{e-1}^2\,{\mathrm d} x\big)\,{\mathrm d} s \\
&\le C(\varepsilon) + C(\varepsilon) \int_{0}^{t} \sup_{y\in[\frac{\varepsilon}{2},k]} e^2(y,s)\,{\mathrm d} s\le C(\varepsilon).
\end{align*}
Combining the estimates for $I_4^{(1)}$--$I_4^{(3)}$, it follows that
\begin{align*}
I_4 \le \dfrac{1}{8} \int_{0}^{t}\int_{0}^{k} \sigma g_\varepsilon \snorm{{\mathrm D}_t u}^2{\mathrm d} x {\mathrm d} s
+ C(\varepsilon)
+C(\varepsilon) \int_{0}^t\sup_{y\in[\varepsilon/2,k]} ( e^2 + u^2)(y,s)
\big(\int_{0}^{k} \sigma g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x\big)\,{\mathrm d} s.
\end{align*}
5. $I_5$
is estimated by using Theorem \ref{thm:priori}\ref{item:prio1}, Lemmas \ref{lemma:eL2}--\ref{lemma:L2Dxu},
and \eqref{eqs:g}--\eqref{eqs:grbd}:
\begin{align*}
I_5
&\le \beta^2 \int_{0}^{t} \sup_{y\in[\frac{\varepsilon}{2},\varepsilon]} \dfrac{r^{2m}}{v}(y,s)
\big(\int_{\frac{\varepsilon}{2}}^{\varepsilon} \dfrac{32}{\varepsilon^2} \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}{\mathrm d} x\big)\,{\mathrm d} s
+ \dfrac{1}{4}\int_{0}^t \int_{\frac{\varepsilon}{2}}^{\varepsilon}\sigma\dfrac{\varepsilon^2}{32}\snorm{g_\varepsilon^{\prime}}^2
\snorm{{\mathrm D}_t u}^2\,{\mathrm d} x{\mathrm d} t \\
&\le \dfrac{1}{4}\int_{0}^t \int_{\frac{\varepsilon}{2}}^{\varepsilon} \sigma g_\varepsilon \snorm{{\mathrm D}_t u}^2\,{\mathrm d} x {\mathrm d} t + C(\varepsilon).
\end{align*}
\smallskip
6. Combining the estimates of $I_1$--$I_5$
together, it follows that
\begin{align*}
&\dfrac{1}{2}\int_{0}^{t}\int_{0}^{k} \sigma g_\varepsilon \snorm{{\mathrm D}_t u}^2\,{\mathrm d} x {\mathrm d} s
+ \dfrac{\beta}{4} \int_{0}^k \Big(\sigma g_\varepsilon\dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}\Big)(x,t)\,{\mathrm d} x\\
&\le C(\varepsilon) + C(\varepsilon)\Big(\int_{0}^{t}\int_{0}^k \sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\,{\mathrm d} x {\mathrm d} s \Big)^{\frac{1}{2}}\\
&\quad
+ C(\varepsilon) \int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},k]}(e+e^2+u^2)(y,s)
\Big(\int_{0}^{k}\big(\sigma g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}\big)(x,s)\,{\mathrm d} x\Big)\,{\mathrm d} s.
\end{align*}
Applying Gr\"onwall's inequality, we have
\begin{align*}
&\dfrac{1}{2}\int_{0}^{t}\int_{0}^{k} \sigma g_\varepsilon \snorm{{\mathrm D}_t u}^2\,{\mathrm d} x {\mathrm d} s
+ \dfrac{\beta}{4} \int_{0}^k
\Big(\sigma g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}\Big)(x,t)\,{\mathrm d} x\\
&\le C(\varepsilon)\Big(1+ \Big(\int_{0}^t\int_{0}^k \sigma^2 g_\varepsilon^2
\snorm{{\mathrm D}_t e}^2\,{\mathrm d} x {\mathrm d} s\Big)^{\frac{1}{2}}\Big)\Big(1+ \big(\int_{0}^t h_\varepsilon(s)\,{\mathrm d} s\big)
\exp\big(\int_{0}^t h_{\varepsilon}(s)\,{\mathrm d} s\big)\Big),
\end{align*}
where $h_\varepsilon(s)\vcentcolon=C(\varepsilon)\sup_{y\in[\frac{\varepsilon}{2},k]}(e+e^2+u^2)(y,s)$.
By Lemmas \ref{lemma:euLinf}--\ref{lemma:eL2}, we obtain \eqref{yinlis.9}.
\end{proof}
\begin{lemma}[$L^{\infty}(0,T;L^2(\varepsilon,k))$--Estimates of ${\mathrm D}_x u$ and ${\mathrm D}_x e$]\label{lemma:euH1}
For each $\varepsilon>0$,
\begin{align}
&\int_{0}^{t}\int_{\varepsilon}^{k} \big(\sigma^2 \snorm{{\mathrm D}_t e}^2+\sigma g_\varepsilon \snorm{{\mathrm D}_t u}^2\big)
\,{\mathrm d} x {\mathrm d} s
+ \sup_{ 0\le s\le t}\int_{\varepsilon}^k \Big(\sigma^2 \dfrac{r^{2m}}{v} \snorm{{\mathrm D}_x e}^2+\sigma \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}\Big)(x,s)\,{\mathrm d} x
\le C(\varepsilon).\label{eqs:euH1}
\end{align}
\end{lemma}
\begin{proof} We divide the proof into six steps.
\smallskip
1. Rewriting the energy equation $\eqref{eqs:LFNS-k}_3$ in terms of the effective
viscosity flux $F$, then multiplying both sides by $\sigma^2(t) g_\varepsilon^2 {\mathrm D}_t e$, and integrating by parts, it follows that
\begin{align*}
&\int_{0}^{t}\int_{0}^{k} \sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\,{\mathrm d} x {\mathrm d} s
+\dfrac{\kappa}{2}\int_{0}^k\Big(\sigma^2 g_\varepsilon^2\dfrac{r^{2m}}{v}\snorm{{\mathrm D}_x e}^2\Big)(x,t)\,{\mathrm d} x\\
&= \kappa\int_{0}^{t} \int_{0}^k \sigma\sigma^{\prime}(s) g_\varepsilon^2 \dfrac{r^{2m}}{v}
\snorm{{\mathrm D}_x e}^2\,{\mathrm d} x {\mathrm d} s
+ m\kappa\int_{0}^{t} \int_{0}^k \sigma^2 g_\varepsilon^2 \dfrac{r^{2m-1}u}{v}
\snorm{{\mathrm D}_x e}^2\,{\mathrm d} x{\mathrm d} s \\
&\quad + \int_{0}^{t}\int_{0}^k \sigma^2 g_\varepsilon^2 \mathcal{A} {\mathrm D}_t e\,{\mathrm d} x {\mathrm d} s -\dfrac{\kappa}{2}\int_{0}^{t} \int_{0}^k \sigma^2 g_\varepsilon^2 \dfrac{r^{2m}}{v^2} {\mathrm D}_x(r^m u) \snorm{{\mathrm D}_x e}^2\,{\mathrm d} x {\mathrm d} s \\
&\quad +\int_{0}^{t}\int_{0}^{k} \sigma^2 g_\varepsilon^2 F {\mathrm D}_x(r^m u) {\mathrm D}_t e\,{\mathrm d} x {\mathrm d} s
- \kappa \int_{0}^t \int_{0}^{k} 2\sigma^2 g_\varepsilon g_\varepsilon^{\prime} \dfrac{r^{2m}}{v}
{\mathrm D}_t e {\mathrm D}_x e\,{\mathrm d} x {\mathrm d} s\\
&\vcentcolon= \sum_{i=1}^6 I_i,
\end{align*}
where $\mathcal{A}\vcentcolon=-(\gamma-1) {\mathrm D}_x(r^m u)- 4m\mu\dfrac{u}{r}{\mathrm D}_x(r^m u)
+ 2m n\mu \dfrac{v}{r^2} u^2$.
By Lemma \ref{lemma:eL2}, we have
\begin{equation*}
I_1 \vcentcolon = \kappa \int_{0}^{t} \int_{0}^k \sigma\sigma^{\prime}(s) g_\varepsilon^2
\dfrac{r^{2m}}{v} \snorm{{\mathrm D}_x e}^2\,{\mathrm d} x {\mathrm d} s
\le \kappa \int_{0}^{1} \int_{\frac{\varepsilon}{2}}^k \dfrac{r^{2m}}{v} \snorm{{\mathrm D}_x e}^2\,{\mathrm d} x {\mathrm d} s \le C(\varepsilon).
\end{equation*}
$I_2$ is estimated by using \eqref{eqs:g}--\eqref{eqs:grbd}
and Lemma \ref{lemma:eL2}:
\begin{align*}
I_2
&\le 2m\kappa\int_{0}^{t} \sup_{y\in[\frac{\varepsilon}{2},k]}\dfrac{u}{r}(y,s)
\big(\int_{0}^k \sigma^2 g_\varepsilon^2 \dfrac{r^{2m}}{v}\,\snorm{{\mathrm D}_x e}^2\,{\mathrm d} x\big) {\mathrm d} s\\
&\le C(\varepsilon) + C(\varepsilon) \int_{0}^{t} \sup_{y\in[\frac{\varepsilon}{2},k]}u^2(y,s)
\big(\int_{0}^k \sigma^2 g_\varepsilon^2 \dfrac{r^{2m}}{v} \snorm{{\mathrm D}_x e}^2\,{\mathrm d} x\big)\,{\mathrm d} s.
\end{align*}
\smallskip
2. $I_3$ is estimated by using Theorem \ref{thm:priori}\ref{item:prio1} and
Lemmas \ref{lemma:eL2}--\ref{lemma:L2Dxu}:
\begin{equation}\label{HEtemp4}
\begin{aligned}
I_3
&=\int_{0}^{t}\int_{0}^k \sigma^2 g_\varepsilon^2
\Big\{-(\gamma-1) {\mathrm D}_x(r^m u) - 4m\mu\dfrac{u}{r}{\mathrm D}_x(r^m u) + 2m n \mu \dfrac{v}{r^2} u^2\Big\} {\mathrm D}_t e\,{\mathrm d} x {\mathrm d} s \\
&\le \dfrac{1}{16} \int_{0}^{t}\int_{0}^{k}\sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\,{\mathrm d} x {\mathrm d} s
+ C(\varepsilon) \int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},k]}(1+u^2)(y,s)
\int_{0}^{k}\sigma^2 g_\varepsilon^2 \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x {\mathrm d} s \\
&\quad + C(\varepsilon)\int_{0}^{T}\int_{0}^{k}\sigma^2 g_\varepsilon^2 u^4\,{\mathrm d} x {\mathrm d} t\\
&\le \dfrac{1}{16}\int_{0}^t \int_{0}^k \sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\, {\mathrm d} x {\mathrm d} t
+C(\varepsilon)\Big(1+\int_{0}^{t}\sup_{y\in[\frac{\varepsilon}{2},k]}u^2(y,s)
\big(\int_{0}^k \sigma^2 g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}{\mathrm d} x\big)\,{\mathrm d} s\Big).
\end{aligned}
\end{equation}
In fact, by Lemmas \ref{lemma:euLinf} and \ref{lemma:uH1int}, it follows that
\begin{align*}
\int_{0}^{t} \sup_{y\in[\frac{\varepsilon}{2},k]} u^2(y,s)
\big(\int_{0}^k \sigma^2 g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}{\mathrm d} x\big)\,{\mathrm d} s
&\le \sup_{ 0\le s \le t}\int_{0}^k\Big(\sigma g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}\Big)(x,s)\,
{\mathrm d} x\int_{0}^{t} \sup_{y\in[\frac{\varepsilon}{2},k]} u^2(y,\widetilde{t}\,)\,{\mathrm d} \widetilde{t}\\
&\le C(\varepsilon) + C(\varepsilon)\Big(\int_{0}^t\int_{0}^k\sigma^2 g_\varepsilon^2\snorm{{\mathrm D}_t e}^2\,{\mathrm d} x {\mathrm d} s\Big)^{\frac{1}{2}}\\
&\le \dfrac{1}{16} \int_{0}^t \int_{0}^k \sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\, {\mathrm d} x {\mathrm d} s + C(\varepsilon).
\end{align*}
Substituting the above estimate into \eqref{HEtemp4} leads to
\begin{align*}
I_3 \le \dfrac{1}{8}\int_{0}^t\int_{0}^k\sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\,{\mathrm d} x {\mathrm d} s+C(\varepsilon).
\end{align*}
\smallskip
3.
$I_4$ is first rewritten in terms of $F$ and then estimated by using Theorem \ref{thm:priori}\ref{item:prio1}:
\begin{equation*}
\begin{aligned}
I_4 \vcentcolon
&=-\dfrac{\kappa}{2\beta}\int_{0}^t \int_{0}^k \sigma^2 g_\varepsilon^2
\Big(F + (\gamma-1)\dfrac{e-v}{v} \Big)\dfrac{\snorm{r^m {\mathrm D}_x e}^2}{v}\,{\mathrm d} x {\mathrm d} s \\
&\le \dfrac{\kappa}{2\beta}\int_{0}^t
\sup_{y\in[\frac{\varepsilon}{2},k]} \snorm{\sqrt{g_\varepsilon} F}(y,s)
\big(\int_{0}^k \sigma^2 g_\varepsilon^{\frac{3}{2}} \dfrac{r^{2m}}{v} \snorm{{\mathrm D}_x e}^2\,{\mathrm d} x)\,{\mathrm d} s + C(\varepsilon) \\
&\quad + C(\varepsilon)\int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},k]}e(y,s)
\big(\int_{0}^k \sigma^2 g_\varepsilon^2 \dfrac{r^{2m}}{v} \snorm{{\mathrm D}_x e}^2\,{\mathrm d} x\big)\,{\mathrm d} s.
\end{aligned}
\end{equation*}
By Lemma \ref{lemma:estF} and the Sobolev embedding theorem: $W^{1,1}(\varepsilon,k) \xhookrightarrow[]{} C^0(\varepsilon,k)$, it follows that
\begin{equation}\label{HEtemp2}
\begin{aligned}
\sup_{y\in[\frac{\varepsilon}{2},k]} \snorm{\sqrt{g_\varepsilon} F}^2(y,s)
&\le C \int_{\frac{\varepsilon}{2}}^k\big( g_\varepsilon F^2 + 2 g_\varepsilon \snorm{F {\mathrm D}_x F} + 2 g_\varepsilon^{\prime} F^2 \big) (x,s)\,{\mathrm d} x \\
&\le \int_{\frac{\varepsilon}{2}}^k \big( C(\varepsilon) F^2 + g_\varepsilon r^{2m} \snorm{{\mathrm D}_x F}^2 \big)\,{\mathrm d} x \\
&\le \int_{\frac{\varepsilon}{2}}^{k} \Big(C(\varepsilon)\dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}
+ g_\varepsilon \snorm{{\mathrm D}_t u}^2 \Big)\,{\mathrm d} x + C(\varepsilon),
\end{aligned}
\end{equation}
so that, using Lemma \ref{lemma:L2Dxu}
and the Cauchy-Schwartz inequality, we have
\begin{align*}
&\dfrac{\kappa}{\beta}\int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},k]} \snorm{g_\varepsilon^{\frac{1}{2}} F}(y,s)
\big(\int_{0}^k \sigma^2 g_\varepsilon^{\frac{3}{2}} \dfrac{r^{2m}}{v} \snorm{{\mathrm D}_x e}^2 {\mathrm d} x\big){\mathrm d} s\\
&\le C(\varepsilon)\int_{0}^t \Big\{ \sigma^2 +\int_{\frac{\varepsilon}{2}}^{k}
\sigma^2 \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}\,{\mathrm d} x
+ \int_{\frac{\varepsilon}{2}}^{k} \sigma^2 g_\varepsilon \snorm{{\mathrm D}_t u}^2\,{\mathrm d} x \Big\} {\mathrm d} s \\
&\quad + C(\varepsilon) \int_{0}^{t}\sigma^2
\Big( \int_{0}^k g_\varepsilon^{\frac{3}{2}} \dfrac{r^{2m}}{v}\snorm{{\mathrm D}_x e}^2\,{\mathrm d} x \Big)^2 {\mathrm d} s\\
&\le \int_{0}^t \int_{0}^{k} \sigma g_\varepsilon \snorm{{\mathrm D}_t u}^2\,{\mathrm d} x {\mathrm d} s + C(\varepsilon)
+C(\varepsilon)\int_{0}^t\Big(\int_{0}^k g_\varepsilon\dfrac{r^{2m}}{v}\snorm{{\mathrm D}_x e}^2\,{\mathrm d} x \Big)
\Big(\int_{0}^k\sigma^2g_\varepsilon^2 \dfrac{r^{2m}}{v}\snorm{{\mathrm D}_x e}^2\,{\mathrm d} x\Big){\mathrm d} s.
\end{align*}
Using the above estimate, Lemma \ref{lemma:uH1int}, and the Cauchy-Schwartz inequality,
one has
\begin{align*}
I_4
&\le C(\varepsilon) + \int_{0}^t \int_{0}^{k} \sigma g_\varepsilon \snorm{{\mathrm D}_t u}^2\,{\mathrm d} x {\mathrm d} s
+ C(\varepsilon)\int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},k]}e(y,s)
\big(\int_{0}^k \sigma^2 g_\varepsilon^2 \dfrac{r^{2m}}{v} \snorm{{\mathrm D}_x e}^2\,{\mathrm d} x\big) {\mathrm d} s\\
&\quad +C(\varepsilon) \int_{0}^t \Big(\int_{0}^k g_\varepsilon\dfrac{r^{2m}}{v}\snorm{{\mathrm D}_x e}^2\,{\mathrm d} x \Big) \Big(\int_{0}^k \sigma^2 g_\varepsilon^2 \dfrac{r^{2m}}{v}\snorm{{\mathrm D}_x e}^2\,{\mathrm d} x\Big) {\mathrm d} s,\\
&\le \dfrac{1}{8}\int_{0}^t \int_{0}^{k} \sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\,{\mathrm d} x {\mathrm d} s + C(\varepsilon)
+ C(\varepsilon)\int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},k]}e(y,s)
\big(\int_{0}^k \sigma^2 g_\varepsilon^2 \dfrac{r^{2m}}{v} \snorm{{\mathrm D}_x e}^2\, {\mathrm d} x\big){\mathrm d} s\\
&\quad +C(\varepsilon) \int_{0}^t \Big(\int_{0}^k g_\varepsilon\dfrac{r^{2m}}{v}\snorm{{\mathrm D}_x e}^2\,{\mathrm d} x \Big)\Big(\int_{0}^k\sigma^2 g_\varepsilon^2\dfrac{r^{2m}}{v}\snorm{{\mathrm D}_x e}^2\,{\mathrm d} x\Big){\mathrm d} s.
\end{align*}
\smallskip
4. $I_5$ is first estimated as follows:
\begin{align}\label{HEtemp3}
\begin{aligned}
I_5
&\le 4\int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},k]} (g_\varepsilon v \snorm{F}^2)(y,s)
\big(\int_{0}^k \sigma^2 g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x\big){\mathrm d} s
+ \dfrac{1}{16}\int_{0}^t \int_{0}^k \sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\,{\mathrm d} x {\mathrm d} s.
\end{aligned}
\end{align}
By the same argument as \eqref{HEtemp2}, we have
\begin{align*}
\sup_{y\in[\frac{\varepsilon}{2},k]}(g_\varepsilon \snorm{F}^2)(y,s)
\le C(\varepsilon) + C(\varepsilon)\int_{\frac{\varepsilon}{2}}^{k} \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x
+ C(\varepsilon)\Big( \int_{0}^{k}g_\varepsilon F^2\,{\mathrm d} x \Big)^{\frac{1}{2}}
\Big(\int_{0}^k g_\varepsilon \snorm{{\mathrm D}_t u}^2\,{\mathrm d} x\Big)^{\frac{1}{2}},
\end{align*}
which, along with Theorem \ref{thm:priori}\ref{item:prio1} and Lemma \ref{lemma:L2Dxu}, implies that
\begin{align*}
&4\int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},k]}(g_\varepsilon v \snorm{F}^2)(y,s)
\big(\int_{0}^k \sigma^2 g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x\big){\mathrm d} s\\
&\le C(\varepsilon)
+ C(\varepsilon)\sup_{0\le s \le t}\int_{0}^k
\Big(\sigma g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}\Big)(x,s)\,{\mathrm d} x\\
&\quad + C(\varepsilon)\Big( \int_{0}^t\int_{0}^{k} \sigma g_\varepsilon F^2 \Big)^{\frac{1}{2}}
\Big(\int_{0}^t\int_{0}^k \sigma g_\varepsilon \snorm{{\mathrm D}_t u}^2\Big)^{\frac{1}{2}}
\sup_{ 0\le s \le t}\int_{0}^k \Big(\sigma g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v}\Big)(x,s)\,{\mathrm d} x.
\end{align*}
Then applying Lemmas \ref{lemma:estF}--\ref{lemma:uH1int} on the above
and using the Cauchy-Schwartz inequality yield
\begin{align*}
&4\int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},k]} (g_\varepsilon v \snorm{F}^2)(y,s)
\big(\int_{0}^k \sigma^2 g_\varepsilon \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{v} {\mathrm d} x\big) {\mathrm d} s\\
&\le C(\varepsilon)
+C(\varepsilon)\Big\{1+\Big(\int_{0}^t\int_{0}^k\sigma^2g_\varepsilon^2\snorm{{\mathrm D}_t e}^2\,{\mathrm d} x{\mathrm d} s\Big)^{\frac{1}{2}}\Big\}^{\frac{1}{2}}
\Big\{ 1+\Big(\int_{0}^t\int_{0}^k \sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\,
{\mathrm d} x {\mathrm d} s\Big)^{\frac{1}{2}} \Big\}\\
&\le \dfrac{1}{16}\int_{0}^t\int_{0}^k \sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\,{\mathrm d} x {\mathrm d} s + C(\varepsilon).
\end{align*}
Substituting the above into \eqref{HEtemp3}, we obtain
\begin{align*}
I_5 \le \dfrac{1}{8}\int_{0}^t\int_{0}^k \sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\,{\mathrm d} x {\mathrm d} s + C(\varepsilon).
\end{align*}
\smallskip
5. $I_6$
is estimated by using Theorem \ref{thm:priori}\ref{item:prio1}, Lemma \ref{lemma:eL2}, and \eqref{eqs:g}--\eqref{eqs:grbd}:
\begin{align*}
I_6
&\le \dfrac{1}{8} \int_{0}^t \int_{\frac{\varepsilon}{2}}^{\varepsilon} \sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\,{\mathrm d} x {\mathrm d} s
+ 8\kappa^2 \int_{0}^t \sup_{y\in[\frac{\varepsilon}{2},\varepsilon]}\dfrac{r^{2m}}{v}(y,s)
\big(\int_{\frac{\varepsilon}{2}}^{\varepsilon} \sigma^2 \snorm{g_\varepsilon^{\prime}}^2 \dfrac{r^{2m}}{v}
\snorm{{\mathrm D}_x e}^2\,{\mathrm d} x\big)\,{\mathrm d} s\\
&\le \dfrac{1}{8} \int_{0}^t \int_{0}^{k} \sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\,{\mathrm d} x {\mathrm d} s + C(\varepsilon).
\end{align*}
\smallskip
6. Combining
the estimates for $I_1$--$I_6$ together, we have
\begin{align}\label{HEtemp5}
&\dfrac{1}{2}\int_{0}^{t}\int_{0}^{k} \sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\,{\mathrm d} x {\mathrm d} s
+\dfrac{\kappa}{2}\int_{0}^k\Big(\sigma^2 g_\varepsilon^2\dfrac{r^{2m}}{v}\snorm{{\mathrm D}_x e}^2\Big)(x,t)\,{\mathrm d} x \nonumber\\
&\le C(\varepsilon) + C(\varepsilon) \int_{0}^t \widetilde{h}_\varepsilon(s)
\big(\int_{0}^k\sigma^2 g_\varepsilon^2 \dfrac{r^{2m}}{v}\snorm{{\mathrm D}_x e}^2\,{\mathrm d} x\big)\,{\mathrm d} s,
\end{align}
where $\widetilde{h}_\varepsilon(s) \vcentcolon
= \int_{0}^k \big(g_\varepsilon\dfrac{r^{2m}}{v}\norm{{\mathrm D}_x e}^2\big)(x,s)\,{\mathrm d} x
+\sup_{y\in[\frac{\varepsilon}{2},k]}(u^2+e)(y,s)$.
Then it follows from Lemmas \ref{lemma:euLinf}--\ref{lemma:eL2}
and the Gr\"onwall inequality that
\begin{align*}
\dfrac{1}{2}\int_{0}^{t}\int_{0}^{k} \sigma^2 g_\varepsilon^2 \snorm{{\mathrm D}_t e}^2\,{\mathrm d} x {\mathrm d} s
+\dfrac{\kappa}{2}\int_{0}^k\Big(\sigma^2 g_\varepsilon^2\dfrac{r^{2m}}{v}\snorm{{\mathrm D}_x e}^2\Big)(x,t)\,
{\mathrm d} x\le C(\varepsilon),
\end{align*}
which, along with Lemma \ref{lemma:uH1int}, yields \eqref{eqs:euH1}.
This completes the proof.
\end{proof}
\begin{corollary}\label{coro:eubd}
For all $(x,t) \in [\varepsilon,k]\times[0,T]$,
\begin{equation}\label{eqs:ubd}
\sigma^{\frac{1}{4}}(t)\snorm{u(x,t)}+\sigma^{\frac{1}{2}}(t) e(x,t)\le C(\varepsilon).
\end{equation}
\end{corollary}
\begin{proof}
First, repeating the same calculation \eqref{eqs:uSob} in the proof of Lemma \ref{lemma:euLinf}, one has
\begin{align*}
\sigma^{\frac{1}{2}}(t)\sup_{y\in[\varepsilon,k]}u^2(y,t)
\le C(\varepsilon) \int_{\varepsilon}^{k} u^2(x,t)\,{\mathrm d} x + C \int_{\varepsilon}^{k}\sigma(t) \dfrac{\snorm{{\mathrm D}_x(r^mu)}^2}{v}(x,t) {\mathrm d} x \le C(\varepsilon),
\end{align*}
where we have used Lemma \ref{lemma:euH1}.
Then, by the Sobolev embedding theorem, Theorem \ref{thm:priori}\ref{item:prio1},
Lemmas \ref{lemma:eL2} and \ref{lemma:euH1}, and \eqref{eqs:g}--\eqref{eqs:grbd},
we have
\begin{align*}
\sigma(t) \sup_{y\in[\varepsilon,k]}(e-1)^2(y,t)
&\le C \int_{\varepsilon}^k \sigma(t) (e-1)^2(x,t)\,{\mathrm d} x
+ C \int_{\varepsilon}^{k} \sigma(t) (\snorm{e-1} \snorm{{\mathrm D}_x e}) (x,t)\,{\mathrm d} x\\
&\le C \sup_{y\in[\varepsilon,k]}\Big( \sigma(t) + \dfrac{v}{r^{2m}}(y,t)\Big)
\int_{\varepsilon}^{k} (e-1)^2 (x,t)\,{\mathrm d} x \\
&\quad\, + \int_{\varepsilon}^k \sigma^2(t)
\Big(\dfrac{r^{2m}}{v} \snorm{{\mathrm D}_x e}^2\Big)(x,t)\,{\mathrm d} x \le C(\varepsilon),
\end{align*}
which implies $\eqref{eqs:ubd}_2$.
\end{proof}
\subsection{Lower bound of the internal energy}\label{subsec:elbd}
We now estimate the lower bound of
$e$.
\begin{lemma}[Lower Bound of the Internal Energy $e$] \label{lemma:ebd}
There exists $C(a)>0$ such that
\begin{equation} \label{eqs:ebd}
e(x,t)\ge C(a)^{-1} \qquad \text{for all $(x,t)\in [0,\infty)\times[0,T]$}. \end{equation}
\end{lemma}
\begin{proof}
Multiplying equation $\eqref{eqs:LFNS-k}_3$ with $-e^{-2}$ to obtain
\begin{equation}\label{eqs:e-1}
{\mathrm D}_t e^{-1}
= (\gamma-1) \dfrac{{\mathrm D}_x(r^m u)}{e v}
-\kappa e^{-2} {\mathrm D}_x\Big(\dfrac{r^{2m}{\mathrm D}_x e}{v}\Big)
+\dfrac{2\mu m}{e^2} {\mathrm D}_x(r^{m-1} u^2)
- \beta \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{e^2v},
\end{equation}
where $\beta:= 2\mu + \lambda$.
To further reduce the equation, we write
\begin{equation*}
- \kappa e^{-2} {\mathrm D}_x\Big( \dfrac{r^{2m}}{v}{\mathrm D}_x e \Big) = \kappa {\mathrm D}_x\Big( \dfrac{r^{2m}}{v}{\mathrm D}_x e^{-1} \Big) - 2\kappa \dfrac{r^{2m}\snorm{{\mathrm D}_x e}^2}{v e^3} .
\end{equation*}
Moreover, using the relation ${\mathrm D}_x(r^{m-1}u^2) = \frac{2u}{r}{\mathrm D}_x(r^m u) - n\frac{vu^2}{r^2}$,
we have
\begin{align*}
-\beta\dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{e^2v} + \dfrac{2\mu m}{e^2} {\mathrm D}_x(r^{m-1} u^2)
&= -\Big(\dfrac{2\mu}{n} + \lambda\Big)\dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{e^2 v} - 2 m \mu \dfrac{v}{e^2} \Big(\dfrac{{\mathrm D}_x(r^m u)}{\sqrt{n} v} - \sqrt{n}\dfrac{u}{r} \Big)^2\\
&\le -\Big(\dfrac{2\mu}{n} + \lambda\Big)\dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{e^2 v}.
\end{align*}
Therefore, using the above two inequalities, \eqref{eqs:e-1} can be reduced to the inequality:
\begin{equation}\label{eqs:eIneq}
{\mathrm D}_t e^{-1} + 2\kappa \dfrac{r^{2m}\snorm{{\mathrm D}_x e}^2}{v e^3}+\Big(\dfrac{2\mu}{n} + \lambda\Big)\dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{e^2 v} \le \dfrac{(\gamma-1){\mathrm D}_x(r^m u)}{e v} + \kappa {\mathrm D}_x\Big( \dfrac{r^{2m}}{v}{\mathrm D}_x e^{-1} \Big) .
\end{equation}
Multiplying \eqref{eqs:eIneq} by $j e^{-j+1}$ for $j\ge 2$ integers and integrating in $x\in[0,k]$ yield
\begin{align*}
&\dfrac{{\mathrm d} }{{\mathrm d} t} \int_{0}^{k} e^{-j} {\mathrm d} x + j\Big(\dfrac{2\mu}{n}+\lambda\Big)\int_{0}^{k}\dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{e^{j+1}v} {\mathrm d} x + 2j\kappa\int_{0}^{k}\dfrac{r^{2m}\snorm{{\mathrm D}_x e}^2}{ve^{j+2}} {\mathrm d} x \\
&\le j (\gamma-1) \int_{0}^{k} e^{-j} \dfrac{{\mathrm D}_x(r^m u)}{v} {\mathrm d} x - j(j-1)\kappa \int_{0}^{k} \dfrac{r^{2m}\snorm{{\mathrm D}_x e^{-1}}^2}{ve^{j-2}} {\mathrm d} x \\
&\le C \int_{0}^k v^{-1} e^{1-j} {\mathrm d} x + \dfrac{j}{2}\Big( \dfrac{2\mu}{n} + \lambda \Big)
\int_{0}^k \dfrac{\snorm{{\mathrm D}_x(r^m u)}^2}{e^{j+1}v} {\mathrm d} x - j(j-1)\kappa \int_{0}^{k} \dfrac{r^{2m}\snorm{{\mathrm D}_x e^{-1}}^2}{ve^{j-2}} {\mathrm d} x.
\end{align*}
Rearranging the above and using Lemma \ref{coro:eubd} and Theorem \ref{thm:priori}\ref{item:prio1}, we have
\begin{align*}
\dfrac{{\mathrm d} }{{\mathrm d} t} \int_{0}^{k} e^{-j} {\mathrm d} x
\le C\int_{0}^k e^{-j} \dfrac{e}{v} {\mathrm d} x \le C(a) \sigma^{-\frac{1}{2}}(t) \int_{0}^k e^{-j} {\mathrm d} x.
\end{align*}
Integrating the above inequality in time, it follows that
\begin{equation*}
\int_{0}^{k} e^{-j}(x,t)\,{\mathrm d} x
\le \sbnorm{e_0^{-1}}_{L^j(0,k)}^j
+ C(a) \int_{0}^{t} \sigma^{-\frac{1}{2}}(\tau) \big(\int_{0}^{k} e^{-j} {\mathrm d} x\big)\,{\mathrm d} \tau.
\end{equation*}
Thus, by the Gr\"onwall inequality, we obtain that, for each integer $j\ge 2$,
\begin{align*}
\sbnorm{e^{-1}(\cdot,t)}_{L^j(0,k)}^j &\le C(a) \sbnorm{e_0^{-1}}_{L^j(0,k)}^j
+ C(a) \Big(\int_{0}^{t} \sigma^{-\frac{1}{2}}(\tau){\mathrm d} \tau \Big) \exp\Big\{C(a)\int_{0}^{t} \sigma^{-\frac{1}{2}}(\tau){\mathrm d}\tau\Big\}\le C(a).
\end{align*}
Letting $j\to \infty$ in the above inequality, we conclude that
$\sbnorm{e^{-1}(\cdot,t)}_{L^\infty} \le C(a) \sbnorm{e_0^{-1}}_{L^\infty}$.
\end{proof}
\section{Global Weak Solutions of the Exterior Problem in the Eulerian Coordinates}
\label{sec:WSCEx}
In this section, the approximate Lagrangian solutions are converted into
the approximate solutions in the Eulerian coordinates
by constructing a set of coordinate transformations.
Then these functions are extended into the entire domain
$(r,t)\in[a,\infty)\times[0,T]$ by using a set of smooth cut-off functions.
Finally, a global weak solution of problem \eqref{eqs:SFNS} and \eqref{eqs:eSFNS} is obtained
via the limit: $k\to\infty$.
\subsection{Coordinate transformation and the Jacobian}\label{subsec:coord}
Let $\{(\widetilde{v}_{a,k},\widetilde{u}_{a,k},\widetilde{e}_{a,k},\widetilde{r}_{a,k})(x,s)\}_{k\in\mathbb{N}}$
be the solutions obtained in the bounded Lagrangian domain for $(x,s)\in[0,k]\times[0,T]$.
For each fixed $(a,k)\in(0,1)\times\mathbb{N}$, the coordinate transformation $(r,t)=\mathcal{T}_{a,k}(x,s)$ is defined as
\begin{equation}\label{eqs:coordtrans}
r=\mathcal{T}_{a,k}^{(1)}(x,s)=\widetilde{r}_{a,k}(x,s),\quad t=\mathcal{T}_{a,k}^{(2)}(x,s)=s \qquad \,\,\,\text{for $(x,s)\in [0,k]\times[0,T]$}.
\end{equation}
Then we see that $\mathcal{T}_{a,k}$ has the image:
$$
\mathcal{T}_{a,k}([0,k]\times [0,T] ) = R_{a,k}
\vcentcolon= \big\{(r,t)\,\vcentcolon\,t\in[0,T], r\in [a,\widetilde{r}_{a,k}(k,t)] \big\}.
$$
Moreover, it follows from a direct calculation that
\begin{equation*}
J\vcentcolon=
{\rm det}(
{\mathrm D}_x\mathcal{T}_{a,k}\,\, \,{\mathrm D}_s\mathcal{T}_{a,k})
= \widetilde{v}_{a,k} (\widetilde{r}_{a,k})^{-m},
\end{equation*}
which, along with Theorem \ref{thm:priori} and Lemma \ref{lemma:rbd}, yields that
\begin{equation}\label{eqs:Jac}
C^{-1}(a)( a^n + C(a) x )^{-\frac{m}{n}} \le J(x,t) \le a^{-m}C(a)
\qquad \text{ for all $(x,t)\in[0,k]\times[0,T]$}.
\end{equation}
Thus, the map $\mathcal{T}_{a,k}\vcentcolon [0,k]\times[0,T]\to R_{a,k}$ is a diffeomorphism so that the inverse map $\mathcal{T}_{a,k}^{-1}$ exists.
Define $(x_{a,k},s_{a,k})(r,t)\vcentcolon= \mathcal{T}_{a,k}^{-1}(r,t)$. Then
\begin{equation*}
\widetilde{r}_{a,k}( x_{a,k}(r,t), t) = r, \quad s_{a,k}(r,t)=t
\qquad\,\, \text{for all $(r,t) \in R_{a,k}$}.
\end{equation*}
Let $\widetilde{\theta}(x,t)\vcentcolon [0,k]\times[0,T]\to {\mathbb R}$ be such that $\widetilde{\theta}\in C^1$.
Define $\theta(r,t)\vcentcolon=\widetilde{\theta}(x_{a,k}(r,t),t)$ for each $(r,t)\in R_{a,k}$.
Then, by the inverse function theorem,
its Eulerian derivative can be computed:
\begin{equation}\label{eqs:coorderiv}
\begin{dcases}
\partial_r \theta(r,t) = \dfrac{\big(\widetilde{r}_{a,k}\big)^m(x_{a,k}(r,t),t)}{\widetilde{v}_{a,k}(x_{a,k}(r,t),t)} {\mathrm D}_x \widetilde{\theta}(x_{a,k}(r,t),t) = r^m \dfrac{{\mathrm D}_x \widetilde{\theta}}{\widetilde{v}_{a,k}}(x_{a,k}(r,t),t), \\
\partial_t \theta(r,t) = {\mathrm D}_t \widetilde{\theta}(x_{a,k}(r,t),t)
- r^m \big(\dfrac{ \widetilde{u}_{a,k}}{\widetilde{v}_{a,k}} {\mathrm D}_x \widetilde{\theta}\big)( x_{a,k}(r,t),t ),
\end{dcases}
\end{equation}
for each $(r,t)\in R_{a,k}$.
Now, for each $(a,k)$, solution $(\widetilde{v}_{a,k}, \widetilde{u}_{a,k}, \widetilde{e}_{a,k})(x,s)$ in $(x,t)\in[0,k]\times[0,T]$ is pulled back to $R_{a,k}$ as:
\begin{equation}\label{eqs:transfunc}
(\overline{v}_{a,k},\overline{u}_{a,k},\overline{e}_{a,k})(r,t) \vcentcolon= ( \widetilde{v}_{a,k}, \widetilde{u}_{a,k}, \widetilde{e}_{a,k} ) ( x_{a,k}(r,t),t )
\qquad \ \text{for $(r,t)\in R_{a,k}$}.
\end{equation}
\subsection{Extension in the Eulerian domain} \label{subsec:Eulext}
We now extend $(\overline{v}_{a,k},\overline{u}_{a,k},\overline{e}_{a,k})$ into the whole Eulerian domain $[a,\infty)\times[0,T]$.
First, let $\chi\in C^{\infty} \vcentcolon {\mathbb R} \to [0,1]$ be such that
$\chi(\zeta)=1$ if $\zeta\le 0$, $\chi(\zeta)=0$ if $\zeta\ge 1$, and $\snorm{\chi^{\prime}} \le 2$ in ${\mathbb R}$.
With this, the following cut-off functions are defined:
\begin{equation}\label{eqs:COfunc}
\varphi_{a,k}(r,t) \vcentcolon= \chi(\dfrac{2r-\widetilde{r}_{a,k}(k,t)}{\widetilde{r}_{a,k}(k,t)}).
\end{equation}
Since $\widetilde{u}_{a,k}(k,t)=0$ for all $t\in[0,T]$, it follows that, for all $(r,t)\in[a,\infty)\times[0,T]$,
\begin{equation}\label{eqs:varphiDeriv}
\partial_r \varphi_{a,k}(r,t)
=\dfrac{2}{\widetilde{r}_{a,k}(k,t)} \chi^{\prime}(\dfrac{2r-\widetilde{r}_{a,k}(k,t)}{\widetilde{r}_{a,k}(k,t)}),
\qquad \partial_t \varphi_{a,k}(r,t) = 0.
\end{equation}
Then
$\varphi_{a,k}\in C^{\infty}([0,\infty)\times[0,T])$.
Moreover, by construction,
\begin{equation*}
\varphi_{a,k}(r,t)=1 \ \text{ if $r\in[0,\frac{\widetilde{r}_{a,k}(k,t)}{2}]$},
\qquad \varphi_{a,k}(r,t)=0 \ \text{ if $r\in[\widetilde{r}_{a,k}(k,t),\infty)$}.
\end{equation*}
Now, by Theorem \ref{thm:priori}\ref{item:prio1},
\begin{align*}
\widetilde{r}_{a,k}(k,t)&= \Big( a^n + n \int_{0}^k v_{a,k}(y,t)\,{\mathrm d} y \Big)^{\frac{1}{n}}
\ge \Big( n \int_{\frac{1}{2}}^k v_{a,k}(y,t)\,{\mathrm d} y \Big)^{\frac{1}{n}}
\ge \Big(\frac{n}{C(T)}\int_{\frac{1}{2}}^k {\mathrm d} y \Big)^{\frac{1}{n}}
\ge \frac{1}{C(T)},
\end{align*}
where the universal constant $C(T)>0$ depends only on $T$ and $n$.
It also follows from Theorem \ref{thm:priori} that
\begin{align*}
\widetilde{r}_{a,k}(k,t) = \Big( a^n + n \int_{0}^k v_{a,k}(y,t)\,{\mathrm d} y \Big)^{\frac{1}{n}}
\ge \frac{k^{\frac{1}{n}}}{C(a)},
\end{align*}
so that, by \eqref{eqs:varphiDeriv},
\begin{equation}\label{eqs:extcut}
\snorm{\partial_r\varphi_{a,k}(r,t)} \le \min\{C(T),C(a)k^{-\frac{1}{n}}\}, \ \
\partial_t \varphi_{a,k}(r,t)=0
\qquad \mbox{for all $(r,t)\in[a,\infty)\times[0,T]$}.
\end{equation}
Using \eqref{eqs:coordtrans}--\eqref{eqs:COfunc}, the extended approximate functions $(\rho_{a,k},u_{a,k},e_{a,k})(r,t)$ in the entire Eulerian domain $(r,t)\in[a,\infty)\times[0,T]$
are defined by \eqref{eqs:extfunc} in \S \ref{subsec:MS}.
\smallskip
The following lemmas show that the extended functions inherit the {\it a-priori} estimates derived
in \S \ref{sec:rhobd}. As before, we still denote $\sigma\vcentcolon=\min\{1,t\}$.
\begin{lemma} \label{lemma:entEstext}
There exists an integer $N(a)=N(a,T,C_0)\in\mathbb{N}$ such that, for all $k\ge N(a)$,
\begin{align}\label{eqs:entEul}
&\sup\limits_{t\in[0,T]}\int_{a}^{\infty}
\Big\{\rho_{a,k}\big(\frac{1}{2}\snorm{u_{a,k}}^2+\psi(e_{a,k})\big)
+G(\rho_{a,k}) \Big\}\,r^m {\mathrm d} r
+ \kappa\int_{0}^{T}\int_{a}^{\infty}
\dfrac{\snorm{\partial_r e_{a,k}}^2}{e_{a,k}^2}\,r^m{\mathrm d} r{\mathrm d} t\nonumber \\
&+\int_{0}^{T}\int_{a}^{\infty} \Big\{\big(\dfrac{2\mu}{n}+\lambda\big) \dfrac{\snorm{\partial_r u_{a,k}+m \frac{u_{a,k}}{r}}^2}{e_{a,k}}
+ \dfrac{2m\mu}{n} \dfrac{\snorm{\partial_r u_{a,k}-\frac{u_{a,k}}{r}}^2}{e_{a,k}}\Big\}
\,r^m{\mathrm d} r{\mathrm d} t \le C(T).
\end{align}
\end{lemma}
\begin{proof}
Throughout the proof, we suppress $(a,k)$ for simplicity and denote:
\begin{equation}\label{aksupp}
\begin{aligned}
&(\rho,u,e,\varphi)(r,t) \equiv (\rho_{a,k},u_{a,k},e_{a,k},\varphi_{a,k})(r,t) && \text{for $(r,t)\in[a,\infty)\times[0,T]$}, \\[1mm]
&(\overline{v},\overline{u},\overline{e})(r,t) \equiv (\overline{v}_{a,k},\overline{u}_{a,k},\overline{e}_{a,k})(r,t)
&& \text{for $(r,t)\in R_{a,k}$}, \\[1mm]
&(\widetilde{v},\widetilde{u},\widetilde{e},\widetilde{r})(x,t) \equiv (\widetilde{v}_{a,k},\widetilde{u}_{a,k},\widetilde{e}_{a,k},\widetilde{r}_{a,k})(x,t)&& \text{for $(x,t)\in[0,k]\times[0,T]$}.
\end{aligned}
\end{equation}
Since $G(\cdot)$ and $\psi(\cdot)$ are convex, and $0\le \varphi \le 1$, it follows
that $G(\rho)\le \varphi G(\overline{v}^{-1})$ and $\psi(e)\le \varphi\,\psi(\overline{e})$.
Using these and the fact that the Jacobian is bounded from \eqref{eqs:Jac},
one can convert the integral in the Eulerian coordinates into the Lagrangian coordinates:
\begin{align*}
&\int_{a}^{\infty} \Big(\rho\big(\frac{1}{2}\snorm{u}^2+\psi(e)\big) + G(\rho) \Big)(r,t)
\,r^m {\mathrm d} r\\
&\le \int_{a}^{\widetilde{r}(k,t)}
\Big(\frac{1}{2}\snorm{\overline{u}}^2+ \psi(\overline{e}) + \overline{v} G(\overline{v}^{-1}) \Big) \overline{v}^{-1}\,r^m {\mathrm d} r
+ \int_{a}^{\widetilde{r}(k,t)}(1-\varphi)\Big(\frac{1}{2}\snorm{\overline{u}}^2+ \psi(\overline{e}) \Big)\,r^m{\mathrm d} r\\
&= \int_{0}^{k} \Big(\frac{1}{2}\snorm{\widetilde{u}}^2 + \psi(\widetilde{e}) + \psi(\widetilde{v})
\Big)(x,t)\,{\mathrm d} x + \int_{0}^{k} \big(1-\varphi(\widetilde{r}(x,t),t)\big)
\Big(\frac{1}{2}\widetilde{v}\snorm{\widetilde{u}}^2+ \widetilde{v} \psi(\widetilde{e}) \Big) (x,t)\,{\mathrm d} x,
\end{align*}
where
$zG(z^{-1})=\psi(z)$ has been used.
By Theorem \ref{thm:priori}\ref{item:prio0}, the first term of the right-hand side is bounded by $C_0>0$.
By the definition of $\widetilde{r}(x,t)$,
\begin{equation*}
\Big( a^n + \frac{n}{C(a)} x \Big)^{\frac{1}{n}} \le \widetilde{r}(x,t)
= \Big( a^n + n \int_{0}^x \widetilde{v}(y,t)\,{\mathrm d} y \Big)^{\frac{1}{n}}
\le \big( a^n + n C(a) x\big)^{\frac{1}{n}}.
\end{equation*}
Thus, for fixed $a>0$, if $k\in\mathbb{N}$ is such that
$k\ge M(a)\vcentcolon= \dfrac{2^n -1}{n}a^n C(a)+ 2^n C(a)^2$, then
\begin{equation}\label{temp:kMa}
\widetilde{r}(k,t)\ge 2 \widetilde{r}(1,t).
\end{equation}
Using this and the support of the cut-off function $\varphi$, it follows that
\begin{align*}
&\int_{0}^k \big(1-\varphi(\widetilde{r}(x,t),t)\big)
\Big(\frac{1}{2}\widetilde{v}\snorm{\widetilde{u}}^2+\widetilde{v}\psi(\widetilde{e}) \Big)(x,t)\,{\mathrm d} x\\
&= \int_{a}^{\widetilde{r}(k,t)} \big(1-\varphi(r,t)\big)
\Big(\frac{1}{2}\snorm{\overline{u}}^2 +\psi(\overline{e})\Big)(r,t)\,r^m {\mathrm d} r
\le \int_{\widetilde{r}(1,t)}^{\widetilde{r}(k,t)}
\Big(\frac{1}{2}\snorm{\overline{u}}^2+\psi(\overline{e})\Big)(r,t)\,r^m {\mathrm d} r\\
&= \int_{1}^{k} \widetilde{v}(x,t)\Big(\frac{1}{2}\snorm{\widetilde{u}}^2 +\psi(\widetilde{e})\Big) (x,t)
\,{\mathrm d} x \le \sup_{y\in[1,k]} \widetilde{v}(y,t) \int_{1}^{k}
\Big(\frac{1}{2}\snorm{\widetilde{u}}^2 +\psi(\widetilde{e})\Big) (x,t)\,{\mathrm d} x \le C(T) C_0,
\end{align*}
where Theorem \ref{thm:priori}\ref{item:prio0}--\ref{item:prio1} has been used.
Next, the dissipation terms in $\eqref{eqs:entEul}$ are considered.
Using \eqref{eqs:extfunc} and the fact that $z\mapsto z^{-2}$ is convex in
$z\in (0,\infty)$ yield
\begin{align*}
\dfrac{\snorm{\partial_r e}^2}{2\snorm{e}^2}
&= \dfrac{\snorm{\varphi\partial_r\overline{e} + (\overline{e}-1) \partial_r \varphi }^2}{2\snorm{\varphi\overline{e}+(1-\varphi)}^2}\le 2^{-1}\big( \varphi \overline{e}^{-2} + (1-\varphi) \big) \snorm{\varphi\partial_r\overline{e} + (\overline{e}-1) \partial_r \varphi }^2 \\
&\le \varphi^2\dfrac{\snorm{\partial_r \overline{e}}^2}{\overline{e}^2} + \dfrac{\snorm{\overline{e}-1}^2}{\overline{e}^2} \varphi \snorm{\partial_r\varphi}^2 + (1-\varphi) \varphi^2 \snorm{\partial_r \overline{e}}^2 + (1-\varphi)\snorm{\partial_r\varphi}^2 \snorm{\overline{e}-1}^2.
\end{align*}
Multiplying both sides by $r^m$ and integrating in $[a,\infty)\times[0,T]$, it follows from \eqref{eqs:extcut} and \eqref{temp:kMa} that, if $k\ge M(a)$, then
\begin{align*}
&\dfrac{1}{2}\int_{0}^T\int_{a}^{\infty} \dfrac{\snorm{\partial_re}^2}{e^2}\,r^m {\mathrm d} r {\mathrm d} t\\
&\le \int_{0}^T \int_{a}^{\infty} \Big\{ \varphi^2\dfrac{\snorm{\partial_r \overline{e}}^2}{\overline{e}^2} + \dfrac{\snorm{\overline{e}-1}^2}{\overline{e}^2} \varphi \snorm{\partial_r\varphi}^2 + (1-\varphi) \Big( \varphi^2 \snorm{\partial_r \overline{e}}^2 + \snorm{\partial_r\varphi}^2 \snorm{\overline{e}-1}^2 \Big) \Big\}\,r^m {\mathrm d} r {\mathrm d} t \\
&\le \int_{0}^T \int_{a}^{\widetilde{r}(k,t)} \dfrac{\snorm{\partial_r \overline{e}}^2}{\overline{e}^2} r^m {\mathrm d} r {\mathrm d} t + \int_{0}^T\int_{\widetilde{r}(1,t)}^{\widetilde{r}(k,t)} \Big\{ \snorm{\partial_r\overline{e}}^2 + C(T) \snorm{\overline{e}-1}^2 + C(a)k^{-\frac{2}{n}} \dfrac{\snorm{\overline{e}-1}^2}{\overline{e}^2} \Big\}\,r^m {\mathrm d} r {\mathrm d} t.
\end{align*}
If $k\ge \max\{C^{\frac{n}{2}}(a)C^{\frac{n}{2}}(T),M(a)\}$, then translating the above integral into
the Lagrangian coordinates and using \eqref{eqs:coorderiv} and Theorem \ref{thm:priori} yield
\begin{align*}
&\int_{0}^T\int_{a}^{\infty} \dfrac{\snorm{\partial_re}^2}{e^2}\,r^m {\mathrm d} r {\mathrm d} t\\
&\le \int_{0}^T \int_{0}^{k} \dfrac{\widetilde{r}^{2m}}{\widetilde{v}} \dfrac{\snorm{{\mathrm D}_x \widetilde{e}}^2}{\widetilde{e}^2} {\mathrm d} x {\mathrm d} t + \int_{0}^T\int_{1}^{k} \Big\{ \dfrac{\widetilde{r}^{2m}}{\widetilde{v}}\snorm{{\mathrm D}_x \widetilde{e}}^2 + C(T)\widetilde{v}\snorm{\widetilde{e}-1}^2 + \dfrac{C(a)}{k^{\frac{2}{n}}} \dfrac{\widetilde{v}\snorm{\widetilde{e}-1}^2}{\widetilde{e}^2} \Big\}{\mathrm d} x {\mathrm d} t\\
&\le C(T)\Big(1+ \sup \big\{ (\widetilde{v}+k^{-\frac{2}{n}}
\dfrac{\widetilde{v}}{\widetilde{e}^{2}})(y,s)\,\vcentcolon\,(y,s)\in[1,k]\times[0,T] \big\}\Big)\le C(T),
\end{align*}
Since $z\to z^{-1}$ is convex in $z\in(0,\infty)$ and $\partial_r u + m \frac{u}{r} = r^{-m}\partial_r(r^m u)$,
then
\begin{align*}
\dfrac{\snorm{\partial_r u + m \frac{u}{r}}^2}{2e}
&\le \frac{1}{2}\big(\varphi\overline{e}^{-1} + (1-\varphi)\big)
\big|\varphi\partial_r \overline{u} + \overline{u} \partial_r\varphi + m \varphi \frac{\overline{u}}{r}\big|^2\\
&\le \varphi^3 \overline{e}^{-1} \snorm{ r^{-m} \partial_r(r^m \overline{u})}^2 + \varphi \snorm{\partial_r \varphi}^2 \overline{e}^{-1}\snorm{\overline{u}}^2
+ (1-\varphi)\big(\varphi^2 \snorm{r^{-m} \partial_r(r^m \overline{u})}^2
+ \snorm{\overline{u}\partial_r \varphi}^2\big).
\end{align*}
By similar arguments as above, we see that, for all $k\ge \max\{ C^{\frac{n}{2}}(a)C^{\frac{n}{2}}(T), M(a) \}$,
\begin{align*}
&\dfrac{1}{2}\int_{0}^T\int_{a}^{\infty} \dfrac{\snorm{r^{-m}\partial_r(r^m \overline{u})}^2}{e} r^m {\mathrm d} r {\mathrm d} t\\
&\le \int_{0}^T\int_{a}^{\widetilde{r}(k,t)}
\dfrac{\snorm{r^{-m}\partial_r( r^m \overline{u})}^2}{\overline{e}} r^m {\mathrm d} r {\mathrm d} t +\dfrac{C(a)}{k^{\frac{2}{n}}}\int_{0}^T\int_{a}^{\widetilde{r}(k,t)}
\dfrac{\snorm{\overline{u}}^2}{\overline{e}}\,r^m {\mathrm d} r {\mathrm d} t \\
&\quad + \int_{0}^T\int_{\widetilde{r}(1,t)}^{\widetilde{r}(k,t)} \big\{ \snorm{r^{-m}\partial_r(r^m \overline{u})}^2
+ C(T)\overline{u}^2 \big\}\,r^m {\mathrm d} r {\mathrm d} t \\
&= \int_{0}^T \int_{0}^{k} \Big\{ \dfrac{\snorm{{\mathrm D}_x(\widetilde{r}^m \widetilde{u})}^2}{\widetilde{v}\widetilde{e}} + \dfrac{C(a)}{k^{\frac{2}{n}}}\dfrac{\widetilde{v}}{\widetilde{e}} \snorm{\widetilde{u}}^2 \Big\}\,{\mathrm d} x {\mathrm d} t
+ \int_{0}^T \int_{1}^{k}\Big\{ \dfrac{\snorm{{\mathrm D}_x(\widetilde{r}^m \widetilde{u})}^2}{\widetilde{v}}
+ C(T) \widetilde{v} \snorm{\widetilde{u}}^2 \Big\}\,{\mathrm d} x {\mathrm d} t\le C(T),
\end{align*}
By the same analysis, we also see that, for all $k\ge \max\{ C^{\frac{n}{2}}(a)C^{\frac{n}{2}}(T), M(a) \}$,
\begin{align*}
\int_{0}^T\int_{a}^{\infty} \dfrac{1}{e}\Bignorm{\partial_r u - \dfrac{u}{r}}^2 r^m {\mathrm d} r {\mathrm d} t
\le C(T).
\end{align*}
Set $N(a)\vcentcolon=
[(\max\{ C^{\frac{n}{2}}(a)C^{\frac{n}{2}}(T), M(a) \}]\in \mathbb{N}$, where $[\cdot]$ is the ceiling function.
Then the proof is completed.
\end{proof}
\begin{lemma}\label{lemma:HREext}
For each $a\in(0,1)$,
\begin{equation}\label{temp:aHREext1}
\begin{split}
&\sup_{k\in\mathbb{N}}\sup_{ 0\le t \le T} \int_{a}^{\infty} \Big\{ \big|(\rho_{a,k}-1,u_{a,k}^2, e_{a,k}-1,\sqrt{\sigma}\partial_r u_{a,k}, \sigma\partial_r e_{a,k})\big|^2 \Big\} (r,t)\, r^m {\mathrm d} r \\
&\, + \sup_{k\in\mathbb{N}}\int_{0}^{T} \int_{a}^{\infty} \Big\{ \big|( \partial_r u_{a,k}, \partial_r e_{a,k}, u_{a,k} \partial_r u_{a,k}, \sqrt{\sigma} \partial_t u_{a,k}, \sigma \partial_t e_{a,k} )\big|^2 \Big\}\,
r^m {\mathrm d} r{\mathrm d} t \le C(a),
\end{split}
\end{equation}
and, for all $(k,r,t)\in\mathbb{N}\times[a,\infty)\times[0,T]$,
\begin{equation}\label{temp:aHREext2}
C^{-1}(a) \le \rho_{a,k}(r,t) \le C(a), \,\, e_{a,k}(r,t) \ge C^{-1}(a),\,\,
\sigma^{\frac{1}{4}}(t)\snorm{u_{a,k}(r,t)}+\sigma^{\frac{1}{2}}(t)e_{a,k}(r,t) \le C(a).
\end{equation}
Moreover, for fixed $\varepsilon\in(0,1]$,
\begin{align}
\sup_{\substack{k\in\mathbb{N}\\ a\in(0,1)}}
\bigg\{&
\sup_{ 0\le t \le T} \int_{\widetilde{r}_{a,k}(\varepsilon,t)}^{\infty} \Big\{ \big|(\rho_{a,k}-1,u_{a,k}^2, e_{a,k}-1,\sqrt{\sigma}\partial_r u_{a,k}, \sigma\partial_r e_{a,k})\big|^2 \Big\} (r,t)\,r^m {\mathrm d} r\bigg.
\nonumber\\
&+ \int_{0}^{T} \int_{\widetilde{r}_{a,k}(\varepsilon,t)}^{\infty}
\Big\{ \big|(\partial_r u_{a,k}, \partial_r e_{a,k},
u_{a,k} \partial_r u_{a,k},\sqrt{\sigma} \partial_t u_{a,k},
\sigma \partial_t e_{a,k} )\big|^2\Big\}\,
r^m {\mathrm d} r{\mathrm d} t\bigg\} \le C(\varepsilon),\quad\label{temp:veHREext1}
\end{align}
and, for any $(r,t)\in [\widetilde{r}(\varepsilon,t),\infty)\times [0,T]$ and $(a,k)\in(0,1)\times\mathbb{N}$,
\begin{equation}\label{temp:veHREext2}
\begin{split}
& C^{-1}(\varepsilon) \le \rho_{a,k}(r,t) \le C(\varepsilon), \quad \ \sigma^{\frac{1}{4}}(t)\snorm{u_{a,k}(r,t)}
+\sigma^{\frac{1}{2}}(t)e_{a,k}(r,t) \le C(\varepsilon).
\end{split}
\end{equation}
\end{lemma}
\begin{proof}
As before, we suppress $(a,k)$ for simplicity and use the notation in \eqref{aksupp}.
Only \eqref{temp:veHREext1}--\eqref{temp:veHREext2} are shown, since \eqref{temp:aHREext1}--\eqref{temp:aHREext2}
follow in the same manner.
First, using the fact that the Jacobian is bounded from \eqref{eqs:Jac}
and construction \eqref{eqs:COfunc}, the integration in the Eulerian coordinates
is converted to the following in the Lagrangian coordinates:
\begin{align*}
\int_{\widetilde{r}(\varepsilon,t)}^{\infty} \snorm{\rho-1}^2\,r^m {\mathrm d} r
\le \int_{\widetilde{r}(\varepsilon,t)}^{\widetilde{r}(k,t)} \snorm{\overline{v}^{-1}-1}^2\,r^m {\mathrm d} r
= \int_{\varepsilon}^{k} \frac{\snorm{\widetilde{v}-1}^2}{\widetilde{v}} {\mathrm d} x \le C(\varepsilon),
\end{align*}
where Theorem \ref{thm:priori} has been used in the last inequality.
By the same method, one can also obtain the corresponding estimates for $\snorm{e-1}^2$ and $\snorm{u}^4$ in $\eqref{temp:veHREext1}_1$.
Now, using \eqref{eqs:coorderiv}, \eqref{eqs:extcut}, and Theorem \ref{thm:priori}, it follows that
\begin{align*}
\int_{\widetilde{r}(\varepsilon,t)}^{\infty} \sigma(t) \snorm{\partial_r u}^2
\,r^m {\mathrm d} r
&= \int_{\widetilde{r}(\varepsilon,t)}^{\infty} \sigma(t) \snorm{\varphi\partial_r \overline{u}
+ \overline{u}\partial_r\varphi}^2
\,r^m {\mathrm d} r\\
&\le \int_{\widetilde{r}(\varepsilon,t)}^{\widetilde{r}(k,t)} \sigma(t) \snorm{\partial_r \overline{u}}\,r^m {\mathrm d} r
+ C\int_{\widetilde{r}(\varepsilon,t)}^{\widetilde{r}(k,t)} \snorm{\overline{u}}^2\,r^m {\mathrm d} r\\
&= \int_{\varepsilon}^{k} \sigma \dfrac{\widetilde{r}^{2m}}{\widetilde{v}} \snorm{{\mathrm D}_x \widetilde{u}}^2\,{\mathrm d} x
+ \int_{\varepsilon}^{k} \widetilde{v} \snorm{\widetilde{u}}^2\,{\mathrm d} x \le C(\varepsilon).
\end{align*}
In the same way, we can also obtain the estimates for $\sigma^2 \snorm{\partial_r e}^2$
and $\snorm{(\partial_r u,\partial_r e, u \partial_r u)}^2$ in \eqref{temp:veHREext1}.
By \eqref{eqs:coorderiv} and Theorem \ref{thm:priori}, it follows that
\begin{align*}
&\int_{0}^{T}\int_{\widetilde{r}(\varepsilon,t)}^{\infty}
\sigma^2(t)\snorm{ \partial_t e}^2(r, t)\,r^m {\mathrm d} r {\mathrm d} t
= \int_{0}^{T}\int_{\widetilde{r}(\varepsilon,t)}^{\infty}\sigma^2(t)
\varphi^2(r,t)\snorm{\partial_t \overline{e}}^2(r, t)\,r^m {\mathrm d} r {\mathrm d} t\\
&\le \int_{0}^{T}\int_{\widetilde{r}(\varepsilon,t)}^{\widetilde{r}(k,t)}\sigma^2(t)
\snorm{ \partial_t \overline{e}}^2(r, t)\,r^m {\mathrm d} r {\mathrm d} t
= \int_{0}^{T}\int_{\varepsilon}^{k}\sigma^2(t) \widetilde{v}(x,t)
\Bignorm{ {\mathrm D}_t \widetilde{e} - \widetilde{r}^m \dfrac{\widetilde{u}}{\widetilde{v}}{\mathrm D}_x \widetilde{e} }^2(x, t)\,{\mathrm d} x {\mathrm d} t\\
&\le \int_{0}^{T}\int_{\varepsilon}^{k}\sigma^2 \widetilde{v} \snorm{ {\mathrm D}_t \widetilde{e} }^2 \, {\mathrm d} x {\mathrm d} t
+ \int_{0}^{T}\int_{\varepsilon}^{k}\sigma^{\frac{3}{2}} \snorm{\sigma^{\frac{1}{4}} \widetilde{u}}^2
\dfrac{\widetilde{r}^{2m}}{\widetilde{v}} \snorm{{\mathrm D}_x\widetilde{e}}^2\, {\mathrm d} x {\mathrm d} t \le C(\varepsilon).
\end{align*}
By the same argument, one can also obtain the estimate for $\sigma\snorm{\partial_t u}^2$ in \eqref{temp:veHREext1}.
Next, for $(r,t)$ such that $t\in[0,T]$ and $r\in [\widetilde{r}(\varepsilon,t),\infty)\cap R_{a,k}$, there exists $(x,s)\in[\varepsilon,k]\times[0,T]$ such that $(r,t)=(\widetilde{r}(x,s), s)$
and the following estimates hold for $(r,t)\in R_{a,k}$:
\begin{align*}
\rho(r,t)
&= \varphi(\widetilde{r}(x,t),t)\widetilde{v}^{-1}(x,t) + 1 -\varphi(\widetilde{r}(x,t),t)
\le C(\varepsilon)\varphi(\widetilde{r}(x,t),t) +1 -\varphi(\widetilde{r}(x,t),t) \le C(\varepsilon),\\
\rho(r,t)
&\ge C(\varepsilon)^{-1}\varphi(\widetilde{r}(x,t),t) +1 -\varphi(\widetilde{r}(x,t),t) \ge C(\varepsilon)^{-1}.
\end{align*}
For $(r,t)$ such that $t\in[0,T]$ and $r\in [\widetilde{r}(\varepsilon,t),\infty)\times [0,T] \backslash R_{a,k}$,
then $r> \widetilde{r}(k,t)$ so that $\varphi(r,t)=0$ and
$\rho(r,t)=\overline{v}^{-1}(r,t)\varphi(r,t) - \varphi(r,t)+1 = 1$,
which leads to the upper and lower bounds of the extended density.
In a similar manner, we can also obtain the bounds for $(u,e)$
listed in \eqref{temp:veHREext2}.
\end{proof}
Throughout \S\ref{subsec:klimpath}--\S \ref{subsec:indepPath},
we will suppress parameter $a\in(0,1)$ and denote
\begin{equation}\label{asuppress}
\begin{aligned}
&(\rho_k,u_k,e_k,\varphi_k)\equiv (\rho_{a,k},u_{a,k},e_{a,k},\varphi_{a,k}), \ \ (\overline{v}_k,\overline{u}_k,\overline{e}_k) \equiv (\overline{v}_{a,k},\overline{u}_{a,k},\overline{e}_{a,k}),\\[1mm]
&(\widetilde{v}_{k},\widetilde{u}_{k},\widetilde{e}_{k},\widetilde{r}_{k}) \equiv (\widetilde{v}_{a,k},\widetilde{u}_{a,k},\widetilde{e}_{a,k},\widetilde{r}_{a,k})
\end{aligned}
\end{equation}
Note that the main aim of these subsections is to prove Theorem \ref{thm:WSCEx} via the limit: $k\to\infty$.
\subsection{Construction of the particle path function}\label{subsec:klimpath}
In this section, we construct the particle path function
$\widetilde{r}(x,t)$.
\begin{lemma}\label{lemma:rkcvg}
There exist both a continuous function $\widetilde{r}(x,t) \vcentcolon [0,\infty)\times[0,T] \to [a,\infty)$
and a subsequence $k\to\infty$ such that
\begin{enumerate}[label=(\roman*),ref=(\roman*),font={\normalfont\rmfamily}]
\item\label{item:rkcvg1} for any compact subset $K\Subset [0,\infty)\times[0,T]$,
\begin{align*}
\lim\limits_{k\to \infty} \sup_{(x,t)\in K} \snorm{\overline{r}_{k}(x,t)-\widetilde{r}(x,t)} = 0,
\end{align*}
where $\overline{r}_k(x,t)$ is defined by
\begin{equation}\label{rrtemp0}
\overline{r}_k(x,t) =\begin{dcases*}
\widetilde{r}_k(x,t) & if $(x,t)\in[0,k]\times[0,T]$,\\
x-k + \widetilde{r}_k(k,t) & if $(x,t)\in[k,\infty)\times[0,T]$.
\end{dcases*}
\end{equation}
\item\label{item:rkcvg2} for each $\varepsilon>0$,
\begin{equation*}
\begin{aligned}
&\snorm{\widetilde{r}(x_1,t)-\widetilde{r}(x_2,t)} \le C(\varepsilon) \snorm{x_1-x_2}
&&\quad \text{for all} \ (x_1,x_2,t)\in [\varepsilon,\infty)^2\times[0,T],\\
&\snorm{\widetilde{r}(x,t_1)-\widetilde{r}(x,t_2)}
\le C(\varepsilon) \snorm{t_1^{\frac{3}{4}}-t_2^{\frac{3}{4}}}
&&\quad \text{for all} \ (x,t_1,t_2)\in [\varepsilon,\infty)\times[0,T]^2.
\end{aligned}
\end{equation*}
\smallskip
\item\label{item:rkcvg3} for each $t\in[0,T]$, the function $x\mapsto\widetilde{r}(x,t)$ is strictly increasing and
\begin{equation*}
n x \psi^{-1}_{-}(\dfrac{C_0}{x}) \le \widetilde{r}^{\,n}(x,t) \le C_0(1 +x)
\qquad\text{for all} \ \ (x,t)\in[0,\infty)\times[0,T].
\end{equation*}
\end{enumerate}
\end{lemma}
\begin{proof}
For each integer $l\in\mathbb{N}$,
let $S_l\vcentcolon= [0,l]\times[0,T]\subset [0,\infty)\times[0,T]$.
Set $x_1,\, x_2,\, x \in [0,l]$ and $t_1,\,t_2,\,t \in[0,T]$. We consider two cases: $k\ge l$ or $k<l$.
\smallskip
Case 1: $k\ge l$. Then $\overline{r}_k(x_i,t)=\widetilde{r}_k(x_i,t)$ for $i=1,\,2$.
It follows from Theorem \ref{thm:priori} and the mean value theorem that
\begin{equation*}
\snorm{\overline{r}_k(x_1,t)- \overline{r}_k(x_2,t)} \le \snorm{x_1-x_2} \sup_{x\in[0,k]}|{\mathrm D}_x \widetilde{r}_k(x,t)|
\le \snorm{x_1-x_2} \sup_{x\in[0,k]}\big|\dfrac{\widetilde{v}_k}{\widetilde{r}_k^{\,m}}\big|
\le C(a) \snorm{x_1-x_2}.
\end{equation*}
Case 2: $k<l$. For this case, we consider three sub-cases:
If $x_1,\, x_2 \in[0,k]$, the same inequality as Case 1 can be obtained by repeating
the same argument.
If $x_1, x_2 \in [k,l]$, then, by \eqref{rrtemp0},
$\snorm{\overline{r}_k(x_1,t)-\overline{r}_k(x_2,t)}= \snorm{x_1-x_2}$.
If $x_1\in [0,k]$ and $x_2\in[k,l]$, then, by \eqref{rrtemp0},
\begin{equation*}
\snorm{\overline{r}_k(x_1,t)-\overline{r}_{k}(x_2,t)} \le \snorm{\widetilde{r}_k(x_1,t)-\widetilde{r}_k(k,t)} + \snorm{x_2-k} \le C(a)(k-x_1)+x_2-k\le C(a) \snorm{x_1-x_2}.
\end{equation*}
\smallskip
In summary, for both cases,
\begin{equation}\label{rrtemp1a}
\sup_{k\in\mathbb{N}} \snorm{\overline{r}_k(x_1,t)-\overline{r}_k(x_2,t)} \le C^{\ast}(a) \snorm{x_1-x_2}.
\end{equation}
Moreover, for both cases $k\ge l$ and $k<l$, it can be shown that, if $t_*=\min\{1,t_1,t_2\}$, then
\begin{equation}\label{rrtemp1b}
\snorm{\overline{r}_k(x,t_1)-\overline{r}_k(x,t_2)} = \Bignorm{\int_{t_1}^{t_2} u_k(x,s)\,{\mathrm d} s }
\le C(a) \Bignorm{\int_{t_1}^{t_2} s^{-\frac{1}{4}}\,{\mathrm d} s } \le C(a) t_*^{-\frac{1}{4}} \snorm{t_1-t_2}.
\end{equation}
Similarly, by \eqref{rrtemp0}, Theorem \ref{thm:priori}, and Lemma \ref{lemma:rbd},
it can be shown that, for each $\varepsilon\in(0,1]$,
\begin{equation}\label{rrtemp2}
\begin{aligned}
&\sup_{k \in \mathbb{N}}\snorm{\overline{r}_k(x_1,t)- \overline{r}_k(x_2,t)} \le C(\varepsilon) \snorm{x_1-x_2}
&& \ \text{ for all $(x_1,x_2,t)\in [\varepsilon,l]^2\times[0,T]$,} \\
&\sup_{k\in \mathbb{N}}\snorm{\overline{r}_k(x,t_1)- \overline{r}_k(x,t_2)} \le C(\varepsilon) \big|t_1^{\frac{3}{4}}-t_2^{\frac{3}{4}}\big|
&& \ \text{ for all $(x,t_1,t_2)\in [\varepsilon,l]\times[0,T]^2$.}
\end{aligned}
\end{equation}
Furthermore, the following bound holds for all $k\in\mathbb{N}$:
\begin{equation}\label{rrtemp3}
\sup_{(x,t)\in S_l}\overline{r}_k(x,t)
= \sup_{(x,t)\in S_l}\Big( a^n + n\int_{0}^{x} \widetilde{v}_k(y,t)\,{\mathrm d} y \Big)^{\frac{1}{n}}
\le \big(a^n + n C(a) l\big)^{\frac{1}{n}}.
\end{equation}
By \eqref{rrtemp1a}--\eqref{rrtemp1b} and \eqref{rrtemp3}, one can apply Proposition \ref{prop:aaExt}
to obtain a subsequence (still denoted) $k \to\infty$
and a continuous function $\widetilde{r}(x,t)\vcentcolon [0,\infty)\times[0,T]\to [a,\infty)$ such that
\begin{equation}\label{rtemp4}
\lim\limits_{k\to\infty} \sup_{(x,t)\in K} \snorm{\widetilde{r}(x,t)-\overline{r}_{k}(x,t)} =0
\qquad \text{ for all}
\ K\Subset [0,\infty)\times[0,T].
\end{equation}
Applying the strong convergence \eqref{rtemp4} for \eqref{rrtemp2}, one obtains Lemma \ref{lemma:rkcvg}\ref{item:rkcvg2}.
\smallskip
Next, let $0<x_1<x_2<\infty$. For all $k \ge 2^n x_2 C^2(a) + \frac{2^n-1}{n}C(a)$, one can verify $2\widetilde{r}_k(x_1,t)\le2\widetilde{r}_k(x_2,t)\le \widetilde{r}_k(k,t)$. It follows that
\begin{equation*}
\int_{\widetilde{r}_k(x_1,t)}^{\widetilde{r}_k(x_2,t)} \rho_k(r,t)\,r^m {\mathrm d} r
= \int_{\widetilde{r}_k(x_1,t)}^{\widetilde{r}_k(x_2,t)} \overline{v}_k^{-1}(r,t)\,r^m {\mathrm d} r
= x_2 - x_1.
\end{equation*}
Since $\rho_k(r,t)\le C(x_1) $ for $r\in[\widetilde{r}_k(x_1,t),\infty)$ from Lemma \ref{lemma:HREext},
\begin{align*}
x_2-x_1 = \int_{\widetilde{r}_k(x_1,t)}^{\widetilde{r}_k(x_2,t)} \rho_k(r,t)\,r^m {\mathrm d} r
\le \dfrac{C(x_1)}{n} \big(\widetilde{r}_k^{\, n}(x_1,t) -\widetilde{r}_k^{\,n}(x_2,t)\big).
\end{align*}
Taking $k\to\infty$ on the above and using \eqref{rtemp4} yields
\begin{equation*}
0 < x_2 - x_1 < \dfrac{C(x_1)}{n} \big(\widetilde{r}^{\,n}(x_1,t) -\widetilde{r}^{\,n}(x_2,t)\big).
\end{equation*}
This shows that $x\mapsto \widetilde{r}(x,t)$ is strictly increasing for each $t\in[0,T]$.
For a fixed point $(x,t)\in[0,\infty)\times[0,T]$, it follows from Lemma \ref{lemma:rbd} that
\begin{align*}
nx \psi_{-}^{-1}(\frac{C_0}{x}) \le \widetilde{r}_k^{\,n}(x,t) \le C_0(1 + x)
\qquad \text{for all $k\ge x$}.
\end{align*}
Taking $k\to\infty$ and using \eqref{rtemp4},
we conclude Lemma \ref{lemma:rkcvg}\ref{item:rkcvg3}.
\end{proof}
\smallskip
\subsection{Compactness results}\label{subsec:compact}
\begin{lemma}\label{lemma:klimholder}
There exist both a continuous function $(u,e)(r,t)\vcentcolon [a,\infty)\times (0,T]\to {\mathbb R}\times [0,\infty)$
and a subsequence $($still denoted$)$ $\{(u_k,e_k)\}_{k\in\mathbb{N}}$ such that,
for each compact subset $K\Subset [a,\infty)\times[0,T]$,
\begin{equation}\label{temp:klimue0}
\lim\limits_{k\to\infty} \sup\limits_{(r,t)\in K} \snorm{(u_k-u,e_k-e)}(r,t) = 0.
\end{equation}
Moreover, for each fixed $\varepsilon>0$, if $t\in[0,T]$ and $r_1,\, r_2 \ge \widetilde{r}(\varepsilon,t)$,
then
\begin{equation}\label{temp:klimue1}
\sigma^{\frac{1}{2}}(t)\snorm{u(r_1,t)-u(r_2,t)} + \sigma(t)\snorm{e(r_1,t)-e(r_2,t)} \le C(\varepsilon) \snorm{r_1-r_2}^{\frac{1}{2}},
\end{equation}
and, if $0<t_1<t_2 \le T$ and $r \ge \sup_{t_1\le t \le t_2}\widetilde{r}(\varepsilon,t)$, then
\begin{equation}\label{temp:klimue2}
\sigma^{\frac{1}{2}}(t_1)\snorm{u(r,t_1)-u(r,t_2)}+ \sigma(t_1)\norm{e(r,t_1)-e(r,t_2)} \le C(\varepsilon) \snorm{t_2-t_1}^{\frac{1}{4}}.
\end{equation}
In addition,
\begin{equation}\label{temp:klimue3}
\begin{dcases*}
\snorm{u(r,t)} \le C(a) \sigma^{-\frac{1}{4}}(t), \ C^{-1}(a) \le e(r,t) \le C(a) \sigma^{-\frac{1}{2}}(t) & if $(r,t)\in[a,\infty)\times[0,T]$,\\
\snorm{u(r,t)} \le C(\varepsilon) \sigma^{-\frac{1}{4}}(t),
\ \ e(r,t) \le C(\varepsilon) \sigma^{-\frac{1}{2}}(t) & if $t\in[0,T]$, $r\ge \widetilde{r}(\varepsilon,t)$.
\end{dcases*}
\end{equation}
\end{lemma}
\begin{proof}
We give our proof only for $u(r,t)$, since the proof is the same for $e(r,t)$.
Let $(r_1,r_2,t)\in [a,\infty)^2\times(0,T]$. By
Lemma \ref{lemma:HREext}, for all $k\in\mathbb{N}$,
\begin{align*}
\norm{u_k(r_2,t)-u_k(r_1,t)}
&\le (r_2 - r_1)^{\frac{1}{2}} r_1^{-\frac{m}{2}}
\Big( \int_{r_1}^{r_2} \snorm{\partial_r u_k}^2 (r,t)\,r^m {\mathrm d} r \Big)^{\frac{1}{2}}\\
&\le \sigma^{-\frac{1}{2}}(t) a^{-\frac{m}{2}} \Big( \int_{a}^{\infty} \sigma(t) \norm{\partial_r u_k}^2 (r,t)\,r^m
{\mathrm d} r \Big)^{\frac{1}{2}} \snorm{r_2-r_1}^{\frac{1}{2}} \\
& \le C(a) \sigma^{-\frac{1}{2}}(t) \snorm{r_2-r_1}^{\frac{1}{2}}.
\end{align*}
Let $r\in [a,\infty)$ and $0<t_1<t_2\le T$, and set $h\vcentcolon= \sqrt{t_2-t_1}$.
By the mean value theorem,
\begin{equation}\label{aholtemp1}
\dfrac{1}{h}\int_{r}^{r+h} \snorm{ u_k(\zeta,t_1) - u_k(\zeta,t_2) }\,{\mathrm d} \zeta
= \snorm{ u_k(\xi,t_1) - u_k(\xi,t_2) }
\qquad \ \text{for some $\xi\in (r,r+h)$}.
\end{equation}
By the triangle inequality and the fundamental theorem of calculus, for all $k\in\mathbb{N}$,
\begin{align*}
&\snorm{u_k(r,t_1) - u_k(r,t_2)}-\snorm{ u_k(\xi,t_1) - u_k(\xi,t_2) }
\le \Bignorm{\int_{r}^{\xi} \big(\partial_r u_k (\zeta,t_1) - \partial_r u_k (\zeta,t_2) \big)\,{\mathrm d} \zeta} \\
&\le \norm{\xi-r}^{\frac{1}{2}} r^{-\frac{m}{2}}
\sum_{i=1}^2 \Big( \int_{r}^{\xi}\snorm{\partial_r u_k(\zeta,t_i)}^2\,\zeta^m {\mathrm d} \zeta \Big)^{\frac{1}{2}} \\
&\le a^{-\frac{m}{2}} \sigma^{-\frac{1}{2}}(t_1) \sqrt{h}
\sum_{i=1}^2 \Big(\int_{a}^{\infty}
\sigma(t_i)\snorm{\partial_r u_k(\zeta,t_i)}^2\,\zeta^m {\mathrm d} \zeta \Big)^{\frac{1}{2}}
\le C(a) \sigma^{-\frac{1}{2}}(t_1) \sqrt{h}.
\end{align*}
Thus, it follows from \eqref{aholtemp1} that, for all $k\in\mathbb{N}$,
\begin{align*}
&\snorm{u_k(r,t_1) - u_k(r,t_2)}\\
&\le C(a) \sigma^{-\frac{1}{2}}(t_1) \sqrt{h}
+ \dfrac{1}{h}\int_{r}^{r+h} \int_{t_1}^{t_2} \snorm{ \partial_t u_k (\zeta,t)}\,{\mathrm d} t {\mathrm d} \zeta \\
&\le C(a) \sigma^{-\frac{1}{2}}(t_1)\sqrt{h}
+ \dfrac{\sigma^{-\frac{1}{2}}(t_1) a^{-\frac{m}{2}}}{h}
\Big(\int_{r}^{r+h} \int_{t_1}^{t_2} {\mathrm d} t {\mathrm d} \zeta\Big)^{\frac{1}{2}}
\Big(\int_{t_1}^{t_2}\int_{a}^{\infty}\sigma(t)\snorm{ \partial_t u_k}^2\,\zeta^m {\mathrm d}\zeta{\mathrm d} t \Big)^{\frac{1}{2}}\\
&\le C(a) \sigma^{-\frac{1}{2}}(t_1) \sqrt{h} + C(a)\sigma^{-\frac{1}{2}}(t_1) h^{-\frac{1}{2}}
\snorm{t_2-t_1}^{\frac{1}{2}}
\le C(a)\sigma^{-\frac{1}{2}}(t_1) \snorm{t_2-t_1}^{\frac{1}{4}}.
\end{align*}
Therefore, we have
\begin{equation}\label{aholtemp2}
\begin{dcases*}
\sup\limits_{k\in \mathbb{N}}\,\snorm{u_k(r_2,t)-u_k(r_1,t)} \le C(a) \sigma^{-\frac{1}{2}}(t) \snorm{r_2-r_1}^{\frac{1}{2}} &\quad if $(r_1,r_2,t)\in[a,\infty)^2\times(0,T]$,\\
\sup\limits_{k\in \mathbb{N}}\,\snorm{u_k(r,t_2)-u_k(r,t_1)}
\le C(a) \sigma^{-\frac{1}{2}}(t_*) \snorm{t_2 - t_1}^{\frac{1}{4}}
&\quad if $(r,t_1,t_2)\in[a,\infty)\times(0,T]^2$,
\end{dcases*}
\end{equation}
where $t_*\equiv \min\{t_1,t_2\}$. Let $S_l=[0,\infty)\times[l^{-1},T]$ for each $l\in\mathbb{N}$.
Then we can apply Proposition \ref{prop:aaExt} with \eqref{aholtemp2} and $S_l$
to obtain a continuous function $u(r,t)\vcentcolon [a,\infty)\times(0,T]\to {\mathbb R}$ and a subsequence (still denoted) $\{u_k\}_{k\in\mathbb{N}}$ such that \eqref{temp:klimue0} holds.
\smallskip
Next, let $\varepsilon\in(0,1]$. Since $x\to \widetilde{r}(x,t)$ is strictly increasing and $(x,t)\mapsto\widetilde{r}(x,t)$ is continuous,
it follows that $ d \vcentcolon= \inf_{t\in[0,T]}\{\widetilde{r}(\varepsilon,t) - \widetilde{r}(\frac{\varepsilon}{2},t)\}>0$.
By Lemma \ref{lemma:rkcvg}, there exists $N_\varepsilon\in\mathbb{N}$ such that
$\sup_{0\le t \le T}\snorm{\widetilde{r}_k(\frac{\varepsilon}{2},t)-\widetilde{r}(\frac{\varepsilon}{2},t)} \le \frac{d}{2}$
if $k\ge N_\varepsilon$, which implies that
\begin{equation}\label{veholtemp1}
\widetilde{r}_k(\frac{\varepsilon}{2},t) < \widetilde{r}(\varepsilon,t) \qquad \text{for each $t\in[0,T]$ and $k\ge N_\varepsilon$}.
\end{equation}
Let $(r_1,r_2,t)$ be a triplet such that $t\in(0,T]$ and $r_1,\, r_2\in [\widetilde{r}(\varepsilon,t),\infty)$.
Then \eqref{veholtemp1} implies that $r_1, r_2 > \widetilde{r}_k(\frac{\varepsilon}{2},t)$ if $k\ge N_\varepsilon$.
By
Lemma \ref{lemma:HREext}, it follows that, for all $k\ge N_\varepsilon$,
\begin{align*}
\snorm{u_k(r_2,t)-u_k(r_1,t)}
\le \dfrac{\sigma^{-\frac{1}{2}}(t)}{\widetilde{r}_k^{\,\frac{m}{2}}(\frac{\varepsilon}{2},t)}
\Big(\int_{\widetilde{r}_k(\frac{\varepsilon}{2},t)}^{\infty} \sigma(t) \snorm{\partial_r u_k}^2(r,t)\,r^m {\mathrm d} r \Big)^{\frac{1}{2}} \snorm{r_2-r_1}^{\frac{1}{2}}
\le C(\varepsilon) \sigma^{-\frac{1}{2}}(t) \snorm{r_2-r_1}^{\frac{1}{2}}.
\end{align*}
Moreover, let $(r,t_1,t_2)\in [a,\infty)\times(0,T]^2$ be a triplet such that
$0<t_1 < t_2\le T$ and $r \ge \sup_{t_1\le t\le t_2}\widetilde{r}(\varepsilon,t)$.
Then, by \eqref{veholtemp1}, $r > \widetilde{r}_k(\frac{\varepsilon}{2},t)\ge C^{-1}(\varepsilon)$ for all $t\in[0,T]$ if $k\ge N_\varepsilon$.
Set $h\vcentcolon= \sqrt{t_2-t_1}$. Then, by the mean value theorem,
\begin{equation*}
\dfrac{1}{h}\int_{r}^{r+h} \snorm{ u_k(\zeta,t_1) - u_k(\zeta,t_2) }\,{\mathrm d} \zeta
= \snorm{ u_k(\xi,t_1) - u_k(\xi,t_2) } \qquad \text{for some $\xi\in (r,r+h)$}.
\end{equation*}
By the triangle inequality,
for all $k\ge N_\varepsilon$,
\begin{align*}
&\snorm{u_k(r,t_1) - u_k(r,t_2)}-\snorm{ u_k(\xi,t_1) - u_k(\xi,t_2) }\\
&\le \sqrt{\snorm{\xi-r}}
r^{-\frac{m}{2}}
\sum_{i=1}^2\Big(\int_{r}^{\xi}\snorm{\partial_r u_k(\zeta,t_i)}^2\,\zeta^m{\mathrm d}\zeta\Big)^{\frac{1}{2}} \\
&\le C(\varepsilon) \sqrt{h}
\sigma^{-\frac{1}{2}}(t_1)
\sum_{i=1}^2 \Big( \int_{\widetilde{r}_k(\frac{\varepsilon}{2},t_i)}^{\infty}
\sigma(t_i)\snorm{\partial_r u_k(\zeta,t_i)}^2 \zeta^m {\mathrm d} \zeta \Big)^{\frac{1}{2}}
\le C(\varepsilon) \sigma^{-\frac{1}{2}}(t_1)\sqrt{h}
\end{align*}
It follows that, for all $k\ge N_\varepsilon$,
\begin{align*}
&\snorm{u_k(r,t_1) - u_k(r,t_2)} \\
&\le C(\varepsilon) \sqrt{h}
\sigma^{-\frac{1}{2}}(t_1)
+ \dfrac{1}{h}\int_{r}^{r+h} \int_{t_1}^{t_2} \snorm{ \partial_t u_k (\zeta,t)}\,{\mathrm d} t {\mathrm d} \zeta \\
&\le C(\varepsilon) \sqrt{h}
\sigma^{-\frac{1}{2}}(t_1) + \dfrac{\sigma^{-\frac{1}{2}}(t_1) r^{-\frac{m}{2}}}{h}\Big(\int_{r}^{r+h} \int_{t_1}^{t_2} {\mathrm d} t {\mathrm d} \zeta\Big)^{\frac{1}{2}}
\Big( \int_{t_1}^{t_2}\int_{\widetilde{r}_k(\frac{\varepsilon}{2},t)}^{\infty} \sigma \snorm{ \partial_t u_k }^2\,\zeta^m
{\mathrm d}\zeta {\mathrm d} t \Big)^{\frac{1}{2}}\\
&\le C(\varepsilon) \sqrt{h}
\sigma^{-\frac{1}{2}}(t_1) + C(\varepsilon) h^{-\frac{1}{2}} \snorm{t_2-t_1}^{\frac{1}{2}} \sigma^{-\frac{1}{2}}(t_1)
\le C(\varepsilon) \sigma^{-\frac{1}{2}}(t_1) \snorm{t_2-t_1}^{\frac{1}{4}}.
\end{align*}
Then
\begin{equation*}
\begin{dcases*}
\sup\limits_{k\ge N_\varepsilon}\snorm{u_k(r_2,t)-u_k(r_1,t)} \le C(\varepsilon) t^{-\frac{1}{2}} \snorm{r_2-r_1}^{\frac{1}{2}} &if $t\in(0,T]$, $r_1, r_2 \ge \widetilde{r}(\varepsilon,t)$,\\
\sup\limits_{k\ge N_\varepsilon}\snorm{u_k(r,t_2)-u_k(r,t_1)} \le C(\varepsilon) \sigma^{-\frac{1}{2}}(t_*)
\snorm{t_2 - t_1}^{\frac{1}{4}} &if $0<t_1<t_2\le T$, $r\ge R(\varepsilon,t_1,t_2)$,
\end{dcases*}
\end{equation*}
where $R(\varepsilon,t_1,t_2)\vcentcolon= \sup_{t_1\le t\le t_2}\widetilde{r}(\varepsilon,t)$.
Taking $k\to\infty$ and using \eqref{temp:klimue0}, we obtain the continuity estimates for $u$ in \eqref{temp:klimue1}--\eqref{temp:klimue2}.
Finally, Lemma \ref{lemma:HREext} and \eqref{temp:klimue0} lead to $\eqref{temp:klimue3}_1$.
In addition, for each $\varepsilon\in(0,1]$, \eqref{veholtemp1} implies that $r\ge \widetilde{r}_k(\frac{\varepsilon}{2},t)$
if $r\ge \widetilde{r}(\varepsilon,t)$ and $k\ge N_\varepsilon$.
Hence, $\norm{u_k(r,t)}\le C(\varepsilon)\sigma^{-\frac{1}{4}}(t)$ for all $(r,t,k)$ such that $k\ge N_\varepsilon$, $t\in[0,T]$,
and $r \ge \widetilde{r}(\varepsilon,t)$.
Letting $k\to\infty$ and using \eqref{temp:klimue0}, we obtain $\eqref{temp:klimue3}_2$.
\end{proof}
\begin{proposition}\label{prop:veDomain}
Let $\varepsilon\in(0,1]$, and let $(r,t)\in[a,\infty)\times(0,T]$ be such that $r> \widetilde{r}(\varepsilon,t)$.
Denote ${\mathrm D}_{\delta}(r,t)\vcentcolon= \{ (\zeta,s) \vcentcolon \sqrt{(\zeta-r)^2 + (s-t)^2} < \delta \}$.
Then there exist $N(\varepsilon,r,t)\in\mathbb{N}$ and $\delta(\varepsilon,r,t)>0$ such that,
for all $\delta\in (0,\delta(\varepsilon,r,t))$ and $ k\ge N(\varepsilon,r,t)$,
\begin{equation*}
{\mathrm D}_\delta(r,t) \subset \{ (\zeta,s)\in[a,\infty)\times[0,T]\,\vcentcolon\,\zeta> \widetilde{r}_k(\varepsilon,s) \}.
\end{equation*}
\end{proposition}
\begin{proof}
Let $\eta\vcentcolon=r-\widetilde{r}(\varepsilon,t)>0$.
Since $(\zeta,s) \mapsto \zeta-\widetilde{r}(\varepsilon,s)$ is continuous,
there exists $\delta(\varepsilon,r,t)>0$ such that, for all $\delta\le \delta(\varepsilon,r,t)$,
if $(\zeta,s)\in {\mathrm D}_\delta(r,t)$, then
$\snorm{\zeta-\widetilde{r}(\varepsilon,s) - r + \widetilde{r}(\varepsilon,t)} < \frac{\eta}{2}$ which implies that
$\zeta> \frac{\eta}{2} + \widetilde{r}(\varepsilon,s)$.
Therefore,
\begin{equation}\label{temp:veDomain}
{\mathrm D}_\delta(r,t) \subset \{ (\zeta,s)\in[a,\infty)\times[0,T]\,\vcentcolon\,\zeta>\frac{\eta}{2} + \widetilde{r}(\varepsilon,s) \}.
\end{equation}
By Lemma \ref{lemma:rkcvg}, $\lim_{k\to\infty} \sup_{s\in[0,T]}\snorm{\widetilde{r}_k(\varepsilon,s)-\widetilde{r}(\varepsilon,s)}=0$.
Thus, there exists $N(\varepsilon,r,t)\in\mathbb{N}$ such that
$\widetilde{r}(\varepsilon,s)+\frac{\eta}{4}> \widetilde{r}_k(\varepsilon,s)$ for all $s\in[0,T]$ and $k\ge N(\varepsilon,r,t)$.
If $(\zeta,s)$ is such that $\zeta> \frac{\eta}{2} + \widetilde{r}(\varepsilon,s)$,
then $\zeta > \frac{\eta}{4} + \widetilde{r}(\varepsilon,s) > \widetilde{r}_k(\varepsilon,s)$ for all $k\ge N(\varepsilon,r,t)$.
We combining this with \eqref{temp:veDomain} to conclude the proof.
\end{proof}
\begin{lemma}\label{lemma:rhoWeak}
There exist $\rho\ge 0$
and a subsequence $($still denoted$)$ $\{\rho_k\}_{k\in\mathbb{N}}$ such that
\begin{equation}\label{temp:rhoWeak0}
(\rho_k-1) \overset{\ast}{\rightharpoonup} (\rho-1)
\qquad \text{in $L^{\infty}\big( 0,T ; L^2( [a,\infty), r^m {\mathrm d} r ) \big)$}.
\end{equation}
Moreover, for each $\varepsilon>0$,
\begin{equation}\label{temp:rhoWeak}
\begin{dcases*}
C^{-1}(\varepsilon)\le \rho(r,t) \le C(\varepsilon)
&\,\, for a.e. $(r,t)\in\{(\zeta,s)\in[a,\infty)\times[0,T]\,\vcentcolon\,\zeta\ge \widetilde{r}(\varepsilon,t)\}$,\\
C^{-1}(a)\le \rho(r,t) \le C(a) &\,\, for a.e. $(r,t)\in[a,\infty)\times[0,T]$.
\end{dcases*}
\end{equation}
\end{lemma}
\begin{proof}
From Lemma \ref{lemma:HREext},
\begin{equation}\label{temp:rho1}
\sup_{k\in\mathbb{N}}\sup_{ 0\le t \le T}\int_{a}^{\infty}\snorm{\rho_{k}-1}^2(r,t)\,{\mathrm d} r {\mathrm d} t
\le a^{-m}\sup_{k\in\mathbb{N}}\sup_{ 0\le t \le T}\int_{a}^{\infty}\snorm{\rho_{k}-1}^2(r,t)\,r^m{\mathrm d} r {\mathrm d} t\le C(a).
\end{equation}
Thus there exist $(f,\rho-1) \in L^{\infty}\big(0,T;L^2([a,\infty),{\mathrm d} r)\big)$ and a subsequence (still denoted) $\{\rho_k\}_{k\in\mathbb{N}}$ such that, as $k\to\infty$,
\begin{equation}\label{temp:rhowc}
\rho_k-1\overset{\ast}{\rightharpoonup} (\rho-1), \quad
r^{\frac{m}{2}} (\rho_k-1) \overset{\ast}{\rightharpoonup} f
\qquad\,\text{ in $L^{\infty}\big(0,T;L^2( [a,\infty), {\mathrm d} r)\big)$}.
\end{equation}
Let $\phi\in C^{\infty}_{\rm c}([a,\infty)\times[0,T])$.
It follows from \eqref{temp:rhowc} that, as $k\to\infty$,
\begin{align*}
\int_{0}^{T} \int_{a}^{\infty} f \phi\,{\mathrm d} r {\mathrm d} t
\leftarrow \int_{0}^{T} \int_{a}^{\infty} r^{\frac{m}{2}}(\rho_k -1 ) \phi\,{\mathrm d} r {\mathrm d} t
\to \int_{0}^{T} \int_{a}^{\infty} (\rho -1 ) r^{\frac{m}{2}} \phi\, {\mathrm d} r {\mathrm d} t.
\end{align*}
By the fundamental lemma of calculus,
$f(r,t)= r^{\frac{m}{2}} (\rho(r,t)-1)$ for {\it a.e.} $(r,t)\in[a,\infty)\times[0,T]$.
By \eqref{temp:rhowc} again, the weak-star convergence \eqref{temp:rhoWeak0} is obtained.
Next, fix $\varepsilon\in(0,1]$ and a point $(r,t)\in[a,\infty)\times(0,T]$ such that $r>\widetilde{r}(\varepsilon,t)$.
Let $\delta(\varepsilon,r,t)>0$ and $N(\varepsilon,r,t)\in\mathbb{N}$ be obtained in Proposition \ref{prop:veDomain}.
For $0<\delta \le \delta(\varepsilon,r,t)$, set
\begin{equation}\label{rhobdtemp3}
\phi_{\delta}^{(r,t)}(\zeta,s)\vcentcolon= \dfrac{\mathbbm{1}_{{\mathrm D}_{\delta}(r,t)}(\zeta,s)}{\snorm{{\mathrm D}_\delta(r,t)}} = \begin{dcases*}
\snorm{{\mathrm D}_\delta(r,t)}^{-1} & if $(\zeta,s)\in {\mathrm D}_\delta(r,t)$,\\
0 & if $(\zeta,s)\in [a,\infty)\times[0,T] \backslash {\mathrm D}_\delta(r,t)$,
\end{dcases*}
\end{equation}
where ${\mathrm D}_\delta(r,t)$ is defined in Proposition \ref{prop:veDomain}. It follows that
$\phi_\delta^{(r,t)}\in L^1\big(0,T;L^2( [a,\infty),{\mathrm d} r )\big)$.
Moreover, by Proposition \ref{prop:veDomain}, for all $k\ge N(\varepsilon,r,t)$ and $\delta\le \delta(\varepsilon,r,t)$,
\begin{equation}\label{rhobdtemp1}
{\mathrm {supp}}(\phi_\delta^{(r,t)}) \subset \{ (\zeta,s)\in[a,\infty)\times(0,T]\,\vcentcolon\,\zeta > \widetilde{r}_k(\varepsilon,s) \}.
\end{equation}
For $0<\delta\le \delta(\varepsilon,r,t)$ and $\eta>0$,
there exists $N_{\eta,\delta}\in\mathbb{N}$ by \eqref{temp:rhowc} such that, for all $k\ge N_{\eta,\delta}$,
\begin{equation}\label{rhobdtemp2}
-\eta + \int_{0}^{T}\int_{a}^{\infty} \rho_{k} \phi_\delta^{(r,t)} {\mathrm d} \zeta {\mathrm d} s \le \int_{0}^{T}\int_{a}^{\infty} \rho\phi_\delta^{(r,t)} {\mathrm d} \zeta {\mathrm d} s \le \eta + \int_{0}^{T}\int_{a}^{\infty} \rho_{k}\phi_\delta^{(r,t)} {\mathrm d} \zeta {\mathrm d} t.
\end{equation}
Let $N_1\vcentcolon= \max\{ N_{\eta,\delta}, N(\varepsilon,r,t) \}$. Then, by \eqref{rhobdtemp1} and Lemma \ref{lemma:HREext},
it follows that, for all $k\ge N_1$, $C^{-1}(\varepsilon) \le \rho_k(\zeta,s)\le C(\varepsilon)$ if $(\zeta,s)\in {\mathrm {supp}}(\phi_\delta^{(r,t)})$.
Using this in \eqref{rhobdtemp2} yields
\begin{align*}
C^{-1}(\varepsilon) - \eta \le \dfrac{1}{\snorm{{\mathrm D}_{\delta}(r,t)}} \iint_{{\mathrm D}_{\delta}(r,t)} \rho(\zeta,s)\,{\mathrm d} \zeta {\mathrm d} s
\le C(\varepsilon) + \eta
\qquad \text{for $\eta>0$ and $\delta\in (0,\delta(\varepsilon,r,t)]$}.
\end{align*}
By the Lebesgue differentiation theorem over $\delta$ and the fact that $\eta>0$ is arbitrarily chosen,
$\eqref{temp:rhoWeak}_1$ is proved.
Finally, we consider the proof for $\eqref{temp:rhoWeak}_2$. For $(r,t)\in[a,\infty)\times[0,T]$ and $\delta>0$,
we use $\phi_\delta^{(r,t)}$ defined in \eqref{rhobdtemp3}.
Then $\eqref{temp:rhoWeak}_2$ can be proved by repeating the same argument as before
and using $C^{-1}(a) \le \rho_k(r,t) \le C(a)$ for all $(k,r,t)\in\mathbb{N}\times[a,\infty)\times[0,T]$
from Lemma \ref{lemma:HREext}.
\end{proof}
We now introduce the following function spaces: for each connected interval $I\subseteq [0,\infty)$, define the space $\widetilde{H}_0^1(I,r^m{\mathrm d} r)$ to be the closure of
\begin{equation}\label{eqs:tildeH1}
\mathcal{D}_0(I)\vcentcolon=\{ \phi\in C^{\infty}(I)\,\vcentcolon\,\exists N>0 \ \text{such that} \ [0,N]\subset I \ \text{and} \ \phi(r)=0 \ \text{for} \ r\in I\cap[N,\infty) \}
\end{equation}
via the $H^1(I,r^m{\mathrm d} r)$--norm. We also denote $\widetilde{H}^{-1}( I, r^m {\mathrm d} r)$ as its dual space.
\begin{lemma}\label{lemma:rhoScvg}
Let $(\rho,u,e)$ be the limit function obtained in {\rm Lemma \ref{lemma:klimholder}}--{\rm \ref{lemma:rhoWeak}}.
Then
\begin{enumerate}[label=(\roman*),ref=(\roman*),font={\normalfont\rmfamily}]
\item\label{item:rhoScvg1}
For all $L\in\mathbb{N}$, $\rho\in C^0\big( [0,T]; \widetilde{H}^{-1}([a,L],r^m {\mathrm d} r) \big)$ and
there exists a further subsequence $($still denoted$)$ $\rho_{k}(r,t)$ such that
\begin{equation*}
\lim\limits_{k\to\infty} \sup_{t\in[0,T]}\sbnorm{\rho_{k}(\cdot,t)- \rho (\cdot,t)}_{\widetilde{H}^{-1}([a,L], r^m {\mathrm d} r)} = 0.
\end{equation*}
\item\label{item:rhoScvg2}
For each $\varepsilon\in(0,1]$, define $\rho^{(\varepsilon)} (r,t) \vcentcolon = \rho(r,t) \chi_\varepsilon(r,t)$ with $\chi_\varepsilon(r,t)$
given by
\begin{equation}\label{temp:chive}
\chi_\varepsilon(r,t) \vcentcolon = \begin{dcases*}
1 & if $t\in[0,T]$ and $r\in [\widetilde{r}(\varepsilon,t),\infty)$,\\
0 & otherwise.
\end{dcases*}
\end{equation}
Then, for all $L\in\mathbb{N}$,
\begin{align*}
\begin{dcases}
\rho^{(\varepsilon)} \in C^0\big( [0,T] ; \widetilde{H}^{-1}([0,L], r^m {\mathrm d} r) \big) ,\\
\sbnorm{\rho^{(\varepsilon)} (\cdot,t_1)- \rho^{(\varepsilon)}(\cdot,t_2)}_{\widetilde{H}^{-1}([0,L],r^m{\mathrm d} r)} \le C(\varepsilon)\snorm{t_1-t_2}
\qquad \text{for $t_1, t_2 \in [0,T]$}.
\end{dcases}
\end{align*}
\end{enumerate}
\end{lemma}
\begin{proof}
Let $\phi\in \widetilde{H}^1_0([a,\infty),r^m {\mathrm d} r)$.
Then, by \eqref{eqs:extfunc} and \eqref{eqs:coorderiv}--\eqref{eqs:extcut},
\begin{align*}
&\int_{a}^{\infty} \big\{\rho_k(r,t_2) - \rho_k(r,t_1)\big\} \phi(r)\,r^m {\mathrm d} r
= \int_{t_1}^{t_2} \int_{a}^{\infty} \partial_t \rho_k (r,t) \phi(r)\,r^m {\mathrm d} r {\mathrm d} t\\
&= \int_{t_1}^{t_2} \int_{a}^{\infty} (\varphi_k \partial_t \overline{v}_k^{-1})(r,t) \phi(r)\,r^m {\mathrm d} r {\mathrm d} t = \int_{t_1}^{t_2} \int_{a}^{\widetilde{r}_k(k,t)} (\varphi_k \partial_t \overline{v}_k^{-1})(r,t) \phi(r)\,r^m {\mathrm d} r {\mathrm d} t\\
&= - \int_{t_1}^{t_2} \int_{0}^{k} (\varphi_k \,\phi)(\widetilde{r}_k(x,t),t) \big(\widetilde{v}_k^{-1} {\mathrm D}_t \widetilde{v}_k
+\widetilde{r}_k^m \widetilde{u}_k{\mathrm D}_x\widetilde{v}_k^{-1}\big)(x,t)\, {\mathrm d} x {\mathrm d} t.
\end{align*}
Integrating by parts and applying the continuity equation $\eqref{eqs:LFNS-k}_1$, the boundary condition
$\widetilde{u}_k(0,t)=\widetilde{u}_k(k,t)$, and Theorem \ref{thm:priori}\ref{item:prio0}--\ref{item:prio1}, we have
\begin{align*}
&\int_{a}^{\infty} \big(\rho_k(r,t_2) - \rho_k(r,t_1)\big) \phi(r)\, r^m {\mathrm d} r\\
&= \int_{t_1}^{t_2} \int_{0}^{k} \widetilde{u}_k (x,t) \big(\partial_r\varphi_k\, \phi
+ \partial_r\phi\,\varphi_k \big)(\widetilde{r}_k(x,t),t)\,{\mathrm d} x {\mathrm d} t \\
&\le C\int_{t_1}^{t_2} \Big( \int_{0}^{k} \dfrac{\snorm{\widetilde{u}_k}^2}{\widetilde{v}_k}{\mathrm d} x\Big)^{\frac{1}{2}}
\Big(\int_{0}^{k}\widetilde{v}_k (x,t)\big(\snorm{\phi}^2 +\snorm{\partial_r\phi}^2\big)(\widetilde{r}_k(x,t))\,{\mathrm d} x \Big)^{\frac{1}{2}} {\mathrm d} t\\
&\le C(a) \bnorm{\phi(\cdot)}_{H^1([a,\infty),r^m{\mathrm d} r)} \snorm{t_2-t_1}.
\end{align*}
Thus, the following equicontinuity inequality holds:
\begin{equation*}
\sup_{k\in\mathbb{N}} \sbnorm{\rho_k(\cdot,t_2)-\rho_k(\cdot,t_1)}_{\widetilde{H}^{-1}([a,\infty),r^m {\mathrm d} r)} \le C(a) \snorm{t_2-t_1}.
\end{equation*}
By the same calculation, $\rho_k \in L^{\infty}(0,T;\widetilde{H}^{-1}([a,L],r^m {\mathrm d} r))$ for all $(k,L)\in\mathbb{N}^2$.
In addition, it follows from Lemma \ref{lemma:HREext} that
\begin{equation*}
\sup_{k\in\mathbb{N}} \sup_{t\in[0,T]} \int_{a}^{\infty} \snorm{\rho_k -1 }^2 (r,t)\, r^m {\mathrm d} r\le C(a).
\end{equation*}
Then we can apply Proposition \ref{prop:scHm1}
to obtain $f$ such that, for all $L\in \mathbb{N}$,
\begin{equation*}
\begin{dcases}
f\in C^0\big( [0,T] ; \widetilde{H}^{-1}([a,L], r^m {\mathrm d} r) \big), \ \ f-1\in L^{\infty}( 0,T; L^2( [a,\infty), r^m {\mathrm d} r ) ),\\
\lim\limits_{k\to \infty}\sup\limits_{t\in[0,T]}\sbnorm{\rho_{k}(\cdot,t)-f(\cdot,t)}_{\widetilde{H}^{-1}( [a,L], r^m {\mathrm d} r)} = 0,\\
\sbnorm{f(\cdot,t_1)-f(\cdot,t_2)}_{\widetilde{H}^{-1}( [a,L], r^m {\mathrm d} r)} \le C(a) \snorm{t_1-t_2}
\qquad \text{for all $t_1,\,t_2 \in[0,T]$.}
\end{dcases}
\end{equation*}
By Lemma \ref{lemma:rhoWeak} and the fundamental lemma of calculus, we see that $\rho=f$ for {\it a.e.} $(r,t)\in[a,\infty)\times[0,T]$.
This completes the proof of \ref{item:rhoScvg1}.
For each $\varepsilon>0$ and $k\in\mathbb{N}$, define $\rho_k^{(\varepsilon)} (r,t) \vcentcolon = \rho_k(r,t) \chi_\varepsilon^k(r,t)$
with
\begin{equation}\label{temp:chivek}
\chi_\varepsilon^k(r,t) \vcentcolon = \begin{dcases*}
1 & if $t\in[0,T]$ and $r\in [\widetilde{r}_k(\varepsilon,t),\infty)$,\\
0 & otherwise.
\end{dcases*}
\end{equation}
If $k\ge M(\varepsilon,a)\vcentcolon= \varepsilon C^2(a) + \frac{2^n -1}{n}a^n C(a)>0$,
then it follows from construction \eqref{eqs:COfunc} that $\varphi_k(\widetilde{r}_k(\varepsilon,t),t) = 1$ for all $t\in[0,T]$.
Thus, by \eqref{eqs:extfunc} and \eqref{eqs:transfunc},
\begin{equation}\label{temp:vekMa}
(\rho_k,u_k,e_k)(\widetilde{r}_k(\varepsilon,t),t) = (\overline{v}_k^{-1},\overline{u}_k,\overline{e}_k)(\widetilde{r}_k(\varepsilon,t),t) = (\widetilde{v}_k^{-1},\widetilde{u}_k,\widetilde{e}_k)(\varepsilon,t).
\end{equation}
Let $\phi\in \widetilde{H}^1_0([0,\infty),r^m {\mathrm d} r)$. Then, by \eqref{temp:vekMa} and the Leibniz rule,
for all $k\ge M(\varepsilon,a)$,
\begin{align}\label{veSctemp0}
&\int_{0}^{\infty} \big(\rho_k^{(\varepsilon)}(r,t_2) - \rho_k^{(\varepsilon)}(r,t_1)\big) \phi(r)\,r^m {\mathrm d} r \nonumber\\
&= \int_{t_1}^{t_2}\int_{\widetilde{r}_k(\varepsilon,t)}^{\infty} \phi \partial_t \rho_k\,r^m {\mathrm d} r {\mathrm d} t
- \int_{t_1}^{t_2} \widetilde{r}_k^{\,m}(\varepsilon,t) \phi(\widetilde{r}_k(\varepsilon,t)) \widetilde{u}_k(\varepsilon,t) \widetilde{v}_k^{-1}(\varepsilon,t)\,{\mathrm d} t\nonumber\\
&=\vcentcolon I_1 + I_2.
\end{align}
Using \eqref{eqs:extfunc} and \eqref{eqs:coorderiv}--\eqref{eqs:transfunc}, $I_1$ can be written as
\begin{align*}
I_1
&= \int_{t_1}^{t_2} \int_{\widetilde{r}_k(\varepsilon,t)}^{\widetilde{r}_k(k,t)} \varphi_k \partial_t \overline{v}_k^{-1} (r,t) \phi(r)\, r^m {\mathrm d} r {\mathrm d} t\\
&= \int_{t_1}^{t_2} \int_{\varepsilon}^{k} \big(\varphi_k \, \phi\big)(\widetilde{r}_k(x,t),t) \big\{ \widetilde{v}_k {\mathrm D}_t \widetilde{v}_k^{-1} - \widetilde{r}_k^m \widetilde{u}_k {\mathrm D}_x \widetilde{v}_k^{-1} \big\}(x,t)\,{\mathrm d} x {\mathrm d} t.
\end{align*}
Integrating by parts and using the boundary condition $\widetilde{u}_k(k,t)=0$, the continuity equation $\eqref{eqs:LFNS-k}_1$, and Theorem \ref{thm:priori}\ref{item:prio0}--\ref{item:prio1},
it follows that
\begin{align*}
I_1 &= \int_{t_1}^{t_2} \int_{\varepsilon}^{k} (\varphi_k\cdot \phi)(\widetilde{r}_k(x,t),t) \big\{ - \widetilde{v}_k^{-1}{\mathrm D}_x(\widetilde{r}_k^m\widetilde{u}_k)-\widetilde{r}_k^m\widetilde{u}_k{\mathrm D}_x\widetilde{v}_k^{-1} \big\}\,{\mathrm d} x {\mathrm d} t\\
&=\int_{t_1}^{t_2} \phi(\widetilde{r}_k(\varepsilon,t)) (\widetilde{r}_k^m\widetilde{u}_k \widetilde{v}_k^{-1})(\varepsilon,t)\,{\mathrm d} t
+ \int_{t_1}^{t_2} \int_{\varepsilon}^{k} \widetilde{u}_k (x,t)
\big(\partial_r\varphi_k\, \phi + \partial_r\phi\,\varphi_k \big)(\widetilde{r}_k(x,t),t)
\,{\mathrm d} x {\mathrm d} t\\
&\le -I_2 + C\int_{t_1}^{t_2}\Big(\int_{\varepsilon}^{k}\dfrac{\snorm{\widetilde{u}_k}^2}{\widetilde{v}_k}{\mathrm d} x\Big)^{\frac{1}{2}} \Big(\int_{\varepsilon}^{k} \widetilde{v}_k (x,t) \big(\snorm{\phi}^2 + \snorm{\partial\phi}^2 \big)(\widetilde{r}_k(x,t))\,
{\mathrm d} x \Big)^{\frac{1}{2}} {\mathrm d} t\\
&\le - I_2 + C(\varepsilon) \sbnorm{\phi}_{H^1([0,\infty),r^m{\mathrm d} r)} \snorm{t_2-t_1}.
\end{align*}
Substituting this back into \eqref{veSctemp0}, it follows that
the following equicontinuity inequality holds:
\begin{equation}\label{veSctemp1}
\sup_{k\in\mathbb{N}} \sbnorm{\rho_k^{(\varepsilon)}(\cdot,t_2)-\rho_k^{(\varepsilon)}(\cdot,t_1)}_{\widetilde{H}^{-1}([0,\infty),r^m {\mathrm d} r)} \le C(\varepsilon) \snorm{t_2-t_1}.
\end{equation}
Let $L\in\mathbb{N}$ be such that $\widetilde{r}_k(1,t)< L$ for all $(k,t)\in\mathbb{N}\times[0,T]$.
By the same derivation, $\rho_k^{(\varepsilon)} \in L^{\infty}( 0,T; \widetilde{H}^{-1}([0,L],r^m {\mathrm d} r) )$ for all $(k,L)\in\mathbb{N}$. Moreover, set $r_{\varepsilon}=\frac{1}{2}\inf_{t\in[0,T]}\widetilde{r}(\varepsilon,t)>0$.
Then, by Lemma \ref{lemma:rkcvg}, there exists $N_\varepsilon\in\mathbb{N}$ such that $ r_{\varepsilon}<\widetilde{r}_k(\varepsilon,t)$
for all $k\ge N_\varepsilon$ and $t\in[0,T]$. From \eqref{veSctemp1} and Lemma \ref{lemma:HREext}, it follows that
\begin{equation*}
\begin{dcases}
\sup_{k\ge N_\varepsilon}\sup\limits_{t\in[0,T]} \int_{r_\varepsilon}^{\infty} \snorm{\rho^{(\varepsilon)}_k -1}^2(r,t)\,r^m {\mathrm d} r \le C(\varepsilon),\\
\sup_{k\ge N_\varepsilon} \sbnorm{\rho_k^{(\varepsilon)}(\cdot,t_2)-\rho_k^{(\varepsilon)}(\cdot,t_1)}_{\widetilde{H}^{-1}([r_\varepsilon,L],r^m {\mathrm d} r)}
\le C(\varepsilon) \snorm{t_2-t_1}.
\end{dcases}
\end{equation*}
Thus, we can apply Proposition \ref{prop:scHm1}
to obtain $f_\varepsilon$
such that, for all $L\in \mathbb{N}$,
\begin{equation}\label{veSctemp2}
\begin{dcases}
f_\varepsilon\in C^0\big( [0,T] ; \widetilde{H}^{-1}( [r_\varepsilon,L], r^m {\mathrm d} r) \big), \ \ f_\varepsilon-1\in L^{\infty}\big( 0,T; L^2( [r_\varepsilon,\infty), r^m {\mathrm d} r) \big),\\
\lim\limits_{k\to \infty}\sup\limits_{t\in[0,T]}\sbnorm{\rho_{k}^{(\varepsilon)}(\cdot,t)-f_\varepsilon(\cdot,t)}_{\widetilde{H}^{-1}( [r_\varepsilon,L], r^m {\mathrm d} r)} = 0,\\
\sbnorm{f_\varepsilon(\cdot,t_1)-f_\varepsilon(\cdot,t_2)}_{\widetilde{H}^{-1}( [r_\varepsilon,L], r^m {\mathrm d} r )} \le C(\varepsilon) \snorm{t_1-t_2}
\qquad \text{for all $t_1,\,t_2 \in[0,T]$.}
\end{dcases}
\end{equation}
Let $\phi\in C^{\infty}_c([a,\infty)\times[0,T])$. Then
\begin{equation*}
\begin{aligned}
&\int_{0}^{T}\int_{r_\varepsilon}^{\infty} \phi\big( \rho^{(\varepsilon)} - f_\varepsilon \big)\,{\mathrm d} r {\mathrm d} t\\
&= \int_{0}^{T}\int_{r_\varepsilon}^{\infty} \phi \chi_\varepsilon ( \rho - \rho_k )\,{\mathrm d} r {\mathrm d} t
+ \int_{0}^{T}\int_{r_\varepsilon}^{\infty} \phi \rho_k ( \chi_\varepsilon - \chi_\varepsilon^k )\,{\mathrm d} r {\mathrm d} t
+ \int_{0}^{T}\int_{r_\varepsilon}^{\infty} \phi ( \rho_k^{(\varepsilon)} - f_\varepsilon )\,{\mathrm d} r {\mathrm d} t \\
&=: \sum_{i=1}^{3} \text{II}_i^k.
\end{aligned}
\end{equation*}
By Lemma \ref{lemma:rhoWeak}, $\text{II}_1^k \to 0$ as $k\to\infty$.
By \eqref{temp:chive}--\eqref{temp:chivek} and Lemma \ref{lemma:rkcvg},
$\chi_{\varepsilon}^{k}\to\chi_\varepsilon$ {\it a.e.} as $k\to\infty$.
Since $\rho_k \le C(a)$ from Lemma \ref{lemma:HREext},
by the dominated convergence theorem,
$\text{II}_2^k \to 0$ as $k\to\infty$.
Finally, $\text{II}_3^k \to 0$ as $k\to\infty$ by \eqref{veSctemp2}.
Therefore,
$\rho^{(\varepsilon)}=f_\varepsilon$ for {\it a.e.} $(r,t)\in[r_\varepsilon,\infty)\times[0,T]$.
Since $\rho^{(\varepsilon)}=\rho^{(\varepsilon)}_k=0$ for $r\in[0,r_\varepsilon]$ and $k\ge N_\varepsilon$,
Lemma \ref{lemma:rhoScvg}\ref{item:rhoScvg2} is proved.
\end{proof}
\begin{lemma}\label{lemma:Relation}
The following relations hold{\rm :}
\begin{equation}\label{temp:relation}
\begin{aligned}
&x = \int_{a}^{\widetilde{r}(x,t)} \rho(r,t)\,r^m {\mathrm d} r && \text{ for a.e. $t\in [0,T]$ and for each $x>0$,}\\
&\widetilde{r}(x,t) = \widetilde{r}_a^0(x) + \int_{0}^{t} u( \widetilde{r}(x,s), s)\,{\mathrm d} s
&& \text{ for all $(x,t)\in[0,\infty)\times[0,T]$,}
\end{aligned}
\end{equation}
where the initial data $\widetilde{r}_a^0(x)$ is the function constructed in {\rm \S \ref{subsec:ffaprox}},
which satisfies
\begin{equation*}
x = \int_{a}^{\widetilde{r}_a^0(x)} \rho_a^0(r)\,r^m {\mathrm d} r \qquad \text{for all $x\in[0,\infty)$}.
\end{equation*}
\end{lemma}
\begin{proof}
Let $(x,t)\in [0,\infty)\times[0,T]$. Then, for $k \ge M(a,x) \vcentcolon=\frac{2^n-1}{n}C(a)a^n + 2^n C(a)^2 x$,
$2\widetilde{r}_k(x,t) \le \widetilde{r}_k(k,t)$.
Thus, by \eqref{eqs:extfunc}, if $k\ge M(a,x)$ and $r\in[a,\widetilde{r}_k(x,t)]$, then $\rho_k (r,t) = \varphi_k \overline{v}_k^{-1} (r,t) + ( 1- \varphi_k(r,t) ) = \overline{v}_k^{-1} (r,t)$.
Using this and \eqref{eqs:Jac}--\eqref{eqs:transfunc}, we have
\begin{equation*}
\int_{a}^{\widetilde{r}_k(x,t)} \rho_k(r,t)\,r^m {\mathrm d} r
= \int_{a}^{\widetilde{r}_k(x,t)} \overline{v}_k^{-1} (r,t)\,r^m {\mathrm d} r
= \int_{0}^{x} \widetilde{v}_k^{-1}(y,t) \widetilde{v}_k (y,t)\,{\mathrm d} y = x.
\end{equation*}
Using the above identity leads to
\begin{align*}
x-\int_{a}^{\widetilde{r}(x,t)} \rho(r,t)\,r^m {\mathrm d} r
= \int_{\widetilde{r}(x,t)}^{\widetilde{r}_k(x,t)} \rho_k (r,t)\,r^m {\mathrm d} r
+\int_{a}^{\widetilde{r}(x,t)} ( \rho_k - \rho)(r,t)\, r^m {\mathrm d} r.
\end{align*}
Let $\chi(t)\in C^{\infty}_c([0,T])$. Multiplying it to the above and integrating in time,
it follows that
\begin{align*}
&\int_{0}^{T} \Big( x-\int_{a}^{\widetilde{r}(x,t)} \rho(r,t)\,r^m {\mathrm d} r \Big) \chi(t) {\mathrm d} t \\
&= \int_{0}^{T}\int_{\widetilde{r}(x,t)}^{\widetilde{r}_k(x,t)} \rho_k (r,t) \chi(t)\,r^m {\mathrm d} r {\mathrm d} t
+ \int_{0}^{T}\int_{a}^{\widetilde{r}(x,t)} (\rho_k -\rho)(r,t) \chi(t)\,r^m {\mathrm d} r {\mathrm d} t\\
&=: I_k^{(1)} + I_k^{(2)}.
\end{align*}
$I_k^{(1)}$ is estimated by using Lemma \ref{lemma:rkcvg} and the upper bound $\rho_k \le C(a)$
from Lemma \ref{lemma:HREext}:
\begin{align*}
\snorm{I_k^{(1)}}
\le \dfrac{C(a)T}{n}\sbnorm{\chi(\cdot)}_{L^\infty}
\sup_{ 0\le t \le T}\snorm{\widetilde{r}_k^n(x,t)-\widetilde{r}^n(x,t)} \to 0 \qquad\text{as $k\to\infty$}.
\end{align*}
For $I_k^{(2)}$, observe that
$\chi(t) r^m \mathbbm{1}_{[a,\widetilde{r}(x,t)]}(r) \in L^{1}(0,T;L^2( [a,\infty), {\mathrm d} r))$, where $\mathbbm{1}_E(r)$ denotes the indicator function for spatial set $E\subset[a,\infty)$.
By Lemma \ref{lemma:rhoWeak}, it follows that
\begin{align*}
\lim\limits_{k\to\infty} I_k^{(2)}
=\lim\limits_{k\to\infty}\int_{0}^{T}\int_{a}^{\infty}(\rho_k-\rho)(r,t) \chi(t)
\mathbbm{1}_{[a,\widetilde{r}(x,t)]}(r)\,r^m {\mathrm d} r {\mathrm d} t = 0.
\end{align*}
Since $|I_k^{(1)}|+|I_k^{(2)}|\to 0$ as $k\to\infty$, then
\begin{align*}
\int_{0}^{T} \Big( x-\int_{a}^{\widetilde{r}(x,t)} \rho(r,t) r^m {\mathrm d} r \Big) \chi(t)\,{\mathrm d} t=0
\qquad \text{for each $\chi(t) \in C_c^{\infty}([0,T])$}
\end{align*}
so that $\eqref{temp:relation}_1$ follows.
Next, by $\eqref{eqs:LFNS-k}_1$ and $\eqref{eqs:LFNS-k}_4$, $\widetilde{r}_k(x,t)$ satisfies
\begin{equation*}
\widetilde{r}_k(x,t) = \widetilde{r}_k(x,0) + \int_{0}^{t} \widetilde{u}_k(x,s)\,{\mathrm d} s
\qquad\text{ for $(x,t)\in[0,k]\times[0,T]$},
\end{equation*}
where, from $\eqref{eqs:LFNS-k}_4$ and \S \ref{subsec:ffaprox}, $\widetilde{r}_k(x,0)$ is given by
\begin{equation*}
\widetilde{r}_k(x,0) = \Big( a^n + n \int_{0}^{x} \widetilde{v}_{a,k}^0(y)\,{\mathrm d} y \Big)^{\frac{1}{n}}
= \Big( a^n + n \int_{0}^{x} \dfrac{1}{\rho_{a,k}^0(\widetilde{r}_a^0(y))} {\mathrm d} y \Big)^{\frac{1}{n}}.
\end{equation*}
Fix $(x,t)\in[0,\infty)\times[0,T]$. If $k \ge M(a,x)$, then, by \eqref{eqs:extfunc},
$u_k(\widetilde{r}_k(x,t),t) = \widetilde{u}_k(x,t)$ and
\begin{equation}\label{temp:rela1}
\widetilde{r}_k(x,t) = \widetilde{r}_k(x,0) + \int_{0}^{t} u_k(\widetilde{r}_k(x,s),s)\,{\mathrm d} s.
\end{equation}
Moreover, by Proposition \ref{prop:r0diff} in \S \ref{subsec:ffaprox},
if $k \ge N(x) \vcentcolon=\frac{2^n-1}{n}C_0a^n+ 2^n C_0^2 x$, then
\begin{align*}
\dfrac{1}{2}\widetilde{r}_a^0(k)
=& \dfrac{1}{2} \Big( a^n + n \int_{0}^{k} \dfrac{1}{\rho_{a,k}^0(\widetilde{r}_a^0(y))} {\mathrm d} y \Big)^{\frac{1}{n}}
\ge \dfrac{1}{2}\big( a^n + n C_0^{-1} k \big)^{\frac{1}{n}} \\
&\ge ( a^n + n C_0 x )^{\frac{1}{n}} \ge \Big( a^n + n \int_{0}^{x} \dfrac{1}{\rho_a^0(\widetilde{r}_a^0(y))} {\mathrm d} y\Big)^{\frac{1}{n}} =\widetilde{r}_a^0(x).
\end{align*}
By the construction of $\rho_k^0(r)$ in \eqref{def:appInitr1},
it follows that, if $k \ge N(x)$, then
\begin{align*}
\widetilde{v}_{a,k}^0(x) = \dfrac{1}{\rho_{a,k}^0(\widetilde{r}_a^0(x))}
= \Big( \rho_a^0(\widetilde{r}_a^0(x)) \varphi_{a,k}^0(\widetilde{r}_a^0(x))
+ 1 - \varphi_{a,k}^0(\widetilde{r}_a^0(x)) \Big)^{-1}
= \dfrac{1}{\rho_a^0(\widetilde{r}_a^0(x))}.
\end{align*}
Thus, by Proposition \ref{prop:r0diff}, for all $k\ge N(x)$,
\begin{align*}
\widetilde{r}_k(x,0)
= \Big( a^n + n \int_{0}^{x} \dfrac{1}{\rho_{a,k}^0(\widetilde{r}_a^0(y))} {\mathrm d} y \Big)^{\frac{1}{n}}
= \Big( a^n + n \int_{0}^{x} \dfrac{1}{\rho_a^0(\widetilde{r}_a^0(y))} {\mathrm d} y \Big)^{\frac{1}{n}}
= \widetilde{r}_a^0(x).
\end{align*}
This shows that, for each $x\in[0,\infty)$, $\widetilde{r}_k(x,0)\to \widetilde{r}_a^0(x)$ as $k\to\infty$.
Moreover, by Lemma \ref{lemma:rkcvg} and inequality \eqref{aholtemp2} obtained in Lemma \ref{lemma:klimholder},
it follows that, for all $(x,s)\in[0,\infty)\times[0,T]$,
\begin{equation*}
\snorm{u_k(\widetilde{r}_k(x,s),s) - u_k(\widetilde{r}(x,s),s)} \le C(a) \snorm{\widetilde{r}_k(x,s)-\widetilde{r}(x,s)} \to 0
\qquad\text{as $k\to\infty$.}
\end{equation*}
By Lemma \ref{lemma:klimholder}, $\snorm{u_k(\widetilde{r}(x,s),s)-u(\widetilde{r}(x,s),s)} \to 0 $ for all $(x,s)\in[0,\infty)\times[0,T]$ as $k\to \infty$.
Combining these limits in \eqref{temp:rela1},
$\eqref{temp:relation}_2$ is shown by the dominated convergence theorem.
\end{proof}
\begin{lemma}\label{lemma:weak}
There exists a subsequence $($still denoted$)$ $(\rho_k,u_k, e_k )_{k\in\mathbb{N}}$ such that, as $k\to \infty$,
\begin{equation*}
\begin{dcases*}
(\sqrt{\sigma}\partial_r u_k, \sigma\partial_r e_k ) \overset{\ast}{\rightharpoonup} ( \sqrt{\sigma}\partial_r u, \sigma\partial_r e ) & in $L^{\infty}\big(0,T;L^2([a,\infty),r^m {\mathrm d} r)\big)$,\\
( \partial_r u_k, \partial_r e_k, \sqrt{\sigma}\partial_t u_k, \sigma \partial_t e_k )
\rightharpoonup ( \partial_r u, \partial_r e, \sqrt{\sigma}\partial_t u, \sigma \partial_t e )
& in $L^{2}\big(0,T;L^2([a,\infty), r^m {\mathrm d} r)\big)$,
\end{dcases*}
\end{equation*}
where $\sigma=\min\{1,t\}$. In addition,
\begin{align}\label{weakstmt1}
&\sup_{t\in[0,T]}\int_{a}^{\infty}
\big((\rho-1)^2 + u^4 + (e-1)^2 + \sigma\snorm{\partial_r u}^2 +\sigma^2 \snorm{\partial_r e}^2 \big)(r,t)\,r^m {\mathrm d} r\nonumber\\
& +\int_{0}^{T}\int_{a}^{\infty} \Big\{ \snorm{\partial_r u}^2 + \snorm{u\partial_r u}^2 + \snorm{\partial_r e}^2 + \sigma\snorm{\partial_t u}^2 + \sigma^2 \snorm{\partial_t e}^2 \Big\}\,r^m {\mathrm d} r {\mathrm d} t \le C(a).
\end{align}
Furthermore, for each $\varepsilon>0$,
\begin{align}\label{weakstmt2}
&\sup_{ 0\le t \le T} \int_{\widetilde{r}(\varepsilon,t)}^{\infty}
\big((\rho-1)^2 + u^4 + (e-1)^2 + \sigma\snorm{\partial_r u}^2
+ \sigma^2 \snorm{\partial_r e}^2 \big)(r,t)\,r^m {\mathrm d} r \nonumber \\
& +\int_{0}^{T} \int_{\widetilde{r}(\varepsilon,t)}^{\infty} \Big\{ \snorm{\partial_r u}^2 + \snorm{u\partial_r u}^2
+ \snorm{\partial_r e}^2 + \sigma\snorm{\partial_t u}^2 + \sigma^2 \snorm{\partial_t e}^2 \Big\}
\,r^m {\mathrm d} r{\mathrm d} t \le C(\varepsilon).
\end{align}
\end{lemma}
\begin{proof}
The proof only for $\sqrt{\sigma} \partial_r u_k \overset{\ast}{\rightharpoonup} \sqrt{\sigma} \partial_r u$
is presented, since the other limits can be shown via the same argument.
By Lemma \ref{lemma:HREext},
\begin{equation*}
\sup_{k\in\mathbb{N}} \sup_{t\in[0,T]} \int_{a}^{\infty} (\sigma \snorm{\partial_r u_k}^2)(r,t)\,{\mathrm d} r
\le a^{-m} \sup_{k\in\mathbb{N}} \sup_{t\in[0,T]} \int_{a}^{\infty}
(\sigma\snorm{\partial_r u_k}^2)(r,t)\,r^m {\mathrm d} r \le a^{-m}C(a).
\end{equation*}
Thus, there exists a subsequence (still denoted) $\{ \partial_r u_k \}_{k\in\mathbb{N}}$ such that
\begin{equation}\label{temp:weak1}
(\sqrt{\sigma}\partial_r u_k,\, r^{\frac{m}{2}}\sqrt{\sigma}\partial_r u_k)\,
\overset{\ast}{\rightharpoonup}\, (w_1,\, w_2) \qquad \text{ in $L^{\infty}\big(0,T; L^2( [a,\infty), {\mathrm d} r )\big)$}.
\end{equation}
By \eqref{temp:weak1}, Lemma \ref{lemma:klimholder},
and the dominated convergence theorem,
we see that the spatial weak derivative of $\sqrt{\sigma} u$ exists and $w_1 = \sqrt{\sigma} \partial_r u$.
As a result,
by \eqref{temp:weak1} and the fundamental lemma of calculus,
$w_2=r^{\frac{m}{2}}\sqrt{\sigma}\partial_r u$ for {\it a.e.} $(r,t)\in[a,\infty)\times[0,T]$.
Then \eqref{temp:weak1} implies that
$\sqrt{\sigma} \partial_r u_k \overset{\ast}{\rightharpoonup} \sqrt{\sigma} \partial_r u$
in $L^{\infty}(0,T;L^2([a,\infty),r^m {\mathrm d} r))$.
By the weak-star lower semi-continuity of the Sobolev norm,
\begin{equation*}
\sbnorm{\sqrt{\sigma}\partial_r u}_{L^{\infty}(0,T;L^2([a,\infty),r^m {\mathrm d} r))}
\le \liminf_{k\to\infty} \sbnorm{\sqrt{\sigma}\partial_r u_k}_{L^{\infty}(0,T;L^2([a,\infty),r^m {\mathrm d} r))} \le C(a).
\end{equation*}
The other terms in \eqref{weakstmt1} can be proved by the same argument.
\smallskip
Next, we show \eqref{weakstmt2}. Only the proof for $\sqrt{\sigma}\partial_r u$ is provided, since the other terms
in \eqref{weakstmt2} can be obtained in the same way.
Let $\phi \in L^{1}(0,T;L^2([a,\infty),r^m {\mathrm d} r))$. Then, by Lemma \ref{lemma:HREext},
\begin{align*}
&\Bignorm{\int_{0}^{T}\int_{\widetilde{r}_k(\varepsilon,t)}^{\infty} \sqrt{\sigma} \partial_r u_k \phi\,r^m {\mathrm d} r {\mathrm d} t
- \int_{0}^{T}\int_{\widetilde{r}(\varepsilon,t)}^{\infty} \sqrt{\sigma} \partial_r u \phi\,r^m {\mathrm d} r {\mathrm d} t }\\
&\le \Bignorm{\int_{0}^T \int_{\widetilde{r}_k(\varepsilon,t)}^{\widetilde{r}(\varepsilon,t)} \sqrt{\sigma} \partial_r u_k \phi\,r^m {\mathrm d} r {\mathrm d} t}
+ \Bignorm{\int_{0}^T \int_{\widetilde{r}(\varepsilon,t)}^{\infty} \sqrt{\sigma} (\partial_r u_k-\partial_r u) \phi\,r^m {\mathrm d} r{\mathrm d} t}\\
&\le C(a)\int_{0}^{T}\Big(\int_{\widetilde{r}_k(\varepsilon,t)}^{\widetilde{r}(\varepsilon,t)}\snorm{\phi}^2\,r^m{\mathrm d} r\Big)^{\frac{1}{2}}
{\mathrm d} t + \Bignorm{\int_{0}^T \int_{\widetilde{r}(\varepsilon,t)}^{\infty} \sqrt{\sigma} (\partial_r u_k-\partial_r u) \phi\, r^m {\mathrm d} r{\mathrm d} t}\\
&=: I_k^{(1)} + I_k^{(2)}.
\end{align*}
$I_k^{(2)}\to 0$ as $k\to\infty$ by \eqref{temp:weak1}.
For $I_k^{(1)}$, define
\begin{equation*}
\chi^{(k,\varepsilon)}(r,t) \vcentcolon = \begin{dcases*}
1 & if $r\in[\widetilde{r}_k(\varepsilon,t),\widetilde{r}(\varepsilon,t)]$ or $r\in[\widetilde{r}(\varepsilon,t),\widetilde{r}_k(\varepsilon,t)]$,\\
0 & otherwise.
\end{dcases*}
\end{equation*}
By
Lemma \ref{lemma:rkcvg}, for each fixed $\varepsilon>0$, $\chi^{(k,\varepsilon)}(r,t) \to 0$ as $k\to\infty$
for {\it a.e.} $(r,t)\in[a,\infty)\times[0,T]$.
Moreover, since $\snorm{\chi^{(k,\varepsilon)} \phi} \le \snorm{\phi}$,
by the dominated convergence theorem, as $k\to\infty$,
\begin{equation*}
I_k^{(1)} \vcentcolon
= C(a)\int_{0}^{T} \Big(\int_{\widetilde{r}_k(\varepsilon,t)}^{\widetilde{r}(\varepsilon,t)} \snorm{\phi}^2\,r^m {\mathrm d} r\Big)^{\frac{1}{2}}
{\mathrm d} t
= C(a) \int_{0}^{T} \Big(\int_{a}^{\infty} \chi^{(\varepsilon,k)}\snorm{\phi}^2\,r^m {\mathrm d} r\Big)^{\frac{1}{2}} {\mathrm d} t \to 0.
\end{equation*}
Let $\chi_\varepsilon$ and $\chi_\varepsilon^k$ be defined in \eqref{temp:chive}--\eqref{temp:chivek}.
Then the limits: $I_k^{(1)}$, $I_k^{(2)}\to 0$ imply the weak-star convergence:
$\chi_{\varepsilon}^k\sqrt{\sigma} \partial_r u_k \overset{\ast}{\rightharpoonup} \chi_\varepsilon \sqrt{\sigma} \partial_r u$
as $k\to\infty$ in $L^{\infty}(0,T;L^2( [a,\infty), r^m {\mathrm d} r ))$.
By the weak-star lower semi-continuity of the Sobolev norm,
it follows that
\begin{align*}
\sup_{t\in[0,T]}\int_{\widetilde{r}(\varepsilon,t)}^{\infty} (\sigma \snorm{\partial_r u}^2)(r,t) r^m {\mathrm d} r
&= \sbnorm{\chi_\varepsilon \sqrt{\sigma} \partial_r u}_{L^{\infty}(0,T;L^2([a,\infty),r^m {\mathrm d} r))}\\
&\le \liminf_{k\to\infty}\sbnorm{\chi_\varepsilon^k \sqrt{\sigma} \partial_r u_k}_{L^{\infty}(0,T;L^2([a,\infty),r^m {\mathrm d} r))} \\
&= \sup_{k\in\mathbb{N}}\sup_{t\in[0,T]}\int_{\widetilde{r}_k(\varepsilon,t)}^{\infty} \sigma \snorm{\partial_r u_k}^2(r,t) r^m {\mathrm d} r \le C(\varepsilon).
\end{align*}
\end{proof}
In \S \ref{subsec:klimEnt}--\S \ref{subsec:indepPath} below,
we denote $(\rho,u,e)(r,t)$ in $[a,\infty)\times[0,T]$ as the limit functions obtained in Lemma \ref{lemma:klimholder}--\ref{lemma:rhoWeak} and keep the simplified notation \eqref{asuppress}
as in \S 5.4.
\subsection{Entropy estimate for the limit solution}\label{subsec:klimEnt}
\begin{lemma}\label{lemma:Ent}
There exists $C(T)>0$ such that
\begin{align}\label{temp:Ent0}
&\esssup\limits_{t\in[0,T]}\int_{a}^{\infty}
\Big(\rho\big(\frac{1}{2}\snorm{u}^2+ \psi(e)\big)+ G(\rho) \Big)(r,t)\, r^m {\mathrm d} r
+ \kappa\int_{0}^{T}\int_{a}^{\infty}\dfrac{\snorm{\partial_r e}^2}{ e^2}\,r^m {\mathrm d} r {\mathrm d} t\nonumber \\
&+\int_{0}^{T}\int_{a}^{\infty} \Big\{ \big( \dfrac{2\mu}{n} + \lambda \big)
\dfrac{1}{e}\Bignorm{\partial_r u+m \dfrac{u}{r}}^2 + \dfrac{2m\mu}{n} \dfrac{1}{e}
\Bignorm{\partial_r u-\dfrac{u}{r}}^2\Big\}\,r^m{\mathrm d} r{\mathrm d} t \le C(T).
\end{align}
\end{lemma}
\begin{proof} We divide the proof into four steps.
\smallskip
1. For each integer $L\ge 2$, let $\phi_L\in C^{\infty}([a,\infty))$ be such that $\phi_L(r)=1$ if $r\in[a,L-1]$ and $\phi_L(r)=0$ if $r\in[L,\infty)$.
Then $\phi_L\in \widetilde{H}^1_0( [a,L], r^m {\mathrm d} r )$, where $\widetilde{H}_0^1$ is the Sobolev space defined in \eqref{eqs:tildeH1}. Moreover,
$\partial_r \psi(e_k) = \frac{e_k-1}{e_k} \partial_r e_k$.
Thus, for each fixed $t\in(0,T]$,
$(\psi(e_k)\phi_L)(\cdot,t) \in \widetilde{H}^1_0( [a,L], r^m {\mathrm d} r)$
for all $(k,t)\in\mathbb{N}\times(0,T]$ and
\begin{align*}
\int_{a}^{\infty} \rho_k \psi(e_k)\phi_L\,r^m {\mathrm d} r - \int_{a}^{\infty} \rho \psi(e) \phi_L\,r^m {\mathrm d} r
&\le \int_{a}^{L}(\rho_k-\rho)\psi(e_k)\phi_L\,r^m {\mathrm d} r
+ \int_{a}^{\infty}\rho ( \psi(e_k) - \psi(e) )\phi_L\,r^m {\mathrm d} r \\
&=\vcentcolon I_k^{(1)} + I_k^{(2)}.
\end{align*}
For all $t\in(0,T]$, $I_k^{(1)}$ is estimated by using Lemmas \ref{lemma:HREext} and \ref{lemma:rhoScvg} as
\begin{align*}
I_k^{(1)}
&\le \sbnorm{\rho_k-\rho}_{\widetilde{H}^{-1}([a,L];r^m {\mathrm d} r)} \sbnorm{ \psi(e_k)\phi_L}_{H^{1}([a,L];r^m {\mathrm d} r)}\\
&= \sbnorm{\rho_k-\rho}_{\widetilde{H}^{-1}} \Big\{ \sbnorm{\phi_L}_{H^1}\sbnorm{\psi(e_k)}_{L^\infty}
+ \sigma^{-1}(t)\Bigbnorm{\dfrac{e_k-1}{e_k}\phi_L}_{\infty} \sbnorm{\sigma\partial_r e_k}_{L^2} \Big\}\\
&\le \sbnorm{\rho_k-\rho}_{\widetilde{H}^{-1}} \big\{ C(a)\sbnorm{\phi_L}_{H^1} + C(a) \sigma^{-1}(t)\big\} \to 0
\qquad\ \text{as $k\to\infty$}.
\end{align*}
From Lemma \ref{lemma:klimholder},
$$
e_k(r,t)\to e(r,t) \qquad \mbox{as $k\to\infty$ for $(r,t)\in[a,\infty)\times[0,T]$},
$$
and, from Lemmas \ref{lemma:klimholder}--\ref{lemma:rhoWeak},
$$
C^{-1}(a)\le \rho(r,t) \le C(a), \,\, C^{-1}(a) \le e_k(r,t),\, e(r,t) \le C(a) \sigma^{-\frac{1}{2}}(t)
$$
for {\it a.e.} $(r,t)\in[a,\infty)\times(0,T]$.
Using these and the dominated convergence theorem, it follows that
\begin{align*}
\lim\limits_{k\to\infty} I_k^{(2)}
\vcentcolon= \lim\limits_{k\to\infty} \int_{a}^{\infty} \rho \phi_n ( \psi(e_k) - \psi(e) )(r,t)\,r^m {\mathrm d} r
= 0 \qquad \text{ for each $t\in(0,T]$}.
\end{align*}
Using Lemma \ref{lemma:entEstext}, the limits: $I_k^{(1)}, I_k^{(2)}\to 0$ then imply that, for each $t>0$,
\begin{align*}
\int_{a}^{\infty} \phi_L(r,t)(\rho \psi(e))(r,t)\,r^m {\mathrm d} r
= \lim\limits_{k\to\infty}\int_{a}^{\infty}\phi_L(r,t)(\rho_k \psi(e_k))(r,t)\,r^m {\mathrm d} r \le C(T).
\end{align*}
Now, by construction, $\phi_L(r)\to 1$ for all $r\in[a,\infty)$ as $L\to \infty$.
Since $\rho \psi(e) >0$, we apply the monotone convergence theorem to obtain
\begin{align*}
\int_{a}^{\infty} (\rho \psi(e))(r,t)\,r^m {\mathrm d} r
=\lim\limits_{L\to\infty} \int_{a}^{\infty} \phi_L(r,t) (\rho \psi(e))(r,t)\,r^m {\mathrm d} r \le C(T)
\qquad \text{ for each $t\in(0,T]$}.
\end{align*}
\smallskip
2. From Lemma \ref{lemma:HREext}, $u_k^2 \in L^2( [a,\infty), r^m {\mathrm d} r )$ and, for all $(k,t)\in\mathbb{N}\times(0,T]$,
\begin{equation*}
\sbnorm{\partial_r u_k^2(\cdot,t)}_{L^2([a,\infty), r^m {\mathrm d} r)} = \sbnorm{u_k}_{L^\infty} \sbnorm{\partial_r u_k(\cdot,t)}_{L^2( [a,\infty), r^m {\mathrm d} r)} \le \sigma^{-\frac{3}{4}}(t)C(a).
\end{equation*}
From these, $ u_k\phi_L\in \widetilde{H}_0^1([a,L],r^m {\mathrm d} r)$ and
\begin{align*}
&\Bignorm{\int_{a}^{\infty} \rho_k u_k^2\phi_L\, r^m {\mathrm d} r - \int_{a}^{\infty} \phi_L \rho u^2\, r^m {\mathrm d} r}\\
&\le \bnorm{\rho_k-\rho}_{\widetilde{H}^{-1}([a,L], r^m{\mathrm d} r)} \sbnorm{u_k^2\phi_L}_{H^1([a,L],r^m {\mathrm d} r)} + \int_{a}^{\infty} \rho ( u_k^2 - u^2 ) \phi_L\, r^m {\mathrm d} r\\
&\le \int_{a}^{\infty} \phi_L \rho ( u_k^2 - u^2)\, r^m {\mathrm d} r
+ C(a)\big(1+\sigma^{-\frac{3}{4}}(t)\big)
\sbnorm{\rho_k-\rho}_{\widetilde{H}^{-1}([a,L],r^m{\mathrm d} r)}.
\end{align*}
Since $u_k(r,t)\to u(r,t)$ from Lemma \ref{lemma:klimholder},
$\snorm{u_k(r,t)}+\snorm{u(r,t)}\le C(a) \sigma^{-\frac{1}{4}}(t)$
from Lemmas \ref{lemma:HREext} and \ref{lemma:klimholder},
and $\rho_k\to \rho$ strongly in $\widetilde{H}^{-1}$ from Lemma \ref{lemma:rhoScvg},
then, by the dominated convergence theorem,
\begin{align*}
\int_{a}^{\infty} \rho u^2\phi_L\,r^m {\mathrm d} r
= \lim\limits_{k\to\infty} \int_{a}^{\infty} \rho_k u_k^2\phi_L\,r^m {\mathrm d} r
\le \limsup\limits_{k\to\infty} \int_{a}^{\infty} \rho_k u_k^2\,r^m {\mathrm d} r \le C(T).
\end{align*}
Then $\sbnorm{(\rho u^2)(\cdot,t)}_{L^1([a,\infty),r^m{\mathrm d} r)}\le C(T)$ follows
from the monotone convergence theorem for $L\to\infty$.
\smallskip
3. Let $\phi\in L^2(0,T;L^2([a,\infty),r^m {\mathrm d} r))$. By Lemma \ref{lemma:HREext},
\begin{align*}
&\Bignorm{\int_{0}^{T}\int_{a}^{\infty} \dfrac{\partial_r e_k}{e_k} \phi\,r^m {\mathrm d} r {\mathrm d} t
- \int_{0}^{T}\int_{a}^{\infty} \dfrac{\partial_r e}{e} \phi\,r^m {\mathrm d} r {\mathrm d} t}\\
&\le C(a) \Big( \int_{0}^{T}\int_{a}^{\infty} \big(e_k^{-1}-e^{-1}\big)^2 \snorm{\phi}^2\,r^m {\mathrm d} r {\mathrm d} t
\Big)^{\frac{1}{2}}
+ \Bignorm{\int_{0}^{T}\int_{a}^{\infty} (\partial_r e_k-\partial_r e)\dfrac{\phi}{e}\,r^m{\mathrm d} r{\mathrm d} t}\\
& =\vcentcolon \textrm{II}_k^{(1)} + \textrm{II}_k^{(2)}.
\end{align*}
Using $e_k\to e$
from Lemma \ref{lemma:klimholder},
and the estimates: $C(a)\le e_k(r,t), e(r,t) \le C(a)\sigma^{-\frac{1}{2}}(t)$ from Lemmas \ref{lemma:HREext} and \ref{lemma:klimholder},
we use the dominated convergence theorem to obtain
\begin{align*}
\lim\limits_{k\to\infty} \textrm{II}_k^{(1)}
\vcentcolon= \lim\limits_{k\to\infty} C(a) \Big( \int_{0}^{T}\int_{a}^{\infty} \big(e_k^{-1}-e^{-1}\big)^2 \snorm{\phi}^2\,r^m {\mathrm d} r {\mathrm d} t \Big)^{\frac{1}{2}} =0.
\end{align*}
For $\textrm{II}_k^{(2)}$, since $e(r,t)\ge C^{-1}(a)$
and $\partial_r e_k \rightharpoonup \partial_r e$ in Lemma \ref{lemma:weak}, we have
\begin{align*}
\lim\limits_{k\to\infty} \textrm{II}_k^{(2)} \vcentcolon
= \lim\limits_{k\to\infty}\Bignorm{\int_{0}^{T}\int_{a}^{\infty} (\partial_r e_k - \partial_r e)
\dfrac{\phi}{e} (r,t)\,r^m {\mathrm d} r {\mathrm d} t} = 0.
\end{align*}
The above limits imply that $\frac{\partial_r e_k}{e_k} \rightharpoonup \frac{\partial_r e}{e}$ as $k\to\infty$
weakly in $L^2(0,T;L^2([a,\infty),r^m {\mathrm d} r))$.
By Lemma \ref{lemma:entEstext} and the weak-star lower semi-continuity of the Sobolev norm,
it follows that
\begin{align*}
\int_{0}^{T}\int_{a}^{\infty} \dfrac{\snorm{\partial_r e}^2}{e^2}\,r^m {\mathrm d} r {\mathrm d} t
\le \liminf_{k\to\infty} \int_{0}^{T}\int_{a}^{\infty} \dfrac{\snorm{\partial_r e_k}^2}{e_k^2}\,r^m {\mathrm d} r {\mathrm d} t
\le C(T).
\end{align*}
By the same argument,
we can also obtain the estimates
for $\sbnorm{(\partial_r u + m \frac{u}{r})e^{-1}}_{L^2}$ and $\sbnorm{(\partial_r u - \frac{u}{r})e^{-1}}_{L^2}$
in \eqref{temp:Ent0}.
\smallskip
4. From Lemmas \ref{lemma:entEstext}--\ref{lemma:HREext} and \ref{lemma:rhoWeak}, we have
\begin{equation}
\begin{dcases}
\rho_k \rightharpoonup \rho \ \text{ weakly in } \ L^2\big(0,T;L^2([a,\infty),r^m {\mathrm d} r)\big)
\qquad \text{ as $k\to\infty$}.\\
\sup\limits_{k\in\mathbb{N}}\esssup_{t\in[0,T]}\int_{a}^{\infty} G(\rho_k)(r,t)\,r^m {\mathrm d} r \le C(T), \\
C^{-1}(a)\le \rho_k(r,t)\le C(a) \qquad \text{for {\it a.e.} \ $(r,t)\in[a,\infty)\times[0,T]$ and for each
$k\in\mathbb{N}$}.
\end{dcases}
\end{equation}
By Proposition \ref{prop:mazur},
$\sbnorm{G(\rho)(\cdot,t)}_{L^1([a,\infty),r^m {\mathrm d} r)}\le C(T)$ for {\it a.e.} $t\in[0,T]$.
\end{proof}
\subsection{Weak forms of the exterior problem}\label{subsec:klimWF}
We now show that $(\rho,u,e)$ is indeed a weak solution of problem \eqref{eqs:SFNS} and \eqref{eqs:eSFNS}.
\begin{lemma}\label{lemma:klimWF}
Let $(\rho,u,e)(r,t)$ be the limit function obtained
in {\rm Lemmas \ref{lemma:klimholder}}--{\rm \ref{lemma:rhoWeak}}.
Then $(\rho,u,e)(r,t)$ is a weak solution of problem \eqref{eqs:SFNS} and \eqref{eqs:eSFNS} as stated in {\rm Definition \ref{def:SWeak}}.
\end{lemma}
\begin{proof} We divide the proof into three steps.
\smallskip
1. We start with the {\it continuity equation}.
For any $\phi\in \mathcal{D}^a$, then there exists an integer $N\in\mathbb{N}$
such that $\phi(r,t) = 0$ for $r\ge N$ and $t\in[0,T]$.
Take $k \in \mathbb{N}$ to be $k\ge M(\phi,a) \vcentcolon= 2^n N^n\max\{C_0,C(a)\}$.
Then, for all $t\in[0,T]$,
\begin{align*}
&\widetilde{r}_k(k,t)= \Big( a^n + \int_{0}^{k} \widetilde{v}_k(x,t) {\mathrm d} x \Big)^{\frac{1}{n}}
\ge \Big( a^n + \frac{k}{C(a)}\Big)^{\frac{1}{n}} \ge ( a^n + 2^n N^n )^{\frac{1}{n}} > 2N,\\
& \widetilde{r}_a^0(k) = \Big( a^n + \int_{0}^{k}\dfrac{1}{\rho_a^0(\widetilde{r}_a^0(x))} {\mathrm d} x \Big)^{\frac{1}{n}}
\ge \Big( a^n + \frac{k}{C_0}\Big)^{\frac{1}{n}} \ge ( a^n + 2^n N^n )^{\frac{1}{n}} > 2N.
\end{align*}
Thus, it follows by construction that, if $(r,t)\in {\mathrm {supp}}(\phi)$ and $k\ge M(\phi,a)$,
\begin{equation}\label{wtemp1}
\begin{split}
\rho_k(r,t) =&\, \overline{v}_k^{-1}\varphi_k(r,t) + 1 - \varphi_k(r,t) = \overline{v}_k^{-1}(r,t), \ \ u_k(r,t) = \overline{u}_k\varphi_k(r,t) = \overline{u}_k(r,t),\\
e_k(r,t) =&\, \overline{e}_k\varphi_k(r,t) + 1 - \varphi_k(r,t) = \overline{e}_k(r,t).
\end{split}
\end{equation}
Moreover, using that $\widetilde{r}_a^0(k)> 2N$
for $k \ge M(\phi,a) \vcentcolon= 2^n N^n \max\{C_0,C(a)\}$,
we see from construction \eqref{def:appInitr1} in \S\ref{subsec:ffaprox} that,
if $r\in {\mathrm {supp}}(\phi(\cdot,0))$, then
\begin{equation}\label{wtemp6}
\begin{split}
\rho_{a,k}^0(r)
=&\, \rho_a^0(r) \varphi_{a,k}^0(r) + 1 - \varphi_{a,k}^0(r)
= \rho_a^0(r),\ \ u_{a,k}^0(r) = u_a^0(r)\varphi_{a,k}^0(r) = u_a^0(r),\\
e_{a,k}^0(r) =&\, e_a^0(r)\varphi_{a,k}^0(r) + 1 - \varphi_{a,k}^0(r) = e_a^0(r).
\end{split}
\end{equation}
With this, if $x\in[0,\infty)$ is such that $\widetilde{r}_a^0(x)\in {\mathrm {supp}}( \phi(\cdot,0) )$, then
it follows from Proposition \ref{prop:r0diff} that,
for all $k\ge M(\phi,a)$,
\begin{equation}\label{wtemp7}
\widetilde{r}_a^0(x) = \Big( a^n + n\int_{0}^{x} \dfrac{1}{\rho_a^0(\widetilde{r}_a^0(y))} {\mathrm d} y \Big)^{\frac{1}{n}}
= \Big( a^n + n\int_{0}^{x} \widetilde{v}_{a,k}^0(y) {\mathrm d} y \Big)^{\frac{1}{n}} = \widetilde{r}_{k}(x,0).
\end{equation}
Thus, from the coordinate transformation \eqref{eqs:coorderiv} and the continuity equation for which the approximate Lagrangian functions satisfy, we obtain that, for all $k\ge M(\phi,a)$,
\begin{align*}
&\int_{a}^{\infty} \Big\{ \rho_k(r,t) \phi(r,t) - \rho_a^0(r)\phi(r,0) \Big\}\, r^m {\mathrm d} r\\
&= \int_{a}^{\widetilde{r}_k(k,t)} \overline{v}_k^{-1} \phi(r,t)\, r^m {\mathrm d} r
- \int_{a}^{\widetilde{r}_a^0(k)} \phi(r,0) \rho_a^0(r)\,r^m {\mathrm d} r \\
&= \int_{0}^{k} \phi(\widetilde{r}_k(x,t),t)\,{\mathrm d} x - \int_{0}^{k} \phi(\widetilde{r}_k(x,0),0)\,{\mathrm d} x \\
&= \int_{0}^{t} \int_{0}^{k} \big( \widetilde{u}_k \partial_r \phi+ \partial_t \phi \big)(\widetilde{r}_k(x,s),s)\,{\mathrm d} x {\mathrm d} s\\
&= \int_{0}^{t}\int_{a}^{\widetilde{r}_k(k,t)} \big(\overline{v}_k^{-1}\overline{u}_k \partial_r \phi
+ \widetilde{v}_k^{-1} \partial_t \phi \big)(r,t)\,r^m {\mathrm d} r {\mathrm d} s\\
&= \int_{0}^{t}\int_{a}^{\infty} \big(\rho_k u_k \partial_r \phi + \rho_k \partial_t \phi \big)(r,t)\,r^m {\mathrm d} r {\mathrm d} s,
\end{align*}
where we have used $\widetilde{r}_k(k,t)\ge 2N$ and \eqref{wtemp1} in the last equality.
Also, by Lemma \ref{lemma:klimholder}--\ref{lemma:rhoScvg}, we have
\begin{equation}\label{temp:klimWF1}
\begin{aligned}
&\lim\limits_{k\to\infty} \sup_{t\in[0,T]}\sbnorm{\rho_k(\cdot, t)
- \rho(\cdot,t)}_{\widetilde{H}^{-1}([a,L], r^m {\mathrm d} r)} = 0 && \text{ for all $L\in\mathbb{N}$,}\\
&\rho_k-1 \overset{\ast}{\rightharpoonup} \rho-1
&& \text{ in $L^{\infty}\big(0,T;L^2([a,\infty), r^m {\mathrm d} r)\big)$,}\\
&u_k(r,t) \to u(r,t) \ \text{ and } \ \rho_k(r,t),\,\rho(r,t) \le C(a) && \text{ for {\it a.e.} $(r,t) \in[a,\infty)\times[0,T]$.}
\end{aligned}
\end{equation}
Using these, we can apply the dominated convergence theorem to obtain
\begin{align*}
\int_{a}^{\infty} \rho(r,t) \phi(r,t)\,r^m {\mathrm d} r
- \int_{a}^{\infty} \rho_a^0(r) \phi(r,0)\,r^m {\mathrm d} r
= \int_{0}^{t}\int_{a}^{\infty} \big(\rho u \partial_r \phi + \rho \partial_t \phi\big)\,r^m {\mathrm d} r {\mathrm d} s.
\end{align*}
\smallskip
2. We show the weak form for the {\it momentum equation}.
Let $\varphi\in \mathcal{D}_0^a$.
Then there exists $N\in\mathbb{N}$ such that $\varphi(r,t) = 0$ for $r\ge N$ and $t\in[0,T]$.
If $k \in \mathbb{N}$ is such that $k \ge M(\varphi,a)$, then, as previously shown,
$\widetilde{r}_k(k,t) > 2N$ for all $t\in[0,T]$, and $\widetilde{r}_a^0(k)>2N$.
By construction, if $(r,t)\in {\mathrm {supp}} (\varphi)$ and $r\in {\mathrm {supp}} (\varphi(\cdot,0))$,
then \eqref{wtemp1}--\eqref{wtemp6} hold so that
\begin{align*}
&\int_{a}^{\infty} (\rho_k u_k)(r,t)\varphi(r,t)\,r^m {\mathrm d} r
- \int_{a}^{\infty} (\rho_a^0 u_a^0)(r) \varphi(r,0)\,r^m {\mathrm d} r\\
&=\int_{a}^{\widetilde{r}_k(k,t)} (\overline{v}_k^{-1} \overline{u}_k) (r,t)\varphi(r,t)\,r^m {\mathrm d} r
- \int_{a}^{\widetilde{r}_a^0(k)} u_{a,k}^0 (r) \varphi(r,0) \rho_{a}^0(r)\,r^m {\mathrm d} r\\
&= \int_{0}^{k} \widetilde{u}_k (x,t)\varphi(\widetilde{r}_k(x,t),t)\,{\mathrm d} x
- \int_{0}^{k} \widetilde{u}_{a,k}^0 (x) \varphi(\widetilde{r}_{k}(x,0),0)\,{\mathrm d} x\\
&= \int_{0}^{t}\int_{0}^{k} \big(\varphi(\widetilde{r}_k(x,s),s) {\mathrm D}_t \widetilde{u}_k
+ \partial_t \varphi (\widetilde{r}_k(x,s),s) \widetilde{u}_k
+ \partial_r \varphi(\widetilde{r}_k(x,s),s) \widetilde{u}_k^2 \big)\,{\mathrm d} x {\mathrm d} s.
\end{align*}
By $\eqref{eqs:LFNS-k}_2$
and the boundary condition: $\varphi(a,t)=0$, $\widetilde{u}_k(k,t)=0$ for $t\in[0,T]$,
it follows that
\begin{align*}
&\int_{a}^{\infty} (\rho_k u_k)(r,t)\varphi(r,t)\,r^m {\mathrm d} r
- \int_{a}^{\infty} (\rho_a^0 u_a^0)(r) \varphi(r,0)\,r^m {\mathrm d} r \\
&= \int_{0}^{t}\int_{0}^{k} \big(\varphi(\widetilde{r}_k(x,s),s) {\mathrm D}_t \widetilde{u}_k
+ \partial_t \varphi(\widetilde{r}_k(x,s),s) \widetilde{u}_k
+ \partial_r \varphi(\widetilde{r}_k(x,s),s) \widetilde{u}_k^2 \big)\,{\mathrm d} x {\mathrm d} s\\
&= \int_{0}^{t}\int_{0}^{k} \big(\widetilde{u}_k(x,s)\,\partial_t \varphi(\widetilde{r}_k(x,s),s)
+ \widetilde{u}_k^2(x,s)\,\partial_r \varphi(\widetilde{r}_k(x,s),s) \big)\,{\mathrm d} x {\mathrm d} s\\
&\quad + \int_{0}^{t}\int_{0}^{k} \widetilde{r}_k^m(x,s)\varphi(\widetilde{r}_k(x,s),s) {\mathrm D}_x\Big( \beta \dfrac{{\mathrm D}_x(\widetilde{r}_k^m \widetilde{u}_k)}{\widetilde{v}_k} - (\gamma-1) \dfrac{\widetilde{e}_k}{\widetilde{v}_k} \Big)\,{\mathrm d} x {\mathrm d} s\\
&= \int_{0}^{t}\int_{0}^{k}\big( \widetilde{u}_k^2(x,s)\,\partial_r \varphi(\widetilde{r}_k(x,s),s)
+ \widetilde{u}_k(x,s)\,\partial_t \varphi(\widetilde{r}_k(x,s),s)\big)\,{\mathrm d} x {\mathrm d} s \\
&\quad+ \int_{0}^{t} \int_{0}^{k} \Big( (\gamma-1)\widetilde{e}_k - \beta \widetilde{r}_k^m {\mathrm D}_x\widetilde{u}_k
- m\beta \dfrac{\widetilde{v}_k\widetilde{u}_k}{\widetilde{r}_k} \Big)
\Big( m \dfrac{\varphi(\widetilde{r}_k(x,s),s)}{\widetilde{r}_k(x,s)}
+ \partial_r \varphi(\widetilde{r}_k(x,s),s) \Big)\,{\mathrm d} x {\mathrm d} s\\
&= \int_{0}^{t}\int_{a}^{\widetilde{r}_{k}(k,t)} \big(\overline{v}_k^{-1} \overline{u}_k^2 \partial_r\varphi + \overline{v}_k^{-1}\overline{u}_k \partial_t \varphi \big)\,r^m {\mathrm d} r {\mathrm d} s \\
&\quad + \int_{0}^{t}\int_{a}^{\widetilde{r}_k(k,t)} \Big((\gamma-1)\overline{v}_k^{-1}\overline{e}_k
-\beta \partial_r \overline{u}_k - m\beta \dfrac{\overline{u}_k}{r} \Big)
\Big( \partial_r \varphi + m \dfrac{\varphi}{r} \Big)\,r^m {\mathrm d} r {\mathrm d} s\\
&= \int_{0}^{t}\int_{a}^{\infty} \big(\rho_k u_k^2 \partial_r\varphi
+ \rho_k u_k \partial_t \varphi \big)\, r^m {\mathrm d} r {\mathrm d} s \\
&\quad + \int_{0}^{t}\int_{a}^{\infty} \Big((\gamma-1)\rho_k e_k
-\beta \partial_r u_k - m\beta \dfrac{u_k}{r} \Big)
\Big( \partial_r \varphi + m \dfrac{\varphi}{r} \Big)\,r^m {\mathrm d} r {\mathrm d} s.
\end{align*}
In addition, from Lemmas \ref{lemma:HREext} and \ref{lemma:klimholder}--\ref{lemma:weak}, we have
\begin{equation}\label{wtemp4}
\begin{aligned}
&\partial_r u_k \rightharpoonup \partial_r u && \text{ in $L^{2}\big(0,T;L^2([a,\infty), r^m {\mathrm d} r)\big)$,}\\
&e_k(r,t) \to e(r,t) && \text{ for $(r,t) \in[a,\infty)\times[0,T]$,}\\
&\sigma^{-\frac{1}{4}}(t) \snorm{u(r,t)} + \sigma^{-\frac{1}{2}}(t) e(r,t) \le C(a)
&& \text{ for $(r,t)\in[a,\infty)\times[0,T]$,} \\
&\sigma^{-\frac{1}{4}}(t) \snorm{u_k(r,t)} + \sigma^{-\frac{1}{2}}(t) e_k(r,t) \le C(a)
&& \text{ for $(k,r,t)\in \mathbb{N}\times[a,\infty)\times[0,T]$.}
\end{aligned}
\end{equation}
Using \eqref{temp:klimWF1} and \eqref{wtemp4},
we conclude Definition \ref{def:SWeak}\ref{item:SWeak2}.
\smallskip
3. We now show the {\it energy equation}.
Let $\phi\in \mathcal{D}^a$. Then, for $k\ge M(\phi,a)$,
\begin{align}\label{wtemp2}
&\int_{a}^{\infty} (\rho_k e_k)(r,t) \phi (r,t)\,r^m {\mathrm d} r
- \int_{a}^{\infty} (\rho_a^0 e_a^0) (r) \phi(r,0)\,r^m {\mathrm d} r \nonumber\\
&= \int_{0}^{k} \widetilde{e}_k(x,t) \phi(\widetilde{r}_k(x,t),t)\,{\mathrm d} x
- \int_{0}^{k} \widetilde{e}_{a,k}^0(x) \phi(\widetilde{r}_{k}(x,0),0)\,{\mathrm d} x \nonumber\\
&= \int_{0}^{t}\int_{0}^{k} \Big\{ \phi(\widetilde{r}_k(x,s),s) {\mathrm D}_t \widetilde{e}_k(x,s)
+ \widetilde{e}_k(x,s)\big( \partial_t \phi + \widetilde{u}_k(x,s) \partial_r \phi\big)
(\widetilde{r}_k(x,s),s)\Big\}\,{\mathrm d} x {\mathrm d} s \nonumber\\
&= \int_{0}^{t} \int_{0}^{k} \phi(\widetilde{r}_k(x,s),s) {\mathrm D}_t \widetilde{e}_k\, {\mathrm d} x {\mathrm d} s
+ \int_{0}^{t} \int_{a}^{\infty} \rho_k e_k ( \partial_t \phi + u_k \partial_r \phi )\,r^m {\mathrm d} r {\mathrm d} s\nonumber \\
& =\vcentcolon I_1+I_2.
\end{align}
Denote $\widetilde{F}_k\vcentcolon= \beta \widetilde{v}_k^{-1}{\mathrm D}_x(\widetilde{r}_k^m \widetilde{u}_k)
- (\gamma-1)\widetilde{v}_k^{-1}\widetilde{e}_k$.
It follows from the momentum equation that ${\mathrm D}_t \widetilde{u}_k = \widetilde{r}_k^{m} {\mathrm D}_x \widetilde{F}_k$.
Then substituting this into $\eqref{eqs:LFNS-k}_3$ yields
\begin{align*}
{\mathrm D}_t \widetilde{e}_k
= {\mathrm D}_x\big( \widetilde{r}_k^m \widetilde{u}_k \widetilde{F}_k \big) - \dfrac{1}{2}{\mathrm D}_t \snorm{\widetilde{u}_k}^2 - 2 m \mu {\mathrm D}_x \big(\widetilde{r}_k^{m-1} \widetilde{u}_k^2\big) + \kappa {\mathrm D}_x\Big( \dfrac{\widetilde{r}_k^{2m}}{\widetilde{v}_k} {\mathrm D}_x \widetilde{e}_k \Big)
\end{align*}
for $I_1$ which, along with ${\mathrm D}_x \widetilde{e}_k(0,t)={\mathrm D}_x \widetilde{e}_k (k,t) = 0$ for all $t\in[0,T]$, leads to
\begin{align*}
I_1
=&\, \int_{0}^{t} \int_{0}^{k} \phi(\widetilde{r}_k(x,s),s) \Big\{ {\mathrm D}_x( \widetilde{r}_k^m \widetilde{u}_k \widetilde{F}_k )
- \dfrac{1}{2}{\mathrm D}_t \snorm{\widetilde{u}_k}^2 \Big\}\,{\mathrm d} x {\mathrm d} s \\
&\,+ \int_{0}^{t} \int_{0}^{k} \phi(\widetilde{r}_k(x,s),s) \Big\{ - 2 m \mu {\mathrm D}_x (\widetilde{r}_k^{m-1} \widetilde{u}_k^2)
+ \kappa {\mathrm D}_x\Big( \dfrac{\widetilde{r}_k^{2m}}{\widetilde{v}_k} {\mathrm D}_x \widetilde{e}_k \Big) \Big\}\,{\mathrm d} x {\mathrm d} s\\
=&\, \int_{0}^{t} \int_{0}^{k} \big\{ -m \beta \widetilde{r}_k^{-1} \widetilde{v}_k \snorm{\widetilde{u}_k}^2
- \beta \widetilde{r}_k^m \widetilde{u}_k {\mathrm D}_x\widetilde{u}_k + (\gamma-1) \widetilde{u}_k \widetilde{e}_k \big\} \partial_r\phi(\widetilde{r}_k(x,s),s)\,{\mathrm d} x {\mathrm d} s \\
&\, + \frac{1}{2}\int_{0}^{t}\int_{0}^{k} \snorm{\widetilde{u}_k}^2\big\{ \partial_t \phi(\widetilde{r}_k(x,s),s)
+ \widetilde{u}_k(x,s) \partial_r \phi(\widetilde{r}_k(x,s),s) \big\}\,{\mathrm d} x {\mathrm d} s \\
&\,- \frac{1}{2}\int_{0}^{k} \snorm{\widetilde{u}_k}^2(x,t) \phi(\widetilde{r}_k(x,t),t)\,{\mathrm d} x
+ \frac{1}{2}\int_{0}^{k} \snorm{\widetilde{u}_{a,k}^0(x)}^2 \phi(\widetilde{r}_k(x,0),0)\,{\mathrm d} x \\
&\,+ \int_{0}^{t} \int_{0}^{k} \Big\{ 2 m \mu \widetilde{r}_k^{-1} \widetilde{v}_k \snorm{\widetilde{u}_k}^2\,\partial_r \phi(\widetilde{r}_k(x,s),s)
-\kappa\widetilde{r}_k^{m} {\mathrm D}_x\widetilde{e}_k\,\partial_r \phi(\widetilde{r}_k(x,s),s) \Big\}
\,{\mathrm d} x {\mathrm d} s.
\end{align*}
From \eqref{wtemp7}, $\widetilde{r}_k(x,0)=\widetilde{r}_a^0(x)$
if $\widetilde{r}_a^0(x)\in {\mathrm {supp}}(\phi(\cdot,0))$ and $k\ge M(\phi,a)$.
By the coordinate transformation \eqref{eqs:coordtrans}--\eqref{eqs:coorderiv},
and \eqref{wtemp1}--\eqref{wtemp6}, it follows that
\begin{align*}
I_1
&= \int_{0}^{t} \int_{a}^{\widetilde{r}_k(k,t)} \Big\{ m(2\mu-\beta) \dfrac{\snorm{\overline{u}_k}^2}{r}
- \beta \overline{u}_k \partial_r\overline{u}_k + (\gamma-1) \overline{u}_k \overline{v}_k^{-1} \overline{e}_k
- \kappa \partial_r \overline{e}_k \Big\} \partial_r\phi\,r^m {\mathrm d} r {\mathrm d} s \\
&\quad + \frac{1}{2}\int_{0}^{t}\int_{a}^{\widetilde{r}_k(k,t)}
\overline{v}_k^{-1} \snorm{\overline{u}_k}^2
\big\{\partial_t \phi + \overline{u}_k \partial_r\phi\big\}\,r^m {\mathrm d} r {\mathrm d} s \\
&\quad - \frac{1}{2}\int_{a}^{\widetilde{r}_k(k,t)}
(\overline{v}_k^{-1}\snorm{\overline{u}_k}^2)(r,t)\phi(r,t)\,r^m {\mathrm d} r
+ \frac{1}{2}\int_{a}^{\widetilde{r}_a^0(k)} (\rho_a^0 \snorm{u_{a}^0}^2)(r) \phi(r,0)\,r^m {\mathrm d} r \\
&= \int_{0}^{t} \int_{a}^{\infty} \Big\{ -m \lambda \dfrac{\snorm{u_k}^2}{r}
- \beta u_k \partial_r u_k + (\gamma-1) \rho_k u_k e_k - \kappa \partial_r e_k \Big\}
\partial_r\phi\,r^m {\mathrm d} r {\mathrm d} s \\
&\quad + \frac{1}{2}\int_{0}^{t}\int_{a}^{\infty} \rho_k \snorm{u_k}^2
\big\{\partial_t \phi + u_k \partial_r \phi\big\}\,r^m {\mathrm d} r {\mathrm d} s \\
&\quad - \frac{1}{2}\int_{a}^{\infty} (\rho_k \snorm{u_k}^2)(r,t) \phi(r,t)\, r^m {\mathrm d} r
+ \frac{1}{2}\int_{a}^{\infty} (\rho_a^0 \snorm{u_a^0}^2)(r) \phi(r,0)\, r^m {\mathrm d} r.
\end{align*}
Define $E_k(r,t) \vcentcolon= (\frac{1}{2}\snorm{u_k}^2 + e_k)(r,t)$. Then the above identity becomes:
\begin{align*}
I_1
=&\,\int_{0}^{t} \int_{a}^{\infty} \Big\{ -m \lambda \dfrac{\snorm{u_k}^2}{r}
- \beta u_k \partial_r u_k + (\gamma-1) \rho_k u_k e_k - \kappa \partial_r e_k \Big\}
\partial_r\phi\,r^m {\mathrm d} r {\mathrm d} s \\
&\, + \int_{0}^{t}\int_{a}^{\infty} \big\{ \rho_k E_k ( \partial_t \phi + u_k \partial_r \phi)
- \rho_k e_k (\partial_t \phi + u_k \partial_r \phi) \big\} \,r^m {\mathrm d} r {\mathrm d} s \\
&\,- \frac{1}{2}\int_{a}^{\infty} (\rho_k \snorm{u_k}^2)(r,t) \phi(r,t)\,r^m {\mathrm d} r
+ \frac{1}{2}\int_{a}^{\infty} (\rho_a^0\snorm{u_a^0}^2)(r) \phi(r,0)\,r^m {\mathrm d} r.
\end{align*}
Denote $P_k\vcentcolon= (\gamma-1)\rho_k e_k$ and $E_a^0\vcentcolon=\frac{1}{2}\snorm{u_a^0}^2+e_a^0$.
Substituting $I_1$ into \eqref{wtemp2}, we have
\begin{equation}\label{wtemp3}
\begin{aligned}
&\int_{a}^{\infty} (\rho_k E_k)(r,t) \phi (r,t)\,r^m {\mathrm d} r
- \int_{a}^{\infty} (\rho_a^0 E_a^0) (r) \phi(r,0) r^m {\mathrm d} r \\
&= \int_{0}^{t} \int_{a}^{\infty} \Big\{ \Big( (\rho_kE_k + P_k) u_k -m \lambda \dfrac{\snorm{u_k}^2}{r}
- \beta u_k \partial_r u_k - \kappa \partial_r e_k \Big)\partial_r\phi
+ \rho_k E_k \partial_t \phi \Big\}\,r^m {\mathrm d} r {\mathrm d} s.
\end{aligned}
\end{equation}
In addition to \eqref{wtemp4}, we also uses the weak convergence from Lemma \ref{lemma:weak}
that $\partial_r e_k \rightharpoonup \partial_r e$ in $L^2(0,T;L^2([a,\infty),r^m {\mathrm d} r))$.
Then \eqref{wtemp3} becomes:
\begin{align*}
&\int_{a}^{\infty} (\rho E)(r,t) \phi (r,t)\,r^m {\mathrm d} r
- \int_{a}^{\infty} (\rho_a^0 E_a^0)(r) \phi(r,0)\,r^m {\mathrm d} r \\
&= \int_{0}^{t} \int_{a}^{\infty} \Big\{ \Big( (\rho E + P ) u -m \lambda \dfrac{\snorm{u}^2}{r}
- \beta u \partial_r u - \kappa \partial_r e \Big) \partial_r\phi
+ \rho E \partial_t \phi \Big\}\,r^m {\mathrm d} r {\mathrm d} s,
\end{align*}
where $P\vcentcolon= (\gamma-1) \rho e$ and $E\vcentcolon= \frac{1}{2}\norm{u}^2 + e$.
\end{proof}
\subsection{Estimates away from the origin, independent of the particle path function}\label{subsec:indepPath}
The main aim of this subsection is to obtain
the uniform estimates, independent of $a\in(0,1)$, which do not involve the particle path
function $\widetilde{r}(x,t)$.
\begin{lemma}\label{lemma:dissip}
For each $\eta\in (a,1)$,
\begin{equation}\label{temp:dissip0}
\begin{dcases*}
\int_{0}^{T} \sup_{r\ge\eta}\,\dfrac{\snorm{u}}{\sqrt{e}}(r,t)\,{\mathrm d} t
\le C(T) \big(\eta^{\frac{2-n}{2}} + \eta^{2-n}\big) \ \ & if $n=2$, $3$,\\
\int_{0}^{T} \sup_{r\ge \eta} \log( \max\{ 1, e^{\pm 1}(r,t)\})\,{\mathrm d} t \le C(T) \eta^{2-n} & if $n=3$,\\
\int_{0}^{T} \sup_{r\ge\eta} \log( \max\{1, e(r,t)\})\,{\mathrm d} t
\le C(T)\big( 1 + \sqrt{|\log\eta|}\big) & if $n=2$.
\end{dcases*}
\end{equation}
\end{lemma}
\begin{proof}
We divide the proof into three steps.
\smallskip
1. Fix a point $r \in [\eta, \infty)$ and a small constant $0 < \delta < 1$. Since $\lim_{r\to\infty} u_k(r,t)=0$
and $\lim_{r\to\infty} e_k(r,t)=1$ for each $t\in[0,T]$ and $k\in\mathbb{N}$, we have
\begin{align*}
\dfrac{u_k}{r^{\delta}\sqrt{e_k}}(r,t)
&= \lim\limits_{r\to\infty}\dfrac{u_k}{r^{\delta}\sqrt{e_k}}(r,t)
- \int_{r}^{\infty} \Big\{ \zeta^{-\delta}\dfrac{\partial_r u_k}{\sqrt{e_k}} -\delta \zeta^{-\delta-1} \dfrac{u_k}{\sqrt{e_k}}
-\dfrac{1}{2}\zeta^{-\delta} u_k e_k^{-\frac{3}{2}}\partial_r e_k \Big\} (\zeta,t)\,{\mathrm d} \zeta\\
&\le \Big( \int_{r}^{\infty} \zeta^{-2\delta-m}{\mathrm d} \zeta \Big)^{\frac{1}{2}}\Big\{ \delta\Big( \int_{r}^{\infty} \dfrac{\snorm{u_k}^2}{\zeta^2 e_k}\,\zeta^m {\mathrm d} \zeta \Big)^{\frac{1}{2}}
+ \Big( \int_{r}^{\infty} \dfrac{\snorm{\partial_r u_k}^2}{e_k}\,\zeta^m {\mathrm d}\zeta \Big)^{\frac{1}{2}} \Big\}\\
&\quad + \dfrac{1}{2}\Big( \int_{r}^{\infty} \dfrac{\snorm{u_k}^2}{\zeta^2 e_k} \zeta^{-\delta+1} {\mathrm d} \zeta \Big)^{\frac{1}{2}} \Big( \int_{r}^{\infty} \dfrac{\snorm{\partial_r e_k}}{e_k^2} \zeta^{-\delta+1} {\mathrm d}\zeta\Big)^{\frac{1}{2}}\\
&\le \dfrac{r^{-\delta} \eta^{\frac{1-m}{2}}}{\sqrt{2\delta+m-1}}\Big\{ \delta\Big( \int_{\eta}^{\infty} \dfrac{\snorm{u}_k^2}{\zeta^2 e_k}(\zeta,t)\,\zeta^m {\mathrm d} \zeta \Big)^{\frac{1}{2}}
+ \Big( \int_{\eta}^{\infty} \dfrac{\snorm{\partial_r u_k}^2}{e_k}\,\zeta^m {\mathrm d}\zeta\Big)^{\frac{1}{2}}\Big\} \\
&\quad +\dfrac{1}{2} r^{-\delta} \eta^{1-m}\Big(\int_{\eta}^{\infty}
\dfrac{\snorm{u_k}^2}{\zeta^2 e_k}\,\zeta^{m} {\mathrm d} \zeta \Big)^{\frac{1}{2}}
\Big( \int_{\eta}^{\infty} \dfrac{\snorm{\partial_r e_k}}{e_k^2}\,\zeta^{m} {\mathrm d} \zeta \Big)^{\frac{1}{2}}
=\vcentcolon r^{-\delta}g_k(\eta,t),
\end{align*}
where we have used the fact that $-2\delta-m+1<0$ and $-\delta+1-m<0$ for $n=2, 3$ since $0<\delta<1$.
Multiplying both sides of the above inequality by $r^{\delta}$ and taking the limit inferior over $k\to\infty$,
it follows by Lemma \ref{lemma:klimholder} that
\begin{equation*}
\dfrac{|u|}{\sqrt{e}}(r,t) \le \liminf\limits_{k\to\infty} g_k(\eta,t)
\qquad\ \text{for all $(r,t)\in[\eta,\infty)\times[0,T]$.}
\end{equation*}
Taking the supremum in $r\in[\eta,\infty)$ and integrating in $t\in[0,T]$, we obtain
\begin{equation}\label{temp:dissip1}
\int_{0}^{T} \sup_{r\ge\eta}\,\dfrac{|u|}{\sqrt{e}}(r,t)\,{\mathrm d} t \le \int_{0}^{T} \liminf_{k\to\infty}g_{k}(\eta,t)\,{\mathrm d} t
\le \liminf_{k\to\infty} \int_{0}^{T} g_{k}(\eta,t)\,{\mathrm d} t,
\end{equation}
where we have used Fatou's lemma since $g_k(\eta,t)>0$.
By definition of $g_k(\eta,t)$
and Lemma \ref{lemma:entEstext}, we obtain that, for all $k\in\mathbb{N}$,
\begin{align*}
\int_{0}^{T} g_{k}(\eta,t) {\mathrm d} t
&\le C\eta^{\frac{1-m}{2}}\Big\{\int_{0}^{T}\Big( \int_{\eta}^{\infty}
\dfrac{\snorm{u}_k^2}{\zeta^2 e_k}\,\zeta^m{\mathrm d} \zeta \Big)^{\frac{1}{2}} {\mathrm d} t
+ \int_{0}^{T}\Big( \int_{\eta}^{\infty}
\dfrac{\snorm{\partial_r u_k}^2}{e_k} \,\zeta^m {\mathrm d} \zeta \Big)^{\frac{1}{2}}{\mathrm d} t \Big\} \\
&\quad + C \eta^{1-m} \int_{0}^{T} \Big( \int_{\eta}^{\infty} \dfrac{\snorm{u_k}^2}{\zeta^2 e_k}\,\zeta^{m} {\mathrm d} \zeta \Big)^{\frac{1}{2}} \Big( \int_{\eta}^{\infty} \dfrac{\snorm{\partial_r e_k}}{e_k^2}\,\zeta^{m}{\mathrm d}\zeta\Big)^{\frac{1}{2}} {\mathrm d} t \\
&\le C\eta^{\frac{1-m}{2}} T^{\frac{1}{2}} \Big\{ \Big( \int_{0}^{T} \int_{\eta}^{\infty}
\dfrac{\norm{u}_k^2}{\zeta^2 e_k}\,\zeta^m {\mathrm d} \zeta {\mathrm d} t \Big)^{\frac{1}{2}}
+ \Big( \int_{0}^{T}\int_{\eta}^{\infty}
\dfrac{\snorm{\partial_r u_k}^2}{e_k}\,\zeta^m {\mathrm d} \zeta {\mathrm d} t \Big)^{\frac{1}{2}} \Big\}\\
&\quad + C \eta^{1-m} \Big( \int_{0}^{T}\int_{\eta}^{\infty}
\dfrac{\snorm{u_k}^2}{\zeta^2 e_k}\,\zeta^{m} {\mathrm d} \zeta {\mathrm d} t \Big)^{\frac{1}{2}}
\Big( \int_{0}^{T} \int_{\eta}^{\infty} \dfrac{\snorm{\partial_r e_k}}{e_k^2}\,
\zeta^{m} {\mathrm d} \zeta {\mathrm d} t \Big)^{\frac{1}{2}}\\
&\le C(T) ( \eta^{\frac{1-m}{2}} + \eta^{1-m}),
\end{align*}
which, together with \eqref{temp:dissip1}, leads to $\eqref{temp:dissip0}_1$.
\smallskip
2. Consider the case: $n=3$. Let $(r,t)\in[\eta,\infty)\times[0,T]$. By $\lim_{r\to\infty} e_k(r,t) =1$, we have
\begin{align*}
\log e_k(r,t)
&= -\int_{r}^{\infty} \dfrac{\partial_r e_k}{e_k}(\zeta,t)\,{\mathrm d} r
\le \Big(\int_{r}^{\infty} \zeta^{-m} {\mathrm d} \zeta\Big)^{\frac{1}{2}}
\Big( \int_{r}^{\infty} \dfrac{\snorm{\partial_r e_k}^2}{e_k^2}(\zeta,t)\,\zeta^m {\mathrm d} \zeta \Big)^{\frac{1}{2}}\\
&\le \dfrac{\eta^{2-n}}{\sqrt{n-2}}\Big( \int_{\eta}^{\infty}
\dfrac{\snorm{\partial_r e_k}^2}{e_k^2}(\zeta,t) \,\zeta^m {\mathrm d} \zeta \Big)^{\frac{1}{2}}
=\vcentcolon h_k(\eta,t).
\end{align*}
Thus, $0 \le \log ( \max\{ 1, e_k \} ) (r,t) \le h_k(\eta,t)$ for all $(r,t)\in[\eta,\infty)\times[0,T]$.
Replacing the term $e_k(r,t)$ with $e_k^{-1}(r,t)$ in the above derivation,
then $0 \le \log ( \max\{ 1, e_k^{-1}\} ) (r,t) \le h_k(\eta,t)$ for all $(r,t)\in[\eta,\infty)\times[0,T]$.
Taking the limit inferior over $k\to\infty$ on both sides of these inequalities,
it follows from Lemma \ref{lemma:klimholder} that
\begin{align*}
\log ( \max\{ 1, e^{\pm 1} \} ) (r,t) \le \liminf_{k\to\infty}h_{k}(\eta,t) \qquad\ \text{for all $(r,t)\in[\eta,\infty)\times[0,T]$.}
\end{align*}
Taking the supremum in $r\ge\eta$ on the above inequality and integrating in $t\in [0,T]$, we have \begin{equation}\label{temp:dissip2}
\int_{0}^{T} \sup_{r\ge\eta} \log ( \max\{ 1, e^{\pm 1}(r,t) \} )\,{\mathrm d} t
\le \int_{0}^{T} \liminf_{k\to\infty}h_{k}(\eta,t)\,{\mathrm d} t
\le \liminf_{k\to\infty} \int_{0}^{T} h_{k}(\eta,t)\,{\mathrm d} t,
\end{equation}
where we have used Fatou's lemma since $h_k(\eta,t)>0$.
By Lemma \ref{lemma:entEstext}, we obtain that, for all $k\in\mathbb{N}$,
\begin{align*}
\int_{0}^{T} h_{k}(\eta,t)\,{\mathrm d} t
\le C \eta^{2-n} \sqrt{T} \Big( \int_{0}^{T}\int_{\eta}^{\infty}
\dfrac{\snorm{\partial_r e_k}^2}{e_k^2}\,\zeta^m {\mathrm d} \zeta {\mathrm d} t\Big)^{\frac{1}{2}}
\le C(T) \eta^{2-n},
\end{align*}
which, together with \eqref{temp:dissip2}, proves $\eqref{temp:dissip0}_2$.
\smallskip
3. We consider the case: $n=2$.
Set $M \vcentcolon= \sup_{k\in\mathbb{N}} \sup_{t\in[0,T]} \widetilde{r}_k(1,t)$
so that, by Lemma \ref{lemma:rbd}, $M\le (2C_0)^{\frac{1}{n}}$.
If $r\ge M$, then
$r > \sup_{t\in[0,T]} \widetilde{r}_k(1,t)$ for all $k\in\mathbb{N}$.
By Lemma \ref{lemma:HREext}, $e_k(r,t)\le C(\varepsilon) \sigma^{-\frac{1}{2}}(t)$ for all $t\in[0,T]$ and $r\in[\widetilde{r}_k(\varepsilon,t),\infty)$.
It follows that there exists a constant $C(T)>1$
such that
\begin{equation}\label{temp:logM1}
\log (\max \{ 1,e_k(r,t) \}) \le \log (C(T) \sigma^{-\frac{1}{2}}(t)) \qquad\,
\text{for all $(r,t)\in[M,\infty)\times[0,T]$.}
\end{equation}
By
\eqref{temp:logM1}, for all $r\in[\eta,M)$,
\begin{equation}\label{temp:logM2}
\begin{aligned}
\log e_k(r,t) & = \log e_k(2M,t) - \int_{r}^{2M}\dfrac{\partial_r e_k}{e_k}(\zeta,t)\,{\mathrm d} \zeta\\
&\le \log (\max\{1,e_k(2M,t)\}) + \Big(\int_{r}^{2M} \zeta^{-1} {\mathrm d} \zeta \Big)^{\frac{1}{2}} \Big(\int_{r}^{2M} \dfrac{\snorm{\partial_r e_k}^2}{e_k^2}(\zeta,t)\,\zeta {\mathrm d} \zeta\Big)^{\frac{1}{2}}\\
&\le \log (C(T) \sigma^{-\frac{1}{2}}(t))
+ \sqrt{\log (C(T)/\eta)} \Big(\int_{\eta}^{\infty}
\dfrac{\snorm{\partial_r e_k}^2}{e_k^2}(\zeta,t)\,\zeta {\mathrm d} \zeta\Big)^{\frac{1}{2}}.
\end{aligned}
\end{equation}
Finally, in a similar way to that for the case $n=3$,
\begin{align*}
\int_{0}^{T} \sup_{r\ge\eta}\log (\max\{ 1, e(r,t) \}) {\mathrm d} t
&\le \dfrac{C^2(T)}{2} + \int_{0}^{T} \sup_{r\in[\eta,M)}\log (\max\{ 1, e(r,t)\})\,{\mathrm d} t \\
&\le C(T) + \sqrt{T\log(C(T)/\eta)} \sup_{k\in\mathbb{N}}\Big( \int_{0}^{T}\int_{\eta}^{\infty} \dfrac{\snorm{\partial_r e_k}^2}{e_k^2}\,\zeta^m {\mathrm d} \zeta {\mathrm d} t\Big)^{\frac{1}{2}}\\
&\le C(T)\big(1+ \sqrt{|\log\eta|}\big).
\end{align*}
\end{proof}
\section{Global Weak Solutions including the Origin}\label{sec:ato0}
We denote $(\rho_a,u_a,e_a)(r,t)$ in $[a,\infty)\times[0,T]$ as the solution
of problem \eqref{eqs:SFNS} and \eqref{eqs:eSFNS}, which is guaranteed by Theorem \ref{thm:WSCEx},
with the modified initial data $(\rho_a^0,u_a^0,e_a^0)(r)$ constructed in \S \ref{subsec:mollify}
for each $a\in(0,1)$. We also denote $\widetilde{r}_a(x,t)$ as the particle path function associated
with $\rho_a$, which satisfies
\begin{equation*}
x= \int_{a}^{\widetilde{r}_a(x,t)} \rho_a(y,t)\,r^m {\mathrm d} r \qquad
\text{for {\it a.e.} $(x,t)\in [0,\infty)\times[0,T]$}.
\end{equation*}
The main aim of this section is to take the limit: $a\searrow 0$ of the solution sequence
$(\rho_a,u_a,e_a)(r,t)$ of problems \eqref{eqs:SFNS} and \eqref{eqs:eSFNS}
to obtain a global spherically symmetric weak solution of the original Cauchy
problem \eqref{eqs:FNS}--\eqref{eqs:CauchyInit}.
\subsection{Uniform inequalities from the entropy estimates} \label{subsec:meas}
We obtain a set of inequalities involving measurable subsets of $[a,\infty)$, which
are uniform in $a\in(0,1)$.
In this section, we denote
\begin{equation}\label{eqs:GpsiH}
G(\zeta) \vcentcolon= 1-\zeta+\zeta \log \zeta,
\quad \psi(\zeta) \vcentcolon= \zeta -1 - \log \zeta,
\quad H(\zeta) \vcentcolon= \zeta \log \zeta,
\end{equation}
which are all convex functions.
Let $G^{-1}_{+}$, $\psi_{+}^{-1}$, and $H_{+}^{-1}$ be the right branch inverses of $G$, $\psi$,
and $H$, respectively.
Then, by Proposition \ref{prop:omega},
the following functions are well defined for $(y,z)\in(0,\infty)^2\vcentcolon$
\begin{equation}\label{eqs:fis}
f_1(y;z)\vcentcolon= y G_{+}^{-1}( \dfrac{z}{y} ), \quad
f_2( y; z) \vcentcolon= y \psi_{+}^{-1}(\dfrac{z}{y}) - z,
\quad f_3(y;z) \vcentcolon= y H_{+}^{-1}(\dfrac{z}{y}).
\end{equation}
Some further properties of $(y,z)\mapsto f_i(y;z)$ for $i=1,2,3$ can be found
in Proposition \ref{prop:omega}.
\begin{lemma}\label{lemma:rhoM}
For any measurable set $E\subset [a,\infty)$ and each $t\in[0,T]$,
\begin{align}
&\sup_{a\in(0,1)}\int_{E} \rho_a(r,t)\,r^m {\mathrm d} r \le \omega_1(E; C(T)),\label{6.3a} \\
&\sup_{a\in(0,1)}\int_{E} (\rho_a e_a)(r,t)\,r^m {\mathrm d} r \le C(T) + \omega_2(E;C(T)),\label{6.3b}
\end{align}
where $\omega_1(E;z) \vcentcolon = f_1 (\int_{E} r^m {\mathrm d} r ; z)$
and
$\omega_2(E;z) \vcentcolon = f_2(\omega_1(E;z);z)
= f_2( f_1(\int_{E}r^m {\mathrm d} r ; z);z)$ for $z>0$.
\end{lemma}
\begin{proof}
Denote $\mathcal{L}_n(E) \vcentcolon= \int_{E} r^m {\mathrm d} r$ for each $E\subseteq [a,\infty)$.
By Jensen's inequality and Theorem \ref{thm:WSCEx}\ref{item:WSCEx1}, it follows that
\begin{align*}
G( \dfrac{1}{\mathcal{L}_n(E)} \int_{E} \rho_a(r,t)\,r^m {\mathrm d} r)
\le \dfrac{1}{\mathcal{L}_n(E)} \int_{E} G(\rho_a(r,t))\,r^m {\mathrm d} r \le \dfrac{C(T)}{\mathcal{L}_n(E)}.
\end{align*}
Taking the right branch inverse $G_{+}^{-1}(z)$ on both sides of the above
and using the fact that $z\mapsto G_{+}^{-1}(z)$ is monotone increasing, it follows that, for all $a\in(0,1)$,
\begin{align*}
\int_{E} \rho_a(r,t)\,r^m {\mathrm d} r
\le \mathcal{L}_n(E) G_{+}^{-1}(\dfrac{C(T)}{\mathcal{L}_n(E)})
= f_1( \mathcal{L}_n(E) ; C(T)) = \vcentcolon \omega_1(E;C(T)).
\end{align*}
Now, fix $t\in[0,T]$ and set $m_{a,t}(E) \vcentcolon= \int_{E} \rho_a(r,t)\,r^m {\mathrm d} r$.
By Jensen's inequality and Theorem \ref{thm:WSCEx}\ref{item:WSCEx1}, it follows that
\begin{align*}
\psi( \dfrac{1}{m_{a,t}(E)} \int_E (\rho_a e_a) (r,t)\,r^m {\mathrm d} r)
\le \dfrac{1}{m_{a,t}(E)} \int_{E} (\rho_a \psi(e_a)) (r,t)\,r^m {\mathrm d} r \le \dfrac{C(T)}{m_{a,t}(E)}.
\end{align*}
Taking $\psi_{+}^{-1}(\cdot)$ on both sides of the above, it follows that
\begin{align*}
\int_E (\rho_a e_a) (r,t)\,r^m {\mathrm d} r \le m_{a,t}(E) \psi_{+}^{-1}(\dfrac{C(T)}{m_{a,t}(E)})
= f_2(m_{a,t}(E); C(T)) + C(T).
\end{align*}
Using \eqref{6.3a}, $m_{a,t}(E) \le \omega_1(E ;C(T))$.
Since $y\mapsto f_2(y;z)$ is increasing in $y$ for fixed $z>0$ from Proposition \ref{prop:omega},
it follows that, for all $a\in(0,1)$,
\begin{align*}
\int_E (\rho_a e_a) (r,t)\,r^m {\mathrm d} r
\le C(T) + f_2(\omega_1(E;C(T)); C(T)) = \vcentcolon C(T) + \omega_2(E;C(T)).
\end{align*}
\end{proof}
\subsection{Existence of particle paths and vacuum interfaces}\label{subsec:alimpath}
\begin{lemma}\label{lemma:alimpath}
There exist both a subsequence $\{a_j\}_{j\in\mathbb{N}}$ with $a_j \searrow 0$ as $j\to \infty$
and a continuous function $\widetilde{r}(x,t)\vcentcolon [0,\infty)\times[0,T]\to [0,\infty)$ such that
\begin{enumerate}[label=(\roman*),ref=(\roman*),font={\normalfont\rmfamily}]
\item\label{item:apath1} for each compact subset $K\Subset (0,\infty)\times[0,T]$,
\begin{align*}
\lim\limits_{j\to \infty} \sup_{(x,t)\in K} \snorm{\widetilde{r}_{a_j}(x,t)-\widetilde{r}(x,t)} = 0.
\end{align*}
\item\label{item:apath2} for each fixed $\varepsilon>0$,
\begin{equation*}
\begin{aligned}
&\snorm{\widetilde{r}(x_1,t)-\widetilde{r}(x_2,t)} \le C(\varepsilon) \snorm{x_1-x_2}
&&\quad \text{for all $(x_1,x_2,t)\in [\varepsilon,\infty)^2\times[0,T]$},\\
&\snorm{\widetilde{r}(x,t_1)-\widetilde{r}(x,t_2)} \le C(\varepsilon) \snorm{t_1^{\frac{3}{4}}-t_2^{\frac{3}{4}}}
&&\quad \text{for all $(x,t_1,t_2)\in [\varepsilon,\infty)\times[0,T]^2$}.
\end{aligned}
\end{equation*}
\item\label{item:apath3} for all $t\in[0,T]$, the function{\rm :} $x\mapsto \widetilde{r}(x,t)$ is strictly increasing
so that $\underline{r}(t) \vcentcolon= \lim_{x\to 0^+} \widetilde{r}(x,t)$ exists for all $t\in[0,T]$
and
\begin{equation*}
\lim_{t\searrow 0} \underline{r}(t) = 0, \qquad t\mapsto \underline{r}(t) \ \text{is upper semi-continuous}.
\end{equation*}
In addition, $\mathring{\mathrm{F}}\vcentcolon= \{(r,t)\in(0,\infty)\times(0,T)\,\vcentcolon\, \underline{r}(t)<r\}$
is an open set.
\item\label{item:apath4} $\widetilde{r}(x,t)$ satisfies the pointwise bound{\rm :}
\begin{equation*}
n x \psi^{-1}_{-}(\dfrac{C_0}{x}) \le r^n(x,t) \le C_0( 1 +x )
\qquad \text{for all $(x,t)\in[0,\infty)\times[0,T]$}.
\end{equation*}
\item\label{item:apath5} if $\underline{r}(t)>0$ for some $t\in(0,T]$, then
\begin{equation*}
\lim\limits_{j\to \infty} \int_{a_j}^{\underline{r}(t)} \rho_{a_j}(r,t)\,r^m {\mathrm d} r =0.
\end{equation*}
In particular, for such $t\in(0,T]$,
\begin{equation*}
\liminf\limits_{j\to \infty} \rho_{a_j} (r,t) = 0 \qquad \text{for a.e. $r\in(0,\underline{r}(t))$}.
\end{equation*}
\end{enumerate}
\end{lemma}
\begin{proof} We divide the proof into six steps.
\smallskip
1. For each integer $i\in\mathbb{N}$, let $S_{i}\vcentcolon= [i^{-1},i]\times[0,T]$.
From Theorem \ref{thm:WSCEx}\ref{item:WSCEx3}--\ref{item:WSCEx4}, it follows that, for all $(x,t)\in S_i$,
\begin{equation}\label{temp:apath1}
\widetilde{r}_a(x,t) = \widetilde{r}_a^0(x) + \int_{0}^{t} u_a( \widetilde{r}(x,s), s)\,{\mathrm d} s
\le \widetilde{r}_a^0(x) + C(i) \int_{0}^{T} \sigma^{-\frac{1}{4}}(s)\,{\mathrm d} s \le C(i)(1+T),
\end{equation}
with $C(i)>0$ a constant independent of $a\in(0,1)$, and
\begin{equation}\label{temp:apath2}
\begin{split}
&\snorm{\widetilde{r}_a(x_1,t)-\widetilde{r}_a(x_2,t)} \le C(i) \snorm{x_1-x_2} \qquad\text{for all $(x_1,x_2,t)\in [i^{-1},i]^2\times[0,T]$},\\
&\snorm{\widetilde{r}_a(x,t_1)-\widetilde{r}_a(x,t_2)} \le C(i) \snorm{t_1^{\frac{3}{4}}-t_2^{\frac{3}{4}}} \qquad\,
\text{for all $(x,t_1,t_2)\in [i^{-1},i]\times[0,T]^2$}.
\end{split}
\end{equation}
In addition, $S_i\subset S_{i+1}$, $\text{Int}(S_i)\neq \emptyset$ for each $i\in\mathbb{N}$, and $(0,\infty)\times[0,T]=\cup_{i\in\mathbb{N}}S_i$.
Thus, by Proposition \ref{prop:aaExt}, it follows that there exists a continuous function $\widetilde{r}(x,t)\vcentcolon(0,\infty)\times[0,T]\to [0,\infty)$ such that,
for all compact subset $K\Subset (0,\infty)\times[0,T]$,
\begin{equation}\label{temp:apath3}
\lim\limits_{j\to\infty} \sup\limits_{(x,t)\in K} \snorm{\widetilde{r}_{a_j}(x,t)-\widetilde{r}(x,t)} = 0,
\end{equation}
which, along with \eqref{temp:apath2} and Theorem \ref{thm:WSCEx}\ref{item:WSCEx4}, implies Lemma \ref{lemma:alimpath}\ref{item:apath2}.
\smallskip
2. Let $0<x_1<x_2<\infty$. It follows from Theorem \ref{thm:WSCEx}\ref{item:WSCEx3}--\ref{item:WSCEx4} that,
for {\it a.e.} $t\in[0,T]$,
\begin{align*}
x_2 - x_1 =\int_{\widetilde{r}_a(x_1,t)}^{\widetilde{r}_a(x_2,t)} \rho_a(r,t)\,r^m {\mathrm d} r
\le C(x_1)\int_{\widetilde{r}_a(x_1,t)}^{\widetilde{r}_a(x_2,t)} r^m {\mathrm d} r
= \dfrac{C(x_1)}{n} \big(\widetilde{r}_a^{\,n}(x_2,t) - \widetilde{r}_a^{\,n}(x_1,t)\big),
\end{align*}
where
we have used that $\rho_a(r,t)\le C(x_1)$ for {\it a.e.} $(r,t)$
such that $r \ge \widetilde{r}_a(x_1,t)$.
Taking $j\to\infty$ and using \eqref{temp:apath3} yield
\begin{equation*}
0 < x_2 - x_1 \le \dfrac{C(x_1)}{n} \big(\widetilde{r}^{\,n}(x_2,t)-\widetilde{r}^{\,n}(x_1,t)\big).
\end{equation*}
This shows that $x\mapsto\widetilde{r}(x,t)$ is strictly increasing for {\it a.e.} $t\in[0,T]$.
Since $(x,t)\mapsto \widetilde{r}(x,t)$ is continuous, then $x\mapsto\widetilde{r}(x,t)$ is strictly increasing
for all $t\in[0,T]$. Since $\widetilde{r}(x,t)\ge 0$,
\begin{equation}\label{temp:apath4}
\underline{r}(t) \vcentcolon= \lim\limits_{x\to 0^+} \widetilde{r}(x,t) \quad \text{ exists for all $t\in[0,T]$}.
\end{equation}
From Theorem \ref{thm:WSCEx}\ref{item:WSCEx3}, it follows that $\widetilde{r}_a(x,0)=\widetilde{r}_a^0(x)$
for each $a\in(0,1)$ and $x\in[0,\infty)$, where $\widetilde{r}_a^{0}(x)$ is the function constructed
in \S \ref{subsec:ffaprox}, satisfying
\begin{equation*}
x=\int_{a}^{\widetilde{r}_a^0(x)} \rho_a^{0}(r)\,r^m {\mathrm d} r\qquad \text{for all $x\in[0,\infty]$}.
\end{equation*}
Then, by the uniform convergence \eqref{temp:apath3} and Proposition \ref{prop:mollify} in \S \ref{subsec:mollify},
we have
\begin{equation*}
x=\int_{0}^{\widetilde{r}(x,0)}\rho_0(r)\,r^m {\mathrm d} r \ge \dfrac{\widetilde{r}^{\,n}(x,0)}{C_0 n}\qquad
\text{for all $x>0$}.
\end{equation*}
This implies the inequality: $\widetilde{r}(x,0) \le (n C_0 x)^{\frac{1}{n}}$.
Given $\delta>0$, let $x_\delta>0$ be such that $\widetilde{r}(x_\delta,0)\le ( n C_0 x_\delta )^{\frac{1}{n}}
\le \frac{\delta}{2}$.
Since $t\mapsto \widetilde{r}(x_\delta,t)$ is continuous in $[0,T]$, there exists $t_\delta>0$ such that,
if $t\in(0,t_\delta)$, then $\widetilde{r}(x_\delta,t) < \widetilde{r}(x_\delta,0) + \frac{\delta}{2}\le
\delta$. By \eqref{temp:apath4}, $\underline{r}(t)=\inf\{ \widetilde{r}(x,t) \vcentcolon x>0 \}$ for each $t\in[0,T]$.
It follows that $\underline{r}(t)\le \widetilde{r}(x_\delta,t) < \delta$ if $t\in(0,t_\delta)$.
This shows that $\lim_{t\searrow 0}\underline{r}(t)=0$.
\smallskip
3. We now show that $t\mapsto \underline{r}(t)$ is upper semi-continuous.
To start with, a function $f\vcentcolon {\mathbb R} \to {\mathbb R}$ is upper semi-continuous if and only if
the set $\{x\in{\mathbb R}\vcentcolon f(x)<y \}$ is open for all $y\in{\mathbb R}$.
Define the sets:
\begin{equation*}
\begin{aligned}
&U_{\zeta} \vcentcolon = \{ s\in(0,T) \vcentcolon \underline{r}(s) < \zeta \} \ &&\text{ for each $\zeta\in(0,\infty)$,} \\
&U_{\zeta}^x \vcentcolon = \{ s\in(0,T) \vcentcolon \widetilde{r}(x,s) < \zeta \} \
&&\text{ for each $(x,\zeta)\in(0,\infty)^2$.}
\end{aligned}
\end{equation*}
Suppose that $t\in U_{\zeta}$.
Since $\underline{r}(t)=\inf_{x>0} \widetilde{r}(x,t)$, there exists $\delta>0$
such that $\zeta > \widetilde{r}(\delta,t) \ge \underline{r}(t)$.
Thus, $t\in U_{\zeta}^{\delta} \subseteq \cup_{x>0} U_{\zeta}^{x}$, which implies that
$U_{\zeta}\subseteq \cup_{x>0} U_{\zeta}^{x}$.
On the other hand, if $t\in U_{\zeta}^{\delta}$ for some $\delta>0$,
then $\zeta>\widetilde{r}(\delta,t)\ge \underline{r}(t)$ so that $t\in U_{\zeta}$,
which implies that $U_{\zeta}= \cup_{x>0}U_{\zeta}^{x}$.
Since $t\mapsto \widetilde{r}(x,t)$ is continuous,
$U_{\zeta}^{x}$ is an open set for each $x>0$.
Therefore, $U_{\zeta}$ is also an open set as it is the union of open sets.
Since this is true for arbitrary $\zeta>0$,
and $\{s\in[0,T]\vcentcolon \underline{r}(s)<\zeta \}=\emptyset$ for $\zeta\le 0$,
$t\mapsto \underline{r}(t)$ is upper semi-continuous.
\smallskip
4. For each $\varepsilon\in(0,1]$, define
$\mathring{\mathrm{F}}_{\varepsilon} \vcentcolon= \{(r,t)\in(0,\infty)\times(0,T)\,\vcentcolon\,\widetilde{r}(\varepsilon,t)< r\}$.
Then $\mathring{\mathrm{F}}_{\varepsilon}$ is an open set, since $t\mapsto \widetilde{r}(\varepsilon,t)$ is continuous in $t\in[0,T]$.
Since $\underline{r}(t)=\inf_{x>0}\widetilde{r}(x,t)$ for all $t\in[0,T]$,
it follows that $\mathring{\mathrm{F}}_{\varepsilon}\subset \mathring{\mathrm{F}}$ for each $\varepsilon>0$.
Furthermore, if $(r,t)\in\mathring{\mathrm{F}}$, then there exists $\varepsilon>0$ such
that $\underline{r}(t) \le \widetilde{r}(\varepsilon,t)<r$, which implies $(r,t)\in\mathring{\mathrm{F}}_{\varepsilon}$
so that $\mathring{\mathrm{F}}=\cup_{\varepsilon>0}\mathring{\mathrm{F}}_{\varepsilon}$.
Thus, $\mathring{\mathrm{F}}$ is open as it is a union of open sets. This concludes the proof for Lemma \ref{lemma:alimpath}\ref{item:apath3}.
\smallskip
5. We show the upper and lower bounds of $\widetilde{r}(x,t)$.
From Theorem \ref{thm:WSCEx}\ref{item:WSCEx3}, we have
\begin{equation*}
nx \psi_{-}^{-1}(\frac{C_0}{x}) \le \widetilde{r}_a^{\,n}(x,t) \le C_0(1 + x) \qquad\ \text{for all $(a,x,t)\in(0,1)\times[0,\infty)\times[0,T]$,}
\end{equation*}
which, along with the uniform convergence in \eqref{temp:apath3}, implies Lemma \ref{lemma:alimpath}\ref{item:apath4}.
\smallskip
6. We now prove Lemma \ref{lemma:alimpath}\ref{item:apath5}.
If $\underline{r}(t)>0$ for some $t\in(0,T]$, then, for each $a\in (0,\underline{r}(t))$,
Theorem \ref{thm:WSCEx}\ref{item:WSCEx3} and \eqref{6.3a} imply that
\begin{align*}
\Bignorm{\int_{a}^{\underline{r}(t)} \rho_a(r,t)\,r^m {\mathrm d} r}
&\le \Bignorm{\int_{a}^{\widetilde{r}_a(x,t)} \rho_a(r,t)\,r^m {\mathrm d} r}
+ \Bignorm{\int_{\widetilde{r}_a(x,t)}^{\widetilde{r}(x,t)} \rho_a(r,t) r^m {\mathrm d} r} + \Bignorm{\int_{\widetilde{r}(x,t)}^{\widetilde{r}(t)}\rho_a(r,t)\,r^m {\mathrm d} r}\\
&\le x + \omega_1\big( [\widetilde{r}(x,t),\widetilde{r}_a(x,t)]; C(T) \big)
+ \omega_1\big( [\underline{r}(t),\widetilde{r}(x,t)]; C(T) \big),
\end{align*}
where $\omega_1(E;C) \vcentcolon= f_1( \int_{E}r^m {\mathrm d} r ; C)$ with
$f_1(y;C) \vcentcolon= y G_+^{-1}(\frac{C}{y})$.
By Proposition \ref{prop:omega}, for each fixed $C>0$, $y\mapsto f_1(y;C)$ is a positive, monotone increasing,
continuous function with $\lim_{y\searrow 0} f_1(y;C) = 0$.
Hence, given $\varepsilon>0$, there exists $\delta^{\prime}>0$ such that
$$
f_1( \frac{\snorm{\widetilde{r}^{\,n}(x,t)-\underline{r}^{\,n}(t)}}{n};C)\le \frac{\varepsilon}{2} \qquad\mbox{for $x\in(0,\delta^{\prime})$}.
$$
Let $\delta\vcentcolon= \min\{ \delta^{\prime}, \frac{\varepsilon}{2}\}$. We have
\begin{equation*}
\int_{a}^{\underline{r}(t)} \rho_a(r,t)\,r^m {\mathrm d} r
\le \varepsilon + f_1( \dfrac{\widetilde{r}_a^{\,n}(x,t)-\widetilde{r}^{\,n}(x,t)}{n}; C(T))
\qquad \mbox{if $x\in (0,\delta)$.}
\end{equation*}
Now, fix $x\in(0,\delta)$. It follows from \eqref{temp:apath3} and $\lim_{y\searrow 0} f_1(y;C) = 0$ that
\begin{equation*}
\limsup\limits_{a_j\searrow 0} \int_{a_j}^{\underline{r}(t)} \rho_{a_j}(r,t)\,r^m {\mathrm d} r
= \varepsilon + \lim\limits_{a_j\searrow 0}f_1(\dfrac{(\widetilde{r}_{a_j}(x,t))^n-(\widetilde{r}(x,t))^n}{n}; C(T)) = \varepsilon.
\end{equation*}
Since this is true for arbitrarily small $\varepsilon>0$, taking $\varepsilon\searrow 0$ yields
\begin{equation}\label{temp:alimpath1}
\limsup\limits_{j\to\infty} \int_{a_j}^{\underline{r}(t)} \rho_{a_j}(r,t)\,r^m {\mathrm d} r =0.
\end{equation}
Since $\rho_{a}(r,t)\ge C^{-1}(a)$ for {\it a.e.} $(r,t)\in[a,\infty)\times[0,T]$ from
Theorem \ref{thm:WSCEx}\ref{item:WSCEx2},
\begin{align*}
\int_{a}^{\underline{r}(t)} \rho_a(r,t)\,r^m {\mathrm d} r
\ge \frac{1}{C(a)}\int_{a}^{\underline{r}(t)} r^m {\mathrm d} r >0 \qquad
\text{for all $a\in(0,\underline{r}(t))$},
\end{align*}
which, together with \eqref{temp:alimpath1},
implies the first statement of Lemma \ref{lemma:alimpath}\ref{item:apath5}.
Since $\rho_a(r,t)>0$ for each $a\in(0,1)$, by Fatou's lemma,
\begin{align*}
\int_{0}^{\infty} \liminf\limits_{a_j\searrow 0}\rho_{a_j}(r,t) \mathbbm{1}_{(a_j,\underline{r}(t))}(r) r^m {\mathrm d} r \le \liminf\limits_{a_j\searrow 0} \int_{a_j}^{\underline{r}(t)} \rho_{a_j}(r,t)\,r^m {\mathrm d} r =0.
\end{align*}
Thus, $\liminf_{j\to\infty}\rho_{a_j}(r,t) \mathbbm{1}_{(a_j,\underline{r}(t))}(r) = 0$ for {\it a.e.} $r\in(0,\underline{r}(t))$.
Now fix such $r\in(0,\underline{r}(t))$.
It follows that $\rho_{a_j}(r,t) = \rho_{a_j}(r,t) \mathbbm{1}_{(a_j,\underline{r}(t))}(r)$ for all $a_j\in(0,r)$.
Taking the limit inferior as $j\to\infty$ on both sides, we obtain the second statement
of Lemma \ref{lemma:alimpath}\ref{item:apath5}.
\end{proof}
In what follows, we denote the fluid region $\mathrm{F}$
as the set:
\begin{align*}
\mathrm{F}\vcentcolon= \{ (r,t)\in(0,\infty)\times[0,T]\,\vcentcolon\, r> \underline{r}(t) \}.
\end{align*}
\subsection{Compactness results in the exterior domain} \label{subsec:alim}
In this subsection, we show the convergence of the approximate solutions in the exterior domain
and obtain several properties of the limit solutions.
\begin{lemma} \label{lemma:alimue}
There exist both a subsequence $a_j\searrow 0$ as $j\to\infty$
and H\"older continuous functions
$(u,e) \vcentcolon \mathrm{F}\backslash\{t=0\}\to \mathbb{R}\times [0,\infty)$
such that, for any compact subset $K\Subset \mathrm{F}\backslash\{t=0\}$,
\begin{equation*}
\lim\limits_{j\to \infty} \sup\limits_{(r,t)\in K}
\snorm{ ( u_{a_j} - u , e_{a_j} - e )(r,t) } = 0.
\end{equation*}
Moreover, for each $\varepsilon>0$, there exists $C(\varepsilon)>0$ such that
\begin{enumerate}[label=(\roman*),ref=(\roman*),font={\normalfont\rmfamily}]
\item If $t\in[0,T]$ and $r\in[\widetilde{r}(\varepsilon,t),\infty)$,
\begin{equation}\label{temp:alimue1}
\sigma^{\frac{1}{4}}(t) \snorm{u(r,t)}
+ \sigma^{\frac{1}{2}}(t) e(r,t)\le C(\varepsilon);
\end{equation}
\item If $t\in(0,T]$ and $r_1,\, r_2 \ge \widetilde{r}(\varepsilon,t)$,
\begin{equation}\label{temp:alimue2}
\sigma^{\frac{1}{2}}(t)\snorm{u(r_1,t)-u(r_2,t)} + \sigma(t)\snorm{e(r_1,t)-e(r_2,t)}
\le C(\varepsilon) \snorm{r_1-r_2}^{\frac{1}{2}};
\end{equation}
\item If $0<t_1<t_2\le T$ and $r\ge \sup_{t_1\le t\le t_2}\widetilde{r}(\varepsilon,t)$,
\begin{equation}\label{temp:alimue3}
\sigma^{\frac{1}{2}}(t_1)\snorm{u(r,t_1)-u(r,t_2)} + \sigma(t_1)\snorm{e(r,t_1)-e(r,t_2)} \le C(\varepsilon) \snorm{t_2-t_1}^{\frac{1}{4}}.
\end{equation}
\item $(u,e)(r,t)$ satisfy
\begin{align}
&\sup_{ 0\le t \le T} \int_{\widetilde{r}(\varepsilon,t)}^{\infty}
\big(\snorm{(u,e-1)}^2 + \snorm{(\sqrt{\sigma}\partial_r u, \sigma\partial_r e)}^2 \big)(r,t) r^m {\mathrm d} r\nonumber\\
&\, +\int_{0}^{T}\int_{\widetilde{r}(\varepsilon,t)}^{\infty}\big(\snorm{( \partial_r u, \partial_r e, u \partial_r u )}^2
+\snorm{(\sqrt{\sigma} \partial_t u, \sigma \partial_t e)}^2\big) r^m {\mathrm d} r {\mathrm d} t\le C(\varepsilon).\label{6.13b}
\end{align}
\end{enumerate}
\end{lemma}
\begin{proof}
We present the proof only for $u(r,t)$, since the case for $e(r,t)$ follows in the same way.
First, fix $\varepsilon\in(0,1)$ and define the domain:
\begin{equation*}
\mathrm{F}_\varepsilon \vcentcolon
=\big\{(r,t)\in (0,\infty)\times[0,T]\, \vcentcolon\, t\in[\varepsilon,T], \,\widetilde{r}(\varepsilon,t) \le r \le \varepsilon^{-1}\big\}.
\end{equation*}
By Lemma \ref{lemma:alimpath}\ref{item:apath4}, there exists $\varepsilon_0>0$ such that
$\text{Int}(\mathrm{F}_\varepsilon)\neq \emptyset$ for each $\varepsilon\in(0,\varepsilon_0)$.
Since $x\mapsto\widetilde{r}(x,t)$ is strictly increasing, then $\mathrm{F}_{\varepsilon_1} \subset \mathrm{F}_{\varepsilon_2}$ if $\varepsilon_1 > \varepsilon_2$.
\smallskip
We claim:
\begin{equation} \label{fltemp}
\mathrm{F}\backslash \{t=0\} = \bigcup\limits_{0<\varepsilon<1} \mathrm{F}_\varepsilon.
\end{equation}
This can be seen as follows:
If $(r,t)\in \mathrm{F}_\varepsilon$ for some $\varepsilon\in(0,1)$, then $r> \underline{r}(t)$ since $ \widetilde{r}(x,t) \searrow \underline{r}(t)$
as $x\searrow 0$ for each $t\in[0,T]$, which implies that $(r,t)\in \mathrm{F}$.
On the other hand, if $(r,t)\in \mathrm{F}\backslash\{t=0\}$, then $\underline{r}(t)< r$ and
there exists $\varepsilon_1>0$ such that $t>\varepsilon_1$.
Since $\lim_{x\searrow 0}\widetilde{r}(x,t) = \underline{r}(t)$ for all $t\in[0,T]$, there exists $\varepsilon_2\in(0,1)$
such that $\underline{r}(t) \le \widetilde{r}(\varepsilon_2,t) \le r$.
Thus, $(r,t)\in \mathrm{F}_\varepsilon$ for $\varepsilon\vcentcolon=\min\{\varepsilon_1,\varepsilon_2\}$.
Then \eqref{fltemp} follows.
\smallskip
Next, since $x\mapsto \widetilde{r}(x,t)$ is strictly increasing for each $t\in[0,T]$,
and $(x,t)\mapsto \widetilde{r}(x,t)$ is continuous, then
${\mathrm D}_{\varepsilon}\vcentcolon=\inf_{t\in[0,T]}\{ \widetilde{r}(\varepsilon,t) - \widetilde{r}(\frac{\varepsilon}{2},t) \}>0$ by the compactness of $[0,T]$.
By Lemma \ref{lemma:alimpath}, there exists $N_{\varepsilon}\in\mathbb{N}$ such that
$\sup_{t\in[0,T]} \snorm{\widetilde{r}_{a_j}(\frac{\varepsilon}{2},t)-\widetilde{r}(\frac{\varepsilon}{2},t)}\le \frac{{\mathrm D}_{\varepsilon}}{2}$ for all $j \ge N_\varepsilon$.
Hence, $\widetilde{r}_{a_j}(\frac{\varepsilon}{2},t) < \widetilde{r}(\varepsilon,t)$
for all $j\ge N_\varepsilon$ and $t\in[0,T]$, which implies that,
if $(r,t)\in\mathrm{F}_\varepsilon$, then $\widetilde{r}_{a_j}(\frac{\varepsilon}{2},t)< r$ for all $j\ge N_\varepsilon$.
Combining this with Theorem \ref{thm:WSCEx}\ref{item:WSCEx4}, we have
\begin{equation}\label{fltemp2}
\begin{dcases*}
\sup\limits_{j\ge N_\varepsilon}\snorm{u_{a_j}(r_1,t)-u_{a_j}(r_2,t)}
\le C(\frac{\varepsilon}{2}) \sigma^{-\frac{1}{2}}(t) \snorm{r_1-r_2}^{\frac{1}{2}} & for all $(r_1,t), (r_2,t)\in \mathrm{F}_\varepsilon$,\\
\sup\limits_{j\ge N_\varepsilon}\snorm{u_{a_j}(r,t_1)-u_{a_j}(r,t_2)} \le C(\frac{\varepsilon}{2}) \sigma^{-\frac{1}{2}}(t_1) \snorm{t_2-t_1}^{\frac{1}{4}}
& for all $(r,t_1), (r,t_2)\in \mathrm{F}_\varepsilon$.
\end{dcases*}
\end{equation}
In addition, it also follows from Theorem \ref{thm:WSCEx}\ref{item:WSCEx4} that
\begin{equation}\label{fltemp3}
\sup_{j\ge N_\varepsilon}\snorm{u_{a_j}(r,t)} \le C(\varepsilon) \sigma^{-\frac{1}{4}}(t)
\le \varepsilon^{-\frac{1}{4}}C(\varepsilon)
\qquad \text{ for all $(r,t)\in \mathrm{F}_{\varepsilon}$}.
\end{equation}
In light of \eqref{fltemp}--\eqref{fltemp3}, we can apply Proposition \ref{prop:aaExt} to obtain a further
subsequence (still denoted) $a_j$ and a continuous function
$u(r,t)\vcentcolon \mathrm{F}\backslash\{t=0\}\to \mathbb{R}$
such that
\begin{equation}\label{fltemp4}
\lim\limits_{j\to\infty} \sup\limits_{(r,t)\in K} \snorm{u_{a_j}(r,t)-u(r,t)}=0 \qquad
\text{for all compact subset $K \Subset \mathrm{F}\backslash\{t=0\}$},
\end{equation}
which together with Theorem \ref{thm:WSCEx}\ref{item:WSCEx4} implies \eqref{temp:alimue1}--\eqref{temp:alimue3}.
Similarly, we can establish similar estimates for $e(r,t)$ in the same way.
The proof for
\eqref{6.13b} follows the same as that of Lemma \ref{lemma:weak}.
\end{proof}
\begin{lemma}\label{lemma:alimrho}
For each $\varepsilon>0$, define
$$
\rho_{a}^{(\varepsilon)}(r,t) \vcentcolon= \rho_a(r,t) \chi_\varepsilon^a(r,t), \quad
\rho^{(\varepsilon)}(r,t) \vcentcolon= \rho(r,t) \chi_\varepsilon(r,t),
$$
where $\chi_\varepsilon^a$ and $\chi_\varepsilon$ are the indicator functions defined as
\begin{equation}\label{6.6indicator}
\chi_\varepsilon(r,t) \vcentcolon = \begin{dcases*}
1 & if $r\in [\widetilde{r}(\varepsilon,t),\infty),$\\
0 & otherwise,
\end{dcases*} \qquad \ \chi_\varepsilon^a(r,t)
\vcentcolon = \begin{dcases*}
1 & if $r\in [\widetilde{r}_a(\varepsilon,t),\infty)$,\\
0 & otherwise.
\end{dcases*}
\end{equation}
Then there exists $C(\varepsilon)>0$ such that the following statements hold{\rm :}
\begin{enumerate}[label=(\roman*), ref=(\roman*),font=\normalfont\rmfamily]
\item\label{item:alimrho1}
There exist both a nonnegative function $\rho(r,t) \in L^2_{\text{\rm loc}}(\mathrm{F},r^m{\mathrm d} r {\mathrm d} t)$
and a subsequence $a_j\searrow 0$ as $j\to\infty$ such that, for each $\varepsilon>0$,
\begin{equation*}
\begin{cases}
\rho_{a_j}^{(\varepsilon)} - 1 \overset{\ast}{\rightharpoonup} \rho^{(\varepsilon)} - 1 \qquad
\text{in $L^{\infty}\big(0,T; L^{2}([0,\infty), r^m{\mathrm d} r)\big)$ \,\, as $j\to\infty$},\\
\sup_{t\in[0,T]}\int_{\widetilde{r}(\varepsilon,t)}^{\infty} \snorm{\rho^{(\varepsilon)}-1}^2(r,t)\,r^m {\mathrm d} r \le C(\varepsilon).
\end{cases}
\end{equation*}
\smallskip
\item\label{item:alimrho2}
For all $L\in\mathbb{N}$ with $X_L\vcentcolon= \widetilde{H}^{-1}( [0,L], r^m {\mathrm d} r)$,
\begin{equation*}
\begin{cases}
\rho^{(\varepsilon)} \in C^{0}([0,T];X_L), \qquad \lim\limits_{j\to\infty}\sup\limits_{t\in[0,T]}\sbnorm{\rho_{a_j}^{(\varepsilon)}(\cdot,t)-\rho^{(\varepsilon)}(\cdot,t)}_{X_L} = 0,\\
\sbnorm{\rho^{(\varepsilon)}(\cdot,t_1)-\rho^{(\varepsilon)}(\cdot,t_2)}_{X_L} \le C(\varepsilon) \snorm{t_1-t_2} \qquad
\text{for all $t_1,t_2 \in[0,T]$}.
\end{cases}
\end{equation*}
\smallskip
\item\label{item:alimrho3}
For all $x>0$,
$x=\int_{\underline{r}(t)}^{\widetilde{r}(x,t)} \rho(r,t)\,r^m {\mathrm d} r$
for $t\in[0,T]$ almost everywhere,
where $\widetilde{r}(x,t)$ and $\underline{r}(t)$ are the functions obtained in {\rm Lemma \ref{lemma:alimpath}}.
\smallskip
\item\label{item:alimrho4}
$C^{-1}(\varepsilon) \le \rho(r,t) \le C(\varepsilon)$ for almost every $t\in[0,T]$ and $r\in[\widetilde{r}(\varepsilon,t),\infty)$.
\end{enumerate}
\end{lemma}
\begin{proof}
We divide the proof into four steps, showing Lemma \ref{lemma:alimrho}\ref{item:alimrho1}--\ref{item:alimrho4}
in chronological order.
\smallskip
1. Proof of Lemma \ref{lemma:alimrho}\ref{item:alimrho1}.
From Theorem \ref{thm:WSCEx}\ref{item:WSCEx4}, we obtain that, for all $j\in\mathbb{N}$,
\begin{align*}
\sup_{t\in[0,T]} \int_{0}^{\infty} \snorm{\rho_{a_j}^{(\varepsilon)}-1}^2\, r^m {\mathrm d} r
\le \dfrac{\widetilde{r}_{a_j}^{\,n}(\varepsilon,t)}{n}
+ \sup_{t\in[0,T]} \int_{\widetilde{r}_{a_j}(\varepsilon,t)}^{\infty} \snorm{\rho_{a_j}-1}^2\,r^m {\mathrm d} r \le C(\varepsilon).
\end{align*}
Then there exists $\widetilde{w}_\varepsilon-1 \in L^{\infty}(0,T; L^2([0,\infty),r^m {\mathrm d} r) )$ such that
\begin{equation}\label{Ftemp0}
\rho_{a_j}^{(\varepsilon)}-1 \overset{\ast}{\rightharpoonup} \widetilde{w}_\varepsilon-1 \qquad
\text{weakly star in $L^{\infty}\big(0,T; L^{2}([0,\infty), r^m{\mathrm d} r)\big)\,\,$ as $j\to\infty$}.
\end{equation}
For each $l \in \mathbb{N}$, denote $\varepsilon_l \vcentcolon= l^{-1}$ and $\widetilde{w}^{(l)} \vcentcolon= \widetilde{w}_{\varepsilon_l}$.
Fix $l,\,q\in\mathbb{N}$ such that $l< q$.
Then $\widetilde{r}_a(\varepsilon_q,t)< \widetilde{r}_a(\varepsilon_l,t)$ for each $(a,t)\in(0,1)\times[0,T]$.
Thus, by construction, it follows that, for all $j\in\mathbb{N}$,
$\chi_{\varepsilon_l}^{a_j}(r,t)=1=\chi_{\varepsilon_q}^{a_j}(r,t)$ if $r\in [\widetilde{r}_{a_j}(\varepsilon_l,t),\infty)$ so that
\begin{equation}\label{Ftemp1}
\rho_{a_j}^{(\varepsilon_q)}(r,t) = \rho_{a_j}^{(\varepsilon_l)}(r,t) \ \ \text{for {\it a.e.} $t\in[0,T]$}, \qquad\,\,\, r\in[\widetilde{r}_{a_j}(\varepsilon_l,t),\infty) \ \ \text{for each $j\in\mathbb{N}$}.
\end{equation}
Let $\phi \in C^{\infty}_{\rm c}(\mathrm{F}^{(l)})$ for
$\mathrm{F}^{(l)}\vcentcolon= \{ (r,t) \vcentcolon t\in[0,T], r\in [\widetilde{r}(\varepsilon_l,t),\infty) \}$.
Then it follows that $\phi \chi_{\varepsilon_l} \in L^{1}(0,T;L^2([0,\infty),r^m{\mathrm d} r))$ and
\begin{align*}
& \int_{0}^{T} \int_{\widetilde{r}(\varepsilon_l,t)}^{\infty} \{\widetilde{w}^{(q)}-\widetilde{w}^{(l)}\} \phi\, r^m {\mathrm d} r {\mathrm d} t\\
&=\int_{0}^{T} \int_{0}^{\infty} \{\widetilde{w}^{(q)}-\rho_{a_j}^{(\varepsilon_q)}\} \chi_{\varepsilon_l} \phi\, r^m {\mathrm d} r {\mathrm d} t + \int_{0}^{T}\int_{\widetilde{r}(\varepsilon_l,t)}^{\widetilde{r}_{a_j}(\varepsilon_l,t)} \{\rho_{a_j}^{(\varepsilon_q)}-\rho_{a_j}^{(\varepsilon_l)}\}
\phi\, r^m {\mathrm d} r {\mathrm d} t \\
&\quad+ \int_{0}^{T}\int_{\widetilde{r}_{a_j}(\varepsilon_l,t)}^{\infty} \{\rho_{a_j}^{(\varepsilon_q)}-\rho_{a_j}^{(\varepsilon_l)}\}
\phi\,r^m {\mathrm d} r{\mathrm d} t
+\int_{0}^{T} \int_{0}^{\infty} \{\rho_{a_j}^{(\varepsilon_l)}-\widetilde{w}^{(l)}\} \chi_{\varepsilon_l} \phi\, r^m {\mathrm d} r {\mathrm d} t \\
&= \sum_{i=1}^4 I_i^{(j)} .
\end{align*}
The limits of $I_1^{(j)}$ and $I_4^{(j)}$ are evaluated by the weak-star convergence \eqref{Ftemp0}.
As $j\to\infty$,
\begin{align*}
I_1^{(j)} + I_4^{(j)}\vcentcolon
= \int_{0}^{T} \int_{0}^{\infty} \{\widetilde{w}^{(q)}-\rho_{a_j}^{(\varepsilon_q)}\} \chi_{\varepsilon_l} \phi\, r^m {\mathrm d} r {\mathrm d} t
+ \int_{0}^{T} \int_{0}^{\infty} \{\rho_{a_j}^{(\varepsilon_l)}-\widetilde{w}^{(l)}\}
\chi_{\varepsilon_l} \phi\, r^m {\mathrm d} r {\mathrm d} t \to 0.
\end{align*}
Using \eqref{Ftemp1}, we also see that
\begin{align*}
I_3^{(j)} \vcentcolon= \int_{0}^{T}\int_{\widetilde{r}_{a_j}(\varepsilon_l,t)}^{\infty} \{\rho_{a_j}^{(\varepsilon_q)}-\rho_{a_j}^{(\varepsilon_l)}\}
\phi\, r^m {\mathrm d} r {\mathrm d} t = 0.
\end{align*}
For $I_2^{(j)}$, observe that
\begin{equation*}
\snorm{\chi_{\varepsilon_l}^{a_j}-\chi_{\varepsilon_l}}(r,t) = \begin{dcases*}
1 & if $r\in [\widetilde{r}_{a_j}(\varepsilon_l,t),\widetilde{r}(\varepsilon_l,t)]$ or $r\in[\widetilde{r}(\varepsilon_l,t),\widetilde{r}_{a_j}(\varepsilon_l,t)]$,\\
0 & otherwise.
\end{dcases*}
\end{equation*}
By the uniform convergence from Lemma \ref{lemma:alimpath},
it follows that $\snorm{\chi_{\varepsilon_l}^{a_j}-\chi_{\varepsilon_l}}(r,t)\to 0$ as $j\to\infty$ for {\it a.e.} $(r,t)\in[0,\infty)\times[0,T]$.
Using the estimates from Theorem \ref{thm:WSCEx}\ref{item:WSCEx4} and the dominated convergence theorem yield
that, as $j\to\infty$,
\begin{align*}
\snorm{I_2^{(j)}}
&\le \int_{0}^{T}\int_{0}^{\infty} \snorm{\chi_{\varepsilon_l}^{a_j}-\chi_{\varepsilon_l}}\snorm{\rho_{a_j}^{(\varepsilon_q)}-\rho_{a_j}^{(\varepsilon_l)}} \snorm{\phi}\,r^m {\mathrm d} r {\mathrm d} t \\
&\le \int_{0}^{T} \Big( \int_{0}^{\infty} \snorm{\chi_{\varepsilon_l}^{a_j}-\chi_{\varepsilon_l}}^2 \snorm{\phi}^2\, r^m {\mathrm d} r \Big)^{\frac{1}{2}}
\big(\sbnorm{\rho_{a_j}^{(\varepsilon_q)}(\cdot,t) -1}_{L^2} + \sbnorm{\rho_{a_j}^{(\varepsilon_l)}(\cdot,t) -1}_{L^2}\big)\,{\mathrm d} t\\
&\le \big( C(\varepsilon_l) + C(\varepsilon_q)\big)\int_{0}^{T} \Big( \int_{0}^{\infty} \snorm{\chi_{\varepsilon_l}^{a_j}-\chi_{\varepsilon_l}}^2
\snorm{\phi}^2\, r^m {\mathrm d} r \Big)^{\frac{1}{2}} {\mathrm d} t \to 0 .
\end{align*}
Combining the convergence for $I_i^{(j)}$ as $j\to\infty$ for $i=1,2,3,4$, it follows that
\begin{align*}
\int_{0}^{T} \int_{\widetilde{r}(\varepsilon_l,t)}^{\infty} \{\widetilde{w}^{(q)}-\widetilde{w}^{(l)}\} \phi\, r^m {\mathrm d} r {\mathrm d} t = 0
\qquad \text{for all $\phi \in C^{\infty}_{\rm c}(\mathrm{F}^{(l)})$.}
\end{align*}
It follows that, if $l\le q \in \mathbb{N}$, then
\begin{equation}\label{Ftemp6}
\widetilde{w}^{(q)}(r,t) = \widetilde{w}^{(l)}(r,t) \qquad
\text{for almost every $t\in[0,T]$ and $r\in [\widetilde{r}(\varepsilon_l,t),\infty)$.}
\end{equation}
In light of above, $\widetilde{w}^{(q)}(r,t)$ is truncated in domain $\mathrm{F}$ for each $q\in\mathbb{N}$ as
\begin{equation*}
w^{(q)}(r,t) \vcentcolon=
\begin{dcases*}
\widetilde{w}^{(q)}(r,t) & if $t\in[0,T]$ and $r\in[\widetilde{r}(\varepsilon_q,t),\infty)$,\\
0 & otherwise.
\end{dcases*}
\end{equation*}
Using this, for each $N\in\mathbb{N}$, define
\begin{equation*}
h_N(r,t) \vcentcolon= w^{(1)}(r,t) + \sum_{i=1}^{N} \{ w^{(i+1)}(r,t) - w^{(i)}(r,t) \} \qquad
\text{for $(r,t)\in \mathrm{F}$.}
\end{equation*}
Fix $(r,t)\in \mathrm{F}$.
Lemma \ref{lemma:alimpath}\ref{item:apath3}--\ref{item:apath4} imply that
$\lim_{x\searrow 0}\widetilde{r}(x,t)=\underline{r}(t)$ and $\lim_{x\to \infty}\widetilde{r}(x,t)=\infty$.
Since $\underline{r}(t)< r$, there exists $q\in\mathbb{N}$ such that $\widetilde{r}(\varepsilon_q,t)\le r < \widetilde{r}(\varepsilon_{q-1},t)$.
Then \eqref{Ftemp6} implies
\begin{equation*}
h_{N}(r,t) = w^{(1)}(r,t) + \sum_{i=1}^N \{ w^{(i+1)}(r,t) - w^{(i)}(r,t) \} = w^{(q)}(r,t) = \widetilde{w}^{(q)}(r,t)
\quad \text{ if $N\ge q+1$.}
\end{equation*}
Thus, for {\it a.e.} $(r,t)\in \mathrm{F}$,
$\lim_{N\to\infty} h_N(r,t) = \widetilde{w}^{(q)}(r,t)$ exists for some $q\in\mathbb{N}$,
and the following function is well-defined almost everywhere in $(0,\infty)\times[0,T]$:
\begin{equation*}
\rho(r,t) \vcentcolon=\begin{dcases*}
\lim\limits_{N\to\infty} h_N(r,t) & for {\it a.e.} $(r,t)\in \mathrm{F}$,\\
0 & for $(r,t)\in (0,\infty)\times[0,T] \backslash \mathrm{F}$.
\end{dcases*}
\end{equation*}
Now let $\varepsilon>0$. Then there exists some integer $q\in\mathbb{N}$ such that $\varepsilon_q \le \varepsilon$ and
\begin{equation*}
\rho^{(\varepsilon)}(r,t) = \chi_\varepsilon(r,t) \rho(r,t) = \chi_\varepsilon(r,t) \widetilde{w}^{(q)}(r,t) \quad \
\text{for {\it a.e.} $(r,t)\in (0,\infty)\times[0,T]$},
\end{equation*}
from the above construction.
Let $\phi \in L^{1}(0,T;L^2([0,\infty);r^m {\mathrm d} r))$. Using $\widetilde{r}_{a_j}(\varepsilon_q,t)< \widetilde{r}_{a_j}(\varepsilon,t)$ for each $j\in\mathbb{N}$, one has
\begin{equation}\label{Ftemp2}
\chi_\varepsilon^{(a_j)}= \chi_{\varepsilon_q}^{(a_j)}\chi_{\varepsilon}^{(a_j)} \qquad
\text{for all $(r,t)\in[0,\infty)\times[0,T]$ and $j\in\mathbb{N}$.}
\end{equation}
Using this, it then follows that
\begin{align*}
\int_{0}^{T}\int_{0}^{\infty} \{ \rho^{(\varepsilon)} - \rho_{a_j}^{(\varepsilon)} \} \phi\, r^m {\mathrm d} r {\mathrm d} t
&= \int_{0}^{T} \int_{0}^{\infty} \chi_\varepsilon \phi \{ \widetilde{w}^{(q)} - \rho_{a_j}^{(\varepsilon_q)} \}\, r^m {\mathrm d} r {\mathrm d} t
+\int_{0}^{T} \int_{0}^{\infty} \{ \chi_{\varepsilon} - \chi_{\varepsilon}^{a_j} \} \rho_{a_j}^{(\varepsilon_q)} \phi\, r^m {\mathrm d} r {\mathrm d} t\\
&= \vcentcolon J_1^{(j)} + J_2^{(j)}.
\end{align*}
By the weak-star convergence for each $q\in\mathbb{N}$ in \eqref{Ftemp0},
$J_1^{(j)}\to 0$ as $j\to\infty$.
For $J_2^{(j)}$, one uses the pointwise bound of $\rho_a$ from Theorem \ref{thm:WSCEx}\ref{item:WSCEx4} to obtain
\begin{equation*}
\snorm{(\chi_\varepsilon - \chi_\varepsilon^{a_j})\rho_{a_j}^{(\varepsilon_q)} \varphi } \le C(\varepsilon_q) \snorm{\varphi} \in L^{1}\big(0,T;L^2( [0,\infty), r^m {\mathrm d} r)\big).
\end{equation*}
Moreover, due to the uniform convergence from Lemma \ref{lemma:alimpath},
$(\chi_\varepsilon - \chi_\varepsilon^{(a_j)})(r,t)\to 0$ as $j\to \infty$ for {\it a.e.} $(r,t)\in [0,\infty)\times[0,T]$.
Hence, by the dominated convergence theorem,
\begin{align*}
J_2^{(j)} \vcentcolon
= \int_{0}^{T} \int_{0}^{\infty} ( \chi_{\varepsilon} - \chi_{\varepsilon}^{a_j} ) \rho_{a_j}^{(\varepsilon_q)} \varphi\, r^m {\mathrm d} r {\mathrm d} t \to 0
\qquad \text{as $j\to\infty$.}
\end{align*}
Combining the limits of $J_1^{(j)}$ and $J_2^{(j)}$ as $j\to\infty$, it follows that there exists
$\rho \in L^2_{\text{loc}}(\mathrm{F})$ such that, for each $\varepsilon>0$,
\begin{equation}\label{Ftemp7}
\rho^{(\varepsilon)}_{a_j} \overset{\ast}{\rightharpoonup} \rho^{(\varepsilon)} \qquad
\text{in $L^{\infty}\big(0,T; L^2([0,\infty),r^m {\mathrm d} r)\big)\,\,$ as $j\to \infty$}.
\end{equation}
This completes the proof
of Lemma \ref{lemma:alimrho}\ref{item:alimrho1}.
\smallskip
2. Proof of Lemma \ref{lemma:alimrho}\ref{item:alimrho2}.
Define $r_\varepsilon\vcentcolon= \frac{1}{2}\inf_{t\in[0,T]}\widetilde{r}(\varepsilon,t)$.
By Lemma \ref{lemma:alimpath}, there exists $N_\varepsilon\in\mathbb{N}$ such that
$r_\varepsilon < \widetilde{r}_{a_j}(\varepsilon,t)$ for all $j\ge N_\varepsilon$ and $t\in[0,T]$.
By Theorem \ref{thm:WSCEx}\ref{item:WSCEx5}, we see that, for each $\varepsilon>0$ and $t_1, t_2 \in [0,T]$,
\begin{align*}
\sup\limits_{j \ge N_{\varepsilon}}\sbnorm{\rho_{a_j}^{(\varepsilon)}(\cdot, t_1)- \rho_{a_j}^{(\varepsilon)}(\cdot,t_2)}_{\widetilde{H}^{-1}([r_\varepsilon,L], r^m {\mathrm d} r)} \le C(\varepsilon)\snorm{t_1-t_2} \qquad \text{for all $L\in\mathbb{N}$.}
\end{align*}
Moreover, it follows that, for all $j\ge N_\varepsilon$ and $L\in\mathbb{N}$,
\begin{align*}
\rho_{a_j}^{(\varepsilon)} \in C^0\big([0,T]; \widetilde{H}^{-1}([r_\varepsilon,L],r^m {\mathrm d} r) \big), \quad \sup_{j \ge N_\varepsilon } \sbnorm{\rho_{a_j}^{(\varepsilon)}-1}_{L^{\infty}(0,T; L^{2}( [r_\varepsilon,\infty), r^m {\mathrm d} r))} \le C(\varepsilon).
\end{align*}
Therefore, by Proposition \ref{prop:scHm1},
there exists $\widetilde{h}_\varepsilon$ such that, for all $L\in\mathbb{N}$,
\begin{equation}\label{Ftemp3}
\begin{cases}
\widetilde{h}_\varepsilon-1 \in L^{\infty}\big(0,T;L^2([r_\varepsilon,\infty),r^m{\mathrm d} r)\big), \quad \widetilde{h}_\varepsilon \in C^0\big([0,T];\widetilde{H}^{-1}([r_\varepsilon,L], r^m {\mathrm d} r)\big),\\
\lim\limits_{j\to \infty} \sup_{t\in[0,T]}\sbnorm{\rho_{a_j}^{(\varepsilon)}(\cdot,t)-\widetilde{h}_\varepsilon(\cdot,t)}_{H^{-1}([r_\varepsilon,L], r^m {\mathrm d} r)} = 0,\\
\sbnorm{\widetilde{h}_\varepsilon(\cdot,t_1)-\widetilde{h}_\varepsilon(\cdot,t_2)}_{H^{-1}([r_\varepsilon,L],r^m{\mathrm d} r)} \le C(\varepsilon) \snorm{t_1-t_2}
\quad \text{ for each $t_1, t_2 \in[0,T]$.}
\end{cases}
\end{equation}
\smallskip
We claim that $\rho^{(\varepsilon)} = \widetilde{h}_\varepsilon$ for {\it a.e.} $(r,t)\in[r_\varepsilon,\infty)\times[0,T]$.
Let $\phi \in C_c^{\infty}( [r_\varepsilon,\infty) \times [0,T])$. Then there exists $L\in\mathbb{N}$ such that ${\mathrm {supp}}(\phi)\subseteq [r_\varepsilon,L]\times[0,T]$. Using \eqref{Ftemp7}--\eqref{Ftemp3}, we obtain that, as $j\to\infty$,
\begin{align*}
&\int_{0}^{T}\int_{r_\varepsilon}^{\infty} (\rho^{(\varepsilon)} - \widetilde{h}_\varepsilon) \phi\, r^m {\mathrm d} r {\mathrm d} t
= \int_{0}^{T}\int_{r_\varepsilon}^{\infty} (\rho^{(\varepsilon)} - \rho^{(\varepsilon)}_{a_j}) \phi\, r^m {\mathrm d} r {\mathrm d} t
+ \int_{0}^{T}\int_{r_\varepsilon}^{L} (\rho_{a_j}^{(\varepsilon)} - \widetilde{h}_\varepsilon) \phi\, r^m {\mathrm d} r {\mathrm d} t \\
&\le \int_{0}^{T}\int_{r_\varepsilon}^{\infty} \{\rho^{(\varepsilon)} - \rho^{(\varepsilon)}_{a_j}\} \phi\, r^m {\mathrm d} r {\mathrm d} t + \int_{0}^{T} \sbnorm{\rho_{a_j}^{(\varepsilon)}(\cdot,t)-\widetilde{h}_{\varepsilon}(\cdot,t)}_{\widetilde{H}^{-1}}\sbnorm{\phi}_{H^1} {\mathrm d} t\to 0.
\end{align*}
This implies that
$\rho^{(\varepsilon)}=\widetilde{h}_\varepsilon$ for {\it a.e.} $(r,t)\in[r_\varepsilon,\infty)\times[0,T]$.
Lemma \ref{lemma:alimrho}\ref{item:alimrho2} then follows since ${\mathrm {supp}}(\rho^{(\varepsilon)}(\cdot,t))=[\widetilde{r}(\varepsilon,t),\infty)$,
and $r_\varepsilon < \widetilde{r}(\varepsilon,t)$ for all $t\in[0,T]$.
\medskip
3. Proof of Lemma \ref{lemma:alimrho}\ref{item:alimrho3}.
Let $\xi\in C^{\infty}_{\rm c}(0,T)$ and $0<y< x$. By \eqref{Ftemp7}, we have
\begin{align*}
&\int_{0}^{T} \xi(t) \Big(\int_{\widetilde{r}_{a_j}(y,t)}^{\widetilde{r}(x,t)} \rho_{a_j}(r,t)\,r^m {\mathrm d} r\Big) {\mathrm d} t
-\int_{0}^{T} \xi(t) \Big(\int_{\widetilde{r}(y,t)}^{\widetilde{r}(x,t)} \rho(r,t)\, r^m {\mathrm d} r\Big){\mathrm d} t\\
&= \int_{0}^{T} \int_{0}^{\widetilde{r}(x,t)} \xi (\chi_y^{a_j}\rho_{a_j} - \chi_{y}\rho)\, r^m {\mathrm d} r {\mathrm d} t
= \int_{0}^{T} \int_{0}^{\infty} (1-\chi_{x})\xi \{\rho_{a_j}^{(y)} - \rho^{(y)}\}\,r^m {\mathrm d} r {\mathrm d} t \to 0
\end{align*}
as $j\to\infty$, where we have used the fact that $(1-\chi_x) \xi \in L^{1}(0,T; L^2([0,\infty),r^m {\mathrm d} r))$.
Then
\begin{equation}\label{Ftemp4}
\lim\limits_{j\to\infty}\int_{0}^{T} \int_{\widetilde{r}_{a_j}(y,t)}^{\widetilde{r}(x,t)} \xi(t) \rho_{a_j}(r,t)\,r^m {\mathrm d} r {\mathrm d} t
= \int_{0}^{T} \xi(t) \Big(\int_{\widetilde{r}(y,t)}^{\widetilde{r}(x,t)}\rho(r,t)\,r^m{\mathrm d} r\Big){\mathrm d} t.
\end{equation}
Moreover, using \eqref{6.3a}, $\lim_{y\searrow 0}\widetilde{r}(y,t)=\underline{r}(t)$, and the dominated convergence theorem,
\begin{equation*}
\lim\limits_{y\searrow 0}\sup\limits_{j\in\mathbb{N}}\Bignorm{\int_{0}^{T}\int_{\underline{r}(t)}^{\widetilde{r}(y,t)} \xi \rho_{a_j}
\,r^m {\mathrm d} r {\mathrm d} t}
\le \lim\limits_{y\searrow 0} \int_{0}^{T} \omega_{1}( [\underline{r}(t),\widetilde{r}(y,t)]; C(T) ) \snorm{\xi(t)}\,{\mathrm d} t = 0.
\end{equation*}
For a given $\delta>0$, the above limit implies that there exists $z_{\delta}>0$ such that, if $y\in(0,z_{\delta})$, then
\begin{equation}\label{Ftemp8}
\sup_{j\in\mathbb{N}}\Bignorm{\int_{0}^{T}\int_{\underline{r}(t)}^{\widetilde{r}(y,t)} \xi(t) \rho_{a_j}(r,t)\, r^m {\mathrm d} r {\mathrm d} t}
\le \delta.
\end{equation}
Now fix $y\in(0,z_\delta)$. By Theorem \ref{thm:WSCEx}\ref{item:WSCEx3}, for {\it a.e.} $t\in[0,T]$,
\begin{align*}
x = \int_{a_j}^{\widetilde{r}_{a_j}(x,t)} \rho_{a_j}(r,t)r^m {\mathrm d} r
= \Big(\int_{a_j}^{\underline{r}(t)} + \int_{\underline{r}(t)}^{\widetilde{r}(y,t)} + \int_{\widetilde{r}(y,t)}^{\widetilde{r}_{a_j}(y,t)} + \int_{\widetilde{r}_{a_j}(y,t)}^{\widetilde{r}(x,t)} + \int_{\widetilde{r}(x,t)}^{\widetilde{r}_{a_j}(x,t)} \Big) \rho_{a_j}\,r^m {\mathrm d} r.
\end{align*}
Multiplying the above equality with $\xi(t)$, then integrating in $t\in[0,T]$ and letting $j\to\infty$,
it follows from \eqref{Ftemp4}--\eqref{Ftemp8} that
\begin{equation}\label{Ftemp9}
-\delta \le \lim\limits_{j\to\infty}\sum_{i=1}^{3} \mathcal{I}_i^{(j)} +\int_{0}^{T} \xi(t) \Big\{ -x + \int_{\widetilde{r}(y,t)}^{\widetilde{r}(x,t)} \rho(r,t) \,r^m {\mathrm d} r \Big\}\,{\mathrm d} t \le \delta.
\end{equation}
Repeating the same proof for Lemma \ref{lemma:alimpath}\ref{item:apath5}, we have
\begin{align*}
\lim\limits_{j\to\infty}\snorm{\mathcal{I}_1^{(j)}} = \lim\limits_{j\to\infty}\Bignorm{\int_{0}^T \xi(t)\big(\int_{a_j}^{\underline{r}(t)} \rho_{a_j}(r,t)\,r^m {\mathrm d} r\big){\mathrm d} t} = 0.
\end{align*}
Moreover, using \eqref{6.3a}, Lemma \ref{lemma:alimpath}, and the dominated convergence theorem,
we see that
\begin{align*}
\lim\limits_{j\to\infty}\snorm{\mathcal{I}_2^{(j)}}+\snorm{\mathcal{I}_3^{(j)}}
&\le\lim\limits_{j\to\infty} \int_{0}^{T} \snorm{\xi(t)} \Big(\int_{\widetilde{r}(y,t)}^{\widetilde{r}_{a_j}(y,t)}+\int_{\widetilde{r}(x,t)}^{\widetilde{r}_{a_j}(x,t)}\Big) \rho_{a_j}(r,t)\,r^m {\mathrm d} r {\mathrm d} t \\
&\le \lim\limits_{j\to\infty} \int_{0}^{T} \snorm{\xi(t)}
\Big\{ \omega_1\big([\widetilde{r}(y,t),\widetilde{r}_{a_j}(y,t)];C(T)\big) + \omega_1\big([\widetilde{r}(x,t),\widetilde{r}_{a_j}(x,t)];C(T)\big) \Big\}\,{\mathrm d} t = 0.
\end{align*}
It follows from \eqref{Ftemp9} that, for each $\delta>0$, there exists $z_{\delta}>0$ such that, if $y\in(0,z_{\delta})$, then
\begin{equation*}
-\delta \le \int_{0}^{T} \xi(t) \Big\{ -x+ \int_{\widetilde{r}(y,t)}^{\widetilde{r}(x,t)} \rho(r,t) r^m {\mathrm d} r \Big\}\,{\mathrm d} t
\le \delta.
\end{equation*}
Let $\{\xi_k\}_{k\in\mathbb{N}} \in C^{\infty}_{\rm c}(0,T)$ be such that $\xi_k\ge 0$,
$\sbnorm{\xi_k}_{L^1([0,T])}=1$, and $\{\xi_k\}_{k\in\mathbb{N}}$ is a sequence of kernels approximating
the Dirac measure.
Then, for all $(k,y)\in\mathbb{N}\times(0,z_{\delta})$,
\begin{align*}
-\delta \le - x + \int_{0}^{T}\xi_k(t)\Big(\int_{\widetilde{r}(y,t)}^{\widetilde{r}(x,t)} \rho(r,t)\,r^m {\mathrm d} r\Big) {\mathrm d} t
\le \delta.
\end{align*}
Since $\rho(r,t)\ge 0$, we apply the monotone convergence theorem for $y\searrow 0$ to obtain
\begin{align*}
-\delta \le - x
+ \int_{0}^{T}\xi_k(t)\Big(\int_{\underline{r}(t)}^{\widetilde{r}(x,t)}\rho(r,t)\,r^m {\mathrm d} r\Big)\,{\mathrm d} t \le \delta.
\end{align*}
Since this is true for arbitrarily small $\delta>0$, it follows that
\begin{align*}
x=\int_{0}^{T}\xi_k(t)\Big(\int_{\underline{r}(t)}^{\widetilde{r}(x,t)} \rho(r,t)\,r^m {\mathrm d} r\Big)\,{\mathrm d} t \qquad \
\text{for each $i\in\mathbb{N}$}.
\end{align*}
Finally, since $\xi_k$ converges to the Dirac measure in the sense of distributions, we take $k\to \infty$ to
conclude that
\begin{align*}
x=\int_{\underline{r}(t)}^{\widetilde{r}(x,t)} \rho(r,t)\,r^m {\mathrm d} r \qquad \ \text{for $t\in[0,T]$ almost everywhere.}
\end{align*}
\medskip
4. The proof for \ref{item:alimrho4}
is the same as that of Proposition \ref{prop:veDomain} and Lemma \ref{lemma:rhoWeak}.
\end{proof}
In light of Lemmas \ref{lemma:alimpath}\ref{item:apath5} and \ref{lemma:alimrho}\ref{item:alimrho3},
we extend $\rho\in L^2_{\text{loc}}(\mathrm{F})$ obtained above by setting $\rho(r,t)=0$
for $(r,t)\in {\mathbb R}^n\times[0,T] \backslash \mathrm{F}$.
\begin{lemma}\label{lemma:alimEnt}
Let $(\rho,u,e)(r,t)$ be the limit function obtained in {\rm Lemmas \ref{lemma:alimue}}--{\rm \ref{lemma:alimrho}}.
Then
\begin{align*}
&\esssup_{t\in(0,T)}\int_{0}^{\infty}\big(\rho \snorm{u}^2 +G(\rho)\big)(r,t)\,r^m {\mathrm d} r \le C(T),\\
&\esssup_{t\in[0,T]}\int_{E} (\rho e)(r,t)\,r^m {\mathrm d} r \le C(T) + \omega_2(E;C(T)) \quad
\text{for bounded measurable $E\subset [0,\infty)$,}
\end{align*}
where $\omega_2(E,z)$ for $z>0$ is the set function given in {\rm Lemma \ref{lemma:rhoM}}.
\end{lemma}
\begin{proof}
Let $\varepsilon\in (0,1)$. Denote $\rho^{(\varepsilon)}$ and $\rho^{(\varepsilon)}_{a_j}$ as in Lemma \ref{lemma:alimrho}.
Let $E\subset [0,\infty)$ be a bounded measurable set, and let $\mathbbm{1}_E(r)$ be the indicator function.
Since $x\to \widetilde{r}(x,t)$ is strictly increasing and $(x,t)\mapsto\widetilde{r}(x,t)$ is a continuous function from Lemma \ref{lemma:alimpath}, combining with the fact that $[0,T]$ is a compact set yields that
${\mathrm D}_\varepsilon =\vcentcolon 2\inf_{t\in[0,T]}\{\widetilde{r}(\varepsilon,t) - \widetilde{r}(\frac{\varepsilon}{2},t)\}>0$.
By Lemma \ref{lemma:alimpath}, there exists $N_{\varepsilon}\in\mathbb{N}$ such that,
if $j\ge N_\varepsilon$, $\sup_{ 0\le t \le T} \snorm{\widetilde{r}_{a_j}(\varepsilon,t)-\widetilde{r}(\varepsilon,t)} \le \frac{d_\varepsilon}{2}$.
Then
\begin{equation}\label{temp:alimEnt2}
\widetilde{r}_{a_j}(\varepsilon,t) > \widetilde{r}(\frac{\varepsilon}{2},t) \qquad\ \text{for all $t\in[0,T]$ and $j\ge N_\varepsilon$}.
\end{equation}
Let $[t_1,t_2]\subset (0,T]$. Then it follows from \eqref{temp:alimEnt2} that, as $j\to\infty$,
\begin{align*}
&\int_{t_1}^{t_2}\int_{\widetilde{r}_{a_j}(\varepsilon,t)}^{\infty}
\rho_{a_j} \snorm{u_{a_j}}^2\mathbbm{1}_E(r)\,r^m {\mathrm d} r {\mathrm d} t
- \int_{t_1}^{t_2}\int_{\widetilde{r}(\varepsilon,t)}^{\infty} \rho \snorm{u}^2\mathbbm{1}_E(r)\,r^m {\mathrm d} r {\mathrm d} t\\
&= \int_{t_1}^{t_2}\int_{\widetilde{r}(\frac{\varepsilon}{2},t)}^{\infty} \rho_{a_j}^{(\varepsilon)}
\snorm{u_{a_j}}^2\mathbbm{1}_E(r)\,r^m {\mathrm d} r {\mathrm d} t
- \int_{t_1}^{t_2}\int_{\widetilde{r}(\frac{\varepsilon}{2},t)}^{\infty}\rho^{(\varepsilon)} \snorm{u}^2\mathbbm{1}_E(r)\,r^m {\mathrm d} r {\mathrm d} t\\
&= \int_{t_1}^{t_2}\int_{\widetilde{r}(\frac{\varepsilon}{2},t)}^{\infty} \rho_{a_j}^{(\varepsilon)}
\big\{ \snorm{u_{a_j}}^2 - \snorm{u}^2\big\}\mathbbm{1}_E(r)\,r^m {\mathrm d} r {\mathrm d} t
+ \int_{t_1}^{t_2} \int_{\widetilde{r}(\frac{\varepsilon}{2},t)}^{\infty}
\mathbbm{1}_E \big\{\rho_{a_j}^{(\varepsilon)}-\rho^{(\varepsilon)}\big\}\snorm{u}^2\mathbbm{1}_E(r)\,r^m {\mathrm d} r {\mathrm d} t\to 0,
\end{align*}
where the first-term convergence follows from the dominated convergence theorem,
Lemma \ref{lemma:alimue}, and the pointwise bound for $\snorm{u_{a_j}}$ and $\rho_{a_j}^{(\varepsilon)}$
from Theorem \ref{thm:WSCEx}\ref{item:WSCEx4}, while the second-term convergence follows from
Lemma \ref{lemma:alimrho}.
Let $L\in\mathbb{N}$ and set $E\equiv [0,L]$. By Theorem \ref{thm:WSCEx}\ref{item:WSCEx1},
\begin{align*}
\int_{t_1}^{t_2}\int_{\widetilde{r}(\varepsilon,t)}^{L} \rho \snorm{u}^2\,r^m {\mathrm d} r {\mathrm d} t
=&\, \lim_{j\to \infty}\int_{t_1}^{t_2}\int_{\widetilde{r}_{a_j}(\varepsilon,t)}^{L}
\rho_{a_j} \snorm{u_{a_j}}^2\,r^m {\mathrm d} r {\mathrm d} t\\
\le&\, \snorm{t_2-t_1} \esssup_{t\in(0,T)}\int_{a_j}^{\infty}
\rho_{a_j} \snorm{u_{a_j}}^2(r,t) r^m {\mathrm d} r = C(T) \snorm{t_2-t_1},
\end{align*}
where $C(T)=C(T,C_0)$ is independent of $a\in(0,1)$, $\varepsilon>0$, and $L\in\mathbb{N}$. Thus,
by the Lebesgue differentiation theorem, it follows that
\begin{align*}
\esssup_{t\in(0,T)} \int_{\widetilde{r}(\varepsilon,t)}^{L} (\rho \snorm{u}^2)(r,t)\,r^m {\mathrm d} r \le C(T).
\end{align*}
Using the monotone convergence theorem with the consecutive limits $L\to\infty$ and then $\varepsilon\searrow 0$,
\begin{align*}
\esssup_{t\in(0,T)} \int_{\underline{r}(t)}^{\infty} (\rho \snorm{u}^2)(r,t)\,r^m {\mathrm d} r
=\lim\limits_{\substack{L\to\infty\\ \varepsilon\to 0^+}}\esssup_{t\in(0,T)}
\int_{\widetilde{r}(\varepsilon,t)}^{L} (\rho \snorm{u}^2)(r,t)\,r^m {\mathrm d} r \le C(T).
\end{align*}
Next, let $[t_1,t_2]\subseteq(0,T]$, and let $E\subset [0,\infty)$ be a bounded measurable set.
Using \eqref{temp:alimEnt2}, the convergence results from Lemma \ref{lemma:alimue}--\ref{lemma:alimrho},
and the bound: $e(r,t)\le C(\varepsilon) \sigma^{-\frac{1}{2}}(t)$ for $r\ge \widetilde{r}(\frac{\varepsilon}{2},t)$ from Lemma \ref{lemma:alimue},
we see that, as $j\to\infty$,
\begin{align*}
&\int_{t_1}^{t_2}\int_{\widetilde{r}_{a_j}(\varepsilon,t)}^{\infty} \rho_{a_j} e_{a_j}\mathbbm{1}_E(r)\,r^m {\mathrm d} r {\mathrm d} t
-\int_{t_1}^{t_2}\int_{\widetilde{r}(\varepsilon,t)}^{\infty} \rho e\,\mathbbm{1}_E(r)\,r^m {\mathrm d} r {\mathrm d} t \\
&=\int_{t_1}^{t_2}\int_{\widetilde{r}(\frac{\varepsilon}{2},t)}^{\infty}\rho_{a_j}^{(\varepsilon)}e_{a_j}\mathbbm{1}_E(r)\,r^m{\mathrm d} r{\mathrm d} t -\int_{t_1}^{t_2}\int_{\widetilde{r}(\frac{\varepsilon}{2},t)}^{\infty} \rho^{(\varepsilon)} e\,\mathbbm{1}_E(r)\,r^m {\mathrm d} r {\mathrm d} t\\
&=\int_{t_1}^{t_2}\int_{\widetilde{r}(\frac{\varepsilon}{2},t)}^{\infty} \rho_{a_j}^{(\varepsilon)}
(e_{a_j}-e)\mathbbm{1}_E(r)\,r^m {\mathrm d} r {\mathrm d} t + \int_{t_1}^{t_2}\int_{\widetilde{r}(\frac{\varepsilon}{2},t)}^{\infty}
\big(\rho_{a_j}^{(\varepsilon)}-\rho^{(\varepsilon)}\big) e\,\mathbbm{1}_E(r)\,r^m {\mathrm d} r {\mathrm d} t \to 0.
\end{align*}
Let $E_j^{\varepsilon}(t)\vcentcolon= E\cap [\widetilde{r}_{a_j}(\varepsilon,t),\infty)$. Then, by \eqref{6.3b},
\begin{equation}\label{aEtemp1}
\begin{aligned}
\int_{t_1}^{t_2}\int_{\widetilde{r}(\varepsilon,t)}^{\infty}\rho e\,\mathbbm{1}_E(r)\, r^m {\mathrm d} r {\mathrm d} t
&= \lim\limits_{j\to\infty} \int_{t_1}^{t_2}\int_{\widetilde{r}_{a_j}(\varepsilon,t)}^{\infty}
\rho_{a_j} e_{a_j} \,\mathbbm{1}_E(r)\,r^m {\mathrm d} r {\mathrm d} t \nonumber\\
&\le \sup\limits_{j\in\mathbb{N}} \int_{t_1}^{t_2}\int_{E_j^\varepsilon(t)}\rho_{a_j} e_{a_j}\,r^m {\mathrm d} r{\mathrm d} t\\
&\le \int_{t_1}^{t_2} \big(C(T) + \omega_2(E_{j}^{\varepsilon}(t);C(T)) \big){\mathrm d} t,
\end{aligned}
\end{equation}
where $\omega_2\big(E_{j}^{\varepsilon}(t);C(T)\big) = f_2(f_1(\int_{E_{j}^{\varepsilon}(t)}r^m {\mathrm d} r;C(T));C(T))$.
Since $E_{j}^{\varepsilon}(t) \subseteq E$ for all $(j,\varepsilon,t)\in\mathbb{N}\times(0,1)\times[0,T]$,
and $f_1(y;C)$ and $f_2(y;C)$ are monotone increasing functions in $y$ (see Proposition \ref{prop:omega}),
then, for all $(j,\varepsilon,t)\in\mathbb{N}\times(0,1)\times[0,T]$,
\begin{align*}
\omega_2(E_{j}^{\varepsilon}(t);C(T))
\le f_2(f_1(\int_{E}r^m {\mathrm d} r;C(T));C(T)) = \omega_2(E;C(T)).
\end{align*}
Using this, \eqref{aEtemp1} becomes
\begin{align*}
\int_{t_1}^{t_2}\int_{\widetilde{r}(\varepsilon,t)}^{\infty} (\rho e)(r,t)\,\mathbbm{1}_E(r)\, r^m {\mathrm d} r {\mathrm d} t
\le \snorm{t_2-t_1} \big(C(T)+ \omega_2(E;C(T))\big).
\end{align*}
Thus, by the Lebesgue differentiation theorem, it follows that
\begin{align*}
\esssup\limits_{t\in[0,T]}\int_{\widetilde{r}(\varepsilon,t)}^{\infty}(\rho e) (r,t)\mathbbm{1}_E(r)\,r^m {\mathrm d} r
\le C(T)+ \omega_2(E;C(T)).
\end{align*}
Since $\rho=0$ in $\{r< \underline{r}(t)\}$ and $\lim_{\varepsilon\searrow 0}\widetilde{r}(\varepsilon,t)= \underline{r}(t)$,
by the monotone convergence theorem,
\begin{align*}
\esssup\limits_{t\in[0,T]}\int_{E}(\rho e)(r,t)\,r^m {\mathrm d} r \le C(T)+ \omega_2(E;C(T)).
\end{align*}
Let $\varepsilon\in(0,1]$ and $r_\varepsilon \vcentcolon=\frac{1}{2} \inf_{t\in[0,T]}\widetilde{r}(\varepsilon,t)$.
By Lemma \ref{lemma:alimpath}, there exists $M_\varepsilon\in\mathbb{N}$ such that
\begin{equation}\label{temp:alimEnt1}
\inf_{t\in[0,T]}\widetilde{r}_{a_j}(\varepsilon,t)> r_\varepsilon \qquad\ \text{for all $j\ge M_\varepsilon$}.
\end{equation}
Moreover, by Theorem \ref{thm:WSCEx}\ref{item:WSCEx3}, $\widetilde{r}_{a_j}(\varepsilon,t)\le \widetilde{r}_{a_j}(1,t)\le C_0$
for all $(j,t)\in\mathbb{N}\times[0,T]$.
By \eqref{temp:alimEnt1}, Theorem \ref{thm:WSCEx}\ref{item:WSCEx1}, Lemma \ref{lemma:alimrho},
and the fact that $G(0)=1$, we have
\begin{equation*}
\begin{dcases}
\rho_{a_j}^{(\varepsilon)} - 1 \rightharpoonup \rho^{(\varepsilon)} - 1
\quad\text{ weakly in $L^{2}\big(0,T; L^2([r_{\varepsilon},\infty),r^m {\mathrm d} r) \big)$ as $j\to\infty$},\\
\sup_{j\ge M_\varepsilon} \esssup_{t\in[0,T]} \int_{r_\varepsilon}^{\infty} G(\rho_{a_j}^{(\varepsilon)})(r,t)\,r^m {\mathrm d} r \le C(T), \\
0 \le \rho_{a_j}^{(\varepsilon)}(r,t) \le C(\varepsilon) \quad \text{ for {\it a.e.}\, $(r,t)\in[r_\varepsilon,\infty)\times[0,T]$
and for each $j\in\mathbb{N}$}.
\end{dcases}
\end{equation*}
Thus, applying Proposition \ref{prop:mazur},
it follows that
\begin{equation*}
\esssup_{t\in(0,T)}\int_{r_\varepsilon}^{\infty} G(\rho^{(\varepsilon)})(r,t)\, r^m {\mathrm d} r \le C(T).
\end{equation*}
By Lemma \ref{lemma:alimpath}\ref{item:apath4},
$r_\varepsilon < \widetilde{r}(\varepsilon,t)\le C_0(1+\varepsilon)^{\frac{1}{n}}\le 2^{\frac{1}{n}}C_0$ for $t\in[0,T]$.
Since $G(\rho^{(\varepsilon)})=G(0)=1$ for $r\in[r_\varepsilon,\widetilde{r}(\varepsilon,t)]$ and $\rho^{(\varepsilon)}=\rho$
for {\it a.e.} $r\in[\widetilde{r}(\varepsilon,t),\infty)$, then, for {\it a.e.} $t\in[0,T]$,
\begin{equation*}
\int_{\widetilde{r}(\varepsilon,t)}^{\infty} G(\rho)(r,t)\,r^m {\mathrm d} r
\le \int_{r_\varepsilon}^{\widetilde{r}(\varepsilon,t)}\,r^m {\mathrm d} r
+ \int_{\widetilde{r}(\varepsilon,t)}^{\infty} G(\rho)(r,t)\,r^m {\mathrm d} r \le C(T).
\end{equation*}
Since the above is true for all $\varepsilon\in(0,1]$,
using $\lim_{\varepsilon\searrow 0} \widetilde{r}(\varepsilon,t)= \underline{r}(t)$, $\rho \equiv 0$ for $r<\underline{r}(t)$,
and the monotone convergence Theorem, it follows that
\begin{equation*}
\esssup_{t\in[0,T]}\int_{0}^{\infty} G(\rho)(r,t)\,r^m {\mathrm d} r
=\lim\limits_{\varepsilon\searrow 0} \esssup_{t\in[0,T]}\int_{\widetilde{r}(\varepsilon,t)}^{\infty} G(\rho)(r,t)\,r^m {\mathrm d} r
\le C(T).
\end{equation*}
\end{proof}
\begin{lemma}\label{lemma:entdis}
Let $(\rho,u,e)(r,t)$ be the limit function obtained in {\rm Lemmas \ref{lemma:alimue}}--{\rm \ref{lemma:alimrho}}. Then for $\eta>0$,
\begin{align*}
&\int_{0}^{T} \sup_{r\ge\underline{r}(t)+\eta}\dfrac{\snorm{u}}{\sqrt{e}}(r,t)\,{\mathrm d} t
\le C(T)\big(\eta^{\frac{2-n}{2}} + \eta^{2-n} \big) && \text{if } \ n=2,\, 3,\\
&\int_{0}^{T} \sup_{r\ge \underline{r}(t)+\eta} \log \big( \max\big\{ 1, e(r,t)\big\} \big)\, {\mathrm d} t
\le C(T)\big(1+\sqrt{|\log\eta|}\big)&& \text{if } \ n=2,\\
&\int_{0}^{T} \sup_{r\ge\underline{r}(t)+\eta} \log \big( \max \big\{ 1, e^{\pm 1}(r,t) \big\} \big)\,{\mathrm d} t
\le C(T) \eta^{2-n} && \text{if } \ n=3.
\end{align*}
\end{lemma}
\begin{proof}
We will only prove the first estimate, as the other two statements follow the exact same proof. Fix $\eta>0$. Since $a_j\searrow 0$ as $j\to\infty$, there exists $N_1\in\mathbb{N}$ such that $\eta > a_j$ for $j\ge N_1$. Let $\varepsilon\in(0,1]$. It follows from Theorem \ref{thm:WSCEx}\ref{item:WSCEx6} that,
\begin{equation*}
\int_{0}^{T} \sup\limits_{r\ge \widetilde{r}(2\varepsilon,t)+\eta} \dfrac{|u_{a_j}|}{\sqrt{e_{a_j}}}(r,t)\, {\mathrm d} t \le \int_{0}^{T} \sup\limits_{r\ge \eta} \dfrac{|u_{a_j}|}{\sqrt{e_{a_j}}}(r,t)\, {\mathrm d} t \le C(T) \big(\eta^{\frac{2-n}{2}} + \eta^{2-n} \big) \qquad \text{for all } \ j\ge N_1.
\end{equation*}
By Lemma \ref{lemma:alimpath}\ref{item:apath1} and the fact that $\varepsilon\mapsto\widetilde{r}(\varepsilon,t)$ is increasing,
there exists $N_2\in\mathbb{N}$ such that $\widetilde{r}(2\varepsilon,t) \ge \widetilde{r}_{a_j}(\varepsilon,t) $ for all $j\ge N_2$.
Thus, by the uniform convergence in Lemma \ref{lemma:alimue} and Fatou's lemma, we have
\begin{equation*}
\int_{0}^{T} \sup\limits_{r\ge \widetilde{r}(2\varepsilon,t)+\eta} \dfrac{|u|}{\sqrt{e}}(r,t)\, {\mathrm d} t
\le C(T) \big( \eta^{\frac{2-n}{2}} + \eta^{2-n} \big).
\end{equation*}
Finally, since $\widetilde{r}(\varepsilon,t) \searrow \underline{r}(t)$ as $\varepsilon \searrow 0$, the desired estimate is obtained by an application of Monotone Convergence theorem with the limit $\varepsilon\searrow 0$.
\end{proof}
\subsection{Weak forms away from the vacuum}\label{subsec:alimWF}
In this subsection, we denote $(\rho,u,e)(r,t)$ on $\mathrm{F}$ and $\widetilde{r}(x,t)$ on $[0,\infty)\times[0,T]$
as the limit functions obtained in Lemmas \ref{lemma:alimpath}--\ref{lemma:alimrho}.
The main aim is to show that they satisfy the weak forms described in Definition \ref{def:WFAV}.
Moreover, we also define a class of test functions $\mathcal{D}_T$ as
\begin{equation*}
\mathcal{D}_T\vcentcolon
=\{ \phi\in C^{1}([0,\infty)\times[0,T])\,\vcentcolon\,\exists L\in\mathbb{N} \text{ such that } \phi(r,t)=0
\text{ for $(r,t)\in[L,\infty)\times[0,T]$} \}.
\end{equation*}
\begin{lemma}\label{lemma:alimID}
Let $(\rho_0,u_0,e_0)(r)$ be the spherically symmetric initial data in {\rm Theorem \ref{thm:WSAV}},
and let $(\rho_a^0,u_a^0 ,e_a^0)(r)$ be the modified initial data constructed in {\rm \S \ref{subsec:mollify}}.
Then, for all $\phi\in \mathcal{D}_T$,
\begin{equation*}
\begin{split}
&\lim\limits_{a\searrow 0} \int_{a}^{\infty} \rho_a^0(r) \phi(r,0)\,r^m {\mathrm d} r
= \int_{0}^{\infty} \rho_0(r) \phi (r,0)\,r^m {\mathrm d} r,\\
&\lim\limits_{a\searrow 0} \int_{a}^{\infty} (\rho_a^0 u_a^0) (r) \phi(r,0)\,r^m {\mathrm d} r
= \int_{0}^{\infty} (\rho_0 u_0)(r) \phi(r,0)\,r^m {\mathrm d} r,\\
&\lim\limits_{a\searrow 0} \int_{a}^{\infty} \Big(\frac{1}{2}\rho_a^0 \norm{u_a^0}^2 +\rho_a^0 e_a^0\Big)(r) \phi(r,0)\, r^m {\mathrm d} r
= \int_{0}^{\infty} \Big(\dfrac{1}{2}\rho_0\snorm{u_0}^2 + \rho_0 e_0 \Big)(r) \phi(r,0)\,r^m {\mathrm d} r.
\end{split}
\end{equation*}
\end{lemma}
\begin{proof}
By construction \eqref{eqs:amolint} in \S \ref{subsec:mollify},
$(\rho_a^0,u_a^0,e_a^0)(r)=(1,0,1)$ for $r\in[0,a]$. Then
we use Proposition \ref{prop:mollify} to obtain
\begin{align*}
&\Bignorm{\int_{a}^{\infty}\rho_a^0(r)\phi(r,0)\,r^m {\mathrm d} r
- \int_{0}^{\infty}\rho_0(r)\phi(r,0)\,r^m {\mathrm d} r}\\
&\quad\le \int_{0}^a \snorm{\phi(r,0)}\,r^m {\mathrm d} r
+\sbnorm{\phi(\cdot,0)}_{L^2([0,\infty),r^m {\mathrm d} r)}
\sbnorm{\rho_a^0 - \rho_0}_{L^2([0,\infty),r^m {\mathrm d} r)} \to 0 \qquad\text{ as $a\searrow 0$},\\[2mm]
&\Bignorm{\int_{a}^{\infty}(\rho_a^0 u_a^0)(r) \phi(r,0)\,r^m {\mathrm d} r
- \int_{0}^{\infty}(\rho_0 u_0)(r) \phi(r,0)\,r^m {\mathrm d} r}\\
&\quad\le \sbnorm{\rho_a^0}_{L^\infty} \sbnorm{\phi(\cdot,0)}_{L^2([0,\infty),r^m{\mathrm d} r)} \sbnorm{u_a^0-u_0}_{L^2([0,\infty),r^m{\mathrm d} r)} \\[1mm]
&\quad\quad\, + \sbnorm{\phi(\cdot,0)}_{L^\infty} \sbnorm{u_0}_{L^2([0,\infty),r^m{\mathrm d} r)} \sbnorm{\rho_a^0-\rho_0}_{L^2([0,\infty),r^m{\mathrm d} r)} \to 0 \qquad\text{ as $a\searrow 0$}.
\end{align*}
Furthermore, it follows from \eqref{eqs:amolint} and Proposition \ref{prop:mollify} that,
as $a\searrow 0$,
\begin{align*}
&\Bignorm{\int_{a}^{\infty} \rho_a^0 \snorm{u_a^0}^2(r) \phi(r,0)\,r^m {\mathrm d} r
- \int_{0}^{\infty} \rho_0 \snorm{u_0}^2(r) \phi(r,0)\,r^m {\mathrm d} r }\\
&\quad \le \sbnorm{\rho_a^0}_{L^{\infty}} \sbnorm{\phi(\cdot,t)}_{L^2([0,\infty),r^m {\mathrm d} r)} \sbnorm{\snorm{u_a^0}^2 - \snorm{u_0}^2}_{L^2([0,\infty),r^m {\mathrm d} r)}\\[1mm]
&\quad\quad + \sbnorm{\phi(\cdot,0)}_{L^{\infty}}\sbnorm{u_0}_{L^4([0,\infty),r^m{\mathrm d} r)}^2
\sbnorm{ \rho_a^0 - \rho_0}_{L^2([0,\infty),r^m {\mathrm d} r)}\to 0 \qquad\text{ as $a\searrow 0$}.\\[2mm]
&\Bignorm{\int_{a}^{\infty}\rho_a^0 e_a^0(r) \phi(r,0)\,r^m {\mathrm d} r
- \int_{0}^{\infty} \rho_0 e_0(r) \phi(r,0)\,r^m {\mathrm d} r}\\
&\quad\le \Bignorm{\int_{0}^{a}\phi(r,0)\,r^m{\mathrm d} r}
+ \Bignorm{\int_{0}^{\infty} \phi(r,0) \big(\rho_a^0 (e_a^0 - e_0) + e_0 (\rho_a^0-\rho_0)\big)\,r^m {\mathrm d} r }\\
&\quad\le \sbnorm{\phi(\cdot,0)}_{L^{\infty}}\frac{a^n}{n} + \sbnorm{\rho_a^0}_{L^{\infty}} \sbnorm{\phi(\cdot,0)}_{L^2([0,\infty),r^m {\mathrm d} r)} \sbnorm{e_a^0-e_0}_{L^2([0,\infty),r^m {\mathrm d} r)} \\[1mm]
&\quad\quad\, + \big(\sbnorm{\phi(\cdot,0)}_{L^{\infty}}\sbnorm{e_0-1}_{L^2([0,\infty),r^m{\mathrm d} r)} + \sbnorm{\phi(\cdot,0)}_{L^2([0,\infty),r^m{\mathrm d} r)}\big)
\sbnorm{\rho_a^0-\rho_0}_{L^2([0,\infty),r^m {\mathrm d} r)} \to 0.
\end{align*}
This completes the proof.
\end{proof}
\begin{lemma}\label{lemma:alimRhoWF}
For any $\phi\in \mathcal{D}_T$,
\begin{align*}
\int_{0}^{\infty} \rho(r,t) \phi(r,t)\,r^m {\mathrm d} r - \int_{0}^{\infty} \rho_0(r)\phi(r,0)\,r^m {\mathrm d} r
= \int_{0}^{t}\int_{0}^{\infty} \big(\rho \partial_t \phi + \rho u \partial_r\phi\big)(r,s)\,r^m {\mathrm d} r {\mathrm d} s.
\end{align*}
\end{lemma}
\begin{proof}
First, from the continuity equation in Theorem \ref{thm:WSCEx},
we obtain that, for all $j\in\mathbb{N}$ and $a_j\in(0,1)$,
\begin{equation}\label{temp:alimRhoWF1}
\int_{a_j}^{\infty} \rho_{a_j}(r,s) \phi(r,s)\,r^m {\mathrm d} r \Big\vert_{s=0}^{s=t}
= \int_{0}^{t}\int_{a_j}^{\infty} \big( \rho_{a_j} \partial_t \phi
+ \rho_{a_j} u_{a_j} \partial_r \phi\big)(r,t)\,r^m {\mathrm d} r {\mathrm d} t.
\end{equation}
Since $\rho\equiv 0$ in $\{ r < \underline{r}(t) \}$ by construction, it follows from Lemmas \ref{lemma:rhoM} and \ref{lemma:alimrho} that
\begin{align*}
\mathcal{I}_1^{j}\vcentcolon
=&\,\Bignorm{\int_{a_j}^{\infty} \rho_{a_j}(r,t) \phi(r,t)\,r^m {\mathrm d} r
- \int_{\underline{r}(t)}^{\infty} \rho(r,t) \phi(r,t)\,r^m {\mathrm d} r}\\
\le\,&\, \Bignorm{\int_{\widetilde{r}_{a_j}(x,t)}^{\infty}\rho_{a_j}(r,t)\phi(r,t)\,r^m{\mathrm d} r
-\int_{\widetilde{r}(x,t)}^{\infty}\rho(r,t)\phi(r,t)\,r^m{\mathrm d} r}
+ \sbnorm{\phi}_{\infty}\int_{\underline{r}(t)}^{\widetilde{r}(x,t)}\rho(r,t)\,r^m {\mathrm d} r \\
\,& +\sbnorm{\phi}_{\infty} \bigg( \int_{a_j}^{\underline{r}(t)} + \int_{\underline{r}(t)}^{\widetilde{r}(x,t)} + \int_{\widetilde{r}(x,t)}^{\widetilde{r}_{a_j}(x,t)} \bigg)\rho_{a_j}(r,t)\,r^m {\mathrm d} r\\
\le &\, \Bignorm{\int_{0}^{\infty}\big(\rho_{a_j}^{(x)}-\rho^{(x)}\big)(r,t) \phi(r,t)\, r^m{\mathrm d} r}
+ \sbnorm{\phi}_{\infty} x + \sbnorm{\phi}_{\infty} \int_{a_j}^{\underline{r}(t)}\rho_{a_j}(r,t)\,r^m {\mathrm d} r\\
\,&+ \sbnorm{\phi}_{\infty}\omega_1( [\underline{r}(t),\widetilde{r}(x,t)];C ) + \sbnorm{\phi}_{\infty}\omega_1([\widetilde{r}(x,t),\widetilde{r}_{a_j}(x,t)];C).
\end{align*}
Since $\widetilde{r}(x,t)\searrow \underline{r}(t)$ as $x\searrow 0$, for any $\delta>0$,
there exists $y\in(0,1)$ such that
\begin{equation*}
\sbnorm{\phi}_{\infty} x + \sbnorm{\phi}_{\infty} \omega_1( [\underline{r}(t),\widetilde{r}(x,t)];C )
\le \delta \qquad\ \text{ for all $0<x<y$.}
\end{equation*}
Then it follows that, for all $x\in (0,y)$,
\begin{equation*}
\mathcal{I}_1^{j}
\le \delta + \Bignorm{\int_{0}^{\infty}\!\!\phi(r,t)\big(\rho_{a_j}^{(x)}-\rho^{(x)}\big)(r,t)\,r^m{\mathrm d} r}
+ \sbnorm{\phi}_{\infty}\int_{a_j}^{\underline{r}(t)}\!\!\rho_{a_j}(r,t)\,r^m {\mathrm d} r +\sbnorm{\phi}_{\infty}\omega_1([\widetilde{r}(x,t),\widetilde{r}_{a_j}(x,t)];C).
\end{equation*}
Fix a point $x\in(0,y)$. Taking the limit $j\to\infty$ on the above inequality, then using
Lemmas \ref{lemma:alimrho}\ref{item:alimrho1} and \ref{lemma:alimpath}\ref{item:apath5},
we have
\begin{align*}
\lim\limits_{j\to\infty}\Bignorm{\int_{a_j}^{\infty} \rho_{a_j}(r,t) \phi(r,t)\,r^m {\mathrm d} r - \int_{\underline{r}(t)}^{\infty} \rho(r,t) \phi(r,t)\,r^m {\mathrm d} r} \le \delta,
\end{align*}
which implies that $\lim_{j\to\infty}\mathcal{I}_1^{j}=0$.
Next, for each $x>0$, we rewrite the term:
\begin{align*}
\mathcal{I}_{2}^{j}
&\vcentcolon=\Bignorm{\int_{0}^{t}\int_{a_j}^{\infty} \rho_{a_j} u_{a_j} \partial_r \phi(r,s)\,r^m {\mathrm d} r {\mathrm d} s
- \int_{0}^{t}\int_{\underline{r}(s)}^{\infty} \rho u\,\partial_r \phi(r,s)\,r^m {\mathrm d} r {\mathrm d} s}\\
&\,\le \sbnorm{\partial_r\phi}_{\infty} \int_{0}^{t}\int_{a_j}^{\widetilde{r}_{a_j}(x,s)}
\snorm{\rho_{a_j}u_{a_j}} \,r^m {\mathrm d} r {\mathrm d} s
+ \sbnorm{\partial_r\phi}_{\infty} \int_{0}^{t} \int_{\underline{r}(s)}^{\widetilde{r}(x,s)}
\snorm{\rho u}\,r^m {\mathrm d} r {\mathrm d} s \\
&\,\quad +\Bignorm{\int_{0}^{t}\int_{\widetilde{r}_{a_j}(x,s)}^{\infty}\rho_{a_j}u_{a_j}
\partial_r\phi\,r^m{\mathrm d} r{\mathrm d} s
- \int_{0}^{t}\int_{\widetilde{r}(x,s)}^{\infty}\rho u\partial_r \phi\,r^m{\mathrm d} r {\mathrm d} s}\\
& =\vcentcolon \mathcal{I}_{2,1}^{j}(x)+\mathcal{I}_{2,2}^{j}(x)+\mathcal{I}_{2,3}^{j}(x).
\end{align*}
Using Lemma \ref{lemma:rhoM}, the Cauchy-Schwartz inequality, and
Theorem \ref{thm:WSCEx}\ref{item:WSCEx3}, we have
\begin{align*}
\mathcal{I}_{2,1}^j(x)
\le \sbnorm{\partial_r\phi}_{\infty}\sqrt{x} \int_{0}^{t}\Big(\int_{a_j}^{\widetilde{r}_{a_j}(x,t)} \rho_{a_j}u_{a_j}^2\,r^m {\mathrm d} r\Big)^{\frac{1}{2}} {\mathrm d} s
\le \sbnorm{\partial_r\phi}_{\infty} C(T) t \sqrt{x}.
\end{align*}
Moreover, using Lemmas \ref{lemma:alimrho}\ref{item:alimrho3} and \ref{lemma:alimEnt} leads to
\begin{align*}
\mathcal{I}_{2,2}^{j}(x)
\le \sbnorm{\partial_r\phi}_{\infty}\int_{0}^{t} \Big( \int_{\underline{r}(s)}^{\widetilde{r}(x,s)}
\rho \snorm{u}^2\,r^m {\mathrm d} r \Big)^{\frac{1}{2}}
\Big( \int_{\underline{r}(s)}^{\widetilde{r}(x,s)} \rho\,r^m {\mathrm d} r \Big)^{\frac{1}{2}} {\mathrm d} s
\le\sbnorm{\partial_r\phi}_{\infty} C(T) t \sqrt{x}.
\end{align*}
Thus, for any $\delta>0$, there exists $y_\delta\in(0,1)$ such that, if $x\in (0, y_{\delta})$,
then $\mathcal{I}_{2}^{j}(x)\le \delta + \mathcal{I}_{2,3}^{j}(x)$.
By Lemma \ref{lemma:alimpath}, $x\to \widetilde{r}(x,t)$ is increasing,
and $(x,t)\mapsto\widetilde{r}(x,t)$ is continuous. Since $[0,T]$ is compact,
$d:=\inf_{t\in[0,T]}\{\widetilde{r}(x,t) - \widetilde{r}(\frac{x}{2},t)\}>0$.
By Lemma \ref{lemma:alimpath}\ref{item:apath1}, there exists $N_x\in\mathbb{N}$
such that $\sup_{ 0\le t \le T} \snorm{\widetilde{r}_{a_j}(x,t)-\widetilde{r}(x,t)}\le \frac{d}{2}$ if $j\ge N_x$.
Combining with the previous assertion, we have
\begin{equation}\label{temp:alimRhoWF2}
\widetilde{r}(\frac{x}{2},t) < \widetilde{r}_{a_j}(x,t) \,\qquad \text{ for each $t\in[0,T]$ and $j\ge N_x$.}
\end{equation}
Then $\chi_x^{a_j}=\chi_{\frac{x}{2}}\chi_x^{a_j}$ for all $j\ge N_x$,
where $\chi_\varepsilon$ and $\chi_\varepsilon^{a}$ are defined in \eqref{6.6indicator}.
Thus, for all $j\ge N_x$,
\begin{align*}
\mathcal{I}_{2,3}^{j}(x)
\le \Bignorm{\int_{0}^{t}\int_{\widetilde{r}(\frac{x}{2},s)}^{\infty}\rho_{a_j}^{(x)}(u_{a_j} - u)\partial_r
\phi\,r^m{\mathrm d} r {\mathrm d} s }
+ \Bignorm{\int_{0}^{t}\int_{\widetilde{r}(\frac{x}{2},s)}^{\infty}\big(\rho_{a_j}^{(x)} - \rho^{(x)}\big) u
\partial_r \phi\,r^m{\mathrm d} r {\mathrm d} s }.
\end{align*}
By the pointwise bound: $ \rho_a^{(x)}(r,t)\le C(x)$ from Theorem \ref{thm:WSCEx}\ref{item:WSCEx4},
and the uniform convergence from Lemma \ref{lemma:alimue}, it follows that, as $j\to \infty$,
\begin{align*}
\Bignorm{\int_{0}^{t}\int_{\widetilde{r}(\frac{x}{2},s)}^{\infty}\rho_{a_j}^{(x)}(u_{a_j} - u)
\partial_r \phi\, r^m{\mathrm d} r {\mathrm d} s }
\le C(x)\int_{0}^{t}\int_{\widetilde{r}(\frac{x}{2},s)}^{\infty}\snorm{u_{a_j} - u}
\partial_r \phi\,r^m{\mathrm d} r {\mathrm d} s \to 0.
\end{align*}
By Lemma \ref{lemma:alimue}, $\chi_{\frac{x}{2}} u\partial_r \phi \in L^{1}(0,T;L^{2}([0,\infty),r^m {\mathrm d} r))$.
Thus, by Lemma \ref{lemma:alimrho}\ref{item:alimrho1},
\begin{align*}
\Bignorm{\int_{0}^{t}\int_{\widetilde{r}(\frac{x}{2},s)}^{\infty}\big(\rho_{a_j}^{(x)} - \rho^{(x)}\big)u
\partial_r \phi\,r^m{\mathrm d} r {\mathrm d} s } \to 0 \qquad\ \text{as $j\to \infty$},
\end{align*}
which implies that $\lim_{j\to\infty}\mathcal{I}_{2,3}^{j}(x)=0$ for each $x>0$.
Now, for any given $\delta>0$, we fix $x\in(0,y_\delta)$. Then $\mathcal{I}_{2}^j=\mathcal{I}_{2,1}^j(x)+\mathcal{I}_{2,2}^j(x)+\mathcal{I}_{2,3}^j(x)
\le \delta + \mathcal{I}_{2,3}^j(x)$.
Taking limit $j\to\infty$ on both sides of this,
we see that $\lim_{j\to\infty} \mathcal{I}_2^{j}\le \delta$.
Since $\delta>0$ is arbitrarily small, we conclude that $\lim_{j\to\infty} \mathcal{I}_2^{j} = 0$.
By the
same argument, we can also show that
\begin{align*}
\mathcal{I}_{3}^{j}
\vcentcolon= \Bignorm{\int_{0}^{t}\int_{a_j}^{\infty}\rho_{a_j}\partial_t \phi\, r^m {\mathrm d} r {\mathrm d} s - \int_{0}^{t}\int_{\underline{r}(s)}^{\infty}\rho \partial_t \phi\, r^m {\mathrm d} r {\mathrm d} s} \to 0 \qquad\ \text{ as $j\to\infty$.}
\end{align*}
Lemma \ref{lemma:alimRhoWF} is proved by substituting these limits in \eqref{temp:alimRhoWF1}
and using Lemma \ref{lemma:alimID}.
\end{proof}
\begin{lemma}\label{lemma:eWF}
Let $\phi\in \mathcal{D}_T\cap C^{1}$ be such that ${\mathrm {supp}}(\phi)\Subset \mathrm{F}$.
Then
\begin{align}
&\int_{0}^{\infty} (\rho u)(r,t) \varphi(r,t)\,r^m {\mathrm d} r
- \int_{0}^{\infty} \rho_0(r) u_0(r) \varphi(r,0)\,r^m {\mathrm d} r\nonumber\\
&\quad = \int_{0}^{t}\int_{0}^{\infty}\Big\{ \rho u \partial_t \varphi + \rho u^2 \partial_r \varphi + \big( P- \beta (\partial_r u + m \dfrac{u}{r})\big)
\big(\partial_r \varphi + m \dfrac{\varphi}{r}\big) \Big\}\,r^m {\mathrm d} r {\mathrm d} s,\label{eqs:WFAVmom}\\[2mm]
&\int_{0}^{\infty} (\rho E)(r,t) \phi (r,t)\, r^m {\mathrm d} r
- \int_{0}^{\infty} (\rho_0 E_0) (r) \phi(r,0)\,r^m {\mathrm d} r
- \int_{0}^{t}\int_{0}^{\infty} \rho E \partial_t \phi\, r^m {\mathrm d} r {\mathrm d} s\nonumber\\
&\quad =\int_{0}^{t} \int_{0}^{\infty} \Big\{ (\rho E + P ) u - 2\mu u \partial_r u
- \lambda u \big(\partial_r u + m \dfrac{u}{r}\big) - \kappa \partial_r e \Big\}
\partial_r \phi\, r^m {\mathrm d} r {\mathrm d} s. \label{eqs:WFAVener}
\end{align}
\end{lemma}
\begin{proof}
We give the proof only for \eqref{eqs:WFAVener}, since \eqref{eqs:WFAVmom} can be derived in the same way.
First, from the energy equation in Theorem \ref{thm:WSCEx}, we have the weak form:
\begin{equation}\label{temp:eWF1}
\begin{aligned}
&\int_{a_j}^{\infty} (\rho_{a_j} E_{a_j})(r,t) \phi (r,t)\, r^m {\mathrm d} r
- \int_{a_j}^{\infty} \Big(\dfrac{1}{2}\rho_{a_j}^0\snorm{u_{a_j}^0}^2 +\rho_{a_j}^0 e_{a_j}^0 \Big)(r)
\phi(r,0) \,r^m {\mathrm d} r\\
&= -\int_{0}^{t} \int_{a_j}^{\infty} \Big\{ 2\mu u_{a_j} \partial_r u_{a_j} +\lambda u_{a_j} \big( \partial_r u_{a_j} + m \dfrac{u_{a_j}}{r} \big) + \kappa \partial_r e_{a_j} \Big\} \partial_r\phi\, r^m {\mathrm d} r {\mathrm d} s\\
&\quad\, + \int_{0}^{t}\int_{a_{j}}^{\infty} \big\{ \rho_{a_j} E_{a_j} \partial_t \phi + (\rho_{a_j} E_{a_j} + P_{a_j}) u_{a_j} \partial_r \phi \big\}\, r^m {\mathrm d} r {\mathrm d} s,
\end{aligned}
\end{equation}
where $P_{a_j}\vcentcolon= (\gamma-1) \rho_{a_j} e_{a_j}$ and
$E_{a_j}\vcentcolon= \frac{1}{2}\snorm{u_{a_j}}^2 + e_{a_j}$.
Now, for each $x\in(0,1)$, we define
$$
\widetilde{\mathrm{F}}_{x}\vcentcolon= \{ (r,t)\,\vcentcolon\, t\in[0,T], \ \widetilde{r}(x,t) < r <x^{-1} \}.
$$
Then $\widetilde{\mathrm{F}}_{x}\cap\{0<t<T\}$ is open for each $x\in(0,1)$,
since $(x,t)\mapsto\widetilde{r}(x,t)$ is continuous.
Moreover, since $x\mapsto \widetilde{r}(x,t)$ is strictly increasing for each $t\in[0,T]$,
it follows that $\widetilde{\mathrm{F}}_{x}\subsetneq \widetilde{\mathrm{F}}_{y}$ if $y < x \in (0,1)$.
Using
$\widetilde{r}(x,t)\searrow \underline{r}(t)$ as $x\searrow 0$ for all $t\in[0,T]$,
we see that $\mathrm{F} = \cup_{x\in(0,1)} \widetilde{\mathrm{F}}_{x}$, so that $\big\{\widetilde{\mathrm{F}}_x\cap\{0<t<T\}\big\}_{x\in(0,1)}$ is an open
covering of $\mathrm{F}\cap\{0<t<T\}$.
Since ${\mathrm {supp}}(\phi) \Subset
\mathrm{F}$ is a compact subset,
there exists $\varepsilon>0$ such that ${\mathrm {supp}}(\phi) \subset \widetilde{\mathrm{F}}_{2\varepsilon}$.
Defining $d:=\inf_{t\in[0,T]}\widetilde{r}(2\varepsilon,t)$, then $d>0$
due to Lemma \ref{lemma:alimpath}\ref{item:apath4},
and there exists $N_1\in\mathbb{N}$ such that $a_j < d$ for $j\ge N_1$.
This implies that $a_j < \widetilde{r}(2\varepsilon,t)$ if $t\in[0,T]$ and $j\ge N_1$.
In addition, by the same argument for \eqref{temp:alimRhoWF2} as in the proof of Lemma \ref{lemma:alimRhoWF},
there exists $N_2\in\mathbb{N}$ such that $\widetilde{r}_{a_j}(\varepsilon,t)< \widetilde{r}(2\varepsilon,t)$ if $j\ge N_2$
and $t\in[0,T]$.
Let $N\vcentcolon= \max\{N_1,N_2\}$. Then, by Lemmas \ref{lemma:alimue}--\ref{lemma:alimrho}
and the dominated convergence theorem, we have
\begin{align*}
&\Bignorm{\int_{a_j}^{\infty}(\rho_{a_j} E_{a_j})(r,t) \phi(r,t)\, r^m {\mathrm d} r
- \int_{\underline{r}(t)}^{\infty} (\rho E)(r,t) \phi(r,t)\,r^m {\mathrm d} r }\\
&\le \int_{\widetilde{r}(2\varepsilon,t)}^{\infty} \rho_{a_j}^{(\varepsilon)}
\Big(\frac{1}{2}\snorm{u_{a_j}^2 - u^2} + \snorm{e_{a_j}-e} \Big)\phi\, r^m {\mathrm d} r + \Bignorm{\int_{\widetilde{r}(2\varepsilon,t)}^{\infty} \big(\rho_{a_j}^{(\varepsilon)}-\rho^{(\varepsilon)}\big)
\big( \dfrac{1}{2}u^2 + e \big) \phi\, r^m {\mathrm d} r} \\[2mm]
&\longrightarrow 0 \qquad\,\, \mbox{as $j\to\infty$}.
\end{align*}
By the same argument, we also obtain
\begin{align*}
\lim\limits_{j\to\infty} \int_{0}^{t}\int_{a_{j}}^{\infty} \big\{ \rho_{a_j} E_{a_j} \partial_t \phi +(\rho_{a_j} E_{a_j} + P_{a_j} ) u_{a_j} \partial_r \phi \big\}\,r^m {\mathrm d} r {\mathrm d} s
= \int_{0}^{t}\int_{\underline{r}(s)}^{\infty}
\big\{ \rho E \partial_t \phi +(\rho E + P ) u \partial_r \phi\big\}\, r^m {\mathrm d} r {\mathrm d} s.
\end{align*}
Next, integrating by parts and using Lemma \ref{lemma:alimue}, we obtain that, for all $j\ge N$,
\begin{align*}
\int_{0}^{t}\int_{a_j}^{\infty} u_{a_j} \partial_r u_{a_j} \partial_r \phi\, r^m {\mathrm d} r {\mathrm d} s
= -\int_{0}^{t}\int_{\widetilde{r}(2\varepsilon,s)}^{\infty}\dfrac{1}{2}\snorm{u_{a_j}}^2\partial_r^2\phi\, r^m{\mathrm d} r{\mathrm d} s
\to -\int_{0}^{t}\int_{\widetilde{r}(2\varepsilon,s)}^{\infty}
\dfrac{1}{2} \snorm{u}^2 \partial_r^2 \phi\, r^m {\mathrm d} r {\mathrm d} s,
\end{align*}
as $j\to\infty$.
From
\eqref{6.13b}, the spatial weak derivative of $u$ exists with
$\partial_r u \in L^2_{\text{loc}}(\mathrm{F})$ and $\partial_r u \in L^2(\widetilde{\mathrm{F}}_\varepsilon)$ for each $\varepsilon\in(0,1)$.
It then follows that, as $j\to\infty$,
\begin{align*}
\int_{0}^{t}\int_{a_j}^{\infty} u_{a_j} \partial_r u_{a_j} \partial_r \phi\, r^m {\mathrm d} r {\mathrm d} s
\to
-\int_{0}^{t}\int_{\widetilde{r}(2\varepsilon,s)}^{\infty} \dfrac{1}{2} \norm{u}^2 \partial_r^2 \phi\,r^m {\mathrm d} r {\mathrm d} s
= \int_{0}^{t}\int_{\underline{r}(s)}^{\infty} u \partial_r u \partial_r \phi\, r^m {\mathrm d} r {\mathrm d} s,
\end{align*}
where we have used ${\mathrm {supp}}(\phi) \Subset \mathrm{F}$.
By the same argument, we also have
\begin{align*}
\lim\limits_{j\to\infty}\int_{0}^{t} \int_{a_j}^{\infty}
\Big\{m\lambda \dfrac{\snorm{u_{a_j}}^2}{r} +\kappa\partial_r e_{a_j}\Big\}\partial_r\phi\, r^m {\mathrm d} r {\mathrm d} s
= \int_{0}^{t} \int_{\underline{r}(s)}^{\infty} \Big\{ m\lambda \dfrac{\snorm{u}^2}{r} + \kappa \partial_r e \Big\} \partial_r\phi\,r^m {\mathrm d} r {\mathrm d} s.
\end{align*}
Combining all the limits established above with \eqref{temp:eWF1}, we obtain Lemma \ref{lemma:eWF}.
\end{proof}
\smallskip
Finally, the weak forms obtained in Theorem \ref{lemma:alimRhoWF}--\ref{lemma:eWF}
are translated into the Eulerian coordinates $({\boldsymbol{x}},t)\in {\mathbb R}^n \times[0,T]$.
Let $\Phi \in C^1([0,T]; C_c^1 ({\mathbb R}^n))$.
Define
\begin{align*}
\rho({\boldsymbol{x}},t)\vcentcolon= \rho(\snorm{{\boldsymbol{x}}},t), \quad \u({\boldsymbol{x}},t)\vcentcolon= u(\snorm{{\boldsymbol{x}}},t)\frac{{\boldsymbol{x}}}{r},
\qquad\, \phi(r,t) \vcentcolon= \int_{S^{n-1}} \Phi( r {\boldsymbol{y}}, t ) {\mathrm d} S_{{\boldsymbol{y}}},
\end{align*}
where $S^{n-1}$ is the $(n-1)$-dimensional sphere,
and ${\mathrm d} S_{{\boldsymbol{y}}}$ is the surface measure on $S^{n-1}$.
Using this, we have
\begin{align*}
\int_{0}^{t}\int_{0}^{\infty} (\rho u)(r,s)\partial_r \phi (r,s)\, r^m {\mathrm d} r {\mathrm d} s
&= \int_{0}^{t}\int_{0}^{\infty} \int_{S^{n-1}} (\rho u)(r,s){\boldsymbol{y}}\cdot\nabla\Phi(r{\boldsymbol{y}},s)\,r^m{\mathrm d} S_{{\boldsymbol{y}}}{\mathrm d} r{\mathrm d} s \\
&= \int_{0}^{t}\int_{{\mathbb R}^n} (\rho u)(\snorm{{\boldsymbol{x}}},s) \dfrac{{\boldsymbol{x}}}{\snorm{{\boldsymbol{x}}}} \cdot\nabla\Phi( {\boldsymbol{x}} ,s)\, {\mathrm d} {\boldsymbol{x}} {\mathrm d} s\\
&= \int_{0}^{t} \int_{{\mathbb R}^n} \rho({\boldsymbol{x}},s) \u({\boldsymbol{x}},s)\cdot \nabla \Phi({\boldsymbol{x}},s)\,{\mathrm d} {\boldsymbol{x}} {\mathrm d} s.
\end{align*}
By the same way, we also have
\begin{align*}
&\int_{0}^{\infty} \rho(r,s)\phi(r,s)\,r^m {\mathrm d} r \Big\vert_{s=0}^{s=t}
= \int_{{\mathbb R}^n} \rho({\boldsymbol{x}},s) \Phi({\boldsymbol{x}},s)\,{\mathrm d} {\boldsymbol{x}} - \int_{\mathbb{R}^n} \rho_0({\boldsymbol{x}}) \Phi({\boldsymbol{x}},0)\,{\mathrm d} {\boldsymbol{x}},\\[2mm]
& \int_{0}^{t}\int_{0}^{\infty}\rho(r,s) \partial_t \phi(r,s)\,r^m {\mathrm d} r {\mathrm d} s = \int_{0}^{t}\int_{{\mathbb R}^n} \rho({\boldsymbol{x}},s) \partial_t \Phi({\boldsymbol{x}},s)\, {\mathrm d} {\boldsymbol{x}} {\mathrm d} s.
\end{align*}
Since $\phi(r,t)$ satisfies the assumption in Lemma \ref{lemma:alimRhoWF},
using the above identities in Lemma \ref{lemma:alimRhoWF},
we obtain the weak form of the continuity equation.
Next, we show the momentum equations.
Let $\Psi\vcentcolon {\mathbb R}^n\times[0,T] \to \mathbb{R}$ be such that
$\Psi\in C^2([0,T];C^2_c({\mathbb R}^n))$ and
${\mathrm {supp}}(\Psi)\Subset\mathcal{F}\vcentcolon
= \{({\boldsymbol{x}},t)\in{\mathbb R}^n\times[0,T]\,\vcentcolon\,\underline{r}(t)<\snorm{{\boldsymbol{x}}}\}$. We define
\begin{equation}\label{temp:S2N1}
\varphi^i(r,t) \vcentcolon= \int_{S^{n-1}} \Psi(r{\boldsymbol{y}},t) y^i\, {\mathrm d} S_{{\boldsymbol{y}}} \quad\,\,\,\text{for $i=1,\dotsc,n$.}
\end{equation}
It can be verified that $\varphi^i$ satisfies the assumption for Lemma \ref{lemma:eWF},
hence \eqref{eqs:WFAVmom} holds with $\varphi^i$.
The derivations for the weak forms in Theorem \ref{thm:WSAV} are similar to that occurring
in the continuity equation, except for the term:
$\iint\{P-\beta(\partial_r u + m \frac{u}{r})\}(\partial_r\varphi^i + m \frac{\varphi^i}{r})\,r^m {\mathrm d} r {\mathrm d} s$.
To deal with this, we use the following identity that can be verified with few lines of calculation
on the spherical coordinate transformation:
\begin{equation*}
\int_{\snorm{{\boldsymbol{x}}}\le r} \partial_{x^i} f ({\boldsymbol{x}})\,{\mathrm d} {\boldsymbol{x}}
= \int_{\snorm{{\boldsymbol{x}}}=r} f ({\boldsymbol{x}}) \dfrac{x^i}{\snorm{{\boldsymbol{x}}}}\,{\mathrm d} S_{{\boldsymbol{x}}}
\qquad \ \text{for $i=1,\dotsc, n$, for $r>0$.}
\end{equation*}
Using this identity, we have
\begin{align*}
\partial_r ( r^m \varphi^i )
&=\partial_r \Big( \int_{S^{n-1}} y^i \Psi(r {\boldsymbol{y}} ,t)\, r^m {\mathrm d} S_{{\boldsymbol{y}}} \Big)
=\partial_r \Big( \int_{\snorm{{\boldsymbol{x}}}=r} \dfrac{x^i}{\snorm{{\boldsymbol{x}}}} \Psi( {\boldsymbol{x}} ,t)\,{\mathrm d} S_{{\boldsymbol{x}}} \Big)\\
&= \partial_r \Big( \int_{0}^{r} \int_{S^{n-1}} \partial_{x^i} \Psi (\zeta{\boldsymbol{y}},t)\,\zeta^m {\mathrm d} S_{{\boldsymbol{y}}}{\mathrm d}\zeta \Big)
= r^m \int_{S^{n-1}} \partial_{x^i} \Psi ( r {\boldsymbol{y}} ,t)\,{\mathrm d} S_{{\boldsymbol{y}}}.
\end{align*}
Take the derivative on \eqref{temp:S2N1} to obtain that $\partial_r \varphi^i(r,t) = \int_{S^{n-1}} y^i y^j\partial_{{\boldsymbol{x}}^j}\Psi(r{\boldsymbol{y}},t) {\mathrm d} S_{{\boldsymbol{y}}}$ so that
\begin{align*}
\int_{0}^{t}\int_{0}^{\infty} P \big(\partial_r \varphi^i + \dfrac{m}{r} \varphi^i \big)\,r^m{\mathrm d} r{\mathrm d} s
= \int_{0}^{t}\int_{0}^{\infty} \int_{S^{n-1}} P \partial_{x^i} \Psi (r{\boldsymbol{y}},s)\,r^m {\mathrm d} S_{{\boldsymbol{y}}}{\mathrm d} r{\mathrm d} s
= \int_{0}^{t}\int_{{\mathbb R}^n} P \partial_{x^i} \Psi({\boldsymbol{x}},s)\, {\mathrm d} {\boldsymbol{x}} {\mathrm d} s.
\end{align*}
Moreover, integrating by parts yields
\begin{equation*}
\begin{aligned}
&\int_{0}^T\int_{0}^{\infty} \big( \partial_r u + m \dfrac{u}{r} \big)\big(\partial_r \varphi^i+ m \dfrac{\varphi^i}{r}\big)\,r^m {\mathrm d} r {\mathrm d} t
= \int_{0}^T\int_{0}^{\infty}
\big(\partial_r u \partial_r \varphi^i + m \dfrac{\varphi^i u}{r^2} \big)\,r^m {\mathrm d} r {\mathrm d} t.
\end{aligned}
\end{equation*}
Using the identity $\partial_{x^j} \u^i = \partial_r(\frac{u}{\snorm{{\boldsymbol{x}}}})\frac{x^i x^j}{\snorm{{\boldsymbol{x}}}} + \frac{u}{\snorm{{\boldsymbol{x}}}}\delta^{ij}$,
we have
\begin{equation*}
\begin{aligned}
\int_{0}^{T}\int_{{\mathbb R}^n} \partial_{x^j} \u^i \partial_{x^j} \Psi {\mathrm d} {\boldsymbol{x}} {\mathrm d} t
&= \int_{0}^{T}\int_{{\mathbb R}^n} \Big( \partial_r\big(\dfrac{u}{\snorm{{\boldsymbol{x}}}}\big)\dfrac{x^i x^j}{\snorm{{\boldsymbol{x}}}} + \dfrac{u}{\snorm{{\boldsymbol{x}}}}\delta^{ij} \Big) \partial_{x^j} \Psi {\mathrm d} {\boldsymbol{x}} {\mathrm d} t\\
&= \int_{0}^{T}\int_{0}^{\infty}\int_{\mathbb{S}^{n-1}} \Big(r\partial_r\big(\dfrac{u}{r}\big)y^i y^j \partial_{x^j}\Psi(r{\boldsymbol{y}},t) + \dfrac{u}{r}\partial_{x^i}\Psi(r{\boldsymbol{y}},t) \Big)\,{\mathrm d} S_{{\boldsymbol{y}}}\,r^m {\mathrm d} r {\mathrm d} t\\
&= \int_{0}^{T}\int_{0}^{\infty} \Big(r^{m+1} \partial_r\big(\dfrac{u}{r}\big) \partial_r \varphi^i + \dfrac{u}{r}\partial_r(r^m \varphi^i) \Big)\,{\mathrm d} r {\mathrm d} t\\
&=\int_{0}^T\int_{0}^{\infty} \big( \partial_r u + m \dfrac{u}{r} \big)\big(\partial_r \varphi^i+ m \dfrac{\varphi^i}{r}\big)\,r^m {\mathrm d} r {\mathrm d} t.
\end{aligned}
\end{equation*}
Similarly, since ${\mathrm {div}}\,\u = \partial_r u + m \frac{u}{r}$, we also have
\begin{equation*}
\begin{aligned}
\int_{0}^{T}\int_{{\mathbb R}^n} {\mathrm {div}} \u \partial_{x^i} \Psi\,{\mathrm d} {\boldsymbol{x}} {\mathrm d} t
= \int_{0}^T\int_{0}^{\infty} \big( \partial_r u + m \dfrac{u}{r} \big)\big(\partial_r \varphi^i+ m \dfrac{\varphi^i}{r}\big)\,r^m {\mathrm d} r {\mathrm d} t.
\end{aligned}
\end{equation*}
Finally, we show the weak form of energy equation. Let $\varPhi\in C^{2}([0,T];C^2_{\rm c}({\mathbb R}^n))$
be a test function such that
${\mathrm {supp}}(\Phi)\Subset \mathcal{F}=\{ ({\boldsymbol{x}},t)\in {\mathbb R}^n\times[0,T]\,\vcentcolon\,\snorm{{\boldsymbol{x}}}>\underline{r}(t) \}$.
Define
\begin{align*}
\phi(r,t) \vcentcolon= \int_{S^{n-1}} \Phi(r{\boldsymbol{y}},t)\,{\mathrm d} S_{{\boldsymbol{y}}}.
\end{align*}
It can be verified that $\phi$ satisfies the assumption of Lemma \ref{lemma:eWF}, so that
\eqref{eqs:WFAVener} holds with $\phi$.
Most of the terms can be derived by the same procedure presented previously,
except for the terms: $u \partial_r u$, $\frac{u^2}{r}$, and $\partial_r e$.
Since ${\mathrm {supp}}(\phi) \Subset \{(r,t)\,\vcentcolon\, r>\underline{r}(t)\}$, it follows that
\begin{align*}
\int_{0}^{t}\int_{\underline{r}(s)}^{\infty} u \partial_r u \partial_r \phi (r,s)\, r^m {\mathrm d} r {\mathrm d} s
&= \int_{0}^{t}\int_{\underline{r}(s)}^{\infty} u \partial_r u
\Big( \int_{S^{n-1}} y^i \partial_{x^i} \Phi(r{\boldsymbol{y}},s) {\mathrm d} S_{{\boldsymbol{y}}} \Big)\,r^m {\mathrm d} r {\mathrm d} s\\
&= \int_{0}^{t}\int_{\snorm{{\boldsymbol{x}}}>\underline{r}(s)} \partial_r \big(\dfrac{u^2}{2}\big) (\snorm{{\boldsymbol{x}}},s) \dfrac{x^i}{\snorm{{\boldsymbol{x}}}} \partial_{x^i} \Phi({\boldsymbol{x}},s)\, {\mathrm d} {\boldsymbol{x}} {\mathrm d} s\\
&=\int_{0}^{t}\int_{\snorm{{\boldsymbol{x}}}>\underline{r}(s)} \partial_{x^i} \big(\dfrac{\snorm{\u}^2}{2}\big)
\partial_{x^i} \Phi({\boldsymbol{x}},s)\,{\mathrm d} {\boldsymbol{x}} {\mathrm d} s\\
&= \dfrac{1}{2} \int_{0}^{t}\int_{\snorm{{\boldsymbol{x}}}>\underline{r}(s)} \nabla \Phi({\boldsymbol{x}},s) \cdot \nabla \snorm{\u}^2\,
{\mathrm d} {\boldsymbol{x}} {\mathrm d} s,
\end{align*}
and with several lines of derivations, we also have
\begin{align*}
&\int_{0}^{t}\int_{\underline{r}(s)}^{\infty}\big\{ \mu u \partial_r u + \lambda u\big( \partial_r u + m \dfrac{u}{r} \big) \big\} \partial_r \phi (r,s)\,r^m {\mathrm d} r {\mathrm d} s\\
&= \int_{0}^{t}\int_{\underline{r}(s)}^{\infty} \int_{S^{n-1}} \big\{ \mu u \partial_r u + \lambda u\big( \partial_r u + m \dfrac{u}{r} \big) \big\} y^i \partial_{x^i} \Phi(r{\boldsymbol{y}},s)\,r^m {\mathrm d} S_{{\boldsymbol{y}}} {\mathrm d} r {\mathrm d} s\\
&= \int_{0}^{t} \int_{\snorm{{\boldsymbol{x}}}>\underline{r}(s)}
\{ \lambda ({\mathrm {div}} \u) \u + \mu (\nabla \u) \cdot \u \} \nabla \Phi ({\boldsymbol{x}},s)\,{\mathrm d} {\boldsymbol{x}} {\mathrm d} s,
\end{align*}
\begin{align*}
\int_{0}^{t}\int_{\underline{r}(s)}^{\infty} \partial_r e\, \partial_r \phi (r,s)\,r^m {\mathrm d} r {\mathrm d} s
&= \int_{0}^{t}\int_{\underline{r}(s)}^{\infty}\int_{S^{n-1}}
\partial_r e (r,s) y^i \partial_{x^i} \Phi (r{\boldsymbol{y}},s)\, r^m {\mathrm d} S_{{\boldsymbol{y}}} {\mathrm d} r {\mathrm d} s \\
&= \int_{0}^{t}\int_{\snorm{{\boldsymbol{x}}}>\underline{r}(s)} \nabla e \cdot \nabla \Phi\, {\mathrm d} x {\mathrm d} s.
\end{align*}
With these, we complete the proof of the weak form of the energy equation.
|
\section{INTRODUCTION}
\label{sec:intro}
The Cosmology Large Angular Scale Surveyor (CLASS) is designed to measure the polarization of the CMB at large angular scales~\cite{essi14,harr16} enabled by rapid front-end polarization modulation~\cite{harr18} and a scanning strategy allowing adequate cross-linking~\cite{mill16}. Observing over $75\%$ of the sky in multiple frequency bands centered near 40, 90, 150, and 220~GHz from a high altitude site in the Atacama desert in Chile, CLASS will be sensitive to both the recombination and reionization~\cite{kami16} signatures in the B-mode signal associated with primordial gravitational waves from inflation. The large angular scale E-mode measurement will improve constraints on the optical depth to reionization~\cite{2018_tau_Duncan}, which will break degeneracies in cosmological parameters, specifically on measurements from small-angular-scale CMB polarization of the sum of neutrino masses~\cite{Neutrino_Allison}. The 40~GHz telescope~\cite{2014_40GHz_Detector_John,appe19} has been observing since June 2016, while the 90~GHz telescope achieved first light in May 2018~\cite{daha18} and is currently being upgraded. A third telescope housing a dichroic receiver sensitive to both 150 and 220~GHz frequency bands~\cite{daha19} (henceforth referred to as HF) began observations in September 2019. Another telescope at 90~GHz is planned.
\begin{figure*}
\centering
\includegraphics[width=0.9\linewidth]{figures/SPIE_W1_Model_optics_lowRes.jpg}
\caption{Ray-trace schematic of the 90~GHz telescope showing the major optical components. Color-coded rays from a selection of feedhorns illustrate the marginal light path and field-of-view of the telescope. The VPM is the first optical element in the path of the incoming light through a co-moving baffle. The co-moving baffle comprises a ``cage" structure (see Figure~\ref{fig:bs_rotation}), which houses the fore-optics and the receiver, and a ``baffle extension." The solid black line shows the boresight pointing. Stray light is terminated on blackened interiors of the baffle. The mirrors image the cold stop onto the central part of the VPM. High density polyethylene (HDPE) lenses focus the light onto feedhorn-coupled dual-polarization detectors. Seven detector modules, each comprising 37 pixels, constitute the focal plane array. The HF telescope optical design is conceptually similar; however, it employs silicon instead of HDPE lenses. The HF focal plane consists of three detector modules, each comprising 85 pixels. }
\label{fig:ray_schematic}
\end{figure*}
The optical characterization of the CLASS 40~GHz telescope during the first two years of its operation (``Era~1'') was described by Xu et al~\cite{2020ApJ...891..134X}. In these proceedings, we present preliminary measurements of the 90~GHz and HF telescope pointing stability over several months during the 2020--2021 observing season. We also present preliminary on-sky measurements of the instrument beam in intensity at 90, 150 and 220~GHz made from dedicated planet observations. Characterization of far sidelobes, constraints on instrumental temperature-to-polarization leakage, polarization angle calibration, and beam window functions will be discussed in a future paper.
This paper is organized as follows. The CLASS instrument and survey scan strategy, including telescope configuration, observing modes, and data used in this analysis are described in Section~\ref{sec:instrument&obs}. Preliminary pointing stability of the 90~GHz and HF telescopes is presented in Section~\ref{sec:pointing}. In Section~\ref{sec:beams}, preliminary beam measurements are presented. Finally, we summarize in Section~\ref{sec:summary}.
\section{INSTRUMENT \& OBSERVATIONS}
\label{sec:instrument&obs}
A schematic of the 90~GHz CLASS telescope is shown in Fig.~\ref{fig:ray_schematic} with corresponding ray-trace and main components rendered. The optical design is similar to the design of the 40~GHz telescope \cite{eime12}. The first optical element in the path of the sky signal is the variable-delay polarization modulator (VPM) \cite{eime12, chus12, harr18}, which modulates the polarized signal at 10~Hz before potential instrument-induced polarization signals. This front-end modulation to frequencies higher than atmospheric and instrumental drifts enables recovery of the largest angular scale modes while suppressing temperature-to-polarization leakage from atmospheric signals, which is key to achieving the science goals of CLASS. The fore-optics comprising the primary and secondary mirrors produce an image of the cold stop near the central portion of the VPM, which is 60~cm in diameter. An ultra-high molecular weight polyethylene vacuum window on the cryogenic receiver followed by a stack of infrared (IR)-blocking filters~\cite{essi14, 2018Cryostat_Jeff} allow microwave light to enter the receiver while keeping out infrared radiation. The 60~cm VPM is significantly under-illuminated. This minimizes edge effects that could produce unwanted sidelobes. Two high-density polyethylene (HDPE) lenses re-image the light from each field onto the focal plane with an effective $f$/$\sim1.8$. The HF telescope optics and receiver design is conceptually similar; however, it employs silicon instead of HDPE lenses. The HF infrared filtering scheme is also different (see Iuliano et al.~\cite{2018Cryostat_Jeff}).
\begin{figure*}[htb]
\centering
\includegraphics[width=1\linewidth]{figures/Boresight_lowRes.jpg}
\caption{Photos of the telescope mount showing the seven boresight rotation angles. This rotation keeps the telescope boresight pointing unchanged while rotating the detector antenna orientation on the sky in steps of $15^{\circ}$ within a $90^{\circ}$ range, thus enabling measurement of the polarization signal projected onto different orientations.}
\label{fig:bs_rotation}
\end{figure*}
The focal planes of the 90~GHz and HF telescopes consist of seven and three detector modules, respectively. Each 90~GHz module has 37 smooth-walled feedhorns~\cite{zeng2010, daha18} that couple the incoming light to microfabricated antenna probes on a 100~mm planar detector wafer~\cite{Rostem:2016}. Each HF module has 85 smooth-walled feedhorns~\cite{daha19}. The feeds are hexagonal close packed. The 90~GHz feeds are individually assembled whereas the HF feeds are fabricated in array format~\cite{doi:10.1063/5.0049526}. The 90~GHz telescope's field-of-view (FOV) is approximately $23^\circ$ in diameter, while the FOV of the HF telescope is approximately $18^\circ$ in diameter. Each pixel is sensitive to both orthogonal states of linear polarization enabled by an orthomode transducer (OMT) on the detector wafer. Each pixel has two transition-edge sensor (TES) bolometers for measuring the power in both polarization states. The focal plane is cooled to $\sim50~mK$ by a dilution refrigerator~\cite{2018Cryostat_Jeff}. The on-sky receiver performance at all four frequencies is presented in Dahal et al.~\cite{2022ApJ...926...33D}.
A three-axis mount, shown in Figure~\ref{fig:bs_rotation}, points the 90 and 40~GHz telescopes. A second identical mount supports the HF telescope. The two mounts can rotate in azimuth, elevation, and boresight independent of each other. During CMB observations, the mounts nominally rotate 720$^{\circ}$ in azimuth at a fixed elevation of 45$^{\circ}$, before reversing and scanning in the opposite direction, thus scanning the sky in large circles. During the majority of CMB observations, the telescope scanned in azimuth at 2$^{\circ}$/s. During the day, the telescope boresight is restricted to point at least 20$^{\circ}$ away from the Sun, increasing the frequency of scan direction reversals and reducing the azimuthal range from the nominal 720$^{\circ}$. The telescopes observe $\sim75\%$ of the sky every day. Complete measurement of Stokes Q and U is enabled by daily rotations of the telescope boresight platform, which rotate the orientation of the detector polarization on the sky. The rotation about the boresight spans $90^{\circ}$ as depicted in Fig.~\ref{fig:bs_rotation}, cycling through angles $-45^{\circ}$, $-30^{\circ}$, $-15^{\circ}$, $0^{\circ}$, $15^{\circ}$, $30^{\circ}$, $45^{\circ}$ each week. This scan strategy enables cross-linking on different time scales with each pixel on the sky being observed with different scanning directions and detector orientations.
\begin{figure*}[!htb]
\centering
\includegraphics[width=1\textwidth]{figures/W1_band_pointing_new.pdf}
\caption{Deviations from the W-band pointing model in arcminutes for telescope azimuth, elevation, and boresight rotation angle over the period of July 2020 through April 2021. Different telescope boresight angles ranging from $-45^{\circ}$ to $+45^{\circ}$ are highlighted in a rainbow color pattern. Diamonds represent deviations inferred from {\textit{dedicated scans}}, and tri-left markers from {\textit{survey scans}}. Only those {\textit{survey scans}} where at least 100 detectors see the Moon are used to derive pointing information. Each data point corresponds to pointing deviation inferred from all detectors that see the Moon during that {\textit{dedicated}} / {\textit{survey scan}} combined. Deviations are inferred separately from scans conducted during the Moon rising and setting, when available. The gray vertical line highlights the start date of a new pointing correction model based on a series of {\textit{dedicated scans}}.}
\label{fig:W1_pointing}
\end{figure*}
\begin{figure*}[htb]
\centering
\includegraphics[width=0.9\textwidth]{figures/G_band_pointing_new.pdf}
\caption{Deviations from the HF pointing model in arcminutes for telescope azimuth, elevation, and boresight rotation angle over the period of January through May 2021. Different telescope boresight angles ranging from $-45^{\circ}$ to $+45^{\circ}$ are highlighted in a rainbow color pattern. Diamonds represent deviations inferred from {\textit{dedicated scans}}, and tri-left markers from {\textit{survey scans}}. Only those {\textit{survey scans}} where at least 100 detectors see the Moon are used to derive pointing information. Each data point corresponds to pointing deviation inferred from all detectors that see the Moon during that {\textit{dedicated}} / {\textit{survey scan}} combined. Deviations are inferred separately from scans conducted during the Moon rising and setting, when available. The gray vertical line highlights the start date of a new pointing correction model based on a series of {\textit{dedicated scans}}.}
\label{fig:HF_pointing}
\end{figure*}
Calibration campaigns are periodically undertaken, in which dedicated observations of bright sources, namely, the Moon, Jupiter, and Venus are performed. Moon scans are used for pointing calibration, whereas planet scans are used for beam characterization since the bright signal from the Moon saturates many of the detectors. The planets are essentially point sources given the instrument beam size at all frequencies. During each calibration run, the telescope scans across the source back and forth azimuthally, maintaining a constant elevation and allowing the source to drift through the FOV in elevation. For the remainder of the paper, a ``scan'' in the context of pointing calibration and beam characterization refers to a set of continuous sweeps conducted across a source as it rises or sets through the array. Scans performed with different boresight rotations help improve the sampling of the spatial beam and provide information about the offset of the center of the array which would otherwise be degenerate with the boresight pointing. In addition to these {\textit{dedicated scans}}, the telescopes scan over the Moon during CMB observations (henceforth referred to as {\textit{survey scans}}). The {\textit{dedicated scans}} are used for updating the pointing model while the {\textit{survey scans}} are only used for tracking pointing deviations from the last updated model, because the Moon sampling is sparse in such scans.
\section{POINTING}
\label{sec:pointing}
The raw time-ordered data from the {\textit{dedicated scans}} are calibrated~\cite{2022arXiv220506901A}, filtered, and downsampled. Moon-centered intensity maps are then generated for each individual detector, which are fitted with two-dimensional Gaussian profiles to provide the pointing model. Details of the data processing will be described in a future paper. The pointing model is updated using the data from each calibration run. The pointing deviations with respect to the model in telescope azimuth, elevation, and boresight angle as a function of time over a representative period of several months during the 2020--2021 observing season are shown in Figures~\ref{fig:W1_pointing}~and~\ref{fig:HF_pointing} for 90~GHz and HF, respectively. The vertical gray lines indicate updates to the pointing model. The mean offset from model and standard deviations of the 90~GHz telescope pointing over this period of several months are $0.1'\pm1.4'$, $0.4'\pm1.3'$, and $0.4'\pm1.7'$ in azimuth, elevation, and boresight rotation angle, respectively. Similarly, the mean offset from model and standard deviations of the HF telescope pointing are $0.6'\pm0.9'$, $-0.5'\pm1.0'$, and $-0.6'\pm1.6'$ in azimuth, elevation, and boresight rotation angle, respectively.
\section{BEAMS}
\label{sec:beams}
{\textit{Dedicated scans}} of Jupiter performed around the time of the Earth's closest approach to Jupiter and during good weather conditions at the telescope site are used to measure the angular response of the detectors on the sky. Each {\textit{dedicated scan}} spans approximately two hours and consists of multiple passes over the planet. The detectors are tuned prior to each scan. For each detector and its corresponding pointing offset, we define a coordinate system centered at the detector, with the y-axis pointing along the local meridian in the receiver coordinate system, which is a spherical coordinate system with array center located at zero longitude on the equator. Using the data from each scan, a per-detector planet map is generated out to a radius of 4$^{\circ}$ and evaluated on criteria such as signal-to-noise ratio ($S/N$) and root-mean-square ($rms$) noise. Only maps meeting certain quality thresholds are retained. Details regarding these selection criteria will be described in a future paper. The retained maps are averaged per detector with $S/N$-weighting to generate average per-detector beam maps. A total of 319, 389, and 211 detectors at 90, 150, and 220~GHz, respectively, detected the signal from Jupiter~\cite{2022ApJ...926...33D}. For each detector, we define an effective beam full width at half maximum (FWHM) as
\begin{equation}
\sigma_\mathrm{eff} = \sqrt{2\frac{\sigma_\mathrm{maj}^{2}\times\sigma_\mathrm{min}^{2}}{\sigma_\mathrm{maj}^{2}+\sigma_\mathrm{min}^{2}}}
\label{equ:effective_fwhm}
\end{equation}
where the FWHM of the beams along the major and minor axes $\sigma_\mathrm{maj}$ and $\sigma_\mathrm{min}$ are obtained from a two-dimensional Gaussian fit to the average per-detector beam maps.
\begin{figure*}
\includegraphics[width=0.95\linewidth]{figures/SPIE_stacked_beams.pdf}
\caption{\textbf{Average instrument beams}. Preliminary peak-normalized beam maps out to 4$^{\circ}$ in radius at 90~GHz (left panel) and 2$^{\circ}$ in radius at 150 and 220~GHz (middle and right panels, respectively) obtained by coadding 209, 385, and 208 average per-detector beam maps at 90, 150, and 220~GHz, respectively. The color bar scale is logarithmic from 1 to $10^{-3}$, and linear from $10^{-3}$ to $-10^{-3}$ in order to display negative values. The map has a resolution of 0.05$^{\circ}$. }
\label{fig:avg_beam_map}
\end{figure*}
The per-detector averaged maps are then $S/N$-weighted and co-added to generate the effective instrument beam map. When co-adding, we excluded detectors with (1) low end-to-end optical efficiency, (2) highly eccentric beams, or (3) an out-of-range voltage bias for the majority of planet scans.
In total, the above selection criteria resulted in 209, 385, and 208 per-detector beam maps being averaged to generate the effective instrument beam maps at 90, 150, and 220~GHz, respectively, as shown in Fig.~\ref{fig:avg_beam_map}. Due to the left-right symmetry of the optics, averaging per-detector maps from across the focal plane circularizes the instrument beam map. Additionally, boresight rotation and observations of the same point rising and setting (i.e., sky rotation) further ensure that the effective beam for CMB observations is nearly circular. Data from the seven boresight orientations contribute with nearly equal weights to the survey map. Therefore, we approximate our effective beams by azimuthally averaged radial beam profiles as shown in Figure~\ref{fig:1D_beam_prof}.
\begin{figure}
\begin{center}
\includegraphics[width=0.65\linewidth]{figures/SPIE_beam_profiles.pdf}
\caption{\textbf{Beam profiles}. Azimuthally averaged 90 (red squares), 150 (blue circles), and 220 (orange triangles) GHz preliminary beam profiles out to 3$^{\circ}$. The errorbars represent 1$\sigma$ uncertainties.}
\label{fig:1D_beam_prof}
\end{center}
\end{figure}
\section{Summary}
\label{sec:summary}
In these proceedings, we presented measurements of the 90~GHz and HF telescope pointing stability over several months during the 2020--2021 observing season. The telescope pointing offset from model and pointing uncertainties in azimuth, elevation, and boresight rotation angle are on the order of a few percent of the respective beam FWHM. We also presented preliminary measurements of the 90, 150, and 220~GHz average instrument beam in intensity using dedicated planet observations. The measured effective beam FWHMs are $0.615\pm0.019^{\circ}$, $0.378\pm0.005^{\circ}$, and $0.266\pm0.008^{\circ}$ at 90, 150, and 220~GHz, respectively. Other studies related to pointing, beams and optical characterization will be discussed in a future paper.
\acknowledgments
We acknowledge the National Science Foundation Division of Astronomical Sciences for their support of CLASS under Grant Numbers 0959349, 1429236, 1636634, 1654494, 2034400, and 2109311. We thank Johns Hopkins University President R. Daniels and the Deans of the Kreiger School of Arts and Sciences for their steadfast support of CLASS. We further acknowledge the very generous support of Jim and Heather Murren (JHU A\&S ’88), Matthew Polk (JHU A\&S Physics BS ’71), David Nicholson, and Michael Bloomberg (JHU Engineering ’64). The CLASS project employs detector technology developed in collaboration between JHU and Goddard Space Flight Center under several previous and ongoing NASA grants. Detector development work at JHU was funded by NASA cooperative agreement 80NSSC19M0005. CLASS is located in the Parque Astron\'omico Atacama in northern Chile under the auspices of the Agencia Nacional de Investigaci\'on y Desarrollo (ANID). We acknowledge scientific and engineering contributions from Max Abitbol, Fletcher Boone, Jay Chervenak, Lance Corbett, David Carcamo, Mauricio D\'iaz, Ted Grunberg, Saianeesh Haridas, Connor Henley, Ben Keller, Lindsay Lowry, Nick Mehrle, Grace Mumby, Diva Parekh, Isu Ravi, Daniel Swartz, Bingjie Wang, Qinan Wang, Emily Wagner, Tiffany Wei, Zi\'ang Yan, Lingzhen Zeng, and Zhuo Zhang. For essential logistical support, we thank Jill Hanson, William Deysher, Miguel Angel D\'iaz, Mar\'ia Jos\'e Amaral, and Chantal Boisvert. We acknowledge productive collaboration with Dean Carpenter and the JHU Physical Sciences Machine Shop team.
Sumit Dahal is supported by an appointment to the NASA Postdoctoral Program at the NASA Goddard Space Flight Center, administered by Oak Ridge Associated Universities under contract with NASA. Zhilei Xu is supported by the Gordon and Betty Moore Foundation through grant GBMF5215 to the Massachusetts Institute of Technology.
|
\section{Introduction}
We consider on what conditions the `Rank and Yank' will not whip anyone in an organization provided that the evaluation data under each criterion are ordinal rankings and the simple majority rule is used as the aggregation rule.
`Rank and Yank', also called `Up or Out policy', `Forced Ranking', or `Vitality Curve' popularized by General Electric and later adopted by many organizations, refers to a forced distribution rating system (FDRS) that rewards the top employees and removes the bttom employees (in certain ratios, for instance, 20:70:10) each period (in a certain frequency, for instance, every year or every 4 year). This policy or system has been extensively investigated by the academia. Topics cover its advanges and disadvantages from various prospectives (see, e.g., Dominick 2009, Giumetti et al. 2015, Stewart et al. 2010, and many others).
In a `Rank and Yank' scenario, the performances of the employees of a given organization are aggregated into a social preference that reflects the organization's general opinion based on which the employees are ranked. It is obvious when the employees under stake are rated with respect to a single criterion and the employees perform same good, then nobody will be whipped by the `Rank and Yank' policy. This is a simple but trivial case. When the employees are rated with respect to multiple criteria, and the evaluation data under each criterion are assumed of ordinal rankings, the situations can be a bit more complicated in the sense that an employee may perform better with respect to one criterion while worse with respect to another. This paper does not touch on such topics as the policy's ethical issues, reasonability or efficiency. Rather, we concentrate on a certain situation of the policy, that is, even though the policy works well but none will be rewarded or punished given that the evaluation data under each criterion are ordinal rankings and the simple majority rule is used. When this phenomenon (i.e., `none to be whipped') is to be observed, our question is, what the conditions will be?
This question is different from two much concerned topics in social choice theory, one of which is known as `social preference transitivity under the majority rule', and the other is `social choice set existence under the majority rule' (for details see, e.g., Arrow 1951, Sen 1970, and Gaertner 2009). This is evidenced by the following illustration where 3 alternatives $\{x_1,x_2,x_3\}$ are evaluated (the terminology $x\sim y$ stands for `$x$ is indifferent to $y$', and $x\succ y$ stands for `$x$ is more preferred to $y$'):
\begin{itemize}
\item On the one hand, if the final aggregation outcome is $x_1\sim x_2\sim x_3$, then the social preference is transitive and the social choice set is not empty. In this case will nobody be whipped by Rank and Yank.
\item One the other hand, if the final aggregation outcome is $x_1\succ x_2\succ x_3\succ x_1$, then the social preference is not transitive and the social choice set does not exist either. However, `none to be whipped' still happens in this particular case.
\end{itemize}
Therefore, the occurence of `none to be whipped' is independent of the transitivity of the social preference or the existence of the social choice set.
We will present two sufficient conditions. To this end, the paper is organized as follows. Section 2 includes our assumptions and notations. Some aggregation outcomes possible for `none to be whipped' are listed in Section 3. Section 4 presents a sufficient condition in terms of the election matrix whose entries represent the numbers of criteria under which one alternative being preferred to another. In Section 5, we first introduce a matrix called preference probability map (PPM), whose entries stand for the probabilities of alternatives occupying ranking positions. We then obtain a sufficient condition in terms of the sum and mean matrix of the PPMs. Some examples are also examined. And Section 6 includes our concluding remark and put forward some possible directions for future research.
\section{Assumption and notation}
We henceforth call the employees as alternatives. The discussion rests on the following \textit{assumptions}.
\begin{itemize}
\item []\textbf{Assumption I} The alternatives are finite, evaluated under multiple criteria and the criteria are dealt with equally;
\item []\textbf{Assumption II} The evaluation data under each criterion are collected as ordinal rankings;
\item []\textbf{Assumption III} The simple majority rule is used as the aggregation rule;
\item []\textbf{Assumption IV} When to rank alternatives with respect to one criterion, those in a tie will occupy common consecutive positions with an equal probability;
\item []\textbf{Assumption V} The evaluation data under different criteria are independent in the sense that, even though an alternative performs good regarding one criterion, he/she may not necessarily perform good regarding another criterion.
\end{itemize}
The \textit{significance} of the above assumptions lie in that:
\begin{itemize}
\item [$\diamond$] The first three assumptions imply that the 'Rank and Yank' problem under consideration can be treated as a social choice problem where {\em the ordinal ranking of the alternatives under a criterion is viewed as provided by a voter};
\item [$\diamond$] The fourth assumption indicates a probability mapping of each ordinal ranking, which will be described in Section 5;
\item [$\diamond$] The last assumption ensures the feasibility of probability addition in the discussion.
\end{itemize}
To conduct the discussion, the following \textit{notations} will be used throughout the paper.
\begin{itemize}
\item [-] $x_i$ denote alternatives (employees), where $i=1,2,\ldots, m$ and $1<m<+\infty$. Accordingly, $X=\{x_1,x_2,\ldots,x_m\}$ stands for the alternative set.
\item [-] $c_j$ denote criteria (voters), where $j=1,2,\ldots, n$ and $1<n<+\infty$. Accordingly, $\{c_1,c_2,\ldots,c_n\}$ stands for the criterion (voter) set.
\item [-] Possible evaluation data of alternatives $x_i$ and $x_k$ with respect to criterion $j$ can be one of $x_i\succ_j x_k$, $x_i\sim_j x_k$ and $x_i\prec_j x_k$, which stand for ``$x_i$ is more preferred to $x_k$", ``$x_i$ is indifferent to $x_k$" and ``$x_i$ is less preferred to $x_k$", respectively. When we speak of social preferences over alternatives $x_i$ and $x_k$, we just use similar notations but without any subscript attached to `$\succ$', `$\sim$' or `$\prec$'.
\item [-] $N(x_i\succ_j x_k)$ and $N(x_i\sim_j x_k)$ denote the numbers of criteria satisfying $x_i\succ_j x_k$ and $x_i\sim_j x_k$, respectively.
\item [-] The majority rule means that $x_i\succ x_k$ iff $N(x_i\succ_j x_k)>N(x_k\succ_j x_i)$, and $x_i\sim x_k$ iff $N(x_i\succ_j x_k)=N(x_k\succ_j x_i)$.
\end{itemize}
We note that the majority rule in this paper belongs to a `simple majority rule' rather than an `absolute majority rule'. For knowledge of the difference, readers may refer to Arrow (1951), Sen (1970), and Gaertner (2009).
\section{Aggregation outcomes leading to `none to be whipped'}
A fundamental problem in our topic is to identify what final aggregation outcomes will lead to the occurrence of `none to be whipped'. When the majority rule does not deliver a distinguishable result, indifference or cyclicity may hold over the whole alternative set. We list some outcomes of this sort as follows.
\begin{itemize}
\item [(1)] The majority rule yields a final aggregation outcome that all alternatives are indifferent to each other. For instance, when to evaluate 4 alternatives, the particular case of this kind refers to the case where the final collective outcome is $x_1\sim x_2\sim x_3\sim x_4$.
\item [(2)] The majority rule yields a final aggregation outcome that cyclicity holds over the whole alternative set. For instance, when to evaluate 5 alternatives, the particular case of this kind can be the case where the final collective outcome is $x_1\succ x_2\succ x_3\succ x_4\succ x_5\succ x_1$.
\item [(3)] The majority rule yields a final aggregation outcome that is similar to an indifference line including some cycles. For instance, the final collective outcome is such like $x_1\sim\{x_2\succ x_3\succ x_4\}$.
\item [(4)] When the majority rule yields a final aggregation outcome that a global cycle occurs with some local indifference and/or cyclicity mixed in the cycle, 'none to be whipped' will occur. For illustration, the cases can be such as
$$\begin{array}{ll}
x_1\succ \{x_2\sim x_3\sim x_4\}\succ x_5\succ x_1,\\
x_1\succ \{x_2\succ x_3\succ x_4\succ x_2\}\succ x_5\succ x_1,\\
x_1\succ \{x_2\sim x_3\sim x_4\}\succ x_5\succ \{x_6\succ x_7\succ x_8\succ x_9\succ x_6\}\succ x_{10}\succ x_1, \mbox{etc}.
\end{array}
$$
\end{itemize}
In this paper, we will make an attempt to formulate mathematically some sufficient conditions relevant to the first three cases.
\section{A sufficient condition in terms of election matrix}
Binary comparison matrices are widely used in social choice literature whose elements can be assumed of various meannings. For our analytical purpose in this section, we use the one called \textit{election matrix} $M=(a_{ik})_{m\times m}$ whose entry $a_{ik}$ represents the number of times alternative $x_i$ is ranked before aternative $x_k$ across the $n$ criteria (Levenglick 1975). With this interpretation of the entries of the election matrix, the following condition becomes apparent.
\textbf{Sufficient condition in terms of election matrix} If the election matrix $M=(a_{ik})_{m\times m}$ is symmetric, i.e., $a_{ik}=a_{ki}$, then none will be whipped by the `Rank and Yank' under the majority rule.
\textbf{Proof} Since $a_{ik}=N(x_i\succ_j x_k)$ hence $a_{ik}=a_{ki}$ means $N(x_i\succ_j x_k)=N(x_k\succ_j x_i)$, we thus have $x_i\sim x_k$, $\forall i,k$, under the majority rule. Therefore, in this case all alternatives are indifferent to each other and none will be whipped by the `Rank and Yank'.
\section{A sufficient condition in terms of preference probability map}
We start with two definitions which will prove to be useful for the representation of ordinal rankings. Then we present a sufficient condition in terms of one of them.
\textbf{Preference map} (Hou 2015a, 2015b; Hou \& Triantaphyllou 2019) The \emph{preference map} is a sort of preference sequence corresponding to individual $j$'s weak ordering over the alternative set $X=\{x_1,x_2,\ldots,x_m\}$, denoted $PM^{(j)}=[PM_i^{(j)}]_{m\times 1}$, such that
$$PM_i^{(j)}=\{|A_i^{(j)}|+1,|A_i^{(j)}|+2,\ldots,|A_i^{(j)}|+|B_i^{(j)}|\},\eqno(1)$$
where $|\centerdot|$ represents the cardinality of a set;
$A_i^{(j)}$ represents the \textit{predominance set} of alternative $x_i$ according to individual $j$'s preference, i.e., $A_i^{(j)}=\{x_q\mid x_q\in X, x_q \succ_j x_i\}$; and $B_i^{(j)}$ represents the \textit{indifference set} of alternative $x_i$, i.e., $B_i^{(j)}=\{x_q\mid x_q\in X,x_q \sim_j x_i\}$.
For example, we consider the orderings of $x_1\succ x_2\succ x_3\succ x_4$, $x_1\sim x_2\succ x_3\sim x_4$, $x_1\succ x_2\sim x_3\sim x_4$ and $x_1\sim x_2\sim x_3\sim x_4$. Their preferences maps are
$$
\begin{array}{c}
{\begin{array}{cc}
{\begin{array}{c} x_1\\
x_2\\
x_3\\
x_4
\end{array}}&
{ \left[\begin{array}{c}
\{1\}\\
\{2\}\\
\{3\}\\
\{4\}\end{array}\right],
\left[\begin{array}{c}\{1,2\}\\
\{1,2\}\\
\{3,4\}\\
\{3,4\}\end{array}\right],
\left[\begin{array}{c}\{1\}\\
\{2,3,4\}\\
\{2,3,4\}\\
\{2,3,4\}\end{array}\right]\mbox{~and~}
\left[\begin{array}{c}\{1,2,3,4\}\\
\{1,2,3,4\}\\
\{1,2,3,4\}\\
\{1,2,3,4\}\end{array}\right],
}
\end{array}}
\end{array}
$$
respectively.
From the definition of the preference map one can see that, a preference map is an `almost vector' whose entries are sets containing the alternatives' possible ranking position or positions, that is, each alternative is mapped to its ranking position or positions. If two alternatives are in a tie, then their corresponding entries include same numbers which indicate that those alternatives in a tie will occupy same ranking positions. This characteristic of the preference map reflects the requirements in Assumption IV that those alternatives ``in a tie will occupy common consecutive positions". In order to reflect exactly the ``probability" information required in Assumption IV, however, we need another concept as described below.
\textbf{Preference possibility map (Hou 2022)} Let $PM=[PM_i]_{m\times 1}$ be a preference map of a weak ordering. The preference possibility map (for short, PPM) corresponding to that weak ordering is defined by a $m\times m$ matrix $[PPM_{i,k}]_{m\times m}$ such that
$$
{PPM_{i,k}}=\begin{cases}
\frac{1}{|PM_i|}, & \text{if $k\in PM_i$},\\
0, & \text{otherwise}.
\end{cases}
\eqno(2)$$
One can see from Eq.(2) that the PPM reflects exactly the requirement in Assumption IV that those alternatives ``in a tie will occupy common consecutive positions with an equal probability". We note that this sort of strategy (i.e., alternatives in a tie will occupy common consecutive positions with an equal probability) can be traced back at least as early as Black (1976) when to assign Borda scores to those alternatives in a tie.
In our previous example, the corresponding PPMs would be
$$
\begin{array}{c}
\begin{array}{c}{\hspace{0.7cm}\begin{array}{ccccc}1\hspace{0.3cm}2\hspace{0.4cm}3\hspace{0.4cm}4
&\hspace{1.0cm}1\hspace{0.7cm}2\hspace{0.7cm}3\hspace{0.7cm}4 &\hspace{1.0cm}1\hspace{0.7cm}2\hspace{0.7cm}3\hspace{0.7cm}4 &\hspace{1.8cm}1\hspace{0.7cm}2\hspace{0.7cm}3\hspace{0.7cm}4
\end{array}}\\
{\begin{array}{cc}
{\begin{array}{c} x_1\\
x_2\\
x_3\\
x_4
\end{array}}&
{ \left[\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & 1 & 0 \\
0 & 0 & 0& 1
\end{array}\right],
\left[\begin{array}{cccc}
1/2 & 1/2 & 0 & 0 \\
1/2 & 1/2 & 0 & 0 \\
0 & 0 & 1/2 & 1/2 \\
0 & 0 & 1/2 & 1/2
\end{array}\right],
\left[\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & 1/3 & 1/3 & 1/3 \\
0 & 1/3 & 1/3 & 1/3 \\
0 & 1/3 & 1/3 & 1/3
\end{array}\right]\mbox{~and~}
\left[\begin{array}{cccc}
1/4 & 1/4 & 1/4 & 1/4 \\
1/4 & 1/4 & 1/4 & 1/4 \\
1/4 & 1/4 & 1/4 & 1/4 \\
1/4 & 1/4 & 1/4 & 1/4
\end{array}\right],
}
\end{array}}
\end{array}
\end{array}
$$
respectively.
From its definition we know that a PPM has the following properties.
\textbf{Property of PPM} Suppose $[PPM_{i,k}]_{m\times m}$ is a PPM. We have
\begin{itemize}
\item $0\leq PPM_{i,k}\leq 1, \forall i,k$;
\item $\sum_{i=1}^{m}PPM_{i,k}=1$;
\item $\sum_{k=1}^{m}PPM_{i,k}=1$.
\end{itemize}
The above property of PPM indicates that, an alternative occupies at least one ranking position, meanwhile a ranking position is occupied at least by one alternative. Actually, this is a matter of fact that a PPM represents an ordinal ranking.
The following definition is crucial for the analysis in this section.
\textbf{Definition 1} Suppose that $[PPM_{i,k}^{(j)}]_{m\times m}$ are PPMs, which correspond to the alternatives' ordinal rankings with respect to criteria $j$, $j=1,2,\ldots,n$. Their \textit{sum matrix} and \textit{mean matrix} are defined by
$$\Sigma PPM=[n_{i,k}]_{m\times m}=\left[\Sigma_{j=1}^{n}PPM_{i,k}^{(j)}\right]_{m\times m}$$
and $$\overline{\Sigma PPM}=[\overline{p}_{i,k}]_{m\times m}=\left[\frac{1}{n}\Sigma_{j=1}^{n}PPM_{i,k}^{(j)}\right]_{m\times m},$$
respectively.
Since $PPM_{i,k}^{(j)}$ is interpreted as the probability of alternative $x_i$ ranking at position $k$ under criterion $j$, and the probabilities across the criteria are assumed independent of each other (Assumption V), thus $\overline{p}_{i,k}$ represents the average probability of the society's ranking alternative $x_i$ at position $k$. From the property of PPM, the following properties hold respectively for mean matrix and sum matrix.
\textbf{Property of mean matrix} Suppose $[\overline{p}_{i,k}]_{m\times m}$ is a mean matrix. We have
\begin{itemize}
\item $0\leq\overline{p}_{i,k}\leq 1, \forall i,k$;
\item $\sum_{i=1}^{m}\overline{p}_{i,k}=1$;
\item $\sum_{k=1}^{m}\overline{p}_{i,k}=1$.
\end{itemize}
\textbf{Property of sum matrix} Suppose $[n_{i,k}]_{m\times m}$ is a sum matrix. We have
\begin{itemize}
\item $n_{i,k}\geq 0, \forall i,k$;
\item $\sum_{i=1}^{m}n_{i,k}=n$;
\item $\sum_{k=1}^{m}n_{i,k}=n$.
\end{itemize}
\textbf{Lemma 1} The number of criteria under which alternative $x_i$ is ranked at position $k$ is
$$n_{i,k}=\Sigma_{j=1}^{n}PPM_{i,k}^{(j)}=n\times \overline{p}_{i,k}.$$
\textbf{Proof} According to classical definition of probability, we know $\overline{p}_{i,k}=\frac{n_{i,k}}{n}$. Given $\overline{p}_{i,k}=\frac{1}{n}\Sigma_{j=1}^{n}PPM_{i,k}^{(j)}$, we have $n_{i,k}=\Sigma_{j=1}^{n}PPM_{i,k}^{(j)}=n\times \overline{p}_{i,k}$. Q.E.D
The following result can be established.
\textbf{Lemma 2} If $\forall i,k\left(\overline{p}_{i,k}=\frac{1}{m}\right)$ holds for the mean matrix, then, none of the alternatives $\{x_1,x_2,\ldots,x_m\}$ will be whipped by the `Rank and Yank'.
\textbf{Proof} Since $\overline{p}_{i,k}$ stands for the social probability of alternative $x_i$ ranking at position $k$, thus when $\forall i,k\left(\overline{p}_{i,k}=\frac{1}{m}\right)$ holds for the mean matrix, any alternative would occupy any position with an identical probability. In this case, we know that either all the alternatives are indifferent to each other, or cyclicity holds over the whole alternative set. Given the condition, therefore, none will be whipped by the `Rank and Yank'. Q.E.D
Can Lemma 2 be generalized even further? We have the following result.
\textbf{Theorem 1 (A sufficient condition)} If the entries of the mean matrix satisfy a dual relation, that is,
$$\forall i(\overline{p}_{i,1}=\overline{p}_{i,n}, \overline{p}_{i,2}=\overline{p}_{i,n-1}, \ldots),\eqno(3)$$
or equivalently, the entries of the sum matrix satisfy the dual relation
$$\forall i(n_{i,1}=n_{i,n}, n_{i,2}=n_{i,n-1}, \ldots),\eqno(4)$$
then, none of the alternatives $\{x_1,x_2,\ldots,x_m\}$ will be whipped by the `Rank and Yank'.
\textbf{Proof} Given the condition of formula (3), we know that
\begin{itemize}
\item If $\forall i,k\left(\overline{p}_{i,k}=\frac{1}{m}\right)$, then from Lemma 2 we know that the theorem holds.
\item Otherwise, from formula (3) we know that the average ranking position of any alternative will be $\frac{m+1}{2}$, which indicates that the aggregation outcome will be an indifference line with or without some cycles in it. Thererefore, in this case none will be whipped. Q.E.D
\end{itemize}
\textbf{Example 1} Consider $X=\{x_1,x_2,x_3,x_4\}$, and the evaluation data under 5 criteria are given as:
$$\begin{array}{ll}
\text{Criterion 1}: & x_1 \succ_1 x_2 \sim_1 x_3 \sim_1 x_4,\\
\text{Criterion 2}: & x_2 \sim_2 x_3 \sim_2 x_4 \succ_2 x_1,\\
\text{Criterion 3}: & x_1 \sim_3 x_2 \succ_3 x_3 \sim_3 x_4,\\
\text{Criterion 4}: & x_3 \sim_4 x_4 \succ_4 x_1 \sim_4 x_2,\\
\text{Criterion 5}: & x_1 \sim_5 x_2 \sim_5 x_3 \sim_5 x_4.
\end{array}
$$
The corresponding PPMs are
$$
\begin{array}{c}
{\begin{array}{cc}
{\begin{array}{c} x_1\\
x_2\\
x_3\\
x_4
\end{array}}&
{ \left[\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & \frac{1}{3} & \frac{1}{3} & \frac{1}{3} \\
0 & \frac{1}{3} & \frac{1}{3} & \frac{1}{3} \\
0 & \frac{1}{3} & \frac{1}{3} & \frac{1}{3}
\end{array}\right],
\left[\begin{array}{cccc}
0 & 0 & 0 & 1 \\
\frac{1}{3} & \frac{1}{3} & \frac{1}{3} & 0 \\
\frac{1}{3} & \frac{1}{3} & \frac{1}{3} & 0 \\
\frac{1}{3} & \frac{1}{3} & \frac{1}{3} & 0
\end{array}\right],
\left[\begin{array}{cccc}
\frac{1}{2} & \frac{1}{2} & 0 & 0 \\
\frac{1}{2} & \frac{1}{2} & 0 & 0 \\
0 & 0 & \frac{1}{2} & \frac{1}{2} \\
0 & 0 & \frac{1}{2} & \frac{1}{2}
\end{array}\right],
\left[\begin{array}{cccc}
0 & 0 & \frac{1}{2} & \frac{1}{2} \\
0 & 0 & \frac{1}{2} & \frac{1}{2} \\
\frac{1}{2} & \frac{1}{2} & 0 & 0 \\
\frac{1}{2} & \frac{1}{2} & 0 & 0
\end{array}\right],
\left[\begin{array}{cccc}
\frac{1}{4} & \frac{1}{4} & \frac{1}{4} & \frac{1}{4} \\
\frac{1}{4} & \frac{1}{4} & \frac{1}{4} & \frac{1}{4} \\
\frac{1}{4} & \frac{1}{4} & \frac{1}{4} & \frac{1}{4} \\
\frac{1}{4} & \frac{1}{4} & \frac{1}{4} & \frac{1}{4}
\end{array}\right].
}
\end{array}}
\end{array}
$$
The sum matrix is
$$
\begin{array}{c}
{\begin{array}{cc}
{\begin{array}{c} x_1\\
x_2\\
x_3\\
x_4
\end{array}}&
{
\left[\begin{array}{cccc}
\frac{7}{4} & \frac{3}{4} & \frac{3}{4} & \frac{7}{4} \\
\frac{13}{12} & \frac{17}{12} & \frac{17}{12} & \frac{13}{12} \\
\frac{13}{12} & \frac{17}{12} & \frac{17}{12} & \frac{13}{12} \\
\frac{13}{12} & \frac{17}{12} & \frac{17}{12} & \frac{13}{12}
\end{array}\right].
}
\end{array}}
\end{array}
$$
Condition (4) is fulfilled. Hence none will be whipped by `Rank and Yank' in this example. Actually, the final aggregation outcome is $x_1 \sim x_2 \sim x_3 \sim x_4$.
\textbf{Example 2} Consider $X=\{x_1,x_2,x_3,x_4\}$, and the evaluation data under 6 criteria are given as:
$$\begin{array}{ll}
\text{Criterion 1}: & x_1 \succ_1 x_2 \succ_1 x_3 \succ_1 x_4,\\
\text{Criterion 2}: & x_1 \succ_2 x_3 \succ_2 x_4 \succ_2 x_2,\\
\text{Criterion 3}: & x_1 \succ_3 x_4 \succ_3 x_2 \succ_3 x_3,\\
\text{Criterion 4}: & x_2 \succ_4 x_3 \succ_4 x_4 \succ_4 x_1,\\
\text{Criterion 5}: & x_3 \succ_5 x_4 \succ_5 x_2 \succ_5 x_1,\\
\text{Criterion 6}: & x_4 \succ_6 x_2 \succ_6 x_3 \succ_6 x_1.
\end{array}
$$
The corresponding PPMs are
\begin{footnotesize}
$$
\begin{array}{c}
{\begin{array}{cc}
{\begin{array}{c} x_1\\
x_2\\
x_3\\
x_4
\end{array}}&
{ \left[\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & 1 & 0 \\
0 & 0 & 0 & 1
\end{array}\right],
\left[\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 \\
0 & 1 & 0 & 0 \\
0 & 0 & 1 & 0
\end{array}\right],
\left[\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 \\
0 & 0 & 0 & 1 \\
0 & 1 & 0 & 0
\end{array}\right],
\left[\begin{array}{cccc}
0 & 0 & 0 & 1 \\
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & 1 & 0
\end{array}\right],
\left[\begin{array}{cccc}
0 & 0 & 0 & 1 \\
0 & 0 & 1 & 0 \\
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0
\end{array}\right],
\left[\begin{array}{cccc}
0 & 0 & 0 & 1 \\
0 & 1 & 0 & 0 \\
0 & 0 & 1 & 0 \\
1 & 0 & 0 & 0
\end{array}\right].
}
\end{array}}
\end{array}
$$
\end{footnotesize}
The sum matrix is
$$
\begin{array}{c}
{\begin{array}{cc}
{\begin{array}{c} x_1\\
x_2\\
x_3\\
x_4
\end{array}}&
{
\left[\begin{array}{cccc}
3 & 0 & 0 & 3 \\
1 & 2 & 2 & 1 \\
1 & 2 & 2 & 1 \\
1 & 2 & 2 & 1
\end{array}\right].
}
\end{array}}
\end{array}
$$
Condition (4) is fulfilled. Hence none will be whipped by `Rank and Yank' in this example. Actually, the final aggregation outcome is $x_1 \sim \{x_2 \succ x_3 \succ x_4\succ x_2\}$.
\textbf{Remark} The condition presented in Theorem 1 is merely a sufficient condition rather than a necessary one. We illustrate by an example whose data are taken from Gaertner (2009, P.110). Consider $X=\{x_1,x_2,x_3\}$, and the evaluation data under 5 criteria are given as:
$$\begin{array}{ll}
\text{Criterion 1}: & x_1 \succ_1 x_2 \succ_1 x_3,\\
\text{Criterion 2}: & x_1 \succ_2 x_2 \succ_2 x_3,\\
\text{Criterion 3}: & x_2 \succ_3 x_3 \succ_3 x_1,\\
\text{Criterion 4}: & x_2 \succ_4 x_3 \succ_4 x_1,\\
\text{Criterion 5}: & x_3 \succ_5 x_1 \succ_5 x_2.
\end{array}
$$
The corresponding PPMs are
$$
\begin{array}{c}
{\begin{array}{cc}
{\begin{array}{c} x_1\\
x_2\\
x_3
\end{array}}&
{ \left[\begin{array}{ccc}
1 & 0 & 0 \\
0 & 1 & 0 \\
0 & 0 & 1
\end{array}\right],
\left[\begin{array}{ccc}
1 & 0 & 0 \\
0 & 1 & 0 \\
0 & 0 & 1
\end{array}\right],
\left[\begin{array}{ccc}
0 & 0 & 1 \\
1 & 0 & 0 \\
0 & 1 & 0
\end{array}\right],
\left[\begin{array}{ccc}
0 & 0 & 1 \\
1 & 0 & 0 \\
0 & 1 & 0
\end{array}\right],
\left[\begin{array}{ccc}
0 & 1 & 0 \\
0 & 0 & 1 \\
1 & 0 & 0
\end{array}\right].
}
\end{array}}
\end{array}
$$
The sum matrix is
$$
\begin{array}{c}
{\begin{array}{cc}
{\begin{array}{c} x_1\\
x_2\\
x_3
\end{array}}&
{
\left[\begin{array}{ccc}
2 & 1 & 2 \\
2 & 2 & 1 \\
1 & 2 & 2
\end{array}\right].
}
\end{array}}
\end{array}
$$
The conditions in Theorem 1 are not fulfilled. However, none will be whipped by `Rank and Yank' since the final aggregation outcome is a cycle, i.e., $x_1 \succ x_2 \succ x_3\succ x_1$.
In the end of this section we note that, according to Black's stratege II (Black 1976) or Young (1974) one can see that, when condition (4) is satisfied, the alternatives have identical Borda counts.
\section{Concluding remarks}
In this paper we posed a question (i.e., conditions for none to be whipped by Rank and Yank) in social choice theory which is independent of two much discussed topics regarding conditions on social preference transitivity and on social choice set existence. We set forth some sufficient conditions. In a sense, the topic of `none to be whiped by Rank and Yank' is related to the stability of the organization. One can see that the condition of dual probability relation (Eq.(3)) specifies a certain balance in the probabilities of alternatives being ranked at positions. It appears that the organization's stability is somewhat related to the balance exhibited in the probabilities of alternatives occupying ranking positions.
Researchers have recognized for over 40 years that the majority cycles are relavant to the stability of political systems (see, e.g., Miller 1983, Ordeshook 1992, and many others). Therefore, the presented results have significance for the discussion of political stability in light of majority cycles. When the evaluation data are cardinal numbers, it is fairly simple to obtain the stability condition. But if the evaluation data are assumed of ordinal rankings, as discussed in this paper, things can be a bit more complicated. The presented conditions are merely sufficient for some particular specific cases. There are still many things that need further investigation. Future research might include, but not limited to, seeking more sufficient conditions, necessary conditions, necessary and sufficient conditions in certain circumstances, and their practical implicatioins to organization stability. In addition, the likelihood of the occurrence of `none to be whipped by Rank and Yank' under those assumptions made in this paper deserves further effort as well. We would like to leave them open to the readers.
\vspace{0.1cm}
|
\section{INTRODUCTION}
\label{sec:intro}
The SPHEREx mission will perform an all-sky spectrophotometric survey in the near-infrared (0.75 $\mu$m. - 5.0 $\mu$m.) to address major scientific questions consistent with all three major themes of NASA's astrophysics program. Specifically, SPHEREx will:
\begin{enumerate}
\item \textit{Probe the origin and destiny of our Universe.} By measuring galaxy redshifts over a large cosmological volume, SPHEREx provides constraints on the physics of inflation, the superluminal expansion of the Universe that took place $10^{-32}$ s after the Big Bang.
\item \textit{Explore whether planets around other stars could harbor life.} SPHEREx provides absorption spectra of early phase planetary systems in the Milky Way to determine the abundance and composition of biogenic ices.
\item \textit{Explore the origin and evolution of galaxies.} SPHEREx will permit a precise determination of the power spectrum of spatial fluctuations in the extragalactic background light intensity through deep images formed near the ecliptic poles.
\end{enumerate}
SPHEREx instrument parameters are listed in table \ref{table:InstrumentParams}. Previous SPIE proceedings \cite{Korngut18} and \cite{Crill20} provide detailed information about the scientific rationale, hardware configuration, and mission design of SPHEREx.
By mid-2022, SPHEREx is in its optical payload integration and testing phase to deliver several laboratory calibration and instrument characterization measurements. Two such measurements are discussed in this paper: absolute spectral and focus calibration. To accurately determine photometric redshifts and resolve ice absorption features in the spectra of planetary systems, precise spectral bandpass profiles for all 25 million pixels in SPHEREx must be determined through spectral calibration. To maximize the fraction of background sky pixels and enable source masking to deeper flux levels for extragalactic background light power spectrum measurements, the SPHEREx point-spread-function (PSF) must be minimized through focus calibration. A system capable of performing these characterization/calibration measurements must meet the requirements listed in Table \ref{table:ControlReqs}.
\begin{table}[h]
\caption{SPHEREx Instrument Parameters}
\label{table:InstrumentParams}
\begin{center}
\begin{tabular}{|l|l|}
\hline
\rule[-1ex]{0pt}{3.5ex} \textbf{Parameter} & \textbf{Value} \\
\hline
\rule[-1ex]{0pt}{3.5ex} Effective Aperture & 20 cm \\
\hline
\rule[-1ex]{0pt}{3.5ex} Pixel Size & 6.2" x 6.2" \\
\hline
\rule[-1ex]{0pt}{3.5ex} Field of View & 2 x (3.5\textdegree x 11.3\textdegree); dichroic\\
\hline
\rule[-1ex]{0pt}{3.5ex} Resolving Power and Wavelength Coverage &
\begin{tabular}{l}
$\lambda = 0.75 - 2.42 \mu$m; $R=41$ \\
$\lambda = 2.42 - 3.82 \mu$m; $R=35$ \\
$\lambda = 3.82 - 4.42 \mu$m; $R=110$ \\
$\lambda = 4.42 - 5.00 \mu$m; $R=130$ \\
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} Arrays &
\begin{tabular}{l}
3x Teledyne Hawaii-2RG 2.5 $\mu$m. \\
3x Teledyne Hawaii-2RG 5.3 $\mu$m. \\
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} Cooling & All-Passive \\
\hline
\rule[-1ex]{0pt}{3.5ex} 2.5 $\mu$m. Array Temperature & $<80K$ \\
\hline
\rule[-1ex]{0pt}{3.5ex} 5.3 $\mu$m. Array Temperature & $<55K$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[h]
\caption{SPHEREx characterization/calibration system requirements.}
\label{table:ControlReqs}
\begin{center}
\begin{tabular}{|l|l|}
\hline
\rule[-1ex]{0pt}{3.5ex} \textbf{Requirement} & \textbf{Description} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 1) Instrument driver framework & \begin{tabular}{l}
A convenient abstraction away from low-level instrument \\
communication details.
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 2) Manual instrument control &
\begin{tabular}{l}
Interfaces to control instruments in a \\
testbed to facilitate the setup of automated \\
measurements.
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 3) Measurement automation &
\begin{tabular}{l}
A script based framework for specifying automated \\
measurements.
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 4) Real-time data visualization and processing & \begin{tabular}{l}
Live plots of multi-channel time-stream data and image \\
feeds. Basic real-time processing such as averaging \\
and histogram display.
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 5) Data archival tools & \begin{tabular}{l}
Mechanisms to archive data products along with \\
associated metadata in a variety of file formats.
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 6) Configurability &
\begin{tabular}{l}
A method to reconfigure all aforementioned \\
components for a variety of measurements.
\end{tabular} \\
\hline
\end{tabular}
\end{center}
\end{table}
The development of a software system meeting the Table \ref{table:ControlReqs} requirements is a problem that is not unique to SPHEREx. Modern experimental laboratories everywhere are faced with the similar challenge of developing such a system. Despite this, there are limited options when it comes to standard, well tested data acquisition and instrument control packages providing these capabilities. Familiar to many is NI's LabView \cite{LabView}. LabView has the advantage of being well tested, widely adopted, and with the ample support expected from any large-scale proprietary software system. However, integrating custom hardware with LabView can be cumbersome and the graphical programming environment lacks the flexibility that a true general purpose programming language provides. In the open source landscape, the PyMeasure project \cite{PyMeasure} provides a Python based measurement automation package. PyMeasure contains a large repository of existing instrument drivers along with a convenient framework for adding new instruments. Measurement automation tasks are easily achieved and measurement control logic is specified through basic Python scripts. In addition, PyMeasure has gathered a very active development community to support further improvements to the package. However, \emph{.csv} is currently the only output file format supported and the package does not have a convenient mechanism for manual instrument control. PyHK \cite{PyHk} is an additional Python based open source data acquisition package with adoption in astrophysics and astronomy instrumentation laboratories such as the BICEP \cite{BICEP} and TIME \cite{TIME} collaborations. Initially developed at Caltech for control and monitoring of cryogenic instruments, PyHK provides an excellent web-based user interface for manually controlling instruments, as well as viewing and interacting with housekeeping data stored over long periods of time. PyHK also contains a robust configuration file based interface for specifying the instruments in use and to customize a web interface for a given experiment. PyHK's limitations come from the fact that the instrument driver framework is not as robust as PyMeasure's and plain text is the only supported output file format.
Following the evaluation outlined above, it was determined that a custom data acquisition and instrument control system would best suit the needs of SPHEREx characterization/calibration. \textbf{SPHERExLabTools}, hereafter referred to as SLT, is the resulting system. SLT meets all of the Table \ref{table:ControlReqs} requirements and presents a framework for a general purpose laboratory data acquisition and instrument control system. SLT adopts the same instrument driver framework and method of specifying automated measurements as PyMeasure while providing a configuration file based interface that is similar to PyHK. In addition, SLT addresses the limitations of PyMeasure and PyHK by providing support for multiple output file formats, while integrating both manual and automated measurement control mechanisms.
The structure of the remainder of this paper is as follows. Section \ref{section:SLT Architecture} describes the architecture and configuration mechanism of SLT, sections \ref{section:Spectral Cal} and \ref{section:Focus Cal} describe the SPHEREx spectral and focus calibration measurements and illustrate how SLT is configured for each measurement. Initial calibration results obtained using SLT are also presented in these sections. Finally, section \ref{section:Conclusion} provides concluding remarks and discusses plans for further development of the SLT system.
\section{SLT Architecture} \label{section:SLT Architecture}
\subsection{Implementation Language and Class Structure}
SLT is implemented in pure Python. This decision was motivated by the recent wide adoption of Python in many scientific communities and the existence of the powerful Python data ecosystem including packages like NumPy \cite{harris2020array}, SciPy \cite{SciPy}, and Pandas \cite{mckinney-proc-scipy-2010}. The implementation of SLT in Python allows us to utilize this ecosystem while working in an environment familiar to many scientists involved in SPHEREx and beyond. In addition, Python's object-oriented nature facilitates the development of organized modular code for which each requirement of Table \ref{table:ControlReqs} can be met through independent configurable modules and/or classes. The SLT class based modular architecture is summarized in Figure \ref{fig:SLT_Architecture} while section \ref{sect:ClassDescription_UI} describes in more detail all of the classes outlined in this figure.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth]{fig/SLTDataFlow.png}
\caption{The SLT modular architecture consists of 6 classes which interact in the following manner: \textbf{1) Procedures} contain scripts which define the logical flow of a measurement; they interact directly with \textbf{2) Instruments} and generate data which they send to \textbf{3) Viewers} for real time graphical display and/or \textbf{4) Recorders} to archive data. Recorders and Viewers sit idle until data is placed on their associated queue, upon which they perform the appropriate archival or display task. \textbf{5) Controllers} provide graphical interfaces to interact directly with Instruments or to start/stop Procedure execution. Since both Controllers and Procedures have the ability to interact directly with Instruments, a locking mechanism is placed on Instruments such that only one controller/procedure thread of execution can access any given instrument driver class at a single instant. Finally, the \textbf{6) Experiment} class provides a top-level wrapper to interpret user configuration files and create instances of all of the previous classes. SLT is a multi-threaded application so many instances of each of these components can execute simultaneously. The color-scheme in this figure indicates the threading structure where each color corresponds to a separate thread of execution.}
\label{fig:SLT_Architecture}
\end{figure}
\subsection{Experiment Control Packages and Configurability}
SLT focuses heavily on configurability to support the diversity of measurements and instruments involved in SPHEREx testing. The manner in which SLT is configured is through the specification of \textbf{\emph{Experiment Control Packages}}. Experiment control packages are Python packages that define the 5 configuration variables described in Table \ref{table:ConfigVariables}. In brief, these configuration variables are read by the SLT Experiment class (6) of Figure \ref{fig:SLT_Architecture}) to command SLT which instrument drivers and graphical interfaces to instantiate, what the output data formats should be, and where to find user-defined measurement automation scripts (i.e. Procedures, sect. \ref{sect:Procedures}). Table \ref{table:ConfigVariables} describes the purpose of each of these configuration variables, while Tables \ref{table:ProcConfig} through \ref{table:CntrlConfig} specify each variable's appropriate contents.
\begin{table}[H]
\caption{SLT Experiment control package configuration variables. Each variable is a python list of \textbf{configuration dictionaries} (see tables \ref{table:ProcConfig} - \ref{table:CntrlConfig}). The numbered elements of Figure \ref{fig:SLT_Architecture} correspond to the variable numbers in this table.}
\label{table:ConfigVariables}
\begin{center}
\begin{tabular}{|l|l|}
\hline
\rule[-1ex]{0pt}{3.5ex} \textbf{Variable} & \textbf{Description} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 1) PROCEDURES &
\begin{tabular}{l}
List of \textbf{procedure configuration dictionaries} (Table \ref{table:ProcConfig}) pointing SLT to the \\
user defined measurement automation scripts i.e. (Procedure classes).
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 2) INSTRUMENT\_SUITE &
\begin{tabular}{l}
List of \textbf{hardware configuration dictionaries} (Table \ref{table:HardwareConfig}) telling SLT \\
which instrument drivers should be instantiated.
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 3) VIEWERS &
\begin{tabular}{l}
List of \textbf{viewer configuration dictionaries} (Table \ref{table:ViewConfig}) defining data visualization \\
tasks.
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 4) RECORDERS &
\begin{tabular}{l}
List of \textbf{recorder configuration dictionaries} (Table \ref{table:RecConfig}) to specify the output file \\
formats that data generated from procedures should be saved to.
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 5) CONTROLLERS &
\begin{tabular}{l}
List of \textbf{controller configuration dictionaries} (Table \ref{table:CntrlConfig}) identifying the \\
instruments and procedures for which graphical control interfaces \\
should be generated.
\end{tabular} \\
\hline
\end{tabular}
\end{center}
\end{table}
\input{config_dicts}
\subsection{Class Descriptions and User Interface} \label{sect:ClassDescription_UI}
\subsubsection{Procedures}
\label{sect:Procedures}
The central element of both the PyMeasure and SLT packages is the \textbf{Procedure} class. As mentioned in Figure \ref{fig:SLT_Architecture}, \textbf{Procedures} are used to define the logical flow of a measurement. Procedures contain the scripts which interact directly with instruments, collect and process data, and send data to Viewers and Recorders. The basic working principle of a procedure in PyMeasure and SLT is the same with SLT modifying the manner in which procedure output data is handled. Procedures in PyMeasure and SLT contain the following components which are illustrated in the flowchart of Figure \ref{fig:ProcedureExecution}:
\begin{enumerate}
\item \textbf{Parameter Objects:} Classes defining the configurable elements of a Procedure. Parameter objects map directly to the graphical elements found in \textit{Procedure Controllers} (sect. \ref{sect:Control}) so that they can be set and modified via a graphical interface.
\item \textbf{startup():} The initial method which executes first when a Procedure is started. In general, startup() is used to perform the initial configuration of instrumentation in a testbed before the real measurement begins.
\item \textbf{execute():} This method always contains the main body of the measurement control code and executes after startup(). This is where data is collected, processed, and sent out for display and archival.
\item \textbf{shutdown():} The final method executing after execute(). Here, all instrumentation is placed in an idle state before the next measurement is run. \\
\item \textbf{emit():} This method is used to send data out for display and archival. In PyMeasure, Python's built-in \textbf{logging} package is used to write data records sent out by emit() to csv log files. Since the logging package only supports plain-text data, the output file formats supported by PyMeasure are inherently limited. In SLT we modify this mechanism to support multiple output file formats and to separate data visualization from data archival. As illustrated in Figure \ref{fig:SLT_Architecture}, all Viewers and Recorders have associated \textbf{queue} objects. When emit() is called in SLT, two arguments are provided: The first is a string that identifies which queue objects data should be sent to and the second is the data itself. Copies of the data are then generated and placed onto every desired queue object. When the Viewer and/or Recorder associated with a given queue object detects that new data is present, the data is retrieved from the queue and the appropriate display/archival task is performed. This approach has the advantage that python queue objects have no type preference so any data type can be input to the queue. Accordingly, data of all shapes and sizes can be supported so long as the appropriate Viewer/Recorder class to handle the data has been implemented. In addition, queues in SLT are independent between each instance of a Viewer or Recorder. This allows information to be displayed in a Viewer but not necessarily saved out with a Recorder and vice/versa. Or, multiple different Viewers and Recorders can be configured to process the same output data. An argument against this approach is that each time emit() is called, copies of the data object must be created for every queue so for large data objects, CPU memory could become a limiting factor. However, with modern CPU memory sizes this limitation effectively does not exist for the vast majority of practical SLT applications (certainly all applications in SPHEREx testing).
\end{enumerate}
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{fig/ProcedureExecution.png}
\caption{SLT Procedure Execution and Structure: Procedures in SLT and PyMeasure operate with the same basic principle, but with different mechanisms for how output data is handled. Shown here is a basic example procedure for the collimator calibration measurement (sect. \ref{section:Focus Cal}). In this procedure, the startup() method moves the microscope into position and opens or closes the shutter. In execute() images are recorded from the ccd camera and the emit() method is used to send images out to the "camera\_stream" recorder and viewer queues. A Procedure Controller is generated to allow measurement operators to configure the microscope position, shutter state, and exposure time.}
\label{fig:ProcedureExecution}
\end{figure}
\subsubsection{Instruments} \label{sect:Instruments}
SLT adopts the same instrument driver framework as PyMeasure. This allows seamless integration of the extensive repository of existing PyMeasure drivers into SLT. Additionally, users of SLT can develop drivers for instruments not already supported by SLT or PyMeasure by following the detailed instructions in the \emph{Adding Instruments} section of the PyMeasure documentation \cite{PyMeasure}.
\subsubsection{Viewers and Recorders} \label{sect:View_Rec}
All viewers and recorders sit idle until data is placed on their associated queues, upon which they retrieve the data object and perform their archival or display task. To date, Viewer classes have been developed for multi-channel timestream and 2-dimensional image data display. Recorder classes to support data output to CSV, HDF5, and MAT file formats as well as Structured Query Language (SQL) databases exist.
\subsubsection{Controllers} \label{sect:Control}
SLT implements two main types of controllers: \textbf{Instrument Controllers} and \textbf{Procedure Controllers}. Instrument Controllers couple directly to instrument drivers and allow manual control of any instrument supported by SLT and PyMeasure through graphical interfaces. Procedure Controllers are used to execute individual procedures as well as several procedures in a sequence. Users can set and modify the \textbf{Parameter Objects} for a given procedure and create \textit{nested for loops} where the parameters of the procedure are modified at each step of the loop.
\subsubsection{User Interface} \label{sect:Interface}
Shown in Figure \ref{fig:SltInterface} is the SLT user interface generated with the example \textit{Experiment Control Package} described in section \ref{section:Focus Cal}. Note that all graphical elements of SLT are implemented using the PyQtGraph graphics and gui library \cite{PyQtGraph}.
\begin{figure}[H]
\centering
\includegraphics[width=\linewidth]{fig/InterfaceBorder.png}
\caption{SLT interface: \textbf{controllers} are embedded in the left panel, \textbf{viewers} in the center, while \textbf{procedures} and their data products are displayed in the right panel. \textbf{Log messages and errors} occurring throughout a measurement are displayed in the bottom panel. Drop-down menus in each panel allow users to switch between the various controllers, viewers, and procedures.}
\label{fig:SltInterface}
\end{figure}
\section{Example: SLT for spherex spectral calibration} \label{section:Spectral Cal}
The SPHEREx spectral calibration measurement aims to deliver per-pixel spectral bandpass response profiles for every pixel in the SPHEREx detectors. This measurement is important as precise knowledge of the SPHEREx spectral response is required to accurately measure galaxy redshifts and ice absorption features in the Milky Way. We accomplish SPHEREx spectral calibration by coupling a broadband light source to a monochromator whose output beam is collimated and input to the SPHEREx optics. The monochromator contains diffraction gratings whose diffraction pattern is reflected through a small slit such that a monochromatic light source of a desired wavelength is generated. The collimated, monochromatic light is then scanned across the entire SPHEREx band in less than 1 nm wavelength steps. SPHEREx detector exposures are recorded at each wavelength. Figure \ref{fig:SpectralCal} provides a block diagram of the measurement setup and Figure \ref{fig:SpecCalBBTestBed} shows the testbed for the spectral calibration measurement of a prototype focal plane assembly model. Table \ref{table:SpecCalConfig} describes the SLT \textit{experiment control package} configuration variables for the measurement while bandpass results for the prototype calibration are shown for a single pixel in Figure \ref{fig:SpecCalBandpass}.
\begin{figure}[H]
\centering
\includegraphics[width=0.7\linewidth]{fig/SpectralCal.png}
\caption{SPHEREx Spectral Calibration Measurement Schematic: Here, SLT drives the monochromator, a set of shutters including a cold shutter inside of the cryostat, and a neutral density filter attenuation wheel. SLT also communicates directly with the detector data acquisition system while logging the cryostat input optical power level and pressure/temperature housekeeping data. SLT uploads all metadata, housekeeping, and detector data to a central SQL server.}
\label{fig:SpectralCal}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.7\linewidth]{fig/SpecCalBBTestBed.png}
\caption{Prototype focal plane array spectral calibration testbed}
\label{fig:SpecCalBBTestBed}
\end{figure}
\begin{table}[H]
\caption{SLT experiment control package for spectral calibration}
\label{table:SpecCalConfig}
\begin{center}
\begin{tabular}{|l|l|}
\hline
\rule[-1ex]{0pt}{3.5ex} \textbf{Variable} & \textbf{Description} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 1) PROCEDURES &
\begin{tabular}{l}
\textbf{\textit{PowerLogProc:}} Continuously logs the input optical power level. \\
\textbf{\textit{SpecCalProc:}} Sets the monochromator wavelength, attenuation wheel \\
position, and communicates with the detector data acquisition system. \\
Calculates the average cryostat temperature and pressure. Sends all \\
instrument metadata, cryo housekeeping, and detector exposures to the \\ \textit{SpecCalSql} recorder.
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 2) INSTRUMENT\_SUITE &
\begin{tabular}{l}
Contains hardware configuration dictionaries for the monochromator, \\ attenuation wheel, power monitor detector readout, cryostat temperature \\ and pressure readout, and the detector data acquisition system.
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 3) VIEWERS &
\begin{tabular}{l}
\textbf{\textit{PowerView:}} Displays the time-stream input optical power level data \\ generated by the \textit{PowerLogProc} procedure.
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 4) RECORDERS &
\begin{tabular}{l}
\textbf{\textit{SpecCalSql:}} Writes data from \textit{SpecCalProc} to a SQL database.
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 5) CONTROLLERS &
\begin{tabular}{l}
Configuration dictionaries to generate manual instrument controllers for all \\ instruments in INSTRUMENT\_SUITE and procedure controllers for all procedures \\ in PROCEDURES.
\end{tabular} \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[H]
\centering
\includegraphics[width=0.65\linewidth]{fig/MonoResponse.png}
\caption{Spectral calibration bandpass results for a single pixel of the prototype focal plane assembly. The spectral information of this sample pixel is obtained by scanning the monochromator from 0.7 to 1.3 um at 2 nm wavelength steps. This 10 hour measurement procedure is performed in a fully automated manner by SLT, running the \textit{SpecCalProc} procedure described in Table \ref{table:SpecCalConfig}.}
\label{fig:SpecCalBandpass}
\end{figure}
\section{Example: SLT for spherex focus calibration} \label{section:Focus Cal}
The SPHEREx focus calibration quantifies telescope focus error by determining the axial displacement of the focal plane assembly hardware from the true focal plane of the telescope. Focus calibration is important to verify that the SPHEREx PSF is such that a sufficient fraction of background sky pixels and source masking can be achieved to produce a spatial power spectrum measurement of the extragalactic background light. We perform this calibration through two measurements: collimator calibration and focus error. The collimator calibration allows us to position a broadband pinhole light-source relative to an off-axis-paraboloid (OAP) such that we generate a collimated beam. We then steer the collimated beam into the SPHEREx optics and move the pinhole through a set of positions so that the beam deviates from collimation. We record SPHEREx exposures at each pinhole position and find the position at which the SPHEREx PSF is minimized. The difference between the initial pinhole position (at which the beam is collimated) and the final pinhole position (at which the SPHEREx PSF is a minimum) determines SPHEREx focus error. We perform the collimator calibration, i.e. positioning of a pinhole to generate a collimated beam, through the following procedure:
\begin{enumerate}
\item An LED is fiber coupled to a microscope mounted on a motorized stage.
\item A flat is placed inside the cryo-chamber in front of the SPHEREx optics to reflect incoming optical signal.
\item The LED source propagates through the collimating optics whose output beam is steered into the flat and reflected back into the collimating optics.
\item The reflected beam is re-imaged onto the surface of a CCD camera in the microscope and the spot size on the camera is calculated.
\item The microscope is scanned through several positions at which a camera spot size is calculated such that a microscope position vs. spot size curve is generated.
\item The minimum of a fit to the spot size vs. microscope position curve is taken to be the "best focus position" of the collimator. This is the position at which the pinhole source for the focus error measurement should be placed to generate a collimated beam.
\end{enumerate}
\begin{figure}[H]
\centering
\includegraphics[width=0.7\linewidth]{fig/FocusCal_CollimatorCalibration.png}
\label{fig:FocusCal_CollimatorCalibration}
\caption{\textbf{Collimator calibration measurement schematic.} Here, SLT drives the motorized microscope position and reads out this position. SLT drives a motorized relay to steer the beam into the cryo chamber, and implements a module for CCD camera readout and basic real-time processing. Cryo housekeeping data is logged and written to a SQL data server along with all position information and CCD camera images. Collimator focus curves can be generated on the fly to facilitate a rapid transition to the focus error measurement.}
\end{figure}
The \textit{SLT experiment control package} utilized for the collimator calibration measurement is described in Table \ref{table:CollimatorCalConfig} and an example collimator calibration focus curve generated by SLT is shown in Figure \ref{fig:Fcurve}. The collimating optics / microscope and motorized relay setup are pictured in Figures \ref{fig:Collimator} and \ref{fig:Relay} respectively.
\begin{table}[H]
\caption{SLT experiment control package for collimator calibration}
\label{table:CollimatorCalConfig}
\begin{center}
\begin{tabular}{|l|l|}
\hline
\rule[-1ex]{0pt}{3.5ex} \textbf{Variable} & \textbf{Description} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 1) PROCEDURES &
\begin{tabular}{l}
\textbf{\textit{CamViewProc:}} Continuously records frames from the CCD to send out \\
to the \textit{CamView} image viewer. \\
\textbf{\textit{CollimatorFocusProc:}} Drives the microscope and relay motors. Reads positions \\
and CCD camera images. Records average cryo temp/pressure. Sends images to \\
\textit{CamView} and images + position information to the \textit{CollimatorFocusRecorder}. \\
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 2) INSTRUMENT\_SUITE &
\begin{tabular}{l}
Contains hardware configuration dictionaries for the CCD camera, the microscope \\
drive motor and microscope position readout, the motorized relay drive motors, \\
cryostat temperature and pressure readout, and the detector data acquisition system.
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 3) VIEWERS &
\begin{tabular}{l}
\textbf{\textit{CamView:}} View live camera images recorded from \textit{CamViewProc} or \\
\textit{CollimatorFocusProc}
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 4) RECORDERS &
\begin{tabular}{l}
\textbf{\textit{CollimatorFocusRecorder:}} Write images + position information from \\
\textit{CollimatorFocusProc} to an HDF5 file.
\end{tabular} \\
\hline
\rule[-1ex]{0pt}{3.5ex} 5) CONTROLLERS &
\begin{tabular}{l}
Configuration dictionaries to generate manual instrument controllers for all \\ instruments in INSTRUMENT\_SUITE and procedure controllers for all procedures \\ in PROCEDURES.
\end{tabular} \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[H]
\centering
\includegraphics[width=0.6\linewidth]{fig/hspot_triangle_area_fcurve.png}
\caption{Collimator calibration focus curve: Black dots represent calculated beam sizes, the blue dash is the quadratic fit to the beam size data, and the red X is the minimum of the fit which is taken to be the "best focus position" of the collimator.}
\label{fig:Fcurve}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.7\linewidth]{fig/CollimatorInLab.png}
\caption{Microscope used in the collimator calibration measurement and off-axis paraboloid collimating mirror.}
\label{fig:Collimator}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.7\linewidth]{fig/Relay.png}
\caption{Motorized relay used to steer the beam into the SPHEREx optics to explore the whole field-of-view during focus calibration.}
\label{fig:Relay}
\end{figure}
\section{conclusion} \label{section:Conclusion}
We have presented a high-level overview of the SPHERExLabTools (SLT) data acquisition and instrument control package utilized in SPHEREx optics integration and testing. SLT builds upon and adapts the existing open source packages, PyMeasure \cite{PyMeasure} and PyHK \cite{PyHk} and presents a framework for a general purpose laboratory data acquisition and instrument control package. The effectiveness of SLT has been demonstrated through the delivery of early SPHEREx calibration data products, including the \textit{focal plane prototype spectral response} \ref{section:Spectral Cal} and \textit{collimator calibration} \ref{section:Focus Cal} measurements. SLT development is ongoing and the package continues to improve. An open source release along with work to integrate key features of SLT into the PyMeasure project are planned, with the long term vision of helping to establish an open-source standard in laboratory data-acquisition and instrument control.
\acknowledgments
This work was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration (80NM0018D0004).
|
\section*{Abstract}
In opinion dynamics, as in general usage, polarisation is subjective. To understand polarisation, we need to develop more precise methods to measure the agreement in society. This paper presents four mathematical measures of polarisation derived from graph and network representations of societies and information-theoretic divergences or distance metrics. Two of the methods, min-max flow and spectral radius, rely on graph theory and define polarisation in terms of the structural characteristics of networks. The other two methods represent opinions as probability density functions and use the Kullback–Leibler divergence and the Hellinger distance as polarisation measures. We present a series of opinion dynamics simulations from two common models to test the effectiveness of the methods. Results show that the four measures provide insight into the different aspects of polarisation and allow real-time monitoring of social networks for indicators of polarisation. The three measures, the spectral radius, Kullback–Leibler divergence and Hellinger distance, smoothly delineated between different amounts of polarisation, i.e.\ how many cluster there were in the simulation, while also measuring with more granularity how close simulations were to consensus. Min-max flow failed to accomplish such nuance.
\section*{Introduction}
Polarisation occurs in society when its members fail to agree on a topic or opinion. In the simplest case, polarisation is a bifurcation of a society into two sub-groups holding disjoint opinions on a given topic. Multi-modal polarisation, or plurality, occurs when individuals cluster their opinions around more than two poles or opinion loci. The opinion dynamics literature defines polarisation and consensus as two mutually exclusive states \cite{COmodelinganalysisP12017Proskurnikov}, in contrast to how we describe societies in terms of \emph{degrees} of polarisation or consensus on a continuum. Part of the reason for viewing polarisation in the opinion dynamics literature as discrete states is because the models used to describe societies and individual interactions are deterministic and view opinion points in space. This approach facilitates computation and analysis but doesn't accurately reflect the subtle variations in how individuals conceive of their opinions. The Martins model \cite{BSbayesianupdating2009Martins} conceptualises an individual's opinion as Gaussian density functions and naturally produces nuanced states of polarisation or consensus. Opinions distributed as functions on a continuum create the need for more sophisticated measures of polarisation or consensus than classification into one of two broad and possibly minimally informative categories. Such measures are the subject of this paper and will provide tools to explore how models, and ultimately real-world societies, fall into polarisation or reach consensus.
We introduce, in this paper, four methods to measure polarisation in a group of individuals. Two of these methods are derived from graph theory and use graph structure to determine the agreement in a simulation. Specifically, these methods examine the graphs representing information flow between agents, i.e. how much of an opinion can be transmitted between agents. The first is the ``min-max'' information flow between individuals, a continuous version of the $k$-edge connectedness of a graph. The second is to calculate the spectral radius of the graph. Finally, the last two methods are based on probability theory using the mean Kullbeck-Liebler (K-L) divergence and Hellinger distance between all individuals. We will compare these four methods over a series of simulations using two models for agent interaction with various initial conditions and parameter values. We discuss how effective these methods are at measuring the polarisation of the simulations and conclude with which one we think is the best in the most general circumstance.
\subsection*{Methods of Measuring Polarisation from the literature}
One of the most commmon methods used in the literature \cite{BCboundedconfidence2002Hegselmann, BCcontinuousdynamics2007Lorenz, CUtvaryopinions2015Yan, COfuzzyHK2021ZhaoAtel, COcirclaropinion2022HanEtal} to measure polarisation in an agent-based simulation model is to count opinion clusters once the simulation reaches a steady state. Counting opinion clusters is a simple method, reflecting the tendency to categorise a distribution of opinions as plurality, polarisation or consensus. Early works of \cite{BCboundedconfidence2002Hegselmann, BCcontinuousdynamics2007Lorenz} manually identified and counted opinion clusters, which can result in subjective and possibly biased results. Supervised and unsupervised clustering algorithms are available but present several technical challenges, and results can vary depending on the algorithm used. These challenges are exacerbated in more complex simulations with opinions (and clusters) represented as density functions.
We illustrate the fundamental problem of relying solely on cluster counts to measure polarisation or consensus for models representing opinion as a density function on a continuum using two cases.
First, consider the situation where a simulation in a steady-state produces $n$ opinion clusters with 85\% of the agents in a single cluster. Even though there are $n$ clusters (where $n$ could potentially be large), indicating polarisation, the fact that 85\% of agents are in a single cluster suggests consensus. Second, consider the case where there are multiple opinion clusters with approximately equal memberships, but the clusters are diffuse, and there is considerable overlap in the `distinct' opinions across clusters. This case also suggests consensus. Cluster counts do not tell a complete story for a model expressing opinions as density functions on a continuum.
A more advanced method of determining polarisation is \cite{Mrelativeagree2002Deffuant} which is used in the more modern work of \cite{Munifiedframework2018Coates}. This method measures polarisation accounting for the dispersion of agents across opinion space. Specifically, if $p_+$ and $p_-$ are the proportion of agents that finished the simulation in the upper and lower half of the opinion space, respectively, then \cite{Mrelativeagree2002Deffuant} defines
\begin{align}
y=p^2_++p^2_-,
\label{eqn:ystat}
\end{align}
as a measure of polarisation. Agents have reached a consensus when $y=0$ or $y=1$, meaning that the simulation has converged either to the centre $y=0$ or one extreme opinion space $y=1$. When $y=0.5$, an equal number of agents are at each extreme, indicating polarisation.
While Eq \ref{eqn:ystat} contains more information about polarisation, it is limited. Eq \ref{eqn:ystat} arbitrarily divides opinion space into two halves. Furthermore, the method does not distinguish between a densely packed consensus and a more diffuse consensus. A diffuse consensus in the upper half of the opinion space will give an equivalent value of $y$ to a denser-packed consensus. This lack of distinction is a problem as it is common for later models like \cite{BSbayesianupdating2009Martins} to create a very diffuse consensus.
\subsubsection*{Modern Measures of Polarisation}
A more recent approach to measuring polarisation comes from \cite{Mquantifypolar2021Musco}. Their focus was on quantifying polarisation through summary measures. They evaluated the effectiveness of various statistical measures, from the naive variance of opinions to their novel contribution to a measure that uses kurtosis and skew. Although relevant, we seek to differ from \cite{Mquantifypolar2021Musco} by presenting measures that have a closer connection to polarisation and are more physically meaningful.
Another novel approach originates from \cite{QPtwitterkavanaugh2019Darwish}, which measures polarisation through methods established in \cite{QPrandomwalk2018GuerraEtl}. The principle behind the measures is based on a random walk probability to travel from one cluster to another, with one method measuring the probabilities of random walks reaching one opinion cluster from another and the other calculating a `lower dimensional distance' between all agents and calculating the average inter and intra distance between clusters. With these measures \cite{QPtwitterkavanaugh2019Darwish} was able to more rigorously demonstrate that individuals in different opinion groups tended to consume different online media and used distinct hashtags. The method we present in this paper has the same potential to more distinctly identify the impacts of polarisation on society because our methods are continuous in a similar way as these probability measures.
A particularly novel approach to quantifying polarisation is to use machine learning algorithms \cite{QPmachinelearn2021Waller}. First communities were identified in an online space \cite{QPmachinelearn2021Waller} used a word-context learning algorithm which develops a vector of association between a particular context and all words. In the case of an online community, the `contexts' are individual users, and the `words' were the message boards the individuals could post on, so the more an individual posted to a message board, the more they became associated with the message board. With the communities identified \cite{QPmachinelearn2021Waller} sought to establish the social dimensions, the communities would lie upon and accomplished this though through a genetic learning algorithm which the authors seeded with the `Conservative' and `Democrat' message boards with these two message boards defining the ends of the social dimensions. With this social dimension, \cite{QPmachinelearn2021Waller} was able to determine how polarised communities were in the online space.
\subsection*{The Graph Theoretic Approaches to Measuring Polarisation}
Social network analysis uses graphs and graph theory to represent and explore social structures and connections between individuals. Opinion dynamics is the study of how individuals connect and influence each other's opinions. We can conceive of any opinion dynamics model as producing graphs representing the social connections and interactions between a group of individuals. Thus it is sensible to approach measuring polarisation using graph theory. A consequence of polarisation is the changes to the graph's structure representing a group's interactions, e.g.\ the isolation of groups of people with different opinions. As a result, the social networks of polarised compared with non-polarised groups have a starkly different graph structure. We can then use graph theory concepts, such as edge connectivity and spectral analysis, to quantify the difference in these networks, thereby quantifying the polarisation occurring. See \nameref{S1_Appendix} for a review of the graph theory concepts.
\subsubsection*{Polarisation Methods based on $k$-edge-connectivity}
One of the prevalent ideas discussed in the graph structure literature \cite{SNanalysis2008Butts} is the concept of connectivity or the degree of connections between individuals represented as nodes connected via edges. The simplest connectivity measure is graph density, defined as the number of edges in a graph compared to the maximum possible number of edges. Density, while convenient and scale-free, only gives the general local connectivity of a graph while neglecting the graph's global features, such as when a graph is partitioned into two or more components, i.e. when a society divides into two or more groups of individuals. The society would be polarised in that case, but if in those groups individuals have many edges between them, measures of density will rate the graph and society as highly connected.
Component connectivity is complimentary to density. We can derive a measure of component connectivity by noting intuitively that a graph with fewer components has greater ``connectivity". Consider a graph or social network $G$ consisting of $n$ individuals represented as nodes on the graph, if $G$ consists of $K(G)$ components then $(n-K(G))/(n-1)$ is a measure of component connectivity. As $K(G)\rightarrow 1$ then the components connectivity approaches $1$ and as $K(G)\rightarrow n$ the component connectivity approaches $0$. In terms of polarisation, as $K(G)\rightarrow n$ individuals are forming more and more disconnected sub-groups, i.e. ``bubbles" in the current colloquialism, and as $K(G)\rightarrow 1$ individuals form fewer components, i.e. are less polarised or more in consensus. Due to its focus on global characteristics, the proposed component connectivity measure does not consider the internal connectivity of components, so a weakly connected graph of size $n$ and a strongly connected graph of size $n$ can have the same measure of component connectivity \cite{SNanalysis2008Butts}. A more refined version of this component connectivity metric is in \cite{SNtheoreticaldim1994Krackhardt} but suffers from the same issue of evaluating weakly connected graphs as equivalent to strongly connected graphs \cite{SNanalysis2008Butts}. In terms of polarisation, component connectivity only measures the degree that society has partitioned itself into distinct components, not the measure of a component's internal cohesion or communication.
A more compelling measure of graph connectivity originates from the concept of cutsets. Consider a graph $G$ consisting of $V$ vertices or nodes and $E$ edges connecting the nodes, $G=[V,E]$, a cutset is a subset of the edges $H\subset E$ or nodes $H\subset V$ such that if we remove $H$, the number of disconnected components in $G$ increases. When $H$ is a subset of nodes, it is known as a vertex cut. Likewise, when $H$ contains edges, it is known as an edge cut \cite{SNanalysis2008Butts}. The minimum cutset of $G$ is the cutset of either edges or nodes with the smallest size. The larger a graph's minimum cutset is, the more ``connected'' the graph. Therefore, we can measure a graph's connectivity by finding $k$ the size of the minimum cutset. The size of the minimum edge cutset is the $k$-edge-connectedness of a graph, and likewise, the $k$-vertex-connectedness is the size of the minimum vertex cutset \cite{SNanalysis2008Butts}. In social network models, we assume that $V$, the nodes or individuals, are fixed, whereas edges are the connections between individuals, and connections are a direct measure of a society's divisiveness. For this reason, we consider $k$-edge connectivity a more reasonable and intuitive measure of polarisation.
\subsubsection*{Spectral Analysis}
One important concept about graphs is their adjacency matrix representations. An adjacency matrix $A$ of a graph $G$ has elements such that $A_{ij} = 1$ if edge $e_{ij}$ exists between agents $i$ and $j$, and $A_{ij} = 0$ when $e_{ij}$ does not exists. If $G$ is a weighted graph then $A_{ij} = f(e_{ij})$ where $f$ the weighting function. We can then investigate network properties using spectral analysis of a graphs adjacency matrix. Spectral analysis uses the eigenvalue decomposition of matrices to summarise and identify characteristics of the network. The spectral radius of a matrix is an important part of spectral analysis. Denoted by $\rho(A)$, the spectral radius of a matrix $A$ is its largest eigenvalue in magnitude. What is important about the spectral radius is its relationship to the connectivity of a graph. We can illustrate this relationship by considering a society that has polarised into $m$ distinct opinion clusters. In a modelling sense this happens when agents share information completely (100\%) inside a cluster and no information (0\%) outside the cluster. We can represent this as a graph that is composed of $m$ complete subgraphs (i.e. fully connected subgraphs). We can then express this graph as an adjacency matrix such that
$$ A = \begin{bmatrix}
\ones{n_1} & 0 & \ldots & 0\\
0 & \ones{n_2} & \ldots & 0\\
\vdots & \vdots & \ddots & \vdots \\
0 & 0 & \ldots & \ones{n_m}
\end{bmatrix}, $$
where $\ones{n_i}$ is the $n_i\times n_i$ matrix of ones and $n_i$ is the size of the $i$th opinion cluster. From Theorem \ref{thm:ones} in \nameref{S2_Appendix} the spectral radius of $A$ is the size of the largest opinion cluster.
It is clear how the largest opinion cluster's size relates to polarisation. If the size of the largest cluster is the total number of individuals in society, then that society is in consensus. So we can then use the fraction of individuals in the largest cluster to measure how close society is to consensus. Theorem \ref{thm:ones} shows that the spectral radius is the largest cluster size when a society is divided distinctly into opinion clusters. The main advantage of the spectral radius is that we can calculate the spectral radius even when opinion clusters aren't distinct and when there is a significant overlap between clusters. So the spectral radius offers us a method to estimate the size of the `largest cluster,' which allows us to use, more broadly, the largest cluster size as a measure for polarisation.
\subsection*{The Information Theoretic Approaches to Measuring Polarisation}
Some opinion dynamics models \cite{BSbayesianupdating2009Martins,BSbayesianupdating2012Martins} views opinions as probability distributions so it follows to use $f$-divergences as ways to quantify differences between two agents' opinions \cite{Fchisquare2014Nielsen,Ftutorial2004Csiszar}. The $f$-divergences measure distance between two probabilistic objects and are a `statistical distance'. For models that consider agent opinions as probability distributions, $f$-divergences like the Kullback–Leibler divergence and the Hellinger distance can provide insights into polarisation. As for models that don't consider agent opinions as probability distributions, like HK bounded confidence, we can interpret agent opinions in a probabilistic way.
\subsubsection*{Kullback–Leibler Divergence}
Kullback–Leibler divergence ($\mathrm{KLD}$) is a measure of the difference between two probability distributions \cite{KLinformationsufficiency1951KullbackLeibler}. The literature uses $\mathrm{KLD}$ to compare models of statistical inference for Bayesian statistics. The continuous version of the K–L divergence is
\begin{align}
\mathrm{KLD} (f||g) = \int_{-\infty}^\infty f(x)\log \left( \frac{f(x)}{g(x)}\right)\ \mathrm{d}x, \label{eqn:DKL}
\end{align}
where $f$ and $g$ are the probability density functions \cite{KLinformationsufficiency1951KullbackLeibler}. Note that $\mathrm{KLD}(f||g)\neq\mathrm{KLD}(g||f)$. The principle is to maximise the $\mathrm{KLD}$ of the posterior and prior distributions, which is equivalent to maximising over the likelihood in Bayesian statistics \cite{KLDbreview1995Chaloner,KLDcomfit2002Spiegelhalter}. Because of $\mathrm{KLD}$'s link to the likelihood in Bayesian statistics, it is sensible to use $\mathrm{KLD}$ as a measure of distance between agent's opinions in the Martins model \cite{BSbayesianupdating2009Martins} due to the model's reliance on Bayesian inference for generating polarisation. Taking the $\mathrm{KLD}$ between two agents' opinions would be treating one agent's opinion as a theoretical prior and the other's opinion as a theoretical posterior, and $\mathrm{KLD}$ would then reveal how much information is required for the prior agent to adopt the posterior agent's opinion. Therefore finding the $\mathrm{KLD}$ between all agents in a simulation to find the mean of all the inter-agent $\mathrm{KLD}$s, i.e.\ the mean $\mathrm{KLD}$, should reveal how `close' agents are in opinion, thereby revealing how polarised the simulated society is.
\subsubsection*{Hellinger Distance}
Similar to $\mathrm{KLD}$ the Hellinger distance is the distance between two probability density functions $f$ and $g$ \cite{Masymptoticsstatistics2000Lucien,Masymptoticstatistics2000Vaart,Mmeasureprob2001Pollard}, except Hellinger distance qualifies as a distance metric \cite{Masymptoticsstatistics2000Lucien} whereas $\mathrm{KLD}$ does not. The basis of the Hellinger distance is the Hellinger affinity \cite{Hellinger1909neue} defined as
$$
\int_X \sqrt{f(x) g(x)}\ \mathrm{d}x.
$$
When $f(x)=g(x)\ \forall x\in X$ the Hellinger affinity is 1 thus the squared Hellinger distance is
\begin{align}
H^2(f,g) = 1 - \int_{X} \sqrt{f(x) g(x)}\ \mathrm{d}x \label{eqn:Hdistance}
\end{align}
\cite{Masymptoticsstatistics2000Lucien,Masymptoticstatistics2000Vaart,Mmeasureprob2001Pollard}. We can interpret the Hellinger distance as the analogue of Euclidean distance from space vector but for probability distributions. Hence it follows to calculate the Hellinger distance between agents in the Martins model like we have suggested with the $\mathrm{KLD}$ and like with $\mathrm{KLD}$ we can find the mean Hellinger distance between every agent pair in a simulation to measure the polarisation of the simulation. Due to the Hellinger distance being a distance metric, the Hellinger distance has several advantages over $\mathrm{KLD}$, chief of which is the Hellinger distance's symmetry, i.e.\ $H(f,g) = H(g,f)$ which halves the number of computations when calculating the mean Hellinger distance.
\section*{Methods}
In this section, we shall discuss how the methods we used to determine the effectiveness of each measure of polarisation. We developed two types of simulations using two distinct models of agent interaction. In both types, we varied core parameters which influenced polarisation in the selected models. We then applied the four measures of polarisation to every simulation. Since these measures rely on the structural elements of the social network to measure polarisation, we shall have agents interact in an open `everyone can talk to every' environment to not bias toward polarisation.
\subsection*{Agent Interaction Models for the Simulations}
We used two interaction models in the simulations for this paper: the Martins model \cite{BSbayesianupdating2009Martins, BSbayesianupdating2012Martins, BSmistrust2021Adams} and the Hegselmann-Krause (H-K) Bounded Confidence model \cite{BCboundedconfidence2002Hegselmann}. The Martins model is the newest and generates complex behaviour in simulations. The H-K Bounded Confidence model is older than the Martins model, but the behaviour it produces in simulations is well understood.
The model first proposed in \cite{BSbayesianupdating2009Martins, BSbayesianupdating2012Martins}, which we shall call the Martins model, is a simple updating rule for agents derived from Bayesian inference. Because the model operates in a Bayesian framework, each agent's opinion is a guess at a true value. An agent's opinion follows a normal distribution where $x$ is the mean of that normal distribution, and $\sigma$ is the standard deviation. The value $x$ is the location of that agent's opinion, and the standard deviation $\sigma$ is an agent's uncertainty in their opinion, their strength of belief.
Eq \ref{eqn:MartinsOther} and \ref{eqn:Martins} describes the how agents update their opinion. Martins creates polarisation through the parameter $p$, where $p$ is the probability an agent shares useful information with another to update their opinion. Essentially $p$ is a global trust rate. The effect of including $p$ is that when two agents interact, their opinions are updated using $0<p^*<1$, which measures how much the two agents trust each other. $p^*$ is affected by how distant their opinions are relative to their uncertainties \cite{BSmistrust2021Adams}.
\begin{align}
&x_i(t+1) = p^* x_i(t) +(1-p^*) \frac{x_i(t)/\sigma_i(t) + x_j(t)/\sigma_j(t)}{1/\sigma_i(t) + 1/\sigma_j(t)}, \label{eqn:MartinsOther}\\
&\sigma_i^2(t+1) = \left(1 - \frac{\sigma_i^2(t)}{\sigma_j^2(t)+\sigma_i^2(t)}\right)\sigma_i^2(t) + p^* (1-p^*) \left(\frac{x_i(t) - x_j(t)}{1+\sigma_j^2(t)/\sigma_i^2(t)}\right)^2, \label{eqn:Martins}
\end{align}
where
$$
p^* = \frac{p \phi\left(x_i(t) - x_j(t),\sqrt{\sigma_i^2(t)+\sigma_j^2(t)}\right)}{p \phi\left(x_i(t) - x_j(t),\sqrt{\sigma_i^2(t)+\sigma_j^2(t)}\right) + 1-p},
$$
and
$$
\phi(\mu,\sigma) = \frac{1}{\sigma\sqrt{2\pi}} e^{\frac{-\mu^2}{2\sigma^2}}
$$
The Martins model, along with it's extension in \cite{BSmistrust2021Adams}, is a compelling explanation of polarisation with $p$ and $p^*$. It also produces novel behaviour. A good measure of polarisation might explain the model's behaviour. In \cite{BSbayesianupdating2009Martins, BSbayesianupdating2012Martins} the model used an unshared uncertainty assumption, where an agent could not share their $\sigma$, we use a variant of the model with that assumption relaxed \cite{BSmistrust2021Adams}.
The Hegselmann-Krause (H-K) Bounded Confidence model \cite{BCboundedconfidence2002Hegselmann} is a well-studied model. It was one of the first models to create polarisation reliably. Each agent has a continuous opinion $x$. An agent $i$ will update their opinion by first taking all agents' opinions in the interval $[x_i-\epsilon,x_i+\epsilon]$, where $x_i$ is the opinion of agent $i$, and $\epsilon$ is a parameter set by the model. Agent $i$'s new opinion is the mean of all the opinions in the interval. The H-K model will serve to calibrate the new polarisation metric.
\subsection*{Applying the Graph Theory Measures}
For the graph-theoretic measures to work appropriately, we need to establish graphs of the interpersonal connection between agents in a simulation. Specifically, we need the induced adjacency matrix of that graph representing the information flow between agents.
The Martins model has a built-in measure of an agent's ability to compromise with other agents, $p^*$. Practically $p^*$ represents how much an agent accepts the opinion of another agent, i.e. influenced by the agent. Thus, an adjacency matrix created from $p^*$ represents the information flow of all agents in a simulation.
The H-K Bounded Confidence model has a more intuitive adjacency matrix with entries equal to $1$ or $0$. If agent $i$ and $j$'s opinions are within $\epsilon$, then the of the $i$th row and $j$th column in the adjacency matrix is $1$ because agents within $\epsilon$ of each others' opinion will have maximum influence on each other due to the updating rules of the H-K Bounded Confidence model. For the same reason, the $i$th row and $j$th column in the adjacency matrix will be $0$ when agent $i$ and $j$'s opinions are outside $\epsilon$ distance of each other.
\subsubsection*{$k$-edge-connectivity/Min-max Flow}
The simplest method to calculate the $k$-edge-connection of a graph is to turn the problem into a series of maximum flow problems. We find the minimum of all the maximum flow problems, hence the min-max flow algorithm. The maximum flow problem is defined as follows. Let $G = [V,E]$ be a weighted digraph. Let there be a source vertex $s\in V$ and a sink vertex $t\in V$. The weight of each edge is its capacity $c \in \mathbb{R}$. A flow is a function $f:E\rightarrow \mathbb{R}$ which satisfies these conditions.
\begin{itemize}
\item Capacity constraint: The flow over an edge must not exceed its capacity $c$.
\item Conservation of flows: The flow entering a vertex must equal the flow leaving the vertex, excluding $s$ and $t$.
\end{itemize}
The maximum flow problem is to route as much flow from $s$ to $t$, which gives the maximum flow rate $f_{max}$. Algorithms to find $f_{max}$ are the Ford–Fulkerson algorithm \cite{MFffalgorthim1956FordFulkerson}, Dinic's algorithm \cite{MFsolution1970Dinitz} and push–relabel algorithm \cite{MFnewapproach1988GoldbergTarjan}. See \cite{MFnewapproach1988GoldbergTarjan} for a more extensive list of algorithms. In this paper we use the MatLab built-in {\tt flow} function to find $f_{max}$.
To apply a maximum flow algorithm to an unweighted and undirected graph $G$, we need to convert $G$ to a weighted digraph. We accomplish this by replacing every edge in $G$ with two directed edges. The two directed edges connect the two previously connected vertices. Lastly, we assign the capacity of the new directed edges to be one. We find the $k$-edge-connectivity of $G$ using maximum flow by first iterating over every pair of vertices. We set one vertex as the source and the other as the sink and find $f_{max}$ for that source and sink pair. The minimum of those $f_{max}$ will be the $k$-edge-connectivity of $G$.
The algorithm to find $k$-edge-connectivity follows from Menger's theorem, which is a special case of the max-flow min-cut theorem \cite{MFffalgorthim1956FordFulkerson}, stating that the number of edge independent paths between two vertices is equal to the minimum set of edge cuts that separate those two vertices. Therefore finding the minimum of the $f_{max}$ derived from the directed version of $G$ will result in $k$-edge-connectivity of $G$ \cite{MFtestinggraphconn1974Tarjan}. Noted in \cite{MFmultiterminal1961GomoryHu} we can improve the algorithm by fixing a vertex and finding the minimum of its maximum flows with all other vertices in the graph.
The adjacency matrices created by the Martins model \cite{BSbayesianupdating2009Martins,BSbayesianupdating2012Martins} are weighted digraphs. Applying the min-max flow algorithm will result in a meaningful connectivity measurement, hence polarisation measurement, even with a non-integer result for `$k$'. H-K bounded confidence model produces an undirected and unweighted graph and gives $k\in\mathbb{Z}$. So there will be no difficulties in applying the min-max flow. Although more efficient algorithms exist for finding $k$-edge-connectivity of an unweighted graph, for consistency, we will still use the min-max flow algorithm since these algorithms won't work on the Martins' adjacency matrix. See \nameref{S3_Appendix} for other potential methods to calculate min-max flow.
\subsubsection*{Largest Cluster size with Spectral Radius}
Determining the spectral radius was simple. After producing the adjacency matrix at a particular time in the simulation, we calculated the spectral radius of the adjacency matrix using the in-built Matlab function {\tt eig}.
\subsection*{Applying the $f$-divergences Measures}
The general approach with the $f$-divergences was to determine the pairwise divergences between all agents and then measure the mean $f$-divergence.
\subsubsection*{Kullback–Leibler Divergence}
Applying Kullback–Leibler divergence to the Martins model is simple. Since the Martins model considers opinions as normal distortions, finding the Kullback–Leibler divergence between two agents is finding the K-L divergence between two Gaussians $f$ and $g$. This simplifies Eq \ref{eqn:DKL} to
\begin{align}
\mathrm{KLD}(f||g) = \logof{\frac{\sigma_j}{\sigma_i}} + \frac{\sigma_i^2+\left(x_i-x_j\right)^2}{2\sigma_j^2} - \frac{1}{2}, \label{eqn:DKLnorm}
\end{align}
where $x_i$ and $\sigma_i$ are the mean and standard deviation for $f$, and $x_j$ and $\sigma_j$ are the mean and standard deviation for $g$ (for deviation see \nameref{S4_Appendix}). Using Eq \ref{eqn:DKLnorm}, we can calculate the $\mathrm{KLD}$ between all possible agent pairs and then calculate the mean $\mathrm{KLD}$. Agents in an H-K bounded confidence model simulation have very definitive opinions (i.e. places where they rank other opinions as 0), creating singularities in K–L divergence; thus, we can't use K–L divergence to measure the polarisation of those simulations. K–L divergence is thereby limited in its applicability which we discuss in the discussion section of this paper.
\subsubsection*{Hellinger Distance}
It is simple to apply Hellinger distance to the Martins model. Since every agent's opinion is essentially a normal distribution, we can find the squared Hellinger distance between two normal distributions, which is
$$
H^2 = 1 - \sqrt{\frac{2\sigma_1\sigma_2}{\sigma_1^2+\sigma_2^2}}e^{\frac{(x_1-x_2)^2}{4(\sigma_1^2+\sigma_2^2)}},
$$
where $x_1$ and $x_2$ are the means and, $\sigma_1$ and $\sigma_2$ are the standard deviations of two normal distributions \cite{Mhellinger2017Kitsos}. Applying the Hellinger distance to the H-K bounded confidence model is less trivial. Although agents don't have probability density functions for opinions, we can consider an agent's opinion as a uniform distribution over $[x_i-\epsilon,x_i+\epsilon]$. The Hellinger affinity of two agents' opinions will be the area of overlap between the two uniform distributions. The squared Hellinger distance is
$$
H^2(f,g)=\begin{cases}
\frac{x_1-x_2}{2\epsilon} \quad &\text{if} \ x_1-x_2 \leq 2\epsilon\\
1 \quad &\text{if} \ x_1-x_2 > 2\epsilon \\
\end{cases}
$$
where without loss of generality we assume that $x_1 > x_2$ (for deviation see \nameref{S5_Appendix}).
Like with $\mathrm{KLD}$, we can find the Hellinger distance between all agents and then find the mean of those Hellinger distances to use as a measure of polarisation. Finding the mean of these $f$-divergences provides a general perspective on the differences between agents, thus providing resistance to outlying agents.
\section*{Results}
This section presents the results of simulations from Martins and the H-K bounded confidence models for different initial conditions and model parameters and analysed polarisation using the measures based on the min-max flow rate, spectral radius, and mean $\mathrm{KLD}$ and Hellinger distance. We varied the parameter $\epsilon$ for the H-K bounded confidence model. We varied the initial $\sigma$ and fixed $p$ at 0.7 for the extended Martin's model. For each set of initial conditions, we ran 100 simulations, each consisting of $n=1000$ agents. Figs \ref{fig1} and \ref{fig2} shows a sample of simulation output for each initial condition.
\begin{figure}[!h]
\begin{adjustwidth}{-2.25in}{0in}
\includegraphics[scale=0.95]{Fig1.jpg}
\caption{{\bf Sample simulations' opinion shifts of agents through time for the HK bounded confidence model under different values of $\epsilon$.} \\
(A) A simulation with $\epsilon = 0.05$. (B) A simulation with $\epsilon = 0.1$. (C) A simulation with $\epsilon = 0.15$. (D) A simulation with $\epsilon = 0.2$. (E) A simulation with $\epsilon = 0.25$. (F) A simulation with $\epsilon = 0.3$.}
\label{fig1}
\end{adjustwidth}
\end{figure}
\begin{figure}[!h]
\begin{adjustwidth}{-2.25in}{0in}
\includegraphics[scale=0.95]{Fig2.jpg}
\caption{{\bf Sample simulations' opinion shifts of agents through time for the extended Martins Model under different initial uncertainty.} \\
(A) A simulation with an initial $\sigma = 0.5$. (B) A simulation with an initial $\sigma = 0.2$. (C) A simulation with an initial $\sigma = 0.14$. (D) A simulation with an initial $\sigma = 0.1$. (E) A simulation with an initial $\sigma = 0.05$.}
\label{fig2}
\end{adjustwidth}
\end{figure}
\subsection*{Cluster counting and $y$}
We calculated the cluster counts and $y$-statistic for every simulation to compare with the new methods we developed in this paper. Tables \ref{table1} and \ref{table2} show the mean cluster count for the simulations. Both tables reveal the relationship we expect to see between the cluster count and the parameter values. The cluster count for a simulation is inversely proportional to both the initial uncertainty for the Martins model and $\epsilon$ for the HK bounded confidence model. Figure \ref{fig3} shows the $y$-statistic for the simulations. We note that the $y$-statistic seems to categorise simulations as either polarised or in consensus. Interestingly in Figure \ref{fig3}B, the $y$-statistic can identify that simulations are closer to consensus for initial uncertainty of 0.14.
\begin{table}[!ht]
\centering
\caption{
{\bf The mean cluster count for each 100 simulations of the HK Bounded confidence under different values of $\epsilon$}}
\begin{tabular}{|l+l|l|l|l|l|l|}
\hline
$\mathbf{\epsilon}$ & 0.05 & 0.1 & 0.15 & 0.2 & 0.25 & 0.3\\ \thickhline
Mean Cluster Count & 7.52 & 3.74 & 2.64 & 1.99 & 1 & 1 \\ \hline
Standard Deviation & 0.6432 & 0.4845 & 0.4824 & 0.1 & 0 & 0 \\ \hline
\end{tabular}
\begin{flushleft} Cluster counts were found using the inbuilt Matlab function {\tt subclust}.
\end{flushleft}
\label{table1}
\end{table}
\begin{table}[!ht]
\centering
\caption{
{\bf The mean estimated opinion cluster count for each 100 simulations of the extended Martins model under different initial uncertainty}}
\begin{tabular}{|l+l|l|l|l|l|}
\hline
{\bf Initial Uncertainty} & 0.5 & 0.2 & 0.14 & 0.1 & 0.05 \\ \thickhline
Mean Cluster Count & 1 & 1 & 1.15 & 2.03 & 4.18 \\ \hline
Standard Deviation & 0 & 0 & 0.3589 & 0.1714 & 0.73 \\ \hline
\end{tabular}
\begin{flushleft} Cluster counts were found using the inbuilt Matlab function {\tt subclust}.
\end{flushleft}
\label{table2}
\end{table}
\begin{figure}[!h]
\begin{adjustwidth}{-2.25in}{0in}
\includegraphics[scale=1]{Fig3.jpg}
\caption{{\bf The $y$-statistic of simulation for the two different models and different parameter values.} \\
(A) HK bounded confidence simulations. (B) Extended Martins model simulations.}
\label{fig3}
\end{adjustwidth}
\end{figure}
\subsection*{The Min-Max flow/Edge Connectivity}
The min-max flow rate is a bounded measure of polarisation which diverges to $0$ at polarisation or $n-1$ at a consensus, where $n$ is the number of agents in the simulation. Finding the min-max flow rate or edge connectivity is the most computationally intense method of measuring polarisation, and measures of min-max flow rate over individual simulations are noisy, regardless of simulation size $n$. Thus averaging over the 100 simulations produced constant behaviour as an illustration of the method's utility.
Fig \ref{fig4} shows the edge connectivity of the H-K bounded confidence model at different values of $\epsilon$. For $\epsilon \leq 0.2$, the edge connectivity diverges to 0 at the steady-state, indicating that the resulting graph is disjoint, and the simulation has polarised. Note some outlying simulations where $\epsilon = 0.2$ converged to consensus.
\begin{figure}[!h]
\includegraphics[scale=1]{Fig4.jpg}
\caption{{\bf Min-max flow, $k$, through time of the simulations that used the H-K Bounded confidence model with different values of $\epsilon$.}}
\label{fig4}
\end{figure}
For values of $\epsilon > 0.2$, the edge connectivity increased to $n$, suggesting consensus (i.e. every node directly connects to every other node). The $\epsilon = 0.25$ took longer to reach consensus (five to ten iterations) compared to the three iterations when $\epsilon = 0.3$. The longer time for the $\epsilon = 0.25$ simulation to reach consensus suggests that when $\epsilon\approx 0.25$ simulations can transition between consensus and polarisation.
Fig \ref{fig5} shows the edge connectivity of the Martins confidence model at different values of initial $\sigma$. For initial $\sigma \leq 0.14$ edge connectivity decreased to 0, which contrasts with Fig \ref{fig4} where edge connectivity reached a local minimum. To further compound this difference, Fig \ref{fig4} shows that the larger values of $\epsilon$ bound the edge connectivity of smaller values. Such a pattern does not exist in Fig \ref{fig5}.
\begin{figure}[!h]
\includegraphics[scale=1]{Fig5.jpg}
\caption{{\bf Min-max flow, $k$, through time of the simulations that used the Martins model with different values of initial uncertainty.}}
\label{fig5}
\end{figure}
For initial $\sigma > 0.2$ edge connectivity increases to 1000, but at initial $\sigma = 0.2$ the progression to 1000 is not monotonic. Eventually, the edge connectivity of all simulations fell to $0$. This behaviour is consistent with the analysis from \cite{BSbayesianupdating2009Martins}, where the model, in the long term, was demonstrated to approach consensus arbitrarily close before fragmenting into opinion clusters. Edge connectivity decreases to a local minimum for the first few thousand interactions before increasing to 1000. Of note is that in the lower 25\% quartile of simulations, with initial $\sigma = 0.2$, edge connectivity continued decreasing and reached 0. Falling to a local minimum when $\sigma = 0.2$ suggests that $\sigma = 0.2$ is close to a bifurcation point between polarisation and consensus.
\subsection*{Spectral Radius}
Figure \ref{fig6} displays the behaviour of the spectral radius in the H-K bounded confidence model. In most cases, simulations reach a steady-state value of spectral radius after three iterations of the simulation, the exception being $\epsilon = 0.25$ and $\epsilon = 0.15$. At $\epsilon = 0.25$, the simulations first converge to a spectral radius of 500, but for some simulations, after 2 - 7 iterations, the spectral radius jumps to 1000. At $\epsilon = 0.15$ the simulations' spectral radius converges at either where $\epsilon = 0.1$ converges or $\epsilon = 0.2$ converges.
\begin{figure}[!h]
\includegraphics[scale=1]{Fig6.jpg}
\caption{{\bf Spectral radius through time of the simulations that used the H-K Bounded Confidence model with different values of $\epsilon$.}}
\label{fig6}
\end{figure}
Fig \ref{fig7} displays the spectral radius of the extended Martins model under several different parameterisations. The results show two phases of behaviour for the simulations' spectral radii. First, the simulations reach steady-state, and second, they begin to fragment in their opinion clusters. There is little variability for initial $\sigma$ of 0.5 or 0.2, where the simulations reached consensus, except when the simulations enter the second phase, where the spectral radii vary greatly. At initial $\sigma$ 0.1 and 0.05, the simulations polarised and took longer to reach the second phase. In the second phase, all simulations seem to drop between the same value in spectral radius, between 150 and 350. Potentially there is some structure to how agents fragment in the extended Martins model.
\begin{figure}[!h]
\includegraphics[scale=1]{Fig7.jpg}
\caption{{\bf Spectral radius through time of the simulations that used the Martins model with different values of initial uncertainty.}}
\label{fig7}
\end{figure}
What is of particular interest is the results for initial $\sigma = 0.14$. In \cite{BSbayesianupdating2009Martins}, initial $\sigma = 0.14$ was the critical value at which the simulations polarised, when one opinion cluster turned into two. The results in Fig \ref{fig7} show that at initial $\sigma = 0.14$, rather than dividing into two even opinion clusters (as implied in \cite{BSbayesianupdating2009Martins}), a small portion of agents (approximately $10$ - $100$) break away to form their cluster while the rest remain in consensus. The spectral radius highlights the continuous transition between consensus and polarisation while counting opinion clusters obfuscates this behaviour, although the $y$-statistic and averaging over 100 simulations can identify this behaviour.
\subsection*{Mean K-L Divergence}
Fig \ref{fig8} shows the behaviour of the mean K-L divergence in the Extended Martins Simulations. During the beginning of the simulations, K-L divergence inversely correlates with the initial $\sigma$. Later in the simulation, K-L divergence grows exponentially. There are two phases to the mean K-L divergence in Fig \ref{fig8}, similar to the spectral radius. In the first phase, K-L divergence grows at a fixed exponential rate which we observe as a linear trend in Fig \ref{fig8}. The next phase has that fixed exponential growth rate decrease. These phases are present in all of the simulations. The transition between the first and second phases appears to happen at the same amount of divergence. For $\sigma \geq 0.2$, there is no exponential growth until the simulations start fragmenting, and then the simulation seems to pick a random rate of exponential expansion.
\begin{figure}[!h]
\includegraphics[scale=1]{Fig8.jpg}
\caption{{\bf Mean K-L divergence through time of the simulations that used the Martins model with different values of initial uncertainty.}}
\label{fig8}
\end{figure}
\subsubsection*{Interpreting the exponential expansion}
There is a link between the rate of exponential expansion and the number of opinion clusters in a simulation since the exponential expansion rate correlates with initial $\sigma$ (except for simulations which reached consensus), which then is inversely correlated with the number of opinion clusters as seen in Fig \ref{fig1}. We shall now develop this further in this section.
The reason for the exponential growth of $\overline{\mathrm{KLD}}$ is because this term
\begin{align}
\frac{\sigma_i^2 + (x_i-x_j)^2}{2\sigma_j^2} \label{eqn:DKLgrowthterm}
\end{align}
in Eq \ref{eqn:DKLnorm}. In the late stages of the Martins model, when agents successfully interact, both agents will half their `variance' (uncertainty squared), causing the $\mathrm{KLD}$ they share with other agents to double. We can use this fact to estimate the number of clusters in a simulation by fitting a linear regression to the $\logof{\overline{\mathrm{KLD}}}$. Where $m$ is the gradient of the linear regression and $n$ is the number of agents in a simulation, we found the general expression for the effective estimated cluster count to be
\begin{align}
\widehat{\psi} = \frac{\logof{n+2}-\logof{n}}{m}. \label{eqn:psi}
\end{align}
\nameref{S6_Appendix} provides a more detailed deviation of Eq \ref{eqn:psi}.
Table \ref{table3} shows the result of applying Eq \ref{eqn:psi} for simulations that generated more than one opinion cluster. These estimates are close to the number of opinion clusters in Table \ref{table2} for their appropriate initial uncertainties. Moreover, the variance is significantly lower than in Table \ref{table2}.
\begin{table}[!ht]
\centering
\caption{
{\bf The mean estimated opinion cluster count for each 100 simulations under different initial uncertainty}}
\begin{tabular}{|l+l|l|l|l|}
\hline
{\bf Initial Uncertainty} & 0.14 & 0.1 & 0.05\\ \thickhline
Mean $\widehat{\psi}$ & 1.1055 & 1.91751 & 3.8231 \\ \hline
Standard Deviation & 0.1644 & 0.1773 & 0.5745 \\ \hline
\end{tabular}
\label{table3}
\end{table}
In practice, Fig \ref{fig8}A shows that simulations which reach consensus have no exponential growth, hence $m=0$, and the derivation of Eq \ref{eqn:psi} no longer applies.
\subsection*{Mean Hellinger Distance}
The mean Hellinger distance behaves similar to the spectral radius. The mean Hellinger distance results mirror the same features found in the spectral radius results. The mean Hellinger distance in general varied less across simulations.
Fig \ref{fig9} shows the mean Hellinger distance across the various H-K bounded confidence simulations. The mean Hellinger distance converges within five time steps of a simulation with expectations for $\epsilon = 0.25$, converging within ten time steps. As discussed in the spectral radius results section, $\epsilon = 0.25$ is close to a tipping point between polarisation and consensus. The mean Hellinger distance varies more when there is less polarisation, excluding when simulations reach consensus. At $\epsilon = 0.05$, there is little variation in mean Hellinger distance between simulations, whereas, at $\epsilon = 0.20$, there is more variation in mean Hellinger distance between simulations. We postulate that when a simulation splinters into many opinion clusters, the average distance between clusters remains constent, whereas, when there are only two clusters, those two clusters can be close or on opposite ends of opinion space.
\begin{figure}[!h]
\includegraphics[scale=1]{Fig9.jpg}
\caption{{\bf Mean Hellinger distance through time of the simulations that used the H-K bounded confidence model with different values of $\epsilon$.}}
\label{fig9}
\end{figure}
Fig \ref{fig10} shows the mean Hellinger distance across the various Martins simulation simulations. The mean Hellinger distance for Martins largely follows the behaviour of spectral radius for Martins. The mean Hellinger distance differs in one way from the spectral radius. No simulations reached a `consensus' with mean Hellinger distance measuring polarisation. The closest a simulation comes to consensus is a mean Hellinger distance of above $0.1$. Initially, simulations with initial $\sigma = 0.5$ began with a mean Hellinger distance below $0.1$, which converged to a value above $0.1$.
\begin{figure}[!h]
\includegraphics[scale=1]{Fig10.jpg}
\caption{{\bf Mean Hellinger through time of the simulations that used the Martins model with different values of initial uncertainty.}}
\label{fig10}
\end{figure}
\section*{Discussion}
This paper investigated four methods of determining polarisation, the min-max flow rate, spectral radius, the mean Kullbeck-Liebler divergence and the Hellinger distance. The spectral radius and min-max flow methods use graph theory concepts to form the basis for measuring polarisation. As a consequence, both have physical interpretations based on network topology. The min-max flow rate is the minimum number of paths between all vertex pairings. The spectral radius relates to the largest size of an opinion cluster, among other graph theory concepts such as the number paths length $k$. As a result, both measure different aspects of the connectivity of a graph, and through their different approaches, we can uncover more understanding of polarisation.
The min-max flow rate is limited as a dichotomous measurement of polarisation and is useful when the network is not in a steady state for tracking the trajectory of polarisation. The min-max flow rate is a non-robust measure, as it is sensitive to outliers. We would need to prune real-world networks to remove outliers to use the min-max flow rate as a measure of polarisation. The advantage of the min-max flow rate is that its definition is conceptually straightforward to interpret. Spectral radius is more versatile than min-max flow rate and is more robust to outliers, making it more practical to apply to real-world networks. As a non-dichotomous measure, the spectral radius can be used to monitor trajectories of polarisation and as a comparison between networks in a stead-state. The drawback of using spectral radius as a measure of polarisation is that its derivation is more challenging and esoteric.
The Kullbeck-Liebler divergence (KLD) is an information-theoretic measure of information loss (or gain). The KLD measures the divergence between two probability density (or mass) functions. In the Martins model, individual opinions are represented as probability density functions, leading to the KLD as a natural measure to consider as the difference or distance between opinions.
The mean KLD is the average of all pairwise KLD for a network. It is important to note that, unlike the min-max flow rate and spectral radius, the mean KLD doesn't account for network topology (min-max flow rate and spectral radius make use of information about the individuals' opinions via the emergent social network structure). Therefore, the mean KLD is limited to circumstances where opinions are probability distributions. What limits the mean KLD further is that it can't handle definite probability distributions. It is impossible to meaningfully calculate the mean KLD for the HK bounded confidence model because it requires the calculation of the KLD between two uniform distributions resulting in an infinite KLD.
We noted that the exponential growth rate of the KLD (Eq \ref{eqn:meanDKL}) after the simulation achieved steady-state is inversely proportional to the number of clusters at steady-state. Thus the slope of the log-linear model of a simulation, $m$, can be used as a measure of polarisation by producing an effective cluster count. The effective number of clusters $\widehat{\psi}$ from \eqref{eqn:psi} agrees with the spectral radius interpreted polarisation, including the measured cluster counts in Table \ref{table2}. For example, given a simulation with $\sigma = 0.14$, $\widehat{\psi} = 1.1055$ indicating that the agents are ``mostly'' in consensus, but that there is still some disagreement. The conclusions of this measure match the conclusions drawn from the spectral radius and what we observe in Fig \ref{fig2} and Table \ref{table2}, where simulations converge to (``mostly'') a single cluster. Compared with the spectral radius, which describes the size of the largest cluster, $\widehat{\psi}$ is more comprehensive as it describes the effective number of clusters for the simulation at steady-state. A weakness of $\widehat{\psi}$ is that it is ill-defined when the actual cluster counts $\psi = 1$. As seen in Fig \ref{fig8} the two types of simulations that reached consensus exhibited no exceptional growth when they first reached steady-state, making $m=0$, which breaks Eq \ref{eqn:psi} making $\widehat{\psi}\rightarrow \infty$. Considering this instability only occurs when a simulation reaches consensus, it is easy to ignore since we can identify consensus visually. Although, simulation types close to consensus will inherit some instability since some of the simulations will fall into consensus through random chance. Of more pressing concern is that $\widehat{\psi}$ tied in with the Martins model and $\mathrm{KLD}$, which could limit the applicability of $\widehat{\psi}$ to more realistic situations if the Martins model does not reflect how individuals share opinion. Still $\widehat{\psi}$ hints at the possibility of a continuous extension to counting clusters.
The mean Hellinger distance is similar to the spectral radius. The only significant deviation from the spectral radius is that the mean Hellinger distance never reached zero in the Martins simulations. Because the Hellinger distance is relative to uncertainty (Eq \ref{eqn:Hdistance}), like the Martins updating rules, it could determine that Martins simulations were never in complete agreement. This novel ability suggests that the Hellinger distance has an advantage over the spectral radius. Where the spectral radius might determine a group to be in complete consensus, the Hellinger distance can correctly determine that the group is not in complete consensus.
The spectral radius of the H-K Bounded confidence in Fig \ref{fig6} is consistent with the results in \cite{BCboundedconfidence2002Hegselmann}, showing that in the homogeneous case, the model forms uniformly spaced opinion clusters, with the space between them being greater than $\epsilon$. For the H-K bounded confidence model, the spectral radius reflects this discrete nature of opinion clusters, Fig \ref{fig6} shows the spectral radius converging to quantised values depending on $\epsilon$.
The spectral radius can track simulations when they fall between two quantised states. We can see in Fig \ref{fig1} that simulations with values of $\epsilon = 0.15$ and $\epsilon = 0.25$ have opinion cluster merge later in the simulation. This is reflected in the spectral radius through Fig \ref{fig6} as the spectral radius at $\epsilon = 0.15$ and $\epsilon = 0.25$ has more variability and stretch over the neighbouring values of $\epsilon$. So the spectral radius can determine bifurcation points in parameter values, i.e.\ parameter values at which two opinion clusters merge into one.
Of particular note is the late stage behaviour of Figs \ref{fig10}, \ref{fig8} and \ref{fig7}. The Martins model is known to fragment at the late stages of a simulation, but the fragmentation doesn't result in complete disunity (where every agent is isolated from all other agents). From Fig \ref{fig7} the simulations drop to a spectral radius between 200 and 400. It would be interesting to investigate the social network in the late stages of the Martins model.
\section*{Conclusion}
This paper has investigated four methods of measuring polarisation. We conclude that the min-max flow rate is the most insufficient method. The main advantage of the method is that it is intuitive. Although the min-max flow rate does reveal some dynamics as a simulation falls into either polarisation or consensus, simulations can only be in consensus or polarisation. Overall the method at most performs equivalently to the $y$-statisitc. The method is also extremely sensitive to the outlying agent, which is a problem in any real-world application. The major complication with the method is the computation time which makes the method less useful. Although, we briefly investigated a method to improve computational efficiency that resulted in a new measure of polarisation (see \nameref{S3_Appendix}).
In contrast, the spectral radius provided a complete picture of polarisation. The method's physical meaning is loose and difficult to understand, but we understand its meaning as the effective largest cluster size in a simulation, i.e.\ what the largest cluster would be if the simulation reached steady-state. As a result, the spectral radius places polarisation on a continuum and can identify when a simulation is close to reaching a consensus. Furthermore, the method can distinguish between different levels of polarisation, i.e.\ three cluster simulations from two cluster simulations. Essentially the spectral radius blends cluster counting methods and the $y$-statistic together. So we consider the spectral radius an effective at measuring polarisation.
The mean K-L divergence is interesting because the results were initially difficult to interpret. The mean K-L divergence diverged to infinity exponentially, but from the exponential expansion rate, we could determine the `number of clusters' $\psi$ of a particular simulation. The estimated cluster number agreed with what we observed with other measures and the raw results. It is clear that $\psi$ is what we can use to measure polarisation, and it is more intuitive to grasp compared to the spectral radius, but the measure has its drawbacks. First, it breaks when a simulation reaches consensus and second, it relies on the K-L divergence and the mechanics of the Martins model, which in real-world applications might not hold. Still, this method promises a way to express opinion cluster number as a continuous value which would be another avenue of research.
The mean Hellinger distance closely resembles the spectral radius but differs from the spectral radius in one crucial way. The Hellinger distance is more sensitive at complete consensus; it had more foresight into the Martins model degeneration from consensus into arbitrarily close opinion clusters than other methods. The mean Hellinger distance has an advantage over the spectral radius with its ability to detect the finer dynamics of the Martins model sooner.
Overall these four methods measure different aspects of polarisation, and the individual measures fail to capture the whole process of polarisation, but together they reveal the complete picture. Depending on the circumstances, certain methods might be more effective than others. The spectral radius is the most general and can be applied in most situations, whereas the mean Hellinger distance and K-L divergence work better when applied to their appropriate niches.
In this paper, we have only looked at simulated societies. Future research would involve applying these measures to real data sets. It is clear from the investigations in this paper that the mean K-L divergence, Hellinger distance and spectral radius hold the most promise
\section*{Supporting information}
\paragraph*{S1 Fig.}
\label{S1_Fig}
{\bf Alternate min-max flow compared with min-max flow for HK bounded confidence.}
\includegraphics[scale=1]{S1Fig.jpg}
\paragraph*{S1 Appendix.}
\label{S1_Appendix}
{\bf Basics of Graph Theory.} A graph $G$ consists of a set of vectors, which we will call nodes, $V$ where $v_i\in V$ is the $i$th node of $G$ and a set of edges $E$ where $e_{ij} \in E$ means that $v_i,v_j \in V$ are connected in $G$. We can define a path $P$ in $G$ to be a set of edges in $G$ such that if you begin at $v_i$ and follow the edges in $P$ you will finish at $v_j$ such that $P = \{e_{i1}, e_{12}, \ldots, e_{nj}\}$. A component is defined as the set of nodes $V_i\subset V$ in $G$ such that paths exist between all nodes of $V_i$, but no path exists to nodes outside of $V_i$. A graph $G$ is disjoint if $G$ has more than one component; likewise, $G$ is `connected' when only one component exists. A directed graph, digraph, is a graph $G$ that has its edges $E$ have a direction, i.e. $v_1$ connects to $v_2$, but $v_2$ doesn't necessarily connect back to $v_1$. A weighted graph is where for a graph $G$ there exists a function $f:E\rightarrow\mathbb{R}$.
\paragraph*{S2 Appendix.}
\label{S2_Appendix}
\begin{theorem}
Let $A$ be a square block diagonal matrix consisting of $m$ square matrices of ones $\ones{n_i}$ where $n_i$ is the dimension of the matrix of ones $i$, and $i=1,\ldots,m$. Then the spectral radius $\rho(A)$ is equal to the dimension of the largest unit matrix in $A$. \label{thm:ones}
\end{theorem}
\begin{proof}
Since $A$ is block diagonal, the eigenvalues of $A$ are the eigenvalues of $\ones{n_1},\ldots,\ones{n_m}$. We know that for a general square matrix of ones $\ones{k}$ \cite{Mcombinatorics2013Stanley} its characteristic equation is
$$
0 = \left(k - \lambda\right)\lambda^{k-1}.
$$
The dimensional values $n_1,\ldots,n_m$ are all eigenvalues of their respective matrix including multiple eigenvalues equal to $0$. Then $n_1,\ldots,n_m$ must be eigenvalues of $A$, the largest of which is $\rho(A)$, which is also the dimension of the largest matrix of ones in $A$.
\end{proof}
\paragraph*{S3 Appendix.}
\label{S3_Appendix}
{\bf Alternate min-max flow for HK bounded confidence.} To simplify the calculation of the min-max flow rate for the HK bounded confidence model, one might consider the following method:
\begin{enumerate}
\item For every agent in a simulation, count the number of agents within $\epsilon$ of the agent's opinion.
\item Minimise over those agent counts.
\end{enumerate}
Such a method counts the degree of each agent if you were to develop an adjacency matrix at that specific time in the simulation. Intuitively the method should be identical to min-max flow, but it produces different results when a simulation enters polarisation (see \nameref{S1_Fig} for an example). This method converges to the smallest cluster size instead of reaching zero. We thus consider this method distinct from the min-max flow method.
\paragraph*{S4 Appendix.}
\label{S4_Appendix}
{\bf Derivation of the KLD of two normal distributions.}
Let $X ~ \sim \mathcal{N}(\mu_i,\,\sigma_i^{2})$ and $Y \sim \mathcal{N}(\mu_j,\,\sigma_j^{2})$, and $f(x)$ and $g(y)$ describe the probability density function for $X$ and $Y$ then
\begin{align*}
\mathrm{KLD}(X,Y) &= \int_{-\infty}^{\infty}f(x) \logof{\frac{f(x)}{g(x)}}\ \mathrm{d}x \\
&= \int_{-\infty}^{\infty}f(x) \left[\logof{\frac{\sigma_j\sqrt{2\pi}}{\sigma_i\sqrt{2\pi}}e^{-\frac{\left(x-\mu_i\right)^2}{2\sigma_i^2}+\frac{\left(x-\mu_j\right)^2}{2\sigma_j^2}}}\right]\ \mathrm{d}x\\
&= \int_{-\infty}^{\infty}f(x) \left[\logof{\frac{\sigma_j}{\sigma_i}}-\frac{\left(x-\mu_i\right)^2}{2\sigma_i^2}+\frac{\left(x-\mu_j\right)^2}{2\sigma_j^2}\right]\ \mathrm{d}x \\
&= \begin{multlined}[t][10cm]\logof{\frac{\sigma_j}{\sigma_i}}\int_{-\infty}^{\infty}f(x)\ \mathrm{d}x - \frac{1}{2\sigma_i^2}\int_{-\infty}^{\infty}\left(x-\mu_i\right)^2 f(x)\mathrm{d}x\\ + \frac{1}{2\sigma_j^2}\int_{-\infty}^{\infty}\left(x-\mu_j\right)^2 f(x)\mathrm{d}x \end{multlined} \\
&= \logof{\frac{\sigma_j}{\sigma_i}}- \frac{1}{2\sigma_i^2}\Var{X} + \frac{1}{2\sigma_j^2}\int_{-\infty}^{\infty}\left(x^2-2\mu_j x+\mu_j^2\right) f(x)\ \mathrm{d}x \\
&= \logof{\frac{\sigma_j}{\sigma_i}}- \frac{1}{2} + \frac{1}{2\sigma_j^2}\int_{-\infty}^{\infty}\left(\left(x-\mu_i\right)^2+2x\left(\mu_i-\mu_j\right) +\mu_j^2-\mu_i^2\right) f(x)\ \mathrm{d}x \\
&= \begin{multlined}[t][10cm] \logof{\frac{\sigma_j}{\sigma_i}}- \frac{1}{2} + \frac{1}{2\sigma_j^2}\left[\int_{-\infty}^{\infty}\left(x-\mu_i\right)^2 f(x)\ \mathrm{d}x+2\left(\mu_i-\mu_j\right)\int_{-\infty}^{\infty}x f(x)\ \mathrm{d}x \right. \\ \left. +\left(\mu_j^2-\mu_i^2\right)\int_{-\infty}^{\infty} f(x)\ \mathrm{d}x\right] \end{multlined} \\
&= \logof{\frac{\sigma_j}{\sigma_i}}- \frac{1}{2} + \frac{1}{2\sigma_j^2}\left[\Var{X}+2\left(\mu_i-\mu_j\right)\E{X} +\mu_j^2-\mu_i^2\right] \\
&= \logof{\frac{\sigma_j}{\sigma_i}}- \frac{1}{2} + \frac{\sigma_i^2+2\left(\mu_i-\mu_j\right)\mu_i +\mu_j^2-\mu_i^2}{2\sigma_j^2} \\
&= \logof{\frac{\sigma_j}{\sigma_i}} + \frac{\sigma_i^2+\left(\mu_i-\mu_j\right)^2}{2\sigma_j^2} - \frac{1}{2}
\end{align*}
\paragraph*{S5 Appendix.}
\label{S5_Appendix}
{\bf Derivation of the H-distance of two uniform distributions.} Let $x_1$ and $x_2$ be the centers of two uniform distributions $f(x)$ and $g(x)$ both with width $2\epsilon$ and, without loss of generality, let $x_1 > x_2$. The Hellinger distance is
\begin{align*}
H^2(f,g) = 1 - \int_{-\infty}^\infty \sqrt{f(x) g(x)}\ \mathrm{d}x.
\end{align*}
There is two distinct cases for the Hellinger affinity. First is when there is no overlap, i.e.\ $x_1 - x_2 > 2\epsilon$, between $f$ and $g$ which means that the Hellinger affinity is zero and hence
\begin{align*}
\int_{-\infty}^\infty \sqrt{f(x) g(x)}\ \mathrm{d}x = 0,
\end{align*}
therefore
\begin{align*}
H^2(f,g) = 1.
\end{align*}
Second is when there is overlap, i.e.\ $x_1 - x_2 \leq 2\epsilon$, and the Hellinger affinity is non-zero. Specifically the Hellinger affinity will be the area of the overlap which is
\begin{align*}
\int_{-\infty}^\infty \sqrt{f(x) g(x)}\ \mathrm{d}x = \frac{2\epsilon + x_2 - x_1}{2\epsilon},
\end{align*}
therefore,
\begin{align*}
H^2(f,g) = \frac{x_1-x_2}{2\epsilon}.
\end{align*}
We can conclude that
$$
H^2(f,g)=\begin{cases}
\frac{x_1-x_2}{2\epsilon} \quad &\text{if} \ x_1-x_2 \leq 2\epsilon\\
1 \quad &\text{if} \ x_1-x_2 > 2\epsilon \\
\end{cases}.
$$
\paragraph*{S6 Appendix.}
\label{S6_Appendix}
{\bf Estimating cluster count from exponential, mean $\mathrm{KLD}$ growth.} Consider a Martins simulation that has reached steady-state, let $\Omega$ be the set of all agents in the simulation, and the simulation has divided into $\psi$ separate opinion clusters such that
$$\Omega = \bigcup_{k=1}^\psi A_k ,$$
where $A_k$ is a set of agents in the $k$th opinion cluster such that
$$\begin{aligned}
&A_k\cap A_l = \emptyset,\ \forall k,l \\
&\left|A_k\right| = \left|A_l\right|,\ \forall k,l
\end{aligned}$$
In steady-state, all agents in a Martins simulations have $\sigma \rightarrow 0$ . Then, according to Eq \ref{eqn:Martins}, agents will only achieve a $p^*=1$ when $x_i = x_j$ i.e. when two agent are in the same cluster. If $x_i \neq x_j$ i.e. when two agent are in the different clusters, then, with $\sigma \rightarrow 0$, $p^*=0$. Also, from Eq \ref{eqn:Martins}, when $p*=1$, agents will halve their $\sigma^2$, which means that Eq \ref{eqn:DKLgrowthterm} will double for a select proportion of agent pairs in the simulation, but not for all agent pairs. Meaning that $\overline{\mathrm{KLD}}$ will proportionally grow by a fixed amount $a$ after a single $p^*=1$ interaction
\begin{align}
\frac{\overline{\mathrm{KLD}}(t+s)}{\overline{\mathrm{KLD}}(t)} = a, \label{eqn:meanDKLPartRec}
\end{align}
where $t$ is an arbitrary number of interactions after the simulation has reached steady-state and $s$ is the number of interactions until a $p^*=1$ interaction occurs.
In a single $p^*=1$ interaction, two agents will halve their `variance' (uncertainty squared), doubling the $\mathrm{KLD}$ between those agents and every other agent in the simulation. All other pairings will maintain the same $\mathrm{KLD}$. The proportional growth of $\overline{\mathrm{KLD}}$, $a$, is dependent only on the pairwise agent's $\mathrm{KLD}$ where $\mathrm{KLD}\neq 0$.
We refer to agent pairings that have a $\mathrm{KLD} = 0$ as non-contributing and those with $\mathrm{KLD} > 0$ as contributing. Therefore,
$$a = (1-q)+2q = 1+q$$
where $q$ is the proportion of contributing pairings that double their $\mathrm{KLD}$. Let $n$ be the number of agents in the simulation and consider an agent inside a cluster, only $n(1-1/\psi)$ agents would generate contributing $\mathrm{KLD}$s since agents inside the hypothetical agent's cluster would generate a $\mathrm{KLD} = 0$. Since two agents will be interacting, we can double this agent count to get the total number $\mathrm{KLD}$s that double from two opinions updating, resulting in $2n(1-1/\psi)$. The total number of contributing pairings will be the total number of possible pairings, $n^2$, minus the non-contributing pairings, i.e.\ pairings which pair agents from the same clusters, $\psi (n/\psi)^2$. Therefore the total number of contributing pairings is $n^2(1-1/\psi)$. It follow then that $q = 2/n$. Thus
$$a = \frac{n+2}{n},$$
and Eq \ref{eqn:meanDKLPartRec} becomes
\begin{align}
\frac{\overline{\mathrm{KLD}}(t+s)}{\overline{\mathrm{KLD}}(t)} = \frac{n+2}{n}. \label{eqn:meanDKLRec}
\end{align}
Now consider a simulation which enters steady state after $t_0$ interactions, without loss of generality let $t_0 = 0$. Let $s_1$ be the number of interactions until the first $p^*=1$ interaction occurs, from Eq \ref{eqn:meanDKLRec} we know that
\begin{align*}
\frac{\overline{\mathrm{KLD}}(t_0+s_1)}{\overline{\mathrm{KLD}}(t_0)} &= \frac{n+2}{n}, \\
\overline{\mathrm{KLD}}(s_1) &= \frac{n+2}{n} K_0,
\end{align*}
where $K_0 = \overline{\mathrm{KLD}}(t_0)$. For the second $p^*=1$ interaction $s_2$, using Eq \ref{eqn:meanDKLRec} again, we can show
\begin{align*}
\frac{\overline{\mathrm{KLD}}(s_1+s_2)}{\overline{\mathrm{KLD}}(s_1)} &= \frac{n+2}{n}, \\
\overline{\mathrm{KLD}}(s_1+s_2) &= \left(\frac{n+2}{n}\right)^{2} K_0.
\end{align*}
So for the general $r$th $p^*=1$ interaction $s_r$ we have
\begin{align}
\overline{\mathrm{KLD}}\left(\sum_{i=1}^r s_i\right) = \left(\frac{n+2}{n}\right)^{r} K_0. \label{eqn:meanDKLPart1}
\end{align}
Let $t$ be the total number of interaction after $t_0$, then
$$
r = \frac{t}{\overline{s}},
$$
where
$$
\overline{s} = \frac{1}{r}\sum_{i=1}^r s_i.
$$
Therefore Eq \ref{eqn:meanDKLPart1} becomes
\begin{align}
\overline{\mathrm{KLD}}\left(t\right) = \left(\frac{n+2}{n}\right)^{t/\overline{s}} K_0, \label{eqn:meanDKLPart2}
\end{align}
and because
$$
\frac{n+2}{n} > 1,
$$
$\overline{\mathrm{KLD}}$ will grow exponentially in $t$.
The simulation chooses agents at random to interact, so $s_1,\ldots,s_r$ will be valued from the same random variable $S$. which will follow a geometric distribution with the probability of success equal to the probability of interaction being a $p^*=1$ interaction, or the probability that for any $k\in 1,\ldots,\psi$ the simulation chooses two agents from a cluster $A_k$. Therefore the probability of success in $S$ is $1/\psi$ which implies $\E{S} = \psi$ and by extension $\E{\overline{s}} = \psi$, hence $\bar{s}$ is an unbiased estimator or $\psi$. Substituting $\E{\bar{s}}=\psi$ Eq \ref{eqn:meanDKLPart2} becomes
\begin{align}
\overline{\mathrm{KLD}}\left(t\right) = \left(\frac{n+2}{n}\right)^{t/\psi} K_0. \label{eqn:meanDKL}
\end{align}
Taking the $\log$ on both sides of Eq \ref{eqn:meanDKL} gives us
$$
\logof{\overline{\mathrm{KLD}}\left(t\right)} = \logof{K_0} +\frac{\logof{n+2}-\logof{n}}{\psi} t,
$$
we can estimate the effective cluster count $\widehat{\psi}$ by fitting a linear model with slope $m$ to simulation data $\logof{\overline{\mathrm{KLD}}\left(t\right)}$ using a least squares estimation. The general expression for the effectively estimated cluster count is
\begin{align*}
\widehat{\psi} = \frac{\logof{n+2}-\logof{n}}{m}.
\end{align*}
|
\section{Introduction}
In this paper, we study the largest common subtree of two random binary trees.
The trees considered here are \textit{leaf-labeled binary trees of size $n$}, which are binary trees with $n$ leaves together with a numbering of the leaves with distinct numbers $1,\ldots,n$.
By considering the trees up to label-preserving isomorphisms, one can define the notion of a \textit{uniform random leaf-labeled binary tree} with $n$ leaves.
Let $\bs T$ and $\bs T'$ be two independent uniform random trees as above. Let $\kappa(\bs T, \bs T')$ be the size of the largest common subtree, which is, the largest subset $A$ of $\{1,\ldots,n\}$ such that $\bs T$ and $\bs T'$ induce the same binary trees with leaf set $A$. See Section~\ref{sec:definitions} for more formal definitions.
It was shown in~\cite{BeSt15} that $\omid{\kappa(\bs T, \bs T')}$ is of order at least $n^{1/8}$ (for the worst case problem for deterministic trees, a better bound of $n^{1/6}$ was found later in~\cite{deterministictrees}). This bound was improved by D. Aldous in~\cite{aldous} to $n^{\beta}$, where $\beta=(\sqrt 3-1)/2\sim 0.366$. An upper bound of order $n^{1/2}$ is also derived in~\cite{Pi21} using the first moment method and improving the upper bound remains an open problem; see Open Problem~2 of~\cite{aldous} (in a related model where the second tree is obtained by a random relabeling of the leaves, it is proved in~\cite{MiSu19} that the order is exactly $n^{1/2}$).
In this paper, we improve the lower bound to order $n^{0.4464}$. More precisely,
\begin{theorem}
\label{thm:main}
For the constant $0.4464<\beta<0.4465$ given by~\eqref{eq:beta} below and for every $\beta'<\beta$, one has $\omid{\kappa(\bs T, \bs T')}=\Omega(n^{\beta'})$.
\end{theorem}
We will also prove the same result for rooted and doubly-rooted trees (Theorem~\ref{thm:induction}).
The structure of this paper is as follows. In the rest of the introduction, an outline of the algorithm and a heuristic analysis is provided. Section~\ref{sec:definitions} provides the definitions and some lemmas regarding splitting the trees at their centroids. The formal algorithm and the proof of the main theorem is then given in Section~\ref{sec:proof}.
\subsection{Outline of the Algorithm}
The proof of Theorem~\ref{thm:main} is by improving the algorithm of~\cite{aldous} to construct a common subtree recursively. The general structure of the algorithm of~\cite{aldous} can be summarized as follows.
\begin{algorithm}[Aldous, 2022]
\label{alg:general}
Given two binary trees $t$ and $t'$ on the same leaf set,
\begin{enumerate}[(i)]
\item \label{alg:general:branch}
Choose branch points $b$ and $b'$ in $t$ and $t'$ respectively and match $b$ to $b'$.
\item \label{alg:general:sort} By splitting at $b$, $t$ is decomposed into three subtrees (see Figure~\ref{fig:splitting}). Choose a \textit{suitable} order for the subtrees and denote their leaf sets by $A_1,A_2$ and $A_3$. By doing the same for $t'$, the leaves are partitioned into $A'_1, A'_2$ and $A'_3$.
\item \label{alg:general:recursion} For each $i$, two binary trees are induced on the leaf set $A_i\cap A'_i$. Recursively, construct a common subtree of these two trees that contains the distinguished leaf corresponding to $b$ (and any other distinguished leaf from the previous stages). Join the three resulting subtrees to obtain a common subtree of $t$ and $t'$.
\end{enumerate}
\end{algorithm}
There are some choices to be made in this general sketch. In~\cite{aldous}, the branch point $b$ in the first stage is selected by choosing three random leaves and considering the branch point separating them. In the next stages, all trees have two distinguished leaves and the branch point is chosen by selecting a third leaf randomly. The crucial facts that allow the analysis of the algorithm are: (1) the distribution of the sizes of the three branches converges to a known distribution as the number of leaves grow, (2) conditioned on the sizes, the distribution of the three leaf sets $(A_1,A_2,A_3)$ is uniform and (3) conditioned on the leaf sets, the three subtrees of $t$ are uniform and independent. Using these facts, a heuristic identity is obtained in~\cite{aldous} for the asymptotic growth (see Subsection~\ref{subsec:heuristic} below). The formal proof is by controlling the convergence and a martingale argument.
Facts (2) and (3) above hold for other choices in the algorithm as well, as long as the probability of choosing a branch point in step~\eqref{alg:general:branch} is a function of the sizes of the three branches. It is suggested in~\cite{aldous} to choose the \textit{centroid} of the trees in step~\eqref{alg:general:branch} of the algorithm and to sort the branches by their sizes in step~\eqref{alg:general:sort}, since this would result in throwing out less leaves in step~\eqref{alg:general:recursion}.
However, there are some issues in continuing the recursion (as partly mentioned in Subsection~4.2 of~\cite{aldous}), which are discussed below.
The first issue is that sorting by size is only possible in the first stage of the recursion. In the second stage, when one splits the subtree with leaf set $A_i\cap A'_i$ into three pieces, one of the pieces contains $b$ and it must be matched to the piece containing $b'$ in the other tree. Also, this piece has two distinguished leaves, namely $b$ and $b_1$ (which are branch points in $t$). This causes another constraint on the third stage of the recursion. In addition, the next branch point should separate $b$ and $b_1$, and hence, the centroid cannot necessarily be chosen. Here, we propose to select the \textit{semi-centroid} of the path connecting $b$ to $b_1$, which will be defined similarly to the centroid. Using this, we obtain a full recursive algorithm involving three types of trees: Those with 0, 1 or 2 distinguished leaves, which are called \textit{non-rooted}, \textit{rooted} and \textit{doubly-rooted} here (non-rooted trees appear only in the first stage). The formal algorithm will be presented in Subsection~\ref{subsec:alg}. In this algorithm, despite what is mentioned in Subsection~4.2 of~\cite{aldous}, the number of constrains do not increase in the next stages. This allows us to analyze the algorithm heuristically (discussed in Subsection~\ref{subsec:heuristic}), which results in a system of equations. By solving these equations numerically, the exponent $\beta$ is obtained. For a rigorous proof, we simplify the proof of~\cite{aldous} significantly
and extend it to the new algorithm. In particular, no martingale argument is needed and the proof is by a simple induction on $n$.
The value of $\beta$ found in this method is less that the upper bound $0.485\ldots$ given in Section~4.2 of~\cite{aldous} for two reasons: One is that sorting by size is not always possible and the other is that the semi-centroid is sometimes used instead of the true centroid (which is different from the centroid with an asymptotically non-vanishing probability). However, by the above arguments, we think that this is the best bound one can expect from using the centroid if one desires that the trees in each stage are uniformly at random and independent. It is guessed by D. Aldous that the bound given by the centroid method would be less than, but close to, the true growth exponent.
\subsection{Heuristic analysis of the algorithm}
\label{subsec:heuristic}
For $i=0,1,2$, let $\bs T_i$ and $\bs T'_i$ be independent random binary trees on $n$ leaves plus $i$ additional distinguished leaves. The above algorithm of splitting at the centroid constructs a common subtree of $\bs T_i$ and $\bs T'_i$ that contains the distinguished leaves. One might expect that the size of the output of the algorithm is asymptotically $\bs Z_i n^{\beta_i}$ for some constant $\beta_i$ and some random variable $\bs Z_i$ (see Conjecture~6 of~\cite{aldous} which is for the case $i=0$). In particular, one might expect that the expected size of the output is asymptotically $C_i n^{\beta_i}$. However, there is no reason to expect $C_0,C_1$ and $C_2$ to be equal.
We follow the heuristic of~\cite{aldous} for splitting $\bs T_0$ at the centroid. The sizes of the three branches are asymptotically $(nX_1, nX_2, nX_3)$, where $(X_1,X_2,X_3)$ has a Dirichlet$(-1/2,-1/2,-1/2)$ distribution conditioned on $X_1\leq X_2\leq X_3\leq \frac 12$. Similarly, the sizes of the three branches of $\bs T'_0$ are asymptotically $(nX'_1,nX'_2,nX'_3)$, independently of $\bs T_0$. Conditioned on knowing the sizes of the branches, the sizes of $A_i\cap A'_i$, $i=1,2,3$ are very concentrated around their means $(nX_1X'_1, nX_2X'_2, nX_3X'_3)$. Therefore, the recursion gives heuristically that
\begin{equation*}
C_0 n^{\beta_0} = \omid{C_1(nX_1X'_1)^{\beta_1} + C_1(nX_2X'_2)^{\beta_1} + C_1(nX_3X'_3)^{\beta_1}}.
\end{equation*}
This gives that $\beta_0=\beta_1=:\beta$ and
\begin{equation}
\label{eq:rec0}
C_0 = C_1 \omid{X_1^\beta}^2 + C_1 \omid{X_2^\beta}^2 + C_1 \omid{X_3^\beta}^2.
\end{equation}
This is different from equation~(2) of~\cite{aldous} since the right hand side involves $C_1$, which corresponds to having a distinguished leaf after splitting (see Figure~\ref{fig:splitting}).
\begin{figure}[t]
\begin{center}
\includegraphics[width=.9\textwidth]{fig}
\end{center}
\caption{Splitting the trees. In the left, three trees are shown. The distinguished leaves $\star$ and $\bullet$ are depicted by $\times$ and $\bullet$ respectively. The number of distinguished leaves of each tree is shown by a number. A branch point (which is the centroid or semi-centroid) is selected and shown by $\circ$. In the right, the branch point is split into three new leaves, which are now considered as distinguished leaves and shown by $\diamond$.}
\label{fig:splitting}
\end{figure}
We can do similar calculations for splitting $\bs T_1$ at the centroid. We will show that the sizes of the branches are asymptotically $(nU_1,nU_2,nU_3)$, where $(U_1,U_2,U_3)$ has a Dirichlet$(-1/2,-1/2,1/2)$ distribution conditioned on $U_1\leq U_2\leq \frac 12$ and $U_3\leq \frac 12$. Here, $U_3$ corresponds to the branch containing the distinguished leaf and we have sorted the other two branches. By a similar heuristic argument, one obtains that $\beta_2=\beta$ and
\begin{equation}
\label{eq:rec1}
C_1 = C_1 \omid{U_1^\beta}^2 + C_1 \omid{U_2^\beta}^2 + C_2 \omid{U_3^\beta}^2.
\end{equation}
In this equation, the coefficient $C_2$ corresponds to the branch with two distinguished leaves (in fact, the heuristic gives $\beta_2\leq\beta$ and the inequality $\beta_2\geq \beta$ follows from splitting $\bs T_2$ discussed below).
Finally, when splitting $\bs T_2$ at the semi-centroid of the two distinguished leaves, the sizes of the three branches are asymptotically $(nV_1,nV_2,nV_3)$, where $(V_1,V_2,V_3)$ has a Dirichlet$(1/2,1/2,-1/2)$ distribution conditioned on $V_1\leq\frac 12$ and $V_2\leq \frac 12$. Here, $V_1$ and $V_2$ correspond to the branches containing the two distinguished leaves, sorting is not possible and there is no restriction on $V_3$. One finds similarly
\begin{equation}
\label{eq:rec2}
C_2 = C_2 \omid{V_1^\beta}^2 + C_2 \omid{V_2^\beta}^2 + C_1 \omid{V_3^\beta}^2.
\end{equation}
In this equation, the coefficient $C_1$ corresponds to the third branch which has only one distinguished leaf. By letting $\alpha:=C_2/C_1$, one can solve $\alpha$ in terms of $\beta$ in either of~\eqref{eq:rec1} and~\eqref{eq:rec2}:
\begin{equation}
\label{eq:beta}
\alpha=\frac{1-\omid{U_1^\beta}^2 - \omid{U_2^\beta}^2}{\omid{U_3^\beta}^2} = \frac{\omid{V_3^\beta}^2}{1-\omid{V_1^\beta}^2-\omid{V_2^\beta}^2}.
\end{equation}
By a monotonicity and continuity argument, there exists a unique solution for $\beta$. Simulation gives that $0.4464<\beta<0.4465$ and $\alpha\approx 0.859$. Although the above arguments are heuristic and don't contain a formal proof, \eqref{eq:beta} is very important for being able to use induction in the proof of the main theorem.
\section{Random Binary Trees}
\label{sec:definitions}
\subsection{The Model}
\label{subsec:model}
A binary tree $T$ is a finite tree in which every vertex has degree 1 or 3. The vertices of degree 1 are called the leaves and the other vertices are called the branch points.
If $B$ is a subset of leaves, then $T$ induces a binary tree with leaf set $B$ naturally, which is denoted by $\restrict{T}{B}$ (first consider the subtree spanned by $B$ and then remove the vertices of degree two).
Call two binary trees with a common leaf set $A$ equivalent if there exists a graph isomorphism between them which is the identity on $A$; i.e., preserves the labels of the leaves. Under this equivalence relation, there are finitely many classes which are called \defstyle{non-rooted labeled binary trees} with leaf set $A$.
Here, by a \defstyle{random binary tree} with leaf set $A$, we mean a random element of this finite set chosen uniformly.
Fix $n\geq 0$.
Define the random binary trees $\bs T_0$ (if $n\neq 0$), $\bs T_1$ and $\bs T_2$ with the uniform distribution on the set of trees with the following leaf sets.
\begin{eqnarray*}
\bs T_0 &\text{with leaf set}& \{1,\ldots,n\},\\
\bs T_1 &\text{with leaf set}& \{1,\ldots,n,\star\},\\
\bs T_2 &\text{with leaf set}& \{1,\ldots,n,\star,\bullet\}.
\end{eqnarray*}
Here, $1,\ldots,n$ are called \textbf{the original leaves} and $\star$ and $\bullet$ are arbitrary elements called \textbf{the distinguished leaves}. So $\bs T_i$ has $i$ distinguished leaves. We say that $\bs T_0, \bs T_1$ and $\bs T_2$ are \defstyle{non-rooted}, \defstyle{rooted} and \defstyle{doubly-rooted} respectively.\footnote{Usually, rooted binary trees are defined such that the root has degree 2. This is in bijection with the above definition when $n>0$, since deleting the distinguished leaf results in a vertex of degree two.}
We say that the \defstyle{size} of $\bs T_i$ is the number of original leaves; i.e., $n$.
Let $\bs T'_i$ be another random binary tree defined similarly and independently of $\bs T_i$.
If $B$ is a subset of the leaves such that $B$ contains the distinguished leaves $\star$ and $\bullet$ (if they are present) and $\restrict{(\bs T_i)}{B}$ is equivalent to $\restrict{(\bs T'_i)}{B}$, then we call the latter a \textbf{common subtree} of $\bs T_i$ and $\bs T'_i$. Define $\kappa(\bs T_i,\bs T'_i)$ to be the size of the largest common subtree of $\bs T_i$ and $\bs T'_i$.
We stress again that the distinguished leaves are contained in the subtrees but are not counted in their sizes.
We recall some of the properties of random binary trees from~\cite{aldous}.
Such trees can be constructed recursively by adding the leaves one by one and attaching them to (the middle of) one of the existing edges chosen uniformly. Since the number of edges is two times the number of leaves minus 3, one finds that the number of binary trees on $n$ given leaves is
\[
c_n:=1\times 3\times 5 \times \cdots \times (2n-5) = \frac{(2n-5)!}{2^{n-3} (n-3)!}.
\]
Stirling's formula gives that $c_n\sim 2^{n-3/2}e^{-n}n^{n-2}$. In particular, this implies
\begin{equation}
\label{eq:c}
\frac{c_{n+2}}{n!}\sim \pi^{-1/2}2^n n^{-1/2},\quad \frac{c_{n+1}}{n!}\sim \pi^{-1/2}2^{n-1} n^{-3/2}.
\end{equation}
A fundamental simple property is the \textit{consistency property} expressed in the following lemma.
\begin{lemma}
\label{lem:consistency}
If $\bs T$ is a uniform random binary tree with leaf set $A$ and $B\subset A$, then $\restrict{\bs T}{B}$ is a uniform binary tree with leaf set $B$. If $B$ is random and independent of $\bs T$, then the same holds when conditioned on $B$.
\end{lemma}
\subsection{Splitting at the Centroid}
\label{subsec:splitting}
It is known that by splitting $\bs T_0$ at the centroid, the sizes of the three branches asymptotically have the Dirichlet$(-1/2,-1/2,-1/2)$ distribution. However, we do not need this since the recursion only involves rooted and doubly-rooted trees (the theorem for non-rooted trees will follow from a simple comparison with rooted trees; see Lemma~\ref{lem:coupling}). We discuss splitting these types of trees in this subsection.
First, we analyze splitting $\bs T_2$. Let $b$ be a branch point on the path connecting the leaves $\star$ and $\bullet$.
By replacing $b$ with three distinct leaves and connecting them to the neighbors of $b$, $\bs T_2$ is split into three binary trees $\bs S_1,\bs S_2$ and $\bs S_3$. Choose the order such that $\star$ and $\bullet$ belong to $\bs S_1$ and $\bs S_2$ respectively.
Note that $S_3$ has one distinguished leaf (corresponding to $b$) and the other two trees have two distinguished leaves (corresponding to $b$ and one of $\star$ and $\bullet$).
Let $\bs l_i$ be the number of original leaves in $\bs S_i$. Assume $b$ is the \textbf{semi-centroid} of $\bs T_2$ from the beginning, which is uniquely defined by the conditions $\bs l_1< \frac n2$ and $\bs l_2\leq\frac n2$.
See Figure~\ref{fig:splitting} for an illustration.
The above decomposition provides a natural bijection, which enables one to analyze the distribution of the three branches: Every partitioning of $\{1,\ldots,n\}$ into three sets $(A_1,A_2,A_3)$ with sizes $(l_1,l_2,l_3)$, such that $0\leq l_1< \frac n2$ and $0\leq l_2\leq \frac n2$, and every triple of trees $(S_1,S_2,S_3)$ with leaf sets $A_1\cup\{\star, b\}$, $A_2\cup\{\bullet,b\}$ and $A_3\cup\{b\}$ respectively (where $b$ is an abstract element here), determine a unique doubly-rooted binary tree. This implies the following lemma straightforwardly.
\begin{lemma}
\label{lem:split2}
By splitting $\bs T_2$ at the semi-centroid,
\begin{enumerate}[(i)]
\item \label{lem:split2:1} For every $l_1,l_2,l_3\in\mathbb N\cup\{0\}$ such that $\sum_i l_i=n$, $l_1< \frac n2$ and $l_2\leq \frac n2$,
\[
\myprob{\forall i: \bs l_i=l_i} = \frac{\binom{n}{l_1\ l_2\ l_3}\;c_{l_1+2}\;c_{l_2+2}\;c_{l_3+1}}{c_{n+2}}.
\]
\item Conditioned on $\forall i: \bs l_i=l_i$, the sets of original leaves of $\bs S_1,\bs S_2$ and $\bs S_3$ are uniformly at random among all possible partitions of $\{1,\ldots,n\}$ with the given sizes.
\item Conditioned on the three leaf sets, $\bs S_1, \bs S_2$ and $\bs S_3$ are independent and uniform random binary trees, respectively having 2, 2 and 1 distinguished leaves.
\end{enumerate}
\end{lemma}
One can deduce the asymptotic distribution of the three branch sizes. Recall that the \defstyle{Dirichlet distribution} with parameters $(a_1,a_2,a_3)$ is the probability measure on the simplex $\{(x_1,x_2,x_3): x_1,x_2,x_3\geq 0, \sum x_i=1\}$ whose density is proportional to $\prod_i x_i^{a_i-1}$. Part~\eqref{lem:split2:1} of the above lemma, together with~\eqref{eq:c}, imply that
\begin{lemma}
\label{lem:dirichlet2}
The tuple $({\bs l_1}/{n}, {\bs l_2}/{n}, {\bs l_3}/{n})$ converges weakly to a random vector $(V_1,V_2,V_3)$ which has a Dirichlet$(1/2,1/2,-1/2)$ distribution conditioned on $V_1\leq\frac 12$ and $V_2\leq \frac 12$.
\end{lemma}
Now, we discuss splitting $\bs T_1$.
Let $b$ be the centroid of $\bs T_1$; i.e., a branch point such that after splitting at $b$, each of the three branches has at most half of the leaves. If there are two such vertices (which have to be neighbors and this happens only when one branch has exactly half of the leaves), choose the one farthest from the distinguished leaf $\star$. By splitting at $b$, one obtains 3 trees which are denoted by $\bs S_1,\bs S_2$ and $\bs S_3$ again (see Figure~\ref{fig:splitting}). Choose the order such that $\bs S_3$ contains $\star$. Note that $S_3$ has two distinguished leaves and the other two trees have one distinguished leaf. Let $\bs k_i$ be the number of original leaves (i.e., excluding $\star$) in $\bs S_i$. Choose the order of $\bs S_1$ and $\bs S_2$ such that $\bs k_1\leq \bs k_2$. For the rare cases in which $\bs k_1=\bs k_2$, choose one of the two orders randomly.
Similarly to the previous case, this decomposition provides an (almost-) bijection with the difference that it is a 2 to 1 map on the cases $\bs k_1=\bs k_2$. This implies the following result.
\begin{lemma}
\label{lem:split1}
By splitting $\bs T_1$ at the centroid,
\begin{enumerate}[(i)]
\item \label{lem:split1:1} For every $k_1,k_2,k_3\in\mathbb N\cup\{0\}$ such that $\sum_i k_i=n$, $k_1\leq k_2<\frac {n+1}2$ and $k_3\leq \frac {n-1}2$,
\[
\myprob{\forall i: \bs k_i=k_i} = \frac{\delta\binom{n}{k_1\ k_2\ k_3}\;c_{k_1+1}\;c_{k_2+1}\;c_{k_3+2}}{c_{n+1}},
\]
where $\delta=1$ if $k_1<k_2$ and $\delta=\frac 12$ otherwise.
\item \label{lem:split1:2} Conditioned on $\forall i: \bs k_i=k_i$, the sets of original leaves of $\bs S_1,\bs S_2$ and $\bs S_3$ are uniformly at random among all possible partitions of $\{1,\ldots,n\}$ with the given sizes.
\item \label{lem:split1:3} Conditioned on the three leaf sets, $\bs S_1, \bs S_2$ and $\bs S_3$ are independent and uniform random binary trees, respectively having 1, 1 and 2 distinguished leaves.
\end{enumerate}
\end{lemma}
One can show that the probability of $\bs k_1=\bs k_2$ converges to zero. Therefore, similarly to the previous case, part~\eqref{lem:split1:1} of the above lemma and~\eqref{eq:c} imply
\begin{lemma}
\label{lem:dirichlet1}
The tuple $({\bs k_1}/{n}, {\bs k_2}/{n}, {\bs k_3}/{n})$ converges weakly to a random vector $(U_1,U_2,U_3)$ which has a Dirichlet$(-1/2,-1/2,1/2)$ distribution conditioned on $U_1\leq U_2\leq \frac 12$ and $U_3\leq \frac 12$.
\end{lemma}
\section{Proof of the Main Theorem}
\label{sec:proof}
\subsection{The Algorithm}
\label{subsec:alg}
As discussed in Subsection~\ref{subsec:heuristic}, the recursion deals only with rooted and doubly-rooted trees (the other case will follow by comparison; see Subsection~\ref{subsec:proof}).
Let $n\geq 0$ and let $T$ and $T'$ be binary trees on the same leaf set, both rooted or both doubly-rooted. Recursively, we construct a common subtree of $T$ and $T'$ containing the additional distinguished leaf (or leaves). For $n=0$, either both of $T$ and $T'$ are trees with a single vertex $\star$ or with a single edge $\star\bullet$. In this case, output the whole tree. For $n\geq 1$, we proceed inductively as follows.
If the trees are doubly-rooted, split them at their semi-centroids $b$ and $b'$ as described in Subsection~\ref{subsec:splitting}. Assume that splitting $T$ results in partitioning $\{1,\ldots,n\}$ into subsets $(A_1,A_2,A_3)$, where the order is specified in Subsection~\ref{subsec:splitting}. Define $(A'_1,A'_2,A'_3)$ similarly. $T$ induces binary trees $\bs R_1$ with leaf set $(A_1\cap A'_1)\cup\{\star,b\}$, $\bs R_2$ with leaf set $(A_2\cap A'_2)\cup\{\bullet,b\}$ and $\bs R_3$ with leaf set $(A_3\cap A'_3)\cup \{b\}$. Define $\bs R'_1,\bs R'_2$ and $\bs R'_3$ similarly. By relabeling $b$ and $b'$, $\bs R_3$ and $\bs R'_3$ can be regarded as rooted trees and the rest as doubly-rooted trees. The induction hypothesis defines output common subtrees for the pairs $(\bs R_i,\bs R'_i)$ for $i=1,2,3$. Join these trees by identifying the three distinguished leaves corresponding to $b$. It is clear that this forms a common subtree of $T$ and $T'$. This is the output of the algorithm.
If $T$ and $T'$ are rooted, split them at their centroids as described in Subsection~\ref{subsec:splitting}. Similarly to the previous case, redefine the subtrees $\bs R_i$ and $\bs R'_i$ for $i=1,2,3$, use the induction hypothesis to define three common subtrees (note that $\bs R_3$ and $\bs R'_3$ are doubly-rooted and the rest are rooted), glue them to find a common subtree of $T$ and $T'$, and output the resulting tree. This finishes the definition of the algorithm.
\subsection{Analysis of the Algorithm}
For $i=1,2$, let $\gamma(\bs T_i,\bs T'_i)$ be the size of the output of the algorithm of Subsection~\ref{subsec:alg} for the trees $\bs T_i$ and $\bs T'_i$ defined in Subsection~\ref{subsec:model}. Note that the dependence of these values on $n$ are not shown in the symbols for better readability.
\begin{theorem}
\label{thm:induction}
Let $\alpha$ and $\beta$ be constants satisfying~\eqref{eq:beta}.
For every $\beta'<\beta$, there exists $\delta>0$ such that for all $n\geq 0$,
\begin{eqnarray}
\label{eq:thm:induction:1}
\omid{\gamma(\bs T_1,\bs T'_1)}&\geq& \delta n^{\beta'},\\
\label{eq:thm:induction:2}
\omid{\gamma(\bs T_2,\bs T'_2)}&\geq& \alpha \delta n^{\beta'}.
\end{eqnarray}
\end{theorem}
\begin{remark}
By similar arguments, one can prove that the rate is exactly $n^{\beta}$; i.e., for every $\beta''>\beta$, one has $\omid{\gamma(\bs T_i,\bs T'_i)}=O(n^{\beta''})$.
\end{remark}
We will prove this theorem by induction on $n$ according to the following motivation.
With the notations of the algorithm, one has
\begin{equation}
\label{eq:gamma}
\gamma(\bs T_1,\bs T'_1) = \sum_i \gamma(\bs R_i,\bs R'_i).
\end{equation}
So, we need to bound $\sum_i \omid{\gamma(\bs R_i,\bs R'_i)}$. We will show that
$\bs R_i$ is a uniform random binary tree on $A_i\cap A'_i$ (rooted or doubly-rooted as specified in the algorithm). Therefore, if the induction hypothesis holds, the conditional expectation of ${\gamma(\bs R_i,\bs R'_i)}$ is bounded from below by $\delta\norm{A_i\cap A'_i}^{\beta'}$ or $\alpha\delta \norm{A_i\cap A'_i}^{\beta'}$. The formal proof is based on this motivation and the following lemma.
\begin{lemma}
\label{lem:induction}
Fix $\beta'\geq 0$ and $\epsilon>0$.
\begin{enumerate}[(i)]
\item Let $A_i$ and $A'_i$ ($i=1,2,3$) be the leaf sets defined in the algorithm when starting from $\bs T_1$ and $\bs T'_1$. Then, for large enough $n$,
\begin{equation}
\label{eq:lem:induction:1}
\forall i: \omid{\norm{A_i\cap A'_i}^{\beta'}} \geq (1-\epsilon)n^{\beta'} \omid{U_i^{\beta'}}^2,
\end{equation}
where $(U_1,U_2,U_3)$ is defined in Lemma~\ref{lem:dirichlet1}.
\item Similarly, if the sets are defined in the algorithm when starting from $\bs T_2$ and $\bs T'_2$, then for large enough $n$,
\begin{equation}
\label{eq:lem:induction:2}
\forall i: \omid{\norm{A_i\cap A'_i}^{\beta'}} \geq (1-\epsilon)n^{\beta'} \omid{V_i^{\beta'}}^2,
\end{equation}
where $(V_1,V_2,V_3)$ is defined in Lemma~\ref{lem:dirichlet2}.
\end{enumerate}
\end{lemma}
\begin{proof}
The proof is almost the same as that of Proposition~5 of~\cite{aldous}.
We only prove the first claim and the second one can be proved similarly. Condition on the event $E$ of $\norm{A_i}=a_i$ and $\norm{A'_i}=a'_i$ for all $i$. Fix $i\in\{1,2,3\}$. By part~\eqref{lem:split1:2} of Lemma~\ref{lem:split1}, $A_i$ and $A'_i$ are independent uniform random subsets of $\{1,\ldots,n\}$ of sizes $a_i$ and $a'_i$. So, $\norm{A_i\cap A'_i}$ has a hypergeometric distribution. Fix $\epsilon>0$ and let $F_{\epsilon}$ be the event that $a_i$ and $a'_i$ are between $\epsilon n$ and $(1-\epsilon)n$. In this regime, the hypergeometric distribution is concentrated around its mean. In particular, Equation~(16) of~\cite{aldous} gives the following: When $\epsilon$ is fixed, for large enough $n$, and for all choices of $a_i$ and $a'_i$ between $\epsilon n$ and $(1-\epsilon)n$,
\[
\omidCond{\norm{A_i\cap A'_i}^{\beta'}}{E} \geq (1-\epsilon) \omidCond{\norm{A_i\cap A'_i}}{E}^{\beta'} = (1-\epsilon) \left(\frac{a_ia'_i}{n} \right)^{\beta'}.
\]
This implies that for large $n$,
\[
\omid{\norm{A_i\cap A'_i}^{\beta'}} \geq (1-\epsilon) \omid{\left(\frac{\norm{A_i}\cdot\norm{A'_i}}n\right)^{\beta'}\identity{F_{\epsilon}}}=(1-\epsilon)n^{\beta'}\omid{\left(\frac {\norm{A_i}}n\right)^{\beta'}\identity{F_{\epsilon}}}^2.
\]
By the convergence in Lemma~\ref{lem:dirichlet1}, one deduces that
\[
\liminf_n n^{-\beta'} \omid{\norm{A_i\cap A'_i}^{\beta'}} \geq (1-\epsilon)\omid{U_i^{\beta'}\identity{G_{\epsilon}}}^2,
\]
where $G_{\epsilon}$ is the event $\epsilon\leq U_i\leq (1-\epsilon)$.
Now the claim is obtained by reducing the value of $\epsilon$ suitably.
\end{proof}
We are now ready to prove Theorem~\ref{thm:induction}.
\begin{proof}[Proof of Theorem~\ref{thm:induction}]
Fix $\beta'<\beta$. Since $\alpha$ and $\beta$ satisfy~\eqref{eq:beta} and the terms in~\eqref{eq:beta} are monotone in $\beta$ (see also~\eqref{eq:rec1} and~\eqref{eq:rec2}), there exists $\epsilon>0$ such that
\begin{eqnarray}
\label{eq:epsilon1}
\omid{U_1^{\beta'}}^2 + \omid{U_2^{\beta'}}^2 + \alpha \omid{U_3^{\beta'}}^2 &\geq & \frac 1{1-\epsilon},\\
\label{eq:epsilon2}
\alpha \omid{V_1^{\beta'}}^2 + \alpha \omid{V_2^{\beta'}}^2 + \omid{V_3^{\beta'}}^2 &\geq & \frac{\alpha}{1-\epsilon}.
\end{eqnarray}
For this value of $\epsilon$, by Lemma~\ref{lem:induction}, there exists $N\in\mathbb N$ such that~\eqref{eq:lem:induction:1} and~\eqref{eq:lem:induction:2} hold for all $n\geq N$. Choose $\delta>0$ such that for all $n< N$, the claims~\eqref{eq:thm:induction:1} and~\eqref{eq:thm:induction:2} of the theorem hold.
Now, we prove the claims by induction on $N$. By the definition of $\delta$, the claims hold for all $n< N$. Now, assume $n\geq N$ and both claims hold for $1,2,\ldots,n-1$. We will prove them for $n$.
First, consider the algorithm for $\bs T_1$ and $\bs T'_1$. As mentioned in~\eqref{eq:gamma}, one has $\omid{\gamma(\bs T_1,\bs T'_1)} = \sum_i \omid{\gamma(\bs R_i,\bs R'_i)}$. By part~\eqref{lem:split1:3} of Lemma~\ref{lem:split1} and the consistency property (Lemma~\ref{lem:consistency}), conditioned on knowing $A_i$ and $A'_i$, $\bs R_i$ and $\bs R'_i$ are independent random binary trees on the leaf set $A_i\cap A'_i$ (which are rooted for $i=1,2$ and doubly-rooted for $i=3$). Therefore, the induction hypothesis implies that
\[
\omid{\gamma(\bs T_1,\bs T'_1)} \geq \delta\omid{\norm{A_1\cap A'_1}^{\beta'}} + \delta\omid{\norm{A_2\cap A'_2}^{\beta'}} + \alpha\delta\omid{\norm{A_3\cap A'_3}^{\beta'}}.
\]
Since $n\geq N$, the definition of $N$ and~\eqref{eq:lem:induction:1} give
\[
\omid{\gamma(\bs T_1,\bs T'_1)} \geq \delta(1-\epsilon)n^{\beta'}\left(\omid{U_1^{\beta'}}^2 + \omid{U_2^{\beta'}}^2 + \alpha \omid{U_3^{\beta'}}^2\right).
\]
By~\eqref{eq:epsilon1}, the right hand side is at least $\delta n^{\beta'}$.
So the first induction claim~\eqref{eq:thm:induction:1} is proved.
Second, consider the algorithm for $\bs T_2$ and $\bs T'_2$. Similarly to the previous case, $\omid{\gamma(\bs T_2,\bs T'_2)} = \sum_i \omid{\gamma(\bs R_i,\bs R'_i)}$ (note that $\bs R_i, \bs R'_i, A_i$ and $A'_i$ are redefined here). Since $\bs R_i$ and $\bs R'_i$ are doubly-rooted for $i=1,2$ and rooted for $i=3$, the induction hypothesis gives similarly
\[
\omid{\gamma(\bs T_2,\bs T'_2)} \geq \alpha\delta\omid{\norm{A_1\cap A'_1}^{\beta'}} + \alpha\delta\omid{\norm{A_2\cap A'_2}^{\beta'}} + \delta\omid{\norm{A_3\cap A'_3}^{\beta'}}.
\]
Since $n\geq N$, the definition of $N$ and~\eqref{eq:lem:induction:2} give
\[
\omid{\gamma(\bs T_2,\bs T'_2)} \geq \delta(1-\epsilon)n^{\beta'}\left(\alpha\omid{V_1^{\beta'}}^2 + \alpha\omid{V_2^{\beta'}}^2 + \omid{V_3^{\beta'}}^2\right).
\]
By~\eqref{eq:epsilon2}, the right hand side is at least $\alpha\delta n^{\beta'}$.
This proves the second induction claim~\eqref{eq:thm:induction:2} and the theorem is proved.
\end{proof}
\subsection{Proof of Theorem~\ref{thm:main}}
\label{subsec:proof}
Theorem~\ref{thm:induction} is the analogous of Theorem~\ref{thm:main} for rooted and doubly-rooted trees. We will deduce Theorem~\ref{thm:main} from the following comparison lemma.
\begin{lemma}
\label{lem:coupling}
There exists a coupling of $(\bs T_0, \bs T'_0)$ and $(\bs T_1, \bs T'_1)$ such that $\kappa(\bs T_0,\bs T'_0)\geq\kappa(\bs T_1,\bs T'_1)$.
\end{lemma}
\begin{proof}
By the consistency property (Lemma~\ref{lem:consistency}),
if one removes the distinguished leaf $\star$ from $\bs T_1$, the resulting tree has the same distribution as $\bs T_0$. This provides a coupling of $(\bs T_0, \bs T'_0)$ and $(\bs T_1, \bs T'_1)$. In this coupling, every common subtree of $\bs T_1$ and $\bs T'_1$ is also a common subtree for $\bs T_0$ and $\bs T_0$ (after removing $\star$). This proves the claim.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:main}]
Subsection~\ref{subsec:alg} provides an algorithm for producing a common subtree of $\bs T_1$ and $\bs T'_1$ with size $\gamma(\bs T_1,\bs T'_1)$. So $\kappa(\bs T_1,\bs T'_1)\geq \gamma(\bs T_1,\bs T'_1)$. Now the claim is implied from Theorem~\ref{thm:induction} and Lemma~\ref{lem:coupling}.
\end{proof}
\begin{remark}
Similarly to the above proof, one can simplify the proof of Theorem~1 of~\cite{aldous} by replacing the submartingale argument with a simple induction. In fact, the induction argument is basically using one step of the submartingale (instead of stopping it at a stopping time).
\end{remark}
|
\section{Introduction}
With the development of science and technology, people gradually turn their attention to the field of data mining. Various data mining techniques are utilized in real life. Machine learning technology has played a big role in it. The most widely known is the "Battle of the Century" between AlphaGo and Li Shishi, known as the man-machine war, and finally ended with the victory of AlphaGo, once again showing people machine learning of power. In the eyes of many people, machine learning is a very abstract concept, and this article will analyze it in the field of recommendation systems.
\section{The process of machine learning}
What is machine learning? In layman's terms, machine learning is to let machines learn and summarize "experience" like humans. Of course, machines cannot accumulate "experience" by going through various things like humans. Instead, let the machine analyze the existing data, summarize the rules, and summarize to form a set of models, which can be applied to real life.
\subsection{Get data}
The first step in machine learning is to obtain data. Machine learning without data is empty talk. The acquisition of data is easier than the later steps, because data is everywhere. There are purchase records of consumers in supermarkets, driving records of vehicles on driving recorders, and viewing records of movie fans in movie theaters. These data information is the premise for us to start machine learning.
\subsection{Clean data}
The data we obtain is irregular, there is a lot of redundant data that is not useful to us, and there may be misinformation that affects our analysis results. We need to preprocess this information before conducting specific analysis, so as not to affect the accuracy of the analysis results. Data recommendation system based on machine learning is very common in this era of ubiquitous data and information, how to obtain and analyze these data has become the research topic of many people. In view of this situation, this paper makes some brief overviews of machine learning-related recommendation systems. By analyzing some technologies and ideas used by machine learning in recommender systems, let more people understand what is big data and what is machine learning.The most important point is to let everyone understand the profound impact of machine learning on our daily life.
Abstract Preprocessing is mainly divided into three aspects: distance measurement, sampling, dimensionality reduction.
The KNN classification (k-NearestNeighbor) used in collaborative filtering recommender systems mainly depends on the distance metric method. The more commonly used distance measurement methods are Euclidean distance, Pearson correlation coefficient, Jaccard coefficient (for binary attributes), etc. Sampling is the main technique for data mining to select relevant data subsets from large data sets. It also plays an important role in the final interpretation step. The most commonly used sampling method is non-replacement sampling. Taken from the population, but it is also permissible to perform substitution sampling, that is, the item does not need to be removed from the population even if it is selected. Usually the ratio of training and test sets is 80/20. The final dimension reduction is to remove some points that are very sparse and have little impact on the result set, reduce the dimension, avoid dimension disaster, and reduce the difficulty of operation.
\subsection{Analyzing data and building models}
In the whole process of machine learning, the most difficult and core is to analyze data. There are many ways to analyze data, each of which has different effects in different practical applications, and needs to be analyzed on a case-by-case basis. Here are some of the more commonly used data analysis methods in recommender systems:
Nearest Neighbor Algorithm (KNN): KNN predicts the label class of unknown samples by storing training records and using them. Such classifiers store the entire training set and only classify new records when they exactly match the training set. Compared with other machine learning algorithms, KNN is the simplest, because KNN does not need to build an explicit model, which is called a lazy learner. Although the KNN method is simple and intuitive, its results are accurate and very easy to improve.
k-means algorithm: The k-means algorithm is a block clustering algorithm that divides the acquired n object data into k uncorrelated subsets (k$\textless$n). It is similar to the expectation-maximization algorithm that deals with mixed normal distributions in that they both try to find the centers of natural clusters in the data. It assumes that object attributes come from spatial vectors, and the goal is to minimize the sum of mean squared errors within each group. The k-means algorithm starts by randomly selecting k center points, and all items are assigned to the class of their closest center node. When items are newly added or removed, the central node of the new cluster needs to be updated, and the membership of the cluster also needs to be updated, and so on and so forth until no items change their cluster membership. The final cluster is very sensitive to the initial center point, and there may also be empty clusters.
Artificial Neural Network (ANN): The artificial neural network algorithm simulates a biological network, which consists of a set of internal connection points and weighted chains, and is a type of pattern matching algorithm. Often used to solve classification and regression problems. ANN is a huge branch of machine learning, there are hundreds of different algorithms, and deep learning is one of its important components. The main advantage of ANN is that it can handle nonlinear classification tasks, and through parallel processing, it can operate in the case of partial network damage. But it is difficult for ANN to provide an ideal network topology for a given problem.When the topology is determined, its performance level is at the lower line of the classification error rate.
Bayesian Classifiers: Bayesian classifiers are a class of algorithms based on the definition of probability and Bayes' theorem, the school of Bayesian statistics that uses probability to represent uncertainty in relationships learned from data. It treats each attribute and class label as a random variable. Given a record with N attributes ($A_1 , A_2 , A_3 , ..., A_N$ ), the goal is to predict a class $C_k$ , by making a , ..., $A_N$ ), find the value of $C_k$ that maximizes the posterior probability for that class. Common Bayesian classifier algorithms include Naive Bayesian Algorithm, Average Single Dependency Estimation (AODE), and Bayesian Network (BBN).
\subsection{Test the model}
The last step in the whole machine learning is to test the model and check the accuracy of the model. This is an important step in measuring the merits of an algorithm. The test data set can be randomly selected from the test or obtained from the test set reserved in advance.
\section{Construction of restaurant recommendation system based on deep learning}
\subsection{Restaurant matching recommendation system based on deep learning}
\subsubsection{Overall Framework}
The overall framework is shown in Figure 3.1. First, the recall set is filtered out according to the given information. Specifically, the user's historical purchase data is used. Usually, a user buys two clothing products at the same time in a short period of time. The probability of matching will increase; secondly, using the text information of the product title, the recall set of the matching product to be predicted is obtained by calculating the cosine similarity. After the recall set is obtained, it is filtered by collocation categories and weighted for fusion according to their respective hit rates. Finally, the features are obtained through the convolutional neural network, and the logistic regression model is used to calculate the matching probability between the recall set and the test set, and use this probability to revise and reorder the fusion recall set.
\begin{figure}[H]
\centering
\includegraphics[width=14cm]{1.png}
\caption{The overall framework of a recommendation system based on deep learning}
\end{figure}
\subsubsection{Implementation of recall set triggering based on MapReduce}
Mapreduce is a distributed computing framework, mainly used to process large-scale data, which was proposed by Google in 2003. Its idea draws on the map function and reduce function of functional programming, and then the open source community Apache implements the Hadoop ecosystem, and its feature of supporting MapReduce makes it rapidly and widely used in engineering practice. After that, Alibaba Group realized the ODPS platform, and its support for MapReduce and elastic cloud computing features also attracted quite a lot of developers.
MapReduce is mainly divided into two stages, namely map and reduce. The map task is mainly responsible for data loading, parsing, transformation and filtering. The reduce task is responsible for processing a subset of the map task output. The entire processing flow is shown in Figure 3.2.
\begin{figure}[H]
\centering
\includegraphics[width=16cm]{2.png}
\caption{MapReduce processing flow}
\end{figure}
The recall set based on user similarity is mainly to obtain the number of people who have purchased any two products at the same time in a short period of time. The formula is expressed as follows:
Among them, (I$\cdot$) is the indicator function, when the function body is true, the count is increased by 1, It is any commodity, Bu is the set of commodities purchased by user u, Tui is the time when user u purchased commodity i, and $\tau$ is the time interval . This formula needs to be solved by two-step MapReduce. The specific solution process is as follows:
$$
{sim}\left(I t_{i}, I t_{j}\right) \approx {freq}\left(I t_{i}, I t_{j}\right)=I_{u \in U}\left(I t_{i} \in B_{u} \cap I t_{j} \in B_{u} \mid T_{u i}-T_{u j}<\tau\right)
$$
(1) Read the original data, use the user id as the key in the map stage, and output the product id and purchase date as the value; (2) in the reduce stage, combine all the products of a user whose purchase time difference does not exceed $\tau$ months,
and use Commodity ids are sorted and output in descending order; (3) In the second map stage, the pair of commodity id pairs are used as keys, and 1 is used as value for output;
(4) In the reduce stage, all the same id pairs are added up, and the final output, Product id pair and the number of purchasers. The construction of recall set based on product similarity is a little more complicated. It uses Natural Language Processing (NLP). To get the text similarity of product titles, the $T_f-I_{df}$ weight of each document word segmentation must be obtained first. The process is shown in Figure 3.3. shown. Among them, $T_f-I_{df}$ (Termfrequency–Inversedocumentfrequency) is a commonly used weighting technique for data mining. Its calculation is divided into two steps, namely the weight of term frequency (TermFrequency, TF) and the weight of inverse document frequency (InverseDocumentFrequency, IDF).
\begin{figure}[H]
\centering
\includegraphics[width=16cm]{3.png}
\caption{$T_f-I_{df}$ calculation process}
\end{figure}
The word frequency is weighted according to the frequency of the word segmentation in the document. The formula is as follows:
$$
\mathrm{w}_{k j}=t f_{k j}
$$
Among them, $tf_{kj}$ is the word frequency of the participle $t_k$ in the document $d_j$.
The inverse document frequency is to solve a problem in the weight of word frequency, specifically, all word segmentations are considered to be equally important. However, if all documents contain the same word, then the word segmentation has no degree of discrimination, and the original document cannot be distinguished from others. Then the inverse document frequency can be expressed as:
$$
\mathrm{i} d f_{k}=\log \frac{m}{n_{k}}
$$
Among them, m is the number of all documents, and $n_k$ is the number of documents in the document set D of the segmented word tk. According to Equation 4-5, if a word appears less frequently, its inverse document frequency is usually higher, and vice versa.
Therefore, the $T_f-I_{df}$ weights can be expressed as:
$$
\mathbf{w}_{k j}=t f_{k j} \times i d f_{k}
$$
After the $T_f-I_{df}$ weight is obtained, the cosine similarity of the two product title documents can be calculated. For the specific cosine similarity, see Section 2.2.2. This calculation requires the use of an inverted index. The specific MapReduce process is as follows:
(1) Input the document id, the word segmentation id and the $T_f-I_{df}$ weight of the word segmentation into the map task, use the word segmentation id as the key, and the document id and $T_f-I_{df}$ weight as the value output; (2) In the reduce stage, calculate any pair of documents with the same word segmentation The product of the $T_f-I_{df}$ weights, take the document pair as the key, and the weight product as the value output; (3) Input the original data with the $T_f-I_{df}$ weight to the map task again, take the document id as the key, and the word segmentation $T_f-I_{df}$weight is value output; (4) calculate the modulo of each document in the reduce phase;
(5) For the data output in steps 2 and 4, perform a left outer join with the document as the key; (6) Input the output of the previous step to the map task, and use the document id pair as the key, $T_f-I_{df}$ weight product, The document modulo is the value output; (7) Calculate the cosine similarity of each document pair in the reduce task.
\subsubsection{Ranking correction based on convolutional neural network}
Convolutional neural network is one of deep learning technologies, and its powerful image processing capability based on GPU has been more and more applied in engineering practice. The specific convolutional neural network is shown in Section 3.4.
This section will use the powerful feature extraction capabilities of convolutional neural networks to extract image features, and then pass them into the logistic regression model for further processing. The process is shown in Figure 3.4.
Firstly, the existing classic neural network model AlexNet is used to extract the image information of each commodity, and the feature is the neuron data of the penultimate layer of the network. From this, the high-dimensional features of the image can be obtained, and then the matching package data, the recall set data and the image feature data are connected respectively, and the difference between the image features of any two products is calculated. Add certain non-package negative sample data to the matching package as the training data for logistic regression. The output model predicts the probability of recalling the set as a matching package, and uses this probability to revise the original ranking.
\begin{figure}[H]
\centering
\includegraphics[width=14cm]{4.png}
\caption{CNN-based ranking correction process}
\end{figure}
The AlexNet model was proposed by Alex on the imagenet image classification challenge, and its network parameters are as high as 600 million. The first five layers are all convolutional layers, and the sixth to eighth layers are fully connected layers.
For sorting correction, firstly, the correction degree of each commodity pair needs to be calculated, which can be expressed as:
$$
{ fix }_{i t 1, i t 2}=\frac{p_{i t 1, i t 2}}{\sum\limits_{{pair } \in \operatorname{Re} c a l l} p_{ {pair }}} \times m
$$
Among them, $p_{i t 1, i t 2}$ is the output frequency of commodity pairs in the recall set in logistic regression, and m is the number of commodity pairs in the whole recall set. After that, the correction degree is standardized, specifically:
$$
\operatorname{rank}_{f i x}=f i x_{i t 1, i t 2}-a v g\left(f i x_{p a i r}\right)
$$
Among them, $avg(\cdot)$ is the mean value. rankfix is the final ranking correction number. In this way, the matching similarity of commodity pairs can be adjusted to a certain extent with the help of image information to a certain extent.
\subsection{Experimental results and analysis}
\subsubsection{Experimental platform and data}
All experiments in this section are based on the ODPS platform for big data processing. The total data flow of the experiment exceeds 3PB, including the MapReduce algorithm framework, Caffe deep learning components, common machine learning and natural language processing algorithm packages, and ODPSSql data processing platform.
First, the hybrid-based recommendation system will use the mobile Tmall real data (AliMobileRec) provided by Ali Group. Data provides user ID, product ID, product category , product location information (with missing values), when and where the user performed on the product (click, favorite, add to cart, purchase) and where and when (with a large number of missing values in geographic location information). There are a total of 5822532780 pieces of behavioral data, 524484376 pieces of behavioral data in the vertical field in the complete set of data, and 14397493 pieces of commodity information data. The specific information is shown in Table 3.1 and Table 3.2. The goal is to build a machine learning model to predict which users purchased which items on the forecast day, and then use the model to recommend items.
\begin{table}[H]
\renewcommand\arraystretch{1.5}
\centering
\caption{Mobile terminal behavior data of the complete set of commodities}
\begin{tabular}{p{3cm} p{3cm} p{8cm} }
\hline
Field Name & Data Type & Field Meaning \\
\hline
User\_id & bigint & User ID \\
Item\_id & bigint & product identification \\
Behavior\_type & bigint & Browse, Favorite, Add to Cart, Buy \\
User\_geohash & string & latitude and longitude \\
Item\_category & bigint & product category \\
Time & string & behavior time \\
\hline
\end{tabular}
\end{table}
\begin{table}[H]
\renewcommand\arraystretch{1.5}
\centering
\caption{Commodity subset data}
\begin{tabular}{p{3cm} p{3cm} p{8cm} }
\hline
Field Name & Data Type & Field Meaning \\
\hline
Item\_id & bigint & product identification \\
User\_geohash & string & Product location space identification \\
Item\_category & bigint & product category \\
\hline
\end{tabular}
\end{table}
\begin{table}[H]
\renewcommand\arraystretch{1.5}
\centering
\caption{Matching packages}
\begin{tabular}{p{3cm} p{3cm} p{8cm} }
\hline
Field Name & Data Type & Field Meaning \\
\hline
Coll\_id & bigint & with package id \\
User\_geohash & string & with a list of id \\
\hline
\end{tabular}
\end{table}
\begin{table}[H]
\renewcommand\arraystretch{1.5}
\centering
\caption{Commodity Information Sheet}
\begin{tabular}{p{3cm} p{3cm} p{8cm} }
\hline
Field Name & Data Type & Field Meaning \\
\hline
Item\_id & bigint & product identification \\
Cat\_id & bigint & product category id \\
Terms & string & Product title segmentation \\
Img\_data & jpg & product picture \\
\hline
\end{tabular}
\end{table}
\begin{table}[H]
\renewcommand\arraystretch{1.5}
\centering
\caption{ User history behavior table}
\begin{tabular}{p{3cm} p{3cm} p{8cm} }
\hline
Field Name & Data Type & Field Meaning \\
\hline
User\_id & bigint & User identification \\
Item\_id & bigint & product identification \\
Create\_at & string & Purchase date \\
\hline
\end{tabular}
\end{table}
Secondly, the data set used by the deep learning-based collocation recommendation system is the TaobaoClothesMatch data set. The problem is defined as the set data of some collocation masters, the product information and the user's historical purchase data, and the collocation list of the predicted target id. The purpose is to recommend the matching information of the purchased products to the user based on this kind of clothing matching information, and then associate the sales. The specific data formats are shown in Table 3.3 to Table 3.5.
Among them, bigint, string, and jpg are all stored data structures, which are long-format integers, strings, and compressed image formats, respectively. The scale of the data is 158318 package samples, 3269988 commodity and image information samples, and 728318569 user historical data samples.
\subsubsection{Experimental results}
(1) Experimental results of recommendation system based on hybrid technology
The recommender system will mainly use the F1 indicator for effect evaluation. The F1 index is used here instead of the commonly used accuracy index because the accuracy index generally needs to give a complete test set, and then obtain the correct proportion of the predicted samples. In the current scenario, it is necessary to construct a test set, that is, a recall set. Using the accuracy rate can only express the proportion of correct samples in the constructed recall set, while selecting F1 can express both the proportion of correct samples in the recall set and the number of correct samples in the real samples. The superiority of the system can be reflected only when both are relatively high, so the F1 index is selected as the evaluation index of the current system.
The data set gives the data from November 18th to December 18th, because the full set of verification data on the 19th cannot be obtained, so the data on December 18th is used as the verification set, and the number of submissions is 105,000 user-goods pairs. A total of 1200 multi-dimensional features are constructed on the ODPS platform. See Section 3.2.2 for the feature construction method.
First, verify the relationship between the number of features and the system effect, using the GDBT model, the positive and negative sampling ratio is 1:8, the model parameters are metric type AUC, the number of trees is 500, the learning rate is 0.05, the maximum number of cotyledons is 32, the maximum tree depth is 11, and the leaves The minimum number of samples for nodes is 500, the proportion of samples collected for training is 0.6, the proportion of features collected for training is 0.6, and the maximum number of splits is 500. The effect is shown in Figure 3.5.
\begin{figure}[H]
\centering
\includegraphics[width=12cm]{5.png}
\caption{The relationship between the number of features and the score}
\end{figure}
Among them, the added feature categories are: time of day and shopping cart features, user features, user product subset features, product subset features, product complete set features, product category features, product competition features, user product category features, and user product complete set features, etc.
Verify the effect of feature-level fusion on the recommendation effect. Due to the limitation of the algorithm platform, only a maximum of 799-dimensional features can be used. Therefore, three sets of features are randomly selected, each with 799 dimensions, and the GBDT model is used for training, and the parameters are the same as above. The specific effects are shown in Table 3.6.
\begin{table}[H]
\renewcommand\arraystretch{1.5}
\centering
\caption{Feature-level fusion effect table}
\begin{tabular}{p{6cm} p{3cm} p{3cm} p{3cm}}
\hline
feature set & Precision & Recall & F1 \\
\hline
feature set 1 & 0.090503845 & 0.054395265 & 0.067950461 \\
feature set 2 & 0.088748303 & 0.053772593 & 0.066968796 \\
feature set 3 & 0.091227602 & 0.054738021 & 0.068421842 \\
feature set 1+feature set 2 & 0.089557052 & 0.053938258 & 0.067326959 \\
feature set 2+feature set 3 & 0.089729976 & 0.054063935 & 0.067473728 \\
feature set 1+feature set 3 & 0.091543545 & 0.055075063 & 0.068773897 \\
feature set 1+feature set 2+feature set 3 & 0.091087625 & 0.05484656 & 0.068467068\\
\hline
\end{tabular}
\end{table}
\begin{table}[H]
\renewcommand\arraystretch{1.5}
\centering
\caption{Effect table of different filtering and feature selection methods}
\begin{tabular}{p{6.5cm} p{2.5cm} p{3cm} p{3cm}}
\hline
Feature or sample filtering method & Precision & Recall & F1 \\
\hline
Linear correlation filter feature & 0.091638095 & 0.054966524 & 0.06871581 \\
LR filtered samples & 0.09212381 & 0.055257866 & 0.069080028 \\
Crawler rules filter samples & 0.09067619 & 0.054389553 & 0.067994515 \\
Random Negative Sampling Filters Samples & 0.090685714 & 0.054395265 & 0.068001657 \\
Linear correlation + LR filtering & 0.092590476 & 0.055537783 & 0.069429963\\
\hline
\end{tabular}
\end{table}
\begin{table}[H]
\renewcommand\arraystretch{1.5}
\centering
\caption{ Effect table of different recall sets}
\begin{tabular}{p{6cm} p{3cm} p{3cm} p{3cm}}
\hline
Recall Set & Precision & Recall & F1 \\
\hline
Cart recall set & 0.08552381 & 0.051299043 & 0.064130947 \\
shopping cart + non-shopping cart & 0.090742857 & 0.054429541 & 0.068044506\\
\hline
\end{tabular}
\end{table}
\begin{table}[H]
\renewcommand\arraystretch{1.5}
\centering
\caption{Fusion effect table of different random negative sampling}
\begin{tabular}{p{6cm} p{3cm} p{3cm} p{3cm}}
\hline
negative sampling set & Precision & Recall & F1 \\
\hline
Random Negative Sampling 1 & 0.090780952 & 0.054452391 & 0.068073072 \\
Random Negative Sampling 2 & 0.090666667 & 0.05438384 & 0.067987374 \\
Random Negative Sampling 3 & 0.090619048 & 0.054355277 & 0.067951666 \\
Three sets of fusion & 0.090952381 & 0.054555218 & 0.06820162 \\
\hline
\end{tabular}
\end{table}
To verify the influence of different filtering methods and feature selection methods on the system, the GDBT model is used, the parameters are the same as above, and the effect is shown in Table 3.7. To verify the impact of different recall sets and weighted fusion on the system, the GDBT model is used, the parameters are the same as above, and the effect is shown in Table 3.8. Among them, weighted fusion refers to the use of a hierarchical mixing strategy, first taking certain data from the recall set with high hits, and then taking certain data from another recall set when the hit is low to a threshold, thereby ensuring the final selection of multiple recall sets. The data hit rate is comparable. In this experiment, after verification and tuning, 88,500 samples were taken from the shopping cart recall set and 16,500 samples were taken from the non-shopping cart recall set during fusion.
To verify the impact of different model fusions on the system, the GDBT model is used, and the parameters are the same as above. The difference between the models is mainly due to different random negative sampling, and the effects are shown in Table 3.9.
In the end, the online (validated on December 19) F1 value obtained by the fusion of different models was 8.11\%.
The above experimental results show that:
$\textcircled{1}$ With the increase of features, the model effect will gradually improve. The effect of the improvement is large at the beginning, and then the effect is gentle, but the generalization error does not decrease. This is why the sliding window sampling is used to increase the number of samples, making the model less likely to overfit , and the features added afterward have not been greatly improved, indicating that a lot of feature information has been redundant.
$\textcircled{2}$ Fusion when the features are good enough is helpful to improve the system effect, but when some single models perform poorly, the fusion will reduce the effect of the system.
$\textcircled{3}$ When the feature has many redundant features, the linear correlation feature selection will greatly improve the system effect because of the redundant features. Using the waterfall hybrid strategy, that is, using the model for data filtering, performs well.
$\textcircled{4}$For recall sets with different hit rates, it is not possible to simply fuse them on average. Instead, a hierarchical hybrid strategy should be used, and different fusion weights should be set according to the hit rate to achieve the effect of improving the model.
$\textcircled{5}$ When the effect of the single model reaches the standard, the fusion model will get a certain improvement in effect, but if the difference between the models is not large, the effect will not be significantly improved.
(2) Experiment results of collocation recommendation system based on deep learning
This experiment uses the MAP index as the evaluation index. This indicator is selected because the collocation set is a relatively unpredictable set, and if the accuracy rate is selected as the evaluation standard, the same problem as the previous recommendation system will occur. If the F1 indicator is selected, the score is likely to be zero or very small, which cannot be Reflect the real system effect. Therefore, each product needs to provide 200 alternative products as its matching products. Here, it is necessary to take care of the accuracy of the recommendation, and to make the accurate products as far as possible to predict in front. Due to the existence of these problems and needs, choose MAP here. indicators as evaluation indicators.
The feature extraction model uses the AlexNet model. Reorder the probability output using a logistic regression model. The logistic regression model parameters are the maximum number of iterations 1000, the convergence error is 0.000001, the L1 is regular, and the coefficient is 1.
First, use the matching package set to evaluate the classification performance of the features extracted by the convolutional neural network and the logistic regression model. The evaluation method uses the cross-validation technique, and a part of the samples are reserved as the validation set. The validation set does not participate in the model training, and the results are shown in the table. 3.10. The experimental results of each part of the whole recommendation system are shown in Table 3.11. The final online MAP value of this experiment was 4.66\%.
The experimental results show that when the samples of the verification set are balanced, the features extracted by the convolutional neural network will produce a better classification effect in the logistic regression model, especially when the training results are close to the verification results, which indicates that the model is not overfitting. Generally speaking, these features have a strong ability to describe products, but when used in the whole system, although the improvement is not large, the main reason for this phenomenon is that the hit rate of the recall set is too low, and the hit rate is too low. Higher recall set triggering techniques should improve system performance.
\begin{table}[H]
\renewcommand\arraystretch{1.5}
\centering
\caption{CNN network extraction feature effect table}
\begin{tabular}{p{3.5cm} p{3.5cm} p{3.5cm} p{3.5cm}}
\hline
& Correct rate & Recall rate & F1 \\
\hline
Validation results & 0.649624425 & 0.250524847 & 0.361600158 \\
training results & 0.650327156 & 0.256068934 & 0.367452118\\
\hline
\end{tabular}
\end{table}
\begin{table}[H]
\renewcommand\arraystretch{1.5}
\centering
\caption{The effect table of each part of the recommendation system}
\begin{tabular}{p{7.5cm} p{7.5cm}}
\hline
method & MAP \\
\hline
User history recall set & 0.02586574 \\
Commodity tfidf similarity recall set & 0.02506066 \\
Fusion of two recall sets & 0.04474209\\
CNN correction &0.0466117\\
\hline
\end{tabular}
\end{table}
\section*{References}
[1] Ma Yani. Design and implementation of scoring system based on intelligent recommendation model [J]. Microcomputer Application, 2019, 35(03): 70-72.
[2] Yu Wei, Xu Dehua. Overview and Prospect of Recommendation Algorithms [J]. Science and Technology and Innovation, 2019, (04): 50-52.
[3] Lin Jianhong. Analysis of the application of artificial intelligence technology in the field of e-commerce [J]. China Business Review, 2019, (02): 19-20.
[4] Huang Maozhou. Algorithms in Recommendation Systems [J]. Technology Wind, 2019, (01): 22-23.
[5] Wang Junshu, Zhang Guoming, Hu Bin. A Review of Recommendation Algorithms Based on Deep Learning [J]. Journal of Nanjing Normal University (Engineering Technology Edition), 2018, 18(04): 33-43.
[6] Zhang Teng, Lin Guimin, Qiu Lida, Liu Chaoming, Wei Yujing. Traffic data analysis and application based on machine learning [J]. Modern Information Technology, 2018, 2(12): 16-18.
[7] Chen Guo, Zhou Zhifeng, Yang Xiaobo, Wang Cheng, Ouyang Chunping. Design and implementation of a product recommendation system based on face recognition [J]. Computer Age, 2018, (11): 52-55.
[8] Shi Hongyi. Overview of Machine Learning [J]. Communication World, 2018, (10): 253-254.
[9] Lin Haining. Recommendation system using Google Tensorflow [J]. Computer Knowledge and Technology, 2018, 14(28): 195-196.
[10] Zhong Wei, Li Zhichen. Research on Network Education System Based on Machine Learning [J]. Journal of Communications, 2018, 39(S1): 135-140.
[11] Li Jinzhong, Liu Guanjun, Yan Chungang, Jiang Changjun. Research Progress and Prospects of Ranking Learning [J]. Chinese Journal of Automation, 2018, 44(08): 1345-1369.
[12] Zhang Quangui, Li Zhiqiang, Zhang Xinxin, Cao Zhiqiang. Recommendation of Deep Joint Learning Integrating Metadata and Attention Mechanism [J]. Computer Application Research, 2019, (11): 1-3.
[13] Kuang Wenbo, Tong Wenjie. On the algorithm model design of video intelligent recommendation [J]. News and Communication, 2018, (15): 4-9.
[14] Xiao Shibo, Guo Xiuying. Research on document personalized recommendation system based on user characteristics [J]. Network New Media Technology, 2018, 7(04): 24-33.
[15] Lin Fei, Zhang Zhan. A heap model based on downsampling to predict the learning outcomes of large-scale network courses [J]. Computer Applications and Software, 2018, 35(07): 131-137.
[16] Wang Xidian, Wang Lei, Long Quan, Xue Yang. Artificial intelligence and its application in network optimization operation and maintenance [J]. Telecommunications Engineering Technology and Standardization, 2018, 31(07): 81-86.
[17] Lu Langlang, Yuan Qingda, Ling Yuanjun, Lin Yunpeng, Wang Haoyu. Design of Intelligent Recommendation Algorithm Based on Latent Semantic Model [J]. Science and Technology Communication, 2018, 10(13): 124-126
\end{spacing}
\end{document}
|
\section{Introduction}
\label{secIntro}
Membrane proteins play an important role in many essential biological processes, such as signaling, cell shape regulation, and the exchange of molecules between the interior and exterior of cells or between intracellular compartments. Membrane proteins spanning the lipid bilayer are characterized by large hydrophobic regions that approximately match up with the thickness of the lipid bilayer hydrophobic core \cite{engelman05,vanmeer08,Phillips_2013,yeagle_2016,sezgin17,Sych_Mely_Romer_2018}. However, distinct membrane proteins often show distinct hydrophobic thicknesses, and transitions in protein conformational state can change the protein's hydrophobic thickness. Moreover, the lipid composition in cell membranes tends to be highly heterogeneous, with distinct lipids often showing distinct unperturbed lipid bilayer thicknesses. As a result, membrane proteins are generally expected to show a (modest) hydrophobic mismatch with the surrounding lipid bilayer, resulting in protein-induced lipid bilayer thickness deformations \cite{Andersen_Koeppe_2007,Phillips_Ursell_Wiggins_Sens_2009,Huang_1986,Dan_Pincus_Safran_1993,Dan_Berman_Pincus_Safran_1994,Wiggins_Phillips_2005,Ursell_Kondev_Reeves_Wiggins_Phillips_2008,mondal11,Kahraman_Koch_Klug_Haselwandter_PRE_2016,Argudo_Bethel_Marcoline_Wolgemuth_Grabe_2017}. Membrane proteins may also curve the lipid bilayer membrane without perturbing the lipid bilayer thickness \cite{canham70,Helfrich_1973,evans74,zimmerberg06,weikl18,young22}. The energy cost of such protein-induced lipid bilayer deformations depends on the protein shape and conformational state, the lipid composition, membrane mechanical properties such as membrane tension, as well as membrane organization, and can thus regulate, or even determine, membrane protein function.
Membrane elasticity theory provides a beautiful framework for the quantitative description of protein-induced lipid bilayer deformations with, at least in the most basic models, all physical parameters being determined directly from experiments \cite{Andersen_Koeppe_2007,Phillips_Ursell_Wiggins_Sens_2009,Huang_1986,Dan_Pincus_Safran_1993,Dan_Berman_Pincus_Safran_1994,Wiggins_Phillips_2005,Ursell_Kondev_Reeves_Wiggins_Phillips_2008,mondal11,Kahraman_Koch_Klug_Haselwandter_PRE_2016,Argudo_Bethel_Marcoline_Wolgemuth_Grabe_2017,canham70,Helfrich_1973,evans74,zimmerberg06,weikl18,young22,fournier99,Rawicz_Olbrich_McIntosh_Needham_Evans_2000}. As a result, membrane elasticity theory yields definite predictions for the energy cost of protein-induced lipid bilayer deformations and, hence, the coupling between lipid bilayer mechanics and membrane protein function, allowing direct comparisons between theoretical predictions and experimental measurements. Over the past two decades, breakthroughs in membrane protein crystallography and, more recently, cryo-electron microscopy have yielded enormous insight into the shape of membrane proteins. However, analytic treatments of the membrane elasticity theory of protein-induced lipid bilayer deformations have largely focused on idealized protein shapes with a circular cross section, and on perturbative solutions for proteins with non-circular cross section. The objective of this article is to develop, describe, test, and apply a boundary value method (BVM) that permits the construction of non-perturbative analytic solutions of protein-induced lipid bilayer deformations for non-circular protein cross sections. This BVM allows for constant as well as variable boundary conditions along the bilayer-protein interface.
\begin{figure}[t!]
\center
\includegraphics[width=\columnwidth]{DeformationSurfaces.pdf}
\caption{Protein-induced lipid bilayer thickness deformations for selected families of protein shapes: (a) Clover-leaf protein cross section with five-fold symmetry, constant protein hydrophobic thickness, and zero bilayer-protein contact slope, (b) polygonal protein cross section with six-fold symmetry, constant protein hydrophobic thickness, and constant bilayer-protein contact slope $U'=0.3$, (c) clover-leaf protein cross section with three-fold symmetry, a five-fold symmetric (sinusoidal) variation in protein hydrophobic thickness, and zero bilayer-protein contact slope, and (d) polygonal protein cross section with seven-fold symmetry, constant protein hydrophobic thickness, and a three-fold symmetric (sinusoidal) variation in the bilayer-protein contact slope. The color map and purple surfaces show the positions of the upper and lower lipid bilayer leaflets, respectively. The bilayer-protein boundaries are color-coded according to their symmetries (see also Fig.~\ref{fig:3} in Sec.~\ref{secDeformationModel}). For panels (a) and (c) we used $\epsilon=0.2$ and $\epsilon=0.3$ in Eq.~(\ref{eqCdefclover}), respectively, and for panels (b) and (d) we used $P=5$ in Eqs.~(\ref{eqDefCpoly}) and~(\ref{eqDefCpoly2}). All bilayer surfaces were calculated using the reference parameter values in Sec.~\ref{secDeformationModel} and the BVM for protein-induced lipid bilayer thickness deformations described in Sec.~\ref{secBVM}.}
\label{fig:1}
\end{figure}
In particular, we consider here four generic modes for breaking the rotational symmetry of protein-induced lipid bilayer thickness deformations, which are illustrated in Fig.~\ref{fig:1}. Inspired by observed molecular structures of membrane proteins \cite{vinothkumar2010structures,forrest2015structural}, we consider two classes of non-circular membrane protein cross sections: Clover-leaf [see Fig.~\ref{fig:1}(a)] and polygonal [see Fig.~\ref{fig:1}(b)] protein shapes. Furthermore, we allow for variations in the bilayer-protein hydrophobic mismatch [see Fig.~\ref{fig:1}(c)] as well as in the bilayer-protein contact slope [see Fig.~\ref{fig:1}(d)] along the bilayer-protein interface. Such variations in the bilayer-protein boundary conditions can arise, on the one hand, as inherent features of the protein structure or, on the other hand, as a result of, for instance, the binding of small peptides, such as spider toxins, or other molecules along the bilayer-protein interface \cite{vinothkumar2010structures,forrest2015structural,Phillips_Ursell_Wiggins_Sens_2009,suchyna04}. For each of these four classes of protein shapes we use the BVM to obtain the energy cost of protein-induced lipid bilayer thickness deformations, and test these results against corresponding numerical solutions obtained through the finite element method (FEM) for bilayer thickness deformations \cite{Kahraman_Klug_Haselwandter_2014,Kahraman_Koch_Klug_Haselwandter_SR_2016,Kahraman_Koch_Klug_Haselwandter_PRE_2016}. Inspired by the BVM, we develop a simple analytic approximation of the bilayer thickness deformation energy associated with the general protein shapes illustrated in Fig.~\ref{fig:1}, and investigate the limits of applicability of this analytic approximation. On this basis, we explore how protein shape couples to protein-induced lipid bilayer thickness deformations, and thus affects protein oligomerization and transitions in protein conformational state.
This article is organized as follows. Section~\ref{secDeformationModel} summarizes the elasticity theory of protein-induced lipid bilayer thickness deformations. In Sec.~\ref{secBVM} we describe in detail the BVM for bilayer thickness deformations, test this BVM against FEM solutions, and discuss how the BVM can be used to calculate protein-induced lipid bilayer thickness deformations, and their associated elastic energy, for general protein shapes. On this basis, we develop in Sec.~\ref{secAnalytic} a simple analytic scheme for estimating the energy of protein-induced lipid bilayer thickness deformations for membrane proteins with non-circular cross sections. In Sec.~\ref{secDependShape} we test this analytic approximation against BVM solutions, and survey the dependence of the bilayer thickness deformation energy on membrane protein shape. In Sec.~\ref{secStability} we explore some implications of these results for the self-assembly of protein oligomers and transitions in protein conformational state. A summary and conclusions are provided in Sec.~\ref{secConclusion}.
\section{Modeling protein-induced bilayer thickness deformations}
\label{secDeformationModel}
The preferred hydrophobic thickness of lipid bilayers depends strongly on the lipid chain length \cite{engelman05,vanmeer08,Phillips_2013,yeagle_2016,sezgin17,Sych_Mely_Romer_2018} while different membrane proteins, and even different conformational states of the same membrane protein, often have distinct hydrophobic thicknesses. For membrane proteins that offer a rigid interface to the lipid bilayer, the lipid bilayer thickness tends to deform in the vicinity of the membrane protein so as to achieve hydrophobic matching at the bilayer-protein interface \cite{Andersen_Koeppe_2007,Phillips_Ursell_Wiggins_Sens_2009,Huang_1986,Dan_Pincus_Safran_1993,Dan_Berman_Pincus_Safran_1994,Wiggins_Phillips_2005,Ursell_Kondev_Reeves_Wiggins_Phillips_2008,mondal11,Kahraman_Koch_Klug_Haselwandter_PRE_2016,Argudo_Bethel_Marcoline_Wolgemuth_Grabe_2017}. The resulting protein-induced lipid bilayer thickness deformations can result in a pronounced dependence of protein conformational state, and protein function, on lipid chain length \cite{Huang_1986,mobashery97,Nielsen_Goulian_Andersen_1998,Perozo_Kloda_Cortes_Martinac_2002,Wiggins_Phillips_2004,yuan17}. The purpose of this section is to summarize the elasticity theory of protein-induced lipid bilayer thickness deformations \cite{Andersen_Koeppe_2007,Phillips_Ursell_Wiggins_Sens_2009,Huang_1986,Dan_Pincus_Safran_1993,Dan_Berman_Pincus_Safran_1994,Wiggins_Phillips_2005,Ursell_Kondev_Reeves_Wiggins_Phillips_2008,mondal11,Kahraman_Koch_Klug_Haselwandter_PRE_2016,fournier99,Haselwandter_Phillips_EPL_2013,Haselwandter_Phillips_PLOS_2013}. We first outline the standard elasticity theory of lipid bilayer thickness deformations (see Sec.~\ref{subsecFieldandEnergy}). We then describe how protein shape couples to lipid bilayer thickness, and discuss the models of protein shape considered in this article (see Sec.~\ref{subsecProteinShape}).
\subsection{Elasticity theory of lipid bilayer deformations}
\label{subsecFieldandEnergy}
Lipid bilayer thickness deformations tend to decay rapidly, with a characteristic decay length $\approx 1$~nm \cite{Ursell_Kondev_Reeves_Wiggins_Phillips_2008,Phillips_Ursell_Wiggins_Sens_2009}. When modeling protein-induced lipid bilayer thickness deformations it is therefore convenient to represent the positions of the two lipid bilayer leaflets in the Monge parameterization of surfaces, $h_\pm=h_\pm(x,y)$, with Cartesian coordinates $(x,y)$ (see Fig.~\ref{fig:2}). It is instructive to express $h_+(x,y)$ and $h_-(x,y)$ in terms of the midplane deformation field $h=h(x,y)$,
\begin{align} \label{eqhdef}
&h = \frac{h_{+} + h_{-}}{2}\,,
\end{align}
and in terms of the thickness deformation field $u=u(x,y)$,
\begin{align} \label{equdef}
&u = \frac{h_{+} - h_{-} - 2a}{2}\,,
\end{align}
where $a$ is one-half the unperturbed lipid bilayer thickness (Fig.~\ref{fig:2}). The value of $a$ depends on, for instance, the chain length of the lipid species under consideration, and can be directly measured in experiments \cite{Rawicz_Olbrich_McIntosh_Needham_Evans_2000,Andersen_Koeppe_2007,Phillips_Ursell_Wiggins_Sens_2009}.
\begin{figure}[t!]{\includegraphics[width=0.96\columnwidth]{DeformationViews.pdf}}
\caption{Notation used for the calculation of protein-induced lipid bilayer thickness deformations in (a) angled and (b) side views. As an example, we consider here a membrane protein with a non-circular (clover-leaf) bilayer-protein boundary curve, $C(\theta)$, constant hydrophobic thickness, $W(\theta)=W_0$, and zero bilayer-protein contact slope, $U^\prime(\theta)=0$. The positions of the upper and lower lipid bilayer leaflets are denoted by $h_+$ and $h_-$, from which the bilayer midplane and bilayer thickness deformation fields $h$ and $u$ can be obtained via Eqs.~(\ref{eqhdef}) and~(\ref{equdef}), respectively. We denote one-half the unperturbed bilayer thickness by $a$, resulting in a hydrophobic mismatch $U=W/2-a$ at the bilayer-protein interface. The unit vectors $\mathbf{\hat{t}}$ and $\mathbf{\hat{n}}$ denote the directions tangential and perpendicular (pointing towards the protein) to the bilayer-protein boundary, respectively.}
\label{fig:2}
\end{figure}
The membrane elasticity theory describing the shape of lipid bilayers \cite{Seifert_1997,safran03,boal12,Phillips_2013} dates back to the classic work of W. Helfrich \cite{Helfrich_1973}, P.~B. Canham \cite{canham70}, E.~A. Evans \cite{evans74}, and H.~W. Huang \cite{Huang_1986}. Interestingly, one finds that the elastic energies governing $h$ and $u$ in Eqs.~(\ref{eqhdef}) and~(\ref{equdef}) decouple from each other to leading order \cite{fournier99,Wiggins_Phillips_2005}. In the most straightforward model of bilayer-protein interactions
\cite{Huang_1986,goulian93,Dan_Pincus_Safran_1993,Dan_Berman_Pincus_Safran_1994,Weikl_Kozlov_Helfrich_1998,Andersen_Koeppe_2007,Phillips_Ursell_Wiggins_Sens_2009,boal12,safran03,Seifert_1997,Kahraman_Koch_Klug_Haselwandter_PRE_2016,Wiggins_Phillips_2005,Ursell_Kondev_Reeves_Wiggins_Phillips_2008,mondal11}, the energy cost of protein-induced lipid bilayer midplane deformations is then captured by the classic Helfrich-Canham-Evans energy \cite{canham70,Helfrich_1973,evans74}, and the energy cost of protein-induced lipid bilayer thickness deformations is given by \cite{Huang_1986,Dan_Pincus_Safran_1993,Dan_Berman_Pincus_Safran_1994,Wiggins_Phillips_2005,Ursell_Kondev_Reeves_Wiggins_Phillips_2008,Andersen_Koeppe_2007,Kahraman_Koch_Klug_Haselwandter_PRE_2016,mondal11}
\begin{equation} \label{eqGArea}
G = \frac{1}{2} \int dx dy \left[K_b \left(2 H \right)^2 + K_t \left(\frac{u}{a} \right)^{2} \right]\,,
\end{equation}
where the integral runs over the (in-plane) lipid bilayer surface, $K_b$ is the lipid bilayer bending rigidity, the mean curvature $H=\frac{1}{2}\nabla^{2} u$, and $K_t$ is the bilayer thickness deformation modulus. The terms $K_b \left(\nabla^{2} u \right)^2$ and $K_t \left(u/a \right)^{2}$ in Eq.~(\ref{eqGArea}) provide lowest-order descriptions of the energy cost of bilayer bending and the compression/expansion of the bilayer hydrophobic core, respectively. For simplicity, we assume in Eq.~(\ref{eqGArea}) that the bilayer is under negligible lateral membrane tension. A nonzero membrane tension could be readily included in Eq.~(\ref{eqGArea}) \cite{Haselwandter_Phillips_PLOS_2013,Kahraman_Koch_Klug_Haselwandter_PRE_2016}. We also assume in Eq.~(\ref{eqGArea}) that the lipids forming the bilayer have zero intrinsic curvature. A nonzero lipid intrinsic curvature could also be included in the formalism employed here \cite{Dan_Pincus_Safran_1993,Dan_Berman_Pincus_Safran_1994}.
Similarly as the unperturbed lipid bilayer thickness $2a$, the effective parameters $K_b$ and $K_t$ in Eq.~(\ref{eqGArea}) characterizing the elastic properties of the bilayer membrane depend on the lipid composition, and can be directly measured in experiments \cite{Rawicz_Olbrich_McIntosh_Needham_Evans_2000,Andersen_Koeppe_2007,Phillips_Ursell_Wiggins_Sens_2009}. Typical values of $K_{b}, K_{t}$, and $a$ for cell membranes are $K_{b}=20$~$k_{B}T$, $K_{t}=60$~$k_{B}T$/nm$^{2}$, and $a=1.6$~nm \cite{Niggemann_Kummrow_Helfrich_1995, Rawicz_Olbrich_McIntosh_Needham_Evans_2000, Ursell_Kondev_Reeves_Wiggins_Phillips_2008, Pan_Tristram-Nagle_Kucerka_Nagle_2008}. Unless stated otherwise, we use here these values of $K_{b}, K_{t}$, and $a$. When studying the dependence of protein-induced bilayer thickness deformations on lipid chain length we follow Refs.~\cite{Rawicz_Olbrich_McIntosh_Needham_Evans_2000,Wiggins_Phillips_2005,Shrestha_Kahraman_Haselwandter_2020} and assume, for simplicity, a linear dependence of $a$ on lipid chain length:
\begin{equation} \label{eqDefm}
a = \frac{1}{2} \left(0.13 m + 1.7\right)\,\text{nm}\,.
\end{equation}
The integer $m$ in Eq.~(\ref{eqDefm}) denotes the lipid chain length (number of carbon atoms comprising each lipid chain), with the approximate range $13 \leq m \leq 22$ for phospholipids in cell membranes \cite{engelman05,vanmeer08,sezgin17,Rawicz_Olbrich_McIntosh_Needham_Evans_2000}. For simplicity, we take $K_b$ and $K_t$ to be independent of $m$ while noting that, in general, $K_b$ and $K_t$ may have a (weak) dependence on $m$ \cite{Rawicz_Olbrich_McIntosh_Needham_Evans_2000}.
The effective parameters $K_{b}, K_{t}$, and $a$ in Eq.~(\ref{eqGArea}) yield the characteristic length scale
\begin{equation} \label{eqDeflambda}
\lambda = \left(\frac{a^{2}K_{b}}{K_{t}}\right)^{1/4}\,,
\end{equation}
which corresponds to the characteristic decay length of bilayer thickness deformations \cite{Wiggins_Phillips_2005}. As alluded to above, we have $\lambda \approx 1$~nm \cite{Ursell_Kondev_Reeves_Wiggins_Phillips_2008,Phillips_Ursell_Wiggins_Sens_2009}. Similarly, the bilayer bending rigidity $K_b$ defines a characteristic energy scale in Eq.~(\ref{eqGArea}). It is therefore convenient to recast the bilayer thickness deformation energy in Eq.~(\ref{eqGArea}) in terms of the characteristic spatial and energy scales, $\lambda$ and $K_b$. Unless stated otherwise, we use here a dimensionless form of Eq.~(\ref{eqGArea}) such that $G\rightarrow \bar G K_{b}$, $x\rightarrow \bar x \lambda$, $y\rightarrow \bar y \lambda$, $u\rightarrow \bar u \lambda$, $a\rightarrow \bar a \lambda$, and $K_{t}\rightarrow \bar K_{t} K_b/\lambda^{2}$, resulting in
\begin{equation} \label{eqGAreaDL}
\bar G = \frac{1}{2} \int d\bar x d\bar y \left[\left(\bar \nabla^{2} \bar u \right)^2 + \bar u^{2} \right]\,,
\end{equation}
where $\bar \nabla=\lambda \nabla$.
We assume that, for a given protein conformational state, the dominant bilayer thickness deformation field $\bar u(\bar x,\bar y)$ minimizes Eq.~(\ref{eqGAreaDL}) subject to suitable boundary conditions \cite{Huang_1986,Dan_Pincus_Safran_1993,Dan_Berman_Pincus_Safran_1994,Wiggins_Phillips_2005,Ursell_Kondev_Reeves_Wiggins_Phillips_2008,Andersen_Koeppe_2007,Kahraman_Koch_Klug_Haselwandter_PRE_2016}. The Euler-Lagrange equation associated with Eq.~(\ref{eqGAreaDL}) is given by
\begin{equation} \label{eqHelmholtz}
\left(\bar \nabla^{2} - \bar \nu_{+} \right) \left(\bar \nabla^{2} - \bar \nu_{-} \right) \bar u = 0
\end{equation}
with $\bar \nu_{\pm} = \pm \mathbf{i}$, where $\mathbf{i}$ is the imaginary unit. To construct the general solution of Eq.~(\ref{eqHelmholtz}) for protein-induced bilayer thickness deformations it is useful to transform $\left(\bar x,\bar y\right)$ to the dimensionless polar coordinates $\left(\bar r,\theta\right)$ with the protein center as the origin of the polar coordinate system. Assuming that protein-induced bilayer thickness deformations are localized about the protein we have $\bar u \to 0$ as $\bar r \to \infty$ \cite{Huang_1986,Nielsen_Goulian_Andersen_1998,Nielsen_Andersen_2000}, in which case Eq.~(\ref{eqHelmholtz}) yields \cite{Haselwandter_Phillips_EPL_2013,Haselwandter_Phillips_PLOS_2013,Kahraman_Koch_Klug_Haselwandter_PRE_2016}
\begin{align} \label{equEigenSuper}
\bar u\left(\bar r,\theta\right) = \bar f^{+}\left(\bar r,\theta\right) + \bar f^{-}\left(\bar r,\theta\right)\,,
\end{align}
where the Fourier-Bessel series
\begin{align} \label{eqEigenFuncs}
\begin{split}
\bar f^{\pm}(\bar r,\theta) = & A^{\pm}_{0}K_{0} \left(\sqrt{\bar \nu_\pm} \bar r \right) + \\
\sum^{\infty}_{n = 1} \bigg[ A^{\pm}_{n} K_{n}&\left( \sqrt{\bar \nu_\pm} \bar r\right) \cos\left( n \theta \right) + B^{\pm}_{n} K_{n}\left( \sqrt{\bar \nu_\pm} \bar r\right) \sin\left( n \theta \right) \bigg]\,,
\end{split}
\end{align}
in which the $K_{n}$ are the modified Bessel functions of the second kind \cite{Abramowitz_Stegun_1964} and the values of the coefficients $A^{\pm}_{0}$, $A^{\pm}_{n}$, and $B^{\pm}_{n}$ are determined by the bilayer-protein boundary conditions.
The bilayer thickness deformation energy in Eq.~(\ref{eqGAreaDL}) is conveniently evaluated for the stationary bilayer thickness deformation field in Eq.~(\ref{equEigenSuper}) by noting that, via Eq.~(\ref{eqHelmholtz}), Eq.~(\ref{eqGAreaDL}) can be transformed to a line integral along the bilayer-protein boundary $\bar C$ \cite{Wiggins_Phillips_2005,Haselwandter_Phillips_PLOS_2013,Kahraman_Koch_Klug_Haselwandter_PRE_2016} (Fig.~\ref{fig:2}). For simplicity, we thereby take the bilayer-protein boundary to be specified by the polar curve $\mathbf{\bar r}=\bar C(\theta) \mathbf{\hat{r}}$, where $\mathbf{\hat{r}}$ is the radial unit vector pointing away from the protein center. We thus have
\begin{equation} \label{eqGLine}
\bar G = \frac{1}{2} \int_0^{2 \pi} d \theta \, \bar l \, \mathbf{\hat{n}} \cdot \left( \bar \nabla \bar u \bar \nabla^{2} \bar u - \bar u \bar \nabla^{3} \bar u \right) \big |_{\bar r=\bar C(\theta)}\,,
\end{equation}
where the line element $\bar l=\sqrt{\left[\bar C(\theta)\right]^{2}+ \left[\bar C'(\theta)\right]^{2}}$, and the unit vector $\mathbf{\hat{n}}$ is normal to the tangent of $\mathbf{\bar r}=\bar C(\theta) \mathbf{\hat{r}}$ and points towards the protein (Fig.~\ref{fig:2}). Note that the term in brackets in Eq.~(\ref{eqGLine}) may be interpreted as a bilayer-protein line tension along the bilayer-protein boundary \cite{Wiggins_Phillips_2004,Wiggins_Phillips_2005,Kahraman_Koch_Klug_Haselwandter_PRE_2016}. The normal vector $\mathbf{\hat{n}}$ in Eq.~(\ref{eqGLine}) is obtained by differentiating the bilayer-protein boundary curve $\mathbf{\bar r}=\bar C(\theta) \mathbf{\hat{r}}$ with respect to $\theta$ and rotating the resulting tangent vector by $\pi/2$ so as to point towards the protein,
\begin{equation} \label{eqndef}
\mathbf{ \hat{n} } = \frac{- \bar C(\theta) \pmb{ \hat{r} } + \bar C'(\theta) \pmb{ \hat{\theta} } } { \bar l } \,,
\end{equation}
where we have noted that the (counterclockwise) angular unit vector $\pmb{\hat{\theta}}=d \mathbf{\hat{r}}/d\theta$ in polar coordinates (Fig.~\ref{fig:2}). Equation~(\ref{eqGLine}) with Eq.~(\ref{eqndef}) allows calculation of $\bar G$ in Eq.~(\ref{eqGAreaDL}) and, hence, $G$ in Eq.~(\ref{eqGArea}) along a one-dimensional curve rather than over a two-dimensional surface, which provides a computationally efficient method for evaluating~$\bar G$.
\subsection{Modeling protein shape}
\label{subsecProteinShape}
The coefficients $A^{\pm}_{0}$, $A^{\pm}_{n}$, and $B^{\pm}_{n}$ in Eq.~(\ref{eqEigenFuncs}) are fixed by the boundary conditions at the bilayer-protein interface. The general mathematical form of these boundary conditions, which encode the key protein properties governing protein-induced lipid bilayer thickness deformations, follows from the calculus of variations \cite{Hilbert1953,vanbrunt04}. Based on previous work on protein-induced bilayer thickness deformations \cite{Huang_1986,Dan_Pincus_Safran_1993,Dan_Berman_Pincus_Safran_1994,Nielsen_Goulian_Andersen_1998,Andersen_Koeppe_2007,Phillips_Ursell_Wiggins_Sens_2009}, we assume that the lipid bilayer thickness deforms in the vicinity of membrane proteins so as to achieve hydrophobic matching at the bilayer-protein interface. We thus have the boundary condition
\begin{equation} \label{eqDefU}
\bar u(\bar r, \theta) \big|_{\bar r=\bar C(\theta)} = \bar U(\theta)\,,
\end{equation}
where the bilayer-protein hydrophobic mismatch
\begin{equation} \label{eqUWa}
\bar U(\theta) = \frac{1}{2}\left[\bar W(\theta) - 2 \bar a\right] \,,
\end{equation}
in which $W(\theta) = \lambda \bar W(\theta)$ is the protein hydrophobic thickness along the bilayer-protein boundary (Fig.~\ref{fig:2}). For large enough magnitudes of $U$, membrane proteins or lipids may expose parts of their hydrophobic regions to water, which would amount to an offset of $\bar W$ in Eq.~(\ref{eqUWa}). For a given membrane protein, $W(\theta)$ can be estimated from the molecular structure of the membrane protein \cite{Andersen_Koeppe_2007,Phillips_Ursell_Wiggins_Sens_2009,Haselwandter_Phillips_PLOS_2013,Kahraman_Koch_Klug_Haselwandter_SR_2016,Kahraman_Koch_Klug_Haselwandter_PRE_2016} and/or computer simulations \cite{Elmore_Dougherty_2003,Argudo_Bethel_Marcoline_Wolgemuth_Grabe_2017}. We explore here protein-induced bilayer thickness deformations for generic models of $W(\theta)$ inspired by the molecular structure of the mechanosensitive channel of large conductance (MscL) \cite{Chang_Spencer_Lee_Barclay_Rees_1998,Doyle_1998,Sukharev_Durell_Guy_2001,Perozo_Cortes_Sompornpisut_Kloda_Martinac_2002}.
In addition to Eq.~(\ref{eqDefU}), it is also necessary to specify boundary conditions on the (normal) derivative of $u$ at the bilayer-protein interface \cite{Hilbert1953,vanbrunt04}. Based on Refs.~\cite{Huang_1986,Nielsen_Goulian_Andersen_1998,Argudo_Bethel_Marcoline_Wolgemuth_Grabe_2017,Wiggins_Phillips_2004,Wiggins_Phillips_2005,Ursell_Kondev_Reeves_Wiggins_Phillips_2008} we generally focus on the fixed-value boundary condition
\begin{align} \label{eqDefUp}
\mathbf{ \hat{n} } \cdot \bar \nabla \bar u(\bar r, \theta) \big|_{\bar r=\bar C(\theta)} = \bar U'(\theta)\,,
\end{align}
but also explore choices for $\bar U'(\theta)$ minimizing the bilayer thickness deformation energy. We allow for constant as well as varying $\bar U'(\theta)$ in Eq.~(\ref{eqDefUp}).
For a (hypothetical) membrane protein with a perfectly circular cross section $\bar C(\theta) =\bar R$, where $\bar R$ is the (dimensionless) protein radius, and constant $\bar U$ and $\bar U'$, the bilayer-protein boundary conditions in Eqs.~(\ref{eqDefU}) and~(\ref{eqDefUp}) are azimuthally symmetric about the protein center, and the resulting protein-induced bilayer thickness deformations also show azimuthal symmetry about the protein center \cite{Huang_1986,Nielsen_Goulian_Andersen_1998,Wiggins_Phillips_2004,Wiggins_Phillips_2005}. Equations~(\ref{eqDefU}) and~(\ref{eqDefUp}) suggest three, not mutually exclusive, modes for protein structures to break rotational symmetry, and to hence endow protein-induced bilayer thickness deformations with a non-trivial structure \cite{Haselwandter_Phillips_PLOS_2013,Haselwandter_Phillips_EPL_2013}. First, the value of $\bar U$ in Eq.~(\ref{eqDefU}) or, second, the value of $\bar U'$ in Eq.~(\ref{eqDefUp}) may vary along the bilayer-protein interface. To explore generic effects of varying $\bar U$ or $\bar U'$ on protein-induced bilayer thickness deformations we consider the bilayer-protein hydrophobic mismatch
\begin{equation} \label{eqUBC}
\bar U(\theta) = \bar U_{0} + \bar \beta \cos(w \theta)
\end{equation}
and the bilayer-protein contact slope
\begin{equation} \label{eqUpBC}
\bar U'(\theta) = \bar U'_{0} + \bar \gamma \cos(v \theta)\,,
\end{equation}
where $\bar U_{0}$ and $\bar U'_{0}$ denote the average bilayer-protein hydrophobic mismatch and bilayer-protein contact slope, $\bar \beta$ and $\bar \gamma$ denote the amplitudes of the perturbations about these average values, and $w$ and $v$ denote the protein symmetries associated with variations in $\bar U$ and $\bar U'$. Unless stated otherwise, we set here $\bar U_{0} \lambda=-0.1$~nm and $\bar \beta \lambda=0.5$~nm in Eq.~(\ref{eqUBC}) for all calculations involving a modulation in the bilayer-protein hydrophobic mismatch, and $\bar U'_{0}=0$ and $\bar \gamma=0.3$ in Eq.~(\ref{eqUpBC}) for all calculations involving a modulation in the bilayer-protein contact slope. For all scenarios considered here in which we keep $\bar U$ or $\bar U'$ constant along the bilayer-protein interface we set, unless stated otherwise, $\bar U \lambda=0.3$~nm or $\bar U'=0$. The values of $U$ employed here are in line with previous work on MscL and gramicidin channels \cite{Ursell_Kondev_Reeves_Wiggins_Phillips_2008,Chang_Spencer_Lee_Barclay_Rees_1998,Perozo_Cortes_Sompornpisut_Kloda_Martinac_2002,Argudo_Bethel_Marcoline_Wolgemuth_Grabe_2017,Nielsen_Goulian_Andersen_1998}. Note that the ranges of $\bar U$ and $\bar U'$ considered here satisfy the constraints $|U| < a$ and $|U'| < 1$ underlying the use of the leading-order energy in Eq.~(\ref{eqGArea}) to describe protein-induced bilayer thickness deformations \cite{Huang_1986,Dan_Pincus_Safran_1993,Dan_Berman_Pincus_Safran_1994,fournier99,Wiggins_Phillips_2005,Ursell_Kondev_Reeves_Wiggins_Phillips_2008,Andersen_Koeppe_2007,Kahraman_Koch_Klug_Haselwandter_PRE_2016}.
\begin{figure}[t!]{\includegraphics{CrossSectionPanels.pdf}}
\caption{Cross sections of cylindrical protein shapes (left-most panels) and (a) clover-leaf and (b) polygonal protein shapes (right panels). The clover-leaf protein cross sections in panel (a) are obtained from Eq.~(\ref{eqCdefclover}) with $\epsilon=0.07$, 0.14, 0.21, 0.28, and 0.35 (left to right) and $s=1$, 2, 3, 4, and 5 (top to bottom), with $\epsilon =0$ yielding a circular protein cross section. Note that the clover-leaf protein cross sections with $s=1$ only show small deviations from the corresponding circular protein cross section obtained with $\epsilon = 0$ in Eq.~(\ref{eqCdefclover}) (dashed curves) for the values of $\epsilon$ considered here. The polygonal protein cross sections in panel (b) are obtained from Eq.~(\ref{eqDefCpoly}) with $P=1$, 2, 3, 4, and 5 (left to right) and $s=4$, 5, 6, 7, and 8 (top to bottom). As a guide to the eye, these polygonal protein cross sections are inscribed in circles obtained with $P = 0$ in Eq.~(\ref{eqDefCpoly}) (dashed curves).}
\label{fig:3}
\end{figure}
Angular variations in $\bar C(\theta)$ along the bilayer-protein boundary $\mathbf{\bar r}=\bar C(\theta) \mathbf{\hat{r}}$ provide, in addition to Eqs.~(\ref{eqUBC}) and~(\ref{eqUpBC}), a third mode for a protein structure to break azimuthal symmetry of protein-induced bilayer thickness deformations about the protein center. Inspired by molecular structures of tetrameric and pentameric MscL \cite{Chang_Spencer_Lee_Barclay_Rees_1998,Liu2009,Haselwandter_Phillips_PLOS_2013,Kahraman_Koch_Klug_Haselwandter_SR_2016} and other membrane proteins \cite{vinothkumar2010structures,forrest2015structural}, we consider here two generic classes of protein shapes breaking rotational symmetry. On the one hand, we consider clover-leaf protein cross sections specified by
\begin{equation} \label{eqCdefclover}
\bar C(\theta) = \bar R \big{[} 1 + \epsilon \cos( s \theta ) \big{]},
\end{equation}
where $\epsilon$ parameterizes the magnitude of deviations from a circular protein cross section, $\epsilon=0$ for circular protein cross sections, and $s$ denotes the symmetry of the boundary curve [see Figs.~\ref{fig:2}(a) and~\ref{fig:3}(a)]. On the other hand, we consider (rounded) polygonal protein cross sections specified by the series
\begin{align}
\begin{split}
&\bar C(\theta) =\\
&\bar A_{\bar R}\sqrt{\Biggr{[}\sum^{P}_{ p = -P }\frac{\cos(sp + 1)\theta}{(sp + 1)^{2}}\Biggr{]}^{2} + \Biggr{[}\sum^{P}_{ p = -P }\frac{\sin(sp + 1)\theta}{(sp + 1)^{2}} \Biggr{]}^{2} }\,,
\end{split} \label{eqDefCpoly}
\end{align}
where larger $P$ yield sharper polygonal corners with $P=0$ for circular protein cross sections, $\bar A_{\bar R}$ is a rescaling factor chosen so as to ensure that the polygons are inscribed in circles of radius $\bar R$, and $s$ denotes the polygonal symmetry \cite{Robert_1994,Haselwandter_Phillips_PLOS_2013} [see Fig.~\ref{fig:3}(b)]. We have
\begin{equation} \label{eqDefCpoly2}
\bar A_{\bar R} = \frac{\bar R}{\sum^{P}_{ p = -P }\frac{1}{(sp + 1)^{2}}}\,.
\end{equation}
Unless stated otherwise, we set $P=5$ in Eqs.~(\ref{eqDefCpoly}) and~(\ref{eqDefCpoly2}). For the scenarios considered here we found that an increase in $P$ beyond $P=5$ only resulted in minor shifts in the bilayer thickness deformation energy. For all the calculations described here we use $\bar R \lambda\approx2.3$~nm in Eqs.~(\ref{eqCdefclover}) and~(\ref{eqDefCpoly2}), which approximately corresponds to the observed size of a closed state of MscL~\cite{Chang_Spencer_Lee_Barclay_Rees_1998,Wiggins_Phillips_2005}.
\section{Boundary value method for bilayer thickness deformations}
\label{secBVM}
In this section we introduce a BVM for bilayer thickness deformations, which allows calculation of protein-induced bilayer thickness deformations, and their associated elastic energy, for general protein shapes. In the following sections we use this BVM to calculate the bilayer thickness deformation field $\bar u(\bar r,\theta)$ in Eq.~(\ref{equEigenSuper}), and the resulting bilayer thickness deformation energy $G$ in Eq.~(\ref{eqGArea}), for the clover-leaf and polygonal protein shapes $\mathbf{\bar r}=\bar C(\theta) \mathbf{\hat{r}}$ in Eqs.~(\ref{eqCdefclover}) and~(\ref{eqDefCpoly}) and the boundary conditions $\bar U(\theta)$ and $\bar U'(\theta)$ in Eqs.~(\ref{eqUBC}) and~(\ref{eqUpBC}). We first provide a general formulation of the BVM for bilayer thickness deformations, and validate this BVM against exact analytic and FEM solutions (see Sec.~\ref{subsecMethod}). We then discuss how the numerical performance of the BVM for bilayer thickness deformations can be improved by employing an adaptive point distribution (APD) that results in a nonuniform distribution of boundary points for non-circular protein cross sections (see Sec.~\ref{subsubsecAPD}).
\subsection{Formulation and validation of the boundary value method}
\label{subsecMethod}
The BVM for bilayer thickness deformations takes the analytic solution for $\bar u(\bar r,\theta)$ in Eq.~(\ref{equEigenSuper}) as its starting point, and assumes that the infinite series in this general solution can be truncated at some finite order $N$:
\begin{align} \label{equEigenSuperN}
\bar u\left(\bar r,\theta\right) \approx \bar f^{+}_N\left(\bar r,\theta\right) + \bar f^{-}_N\left(\bar r,\theta\right)\,,
\end{align}
where
\begin{align} \label{eqEigenFuncsN}
\begin{split}
\bar f^{\pm}_N(\bar r,\theta) = & A^{\pm}_{0}K_{0} \left(\sqrt{\bar \nu_\pm} \bar r \right) + \\
\sum^{N}_{n = 1} \bigg[ A^{\pm}_{n} K_{n}&\left( \sqrt{\bar \nu_\pm} \bar r\right) \cos\left( n \theta \right) + B^{\pm}_{n} K_{n}\left( \sqrt{\bar \nu_\pm} \bar r\right) \sin\left( n \theta \right) \bigg]\,.
\end{split}
\end{align}
The solution in Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}) contains the $4N+2$ unknown constants $A^{\pm}_{0}$, $A^{\pm}_{n}$, and $B^{\pm}_{n}$. In the BVM for bilayer thickness deformations, we fix these coefficients by imposing the boundary conditions in Eqs.~(\ref{eqDefU}) and~(\ref{eqDefUp}) at $2N+1$ boundary points along the bilayer-protein interface. For now, we take these boundary points to be uniformly distributed along the bilayer-protein interface, with a constant arc length separating adjacent boundary points along the bilayer-protein interface. We return to the distribution of boundary points in Sec.~\ref{subsubsecAPD}.
From the boundary conditions in Eqs.~(\ref{eqDefU}) and~(\ref{eqDefUp}) we have
\begin{eqnarray} \label{eqBCmU}
\bar u(\bar r, \theta_j) \big|_{\bar r=\bar C\left(\theta_j\right)} &=& \bar U\left(\theta_j\right)\,,\\ \label{eqBCmUp}
\mathbf{ \hat{n} } \cdot \bar \nabla \bar u(\bar r, \theta_j) \big|_{\bar r=\bar C\left(\theta_j\right)} &=& \bar U'\left(\theta_j \right)\,,
\end{eqnarray}
in which $j=1,2,\dots,2N+1$ denote the boundary points along the bilayer-protein interface, where $\bar C(\theta)=\bar R$ for proteins with a circular cross section, $\bar C(\theta)$ is as in Eq.~(\ref{eqCdefclover}) for clover-leaf protein shapes, and $\bar C(\theta)$ is as in Eq.~(\ref{eqDefCpoly}) for polygonal protein shapes (Fig.~\ref{fig:3}). Equations~(\ref{eqBCmU}) and~(\ref{eqBCmUp}) amount to a linear system of equations
\begin{equation} \label{eqmeq}
\mathbf{A} \mathbf{x} = \mathbf{b}\,,
\end{equation}
where the vector $\mathbf{x}$ has dimension $4N+2$ and contains the unknown constants $A^{\pm}_{0}$, $A^{\pm}_{n}$, and $B^{\pm}_{n}$, the $4N+2$ components of the vector $\mathbf{b}$ contain the boundary conditions on the right-hand sides of Eqs.~(\ref{eqBCmU}) and~(\ref{eqBCmUp}), and $\mathbf{A}$ is a square matrix of order $4N+2$ that has the coefficients of the constants $A^{\pm}_{0}$, $A^{\pm}_{n}$, and $B^{\pm}_{n}$ on the left-hand sides of Eqs.~(\ref{eqBCmU}) and~(\ref{eqBCmUp}) as its entries. Equation~(\ref{eqmeq}) can be solved efficiently using the extensive numerical methods available for the solution of matrix equations. We employed here LU decomposition with partial pivoting to solve Eq.~(\ref{eqmeq}) for $\mathbf{x}$ \cite{Atkinson_1988,Trefethen_1997}.
\begin{figure}[t!]{\includegraphics{uniformVSadaptive.pdf}}
\caption{Percentage difference between the exact bilayer thickness deformation field along the bilayer-protein boundary and the bilayer thickness deformation field obtained from the BVM solution, $\eta_{b'}$ in Eq.~(\ref{eqetabdef}), as a function of the number of terms in the Fourier-Bessel series in Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}) for (a) uniformly distributed points along the bilayer-protein boundary and (b) the boundary point distributions implied by the APD method (see Sec.~\ref{subsubsecAPD}). For both panels we considered three-fold clover-leaf protein shapes ($s=3$) in Eq.~(\ref{eqCdefclover}) with the indicated values of $\epsilon$, $\bar R \lambda\approx2.3$~nm, and the constant $\bar U \lambda=0.3$~nm and $\bar U'=0$. In panel (b) we used, for ease of comparison, the same gap factor $\Omega=0.25$ in Eq.~(\ref{eqelldef}) for all curves.}
\label{fig:4}
\end{figure}
To quantify numerical errors in our BVM solutions it is useful to compute, based on the calculated $A^{\pm}_{0}$, $A^{\pm}_{n}$, and $B^{\pm}_{n}$ in Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}), the values of $\bar u(\bar r,\theta)$ along the bilayer-protein boundary for a given set of reference points distinct from the boundary points employed for the BVM solution in Eq.~(\ref{eqmeq}). We compile these computed boundary values of $\bar u(\bar r,\theta)$ in a vector $\mathbf{\tilde{b}}$, and compare $\mathbf{\tilde{b}}$ to the corresponding exact boundary values $\mathbf{b'}$ mandated by the boundary conditions in Eqs.~(\ref{eqDefU}) and~(\ref{eqDefUp}),
\begin{equation} \label{eqetabdef}
\eta_{b'} = 100 \times \frac{ ||\mathbf{ \tilde{b} - \mathbf{b'}}||_{L^2} }{ ||\mathbf{b'}||_{L^2} }\,,
\end{equation}
where $||\cdot||_{L^2}$ is the $L^{2}$ norm \cite{Atkinson_1988}. For all the results shown in this article, we used vectors $\mathbf{b'}$ and $\mathbf{\tilde{b}}$ with 800 components [400 components each for $\bar U(\theta)$ and $\bar U'(\theta)$] in Eq.~(\ref{eqetabdef}), which we chose for a given protein shape so as to yield reference points with a uniform spacing in $\theta$ over the interval $0 \leq \theta \leq 2 \pi$. Figure~\ref{fig:4}(a) shows $\eta_{b'}$ in Eq.~(\ref{eqetabdef}) for the clover-leaf shapes in Eq.~(\ref{eqCdefclover}) with $s=3$ and various values of $\epsilon$. As expected, we find that $\eta_{b'}$ tends to decrease with increasing $N$ in Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}), indicating that a greater accuracy of BVM solutions is obtained at larger $N$. The local minima of $\eta_{b'}$ in Fig.~\ref{fig:4}(a) correspond to values of $N$ that are multiples of $s$, which suggests that the accuracy of the BVM is improved if $N$ matches the protein symmetry. Figure~\ref{fig:4}(b) indicates that the convergence of BVM solutions with increasing $N$ can be improved substantially through an APD that allows for a nonuniform distribution of boundary points, which we discuss in Sec.~\ref{subsubsecAPD}.
We performed our BVM calculations in C++ using the arbitrary precision numerical library \textit{Arb} \cite{Johansson_arb_2017}. Unless stated otherwise, we allowed for sufficient numerical precision so that the boundary error $\eta_{b'} \leq 0.1 \%$ in Eq.~(\ref{eqetabdef}) and we obtained changes in $G$ and $\eta_{b'}$ of no more than $10^{-5}\%$ as the numerical precision was increased. We generated all figures in \textit{MATLAB} \cite{MATLAB_2021}. To speed up our calculations, we multi-threaded some of the source code of the \textit{Arb} library \cite{dagum_omp_1998}. Appendix~\ref{AppA} provides a more in-depth description of our computational implementation of the BVM, and discusses possible issues with the numerical solution of Eq.~(\ref{eqmeq}) arising from floating point errors and numerical instabilities. For the polygonal protein shapes considered here, with $P=5$ in Eq.~(\ref{eqDefCpoly}) with Eq.~(\ref{eqDefCpoly2}), and for clover-leaf protein shapes with large $s$ and/or $\epsilon$ in Eq.~(\ref{eqCdefclover}) we found it convenient to perform the BVM calculations with numerical precision greater than double precision (64 bits). In Appendix~\ref{AppB} we illustrate the extent to which double precision calculations could be used to approximate the BVM results described here.
In our BVM calculations we evaluate the bilayer thickness deformation energy $G$ in Eq.~(\ref{eqGArea}) by numerically computing Eq.~(\ref{eqGLine}) using the same 400 reference points employed to calculate $\eta_{b'}$ in Eq.~(\ref{eqetabdef}). To this end, we approximate $\bar G$ in Eq.~(\ref{eqGLine}) through Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}),
\begin{equation} \label{eqGBVM}
\bar G \approx \frac{\mathbf{i}}{2}\int^{2\pi}_{ 0 } d \theta \bar l
\left( \bar U' - \bar U\mathbf{\hat{n}}\cdot \bar \nabla \right) \left( \bar f^{+}_{N} - \bar f^{-}_{N} \right) \big |_{\bar r=\bar C(\theta)}\,,
\end{equation}
where we have used Eq. (\ref{eqHelmholtz}) and the boundary conditions in Eqs.~(\ref{eqDefU}) and~(\ref{eqDefUp}). We have confirmed that, within the numerical accuracy used here, identical results for $\eta_{b'}$ and $\bar G$ are obtained with more than 400 reference points. Note that $\bar l$ in Eq.~(\ref{eqGBVM}) is real and that Eq.~(\ref{eqGBVM}) must evaluate to a real number, which means that the remaining terms in the integrand in Eq.~(\ref{eqGBVM}) evaluate to a purely imaginary number.
\begin{figure}[t!]{\includegraphics{FiniteSeriesBenchmarks.pdf}}
\caption{Percentage difference between exact analytic and FEM (red curves) or BVM (blue curves) solutions for the bilayer thickness deformation energy, $\eta_{G}$ in Eq.~(\ref{eqetaGdef}), as a function of the average edge size $\langle L \rangle$ used in the FEM solution (upper axes) or the number of terms in Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}) used in the BVM solution (lower axes) for (a) a cylindrical protein with $\bar R \lambda=2.3$~nm and $\bar U \lambda=0.3$~nm and (b) a crown-shaped protein with $\bar R \lambda=2.3$~nm, $\bar U_0 \lambda=-0.1$~nm, $\bar \beta \lambda=0.5$~nm, and $w=5$ in Eq.~(\ref{eqUBC}). We set $\bar U'=0$ for both panels.}
\label{fig:5}
\end{figure}
We validated the BVM for bilayer thickness deformations against exact analytic solutions obtained for proteins with circular cross sections \cite{Huang_1986,Nielsen_Goulian_Andersen_1998,Wiggins_Phillips_2005,Haselwandter_Phillips_PLOS_2013,Haselwandter_Phillips_EPL_2013,Kahraman_Koch_Klug_Haselwandter_PRE_2016} and against FEM solutions \cite{Kahraman_Klug_Haselwandter_2014,Kahraman_Koch_Klug_Haselwandter_PRE_2016,Kahraman_Koch_Klug_Haselwandter_SR_2016} (see Fig.~\ref{fig:5}). In particular, we consider in Fig.~\ref{fig:5} cylindrical membrane proteins with constant $U$ and $U'$ [see Fig.~\ref{fig:5}(a)], for which the exact analytic solution of bilayer thickness deformations simply amounts to the zeroth-order terms in Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}) \cite{Huang_1986,Nielsen_Goulian_Andersen_1998,Wiggins_Phillips_2005}. Furthermore, we consider in Fig.~\ref{fig:5} crown-shaped membrane proteins with circular cross section, constant $U'$, and the periodically varying $U(\theta)$ in Eq.~(\ref{eqUBC}) [see Fig.~\ref{fig:5}(b)], for which the exact analytic solution is obtained at order $N=w$ in Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}) \cite{Haselwandter_Phillips_PLOS_2013,Haselwandter_Phillips_EPL_2013,Kahraman_Koch_Klug_Haselwandter_PRE_2016}. We quantified the level of agreement between the BVM and FEM solutions and the corresponding exact analytic solutions through the percentage difference in the calculated bilayer thickness deformation energy $G$ in Eq.~(\ref{eqGArea}),
\begin{equation} \label{eqetaGdef}
\eta_{G} = 100 \times \Biggr{|}\frac{G - G_\mathrm{analy} }{G_\mathrm{analy}}\Biggr{|}\,,
\end{equation}
where $G_\mathrm{analy}$ denotes the analytic solution \cite{Huang_1986,Nielsen_Goulian_Andersen_1998,Wiggins_Phillips_2005,Haselwandter_Phillips_PLOS_2013,Haselwandter_Phillips_EPL_2013,Kahraman_Koch_Klug_Haselwandter_PRE_2016} and $G$ denotes the corresponding BVM or FEM solutions. We found, as expected, excellent numerical agreement between the BVM and the aforementioned exact analytic solutions for $N=0$ [Fig.~\ref{fig:5}(a)] or $N \geq w$ [Fig.~\ref{fig:5}(b)] within floating point error. The FEM solutions in Fig.~\ref{fig:5} are, up to their expected numerical precision \cite{Kahraman_Koch_Klug_Haselwandter_PRE_2016}, in good agreement with the exact analytic and BVM solutions, with the agreement improving with decreasing average edge size $\langle L \rangle$ in the FEM grid. For both cylindrical and crown-shaped membrane proteins, we have $\eta_{G}\approx 0.01 \%$ for $\langle L\rangle \approx 0.1$~nm in the FEM solutions in Fig.~\ref{fig:5}.
\subsection{Nonuniform boundary point distributions}
\label{subsubsecAPD}
As illustrated in Figs.~\ref{fig:4} and~\ref{fig:5}, the BVM can provide a highly accurate method for calculating protein-induced bilayer thickness deformations. However, Fig.~\ref{fig:4}(a) also shows that, for large enough deviations from a circular protein cross section, accurate BVM solutions require a large number of terms in the Fourier-Bessel series in Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}). The numerical performance of the BVM can be improved substantially, for non-circular protein cross sections, by choosing suitable, nonuniform boundary point distributions. In particular, we found that boundary point distributions that assign more points to, as viewed from the lipid bilayer, concave boundary regions yield a more rapid convergence of $G$ with increasing $N$. This can be understood intuitively by noting that, in the concave regions of a boundary curve, different sections of the boundary curve can produce overlapping bilayer thickness deformation fields, inducing protein self-interactions. One expects that higher-order terms in the Fourier-Bessel series in Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}) are required to capture such interactions \cite{Haselwandter_Phillips_EPL_2013,Kahraman_Koch_Klug_Haselwandter_PRE_2016}.
\begin{figure}[t!]{\includegraphics{AdaptivePointCrossSections.pdf}}
\caption{Illustration of the APD method used to increase the numerical efficiency of BVM solutions for (a) a three-fold clover-leaf protein shape ($s=3$) and (b) a four-fold clover-leaf protein shape ($s=4$) in Eq.~(\ref{eqCdefclover}). The boundary points used for the BVM solutions are indicated by blue dots. In panel (a) we set $\epsilon=0.38$, $N=31$, and $\Omega=0.62$ for the gap length $\bar \ell$ in Eq.~(\ref{eqelldef}). In panel (b) we set $\epsilon=0.30$, $N=42$, and $\Omega=0.72$. For both panels we set $\bar R \lambda\approx 2.3$~nm. To achieve an approximately periodic distribution of boundary points for even $s$, we duplicated in panel (b) the boundary point at the right-most apex, and slightly offset the resulting two boundary points along the bilayer-protein interface (see main text). The values of $N$ (numbers of boundary points) considered here were chosen for illustrative purposes. We generally employ values of $N$ greater than those considered here so as to meet the numerical precision criteria imposed here (see main text).}
\label{fig:6}
\end{figure}
To assign more boundary points to the concave boundary regions of clover-leaf and (finite-$P$) polygonal protein shapes, we employ an APD of the BVM boundary points. In the APD method, we distribute the $2N+1$ boundary points such that boundary points are always assigned to the apex points along the bilayer-protein boundary curves furthest away from the protein center (see Fig.~\ref{fig:6}). We distribute the remaining boundary points along the sections of the bilayer-protein boundary curves that are an arc length $\bar l\geq \bar \ell$ away from the apex points, such that these points are uniformly spaced in arc length, with the gap length
\begin{equation} \label{eqelldef}
\bar \ell = \Omega \frac{\bar \Gamma}{2s} \,,
\end{equation}
where the gap factor $\Omega$ satisfies $0 < \Omega < 1$, $\bar \Gamma$ is the (dimensionless) protein circumference, and $s$ is the symmetry of the clover-leaf or polygonal protein shape [see also Eqs.~(\ref{eqCdefclover}) and~(\ref{eqDefCpoly})] (Fig.~\ref{fig:6}). For even values of $s$, we consider in our APD method the general solution in Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}) for $N = s M/2$ with integers $M \geq 3$. For odd values of $s$, we allow in Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}) for $N = s M/2$ for even integers $M \geq 3$, and for $N = \left(s M-1\right)/2$ for odd integers $M \geq 3$. To achieve an approximately periodic distribution of boundary points for even $s$ and for odd $s$ with even $M$, we found it convenient to duplicate one of the apex boundary points, with a slight offset in the duplicated boundary points by an equal arc length distance from the apex [see Fig.~\ref{fig:6}(b)]. For greater numerical accuracy, this distance from the apex could be optimized so as to reduce the boundary error $\eta_{b'}$ in Eq.~(\ref{eqetabdef}), but we found it sufficient here to set it equal to one-half the arc length spacing between the boundary points in the concave boundary regions. Unless stated otherwise, we used the APD method for all BVM calculations described in this article, fixing $\Omega$ in Eq.~(\ref{eqelldef}) and $N$ in Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}) such that the boundary error $\eta_{b'} \leq 0.1 \%$ in Eq.~(\ref{eqetabdef}) and we obtained changes in $G$ and $\eta_{b'}$ of no more than $10^{-5}\%$ as the numerical precision was increased.
As illustrated in Fig.~\ref{fig:4}(b) for clover-leaf shapes, the APD method employed here improves considerably the convergence of the BVM with increasing $N$, particularly for proteins that show substantial deviations from a circular cross section. As a result, a given numerical accuracy of BVM solutions can be achieved with smaller $N$. We note that, for proteins with (discrete) rotational symmetry, the Fourier-Bessel series in Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}) must show the same symmetry. Indeed, in our BVM calculations we find that, within the numerical precision employed here, the coefficients of terms in Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}) that break the protein symmetry take values equal to zero. On this basis one could, for a given protein symmetry, further improve the numerical efficiency of the BVM by using the protein symmetry to remove some of the terms in the Fourier-Bessel series in Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}). For the scenarios considered here the BVM was efficient enough so as not to require such further refinement.
\begin{figure}[t!]{\includegraphics{GvsEpsilon.pdf}}
\caption{Comparing BVM and FEM solutions for the elastic energy of protein-induced bilayer thickness deformations. (a) Bilayer thickness deformation energy, $\bar G$ in Eq.~(\ref{eqGAreaDL}), obtained using BVM and FEM solutions for $\bar u$ in Eq.~(\ref{equdef}) and (b) corresponding percentage difference between the BVM and FEM solutions for $\bar G$, $\mu_{G}'$ in Eq.~(\ref{eqmuGdef}), for the clover-leaf protein shapes in Eq.~(\ref{eqCdefclover}) as a function of $\epsilon$ with the indicated values of $s$, $\bar R\lambda\approx2.3$~nm, $\bar U \lambda=0.3$~nm, and $\bar U'=0$. For the FEM solutions we employed an average edge size $\langle L \rangle \approx 0.1$ nm.}
\label{fig:7}
\end{figure}
Figure~\ref{fig:7} illustrates the calculation of the bilayer thickness deformation energy, $G$ in Eq.~(\ref{eqGArea}), using the BVM with APD for the clover-leaf protein shapes in Eq.~(\ref{eqCdefclover}) with various protein symmetries $s$ and deviations from a circular cross section, $\epsilon$. As expected \cite{Haselwandter_Phillips_PLOS_2013,Kahraman_Klug_Haselwandter_2014}, we find in Fig.~\ref{fig:7}(a) that $G$ increases with increasing $s$ and $\epsilon$. We also show in Fig.~\ref{fig:7} the corresponding results obtained from the FEM with an average edge size $\langle L \rangle\approx 0.1$~nm. In Fig.~\ref{fig:7}(b) we quantify the agreement between our BVM and FEM results through the percentage difference in $G$,
\begin{equation} \label{eqmuGdef}
\mu_{G}' = 100 \times \Biggr{|}\frac{G_\mathrm{BVM} - G_\mathrm{FEM}}{G_\mathrm{FEM} }\Biggr{|}\,,
\end{equation}
where $G_\mathrm{BVM}$ and $G_\mathrm{FEM}$ correspond to the values of $G$ in Eq.~(\ref{eqGArea}) obtained through the BVM and the FEM \cite{Kahraman_Klug_Haselwandter_2014,Kahraman_Koch_Klug_Haselwandter_PRE_2016,Kahraman_Koch_Klug_Haselwandter_SR_2016}, respectively. We find in Fig.~\ref{fig:7} that the BVM and FEM solutions yield excellent agreement for the energy of protein-induced bilayer thickness deformations for non-circular as well as circular protein cross sections, with the level of agreement between BVM and FEM solutions being in line with the accuracy of the FEM solutions expected from Fig.~\ref{fig:5}.
\section{Analytic approximation of the bilayer thickness deformation energy}
\label{secAnalytic}
For membrane inclusions with circular cross section, the solution for the thickness deformation field $\bar u(\bar r,\theta)$ in Eq.~(\ref{equEigenSuper}) with Eq.~(\ref{eqEigenFuncs}) and the bilayer thickness deformation energy in Eq.~(\ref{eqGLine}) yield exact analytic expressions for the energy of protein-induced bilayer thickness deformations for arbitrary (angular) variations in the bilayer-protein boundary conditions \cite{Huang_1986,Dan_Pincus_Safran_1993,Dan_Berman_Pincus_Safran_1994, Haselwandter_Phillips_PLOS_2013,Haselwandter_Phillips_EPL_2013,Kahraman_Klug_Haselwandter_2014,Kahraman_Koch_Klug_Haselwandter_PRE_2016}. The purpose of this section is to develop, on this basis, a simple analytic scheme for estimating the energy of protein-induced bilayer thickness deformations for membrane proteins with non-circular cross sections. In Sec.~\ref{secDependShape} we show that, for many protein shapes, these simple analytic estimates agree remarkably well with the corresponding BVM solutions.
For a single membrane inclusion with circular cross section and arbitrary (angular) variations in $U(\theta)$ and $U'(\theta)$, the exact solution of the Euler-Lagrange equation in Eq.~(\ref{eqHelmholtz}) is given by Eq.~(\ref{equEigenSuper}) with Eq.~(\ref{eqEigenFuncs}), and the corresponding bilayer thickness deformation energy follows from Eq.~(\ref{eqGLine}) \cite{Haselwandter_Phillips_PLOS_2013,Haselwandter_Phillips_EPL_2013,Kahraman_Klug_Haselwandter_2014,Kahraman_Koch_Klug_Haselwandter_PRE_2016}. For the choices for $U(\theta)$ and $U'(\theta)$ in Eqs.~(\ref{eqUBC}) and~(\ref{eqUpBC}), one thus finds the bilayer thickness deformation energy
\begin{equation}\label{eqGanalytic}
\begin{split}
\bar G_\mathrm{analy} = \pi \bar R_\mathrm{analy} &(\bar \nu_+ - \bar \nu_-) \bigg{[} \bar U^2_0 \bar E_0 + \bar U'^2_0 \bar F_0 + \bar U_0 \bar U'_0 \bar H_0 \\
+ \frac{1}{2} \bigg( \bar \beta^2 & \bar E_w + \bar \gamma^2 \bar F_v + \delta_{wv} \bar \beta \, \bar \gamma \bar H_w \bigg) \bigg{]}\bigg{\rvert}_{\bar r = \bar R_{\mathrm{analy}}}
\end{split}
\end{equation}
with $v > 0$ and $w > 0$, where $\bar R_\mathrm{analy}$ is the radius of the circular protein cross section, $\delta_{wv}$ is the Kronecker delta, and we have defined
\begin{equation}\label{eqDEFH}
\begin{split}
\bar D_q &= K_q(\sqrt{\bar \nu_+} \bar r) \partial_{\bar r} K_q(\sqrt{\bar \nu_-} \bar r) - K_q(\sqrt{\bar \nu_-} \bar r) \partial_{\bar r} K_q(\sqrt{\bar \nu_+} \bar r)\,,\\
\bar E_q &= \frac{\left[\partial_{\bar r} K_q(\sqrt{\bar \nu_+}\bar r) \right] \left[ \partial_{\bar r} K_q(\sqrt{\bar \nu_-} \bar r)\right]}{\bar D_q}\,,\\
\bar F_q &= \frac{K_q(\sqrt{\bar \nu_+} \bar r) K_q(\sqrt{\bar \nu_-} \bar r)}{\bar D_q}\,,\\
\bar H_q &= \frac{K_q(\sqrt{\bar \nu_+} \bar r) \partial_{\bar r} K_q(\sqrt{\bar \nu_-} \bar r) + K_q(\sqrt{\bar \nu_-} \bar r) \partial_{\bar r} K_q(\sqrt{\bar \nu_+} \bar r)}{\bar D_q}\,,
\end{split}
\end{equation}
where $q=0,1,\dots$, $K_q$ denotes the $q^\mathrm{th}$ order modified Bessel function of the second kind, and $\partial_{\bar r}$ denotes the partial derivative with respect to $\bar r$. The zeroth order terms in Eq.~(\ref{eqGanalytic}) are the contributions to $\bar G_\mathrm{analy}$ due to the constant $\bar U_0$ and $\bar U_0^\prime$ in Eqs.~(\ref{eqUBC}) and~(\ref{eqUpBC}), while the remaining terms encapsulate the effects of the variations in $U(\theta)$ and $U'(\theta)$ in Eqs.~(\ref{eqUBC}) and~(\ref{eqUpBC}) on $\bar G_\mathrm{analy}$. We use here Eq.~(\ref{eqGanalytic}) to analytically estimate the bilayer thickness deformation energy of membrane proteins with non-circular cross sections. To this end, we choose $\bar R_\mathrm{analy}$ in Eq.~(\ref{eqGanalytic}) such that the circumference of the circular membrane inclusion considered in Eq.~(\ref{eqGanalytic}) is equal to the circumference of the membrane protein under consideration,
\begin{equation} \label{Ranaly}
\bar R_\mathrm{analy} = \frac{\bar \Gamma}{2 \pi}\,,
\end{equation}
where, for the clover-leaf and polygonal boundary curves in Eqs.~(\ref{eqCdefclover}) and~(\ref{eqDefCpoly}) with Eq.~(\ref{eqDefCpoly2}), the protein circumference $\Gamma$ follows from $\bar \Gamma= \int^{2\pi}_{0} d\theta \bar l$, where, as in Eq.~(\ref{eqGLine}), $\bar l$ is the (dimensionless) line element.
\begin{figure}[t!]{\includegraphics{ColorMap.pdf}}
\caption{Color maps of the bilayer thickness deformation footprints due to clover-leaf protein shapes with (a) $\bar R=1$ and (b) $\bar R=10$ in Eq.~(\ref{eqCdefclover}) for $s=5$, $\epsilon=0.2$, $\bar U \lambda=0.3$~nm in Eq.~(\ref{eqUBC}), and $\bar U'=0$. Panels (c) and (d) show the mean curvature in units of $1/\lambda$, $\bar H = \lambda H$, associated with the thickness deformation fields in panels (a) and (b), respectively, while panels (e) and (f) show the corresponding mean curvature maps obtained for $\bar U \lambda=-0.3$~nm in Eq.~(\ref{eqUBC}) rather than $\bar U \lambda=0.3$~nm. We set $2\bar a \lambda=3.2$~nm for all panels. All results were obtained through the BVM.}
\label{fig:8}
\end{figure}
\begin{figure}[t!]{\includegraphics{LambdaVStheta.pdf}}
\caption{Line tension along the bilayer-protein boundary, $\bar \Lambda$ in Eq.~(\ref{eqdefLambda}), as a function of $\theta$ for (a) the protein shape in Fig.~\ref{fig:8}(a) and (b) the protein shape in Fig.~\ref{fig:8}(b), calculated using the same parameter values as in Fig.~\ref{fig:8}. The red dashed lines show the average of $\bar \Lambda(\theta)$ in Eq.~(\ref{eqdefLambda}) over the interval $0 \leq \theta \leq 2 \pi/5$, $\langle \bar \Lambda \rangle$. The yellow dashed lines show $\bar \Lambda_\mathrm{analy}=\bar G_\mathrm{analy}/\bar \Gamma$, where $\bar G_\mathrm{analy}$ is given by Eq.~(\ref{eqGanalytic}) and $\bar \Gamma$ is the protein circumference in Eq.~(\ref{Ranaly}).}
\label{fig:9}
\end{figure}
The analytic estimate of the thickness deformation energy in Eq.~(\ref{eqGanalytic}) captures, for the choice of $\bar R_\mathrm{analy}$ in Eq.~(\ref{Ranaly}), effects related to the overall shape of membrane proteins. However, Eq.~(\ref{eqGanalytic}) does not capture effects due to strong local variations in the protein cross section. For instance, the clover-leaf shapes in Eq.~(\ref{eqCdefclover}) can give, for large enough $\epsilon$ and $s$, protein cross sections with pronounced invaginations. If the protein size $R$ is comparable to the decay length of bilayer thickness deformations, $\lambda$ in Eq.~(\ref{eqDeflambda}), such protein invaginations can yield overlaps in the protein-induced lipid bilayer thickness deformations due to different portions of the bilayer-protein interface, resulting in protein self-interactions [see Fig.~\ref{fig:8}(a)]. As $\bar R$ is increased, these overlaps in protein-induced bilayer thickness deformations become less pronounced [see Fig.~\ref{fig:8}(b)]. Depending on the value of $\bar R$, one thus obtains distinct distributions of the mean curvature of $\bar u$ about the protein [see Figs.~\ref{fig:8}(c,d)], which also depend on the value and sign of $\bar U$ [see Figs.~\ref{fig:8}(e,f)].
To quantify the protein self-interactions suggested by Fig.~\ref{fig:8} it is useful to define, based on Eq.~(\ref{eqGLine}), the line tension along the bilayer-protein interface,
\begin{equation} \label{eqdefLambda}
\bar \Lambda \equiv \left[\bar U'(\theta) \bar \nabla^{2} \bar u - \bar U(\theta) \mathbf{\hat{n}} \cdot \bar \nabla^{3} \bar u \right] \big |_{\bar r=\bar C(\theta)}\,,
\end{equation}
where we used Eqs.~(\ref{eqDefU}) and~(\ref{eqDefUp}). In Figs.~\ref{fig:9}(a) and~\ref{fig:9}(b) we compare, for the protein shapes in Figs.~\ref{fig:8}(a) and~\ref{fig:8}(b) with constant $\bar U > 0$ and $\bar U^\prime = 0$, the line tensions $\bar \Lambda$ in Eq.~(\ref{eqdefLambda}) and their average values $\langle\bar\Lambda\rangle$ to the corresponding (constant) $\bar \Lambda$ associated with $\bar G_\mathrm{analy}$ in Eq.~(\ref{eqGanalytic}), which we denote by $\bar \Lambda_{\mathrm{analy}}$. As expected, Fig.~\ref{fig:9} shows that the variations in $\bar \Lambda$ are more pronounced for smaller clover-leaf protein shapes. We also find in Fig.~\ref{fig:9} that $\langle\bar\Lambda\rangle < \bar \Lambda_{\mathrm{analy}}$, with a larger $\big|\langle\bar\Lambda\rangle - \bar \Lambda_{\mathrm{analy}}\big|$ for smaller $\bar R$ in Fig.~\ref{fig:9}.
Interestingly, we can have $\bar \Lambda < 0$ in Fig.~\ref{fig:9}(a) for the smaller clover-leaf protein shape in Fig.~\ref{fig:8}(a), while $\bar \Lambda > 0$ in Fig.~\ref{fig:9}(b) for the larger clover-leaf protein shape in Fig.~\ref{fig:8}(b). The regime with $\bar \Lambda < 0$ in Fig.~\ref{fig:9}(a) can be understood by noting that, with a constant $\bar U > 0$ and $\bar U^\prime = 0$, $\bar \Lambda$ in Eq.~(\ref{eqdefLambda}) is directly proportional to the change in the mean curvature of $\bar u$ at the protein boundary, in the direction perpendicular to the protein-bilayer boundary and into the bilayer $(-\mathbf{\hat{n}})$. For the points along the clover-leaf boundary closest and furthest away from the protein center, $\mathbf{\hat{n}}$ is anti-parallel with the radial direction $\mathbf{\hat{r}}$. For the points along the clover-leaf boundary shape in Figs.~\ref{fig:8}(a) and~\ref{fig:8}(b) furthest away from the protein center the mean curvature is negative with the sign convention used here and decreases in magnitude as one radially moves away from the protein boundary, yielding $\bar \Lambda > 0$ [Figs.~\ref{fig:8}(c,d)]. In contrast, for the points along the clover-leaf boundary shape closest to the protein center in Fig.~\ref{fig:8}(a) [but not Fig.~\ref{fig:8}(b)], the mean curvature is approximately zero at the protein boundary and decreases as one radially moves away from the protein boundary, yielding $\bar \Lambda < 0$. With a different sign convention for the mean curvature or a protein with a constant $\bar U<0$ rather than $\bar U>0$, analogous considerations apply [Figs.~\ref{fig:8}(e,f)]. Thus, protein self-interactions can effectively lower the energy cost of protein-induced lipid bilayer thickness deformations, in analogy to the energetically favorable bilayer-thickness-mediated protein interactions found for identical membrane proteins in close enough proximity \cite{Dan_Pincus_Safran_1993,Dan_Berman_Pincus_Safran_1994,Ursell2007,harroun99,goforth03,botelho06,Phillips_Ursell_Wiggins_Sens_2009,grage11,Haselwandter_Phillips_EPL_2013,Kahraman_Klug_Haselwandter_2014,Haselwandter_Wingreen_2014,milovanovic15,Kahraman_Koch_Klug_Haselwandter_PRE_2016,Kahraman_Koch_Klug_Haselwandter_SR_2016,pollard18}.
\section{Dependence of bilayer thickness deformation energy on protein shape}
\label{secDependShape}
In this section we survey the dependence of the bilayer thickness deformation energy in Eq.~(\ref{eqGArea}) on the shape of membrane proteins. In particular, we allow for three distinct, not mutually exclusive, modes for breaking rotational symmetry about the protein center (see also Sec.~\ref{subsecProteinShape}). In Sec.~\ref{secShapeConst} we take the bilayer-protein boundary conditions to be constant along the protein circumference, but allow for protein cross sections that break rotational symmetry about the protein center. In Sec.~\ref{secShapeU} we explore the effect of variations in the protein hydrophobic thickness on protein-induced bilayer thickness deformations. Finally, in Sec.~\ref{secShapeUp} we study protein-induced bilayer thickness deformations for proteins that show variations in the bilayer-protein contact slope along the bilayer-protein boundary. To test the analytic approximation of the bilayer thickness deformation energy described in Sec.~\ref{secAnalytic} we compare, for all three scenarios considered in Secs.~\ref{secShapeConst}--\ref{secShapeUp}, our BVM results to the corresponding analytic estimates by computing the signed percent error
\begin{equation} \label{eqxiG}
\xi_{G} = 100\times\frac{\bar G_{\mathrm{analy}} - \bar G}{\bar G}\,,
\end{equation}
where $\bar G$ is the thickness deformation energy in Eq.~(\ref{eqGAreaDL}) obtained through the BVM and the corresponding analytic estimate $\bar G_{\mathrm{analy}}$ is given by Eq.~(\ref{eqGanalytic}) with Eq.~(\ref{eqDEFH}).
\subsection{Constant bilayer-protein boundary conditions}
\label{secShapeConst}
\begin{figure}[t!]{\includegraphics{G_xi_vsR.pdf}}
\caption{Bilayer thickness deformation energy $\bar G$ in Eq.~(\ref{eqGAreaDL}) calculated using the BVM (see Sec.~\ref{secBVM}) as a function of protein size $\bar R$ for clover-leaf protein shapes with (a) $s=3$, (b) $s=4$, and (c) $s=5$ in Eq.~(\ref{eqCdefclover}) with the indicated values of $\epsilon$, and (d) polygonal protein shapes with the indicated values of $s$ and $P=5$ in Eq.~(\ref{eqDefCpoly}). For all panels we set $\bar U \lambda=0.3$~nm and $\bar U'=0$. The insets show the signed percent error $\xi_{G}$ in Eq.~(\ref{eqxiG}) for the corresponding analytic approximations $\bar G_{\text{analy}}$ in Eq.~(\ref{eqGanalytic}).
}
\label{fig:10}
\end{figure}
\begin{figure}[t!]{\includegraphics{G_xi_vsm.pdf}}
\caption{Bilayer thickness deformation energy $\bar G$ in Eq.~(\ref{eqGAreaDL}) calculated using the BVM (see Sec.~\ref{secBVM}) as a function of lipid chain length $m$ in Eq.~(\ref{eqDefm}) for (a) clover-leaf protein shapes with $\epsilon = 0.3$ and the indicated values of $s$, (b) clover-leaf protein shapes with $s=5$ and the indicated values of $\epsilon$, and (c) polygonal protein shapes with the indicated values of $s$ and $P=5$ in Eq.~(\ref{eqDefCpoly}). For all panels we set $\bar U' = 0$ and $\bar W \lambda=3.8$~nm in Eq.~(\ref{eqUWa}), and $\bar R \lambda\approx2.3$~nm. The insets show the signed percent error $\xi_{G}$ in Eq.~(\ref{eqxiG}) for the corresponding analytic approximations $\bar G_{\text{analy}}$ in Eq.~(\ref{eqGanalytic}). We always have $\left|\bar U \right| > 0$ for the $m$-discretization used here.
}
\label{fig:11}
\end{figure}
In Fig.~\ref{fig:10} we consider the energy of protein-induced bilayer thickness deformations for clover-leaf [see Figs.~\ref{fig:10}(a,b,c)] and polygonal [see Fig.~\ref{fig:10}(d)] protein shapes as a function of protein size $\bar R$ with a constant $\bar U \neq 0$ and $\bar U'=0$. Previous work on the lipid bilayer thickness deformations induced by proteins with circular cross section \cite{Wiggins_Phillips_2004,Wiggins_Phillips_2005} suggests that, for $\bar R \gg 1$, $\bar G$ increases approximately linearly with $\bar R$. We find in Fig.~\ref{fig:10} that we also approximately have $\bar G \propto \bar R$ for non-circular protein cross sections, with the (positive) constant of proportionality depending on the protein shape. The analytic estimates $\bar G_{\mathrm{analy}}$ obtained from Eq.~(\ref{eqGanalytic}) match $\bar G$ in Fig.~\ref{fig:10} within approximately $10\%$, with particularly small magnitudes of the signed percent error $\xi_{\bar G}$ for the polygonal protein shapes in Fig.~\ref{fig:10}(d). Note that for protein sizes $R$ comparable to the decay length $\lambda$ we generally have $\xi_{\bar G} > 0$ in Fig.~\ref{fig:10}, indicating that protein self-interactions tend to lower the energy cost of protein-induced bilayer thickness deformations in Fig.~\ref{fig:10}.
The energy cost of protein-induced bilayer thickness deformations depends crucially on the unperturbed lipid bilayer thickness, which can be varied by changing the lipid chain length $m$ in Eq.~(\ref{eqDefm}) \cite{Rawicz_Olbrich_McIntosh_Needham_Evans_2000,Perozo_Kloda_Cortes_Martinac_2002,Andersen_Koeppe_2007,Phillips_Ursell_Wiggins_Sens_2009}. In Fig.~\ref{fig:11} we plot $\bar G$ for clover-leaf [see Figs.~\ref{fig:11}(a,b)] and polygonal [see Fig.~\ref{fig:11}(c)] protein shapes as a function of the lipid chain length $m$ with $\bar U'=0$. We used a protein hydrophobic thickness $\bar W \lambda= 3.8$~nm, which matches the unperturbed lipid bilayer thickness for $m\approx16$. In Fig.~\ref{fig:11}(a) we consider clover-leaf protein shapes with different symmetries $s$ and the same value of $\epsilon$, while in Fig.~\ref{fig:11}(b) we consider clover-leaf protein shapes with different values of $\epsilon$ and the same symmetry $s$. Similarly as in Fig.~\ref{fig:10} we find that deviations from a circular protein cross section increase $\bar G$. Furthermore, similarly as in Fig.~\ref{fig:10}, the dependence of $\bar G$ on $m$ in Fig.~\ref{fig:11} is very well captured by the analytic approximation $\bar G_{\mathrm{analy}}$ in Eq.~(\ref{eqGanalytic}), suggesting that the increase in $\bar G$ for clover-leaf and polygonal protein cross sections compared to circular protein cross sections results primarily from the increase in the length of the bilayer-protein boundary $\bar \Gamma$ due to deviations from a circular protein cross section.
\begin{figure}[t!]{\includegraphics{uvsr_DU.pdf}}
\caption{Bilayer thickness deformation profile $\bar u$ due to a protein with a circular cross section as a function of the radial distance from the protein center, $\bar r=r/\lambda$, obtained from the exact analytic solution in Eq.~(\ref{equEigenSuper}) with Eq.~(\ref{eqEigenFuncs}) for the indicated values of $\bar U^\prime$. We set $\bar U \lambda=0.3$~nm and $\bar R \lambda=2.3$~nm.}
\label{fig:12}
\end{figure}
\begin{figure}[t!]{\includegraphics{G_xi_vsDU.pdf}}
\caption{Bilayer thickness deformation energy $\bar G$ in Eq.~(\ref{eqGAreaDL}) calculated using the BVM (see Sec.~\ref{secBVM}) as a function of the bilayer-protein contact slope $\bar{U'}$ for (a) clover-leaf protein shapes with $\epsilon = 0.3$ and the indicated values of $s$, (b) clover-leaf protein shapes with $s=5$ and the indicated values of $\epsilon$, and (c) polygonal protein shapes with the indicated values of $s$ and $P=5$, and cylindrical protein shapes with a circular cross section of radius $\bar R$. For all panels we set $\bar R \lambda=2.3$~nm and $\bar U \lambda=0.3$~nm. The insets show the signed percent error $\xi_{G}$ in Eq.~(\ref{eqxiG}) for the corresponding analytic approximations $\bar G_{\text{analy}}$ in Eq.~(\ref{eqGanalytic}).
}
\label{fig:13}
\end{figure}
In Figs.~\ref{fig:10} and~\ref{fig:11} we set, in line with previous work on gramicidin channels and MscL \cite{Huang_1986,Wiggins_Phillips_2004,Wiggins_Phillips_2005}, $U^\prime = 0$. However, the most suitable choice for the boundary conditions on the gradient of $u$ at the bilayer-protein interface has been a matter of debate \cite{Nielsen_Goulian_Andersen_1998,Nielsen_Andersen_2000,Argudo_Bethel_Marcoline_Wolgemuth_Grabe_2017}, and may depend on the particular membrane protein and lipid species under consideration. In particular, $U^\prime$ may differ from zero or vary along the bilayer-protein interface, or the gradient of $u$ at the bilayer-protein interface may satisfy natural boundary conditions with $U^\prime$ being adjusted so as to minimize the bilayer thickness deformation energy. As illustrated in Fig.~\ref{fig:12} for a membrane protein with circular protein cross section and constant $\bar U \lambda=0.3$~nm, the value of $U^\prime$ can have a substantial effect on the shape of protein-induced lipid bilayer thickness deformations. In particular, for $U^\prime \approx 0.3$ protein-induced bilayer thickness deformations are seen to decay rapidly in Fig.~\ref{fig:12}.
Plotting $\bar G$ as a function of $U^\prime$ (see Fig.~\ref{fig:13}), we find that $\bar G$ is minimal for $U^\prime \approx 0.28$ with, as suggested by $\bar G_{\mathrm{analy}}$ in Eq.~(\ref{eqGanalytic}), an approximately quadratic dependence of $\bar G$ on $U^\prime$. Allowing for non-circular protein cross sections we find that the optimal $U^\prime$ depends strongly, for large enough $\epsilon$, on the symmetry of clover-leaf protein shapes [see Figs.~\ref{fig:13}(a,b)], but only weakly on the symmetry of polygonal protein shapes [see Fig.~\ref{fig:13}(c)]. Note that, for clover-leaf protein shapes, the optimal $U^\prime$ tend to shift towards $U^\prime \approx 0$ compared to circular protein cross sections. This can be understood by noting that, for clover-leaf protein shapes, the reduction in the size of the membrane footprint brought about by $U \neq 0$ competes with contributions to the bilayer thickness deformation energy due to protein self-interactions. Conversely, polygonal protein shapes only show weak self-interactions, resulting in minor shifts in the optimal $U^\prime$ compared to proteins with circular cross section. Finally, we note that the analytic estimates $\bar G_{\mathrm{analy}}$ in Eq.~(\ref{eqGanalytic}) tend to become less accurate for large $U^\prime$, with up to $\left|\xi_G \right| \approx 60\%$ for the clover-leaf and polygonal shapes considered here [Fig.~\ref{fig:13}(insets)].
\subsection{Variations in protein hydrophobic thickness}
\label{secShapeU}
\begin{figure*}[t!]{\includegraphics{G_xi_vsw_wide.pdf}}
\caption{Bilayer thickness deformation energy $\bar G$ in Eq.~(\ref{eqGAreaDL}) calculated using the BVM (see Sec.~\ref{secBVM}) as a function of the periodicity in protein hydrophobic thickness, $w$ in Eq.~(\ref{eqUBC}), for (a) the clover-leaf protein shapes in Eq.~(\ref{eqCdefclover}) with $s=2$ and the indicated values of $\epsilon$, (b) the clover-leaf protein shapes in Eq.~(\ref{eqCdefclover}) with $s=3$ and the indicated values of $\epsilon$, and (c) the polygonal protein shapes in Eq.~(\ref{eqDefCpoly}) with the indicated values of $s$ and $P=5$. For all panels we set $\bar R \lambda=2.3$~nm, $\bar U_{0} \lambda=-0.1$~nm, $\bar \beta \lambda=0.5$~nm, and $\bar U'=0$. The red dashed lines indicate the asymptotic scaling $\sim w^{3}$. The insets show the signed percent error $\xi_{G}$ in Eq.~(\ref{eqxiG}) for the corresponding analytic approximations $\bar G_{\text{analy}}$ in Eq.~(\ref{eqGanalytic}). In panel (d) we show color maps of the protein-induced bilayer thickness deformations associated with $\epsilon=0.4$ in panel (a) at (i) $w=2$ and (ii) $w=3$, with $\epsilon=0.4$ in panel (b) at (iii) $w=3$ and (iv) $w=4$, and with $s=4$ in panel (c) at (v) $w=2$ and (vi) $w=3$.
}
\label{fig:14}
\end{figure*}
Membrane proteins are, in general, expected to show variations in their hydrophobic thickness along the bilayer-protein boundary \cite{Krepkiy2009,Sonntag2011}. For oligomeric membrane proteins, variations in protein hydrophobic thickness are expected to be periodic so as to reflect the protein symmetry. We employ here the sinusoidal variations of $U(\theta)$ in Eq.~(\ref{eqUBC}) as a generic model of variations in protein hydrophobic thickness, in which we denote the periodicity of $U(\theta)$ by $w$. We focus, for now, on zero bilayer-protein contact slopes, $U^\prime = 0$ in Eq.~(\ref{eqUpBC}), but return to the effects of angular variations in $U^\prime$ in Sec.~\ref{secShapeUp}.
Figure~\ref{fig:14} shows that variations in $U(\theta)$ can have a strong impact on the energy cost of protein-induced bilayer thickness deformations, for non-circular as well as circular protein cross sections. The analytic estimate $\bar G_{\mathrm{analy}}$ in Eq.~(\ref{eqGanalytic}) is seen to approximately capture $\bar G$ in Fig.~\ref{fig:14}, but tends to become less accurate as the protein cross section exhibits greater deviations from a circular shape, with up to $\left|\xi_G \right| \approx 50\%$ for the clover-leaf and polygonal protein shapes considered here [Fig.~\ref{fig:14}(insets)]. Note that, for large enough $w$, $\bar G$ in Fig.~\ref{fig:14} scales approximately as $w^3$ for all protein cross sections considered. This can be understood from $\bar G_{\mathrm{analy}}$ in Eq.~(\ref{eqGanalytic}) by noting that $\bar E_{w} \sim w^{3}$ at large $w$, and $\bar \gamma = 0$ if $U^\prime =0$.
While, broadly speaking, variations in protein hydrophobic thickness are seen to increase $\bar G$ in Fig.~\ref{fig:14} for all protein cross sections considered, the interplay of $U(\theta)$ and the shape of the protein cross section can yield comparatively favorable or unfavorable scenarios. For instance, depending on whether adjacent regions of the bilayer-protein boundaries in clover-leaf protein shapes yield bilayer thickness deformations of the same sign [see panels (i) and (iii) in Fig.~\ref{fig:14}(d)] or distinct signs [see panels (ii) and (iv) in Fig.~\ref{fig:14}(d)], protein self-interactions can decrease or increase the energy of protein-induced bilayer thickness deformations. For polygonal protein shapes, we find that scenarios for which the maxima or minima of $U(\theta)$ coincide with the corners of the polygonal shapes [see panel (v) in Fig.~\ref{fig:14}(d)] tend to be unfavorable from an energetic perspective, as compared to scenarios for which the extrema of $U(\theta)$ tend to occur along the polygonal faces [see panel (vi) in Fig.~\ref{fig:14}(d)]. However, compared to the clover-leaf protein shapes considered in Fig.~\ref{fig:14}, the bilayer thickness deformation energy associated with the polygonal protein shapes in Fig.~\ref{fig:14} depends only weakly on the interplay between $U(\theta)$ and the shape of the protein cross section.
\subsection{Variations in bilayer-protein contact slope}
\label{secShapeUp}
\begin{figure*}[t!]{\includegraphics{G_xi_vsv_wide.pdf}}
\caption{Bilayer thickness deformation energy $\bar G$ in Eq.~(\ref{eqGAreaDL}) calculated using the BVM (see Sec.~\ref{secBVM}) as a function of the periodicity in the bilayer-protein contact slope, $v$ in Eq.~(\ref{eqUpBC}), for (a) the clover-leaf protein shapes in Eq.~(\ref{eqCdefclover}) with $s=2$ and the indicated values of $\epsilon$, (b) the clover-leaf protein shapes in Eq.~(\ref{eqCdefclover}) with $s=3$ and the indicated values of $\epsilon$, and (c) the polygonal protein shapes in Eq.~(\ref{eqDefCpoly}) with the indicated values of $s$ and $P=5$. For all panels we set $\bar R \lambda=2.3$~nm, $\bar U \lambda =0.3$~nm, $\bar U'_{0}=0$, and $\bar \gamma =0.3$. The red dashed lines indicate the asymptotic scaling $\sim v$. The insets show the signed percent error $\xi_{G}$ in Eq.~(\ref{eqxiG}) for the corresponding analytic approximations $\bar G_{\text{analy}}$ in Eq.~(\ref{eqGanalytic}). In panel (d) we show color maps of the protein-induced bilayer thickness deformations associated with $\epsilon=0.4$ in panel (a) at (i) $v=2$ and (ii) $v=4$, with $\epsilon=0.4$ in panel (b) at (iii) $v=3$ and (iv) $v=4$, and with $s=4$ in panel (c) at (v) $v=2$ and (vi) $v=4$.
}
\label{fig:15}
\end{figure*}
Similarly as the variations in $U(\theta)$ considered in Sec.~\ref{secShapeU}, $U^\prime(\theta)$ in Eq.~(\ref{eqUpBC}) will generally vary along the bilayer-protein interface. Such variations could come about, for instance, through the protein structure or the binding of peptides to some sections of the bilayer-protein interface \cite{Phillips_Ursell_Wiggins_Sens_2009,suchyna04}. Alternatively, if the (normal) gradient of $\bar u$ obeys natural boundary conditions at the bilayer-protein interface, a non-circular protein cross section or variations in $U(\theta)$ may effectively induce variations in $U^\prime(\theta)$. We employ here the simple model of $U^\prime(\theta)$ in Eq.~(\ref{eqUpBC}) to explore the effect of variations in $U^\prime(\theta)$ on the energy cost of protein-induced lipid bilayer thickness deformations. For simplicity we thereby use a constant~$\bar U > 0$.
Figure~\ref{fig:15} illustrates the impact of variations in $U^\prime(\theta)$ on the energy cost of protein-induced lipid bilayer thickness deformations. Similarly as in Fig.~\ref{fig:14}, the analytic estimate $\bar G_{\mathrm{analy}}$ in Eq.~(\ref{eqGanalytic}) is seen to approximately capture $\bar G$ in Fig.~\ref{fig:15}, but tends to become less accurate with increasing deviation of the protein cross section from a circular shape, with up to $\left|\xi_G \right| \approx 40\%$ for the clover-leaf and polygonal shapes considered here [Fig.~\ref{fig:15}(insets)]. Note that, for large enough $v$, $\bar G$ in Fig.~\ref{fig:14} scales approximately linearly with $v$, independent of the protein cross section considered. This can be understood from $\bar G_{\mathrm{analy}}$ in Eq.~(\ref{eqGanalytic}) by noting that $\bar F_{v} \sim v$ at large $v$. In analogy to Fig.~\ref{fig:14} we find in Fig.~\ref{fig:15} that, broadly speaking, variations in $U^\prime(\theta)$ increase $\bar G$ for all protein cross sections considered. However, the interplay of $U^\prime(\theta)$ and the shape of the protein cross section can yield, similarly as in Fig.~\ref{fig:14}, comparatively favorable or unfavorable scenarios. In particular, for the clover-leaf protein shapes in Figs.~\ref{fig:15}(a) and~\ref{fig:15}(b), it tends to be energetically favorable for the minima of $\bar U'(\theta)$ to coincide with the minima of $\bar C(\theta)$, so as to make protein self-interactions more favorable, and the maxima of $\bar U'(\theta)$ to coincide with the maxima of $\bar C(\theta)$, so as to reduce the protein's membrane footprint. This configuration is achieved, for instance, when $v=s$ [see panels (i) and (iii) in Fig.~\ref{fig:15}(d)]. Conversely, it tends to be energetically unfavorable for the minima of $\bar U'(\theta)$ to coincide with the maxima of $\bar C(\theta)$, and vice versa, or for $\bar U'(\theta)$ and $\bar C(\theta)$ to be out of phase [see panels (ii) and (iv) in Fig.~\ref{fig:15}(d)]. For the polygonal protein shapes in Fig.~\ref{fig:15}(c), particularly favorable configurations tend to be achieved when the minima of $\bar U'(\theta)$ fall on the polygonal faces, rather than on the corners of the polygonal shapes [see panels (v) and (vi) in~Fig.~\ref{fig:15}(d)].
\section{Transitions in protein organization and shape}
\label{secStability}
Section~\ref{secDependShape} demonstrates that protein-induced lipid bilayer thickness deformations show a strong dependence on protein shape, and that changes in protein shape can bring about changes in the bilayer thickness deformation energy $>10$~$k_B T$ in magnitude. In the present section we suggest possible implications of these results for the biophysical properties of membrane proteins. In particular, Sec.~\ref{secStabilityOligomers} explores the energetic contribution of lipid bilayer thickness deformations to the self-assembly of protein monomers into protein oligomers, and how changes in bilayer-protein interactions could destabilize protein oligomers. In Sec.~\ref{secStabilityShape} we investigate the effect of lipid bilayer thickness deformations on transitions in protein conformational state that involve substantial changes in protein shape.
\subsection{Self-assembly of protein oligomers}
\label{secStabilityOligomers}
\begin{figure}{\includegraphics{oligomerization.pdf}}
\caption{Difference between the lipid bilayer thickness deformation energies associated with protein oligomers of symmetry $s$ and their corresponding $s$ monomers, $\Delta \bar G$, calculated using the BVM (see Sec.~\ref{secBVM}) as a function of (a) the lipid chain length $m$ in Eq.~(\ref{eqDefm}) and (b) the (constant) bilayer-oligomer contact slope $U^\prime$ in Eq.~(\ref{eqDefUp}) for a variety of clover-leaf (solid curves) and polygonal (dashed curves) shapes of the protein oligomers. We took the protein monomers to have circular cross sections with $U^\prime = 0$ and used the indicated values of $s$, with $\epsilon = 0.3$ for the clover-leaf oligomer shapes in Eq.~(\ref{eqCdefclover}) and $P=5$ for the polygonal oligomer shapes in Eq.~(\ref{eqDefCpoly}). We set $\bar R \lambda = 1$~nm for the monomer radii, and used identical cross-sectional areas of the oligomers and their corresponding monomers. We set $U^\prime = 0$ in panel (a), $2\bar a\lambda = 3.2$~nm in panel (b), and used $\bar W \lambda=3.8$~nm for the protein monomers and oligomers in all panels. The schematics in the insets illustrate transitions between monomers and oligomers for selected oligomeric shapes. The plots in the insets show the difference in the oligomerization energies obtained from the analytic approximation $\bar G_{\mathrm{analy}}$ in Eq.~(\ref{eqGanalytic}) and the BVM, $\Delta\bar G_{\xi}=\Delta \bar G_{\mathrm{analy}}-\Delta \bar G$, for each curve in the main~panels.}
\label{fig:16}
\end{figure}
Complex molecular architectures of membrane proteins often arise from self-assembly of small protein subunits (monomers) into protein oligomers \cite{vinothkumar2010structures,forrest2015structural}. The thermodynamic competition between different oligomeric states of membrane proteins depends crucially on how the energy per protein subunit changes with protein oligomeric state \cite{kahraman2016thermodynamic}. If the hydrophobic thickness of the protein oligomers or monomers differs from the unperturbed hydrophobic thickness of the surrounding lipid bilayer, one set of contributions to the oligomerization energy arises from protein-induced bilayer thickness deformations. We illustrate here how these contributions to the oligomerization energy can be calculated through the BVM for protein-induced bilayer thickness deformations. For simplicity, we thereby consider a protein oligomer of symmetry $s$ with a clover-leaf or polygonal cross section, and take the $s$ (identical) competing protein monomers to have circular cross sections with the same total area as the protein oligomer and no interactions between the monomers. Furthermore, we assume that the protein oligomers and monomers show constant values of $U$ and $U^\prime$ along the bilayer-protein interface, with identical $U$ for the protein oligomers and monomers and $U^\prime = 0$ for the protein monomers. These assumptions could easily be lifted to describe more complex scenarios.
Figure~\ref{fig:16} shows the difference between the (dimensionless) bilayer thickness deformation energies associated with protein oligomers and their corresponding monomers, $\Delta \bar G$, as a function of the lipid chain length $m$ [see Fig.~\ref{fig:16}(a)] and the bilayer-oligomer contact slope $U^\prime$ [see Fig.~\ref{fig:16}(b)] for a variety of shapes of the oligomer cross section. The insets in Fig.~\ref{fig:16} show the differences in the oligomerization energies obtained from the analytic approximation $\bar G_{\mathrm{analy}}$ in Eq.~(\ref{eqGanalytic}) and the BVM, $\Delta\bar G_{\xi}=\Delta \bar G_{\mathrm{analy}} - \Delta \bar G$. Equation~(\ref{eqGanalytic}) is seen to provide, for modest magnitudes of $U$ and $U^\prime$, good estimates of the oligomerization energy. We generally have $\Delta G < 0$ in Fig.~\ref{fig:16}(a), indicating that protein-induced lipid bilayer thickness deformations support oligomerization. This can be understood from $\bar G_{\mathrm{analy}}$ by noting that the protein oligomers in Fig.~\ref{fig:16}(a) have a smaller circumference than their corresponding monomers. Interestingly, Fig.~\ref{fig:16}(b) shows that $\Delta G$ can become positive for large enough magnitudes of $U^\prime$ for the protein oligomer, thus destabilizing the protein oligomer. Such a change in $U^\prime$ could be achieved, for instance, through a transition in the conformational state of the oligomer or the binding of peptides to the oligomer \cite{Phillips_Ursell_Wiggins_Sens_2009,suchyna04}. Figure~\ref{fig:16} thus suggests that protein-induced bilayer thickness deformations could assist both in the assembly and disassembly of protein oligomers, and contribute $>10$~$k_B T$ to the energy budget of oligomer assembly or disassembly.
\subsection{Transitions in protein conformational state}
\label{secStabilityShape}
\begin{figure*}[t!]{\includegraphics{shape_change_wide.pdf}}
\caption{Difference between the lipid bilayer thickness deformation energies associated with the final and initial protein shapes indicated in the insets, $\Delta \bar G$, calculated using the BVM (see Sec.~\ref{secBVM}) as a function of (a,b) the lipid chain length $m$ in Eq.~(\ref{eqDefm}) and (c,d) the (constant) bilayer-oligomer contact slope $U^\prime$ in Eq.~(\ref{eqDefUp}). The values of $\epsilon$ associated with each clover-leaf shape in Eq.~(\ref{eqCdefclover}) are indicated in the insets, while for the polygonal protein shapes we set $P=5$. We set $U^\prime = 0$ in panels (a,b) and $2\bar a\lambda=3.2$~nm in panels (c,d), and used $\bar W \lambda=3.8$~nm for all panels. The cross sections of all protein shapes considered here have area $\pi \bar R^{2}$ with $\bar R \lambda= 2.3$~nm. The plots in the insets show the differences in the protein transition energies obtained from the analytic approximation $\bar G_{\mathrm{analy}}$ in Eq.~(\ref{eqGanalytic}) and the BVM, $\Delta\bar G_{\xi}=\Delta \bar G_{\mathrm{analy}} - \Delta \bar G$, for each curve in the main panels.}
\label{fig:17}
\end{figure*}
To perform their biological functions, membrane proteins often have to transition between different conformational states. Such transitions in protein conformational state can be accompanied by changes in the cross-sectional shape of membrane proteins producing, in turn, changes in protein-induced lipid bilayer deformations. Membrane proteins can thus be regulated by lipid bilayer properties, such as the bilayer hydrophobic thickness \cite{Perozo_Kloda_Cortes_Martinac_2002,rusinova2021mechanisms,Phillips_Ursell_Wiggins_Sens_2009,Andersen_Koeppe_2007}. We illustrate here how the BVM can be used to calculate the contribution of lipid bilayer thickness deformations to the energy difference between two protein states with distinct cross-sectional shapes. For simplicity, we thereby take the two states of the membrane protein to show identical $U$ and $U^\prime$ that are constant along the bilayer-protein interface, and to have cross-sectional shapes with the same area. These assumptions could easily be lifted to provide detailed models of specific conformational transitions in membrane proteins, which may also involve more than just two protein states.
Figure~\ref{fig:17} shows the difference between the lipid bilayer thickness deformation energies associated with the final and initial protein shapes indicated in the insets, $\Delta \bar G$, as a function of the lipid chain length $m$ [see Figs.~\ref{fig:17}(a) and~\ref{fig:17}(b)] and the bilayer-protein contact slope $U^\prime$ [see Figs.~\ref{fig:17}(c) and~\ref{fig:17}(d)]. The insets in Fig.~\ref{fig:17} show the corresponding differences in the protein transition energies obtained from the analytic approximation $\bar G_{\mathrm{analy}}$ in Eq.~(\ref{eqGanalytic}) and the BVM, $\Delta\bar G_{\xi}=\Delta \bar G_{\mathrm{analy}} - \Delta \bar G$. Similarly as in Fig.~\ref{fig:16}, Eq.~(\ref{eqGanalytic}) is seen to provide, for modest magnitudes of $U$ and $U^\prime$, good estimates of $\Delta \bar G$ in Fig.~\ref{fig:17}. In Figs.~\ref{fig:17}(a) and~\ref{fig:17}(c) we consider idealized scenarios in which the initial protein shape shows a circular cross section, while the final state corresponds to a clover-leaf or polygonal shape. We find that bilayer thickness deformations generally inhibit such transitions in protein shape, $\Delta \bar G \geq 0$, which is easily understood from $\bar G_{\mathrm{analy}}$ in Eq.~(\ref{eqGanalytic}) by noting that these transitions in protein shape are accompanied by an increase in protein circumference. In Fig.~\ref{fig:17}(b) we study $\Delta \bar G$ for transitions between proteins with non-circular cross sections. We thereby arranged the initial and final protein states such that $\Delta \bar G \leq 0$. Similarly as in Figs.~\ref{fig:17}(a) and~\ref{fig:17}(c), the results in Fig.~\ref{fig:17}(b) can be understood by noting that the transitions in protein shape in Fig.~\ref{fig:17}(b) are accompanied by a decrease in protein circumference. Note, in particular, that the energetically favorable protein shapes in Fig.~\ref{fig:17}(b) tend to correspond to polygonal protein shapes or clover-leaf shapes with small $\epsilon$.
Finally, we consider in Fig.~\ref{fig:17}(d) scenarios where the sign of $\Delta \bar G$ does not necessarily follow from the relative protein circumferences of the initial and final protein shapes, and may not be captured by $G_{\mathrm{analy}}$ in Eq.~(\ref{eqGanalytic}) for all the values of $U$ and $U^\prime$ considered here. In particular, we chose the initial and final protein shapes associated with the dotted teal, green, and red curves in Fig.~\ref{fig:17}(d) so that their circumferences lie within $1\%$ of each other, and the remaining protein shapes so that the final protein shape has a circumference that is substantially smaller than that of the initial state, by at least $6\%$. The former sets of protein shapes yield a change of sign in $\Delta \bar G$ with $U^\prime$. Three of the latter sets of protein shapes, corresponding to the teal, purple, and dotted purple curves in Fig.~\ref{fig:17}(d), always yield $\Delta \bar G \leq 0$, which can again be understood from $\bar G_{\mathrm{analy}}$ in Eq.~(\ref{eqGanalytic}), while the fourth, corresponding to the pink curve in Fig.~\ref{fig:17}(d), can yield a change of sign in $\Delta \bar G$ with $U^\prime$. In analogy to Fig.~\ref{fig:16}(b), this suggests that, for suitable protein shapes, modification of $U^\prime$ in a given (stable) protein conformational state through, for instance, peptide binding \cite{Phillips_Ursell_Wiggins_Sens_2009,suchyna04} could trigger a change in the protein conformational state mediated by bilayer thickness deformations. We note, however, that for the protein shapes in Fig.~\ref{fig:17}(d) $\Delta G$ exceeds zero by not more than a few $k_B T$.
\section{Summary and conclusions}
\label{secConclusion}
Employing protein-induced lipid bilayer thickness deformations as a test case \cite{Andersen_Koeppe_2007,Phillips_Ursell_Wiggins_Sens_2009,Huang_1986,Dan_Pincus_Safran_1993,Dan_Berman_Pincus_Safran_1994,Wiggins_Phillips_2005,Ursell_Kondev_Reeves_Wiggins_Phillips_2008,mondal11,Kahraman_Koch_Klug_Haselwandter_PRE_2016,Argudo_Bethel_Marcoline_Wolgemuth_Grabe_2017,canham70,Helfrich_1973,evans74,zimmerberg06,weikl18,young22,fournier99,Rawicz_Olbrich_McIntosh_Needham_Evans_2000}, we have described here a BVM that permits the construction of non-perturbative analytic solutions of protein-induced lipid bilayer deformations for non-circular membrane protein cross sections. In addition to the membrane protein cross section, our BVM allows for a breaking of rotational symmetry about the protein center through angular variations in the protein hydrophobic thickness and the bilayer-protein contact slope along the bilayer-protein interface. Our BVM reproduces available analytic solutions for membrane proteins with circular cross section \cite{Huang_1986,Nielsen_Goulian_Andersen_1998,Wiggins_Phillips_2005,Haselwandter_Phillips_PLOS_2013,Haselwandter_Phillips_EPL_2013,Kahraman_Koch_Klug_Haselwandter_PRE_2016} and yields, for membrane proteins with non-circular cross section, excellent agreement with numerical, finite element solutions. Based on these BVM solutions, we formulated a simple analytic approximation of the lipid bilayer thickness deformation energy associated with general protein shapes [see Eq.~(\ref{eqGanalytic}) with Eq.~(\ref{eqDEFH})]. We find that, for modest deviations from rotational symmetry, this analytic approximation is in good agreement with BVM solutions. These results suggest that, to a first approximation, the effect of membrane protein shape on the energy of bilayer thickness deformations can be understood in terms of the circumference associated with non-circular protein cross sections.
Through our BVM and analytic approximation of the lipid bilayer thickness deformation energy, we surveyed the dependence of protein-induced lipid bilayer thickness deformations on protein shape. We find that protein shape tends to have a large effect on the energy of protein-induced lipid bilayer thickness deformations, typically shifting the bilayer deformation energy by more than 10~$k_B T$. A limitation of the BVM described here arises for protein shapes that show extreme deviations from circular symmetry, in which case BVM solutions tend to involve a large number of terms and, hence, become increasingly intractable. In such cases it may be advisable to modify the APD method for the distribution of boundary points employed here, so as to reduce the number of terms required in the lipid bilayer thickness deformation field in Eq.~(\ref{equEigenSuperN}) with Eq.~(\ref{eqEigenFuncsN}). While we have focused here on bilayer thickness deformations, an approach analogous to the BVM employed here could be used to construct non-perturbative analytic solutions for other modes of membrane deformation, such as bilayer midplane deformations \cite{canham70,Helfrich_1973,evans74,zimmerberg06,weikl18,young22}. Furthermore, it would be interesting to construct BVM solutions for membrane proteins embedded in bilayers with heterogeneous lipid composition \cite{Leibler1987,SensSafran2000,Schaffer2004,Shrestha_Kahraman_Haselwandter_2020,shrestha22}.
We find that protein self-interactions provide an important motif for the energy of protein-induced lipid bilayer thickness deformations. Such self-interactions arise for invaginations in the protein cross section, from overlaps in the bilayer deformations induced at different sections of the bilayer-protein interface. The basic phenomenology of membrane protein self-interactions can be understood by drawing analogies to bilayer-thickness-mediated interactions between proteins \cite{Dan_Pincus_Safran_1993,Dan_Berman_Pincus_Safran_1994,Ursell2007,harroun99,goforth03,botelho06,Phillips_Ursell_Wiggins_Sens_2009,grage11,Haselwandter_Phillips_EPL_2013,Kahraman_Klug_Haselwandter_2014,Haselwandter_Wingreen_2014,milovanovic15,Kahraman_Koch_Klug_Haselwandter_PRE_2016,Kahraman_Koch_Klug_Haselwandter_SR_2016,pollard18}. In particular, membrane protein self-interactions can effectively lower the energy cost of protein-induced lipid bilayer thickness deformations for proteins with constant bilayer-protein hydrophobic mismatch and zero bilayer-protein contact slope. For non-zero bilayer-protein contact slopes, or for variations in the bilayer-protein hydrophobic mismatch or in the bilayer-protein contact slope along the bilayer-protein interface, protein self-interactions can yield dramatic shifts in the bilayer thickness deformation energy. Thus, the interplay between the cross-sectional shape of membrane proteins, protein hydrophobic thickness, and bilayer-protein contact slope yields a rich energy landscape of protein-induced lipid bilayer thickness deformations. Interestingly, the hydrophobic thickness or bilayer-protein contact slope of membrane proteins may be modified in cells through, for instance, protein mutations, changes in lipid composition, or the binding of peptides at the bilayer-protein interface, while protein oligomerization and transitions in protein conformational state tend to modify the cross-sectional shape of membrane proteins. The results described here therefore suggest general physical mechanisms for how protein shape couples to the function, regulation, and organization of membrane proteins.
\acknowledgements
This work was supported by NSF Grants No.~MCB-2202087, No.~DMR-2051681, and No.~DMR-1554716, a USC Graduate School DIA Fellowship, and the USC Center for Advanced Research Computing. We thank O.~Kahraman for helpful discussions on the finite element method, M.~Olguin for useful discussions on code optimization, and F.~Pinaud for helpful discussions on the self-assembly of membrane protein complexes.
|
\section{Introduction}
The Internet of Things (IoT) is poised to transform our lives and unleash enormous economic benefit. However, inadequate data security and trust are seriously limiting its adoption. Blockchain, a distributed tamper-resistant database, has native resistance to data tamper by which it is practically impossible to modify the data retroactively once a transaction has been recorded. We believe such tamper-resistance property can be of significant value in ensuring trust in IoT systems.
\; We aim at demonstrating the tamper-resistant capability of Blockchain in securing trust in IoT systems. An IoT testbed was constructed in our lab, and Ethereum based Blockchain was built in the testbed. A number of tamper-resistance experiments have been carried out and analyzed to examine the process of block validation in Blockchain on both full node and light node. Our demonstrations reveal that Blockchain has a dedicated tamper-resistant capability, which can be applied to IoT to ensure trusted data collection and sharing.
With this features used within P2P network, precisely by design, a Blockchain-based database can therefore constitute a trust-free decentralized consensus system. Note that trust-free means conventional \engordnumber{3} party as an arbitral body is filled in by common cryptographical theorems. This promotes Blockchain to be a suitable role for data recording, storage and identity management, especially for those sensitive data [2].
The connection between Blockchain and IoT has no longer been futuristic [4-9]. With the research running, scientists have handed out Blockchain-based IoT.
There is no doubt that Blockchain and IoT are the two hot tags in the field of science and technology. IoT, including sensors, vehicles and other moving objects, basically contains any use of embedded electronic components with the outside world communication equipment, in particular, it makes use of the IP protocol,
as shown in Fig. 1.
\begin{figure}[htbp]
\centering
\includegraphics[height=2.8cm]{111.pdf}
\caption{Trusted digital Supply Chain}
\label{fgr:example}
\end{figure}
\section{Core Concept of Tamper-resistant}
\subsection{Key components}
We now explain the key components of tamper-resistant.
\; First of all, we point out the difference between the so-called 51\% attack and chain reorganization. Nodes will never import a block that fails validation. A 51\% attack consists of a 51\% of the miners forking off from a block in the past and creating a new chain that eventually beats the current canonical chain in total difficulty.
\; A reorganization is done only if a block being imported on a side fork leads to a higher total difficulty on that particular fork than the canonical fork. The blocks nonetheless still need to be valid. Note that triggering re-organization function is based on TotalDifficult in Proof-of-Work [2].
\; Synchronization happens all the time across each node starting from an identical genesis block in a chain. Each node runs a validation function to validate each incoming block. No block will be accepted unless passing the validation.
\subsection{Validation procedures}
Throughout the validation procedures, the follow conditions are evaluated:
\begin{compactitem}
\item if \emph{StateRoot} $\in$ local levelDB, throw errors;
\item if \emph{ParentBlock} $\notin$ local levelDB, throw errors;
\item if \emph{StateRoot}$|$$_{\emph{ParentBlock}}$ $\notin$ local levelDB, throw errors;
\item if \emph{Validate(Header)} where \{\emph{nonce, difficulty, mixDigest\ldots}\} $\in$ \emph{Header} not passed, throw errors;
\item if \emph{Validate(UncleHeader)} not passed, throw errors;
\item if \emph{Validate(GasUsed)} not passed $||$ \emph{Validate(bloom)} not passed, throw errors;
\item if \emph{TxHash} $\not=$ \emph{Hash(Txs)} $||$ \emph{ReceiptHash} $\not=$ \emph{Hash(Receipts)}, throw errors;
\item if \emph{StateRoot} $\not=$
\emph{StateRoot}$|$ $ \xLongleftarrow{Txs} \emph{CurrentStateRoot}$, throw errors.
\end{compactitem}
The last validation points out that since there is no transaction sent to those normal nodes, the state root after a state change will never equal to the state root of the header of the incoming block coming from an abnormal node.
A simple fraud of database does not account for any PoW computation. However, not only \emph{difficulty}, \emph{epochDataset} and \emph{mixhash}, but also the \emph{HeaderHash} is involved in calculating the targeting \emph{nonce}. It means arbitrary but sufficient amount of computation should be carried out on the tampered block. Even if sufficient amount of computation is satisfied, the record can be recovered by the canonical chain with the fastest speed effort, unless the hacker has control over 51\% computational power among the whole network.
\section{Tamper-resistance Demonstrations}
\subsection{System setup}
\subsubsection*{Hardware setup (shown in Fig.1)}
\begin{compactitem}
\item two workstations as mining nodes, shown in Fig.1(a);
\item three Raspberry Pi 3 B+, shown in Fig.1(b), attached with IoT sensors as end-point nodes, which are only allowed to look up and upload data without mining.
\begin{figure}[htb]
\centering
\subfloat[]{%
\includegraphics[scale=0.17]{testbed2.pdf}} \hfill
\subfloat[]{%
\includegraphics[width=.24\textwidth]{testbed1.pdf}}
\caption{Overview of testbed}
\label{fgr:example}
\end{figure}
\end{compactitem}
\vspace{1mm}
\subsubsection*{Software setup}
\begin{compactitem}
\item Ubuntu 14.04 Trusty and Mac OS on mining nodes;
\item Raspbian on Raspberry Pi;
\item Golang Ethereum 1.5.9 for Blockchain [3];
\item Golang on hacking, tampering and logging;
\item Python on data receiving and encapsulation;
\item Javascript on Blockchain processing via web3 API [3].
\end{compactitem}
\vspace{2mm}
\subsubsection*{System Implementation}
\
The Raspberry Pi IoT devices equipped with temperature sensors measure room temperatures in the lab every 30 minutes. An Ethereum based Blockchain is built in the testbed. The IoT measurement data are encapsulated and uploaded to a pre-built contract in the Blockchain.
\subsection{Hacking scenarios and analysis}
We demonstrate several distinct scenarios where blocks are tampered with fake solution to PoW. Scenario analysis and experimental results are presented.
\vspace{1mm}
\noindent\emph{1) \ Non-mining node hacked:}
\
When a non-mining node is hacked, and the total difficulty is smaller than that of a normal block, the canonical chain always chooses the block with larger total difficulty in the context of a valid block. Thus this tampered block will be seen as an uncle block, the canonical chain will be recovered back by those normal blocks from other normal nodes.
\
When the total difficulty is greater than that of a normal block, other normal nodes will not accept the incoming blocks since the tampered block fails to pass the PoW validation during the synchronization. Once a future block whose total difficulty transcends that of the tampered block is generated by one of the normal mining nodes, this tampered block will be seen as an uncle block, the canonical chain will be recovered back by those normal blocks from other normal nodes.
\vspace{1mm}
\noindent\emph{2) \ Mining node hacked:}
\
When a mining node hacked, and the total difficulty is smaller than that of a normal block, the results are the same as that of the non-mining node hacked case. Once this node starts mining, the uncle block will be broadcast at the same time for validation, contributing to errors throwing on other normal nodes shown as in Fig.2.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.28]{222.pdf}
\caption{Bad Block with tampered uncle blocks}
\label{fgr:example}
\end{figure}
\
When the total difficulty is greater than that of a normal block, it ends up being insufficient computational power. As a result, this tampered block will not be accepted by any other adversaries, and this mining node will be removed as a bad peer by other normal nodes, shown in Fig.3.
\begin{figure}[htbp]
\centering
\includegraphics[height=1cm]{333.pdf}
\caption{Insufficient computation leading to incorrect nonce}
\label{fgr:example}
\end{figure}
\noindent\emph{3) \ Tampered Block on light node:}
\
We now investigate the scenario that a light node is hacked, and the PoW verification passes and the height of this light node is greater than that of those mining nodes. In this scenario, there will be no suitable peers available for this light node and any transactions sent from local will be broadcast to null, until the height is transcended by canonical chain, shown in Fig.4.
\begin{figure}[htb]
\centering
\subfloat[No suitable peer available when fetching data]{%
\includegraphics[scale=0.29]{hack_light.pdf}}
\\
\subfloat[No suitable peer available when sending transactions]{%
\includegraphics[scale=0.3]{hack_morelightsendtx.pdf}}
\end{figure}
\section{A Practical Demonstration}
\; We now demonstrate the effectiveness of the tamper-resistant property of Blockchain in protecting the IoT data records. In our Blockchain secured IoT testbed, one Raspberry Pi device was hacked, its temperature record was changed from the original measure of 34$^\circ$C to -4$^\circ$C.
\par As soon as the tampering happened, the blockchain noticed this anomaly with a broken chain in the hacked node, which signals a tampering action.
\par Next, when the blockchain is synchronized, the abnormal block is automatically recovered back to the major one through the canonical chain. As a result the tampered record of -4$^\circ$C has been replaced by the original record of 34$^\circ$C. This is the chain reorganization process of the Blockchain. A log is generated to record which content had been changed unexpectedly, and this log is automatically uploaded onto Blockchain for future reference.
\par This demonstrated that the Blockchain can be applied to IoT to secure data records.
|
\section{Introduction}
Entanglement is widely recognized as one of the most important resources in quantum information processing. Maximally entangled states have been applied to various information-processing tasks, including quantum communication channels \cite{CHB}, quantum cryptography \cite{AE}, quantum teleportation \cite{GBC}, and so on. Entanglement detection is one of the open questions in quantum information theory. Many entanglement detection criteria have been proposed, such as the partial transposition criterion \cite{pt,pt2}. The computable cross-norm or realignment (CCNR) criterion \cite{pp,oo} is a strong criterion which is powerful enough to detect some bound entangled states.
Quantum process tomography is a technique employed to fully characterize unknown quantum channels \cite{ov,og,of,oh,GP,os,DW,JB,ob,oc,od,oe,ou,oa,qpt1,qpt2,qpt3,qpt4,qpt5,qpt6,qpt7,qpt8,qpt9,qpt10,qpt11,qpt12,qpt13}.
Complete information of a quantum channel $\$$ in Eq. (\ref{Kraus}) is to determine a set of operation elements $\{K_n\}$ for $\$$, such that for an arbitrary input state $\sigma$ one can obtain the corresponding output state $\$(\sigma)$.
The currently known methods for completely obtaining channel information fall into three categories: standard quantum-process tomography (SQPT) \cite{og,of}, ancilla-assisted quantum process tomography (AAQPT) \cite{oh,GP,os,DW,JB} and direct characterization of quantum dynamics (DCQD) \cite{ob,oc,od,oe}. All SQPT methods are based on multiple comparisons between the output state of the channel and the appropriate input state to analyze the process impact of the identification channel operate. The AAQPT methods usually establish a corresponding relationship between the channel information and the quantum state with auxiliary systems. This correspondence allows us to derive the complete information of the channel, see Fig. \ref{fig1}. The methods of DCQD \cite{ou} are different from SQPT and AAQPT, which do not need to perform any state tomography.
\begin{figure}
\begin{center}
\includegraphics[scale=0.45]{Fig1.eps}
\caption{The scheme of ancilla-assisted process tomography. In order to extract the complete information of quantum channel $\$$ in system $A$, an auxiliary system $B$ has been used to assist. If the input state $\varrho_{AB}$ is of faithfulness, the complete information of channel $\$$ can then be obtained from the output state $\varrho_{AB}^{out}$.} \label{fig1}
\end{center}
\end{figure}
The way of extracting channel information in Refs. \cite{GP,os} clearly belongs to AAQPT. They proposed a novel bipartite quantum states property named \textit{faithfulness} \cite{GP}, which indicated whether the complete information of any quantum channel can be obtained or not. The sufficient and necessary conditions of a bipartite state with faithfulness have been given in Refs. \cite{GP,os}. If the input state is of faithfulness, the corresponding output state carries complete information of the channel, and thus the input state is useful for AAQPT. In Ref. \cite{GP}, the authors showed that there exist separable states which can also be used for AAQPT. Interestingly, one may ask the related question, \textit{are all entangled states useful for AAQPT?} We will focus on this question.
In this paper, we find that not all entangled states are useful for AAQPT, there exist some entangled states which are useless. We first present that the realignment operator $\mathcal{R}$ in the CCNR criterion is related to the question whether a bipartite state is faithful or not. The relationship between the realignment operator and the faithfulness of a bipartite state is explained. Then we show the process of extracting the complete information of an unknown channel by the realignment operation. Based on this relationship, we present two examples of entangled states which are entangled but they cannot be used for AAQPT. Last but not least, experimental verification has also been performed on the IBM platform \textit{ibmq\_athens}.
\section{Entanglement is not sufficient for fully characterize quantum channels}
The output state $\$(\varrho)$, corresponding to an input state $\varrho$, of a linear, trace-preserving and completely positive map $\$$ \cite{oa} can be written in a so-called Kraus form \cite{om},
\begin{equation}\label{Kraus}
\varrho \rightarrow \$(\varrho)=\sum_n K_n \varrho K_n^\dag,
\end{equation}
where $\{K_n\}$ are operators on the Hilbert space $\mathcal{H}$ of the quantum system and satisfy the completeness condition $\sum_n K_n^\dag K_n=I$.
The Choi-Jamio{\l}kowski isomorphism states that every superoperator $\$$ on $\mathcal{H}$ has a one-to-one correspondence to a state on $\mathcal{H}\otimes \mathcal{H}$ given by \cite{TJ,MD,or}
\begin{eqnarray}
S_\$:=\$\otimes I (|\Phi^+\rangle\langle\Phi^+|),\label{S}
\end{eqnarray}
where $|\Phi^+\rangle=\sum_i|ii\rangle$ is the unnormalized maximally entangled state. The inverse relation is
\begin{eqnarray}
\$(\varrho)={\rm Tr}_B[(I\otimes \varrho^T)S_\$],
\end{eqnarray}
where ${\rm Tr}_B$ is the partial trace for the second $\mathcal{H}$, and $T$ is transposition.
From the Choi-Jamio{\l}kowski isomorphism, one can see that the state $S_\$$ in Eq. (\ref{S}) contains complete information of the quantum channel $\$$. One natural question is that: is maximally entangled state $|\Phi^+\rangle$ necessary? i.e., can we use another density matrix $\varrho_{_{AB}}$ on $\mathcal{H}\otimes \mathcal{H}$ such that $\$\otimes I (\varrho_{_{AB}})$ contains complete information of the quantum channel $\$$ as well? As shown in Fig. \ref{fig1} , Ref. \cite{GP} proved that $\$\otimes I (\varrho_{_{AB}})$ contains complete information of the quantum channel $\$$ if and only if all the singular values of $\check{R}(\varrho_{_{AB}})$ are nonzero (i.e. $\check{R}(\varrho_{_{AB}})^{-1}$ exists), where
\begin{eqnarray}
\check{R}(\varrho_{_{AB}}):=(\varrho_{_{AB}}^{T_B}E)^{T_A},\label{checkR}
\end{eqnarray}
with $E=\sum_{ij}|ij\rangle\langle ji|$ being the swap operator.
Interestingly, we find that the realignment operation $\mathcal{R}$ in the CCNR criterion is related to $\check{R}$. Thus, it has the similar consequence as $\check{R}$ in imprinting complete information of quantum channel on output states.
\textit{Theorem 1.}
The singular values of $\check{R}(\varrho_{_{AB}})$ are exactly the same as the ones of $\mathcal{R}(\varrho_{_{AB}})$. Thus, $\$\otimes I (\varrho_{_{AB}})$ contains complete information of the quantum channel $\$$, if and only if all the singular values of $\mathcal{R}(\varrho_{_{AB}})$ are nonzero (i.e. $\mathcal{R}(\varrho_{_{AB}}))^{-1}$ exists).
\textit{Proof.---}
Consider a general bipartite finite-dimensional state, $\varrho_{_{AB}}$ given by
\begin{eqnarray}
\varrho_{_{AB}}=\sum_{i,j,k,l}\varrho_{ij,kl}|i\rangle\langle j|\otimes|k\rangle\langle l|.
\end{eqnarray}
According to Eq. (\ref{checkR}), $\check{R}(\varrho_{_{AB}})$ can be expressed as
\begin{eqnarray}
\check{R}(\varrho_{_{AB}})=(\varrho_{_{AB}}^{T_B}E)^{T_A}=\sum_{i,j,k,l}\varrho_{ij,kl}|k\rangle\langle i|\otimes|l\rangle\langle j|.
\end{eqnarray}
For the realignment operation, we can obtain $\mathcal{R}(\varrho_{_{AB}})$ based on the definition of the realignment operation in Ref. \cite{pp} (here, we use a equivalent realignment method compared with the original one in Ref. \cite{pp}),
\begin{eqnarray}
\mathcal{R}(\varrho_{_{AB}})=\sum_{i,j,k,l}\varrho_{ij,kl}|i\rangle\langle k|\otimes|j\rangle\langle l|.
\end{eqnarray}
We can find the following relationship between $\check{R}(\varrho_{_{AB}})$ and $\mathcal{R}(\varrho_{_{AB}})$,
\begin{eqnarray}
[\check{R}(\varrho_{_{AB}})]^{T_{AB}}=\mathcal{R}(\varrho_{_{AB}}),
\end{eqnarray}
where $T_{AB}$ stand for a transpose of the A and B part of the matrix $\check{R}(\varrho_{_{AB}})$. Any matrix $\check{R}(\varrho_{_{AB}})$ can be expressed as the form of singular decomposition
\begin{eqnarray}
\check{R}(\varrho_{_{AB}})=U\Sigma V^{\dagger},
\end{eqnarray}
which $U$ and $V$ are unitary matrices, $\Sigma$ is a diagonal matrix with singular values of $\check{R}(\varrho_{_{AB}})$. Thus, $\mathcal{R}(\varrho_{_{AB}})$ can be expressed as
\begin{eqnarray}
\mathcal{R}(\varrho_{_{AB}})&=&[\check{R}(\varrho_{_{AB}})]^{T_{AB}}=[U\Sigma V^{\dagger}]^{T_{AB}}\nonumber\\
&=&V^\ast\Sigma^T U^T=V^\ast\Sigma U^T.
\end{eqnarray}
Therefore, $\check{R}(\varrho_{_{AB}})$ and $\mathcal{R}(\varrho_{_{AB}})$ have exactly the same singular values. $\$\otimes I (\varrho_{_{AB}})$ contains complete information of the quantum channel $\$$, if and only if all the singular values of $\mathcal{R}(\varrho_{_{AB}})$ are nonzero (i.e. $\mathcal{R}(\varrho_{_{AB}}))^{-1}$ exists). \hfill $\square$
Based on Theorem 1, we can easily prove the following theorem by using the realignment operation.
\textit{Theorem 2.} Not all entangled states are useful for ancilla-assisted quantum process tomography. There exist entangled states which cannot be used for extracting complete information of quantum channels.
\textit{Proof.---}
To prove this theorem, we will show that there exist entangled states $\varrho_e$ such that $\mathcal{R}(\varrho_e)$ have at least one singular value being zero.
For the first example, consider the following two-qutrit state $\sigma_E$,
\begin{eqnarray}
\sigma_E&=&p\frac{(|00\rangle+|11\rangle)(\langle00|+\langle11|)}{2} \nonumber \\
&&+(1-p)\frac{(|00\rangle+|22\rangle)(\langle00|+\langle22|)}{2},
\end{eqnarray}
which is entangled when $0\leq p\leq1$. One can easily obtain
\begin{equation}
\mathcal{R}(\sigma_E)=\mathrm{diag}\bigg\{ \frac{1}{2},\frac{p}{2},\frac{1-p}{2},\frac{p}{2},\frac{p}{2},0,\frac{1-p}{2},0,\frac{1-p}{2}\bigg\},
\end{equation}
then $\mathcal{R}(\sigma_E)$ has at least two zero singular values. Thus, the entangled state $\sigma_E$ cannot be used for extracting complete information of quantum channels on qutrit states, since for all values of $p\in[0,1]$ the value $\mathcal{R}(\sigma_E)^{-1}$ does not exist.
Another example of entangled state, we review a 3$\times$3 bound entangled state \cite{bound},
\begin{eqnarray}
\rho=\frac{1}{8a+1}
\left(
\begin{array}{ccccccccc}
a & 0 & 0 & 0 & a & 0 & 0 & 0 & a\\
0 & a & 0 & 0 & 0 & 0 & 0 & 0 & 0\\
0 & 0 & a & 0 & 0 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & a & 0 & 0 & 0 & 0 & 0\\
a & 0 & 0 & 0 & a & 0 & 0 & 0 & a\\
0 & 0 & 0 & 0 & 0 & a & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 0 & b & 0 & c\\
0 & 0 & 0 & 0 & 0 & 0 & 0 & a & 0\\
a & 0 & 0 & 0 & a & 0 & c & 0 & b
\end{array}
\right),
\end{eqnarray}
where $0<a<1$, $b=(1+a)/2$, and $c=\sqrt{1-a^2}/2$.
After the realignment operation,
one can find that the $\mathcal{R}(\rho)$ has one singular value of zero, thus $\mathcal{R}(\rho)^{-1}$ does not exist. According to the previous derivation, this state $\rho$ cannot obtain the complete information of the channel although the state is an entangled state.
Therefore, not all entangled states are useful for AAQPT. There exist entangled states which cannot be used for extracting complete information of quantum channels.
\hfill $\square$
In Ref. \cite{GP}, it has been shown that entanglement is not necessary for extracting complete information of quantum channels. Surprisingly, entanglement is not sufficient either as proved in Theorem 2. Thus, one can conclude that entanglement is neither necessary nor sufficient for extracting complete information of quantum channels.
\section{Extracting complete information of quantum channels from $\mathcal{R}(\varrho_{_{AB}})$}
In the above section, Theorem 1 only proved that one can obtain the complete information of quantum channels from $\mathcal{R}(\varrho_{_{AB}})$ if and only if $[\mathcal{R}(\varrho_{_{AB}})]^{-1}$ exists, but the method for extracting complete information of quantum channels from $\mathcal{R}(\varrho_{_{AB}})$ is not given.
In this section, we give a general method for extracting complete information of quantum channels from $\mathcal{R}(\varrho_{_{AB}})$.
The method is derived as follows.
For any input state $\sigma$, we have the corresponding output state $\$(\sigma)$,
\begin{eqnarray}
\$(\sigma)=\sum_n K_n \sigma K_n^\dag.\label{M16}
\end{eqnarray}
As introduced in Ref. \cite{GP}, we define $|\sigma\rangle$ as follows,
\begin{eqnarray}
|\sigma\rangle:=\sigma\otimes I \sum_i|ii\rangle= I\otimes \sigma^T\sum_i|ii\rangle.\label{rangle}
\end{eqnarray}
In the same way, we can also define $\langle\sigma|$ as follows,
\begin{eqnarray}
\langle\sigma|:=\sum_i\langle ii|\sigma\otimes I.
\end{eqnarray}
Now we can express the realignment operator $\mathcal{R}$ as \cite{zhang,zhang2}
\begin{eqnarray}
\mathcal{R}(\sigma_1\otimes\sigma_2)=|\sigma_1\rangle\langle\sigma_2^*|.
\end{eqnarray}
Thus, according to Eq. (\ref{rangle}), $|\$(\sigma)\rangle$ can be expressed as
\begin{eqnarray}
|\$(\sigma)\rangle
&=&\sum_n K_n \sigma K_n^\dag\otimes I \sum_i|ii\rangle\nonumber\\
&=&\sum_n K_n \sigma\otimes (K_n^\dag)^T\sum_i|ii\rangle\nonumber\\
&=&(\sum_n K_n\otimes K_n^\ast) (\sigma\otimes I)\sum_i|ii\rangle\nonumber\\
&=&\sum_n K_n\otimes K_n^\ast|\sigma\rangle,\label{M}
\end{eqnarray}
where the second equation holds since $K_n^\dag\otimes I \sum_i|ii\rangle=I \otimes (K_n^\dag)^T\sum_i|ii\rangle$.
For an arbitrary finite-dimensional bipartite state $\varrho_{_{AB}}$, we can expand it under local orthonormal operator bases \cite{os,zhang,LOO},
\begin{eqnarray}
\varrho_{_{AB}}=\sum_k\lambda_k G_k^A\otimes G_k^B,\label{LOO}
\end{eqnarray}
which $\{G_k^A\}$ and $\{G_k^B\}$ satisfy the following orthogonal conditions,
\begin{eqnarray}
{\rm Tr}(G_k^A G_{k^\prime}^A)={\rm Tr}(G_k^B G_{k^\prime}^B)=\delta_{k{k^\prime}}.
\end{eqnarray}
After the realignment operation, it can be expressed as
\begin{eqnarray}
\mathcal{R}(\varrho_{_{AB}})=\sum_k\lambda_k |G_k^A\rangle \langle (G_k^B)^*|.
\end{eqnarray}
Based on Eq. (\ref{LOO}), $\$\otimes I (\varrho_{_{AB}})$ can be represented as
\begin{eqnarray}
\$\otimes I (\varrho_{_{AB}})=\sum_k\lambda_k \$(G_k^A)\otimes G_k^B.
\end{eqnarray}
Therefore, $\mathcal{R}(\$\otimes I (\varrho_{_{AB}}))$ can be written as
\begin{eqnarray}
\mathcal{R}\big(\$\otimes I (\varrho_{_{AB}})\big)&=&
\sum_k\lambda_k |\$(G_k^A)\rangle\langle (G_k^B)^*|\nonumber\\
&=&\sum_k\lambda_k (\sum_n K_n\otimes K_n^\ast) |G_k^A\rangle\langle (G_k^B)^*|\nonumber\\
&=&M\mathcal{R}(\varrho_{_{AB}}),
\end{eqnarray}
where we have defined
\begin{equation}\label{}
M:=\sum_n K_n\otimes K_n^\ast,\label{Mdefine}
\end{equation}
and used \cite{GP}
\begin{equation}\label{}
|\$(G_k^A)\rangle=|\sum_n K_n G_k^A K_n^\dag\rangle=\sum_n K_n\otimes K_n^\ast |G_k^A\rangle,
\end{equation}
and $M$ has complete information of the quantum channel.
\begin{figure*}[hbtp]
\begin{center}
\includegraphics[scale=1]{Fig2.eps}
\caption{ The initial state of $q_0$ , $q_1$ and $q_2$ is $|0\rangle$. $H$ is the Hadamard Gate ($H$-Gate) and $I$ stands for the identity gate, which has no theoretical influence on the input state $\varrho_{q_0q_1}$, but practically, it will have some interference to cause some errors in the input state. The quantum channel $\$$ has been realized by an
auxiliary qubit $q_2$. By the CNOT gate $U^{cnot}_{q_2q_0}$, we can realize the channel (\ref{Pauli}) after tracing the qubit $q_2$.} \label{fig2}
\end{center}
\end{figure*}
Therefore, if and only if all singular values of $\mathcal{R}(\varrho_{_{AB}})$ are nonzero \big(i.e. $\mathcal{R}(\varrho_{_{AB}})^{-1}$ exists\big), one can obtain $M$ which contains complete information of the channel, and it can be expressed as
\begin{eqnarray}
M=\mathcal{R}\big(\$\otimes I (\varrho_{_{AB}})\big)\mathcal{R}(\varrho_{_{AB}})^{-1}.\label{M27}
\end{eqnarray}
When the arbitrary input state is $\sigma$ and the corresponding output state is $\$(\sigma)$, they have the following relationship based on Eq. (\ref{M}),
\begin{eqnarray}
|\$(\sigma)\rangle=M|\sigma\rangle=\mathcal{R}\big(\$\otimes I (\varrho_{_{AB}})\big)\mathcal{R}\big(\varrho_{_{AB}}\big)^{-1}|\sigma\rangle.\label{M28}
\end{eqnarray}
By solving $M$, we can get the matrix information of $\$(\sigma)$ after transformation.
Because of $|\sigma\rangle=\sigma\otimes I \sum_i|ii\rangle$, we can get $\sigma$ reversely,
\begin{eqnarray}
\sigma={\rm Tr}_B\big(|\sigma\rangle\sum_i\langle ii|\big).
\end{eqnarray}
Similarly, for $|\$(\sigma)\rangle=M|\sigma\rangle$, we can also get $\$(\sigma)$,
\begin{eqnarray}
\$(\sigma)={\rm Tr}_B\big(|\$(\sigma)\rangle\sum_i\langle ii|\big)={\rm Tr}_B\big(M|\sigma\rangle\sum_i\langle ii|\big).\label{output}
\end{eqnarray}
Thus, for an arbitrary input state $\sigma$, we first get $M$ and $|\sigma\rangle$, then based on Eq. (\ref{output}) we can obtain the output state
$\$(\sigma)$.
\section{quantum circuits simulation}
In this section, we will design a quantum circuit, and compare the experimental result of the channel information obtained by the realignment operator $\mathcal{R}$ with the theoretical result of this channel. We realize this quantum circuit on IBM quantum processor \textit{ibmq\_athens}. It is a five-qubit quantum system, we only use three qubits of them.
The quantum circuit is shown in Fig. \ref{fig2}. We denote $\varrho_{q_0q_1}$ and $\varrho _{q_0q_1}^{out}$ as input and output states, respectively. The input state $\varrho_{q_0q_1}$ is a two-qubit state composed of $q_0$ and $q_1$. The initial states of $q_0$ and $q_1$ are $|0\rangle$. The quantum channel $\$$ has been realized by an
auxiliary qubit $q_2$, which can be expressed as follows,
\begin{eqnarray}
\$(\sigma)=\sum_{n=1}^2 K_n \sigma K_n^\dag,\label{Pauli}
\end{eqnarray}
where $K_1=I/\sqrt{2}$ and $K_2=\sigma_x/\sqrt{2}$. Based on Eq. (\ref{Mdefine}), one can get the theoretical value of $M$ which contains complete information of the channel $\$$,
\begin{eqnarray}
M&=&\sum_{n=1}^2 K_n\otimes K_n^\ast
=\frac{I}{\sqrt{2}}\otimes \frac{I^\ast}{\sqrt{2}}+\frac{\sigma_x}{\sqrt{2}} \otimes \frac{\sigma^\ast_x}{\sqrt{2}}
\label{thM}
\end{eqnarray}
From Fig. \ref{fig2}, one can calculate the theoretical results of density matrices for the input state $\varrho_{q_0q_1}$ and the output state $\varrho _{q_0q_1}^{out}$. It is worth noticing that $|\psi_{in}\rangle
=U^{cnot}_{q_0q_1}|+\rangle\otimes|0\rangle
=\frac{1}{\sqrt{2}}(|00\rangle+|11\rangle)$ and
\begin{eqnarray}
\varrho_{q_0q_1}=|\psi_{in}\rangle\langle \psi_{in}|.\label{rhoin}
\end{eqnarray}
For the input state after the $H$ Gate and the CNOT Gate, we have some $I$ operations. Here, $I$ Gates have no theoretical influence on the input state $\varrho_{q_0q_1}$. But for practical quantum computation, it will have some interference to cause some errors in the input state. As mentioned in the previous introduction of the quantum circuit, we reduce the purity of the input state to verify whether $\mathcal{R}$ operation still has a good effect of channel information extraction when the purity of the input state is not high.
From Fig. \ref{fig2}, the output state $\varrho _{q_0q_1}^{out}$ can be calculated as
\begin{eqnarray}
\varrho _{q_0q_1}^{out}
&=&{\rm Tr}_{q_2}|\psi_{out}\rangle\langle\psi_{out}|\nonumber\\
&=&\frac{1}{2}|\psi_{in}\rangle\langle\psi_{in}|+\frac{1}{2}\sigma_x\otimes I|\psi_{in}\rangle\langle\psi_{in}|\sigma_x\otimes I
\end{eqnarray}
where $|\psi_{out}\rangle
=U^{cnot}_{q_2q_0}|\psi_{in}\rangle\otimes|+\rangle
=1/\sqrt{2}|\psi_{in}\rangle\otimes |0\rangle+1/\sqrt{2}\sigma_x\otimes I|\psi_{in}\rangle\otimes |1\rangle$.
Our quantum circuit has been executed on IBM quantum processor \textit{ibmq\_athens} for $10,240$ times, which were divided into 10 batches. For each batch, the density matrices of the experimental input and output states have been obtained based on quantum state tomography, and the error bars of the following fidelities depicted thrice of the standard deviation of those 10 batches.
Thus, we obtained the practical channel information $M$ through Eq. (\ref{M27}) by using the practical density matrices of $\varrho_{q_0q_1}$ and $\varrho _{q_0q_1}^{out}$. In addition, we can calculate the fidelity of the input state between the theoretical value and the practical value. The fidelity is defined as follows \cite{of},
\begin{equation}
F(\varrho,\tilde{\varrho}):= {\rm Tr}\sqrt{\varrho^{1/2}\tilde{\varrho}\varrho^{1/2}}.
\end{equation}
From Eq. (\ref{rhoin}), one can figure out that
\begin{equation}
F_{\varrho_{in}}=0.974\pm0.011.
\end{equation}
The fidelity of output state between the theoretical value and the practical value can also be obtained,
\begin{equation}
F_{\varrho_{out}}=0.954\pm0.027.
\end{equation}
In order to compare the theoretical value of $M$ in Eq. (\ref{thM}) and the practical value of $M$ through Eq. (\ref{M27}), we use $|0\rangle$, $|1\rangle$, $|+\rangle$, $|-\rangle$, $|L\rangle$ and $|R\rangle$ as input states, where $|+/-\rangle=(|0\rangle\pm|1\rangle)/\sqrt{2}$ and $|L/R\rangle=(|0\rangle\pm i|1\rangle)/\sqrt{2}$. We calculate the fidelities between the output states from the theoretical value $M$ and the practical value $M$,
\begin{eqnarray}
F_{out_{|0\rangle}}&=&0.999\pm0.004;\
F_{out_{|1\rangle}}=0.997\pm0.007;\nonumber\\
F_{out_{|+\rangle}}&=&0.968\pm0.050;\
F_{out_{|-\rangle}}=0.974\pm0.050;\nonumber\\
F_{out_{|L\rangle}}&=&0.998\pm0.002;\
F_{out_{|R\rangle}}=0.998\pm0.005.\nonumber
\end{eqnarray}
Although the actual measured value deviates from the theoretical value due to systematic or measurement errors when $\mathcal{R}$ is used to obtain information for the channel, its fidelity is still good to a certain extent. And the error is understandable.
\section{Discussions and conclusions}
We discuss some other possible roles of the operator $\check{R}$ in the Ref. \cite{GP}. It is well known that the realignment operation $\mathcal{R}$ is often used as the criterion of entanglement of quantum states. If a state is separable then the sum of all singular values of $\mathcal{R}(\varrho_{AB})$ is less than 1, and it is entangled when the sum of the singular values is greater than 1 \cite{pp,oo,op}. From Theorem 1, since the singular values of quantum states are same when they are acted on by $\check{R}$ and $\mathcal{R}$, we conclude that $\check{R}$ can also be used as a criterion for entanglement of quantum states.
Moreover, we have presented an example of 3$\times$3 bound entangled state, which is not faithful. Is it possible to prove that all the bound entangled states are unfaithful? If it is true, we can see that the set of bound entangled states is the subset of all unfaithful states. This is an interesting open question for our further research.
In Ref. \cite{os}, the authors present another necessary and sufficient condition for faithful states: A state $\sigma$ of $AB$ may be used to perform AAQPT if and
only if the Schmidt number of $\sigma$ is $d_A^2$, where $d_A$ is the dimension of the state space of system $A$. We can see that $\mathcal{R}(\sigma)$ is equivalent to the Schmidt
decomposition of $\sigma$ and the Schmidt number of $\sigma$ is equal to the number of the number of nonzero singular values of $\mathcal{R}(\sigma)$. Therefore, our condition ($\mathcal{R}(\sigma)^{-1}$ exists) is equivalent to the condition shown in Ref. \cite{os} as well.
In conclusion, it is found in Ref. \cite{GP} that the input state $\varrho_{AB}$ is faithful if $\check{R}(\varrho_{AB})$ is invertible, i.e., none of the singular values of $\check{R}(\varrho_{AB})$ are zero. In this paper, we find that $\mathcal{R}(\varrho_{AB})$ and $\check{R}(\varrho_{AB})$ have equal consequence on the solution of the singular value. By using $\mathcal{R}(\varrho_{AB})$, we can also get the complete channel information just like $\check{R}(\varrho_{AB})$. Furthermore, through this property, we use two entangled state to prove that not all entangled states are useful for AAQPT.
\section*{ACKNOWLEDGMENTS}
This work is supported by the National Natural Science Foundation of China (Grant No. 11734015), and K.C. Wong Magna Fund in Ningbo University.
|
\section{Introduction}
\label{sec:introduction}
String transformations are used in data transformations~\cite{gulwani2011flashfill}, sanitization of untrusted inputs~\cite{Bek2011,LorisVMCAI2013}, and many other domains~\cite{programmable-string-robustness}.
Because in these domains bugs may cause serious security vulnerabilities~\cite{Bek2011}, there has been increased interest
in building tools that can help programmers verify~\cite{Bek2011,LorisVMCAI2013} and synthesize~\cite{gulwani2011flashfill,symmetriclenses,bijectivelenses} string transformations.
Techniques for \textit{verifying} string transformations rely on automata-theoretic approaches that provide powerful decidability properties~\cite{Bek2011}.
On the other hand, techniques for \textit{synthesizing} string transformations rely on domain-specific languages (DSLs)~\cite{gulwani2011flashfill,symmetriclenses}.
These DSLs are designed to make synthesis practical and have to give up the closure and decidability properties enabled by automata-theoretic models.
The disconnect between the two approaches raises a natural question: \textit{Can one synthesize automata-based models and therefore retain and leverage their elegant properties?}
Transducers are a natural automata-based formal model for synthesizing string transformations.
A \textit{finite state transducer} (FT) is an automaton where each transition reads an input
character and outputs a string of output characters.
For instance, Figure~\ref{fig:escape_quotes} shows a transducer that `escapes' instances of the \texttt{"} character.
So, on input \texttt{a"\textbackslash "a}, the transducer outputs the string \texttt{a\textbackslash "\textbackslash \textbackslash "a}.
FTs have found wide adoption in a variety of domains~\cite{mohri1997finite,LorisVMCAI2013} because of their many desirable properties (e.g., decidable equivalence check and closure under composition~\cite{cacmSFA}).
There has been increasing work on building SMT solvers for strings that support transducers; the Ostrich tool \cite{ostrichATVA2020} allows a user to write programs in SMT where string-transformations are modelled using transducers.
One can then write constraints over such programs and use an SMT solver to automatically check for satisfiability or prove unsatisfiability of those constraints.
For example, given a program like the following:
\begin{verbatim}
y = escapeQuotes(x)
z = escapeQuotes(y)
assert(y==z) //Checking idempotence
\end{verbatim}
one can use Ostrich to write a set of constraints and use them to prove whether the assertion holds.
However, to do so, one needs to first write a transducer $T$ that implements the function \verb|escapeQuotes|.
However, writing transducers by hand is a cumbersome and error-prone task.
\begin{figure}[tb]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\resizebox{0.44\textwidth}{0.24\textwidth}{
\begin{tikzpicture}[shorten >=1pt, node distance=2.5cm, auto]
\node[state, accepting, initial] (q0) {$q_0$};
\node[state, accepting, right of=q0] (q1) {$q_1$};
\path[->]
(q0) edge[loop above] node{\texttt{a} $\rightarrow$ \texttt{a}} (q0)
(q0) edge[loop below] node{\texttt{"} $\rightarrow$ \texttt{\textbackslash "}} (q0)
(q0) edge node[above] {\texttt{\textbackslash} $\rightarrow$ \texttt{\textbackslash}} (q1)
(q1) edge[bend left=55] node{\texttt{a} $\rightarrow$ \texttt{a}} (q0)
(q1) edge[bend left=20] node{\texttt{"} $\rightarrow$ \texttt{"}} (q0)
(q1) edge[bend right=45, above] node {\texttt{\textbackslash} $\rightarrow$ \texttt{\textbackslash}} (q0)
;
\end{tikzpicture}
}
\caption{Transducer EscapeQuotes}
\label{fig:escape_finite_correct}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\vspace{1mm}
\centering
\fbox{%
\scriptsize
\parbox{0.95\textwidth}{%
\textbf{Examples:}
\{\texttt{a"a} $\mapsto$ \texttt{a\textbackslash"a}, \texttt{a\textbackslash \textbackslash a} $\mapsto$ \texttt{a\textbackslash \textbackslash a}, \texttt{a\textbackslash a} $\mapsto$ \texttt{a\textbackslash a}, \texttt{a\textbackslash "a} $\mapsto$ \texttt{a\textbackslash "a}, \texttt{\textbackslash} $\mapsto$ \texttt{\textbackslash}\} \\
\textbf{Types:}
\texttt{[a"]$^\ast$\textbackslash?|([a"]$^\ast$\textbackslash[a"\textbackslash][a"]$^\ast$)$^\ast$} $\to$
\texttt{a$^\ast$\textbackslash?|(a$^\ast$\textbackslash[a"\textbackslash]a$^\ast$)$^\ast$} \\
\textbf{Distance:} At most 1 edit per input character
}%
}
\caption{Specification to synthesize EscapeQuotes}
\label{fig:specification_box}
\end{subfigure}
\caption{Simplified version of EscapeQuotes from~\cite{Bek2011}.}
\label{fig:escape_quotes}
\end{figure}
In this paper, we present a technique for synthesizing transducers from high-level specifications.
We use three different specification mechanisms to quickly yield desirable transducers: input-output examples, input-output types, and input-output distances.
When provided with the specification in Figure~\ref{fig:specification_box}, our approach yields the transducer in Figure~\ref{fig:escape_quotes}.
While none of the three specification mechanisms are effective in isolation, they work well altogether.
Input-output examples are a natural specification mechanism that is easy to provide, but only capture finitely many inputs.
Similarly, input-output types are a natural way to prevent a transducer from generating undesired strings and can often be obtained from function/API specifications.
Last, input-output distances are a natural way to specify how much of the input string should be preserved by the transformation.
We show that if the size of the transducers is fixed, all such specifications can be encoded as a set of constraints
whose solution directly provides a transducer.
While the constraints for examples are fairly straightforward, we introduce two new ideas for encoding types and distances.
To encode types and distances, we show that one can use constraints to ``guess'' the simulation relation and the invariants necessary to prove that
the transducer has the given type and respects the given distance constraint.
Because our constraint-based approach is based on decision procedures and is modular, it can support more complex models of transducers:
(\emph{i})~ Symbolic Finite Transducers (s-FTs), which support large alphabets~\cite{LorisCAV17}, and
(\emph{ii})~ FTs with lookahead, which can express functions that otherwise require non-determinism.
In addition, closure properties of transducers allow us to reduce repair problems for string transformations to our synthesis problem.
\noindent \vspace{1mm}\textit{Contributions:}
We make the following contributions.
\begin{itemize}[topsep=2pt]
\item A constraint-based synthesis algorithm for synthesizing transducers from complex specifications (Sec.~\ref{sec:encoding}).
\item Extensions of our synthesis algorithm to more complex models---e.g., symbolic transducers and transducers with lookahead---and problems---e.g., transducer repair---that showcase the flexibility of our approach and the power of working with transducers, which enjoy strong theoretical properties---unlike domain-specific languages (Sec.~\ref{sec:extensions}).
\item \textsc{astra}\xspace: a tool that can synthesize and repair transducers and compares well with a state-of-the-art tool for synthesizing string transformations (Sec.~\ref{sec:eval}).
\end{itemize}
\noindent Proofs and additional results are available in the appendix.
\section{Transducer Synthesis Problem}
\label{sec:transducer-synthesis-problem}
In this section, we define the transducer synthesis problem.
A \textit{deterministic finite automaton} (DFA) over an alphabet $\Sigma$ is a tuple $D = (\setcustomstate{D}, \transition{D}, \initstate{D}, \finalstates{D})$: $\setcustomstate{D}$ is the set of states, $\transition{D}: \setcustomstate{D} \times \Sigma \rightarrow \setcustomstate{D}$ is the transition function, $\initstate{D}$ is the initial state, and $\finalstates{D}$ is the set of final states.
The extended transition function $\extendedtransition{D} : \setcustomstate{D} \times \Sigma^\ast \rightarrow \setcustomstate{D}$ is defined as $\extendedtransition{D}(q, \varepsilon) = q$ and $\extendedtransition{D}(q, au) = \extendedtransition{D}(\transition{D}(q, a), u)$.
We say that $D$ accepts a string $w$ if $\extendedtransition{D}(\initstate{D}, w) \in \finalstates{D}$.
The \textit{regular language} $\mathcal{L}(D)$ is the set of strings accepted by a DFA $D$.
A total \textit{finite state transducer} (FT) is a tuple $T = (\setcustomstate{T}, \delta_{T}^{\mathit{st}}, \delta_{T}^{\mathit{out}}, \initstate{T})$,
where
$\setcustomstate{T}$ are states and $\initstate{T}$ is the initial state.
Transducers have two transition functions:
$\delta_{T}^{\mathit{st}}: \trnsstate \times \Sigma \rightarrow \trnsstate$ defines the target state, while $\delta_{T}^{\mathit{out}}: \trnsstate \times \Sigma \rightarrow \Sigma^\ast$
defines the output string of each transition.
The extended function for states $\delta_{T}^{\mathit{st*}}$ is defined analogously to the extended transition function for a DFA.
The extended function for output strings is defined as
$\delta_{T}^{\mathit{out*}}(q, \varepsilon) =\varepsilon$ and $\delta_{T}^{\mathit{out*}}(q, au)= \delta_{T}^{\mathit{out*}}(q, a)\cdot \delta_{T}^{\mathit{out}}(\delta_{T}^{\mathit{st*}}(q, a), u)$.
Given a string $w$ we use $T(w)$ to denote $\delta_{T}^{\mathit{out*}}(\initstate{T}, w)$, i.e., the output string generated by $T$ on $w$.
Given two DFAs $P$ and $Q$, we write $\hoaretriple{P}{T}{Q}$ for a transducer $T$ iff for every string $s$ in $\DFAlanguage{P}$, the output string $T(s)$ belongs to $\DFAlanguage{Q}$.
An \textit{edit operation} on a string is either an insertion/deletion of a character, or a replacement of a character with a different one.
For example, editing the string \stringfont{ab} to the string \stringfont{acb} requires one edit operation, which is inserting a \stringfont{c} after the \stringfont{a}.
The \textit{edit distance} $\eddistfunc{s}{t}$ between two strings $s$ and $t$ is the number of edit-operations required to reach $t$ from $s$.
We use $\lenfunc{w}$ to denote the length of a string $w$.
The \textit{mean edit distance} $\meaneddist{s}{t}$ between two strings $s$ and $t$ is defined as $\eddistfunc{s}{t} / \textit{len}(s)$.
For example, the mean edit distance from \stringfont{ab} to \stringfont{acb} is $1/2=.5$.
We can now formulate the transducer synthesis problem.
We assume a fixed alphabet $\Sigma$.
If the specification requires that $s$ is translated to $t$, we write that as ${\inputexample \mapsto \outputexample}$.
\begin{problemstatement}[Transducer Synthesis] \sloppypar The transducer synthesis problem has the following inputs and output:\newline
\textbf{Inputs}
\begin{itemize}[topsep=1pt]
\item Number of states $k$ and upper bound $l$ on the length of the output of each transition.
\item Set of input-output examples $E=[\overline{\inputexample \mapsto \outputexample}]$.
\item Input-output types $P$ and $Q$, given as DFAs.
\item A positive upper bound $d\in\mathbb{Q}$ on the mean edit distance.
\end{itemize}
\textbf{Output} A total transducer $T = (\setcustomstate{T}, \delta_{T}^{\mathit{st}}, \delta_{T}^{\mathit{out}}, \initstate{T})$ with $k$ states such that:
\begin{itemize}[topsep=1pt]
\item Every transition of $T$ has an output with length at most $l$, i.e., $\forall \trnsstate \in \settrnsstate, \mathit{a} \in \Sigma.\ \lenfunc{\delta_{T}^{\mathit{out}}(q, a)} \leq l$.
\item $T$ is consistent with the examples: $\forall {\inputexample \mapsto \outputexample}\in E.\ T(s) = t$.
\item $T$ is consistent with input-output types, i.e., $\hoaretriple{P}{T}{Q}$.
\item For every string $w \inP$, $\meaneddist{w}{T(w)} \leq d$.
\end{itemize}
\end{problemstatement}
The synthesis problem that we present here is for FTs, and in Section~\ref{sec:encoding}, we provide a sound algorithm to solve it using a system of constraints.
One of our key contributions is that
our encoding can be easily adapted to synthesizing richer models than FTs (e.g., symbolic transducers~\cite{cacmSFA} and transducers with regular lookahead), while still using the same encoding building blocks (Section~\ref{sec:extensions}).
\section{Constraint-based Transducer Synthesis}
\label{sec:encoding}
In this section, we present a way to generate constraints to solve the transducer synthesis problem defined in Section~\ref{sec:transducer-synthesis-problem}.
The synthesis problem can then be solved by invoking a \textit{Satisfiability Modulo Theories} (SMT) solver on the constraints.
We use a constraint encoding, rather than a direct algorithmic approach because of the multiple objectives to be satisfied.
Synthesizing a transducer that translates a set of input-output examples is already an NP-Complete problem~\cite{minimaltransducersynthesis}.
On top of that, we also need to handle input-output types and distances.
Our encoding is divided into three parts, one for each objective, which are presented in the following subsections.
This division makes our encoding very modular and programmable. In Section~\ref{sec:extensions} we show how it can be adapted to different transducer models and problems.
We include a brief description of the size of the constraint encoding in Section~\ref{sec:size_encoding} of the appendix.
\subsection{Representing Transducers}
\label{sec:representingTransducers}
\looseness-1 The transducer we are synthesizing has $k$ (part of the problem input) states $Q_T = \allowbreak{} \{q_0, ..., \allowbreak{} q_{k-1}\}$.
We often use $\initstate{T}$ as an alternative for $q_0$, the initial state of $T$.
We illustrate how a transition $q_1 \xrightarrow{\texttt{a}/\texttt{bc}} q_2$ is represented in our encoding.
The target state is captured using an uninterpreted function $\normalfont{\texttt{d}^\texttt{st}}:Q_T\times \Sigma \to Q_T$, e.g., $\normalfont{\texttt{d}^\texttt{st}}(q_1,\texttt{a}) = q_2$.
Representing the output of the transition is trickier because its length is not known a priori.
The synthesis problem however provides an output bound $l$, which allows us to limit the number of characters that may appear in the output.
We use an uninterpreted function $\normalfont{\texttt{d}^\texttt{out}_\texttt{ch}}:Q_T\times \Sigma \times \{0,\ldots,l{-}1\} \to \Sigma$
to represent each character in the output string; in our example,
$\normalfont{\texttt{d}^\texttt{out}_\texttt{ch}}(q_1,\texttt{a},0) = \texttt{b}$ and
$\normalfont{\texttt{d}^\texttt{out}_\texttt{ch}}(q_1,\texttt{a},1) = \texttt{c}$.
Since an output string's length can be smaller than $l$, we use an additional uninterpreted function $\normalfont{\texttt{d}^\texttt{out}_\texttt{len}}{}:Q_T\times \Sigma \to \{0,\ldots,l\}$ to model the length of a transition's output; in our example
$\normalfont{\texttt{d}^\texttt{out}_\texttt{len}}(q_1,\texttt{a})=2$.
We say an assignment to the above variables extends to a transducer $T$ for the transducer $T$ obtained by instantiating $\delta^{st}$ and $\delta^{out}$ as described above.
\subsection{Input-output Examples}
\fbox{%
\parbox{0.47\textwidth}{%
\textbf{Goal:} For each input output-example ${\inputexample \mapsto \outputexample}\in E$, $T$ should translate $s$ to $t$.
}%
}
\vspace{1ex}
Translating $s$ to the correct output string means that $\delta_{T}^{\mathit{out*}}(\initstate{T}, s) = t$.
Generating constraints that capture this behavior of $T$ on an example is challenging because we do not know a priori what parts of $t$ are produced by what steps of the transducer's run.
Suppose that we need to translate $s=a_0a_1$ to $t=b_0b_1b_2$.
A possible solution is for the transducer to have the run $q_0 \xrightarrow{a_0/b_0} q1 \xrightarrow{a_1/b_1b_2} q_2$.
Another possible solution might be to instead have $q_0 \xrightarrow{a_0/b_0b_1} q1 \xrightarrow{a_1/b_2} q_2$.
Notice that the two runs traverse the same states but produce different parts of the output strings at each step.
Intuitively, we need a way to ``track'' how much output the transducer has produced before processing the $i$-th character in the input and what state it has landed in.
For every input example ${\inputexample \mapsto \outputexample}$ such that $s=a_0\cdots a_n$ and $t=b_0\cdots b_m$, we introduce an uninterpreted function $\normalfont{\texttt{config}}_s: \{0,\ldots,n\} \rightarrow \{0,\ldots,m\} \times \settrnsstate$ such that
$\normalfont{\texttt{config}}_s(i)=(j,\trnsstate)$ iff after reading $a_0\cdots a_{i-1}$, the transducer $T$ has produced the output $b_0\cdots b_{j-1}$ and reached state $\trnsstate$---i.e., $\delta_{T}^{\mathit{out*}}(q_0,a_0\cdots a_{i-1})=b_0\cdots b_{j-1}$ and $\delta_{T}^{\mathit{st*}}(q_0,a_0\cdots a_{i-1}) = \trnsstate$.
We describe the constraints that describe the behavior of $\normalfont{\texttt{config}}_s$.
Constraint~\ref{eq:examples1} states that a configuration must start at the initial state and be at position 0 in the output.
\begin{equation}
\label{eq:examples1}
\normalfont{\texttt{config}}_s(0) = (0, \initstate{T})
\end{equation}
Constraint~\ref{eq:examples2} captures how the configuration is updated when reading the $i$-th character of the input.
For every $0\leq i < n$, $0 \leq j < m$, $c\in\Sigma$, and $q_T\in Q_T$:
\begin{equation}
\label{eq:examples2}
\begin{split}
& \normalfont{\texttt{config}}_s(i) = (j, \trnsstate)
\wedge
a_i = c
\Rightarrow \\
&
\quad\quad[
\bigwedge_{0 \leq z < l} (\normalfont{\texttt{d}^\texttt{out}_\texttt{ch}}(\trnsstate, c,z) = b_{j+z} \vee z\geq \normalfont{\texttt{d}^\texttt{out}_\texttt{len}}(\trnsstate,c)) \wedge\\
&
\quad\quad\normalfont{\texttt{config}}_s(i + 1) = (j + \normalfont{\texttt{d}^\texttt{out}_\texttt{len}}(\trnsstate,c), \normalfont{\texttt{d}^\texttt{st}}(\trnsstate, c))
]
\end{split}
\end{equation}
Informally, if the $i$-th character is $c$ and the transducer has reached state $\trnsstate$ and produced the characters $b_0\cdots b_{j-1}$ so far, the transition reading $c$ from state $q_T$ outputs characters $b_{j}\cdots b_{j+f-1}$, where $f$ is the output length of the transition.
The next configuration is then $(j+f, \normalfont{\texttt{d}^\texttt{st}}(\trnsstate, c))$.
Finally, Constraint~\ref{eq:examples3} forces $T$ to be completely done with generating $t$ when $s$ has been entirely read.
Recall that $\stringlength{s} = n$ and $\stringlength{t} = m$.
\begin{equation}
\label{eq:examples3}
\bigvee_{\trnsstate \in \settrnsstate} \normalfont{\texttt{config}}_s(n)
= (m, \trnsstate)
\end{equation}
\noindent The constraint encoding for examples is sound and complete (Proofs in~\ref{sec:examples_proofs}).
\subsection{Input-Output Types}
\fbox{%
\parbox{0.47\textwidth}{%
\textbf{Goal:} $T$ should satisfy the property $\hoaretriple{P}{T}{Q}$.
}%
}
\vspace{1ex}
Encoding this property using constraints is challenging because it requires enforcing
that when $T$ reads one of the (potentially) infinitely many strings in $P$ it always outputs
a string in $Q$.
To solve this problem, we draw inspiration from how one proves that the property
$\hoaretriple{P}{T}{Q}$ holds---i.e., using a simulation relation that relates runs over $P$, $T$ and $Q$.
Intuitively, if $P$ has read some string $w$, we need to be able to encode the behavior of $T$ in terms
of $w$, i.e., what state of $T$ this transducer is in after reading $w$ and what output string $w'$ it produced.
Further, we also need to be able to encode in which state $Q$ would be after reading the output string $w'$.
We do this by introducing a function \normalfont{\texttt{sim}}{}: $\setsrcstate \times \settrnsstate \times \settrgstate \rightarrow \{0, 1\}$,
which preserves the following invariant: $\simulation{\srcstate}{\trnsstate}{\trgstate}$ holds if there exist strings $w,w'$ such that
$\extendedtransition{P}(\initstate{P}, w) = q_{P}$,
$\delta_{T}^{\mathit{st*}}(\initstate{T}, w) = \trnsstate$,
$\delta_{T}^{\mathit{out*}}(\initstate{T}, w) = w'$,
and $\extendedtransition{Q}(\initstate{Q}, w') = q_{Q}$.
Constraint~\ref{eq:types1} states the initial condition of the simulation---i.e., $P$, $T$, and $Q$ are in their initial states.
\begin{equation}
\label{eq:types1}
\simulation{\initstate{P}}{\initstate{T}}{\initstate{Q}}
\end{equation}
Constraint~\ref{eq:types2} encodes how we advance the simulation relation for states $\srcstate,\trnsstate,\trgstate$ and for a character $c\in\Sigma$, using free variables $c_0 \ldots, c_{l-1}$ and $\trgstate^0 \ldots, \allowbreak{} \trgstate^{l}$ that are separate for each combination of $\srcstate,\trnsstate,\trgstate$, and $c$:
\begin{equation}
\label{eq:types2}
\begin{split}
\simulation{\srcstate}{\trnsstate}{\trgstate} \Rightarrow &
\bigwedge_{\mathclap{0 \leq z \leq l}} (\normalfont{\texttt{d}^\texttt{out}_\texttt{len}}(\trnsstate, c) = z \Rightarrow \\
& \quad [\bigwedge_{\mathclap{0 \leq x < z}} \normalfont{\texttt{d}^\texttt{out}_\texttt{ch}}(\trnsstate, c, x) {=} c_x] \wedge \\
& \quad [\trgstate^0 {=} \trgstate \wedge \bigwedge_{\mathclap{1 \leq x < z}} \trgstate^x {=}
\normalfont{\texttt{d}_{\outputtype}}(\trgstate^{x-1},c_{x-1})]
\wedge \\
& \quad \simulation{\transition{P}(\srcstate, \mathit{c})}{\normalfont{\texttt{d}^\texttt{st}}(\trnsstate, c)}{\trgstate^z})
\end{split}
\end{equation}
Intuitively, if $\simulation{\srcstate}{\trnsstate}{\trgstate}$ and we read a character $c$, $P$ moves to $\transition{P}(\srcstate, \mathit{c})$
and $T$ moves to $\normalfont{\texttt{d}^\texttt{st}}{}(\srcstate, \mathit{c})$.
However, we also need to advance $Q$ and the $\normalfont{\texttt{d}^\texttt{out}_\texttt{len}}{}$ symbols produced by $\normalfont{\texttt{d}^\texttt{out}_\texttt{ch}}{}$.
We hard-code the transition relation $\transition{Q}$ in an uninterpreted function $\normalfont{\texttt{d}_{\outputtype}}: \settrgstate \times \Sigma \rightarrow \settrgstate$, and apply it to compute the output state reached when reading the output string.
E.g., if $\normalfont{\texttt{d}^\texttt{out}_\texttt{len}}{}(\trnsstate, c) = 2$ and $\normalfont{\texttt{d}^\texttt{out}_\texttt{ch}}(\trnsstate, c, 0)=c_0$ and $\normalfont{\texttt{d}^\texttt{out}_\texttt{ch}}(\trnsstate, c, 1)=c_1$, the next state in $Q$ is $\normalfont{\texttt{d}_{\outputtype}}(\normalfont{\texttt{d}_{\outputtype}}(\trgstate, c_0), c_1)$.
Lastly, Constraint~\ref{eq:types3} states that if we encounter a string in $\DFAlanguage{P}$---i.e., $P$ is in a state $\srcstate \in \finalstates{P}$---the relation does not contain a state $\trgstate \notin \finalstates{Q}$. Since $Q$ is deterministic, this means that $Q$ accepts $T$'s output.
\begin{equation}
\label{eq:types3}
\bigwedge_{\srcstate \in \finalstates{P}}
\bigwedge_{\trgstate \notin \finalstates{Q}}
\neg \simulation{\srcstate}{\trnsstate}{\trgstate}
\end{equation}
\noindent Soundness and completeness of the type constraints are proven in~\ref{sec:types_proofs}.
\subsection{Input-output Distance}
\fbox{%
\parbox{0.47\textwidth}{%
\textbf{Goal:} The mean edit distance between any input string $w$ in $\DFAlanguage{P}$ and the output string $T(w)$ should not exceed $d$.
}%
}
\vspace{1mm}
Capturing the edit distance for all the possible inputs in the language of $P$ and the corresponding outputs produced by the transducer
is challenging because these sets can be infinite.
Furthermore, exactly computing the edit distance between an input and an output string may involve comparing characters appearing on different transitions in the transducer run.
For example, consider the transducer shown in Figure~\ref{fig:distance_example1} and suppose that we are only interested in strings in the input type $P = \texttt{a(ba)}{\ast}\texttt{a}$.
The first transition from $q_0$ deletes the \texttt{a}, therefore making 1 edit.
This transducer has a cycle between states $q_1$ and $q_2$, which can be taken any number of times.
Each iteration, locally, would require that we make 2 edits: one to change the \texttt{b} to \texttt{a}, and the other to change the \texttt{a} to \texttt{b}.
However, the total number of edits made over any string in the input type $P = \texttt{a(ab)}{\ast}\texttt{a}$ by this transducer is 1, because the transducer changes strings of the form $\texttt{a(ba)}^n\texttt{a}$ to be of the form $\texttt{(ab)}^n\texttt{a}$.
Looking at the transitions in isolation, we are prevented from deducing that the edit distance is always 1 because the first transition delays outputting a character.
If there was no such delay, as is the case for the transducer in Figure~\ref{fig:distance_example2}, which is equivalent on the relevant input type to the one in Figure~\ref{fig:distance_example1}, then this issue would not arise.
We take inspiration from Benedikt et al.~\cite{Benedikt11} and focus on the simpler problem of synthesizing a transducer that has `aggregate cost' that satisfies the given objective.\footnote{Benedikt et al.~\cite{Benedikt11} studied a variant of the problem where the distance is bounded by some finite constant.
Their work shows that when there is a transducer between two languages that has some bounded global edit distance, then there is also a transducer that is bounded (but with a different bound) under a local method of computing the edit distance---i.e., one where the computation of the edit distance is done transition by transition.}
For a transducer $T$ and string $s = a_0\ldots a_n$, let $\initstate{T} \xrightarrow{a_0 / y_0} \trnsstate^1
\ldots
\trnsstate^{n} \xrightarrow{a_n / y_n} \trnsstate^{n+1}$ be the run of $s$ on $T$.
Then, the \textit{aggregate cost} of $T$ on $s$ is the sum of the edit distances $\eddistfunc{a_i}{y_i}$ over all indices $0 \leq i \leq n$.
The \textit{mean aggregate cost} of $T$ on $s$ is the aggregate cost divided by $\lenfunc{s}$, the length of $s$.
It follows that if $T$ has a mean aggregate cost lower than some specified $d$ for every string, then it also has a mean edit distance lower than $d$ for every string.
However, the mean aggregate cost overapproximates the edit distance, e.g., the transducer in Figure~\ref{fig:distance_example1} has mean aggregate cost 1, while the mean edit distance when considering only strings in $P = \texttt{a(ab)}{\ast}\texttt{a}$ is less than $1/2$.
For this reason, if the mean edit distance objective was set to $1/2$, our constraint encoding can only synthesize the transducer in Figure~\ref{fig:distance_example2}, and not the equivalent one in Figure~\ref{fig:distance_example1}.
\begin{figure}[t]
\captionsetup[subfigure]{justification=centering}
\centering
\begin{subfigure}{0.23\textwidth}
\centering
\resizebox{0.98\textwidth}{0.45\textwidth}{
\begin{tikzpicture}[shorten >=1pt, node distance=1.75cm, auto]
\node[state, accepting, initial] (q0) {$q_0$};
\node[state, accepting, right of=q0] (q1) {$q_1$};
\node[state, accepting, above of=q1] (q2) {$q_2$};
\node[state, accepting, right of=q1] (q3) {$q_3$};
\path[->]
(q0) edge node {$\texttt{a} {\rightarrow} \epsilon$} (q1)
(q1) edge[bend left=30] node {$\texttt{b} {\rightarrow} \texttt{a}$} (q2)
(q2) edge[bend left=30] node {$\texttt{a} {\rightarrow} \texttt{b}$} (q1)
(q1) edge node {$\texttt{a} {\rightarrow} \texttt{a}$} (q3)
;
\end{tikzpicture}
}
\caption{Transducer with delayed output}
\label{fig:distance_example1}
\end{subfigure}
~
\begin{subfigure}{0.23\textwidth}
\centering
\resizebox{0.98\textwidth}{0.45\textwidth}{
\begin{tikzpicture}[shorten >=1pt, node distance=1.75cm, auto]
\node[state, accepting, initial] (q0) {$q_0$};
\node[state, accepting, right of=q0] (q1) {$q_1$};
\node[state, accepting, above of=q1] (q2) {$q_2$};
\node[state, accepting, right of=q1] (q3) {$q_3$};
\path[->]
(q0) edge node {$\texttt{a} {\rightarrow} \texttt{a}$} (q1)
(q1) edge[bend left=30] node {$\texttt{b} {\rightarrow} \texttt{b}$} (q2)
(q2) edge[bend left=30] node {$\texttt{a} {\rightarrow} \texttt{a}$} (q1)
(q1) edge node {$\texttt{a} {\rightarrow} \epsilon$} (q3)
;
\end{tikzpicture}
}
\caption{Transducer without delay}
\label{fig:distance_example2}
\end{subfigure}
\caption{Transducers with and without delay.}
\end{figure}
Our encoding is complete for transducers in which the aggregate cost coincides with the actual edit distance.
We leave the problem of being complete with regards to global edit distance as an open problem.
In fact, we are not even aware of an algorithm for \textit{checking} (instead of synthesizing) whether a transducer satisfies a mean edit distance objective.\footnote{The mean edit distance is similar to mean payoff~\cite{DBLP:reference/mc/BloemCJ18}, which discounts a cost by the length of a string and looks at the behavior of a transducer in the limit. Our distance is different because 1) it looks at finite-length strings, and 2) it requires computing the edit distance, which cannot be done one transition at a time.}
In Section~\ref{sec:regular-look-ahead}, we present transducers with lookahead, which can mitigate this source of incompleteness.
Furthermore, our evaluation shows that using the aggregate cost and enabling lookahead are both effective techniques in practice.
We can now present our constraints.
First, we provide constraints for the edit distance of individual
transitions (recall that transitions are being synthesized and we therefore
need to compute their edit distances separately).
Secondly, we provide constraints that implicitly compute state invariants to capture the aggregate cost between
input and output strings at various points in the computation.
We are given a rational number $d$ as an input to the problem, which is the allowed distance bound.
\mypar{Edit Distance of Individual Transitions.}
To compute the edit distance between the input and the output of each transition, we introduce a function \normalfont{\texttt{ed}}{}: $\settrnsstate \times \Sigma \rightarrow \mathbb{Z}$.
For a transition from state $\trnsstate$ reading a character $c$, $\normalfont{\texttt{ed}}{}(\trnsstate,c)$ represents the edit distance between $c$ and $\delta_{T}^{\mathit{out}}(\trnsstate, c)$. Notice that this quantity is bounded by the output bound $l$.
The constraints to encode the value of this function are divided into two cases:
i) the output of the transition contains the input character $c$ (Constraint \ref{eq:distance1}),
ii) the output of the transition \textit{does not} contain the input character $c$ (Constraint \ref{eq:distance2}).
In both cases, the values are set via a simple case analysis on whether the length of the output is 0 (edit distance is 1) or not (the edit distance is related to the length of the output).
\begin{equation}
\label{eq:distance1}
\begin{split}
& [\bigvee_{\mathclap{0 \leq z < \normalfont{\texttt{d}^\texttt{out}_\texttt{len}}(\trnsstate, c)}} \normalfont{\texttt{d}^\texttt{out}_\texttt{ch}}(\trnsstate, c, z) = c]
\Rightarrow \\
& \quad [
\normalfont{\texttt{d}^\texttt{out}_\texttt{len}}(\trnsstate, c) = 0 \Rightarrow
\normalfont{\texttt{ed}}(\trnsstate, c) = 1 \wedge \\
& \quad \normalfont{\texttt{d}^\texttt{out}_\texttt{len}}(\trnsstate, c) \neq 0 \Rightarrow
\normalfont{\texttt{ed}}(\trnsstate, c) = \normalfont{\texttt{d}^\texttt{out}_\texttt{len}}(\trnsstate, c) - 1 ]
\end{split}
\end{equation}
\begin{equation}
\label{eq:distance2}
\begin{split}
& [\bigwedge_{\mathclap{0 \leq z < \normalfont{\texttt{d}^\texttt{out}_\texttt{len}}(\trnsstate, c)}} \normalfont{\texttt{d}^\texttt{out}_\texttt{ch}}(\trnsstate, c, z) \neq c]
\Rightarrow \\
& \quad [
\normalfont{\texttt{d}^\texttt{out}_\texttt{len}}(\trnsstate, c) = 0 \Rightarrow
\normalfont{\texttt{ed}}(\trnsstate, c) = 1 \wedge \\
& \quad \normalfont{\texttt{d}^\texttt{out}_\texttt{len}}(\trnsstate, c) \neq 0 \Rightarrow
\normalfont{\texttt{ed}}(\trnsstate, c) = \normalfont{\texttt{d}^\texttt{out}_\texttt{len}}(\trnsstate, c) ]
\end{split}
\end{equation}
\mypar{Edit Distance of Arbitrary Strings.}
Suppose that $T$ has the transitions $q_0 \xrightarrow{\texttt{a}/\texttt{a}} q1 \xrightarrow{\texttt{a}/\texttt{bc}} q_2$, and the specified mean edit distance is $d=0.5$.
The edit distance is 0 for the first transition and 2 for the second one.
For the input string $\texttt{aa}$, the mean aggregate cost is $2 / 2$, which means that the specification is not satisfied.
In general, we cannot keep track of every input string in the input type and look at its length and the number of edits that were made over it.
So, how can we compute the mean aggregate cost over any input string?
The first part of our solution is to scale the edit distance over a single transition depending on the specified mean edit distance.
This operation makes it such that an input string is under the edit distance bound if the sum of the weighted edit distances of its transitions is $\geq 0$.
The invariant we need to maintain is that the sum of the weights at any stage of the run gives us where we are with regard to the mean aggregate cost.
For each transition we compute the difference between the edit distance over the transition and the specified mean edit distance $d$.
We introduce the uninterpreted function $\normalfont{\texttt{wed}} : \settrnsstate \times \Sigma \rightarrow \mathbb{Q}$, which stands for weighted edit distance.
For a transition at $\trnsstate$ reading a character $c$, the weighted edit distance is given by $\normalfont{\texttt{wed}}(\trnsstate, c) = d - \normalfont{\texttt{ed}}(\trnsstate, c)$.
The sum of the weights of all transitions tells us the cumulative difference.
Going back to our example, the weighted edit distances of the two transitions are $\normalfont{\texttt{wed}}(q_0,\texttt{a})=0.5$ and $\normalfont{\texttt{wed}}(q_1,\texttt{a})=-1.5$, making the cumulative distance $-1$ and implying that the specification is violated.
We can now compute the mean edit distance over a run without keeping track of the length of the run and the number of edits performed over it.
We still need to compute the weighted edit distance for every string in the possibly infinite language $\DFAlanguage{P}$.
Building on the idea of simulation from the previous section, we introduce a new function called $\normalfont{\texttt{en}}: \setsrcstate \times \settrnsstate \times \settrgstate \rightarrow \mathbb{Q}$, which tracks an upper bound on the sum of the distances so far at that point in the simulation.
This function is similar to a \textit{progress measure}, which is a type of invariant used to solve \textit{energy games} \cite{progressmeasure}, a connection we expand on in Section~\ref{sec:related_work}.
In particular, we already know that if there exist strings $w,w'$ such that
$\extendedtransition{P}(\initstate{P}, w) = q_{P}$,
$\delta_{T}^{\mathit{st*}}(\initstate{T}, w) = \trnsstate$,
$\delta_{T}^{\mathit{out*}}(\initstate{T}, w) = w'$,
and $\extendedtransition{Q}(\initstate{Q}, w') = q_{Q}$, then we have $\simulation{\srcstate}{\trnsstate}{\trgstate}$.
Let this run over $T$ be denoted by $\initstate{T} \xrightarrow{a_0/y_0} \trnsstate^1 \ldots \trnsstate^{n-1} \xrightarrow{a_{n-1}/y_{n-1}} \trnsstate$, where $w = a_0\cdots a_{n-1}$, $w' = y_0\cdots y_{n-1}$, and $\trnsstate = \trnsstate^n$.
We have that $\normalfont{\texttt{en}}(\srcstate, \trnsstate, \trgstate) \geq \sum_{i = 0}^{n-1} \normalfont{\texttt{wed}}(\trnsstate^i, a_i)$.
The \normalfont{\texttt{en}}{} function is a budget on the number of edits we can still perform.
At the initial states, we start with no `initial credit' and the energy is 0.
\begin{equation}
\label{eq:distance4}
\normalfont{\texttt{en}}(\initstate{P}, \initstate{T}, \initstate{Q}) = 0
\end{equation}
Constraint~\ref{eq:distance5} bounds the energy budget according to the weighted edit distance of a transition by computing the minimum budget required at any point to still satisfy the distance bound. For each combination of $\srcstate,\trnsstate,\trgstate$, and $c \in \Sigma$, the constraint uses free variables $c_0, \ldots, c_{l}$ and $\trgstate^0, \ldots, \trgstate^{l-1}$:
\begin{equation}
\label{eq:distance5}
\begin{split}
\hspace{-18mm}&\bigwedge_{\mathclap{0 \leq z < l}} (\normalfont{\texttt{d}^\texttt{out}_\texttt{len}}(\trnsstate, c) {=} z \Rightarrow \\
& [\bigwedge_{\mathclap{0 \leq x < z}} \normalfont{\texttt{d}^\texttt{out}_\texttt{ch}}(\trnsstate, c, x) {=} c_x] {\wedge}
[\trgstate^0 {=} \trgstate \wedge \bigwedge_{\mathclap{1 \leq x < z}} \trgstate^x {=}
\normalfont{\texttt{d}_{\outputtype}}(\trgstate^{x-1},c_{x-1})] \wedge \\
& \energyfunc{\srcstate}{\trnsstate}{\trgstate} \geq \energyfunc{\transition{P}(\srcstate, c)}{\normalfont{\texttt{d}^\texttt{st}}(\trnsstate, c)}{\trgstate^z} {-}
\normalfont{\texttt{wed}}(\trnsstate, c))
\end{split}
\end{equation}
In our example, Constraint~\ref{eq:distance5} encodes that the energy at $q_0$ can be 1 less than that at $q_1$, but that the energy at $q_1$ needs to be $3$ greater than at $q_2$ since we need to spend 3 edit operations over the second transition.
At any point during a run, the transducer is allowed to go below the mean edit distance and then `catch up' later because we only care about the edit distance when the transducer has finished reading a string in $\DFAlanguage{P}$.
Therefore, when we reach a final state of $P$, the transducer should not be in `energy debt'.
\begin{equation}
\label{eq:distance6}
\bigwedge_{\srcstate \in \finalstates{P}} \simulation{\srcstate}{\trnsstate}{\trgstate} \Rightarrow \energyfunc{\srcstate}{\trnsstate}{\trgstate} \geq 0
\end{equation}
\noindent The encoding presented in this section is sound (Proofs in~\ref{sec:distances_proofs}).
\section{Richer Models and Specifications}
\label{sec:extensions}
We extend our technique to more expressive models (Sections~\ref{sec:symbolic_extension} and~\ref{sec:regular-look-ahead}) and show how our synthesis approach can be used not only to synthesize transducers, but also to repair them (Section~\ref{sec:repair}).
Furthermore, in Appendix~\ref{sec:distances}, we describe an encoding of an alternative distance measure.
\subsection{Symbolic Transducers}
\label{sec:symbolic_extension}
Symbolic finite automata (s-FA) and transducers (s-FT) extend their non-sym\-bo\-lic counterparts by allowing transitions to carry predicates and functions to represent (potentially infinite) sets of input characters and output strings.
Figure~\ref{fig:symbolic_transducer} shows an s-FT that extends the escapeQuotes transducer from Figure~\ref{fig:escape_finite_correct} to handle alphabetic characters.
The bottom transition from $q_0$ reads a character \texttt{"} (bound to the variable $x$) and outputs the string \texttt{\textbackslash "} (i.e., a \texttt{\textbackslash} followed by the character stored in $x$).
Symbolic finite automata (s-FA) are s-FTs with no outputs.
To simplify our exposition, we focus on s-FAs and s-FTs that only operate over ASCII characters that are ordered by their codes.
In particular, all of our predicates are unions of intervals over characters (i.e., $ x \neq \texttt{\textbackslash}$ is really the union of intervals [\texttt{NUL}-\texttt{[}] and [\texttt{]}-\texttt{DEL}]); we often use the predicate notation instead of explicitly writing the intervals for ease of presentation.
Furthermore, we only consider two types of output functions: constant characters and offset functions of the form $x+k$ that output the character obtained by taking the input $x$ and adding a constant $k$ to it---e.g., applying $x+(-32)$ to a lowercase alphabetic letter gives the corresponding uppercase letter.
In the rest of the section, we show how we can solve the transducer synthesis problem in the case where $P$ and $Q$ are s-FAs and the goal is to synthesize an s-FT (instead of an FT) that meets the given specification.
Intuitively, we do this by `finitizing' the alphabet of the now symbolic input-output types, synthesizing a finite transducer over this alphabet using the technique presented in Section~\ref{sec:encoding}, and then extracting an s-FT from the solution.
\mypar{Finitizing the Alphabet.}
The idea of finitizing the alphabet of s-FAs is a known one~\cite{cacmSFA} and is based on the concept of $\mathit{minterms}$, which is the set of maximal satisfiable Boolean combinations of the predicates appearing in the s-FAs.
For an s-FA $M$, we can define its set of predicates as: $\mathit{Predicates}(M) = \{\phi \mid q \xrightarrow{\phi} q' \in \transition{M} \}$.
The set of minterms $\mathit{mterms}(M)$ is the set of satisfiable Boolean combinations of all the predicates in $\mathit{Predicates}(M)$.
For example, for the set of predicates over the s-FT escapeQuotes in Figure~\ref{fig:symbolic_transducer}, we have that $\mathit{mterms}(\text{escapeQuotes}) =\{x \neq \texttt{"} \wedge x \neq \texttt{\textbackslash},
x = \texttt{"},
x = \texttt{\textbackslash}\}$.
The reader can learn more about minterms in \cite{cacmSFA}.
We assign each minterm a representative character, as indicated in Figure~\ref{fig:minterms}, and then construct a finite automaton from the resulting finite alphabet $\Sigma$.
For a character $c \in \Sigma$, we refer to its corresponding minterm by $\mintermpred{c}$.
In the other direction, for each minterm $\psi \in \mathit{minterms}(M)$, we refer to its uniquely determined representative character by $\witnessshort{\psi}$.
\begin{figure}[t]
\captionsetup[subfigure]{justification=centering}
\centering
\begin{subfigure}{0.23\textwidth}
\centering
\resizebox{0.98\textwidth}{0.45\textwidth}{
\begin{tikzpicture}[shorten >=1pt, node distance=3cm, auto]
\node[state, accepting, initial] (q0) {$q_0$};
\node[state, accepting, right of=q0] (q1) {$q_1$};
\path[->]
(q0) edge[loop above] node{$ x \neq \texttt{"} \wedge x \neq \texttt{\textbackslash}$ $\rightarrow$ $x$} (q0)
(q0) edge[loop below] node{$ x = \texttt{"}$ $\rightarrow$ \texttt{\textbackslash}$x$} (q0)
(q0) edge node[above] {$ x = \texttt{\textbackslash}$ $\rightarrow$ $x$} (q1)
(q1) edge[bend left] node{$ x \neq \texttt{\textbackslash}$ $\rightarrow$ $x$} (q0)
(q1) edge[bend right, above] node {$ x = \texttt{\textbackslash}$ $\rightarrow$ $x$} (q0)
;
\end{tikzpicture}
}
\caption{escapeQuotes s-FT}
\label{fig:symbolic_transducer}
\end{subfigure}
~
\begin{subfigure}{0.23\textwidth}
\centering
\resizebox{0.98\textwidth}{0.45\textwidth}{
\begin{tikzpicture}[shorten >=1pt, node distance=2.5cm, auto]
\node[state, accepting, initial] (q0) {$q_0$};
\node[state, accepting, right of=q0] (q1) {$q_1$};
\path[->]
(q0) edge[loop above] node{\texttt{a} $\rightarrow$ \texttt{a}} (q0)
(q0) edge[loop below] node{\texttt{"} $\rightarrow$ \texttt{\textbackslash "}} (q0)
(q0) edge node[above] {\texttt{\textbackslash} $\rightarrow$ \texttt{\textbackslash}} (q1)
(q1) edge[bend left=55] node{\texttt{a} $\rightarrow$ \texttt{a}} (q0)
(q1) edge[bend left=20] node{\texttt{"} $\rightarrow$ \texttt{"}} (q0)
(q1) edge[bend right=45, above] node {\texttt{\textbackslash} $\rightarrow$ \texttt{\textbackslash}} (q0)
;
\end{tikzpicture}
}
\caption{$\finite{escapeQuotes}$}
\label{fig:escape_finite}
\end{subfigure}
\vspace{1ex}
\begin{subfigure}{0.47\textwidth}
\centering
\fbox{%
\scriptsize
\parbox{0.95\textwidth}{%
\textbf{minterms:}
$ [x \neq \texttt{"} \wedge x \neq \texttt{\textbackslash}] $,
$ [x = \texttt{"}] $,
$ [x = \texttt{\textbackslash}] $ \\
\textbf{witness char:}
$ \witnessshort{[x \neq \texttt{"} \wedge x \neq \texttt{\textbackslash}]} {=} \texttt{a} $,
$ \witnessshort{[x = \texttt{"}]} {=} \texttt{"} $,
$ \witnessshort{[x = \texttt{\textbackslash}]} {=} \texttt{\textbackslash} $
}%
}
\caption{Set of minterms and their witness elements}
\label{fig:minterms}
\end{subfigure}
\caption{Example of Finitization}
\label{fig:minterm_reduction}
\end{figure}
For an s-FA $M$, we denote its corresponding FA over the alphabet $\mathit{mterms}(M)$ with $\finite{M}$.
Given an s-FA $M$, the set of transitions of $\finite{M}$ is defined as follows:
\[
\transition{\finite{M}} {=} \{q \xrightarrow{\witnessshort{\psi}} q' {\mid} q \xrightarrow{\phi} q' \wedge \psi \in \mathit{mterms}(M) \wedge \text{IsSat}
(\psi \land \phi)\}
\]
This algorithm replaces a transition guarded by a predicate $\phi$ in the given s-FA with a set of transitions consisting of the witnesses of the minterms where $\phi$ is satisfiable.
In interval arithmetic this is the set of intervals that intersect with the interval specified by $\phi$.
The transition from $q_1$ guarded by the predicate $[x \neq \texttt{\textbackslash}]$ in Figure~\ref{fig:symbolic_transducer} intersects with 2 minterms $[x \neq \texttt{"} \wedge x \neq \texttt{\textbackslash}]$ and $[x = \texttt{"}]$.
As a result, we see that this transition is replaced by two transitions in Figure~\ref{fig:escape_finite}, one that reads \texttt{"} and another that reads \texttt{a}.
\mypar{From FTs to s-FTs.}
Once we have synthesized an FT $T$, we need to extract an s-FT from it.
There are many s-FTs equivalent to a given FT and here we present one way of doing this conversion which is used in our implementation.
Let the size of an interval $I$ (the number of characters it contains) be given by $\mathit{size}(I)$, and the offset between 2 intervals $I_1$ and $I_2$ (i.e. the difference between the least elements of $I_1$ and $I_2$) be given by $\mathit{offset}(I_1,I_2)$.
Suppose we have a transition $q \xrightarrow{c/y_0\cdots y_n} q'$, where $c, y_i \in \Sigma$.
Then, we construct a transition $q \xrightarrow{\mintermpred{c}/f_0\cdots f_n} q'$, where for each $y_i$, the corresponding function $f_i$ is determined by the following rules ($x$ always indicates variable bound to the input predicate):
\begin{enumerate}[topsep=1pt]
\item If $c = y_i$, then $f_i = (x)$, i.e. the identity function.
\item If $\mintermpred{c}$ and $\mintermpred{y_i}$ consist of single intervals $I_1$ and $I_2$, respectively, such that $\mathit{size}(I_1) = \mathit{size}(I_2)$ , then $f_i = (x + \mathit{offset}(I_1,I_2))$.
For instance, if the input interval is \texttt{[a-z]} and the output interval is \texttt{[A-Z]}, then the output function is $(x + (-32))$, which maps lowercase letters to uppercase ones.
\item Otherwise $f_i = y_i$---i.e., the output is a character in the output minterm.
%
\end{enumerate}
While our s-FT recovery algorithm is sound, it may apply case 3 more often than necessary and introduce many constants, therefore yielding a transducer that does not generalize well to unseen examples.
However, our evaluation shows that our technique works well in practice.
We provide a sketch of the proof of soundness of this algorithm in Appendix~\ref{sec:sft_synthesis_soundness}.
\subsection{Synthesizing Transducers with Lookahead}
\label{sec:regular-look-ahead}
Deterministic transducers cannot express functions where the output at a certain transition depends on future characters in the input.
Consider the problem of extracting all substrings of the form \texttt{<x>} (where $\texttt{x} \neq \texttt{<}$) from an input string.
This is the \emph{getTags} problem from \cite{POPL12SFT}.
A deterministic transducer cannot express this transformation because when it reads \texttt{<} followed by \texttt{x} it has to output \texttt{<x} if the next character is a \texttt{>} and nothing otherwise.
However, the transducer does not have access to the next character!
Instead, we extend our technique to handle deterministic transducers with lookahead, i.e., the ability to look at the string suffix when reading a symbol.
Formally, a \textit{Transducer with Regular Lookahead} is a pair $(T, R)$ where $T$ is an FT with $\Sigma_T = Q_{\lookaheadaut} \times \Sigma$, and $R$ is a total DFA with $\Sigma_{R} = \Sigma$.
The transducer $T$ now has another input in its transition function, although it still only outputs characters from $\Sigma$, i.e., $\delta_{T}^{\mathit{out}}:\settrnsstate \times (Q_{\lookaheadaut} \times \Sigma) \rightarrow \Sigma$, and $\delta_{T}^{\mathit{st}}:\settrnsstate \times (Q_{\lookaheadaut} \times \Sigma) \rightarrow \settrnsstate$.
The semantics is defined as follows.
Given a string $w = a_0\cdots a_{n}$, we define a function $\lookahead_w$ such that $\lookahead_w(i)=\transition{R} (\initstate{R}, a_{n}\cdots a_{i+1})$.
In other words, $\lookahead_w(i)$ gives the state reached by $R$ on the reversed suffix starting at $i + 1$.
At each step $i$, the transducer $T$ reads the symbol $(a_i,\lookahead_w(i))$.
The extended transition functions now take as input a lookahead word, which is a sequence of pairs of lookahead states and characters, i.e., from $(Q_{\lookaheadaut} \times \Sigma)^\ast$.
To synthesize transducers with lookahead, we introduce uninterpreted functions $\normalfont{\texttt{d}_\texttt{\lookaheadaut}}$ for the transition function of $R$, and $\normalfont{\texttt{look}}_{w}$ for the $r$-values of $w$ on $R$.
Additionally, we introduce a bound $k_R$ on the number of states in the lookahead automaton $R$ as our synthesis algorithm has to synthesize both $T$ and $R$ at the same time.
Appendix~\ref{sec:lookahead_constraints} provides the modified constraints needed to encode input-output types and input-output examples to use lookahead.
Part of the transducer with lookahead we synthesize for the getTags problem is shown in Figure~\ref{fig:get_tags_lookahead}.
Notice that there are 2 transitions out of $q_1$ for the same input but different lookahead state: the string \texttt{<x} is outputted when the lookahead state is $r_1$.
\begin{figure}[t]
\captionsetup[subfigure]{justification=centering}
\centering
\begin{subfigure}{0.23\textwidth}
\centering
\resizebox{0.98\textwidth}{0.45\textwidth}{
\begin{tikzpicture}[shorten >=1pt, node distance=3cm, auto]
\node[state, accepting, initial] (q0) {$q_0$};
\node[state, accepting, right of=q0] (q1) {$q_1$};
\path[->]
(q0) edge node {$x = \texttt{<}, r_0 \rightarrow \epsilon$} (q1)
(q1) edge[loop above] node {$ x \neq \texttt{<} \wedge x \neq \texttt{>}, r_1 \rightarrow \texttt{<x}$} (q1)
(q1) edge[loop below] node {$ x \neq \texttt{<} \wedge x \neq \texttt{>}, r_0 \rightarrow \epsilon$} (q1)
;
\end{tikzpicture}
}
\caption{Subset of transitions in $T$}
\label{fig:transducer_lookahead}
\end{subfigure}~
\begin{subfigure}{0.23\textwidth}
\centering
\resizebox{0.98\textwidth}{0.45\textwidth}{
\begin{tikzpicture}[shorten >=1pt, node distance=2.5cm, auto]
\node[state, accepting, initial] (r0) {$r_0$};
\node[state, accepting, right of=q0] (r1) {$r_1$};
\path[->]
(r0) edge[loop above] node{$ x \neq \texttt{<} \wedge x \neq \texttt{>}$} (r0)
(r0) edge[loop below] node{$ x = \texttt{<}$} (r0)
(r0) edge node[below] {$ x = \texttt{>}$} (r1)
(r1) edge[loop above] node {$ x \neq \texttt{<} \wedge x \neq \texttt{>}$} (r1)
(r1) edge[bend right=20] node[above] {$ x = \texttt{>}$} (r0)
(r1) edge[bend left=30] node[below] {$ x = \texttt{<}$} (r0)
;
\end{tikzpicture}
}
\caption{Lookahead automaton $R$}
\label{fig:lookahead_aut}
\end{subfigure}
\caption{Regular lookahead for getTags}
\label{fig:get_tags_lookahead}
\end{figure}
\mypar{Lookahead and aggregate cost:}
Lookahead can help representing transducers, even deterministic ones, in a way that has lower aggregate cost---i.e., the aggregate cost better approximates the actual edit distance.
Suppose that we want to synthesize a transducer that translates the string \texttt{abc} to \texttt{ab} and the string \texttt{abd} to \texttt{bd}.
This translation can be done using a deterministic transducer with transitions $q_0 \xrightarrow{a/\epsilon} q_1 \xrightarrow{b/\epsilon} q_2$, followed by two transitions from $q_2$ that choose the correct output based on the next character.
Such a transducer would have a high aggregate cost of 4, even though the actual edit distance is 1.
In contrast, using lookahead we can obtain a transducer that can output each character when reading it; this transducer will have aggregate cost 1 for either string.
We conjecture that for every transducer $T$, there always exists an equivalent transducer with regular lookahead $(T',R)$ for which the edit distance computation for aggregate cost coincides with the actual edit distance of $T$.
\subsection{Transducer Repair}
\label{sec:repair}
In this section, we show how our synthesis technique can also be used to ``repair'' buggy transducers.
The key idea is to use the closure properties of automata and transducers---e.g., closure under union and sequential compositions~\cite{cacmSFA}---to reduce repair problems to synthesis ones.
The ability to algebraically manipulate transducers and automata is one of the key aspects that distinguishes our work from other synthesis works that use domain-specific languages~\cite{gulwani2011flashfill,symmetriclenses}.
We describe two settings in which we can repair an incorrect transducer $T_{\textit{bad}}$:
\noindent \textbf{1.} Let $\hoaretriple{P}{T_{\textit{bad}}}{Q}$ be an input-output type violated by $T_{\textit{bad}}$ and let $\outputlang{P}{T_{\textit{bad}}}$ be the finite automaton describing the set of strings $T_{\textit{bad}}$ can output when fed inputs in $P$ (this is computable thanks to closure properties of transducers).
We are interested in the case where $\outputlang{P}{T_{\textit{bad}}}\setminus Q\neq \emptyset$---i.e., $T_{\textit{bad}}$ can produce strings that are not in the output type.
\noindent \textbf{2.} Let $[\overline{\inputexample \mapsto \outputexample}]$ be a set of input-output examples. We are interested in the case where there is some example ${\inputexample \mapsto \outputexample}$ such that $T_{\textit{bad}}(s) \neq t$.
We describe two ways in which we repair transducers.
\mypar{Repairing from the Input Language.}
This approach synthesizes a new transducer for the inputs on which $T_{\textit{bad}}$ is incorrect.
Using properties of transducers, we can compute an automaton describing the exact set of inputs $\inputtype_{\textit{bad}}\subseteq P$ for which $T_{\textit{bad}}$ does not produce an output in $Q$ (see pre-image computation in~\cite{LorisCAV17}).
Let $restrict(T,L)$ be the transducer that behaves as $T$ if the input is in $L$ and does not produce an output otherwise (transducers are closed under restriction~\cite{LorisCAV17}).
If we synthesize a transducer $T_1$ with type $\hoaretriple{\inputtype_{\textit{bad}}}{T_1}{Q}$, then the transducer $restrict(T_1,\inputtype_{\textit{bad}})\cup restrict(T_{\textit{bad}},P\setminus\inputtype_{\textit{bad}})$ satisfies the desired input-output type (transducers are closed under union).
\mypar{Fault Localization from Examples.}
We use this technique when $T_{\textit{bad}}$ is incorrect on some example.
We can compute a set of ``suspicious'' transitions by taking all the transitions traversed when $T(s) \neq t$ for some ${\inputexample \mapsto \outputexample}\in E$ (i.e., one of these transitions is wrong) and removing all the transitions traversed when $T(s) = t$ for some ${\inputexample \mapsto \outputexample}\in E$ (i.e., transitions that are likely correct).
Essentially, this is a way of identifying $\inputtype_{\textit{bad}}$ when $T_{\textit{bad}}$ is wrong on some examples.
We can also use this technique to limit the transitions we need to synthesize when performing repair.
\section{Evaluation}
\label{sec:eval}
We implemented our technique in a Java tool \textsc{astra}\xspace (Automatic Synthesis of TRAnsducers), which uses Z3 \cite{z3_2008} to solve the generated constraints.
We evaluate using a 2.7 GHz Intel Core i5, RAM 8 GB, with a 300s timeout.
\subsection*{Q1: Can \textsc{astra}\xspace synthesize practical transformations?}
\label{sec:eval-synthesis}
\textit{Benchmarks.}
Our first set of benchmarks is obtained from Optician~\cite{symmetriclenses,bijectivelenses}, a tool for synthesizing lenses, which are bidirectional programs used for keeping files in different data formats synchronized.
We adapted 11 of these benchmarks to work with \textsc{astra}\xspace (note that we only synthesize one-directional transformations), and added one additional benchmark extrAcronym2, which is a harder variation (with a larger input type) of extrAcronym.
We excluded benchmarks that require some memory, e.g., swapping words in a sentence, as they cannot be modeled with transducers.
Our second set of benchmarks (Miscellaneous) consists of 6 problems we created based on file transformation tasks (unixToDos, dosToUnix and CSVSeparator), and s-FTs from the literature--escapeQuotes from \cite{bekonline}, getTags and quicktimeMerger from \cite{POPL12SFT}.
All of the benchmarks require synthesizing s-FTs and getTags requires synthesizing an s-FT with lookahead (details in Table~\ref{table:synthbenchmarks}).
To generate the sets of examples, we started with the examples that were used in the original source when available.
In 5 cases, \textsc{astra}\xspace synthesized a transducer that was not equivalent to the one synthesized by Optician, even though it did meet the specification.
In these cases, we used \textsc{astra}\xspace to synthesize two different transducers that met the specification, computed a string on which the two transducers differed, and added the desired output for that string as a new example.
We repeated this task until \textsc{astra}\xspace yielded the desired transducer and we report the time for such sets of examples.
The number of examples varies between 1 and 5.
The ability to perform equivalence checks of two transducers is yet another reason why synthesizing transducers is useful and is in some ways preferable to synthesizing programs in a DSL.
For each benchmark we chose a mean edit distance of 0.5 when the transformation could be synthesized with this distance and of 1 otherwise.
\begin{table*}[!htbp]\centering
\caption{Metrics of \textsc{astra}\xspace's performance on the set of synthesis benchmarks. The right-most set of columns gives the synthesis time for \textsc{astra}\xspace and Optician (under 2 different configurations).
The middle set of columns gives the sizes of the parameters to the synthesis problem.
In particular, $\setsrcstate$ and $\settrgstate$ denote the number of input and output states and $\transition{P}$ and $\transition{Q}$ denote the number of transitions in the input and output types, respectively.
A \xmark represents a benchmark that failed.
---\ stands in for data that is not available; this is because we only re-ran Optician on the benchmarks that were already encoded in its benchmark set, plus a few additional ones for comparing between the tools that we wrote ourselves.}
\label{table:synthbenchmarks}
\small
\begin{tabular}{c|l|rrrrrrrrr|rrr}\toprule
\multirow{2}{*}[-0.4ex]{} & & & & & & & & & & & & \quad Time(s) & \\
& Benchmark & $\setsrcstate$ & $\settrgstate$ &
$\transition{P}$ & $\transition{Q}$ & $\Sigma$ & $E$ &$k$ &$l$ & $d$ &\textsc{astra}\xspace &Optician & Optician-re \\\midrule
\multirow{12}{*}[-0.4ex]{\rotatebox{90}{Optician}} & extrAcronym &6 &3 &10 &3 &3 &2 &1 &1 & .5 &0.11 &0.05 & \xmark \\
& extrAcronym2 &6 &3 &16 &3 &3 &3 &2 &1 & 1 &0.42 & --- & --- \\
& extrNum &15 &13 &17 &12 &3 &1 &1 &1 & 1 &0.93 &0.05 &0.07 \\
& extrQuant &4 &3 &8 &5 &2 &1 &2 &1 & 1 &0.19 &0.09 & \xmark \\
& normalizeSpaces &7 &6 &19 &10 &2 &2 &2 &1 & 1 &0.46 &16.64 &\xmark \\
& extrOdds &15 &9 &29 &13 &5 &3 &3 &2 & 1 &15.87 &0.12 &\xmark \\
& capProb &3 &3 &3 &3 &2 &2 &2 &1 & 1 &0.05 &0.05 &\xmark \\
& removeLast &6 &3 &8 &3 &3 &3 &2 &1 & .5 &0.21 &0.15 &0.07 \\
& sourceToViews &18 &7 &26 &15 &5 &3 &3 &2 & 1 &50.92 &0.06 &\xmark \\
& normalizeNamePos &19 &7 &35 &24 &13 & 1 & 6 & 2 & 1 & \xmark &0.05 &0.10 \\
& titleConverter &22 &13 &41 &41 &15 & 1 & 3 & 1 & 1 & \xmark &0.07 &\xmark \\
& bibtextToReadable &14 &11 &41 &35 &12 & 1 & 5 & 1 & 1 & \xmark &0.64 &0.15 \\
\hline
\multirow{6}{*}[-0.4ex]{\rotatebox{90}{Miscellaneous}} & unixToDos &5 &7 &17 &19 &4 &4 &2 &2 & .5 &1.24 & --- & --- \\
& dosToUnix &7 &5 &19 &17 &4 &4 &2 &1 & .5 &0.41 & --- & --- \\
& CSVSeparator &5 &5 &9 &9 &4 &1 &1 &1 & 1 &0.142 & --- & --- \\
& escapeQuotes &2 &2 &6 &5 &3 &5 &2 &2 & 1 &0.188 & \xmark & \xmark \\
& quicktimeMerger &7 &3 &9 &3 &2 &2 &1 &1 & .5 &0.075 & --- & --- \\
& getTags & 3& 3& 9& 4& 3& 5& 2& 2& 1& 0.95& \xmark & \xmark \\
\bottomrule
\end{tabular}
\end{table*}
\mypar{Effectiveness of \textsc{astra}\xspace.}
\textsc{astra}\xspace can solve $15/18$ benchmarks (13 in <1s and 2 under a minute) and times out on 3 benchmarks where both $P$ and $Q$ are big.
While the synthesized transducers have at most 3 states, we note that this is because \textsc{astra}\xspace synthesizes total transducers and then restricts their domains to the input type $P$.
This is advantageous because synthesizing small total transducers is easier than synthesizing transducers that require more states to define the domain.
For instance, when we restrict the solution of extrAcronym2 to its input type, the resulting transducer has 11 states instead of the 2 required by the original solution!
\mypar{Comparison with Optician.}
We do not compare \textsc{astra}\xspace to tools that only support input-output examples.
Instead, we compare \textsc{astra}\xspace to Optician on the set of benchmarks common to both tools.
Like \textsc{astra}\xspace, Optician supports input-output examples and types, but the types are expressed as regular expressions.
Furthermore, Optician also attempts to produce a program that minimizes a fixed information theoretical distance between the input and output types~\cite{symmetriclenses}.
Optician is faster when the number of variables in the constraint encoding increases, while \textsc{astra}\xspace is faster on the normalizeSpaces benchmark.
Optician, which uses regular expressions to express the input and output types, does not work so well with unstructured data.
To confirm this trend, we wrote synthesis tasks for the escapeQuotes and getTags benchmarks in Optician and it was unable to synthesize those as well---e.g., escapeQuotes requires replacing every \texttt{"} character with \texttt{\textbackslash "}.
To further look at the reliance of Optician on regular expressions, we converted the regular expressions used in the lens synthesis benchmarks to automata and then back to regular expressions using a variant of the state elimination algorithm that acts on character intervals.
This results in regular expressions that are not very concise and might have redundancies.
Optician could only solve 4/11 benchmarks that it was previously synthesizing (Optician-re in Table~\ref{table:synthbenchmarks}).
\textbf{Answer to Q1:} \textsc{astra}\xspace can solve real-world benchmarks and has performance comparable to that of Optician for similar tasks. Unlike Optician, \textsc{astra}\xspace does not suffer from variations in how the input and output types are specified.
\subsection*{Q2: Can \textsc{astra}\xspace repair transducers in practice?}
\label{sec:eval-repair}
\textit{Benchmarks.}
We considered the benchmarks in Table~\ref{table:repairbenchmarks}.
The only pre-existing benchmark that we found was escapeQuotes, through the interface of the Bek programming language used for verifying transducers \cite{bekonline}.
We generated 11 additional faulty transducers to repair in the following two ways:
(\emph{i})~ Introducing faults in our synthesis benchmarks: We either replaced the output string of a transition with a constant character, inserted an extra character, or deleted a transition altogether.
(\emph{ii})~ Incorrect transducers: We intentionally provided fewer input-output examples and used only example-based constraints on some of our synthesis benchmarks.
All the benchmarks involve s-FTs.
Three benchmarks are wrong on both input-output types and examples and the rest are only wrong on examples.
Additionally, we note that to repair a transducer, we need the ``right'' set of minterms.
Typically, the set of minterms extracted from the transducer predicates is the right one, but in the case of the escapeBrackets problems, \textsc{astra}\xspace needs a set of custom minterms we provide manually---i.e., repairing the transducer requires coming up with a new predicate.
We are not aware of another tool that solves transducer repair problems and so do not show any comparisons.
\begin{table*}[htbp!]\centering
\caption{Metrics of \textsc{astra}\xspace's performance on the set of repair benchmarks. The two right-most columns give the synthesis time without and with the use of templates.
Default is the case where a new transducer is synthesized for $\inputtype_{\textit{bad}}$ and Template is the case where a partial solution to the solver is provided.
The $\transition{T_{\textit{bad}}}$ column gives the number of transitions that were localized by the fault-localization procedure as a fraction of the total number of transitions in the transducer.
The other columns that describe the parameters of the synthesis problem in the default case are the same as for Table~\ref{table:synthbenchmarks}. }
\label{table:repairbenchmarks}
\small
\begin{tabular}{c|l|rrrrrrrrrr|rr}\toprule
\multirow{2}{*}[-0.4ex]{} & & & & & & & & & & & & {Time(s)} & \\
& Benchmark & $\setsrcstate$ & $\settrgstate$ &
$\transition{P}$ & $\transition{Q}$ & $\Sigma$ & $E$ &$k$ &$l$ & $d$ & $\transition{T_{\textit{bad}}}$ &Default &Template \\\midrule
\multirow{7}{*}[-0.4ex]{\rotatebox{90}{Fault injected}} & swapCase1 &2 &1 &6 &3 &3 &2 &1 &1 & 1 &3/3 &0.04 &0.02 \\
& swapCase2 &2 &1 &4 &3 &3 &2 &1 &1 &1 & 1/2 & \xmark & \xmark \\
& swapCase3 &2 &1 &6 &3 &3 &2 &1 &1 & 1 &1/3 &0.06 &0.05 \\
& escapeBrackets1 &2 &6 &16 &36 &8 &4 &1 &4 & 4 &1/3 &0.69 &0.42 \\
& escapeBrackets2 & 1& 6& 1& 7& 6& 5& 1& 4& 4 & 1/2 & \xmark & \xmark \\
& escapeBrackets3 &2 &7 &8 &36 &9 &5 &1 &4 & 4 &2/3 &1.12 &0.34 \\
& caesarCipher &2 &1 &4 &2 &3 &1 &1 &1 &1 &1/1 & \xmark & \xmark \\
\hline
\multirow{4}{*}[-0.4ex]{\rotatebox{90}{Synth.}} & extrAcronym2 &11 &3 &30 &3 &3 &3 &2 &1 & 1 &12/30 &0.59 &10.15 \\
& capProb &3 &3 &3 &3 &2 &2 &2 &1 & 1 &3/3 &0.04 &0.04 \\
& extrQuant &8 &3 &16 &5 &2 &1 &2 &1 & 1 &5/10 &0.37 &0.51 \\
& removeLast &6 &3 &8 &3 &3 &2 &2 &1 & .5 &7/8 &0.40 &1.08 \\
\hline
& escapeQuotes & 3& 2& 9& 5& 3 & 5 & 2 & 1 & 1 & 3/5& 0.17 & 0.10\\
\bottomrule
\end{tabular}
\end{table*}
\mypar{Effectiveness of \textsc{astra}\xspace.}
We indicate the number of suspicious transitions identified by our fault localization procedure (Section~\ref{sec:repair}) in the column labeled $\transition{T_{\textit{bad}}}$.
In many cases, \textsc{astra}\xspace can detect 50\% of the transitions or more as being likely correct, therefore reducing the space of unknowns.
We compare 2 different ways of solving repair problems in \textsc{astra}\xspace.
One uses the repair-from-input approach described in Section~\ref{sec:repair} (Default in Table~\ref{table:repairbenchmarks}).
The second approach involves using a `template', where we supply the constraint solver with a partial solution to the synthesis problem, based on the transitions that were localized as potentially buggy (Template in Table~\ref{table:repairbenchmarks}).
\textsc{astra}\xspace can solve 9/$12$ repair benchmarks (all in less than 1 second).
The times using either approach are comparable in most cases.
While one might expect templates to be faster, this is not always the case because the input-output specification for the repair transducer is small, but providing a template requires actually providing a partial solution, which in some cases happens to involve many constraints.
\textbf{Answer to Q2:} \textsc{astra}\xspace can repair transducers with varying types of bugs.
\section{Related Work}
\label{sec:related_work}
\textit{Synthesis of string transformations.}
String transformations are one of the main targets of program synthesis.
Gulwani showed they could be synthesized from input-output examples \cite{gulwani2011flashfill} and introduced the idea of using a DSL to aid synthesis.
Optician extended the DSL-based idea to synthesizing lenses \cite{bijectivelenses,symmetriclenses}, which are programs that transform between two formats.
Optician supports not only examples but also input-output types.
While DSL-based approaches provide good performance, they are also monolithic as they rely on the structure of the DSL to search efficiently.
\textsc{astra}\xspace does not rely on a DSL and can synthesize string transformations from complex specifications that cannot be handled by DSL-based tools.
Moreover, transducers allow applying verification techniques to the synthesized programs (e.g., checking whether two solutions are equivalent).
One limitation of transducers is that they do not have `memory', and consequently \textsc{astra}\xspace cannot be used for data-transformation tasks where this is required---e.g., mapping the string \texttt{Firstname Lastname} to \texttt{Lastname, Firstname}---something Optician can do.
We remark that there exist transducer models with such capabilities~\cite{streaming-string-transducers} and our work lays the foundations to handle complex models in the future.
\vspace{1mm}\noindent\textit{Synthesis of transducers.}
Benedikt et al. studied the `bounded repair problem', where the goal is to determine whether there exists a transducer that maps strings from an input to an output type using a bounded number of edits~\cite{Benedikt11}.
Their work was the first to identify the relation between solving such a problem and solving games, an idea we leverage in this paper.
However, their work is not implemented, cannot handle input-output examples, and therefore shies away from the source of NP-Completeness.
Hamza et al. studied the problem of synthesizing minimal non-deterministic Mealy machines (transducers where every transition outputs exactly one character), from examples \cite{minimaltransducersynthesis}.
They prove that the problem of synthesizing such transducers is NP-complete and provide an algorithm for computing minimal Mealy machines that are consistent with the input-output examples.
\textsc{astra}\xspace is a more general framework that incorporates new specification mechanisms, e.g., input-output types and distances, and uses them all together.
Mealy machines are also synthesized from temporal specifications in reactive synthesis and regular model checking, where they are used to represent parameterized systems~\cite{safety-parametrized-synthesis,lin2016liveness}.
This setting is orthogonal to ours as the specification is different and the transducer is again only a Mealy machine.
The constraint encoding used in \textsc{astra}\xspace is inspired by the encoding presented by Daniel Neider for computing minimal separating DFA, i.e. a DFA that separates two disjoint regular languages \cite{neider2012}.
\textsc{astra}\xspace's use of weights and energy to specify a mean edit distance is based on energy games \cite{DBLP:conf/emsoft/ChakrabartiAHS03}, a kind of 2-player infinite game that captures the need for a player to not exceed some available resource.
One way of solving such games is by defining a \emph{progress measure} \cite{progressmeasure}.
To determine whether a game has a winning strategy for one of the players, it can be checked whether such a progress measure exists in the game.
We showed that the search for such a progress measure can be encoded as an SMT problem.
\section{Conclusion}
We presented a technique and a tool (\textsc{astra}\xspace) for synthesizing different types of transducers from types, examples, and distance functions, and showed \textsc{astra}\xspace's ability to solve a variety of practical problems.
\textsc{astra}\xspace uses SMT solvers and its performance is affected by input components that result in large constraints (e.g., states in the desired transducer).
Because \textsc{astra}\xspace synthesizes transducers instead of programs in arbitrary DSLs, its output can be analyzed using transducer algorithms (e.g., equivalence and pre-post analysis).
Because of this property, our approach could be beneficial in learning invariants of string-manipulating programs, where a transducer is the formalism of choice, e.g., in the Ostrich tool \cite{ostrichATVA2020}.
In terms of improvements to our technique, aside from optimizing the SMT encoding to improve scalability, our approach could be strengthened by devising ways to effectively `guess' the number of states required for a transducer to work on the given inputs.
We leave these directions for future work.
\section*{Acknowledgements}
This work was funded by the National Science Foundation under grants 1763871, 1750965, 1918211, and 2023222, Facebook and a Microsoft Research Faculty Fellowship.
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{sec:intro}
Space-based Large Scale Structure (LSS) spectroscopic surveys made possible by the Roman Space Telescope \cite{spergel2013widefield} and Euclid \cite{laureijs2011euclid} missions will soon be available. These surveys will make use of slitless spectroscopy to measure galaxy redshifts meaning that, while they have the potential to constrain the cosmological parameters with sub percent levels of accuracy, this requires a thorough understanding of potential sources of systematic error arising from the slitless method. One possible systematic effect, that we study in this paper, is the contamination of the sample by interlopers resulting from line confusion.
To estimate the redshifts of a set of galaxies using slitless spectroscopy, we measure the wavelengths of one or more emission lines for each. In this paper we will primarily be concerned with samples where only one emission line is used to estimate the redshift. Generally, emission lines with different rest-frame wavelengths $\lambda_1$ and $\lambda_2$, from two different galaxies with true redshifts $z_1 = \lambda_{\rm obs} / \lambda_{1}-1$ and $z_2 = \lambda_{\rm obs} / \lambda_{\rm 2}-1$, can appear at the same observed wavelength, $\lambda_{\rm obs.}$. The measured redshift for a galaxy given only the observed position of this single line will depend on the rest-frame wavelength assumed. In the presence of a secondary emission line, one can distinguish between different options. However, if we are measuring redshifts from a single emission line, we are prone to line misidentification, if we assume, for a sample of galaxies, that the observed line has the same rest-frame wavelength for all galaxies. The galaxies in the sample for which the lines are misidentified resulting in them being assigned the wrong redshift, are called interlopers. We refer to the galaxies whose observed emission line matches the assumed rest-frame wavelength when measuring redshifts as the target galaxies.
For Roman, for example, it is likely that many redshifts will be measured using 2 or more emission lines \cite{wang_high_2022}. At $2<z<3$, [O{\sc\,iii}]\ is the primary line for Roman observations, which is actually a doublet. Resolving this doublet as two separate lines, which is possible at given the resolution of Roman, or observing a secondary line from [O{\sc\,ii}]\ emission would allow a secure redshift to be measured. However, to provide a bonus sample to higher number densities, it will be useful to analyse galaxies where only one unresolved line is observed, assumed to be [O{\sc\,iii}], but where this identification is less secure. For this, we will have interlopers, where the H$\beta$ line is mistaken for [O{\sc\,iii}]. The increased size of this bonus sample will help reduce the statistical errors on measured cosmological parameters, at the cost of increasing the systematic errors. It is these systematic errors, and methods to mitigate them, that we consider in this paper.
In general, there are three different ways that a catalogue can be contaminated:
\begin{enumerate}
\item First, due to noise spikes being misidentified as emission lines. The noise is uncorrelated with all the other galaxies, and has a distribution different from that of the true observed lines. Thus, this population of contaminants Poisson samples a volume that has a window different from that of the target galaxy population. It can trivially be corrected in a clustering measurement as it is absorbed into (and is cancelled by) part of the expected density (often quantified by a random catalog) and simply adds to the shot noise and changes the amplitude of the signal measured.
\item Second, misidentified emission lines can arise from interloper galaxies displaced by a large radial distance. In this case, correlations with the target galaxy population are small and we only need to model the auto-correlation of the interloper sample (e.g. \cite{Farrow_2021,Grasshorn_Gebhardt_2019,Gong_2021,pullen_interloper_2016}).
Because the displacement is large, there is a significant change in $H(z)$ between the true redshift of the interlopers and the incorrect redshift inferred, such that the Baryon Acoustic Oscillation (BAO) signal in the auto-correlation of the interlopers is shifted with respect to its true value. This needs to be included in any analysis but, as the shift is known from the line emission wavelengths, this contribution to the clustering of the contaminated sample can be easily modeled provided the fraction of interlopers is known.
\item The third class of contaminant is a more pernicious type of interloper, which is one only shifted by a small radial displacement by the incorrect rest-frame wavelength assignment. Such interlopers will be correlated with other such objects {\em and} with the target galaxy sample. This type of interloper can occur for samples obtained using the Roman Space Telescope at high redshifts, where the target is the primary [O{\sc\,iii}]\ line, and $\rm H \beta$ line is the source of interlopers. The impact of these interlopers on the BAO peak position was studied by \cite{massara2020line}. They showed that misidentifying $\rm H \beta$ line as [O{\sc\,iii}]\ line in an Emission Line galaxy will lead to underestimating its distance by $\sim 90\,h^{-1}\,{\rm Mpc}$. This would distort the shape of the galaxy correlation function and therefore introduce a shift in the BAO peak position. We call this class of interlopers, which are displaced by $\lesssim 150 \,h^{-1}\,{\rm Mpc}$, "small-displacement" interlopers (see Section~\ref{sec:results}).
\end{enumerate}
We focus on the third class of contaminant. Following \cite{massara2020line}, we present a model for the monopole and the quadrupole moments of the correlation function of a Roman-like catalogue, with an [O{\sc\,iii}]\ target line sample contaminated by $\rm H\beta$ interlopers. With this model, we investigate how to find an unbiased estimation of the isotropic and anisotropic dilation parameters $\alpha$ and $\epsilon$ from the BAO signal, as well as the fraction of interlopers $f_{\rm i}$. Our method is based on standard BAO modeling, additionally allowing for the interlopers. The model correlation function has two terms: The galaxy auto-correlation that we estimate using CAMB \cite{lewis_efficient_2000}, and a non-negligible cross-correlation between the galaxies and the interlopers that we consider a number of ways of estimating. If we can measure this from the contaminated auto-correlation, or use another way of accurately calibrating this statistic, our pipeline gives an unbiased estimation of all the three parameters. If we do not take into account the interlopers, the measurements would be highly biased.
The plan of our paper is as follows:
In Section~\ref{subsec:displacement}, we derive the interloper shift value.
In Section~\ref{sec:simulations}, we describe simulations that were used for our analysis.
In Section~\ref{sec:Modelling_the_contaminated_correlation_function}, we present our model for the correlation function of a catalogue contaminated by some small-displacement interlopers. We show the outcomes of our pipeline in Section~\ref{sec:results}. We discuss our results and their implications in Section~\ref{sec:discussions}.
\begin{figure}[t]
\centering
\includegraphics[width=1\textwidth,angle=0]{plots/displacement.png}
\caption{
Displacement of interlopers as a function of their true redshift as shown in Eq. \ref{eq:displacement}.
}
\label{fig:displacement}
\end{figure}
\section{Displacement}\label{subsec:displacement}
Let us assume that in a survey, potential interlopers are at redshift $z^{\rm true}$, and they emit a primary emission line with rest-frame wavelength $\lambda_{\rm emit}^{\rm true}$, that is observed at wavelength $\lambda_{\rm obs}$. The observed wavelength can be misidentified for another photon as if it was emitted with rest-frame wavelength $\lambda_{\rm emit}^{\rm false}$ at redshift $z^{\rm false}$. This redshift misidentification would effectively lead us to a wrong estimation of the comoving proper distance to the galaxies by $\Delta d$:
\begin{subequations}
\label{eq:y}
\begin{align}
\label{eq:y:1}
\Delta d =
d^{\rm true} - d^{\rm false}
= \int_{z^{\rm false}}^{z^{\rm true}} \frac{cdz}{H(z)}
\\
\label{eq:y:2}
\approx
\frac{c}{H(z^{\rm true})} (z^{\rm true} -z^{\rm false})\,.
\\
\label{eq:y:3}
\approx
\frac{c}{H(z^{\rm true})} \left[1 - \frac{\lambda^{\rm true}_{\rm emit}}{\lambda^{\rm false}_{\rm emit}} \right] (1 + z^{\rm true})\,.
\end{align}
\end{subequations}
In this paper we consider $\rm H \beta$ interlopers misidentified as [O{\sc\,iii}]\ emitters. While [O{\sc\,iii}]\ is a doublet, here we assume that $\rm H \beta$ is misidentified as the primary line in the doublet at $\lambda^{\rm OIII}_{\rm false} = 500.7 \, \rm nm$, rather than the secondary line at $\lambda^{\rm OIII}_{\rm false} = 495.9 \, \rm nm$. The rest-frame wavelength of $\rm H \beta$ is $\lambda^{\rm H\beta}_{\rm true} = 486.1 \,\rm nm$, leading to an offset between the true and wrongly inferred interloper comoving position equal to
\begin{equation}
\label{eq:displacement}
\Delta d \approx 87.41 \frac{1+z^{\rm true}}{\sqrt{\Omega_{\Lambda} + \Omega_{\rm m} (1+z^{\rm true})^3}}\, \,h^{-1}\,{\rm Mpc}.
\end{equation}
Figure~\ref{fig:displacement} displays the distance $\Delta d$ as a function of the true galaxy redshift and with the parameters $\Omega_\Lambda$ and $\Omega_{\rm m}$ evaluated at the cosmology of the simulations used in this paper and described in Section~\ref{sec:simulations}. In that cosmology, the displacement $\Delta d$ is equal to $97 \,h^{-1}\,{\rm Mpc}$ at $z = 1$ (the redshift of the N-body snapshots considered in this work), and $90, 85\,h^{-1}\,{\rm Mpc}$ at $z = 1.7, 2.1$, which are the redshifts of interest for the Roman Space Telescope.
Note that measurements stemming from rest-frame wavelength differences cannot be used as standard rulers in the same way as BAO - what is fixed is the difference in redshift, not the distance. The displacement in physical units depends on the assumed fiducial cosmology used to convert redshifts to distances not the true cosmology, and we do not have to vary the displacement when fitting different cosmological models to the data.
\begin{figure}[ht]
\centering
\includegraphics[width=1\textwidth,angle=0]{plots/monopole_all_fractions.png}
\caption{
Monopole of the correlation function as calculated from the mean of 1000 mocks that are contaminated by different fractions of interlopers (different colours), and different interloper displacements: $85$ (top row), $90$ (middle row), and $97\,h^{-1}\,{\rm Mpc}$ (bottom row).
As can be seen, the BAO peak is skewed towards $\Delta d$, and amplified more as the number of interlopers increase.
}
\label{fig:mono_general}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=1\textwidth,angle=0]{plots/quadrupole_all_fractions.png}
\caption{
Quadrupole of the correlation function as calculated from the mean of 1000 mocks that are contaminated by different fractions of interlopers (different colours), and different interloper displacements: $85$ (top row), $90$ (middle row), and $97\,h^{-1}\,{\rm Mpc}$ (bottom row).
As can be seen, the quadrupole peaks at $\Delta d$, which demonstrates the anisotropy caused by the shift of interlopers along the LOS.
}
\label{fig:quad_general}
\end{figure}
\section{Simulations}
\label{sec:simulations}
In this study, our focus is on building and testing a model framework that describes the effect of interloper galaxies rather than predicting the actual expected fraction of interlopers.
While some progress have been made in modelling the population of [O{\sc\,iii}] \ emitters (e.g., \cite{Zhai_2019}),
we need a thorough comparison with observations to know if we have the accuracy required. The population of $\rm H \beta$ emitters is less well understood.
Therefore, we will use halos in place of galaxy catalogs to test our model, for simplicity. We employ the halo catalogs from $1,000$ N-body simulations of the Quijote suite \cite{Quijote_2020}. These simulations have been run in a box with size $1 h^{-1}$Gpc and follow the evolution of $512^3$ cold dark matter particles in the fiducial cosmology of the suite, which is a flat $\rm \Lambda CDM$ cosmology with parameters $\Omega_m = 0.3175$, $\Omega_b = 0.049$, $n_s = 0.9624$, $h = 0.6711$, $\sigma_8 = 0.834$, $M_\nu = 0.0$ eV. Halos have been identified using the Friends-of-Friends algorithm with linking length parameter equal to 0.2, and only those with 20 or more cold dark matter particles have been stored in the catalogue. This means that considered halos have minimum mass equal to $1.31 \cdot 10^{13}h^{-1}$M$_{\odot}$.
Even though [O{\sc\,iii}]\ emission galaxies are the primary target at redshifts beyond $z=1.8$ in the High-Latitude Survey of the Roman telescope, we use halo catalogs at redshift $z=1$ to have a larger number density of objects, $\bar{n}_h = 2\cdot 10^{-4}h^3{\rm Mpc}^{-3}$, consistent with the expected [O{\sc\,iii}]\ mean number density in Roman between redshifts $1.8-2.8$ \cite{Zhai_2019,massara2020line}. At the redshift of the simulation, interlopers appear to be displaced by $97 \,h^{-1}\,{\rm Mpc}$ along the line of sight. To simulate interlopers at higher redshift, we also consider displacements equal $90$ or $85 \,h^{-1}\,{\rm Mpc}$ that correspond to catalogs at redshifts $z=1.7$ and $z=2.1$, respectively. We build the contaminated halo catalogs by randomly selecting objects that will be interlopers and displacing them at the wrongly inferred distance along the $z$ direction, that we assume to be the line-of-sight (hereafter, LOS). We implement four different percentages of interlopers (2, 5, 10, 15 \%) and three different values for the displacement, obtaining 12 different contaminated halo catalogs for each of the initial uncontaminated ones.
We measure the correlation function from all these mocks, by making use of Nbodykit \cite{Nbodykit_Hand_2018} adopting the Landy-Szalay estimator \cite{landy_szalay_1993}.
To build a better understanding of the problem, the monopole and quadrupole of this measured correlation function are shown in Figure~\ref{fig:mono_general} and Figure~\ref{fig:quad_general}, respectively.
In these plots, we show catalogues contaminated by different fractions of interlopers in different colours, and with different displacements in different panels: $\Delta d = 85$ (top), $90$ (middle), and $97 \,h^{-1}\,{\rm Mpc}$ (bottom).
The effects of the interlopers are clear to see in both moments. The amplitude of the small-scale monopole is suppressed by the presence of interlopers, and the suppression increases with the fraction of interlopers. The missing small-scales correlation is converted into larger scales, and in particular it enhances the amplitude of the correlation around the displacement scale, which is close to the BAO peak position. As a result, the BAO peak appears to be enhanced, broadened and shifted towards smaller separations. The quadupole is also modified by the interlopers: Its amplitude is enhanced on all scales with increasing interloper fraction. Moreover, it also exhibits a peak around the scales corresponding to the displacement of the interlopers, as a consequence of the increased number of interloper-target galaxy pairs along the LOS.
\section{Modelling the Contaminated Correlation Function}
\label{sec:Modelling_the_contaminated_correlation_function}
In this Section, we show how to model the monopole and quadrupole moments of the correlation function for a catalogue that is contaminated by interlopers misplaced by relatively small displacements such that the interloper-galaxy cross-correlation term is not negligible. We build a model for the auto-correlation function of a contaminated catalogue in Section~\ref{subsec:correlation_function}, and in Section~\ref{subsec:cross_correlation} we introduce our method to estimate the interloper-target cross-correlation function. Finally, in Section~\ref{subsec:Model_and_Parameters}, we discuss our completed model.
\subsection{Correlation Function with Interlopers}\label{subsec:correlation_function}
We consider a catalogue contaminated by a given fraction of interlopers, $f_{\rm i}$, and define $\xi_{\rm gg}$ the auto-correlation function of the target galaxies, $\xi_{\rm gi}$ the cross-correlation between these galaxies and interlopers and $\xi_{\rm ii}$ the auto-correlation function of the interlopers. Throughout this paper, indices "$\rm i$", and "$\rm g$", stand for interlopers, and galaxies respectively.
We denote the observed comoving distance of an object with $\Vec{x}$: for target galaxies, this is the same as their true position, whereas for interlopers it is the wrong measured position. We denote the true position of an interloper with $\Vec{y}$. Following \cite{Pullen2015}, the contaminated galaxy overdensity is given by
\begin{equation}
\label{eq:total_overdensity}
\delta (\Vec{x}) =
(1 - f_{\rm i}) \delta_{\rm g} (\Vec{x}) + f_{\rm i} \delta_{\rm i}(\Vec{x})\,.
\end{equation}
We can use this equation to calculate the contaminated correlation function at any observed separation $\Vec{r} = \Vec{x}_1 - \Vec{x}_2$,
\begin{equation}
\label{eq:xi_tt_general}
\begin{split}
\xi(\Vec{x}_1-\Vec{x}_2; f_{\rm i}) &=
\langle
\delta(\Vec{x}_1) \delta^{*}(\Vec{x}_2)
\rangle
\\
&=\langle
\left[(1 - f_{\rm i}) \delta_{\rm g} (\Vec{x}_1) + f_{\rm i} \delta_{\rm i}(\Vec{x}_1)
\right]
\left[
(1 - f_{\rm i}) \delta^{*}_{\rm g} (\Vec{x}_2) + f_{\rm i} \delta^{*}_{\rm i}(\Vec{x}_2)
\right]
\rangle\\
&= (1-f_{\rm i})^2\xi_{\rm gg}(\Vec{r})+f_{\rm i}^2\xi_{\rm ii}(\Vec{r})+2f_{\rm i}(1-f_{\rm i})\xi_{\rm gi}(\Vec{r})\,.
\end{split}
\end{equation}
This equation can be significantly simplified if we can make two key assumptions.
First, we assume that the displacement $\Delta d$ in Eq.~\ref{eq:displacement} between the true position of interlopers and the wrongly inferred one is a constant within a redshift bin when calculating the correlation function at a fixed cosmology. Indeed, we only consider correlation functions at separations $r$ smaller than $150\,h^{-1}\,{\rm Mpc}$ in the following analysis and in this range the displacement only varies by at most 0.8\% (see Figure~\ref{fig:displacement}). Therefore, we can assume that each object in the pair is displaced by the same amount. Moreover, it is commonly assumed that within a redshift bin $H(z)$ is a constant, therefore, we can assume that the displacement is fixed across the redshift bin considered for all the pairs.
Thus, the wrongly inferred position of each interloper can be written in terms of a constant displacement along the LOS $\hat{z}$: $\vec{x}_{\rm i} = \vec{y}_{\rm i} - \Delta d \, \hat{z}$, where $\vec{y}_{\rm i}$ is the true position.
The interloper auto-correlation function can then be written as $\xi_{\rm ii}(\Vec{x}_1 - \Vec{x}_2) = \xi_{\rm ii}(\Vec{y}_1 - \Vec{y}_2)$ (where we have assumed that the LOS is the same for both galaxies in a pair), because two point correlators are invariant under translations -- they depend only on the relative distance between two objects. This means that the correlation function of the interlopers at the wrong redshifts is equal to the correlation function of the interlopers at their true position.
Interlopers and targets belong to two different populations of galaxies and have in principle different bias schemes. Consequently, we either need calibration data to measure $\xi_{\rm gi}(\Vec{r})$ and $\xi_{\rm ii}(\Vec{r})$ (we discuss this particular case in Section~\ref{sec:discussions}), or we need a way to model this function. On the other hand, we note that since displacements are small, interlopers and targets both trace a matter field with similar amplitude of clustering. It is also likely that interlopers have similar properties and intrinsic clustering as the target galaxies, meaning that they share the same galaxy bias.
We thus make a second assumption: targets and interlopers share the same galaxy bias. In this case, together with the assumption of constant displacement for the interlopers, we have:
$\xi_{\rm ii}(\Vec{r}) = \xi_{\rm gg}(\Vec{r})$.
We assume that this is true in the rest of the paper. Our results and conclusions will be similar if this is not true, but instead we are able to accurately estimate or model the bias scheme of the interlopers. Substituting into Eq.~\ref{eq:xi_tt_general} yields
\begin{equation}
\label{eq:xi_tt}
\xi(\Vec{r};f_{\rm i}) = (1+2f_{\rm i}^2-2f_{\rm i})\xi_{\rm gg}(\Vec{r})+2f_{\rm i}(1-f_{\rm i})\xi_{\rm gi}(\Vec{r})\,.
\end{equation}
In the next Section, we will further consider methods to estimate the cross-correlation between interlopers and galaxies, $\xi_{\rm gi}(r)$.
\begin{figure}[t]
\centering
\includegraphics[width=1\textwidth,angle=0]{plots/compare_cross_norsd_97_frac10_n1000.png}
\caption{The importance of accurately modelling the small-scale cross-correlation in order to match the large-scale clustering around the BAO feature.
Monopole (top) and quadrupole (bottom) of the galaxy-interloper cross-correlation function correlation in real-space, calculated from mapping of auto-correlation described in Eq.~\ref{eq:cross_equation}. We considered four different auto-correlations described in more detail in text. The interloper displacement, $\Delta d$, used in mapping equals $97$ for $z = 1$.
}
\label{fig:cross_correlation}
\end{figure}
\subsection{Interloper-Target Cross-correlation}
\label{subsec:cross_correlation}
As mentioned before, the cross-correlation between galaxies and interlopers plays a key role in determining the shape of the contaminated correlation function.
While the galaxy-galaxy and interloper-interloper terms remain unchanged after misplacing interlopers along the LOS, the galaxy-interloper cross-correlation changes greatly. This change can easily be understood considering that $\xi_{\rm gi}$ at $\vec{r} = \Vec{x}_1-\Vec{x}_2$, which is the separation between the two objects after misplacing interlopers by $\Delta d$, should be equal to $\xi_{\rm gi}$ at the true separation $\vec{r}\,' = \Vec{x}_1-\Vec{y}_2$. Thus,
\begin{equation}
\label{eq:cross_equation}
\xi_{\rm gi}(\Vec{r})=\xi_{\rm g i}(\vec{r}\,') \,.
\end{equation}
There is a simple trigonometry transformation between $\vec{r}$ and $\vec{r}\,'$
\begin{equation}
\label{eq:polar_coordinates}
\begin{split}
r'(r,\mu) = \sqrt{r^2+\Delta d^2-2 \, \Delta d \, r\, \mu} \,,
\\
\mu'(r,\mu) = \frac{r\mu-\Delta d}{\sqrt{r^2+\Delta d^2-2 \, \Delta d \, r \, \mu}}\,.
\end{split}
\end{equation}
where we have defined $r= |\vec{r}|$ being the modulus of the separation vector and $\mu = {\vec{r}}\cdot\hat{z}/r$ being the cosine of the angle between the separation vector and the LOS.
The auto-correlation to cross-correlation mapping effectively pushes pair separations to larger scales. Nearby objects are more correlated than more distant objects -- the galaxy correlation function roughly decays by $r^2$, so the displacement of interlopers effectively moves the strong correlation between nearby objects around $<20\,h^{-1}\,{\rm Mpc}$ to larger scales in the galaxy-interloper cross-correlation, leaving a peak. The location of the center of this peak is at the interloper displacement. We show this in Figure~\ref{fig:cross_correlation} with blue points obtained by measuring the cross-correlation between galaxies and interlopers in a Quijote simulation box at $z=1$, after misplacing interlopers by $97\,h^{-1}\,{\rm Mpc}$.
In practice, if we have a small survey with higher signal-to-noise ratio (calibration data), it will be possible to estimate which galaxies would appear as interlopers in a survey with lower signal-to-noise ratio. Using calibration data we can directly measure the cross-correlation between galaxies and interlopers at their \emph{true} distance, and then infer the cross-correlation component at the wrong distance via Eq.~\ref{eq:polar_coordinates}.
If the calibration data is not available, there are two other possible approaches that we discuss in the following to estimate the cross-correlation assuming that the bias scheme of interlopers is the same as main target galaxies: analytical and observational approaches.
\emph{Analytical approach:} We considered taking the Fourier transform of the analytical power spectrum from CAMB linear and non-linear (from HALOFIT \cite{smith_stable_2003, takahashi_revising_2012}) to estimate the auto-correlation function.
The results from the linear CAMB model are shown in red in Figure~\ref{fig:cross_correlation}. The top and bottom panels of this plot show that the linear CAMB model roughly describes the shape of the measured monople and quadrupole of the cross-correlation, but it underestimates the peak height by $\sim 30\%$. This is because the prediction of CAMB for the monopole at small scales is significantly lower than the non-linear reality, and, as we discussed earlier, this directly impacts the height of the peak at $97 \,h^{-1}\,{\rm Mpc}$. We have also used HALOFIT to create a non-linear model. The corresponding results are shown in green in the same figure. The outcome remains the same with a slightly better prediction of the height. In this case, while the displacement depends on the fiducial cosmology and it determines the position of the peak in the target-interloper correlation function, the shape of the cross-correlation depends on the true cosmology since it is a remapping of the auto-correlation of target galaxies via Eq.~\ref{eq:Cartesian_to_polar}. This can be easily implemented in the model for the correlation function. The inability of this model to accurately fit the cross-correlation function led us to consider our second approach, or the observational approach.
\emph{Observational approach:}
If we assume that interlopers and targets are drawn from the same population of galaxies, we can estimate their cross-correlation directly from the measured auto-correlation of the contaminated catalogue As shown in Figure~\ref{fig:cross_correlation}, the monopole of cross-term vanishes on small scales ${r}<40\,h^{-1}\,{\rm Mpc}$. Therefore, one can rewrite Eq.~\ref{eq:xi_tt} as
\begin{equation}
\label{eq:xi_tt_approx}
\xi(\Vec{r};f_{\rm i}) \approx (1+2f_{\rm i}^2-2f_{\rm i})\,\xi_{\rm gg}(\Vec{r}) \quad {\rm for \, small \, r,}
\end{equation}
indicating that the contaminated and uncontaminated catalogs have small-scale correlation functions that differ by an overall amplitude only, and this amplitude depends on the fraction of interlopers in the contaminated catalogue We can therefore infer the galaxy-galaxy small-scale correlation function directly from the measured contaminated correlation function, modulo an overall normalization. The large-scale feature in the cross-correlation seen in Figure~\ref{fig:cross_correlation} is mainly determined by the small-scale correlation of galaxies and interlopers at the true position, or equivalently the galaxy auto-correlation since we assumed that galaxies and interlopers have the same bias (see Eq.~\ref{eq:cross_equation}). Therefore, using Eqs.~\ref{eq:xi_tt_approx} and~\ref{eq:cross_equation} we can estimate the cross-correlation term on the scales of interest ($r<150\,h^{-1}\,{\rm Mpc}$) directly from the measured small-scale contaminated correlation function as
\begin{equation}
\label{eq:cross_equation_approx}
\xi_{\rm gi}(\Vec{r})=\frac{\xi(\vec{r}\,';f_{\rm i})}{{(1+2f_{\rm i}^2-2f_{\rm i})}} \,,
\end{equation}
where the mapping between $\Vec{r}$ and $\vec{r}\,'$ is described in Eq.~\ref{eq:polar_coordinates}. Using this observational method, the model for the cross-correlation is already built from a correlation function in the true cosmology but measured in the fiducial one, and we will not need to account for differences between the two cosmologies when performing the fit.
Using this method to determine the cross-correlation, one can immediately see the improvement in the peak height estimation of the monopole and quadrupole moments of the cross-correlation function in Figure~\ref{fig:cross_correlation} (orange line). Note that this method assumes that interlopers and targets have the same bias scheme, and this assumption is satisfied exactly in our mocks.
On the other hand, the general shape of the monopole of the cross-correlation, especially at larger scales, deviates from that of the measurement. We believe that this is due to the approximation that we made to derive Eq.~\ref{eq:xi_tt_approx}; that of a vanishing cross-correlation on small scales. The quadrupole is less affected by this approximation. These effects do not alter the measured parameters (see Section~\ref{subsec:Model_and_Parameters}). Finally, using this approximation, we can use Eq.\ref{eq:cross_equation_approx} for the galaxy-interloper cross-correlation and rewrite Eq.~\ref{eq:xi_tt}
\begin{equation}
\label{eq:xi_final}
\xi(\Vec{r};f_{\rm i}) = (1+2f_{\rm i}^2-2f_{\rm i})\xi_{\rm gg}(\Vec{r})+\frac{2f_{\rm i}(1-f_{\rm i})}{(1+2f_{\rm i}^2-2f_{\rm i})}
\times \mathcal{M}[\xi(f_i)](\vec{r})\,,
\end{equation}
where for simplicity, we use $\mathcal{M}$ to denote the mapping of scales in Eq.~\ref{eq:polar_coordinates} that describes the deformation due to interlopers.
\subsection{Building a Cosmological Model}\label{subsec:Model_and_Parameters}
Thus far, we have introduced a model for the contaminated correlation function that consists of the galaxy-galaxy correlation function term, $\xi_{\rm gg}(\Vec{r})$, and the cross-correlation term, $\xi_{\rm gi}(\Vec{r};f_{\rm i})$.
There is still one part missing: How does our model depend on the assumptions about the fiducial cosmology? In this Section, we will address this question based on the methodology and notation of Ref.~\cite{Padmanabhan_White_2008} and \cite{Xu_2013}.
To measure the correlation function from data, we need to calculate the separation between pairs by assuming a fiducial cosmology to translate from redshifts to distances. The BAO peak in the correlation function can be distorted if the assumed cosmology is not the same as the true cosmology. This distortion can be
parameterized by Alcock–Paczynski parameters (hereafter AP; \cite{alcock_evolution_1979}): $\alpha$, the isotropic dilation parameter, and $\epsilon$, the anisotropic warping parameter:
\begin{equation}
\begin{split}
\alpha = \left[
\frac{D_{\rm A, \rm true}^2(z) H_{\rm fid}(z)}{D_{\rm A, fid}^2(z) H_{\rm true}(z)}
\right]^{1/3}
\frac{r_{\rm s,fid}}{r_{\rm s, true}}\,,\\
1 + \epsilon = \left[\frac{D_{\rm A, fid}(z) H_{\rm fid}(z)}{D_{\rm A, true}(z) H_{\rm true}(z)}\right]^{1/3}\,.
\end{split}
\end{equation}
And we can relate the fiducial and true distances using $\alpha$ and $\epsilon$ using the following equations:
\begin{equation}
\begin{split}
r_{\parallel, \rm true} = \alpha (1+\epsilon)^2 r_{\parallel}\,, \\
r_{\perp, \rm true} = \alpha (1+\epsilon)^{-1} r_{\perp}\,,
\end{split}
\end{equation}
where we have dropped subscript "fid" for simplicity. The separation between two objects can be described in Cartesian $(r_{\perp}, r_{\parallel})$ or polar coordinates $(r, \mu)$, which are related by
\begin{equation}
\label{eq:Cartesian_to_polar}
\begin{split}
r^2 = r_{\perp}^2 + r_{\parallel}^2\,,\\
\mu = {\rm cos(\theta)} = \frac{r_{\parallel}}{r}\,.
\end{split}
\end{equation}
Therefore Eq.~\ref{eq:Cartesian_to_polar} for the true separations would be
\begin{equation}
\begin{split}
r_{\rm true} = \alpha \sqrt{(1+\epsilon)^4 r_{\perp}^2 +(1+\epsilon)^{-2} r_{\parallel}^2}\\
\mu_{\rm true} = \frac{ (1+\epsilon)^2 r_{\parallel}}{\sqrt{(1+\epsilon)^4 r_{\perp}^2 +(1+\epsilon)^{-2} r_{\parallel}^2}}
\end{split}
\end{equation}
The true correlation function can be decomposed into its Legendre multipoles. In the following, we assume that monopole, quadrupole, and hexadecapole would suffice in the summation below, as the higher order moments can be ignored.
\begin{equation}\label{eq:Legendre}
\xi(r_{\rm true}, \mu_{\rm true})
=
\sum_{l = 0,2,4} \xi_{l}(r_{\rm true}) \, \mathcal{L}_{l}(\mu_{\rm true})
\end{equation}
What we measure in reality, assuming a wrong fiducial cosmology, is shifted multipole moments of the correlation function\,,
\begin{equation}
\xi_{l} (r; \alpha,\epsilon)= \frac{2l+1}{2}\int_{-1}^{1} d\mu L_l(\mu)\, \xi(r_{\rm true}(\alpha,\epsilon),\mu_{\rm true}(\alpha,\epsilon))\,.
\end{equation}
For a BAO-only measurement, we are not interested in any feature of the correlation function other than where the BAO is. Typically, one adds an additive polynomial term $A(r)$ to account for the scale-dependent bias and redshift space distortions, and a multiplicative $B(r)$ factor to adjust the amplitude,
\begin{equation}\label{eq:xi_model_uncontaminate}
\xi_{l}^{\rm model}(r; \alpha, \epsilon) = B(r)\xi_{l}(r; \alpha, \epsilon) + A(r))
\end{equation}
Using the right form of polynomials is important, because even though $A(r)$ and $B(r)$ are nuisance parameters not containing any BAO information, they may bias our estimation of the BAO peak if not properly fitted for. We use three orders of polynomials: $A(r) = a_1/r^2 + a_2/r + a_3$, and a constant for the amplitude $B(r) = B_0$.
Substituting Eq.~\ref{eq:xi_model_uncontaminate} into Eq.~\ref{eq:xi_final} gives our final model for the monopole and quadrupole of a catalogue contaminated by some fraction of interlopers $f_{\rm i}$
\begin{equation}
\begin{split}
\label{eq:xi_model_0}
\xi_{0}^{\rm model}(r;\{a_1, a_2, a_3, B_0, \alpha, \epsilon, f_{\rm i}\}) \\
&=
(1+2f_{\rm i}^2-2f_{\rm i})
\times
\left[
(\frac{a_1}{r^2} + \frac{a_2}{r} + a_3) +
B_0 \, \xi^{\rm}_{0}(r; \alpha,\epsilon))
\right]
\\
&+
\frac{2f_{\rm i}(1-f_{\rm i})}{(1+2f_{\rm i}^2-2f_{\rm i})}
\times \mathcal{M}_{0}[\xi^{\rm measured}](r)\,,
\end{split}
\end{equation}
\begin{equation}
\begin{split}
\label{eq:xi_model_2}
\xi_{2}^{\rm model}(r;\{w_1, w_2, w_3, B_0, \alpha, \epsilon, f_{\rm i}\}) \\
&=
(1+2f_{\rm i}^2-2f_{\rm i})
\times
\left[
(\frac{w_1}{r^2} + \frac{w_2}{r} + w_3) +
B_0 \, \xi^{\rm}_{2}(r; \alpha,\epsilon))
\right]
\\
&+
\frac{2f_{\rm i}(1-f_{\rm i})}{(1+2f_{\rm i}^2-2f_{\rm i})}
\times \mathcal{M}_{2}[\xi^{\rm measured}](r)\,.
\end{split}
\end{equation}
Throughout this paper, when we refer to a monopole only fit, we use Eq.~\ref{eq:xi_model_0} assuming $\epsilon = 0$, which simplifies $\xi^{\rm}_{0}(r; \alpha,\epsilon)$ to just $\xi^{\rm}_{0}(\alpha r)$. In that case, there are six free parameters that we fit for. When we refer to simultaneous monopole and quadrupole fit, we use Eq.~\ref{eq:xi_model_0} and \ref{eq:xi_model_2} together. In that case, there are four additional free parameters (ten in total) that should be fitted for.
In the following, we describe the method used to obtain the templates of the monopole and quadrupole, $\xi_0(r)$ and $\xi_2(r)$.
We model the matter power spectrum in real space as
\begin{equation}
P(k,\mu) = \left[ P_{\rm lin}(k) - P_{\rm smooth}(k)\right] e^{-\frac{1}{2} k^2 (\mu^2 \Sigma_{\parallel}^2 + (1- \mu^2) \Sigma_{\perp}^2)} + P_{\rm smooth}(k)\,.
\end{equation}
where $P_{\rm lin}(k)$ is the linear matter power spectrum obtained from CAMB and $P_{\rm smooth}(k)$ is the
no-wiggle counterpart computed as in \cite{vlah_perturbation_2016}, where the BAO oscillations have been removed. The Gaussian term accounts for the damping of the BAO wiggles due to non-linear evolution, with $\Sigma_{\perp}=4.8\,h^{-1}\,{\rm Mpc}$ and $\Sigma_{\parallel}=7.3 \,h^{-1}\,{\rm Mpc}$ at $z=1$ and the considered cosmology. In redshift space, the power spectrum is distorted by the Kaiser \cite{kaiser_clustering_1987} and the Finger-of-God (FoG) \cite{jackson_critique_1972} effects, which lead us to the following equation for the redshift space power spectrum \cite{blake_wigglez_2011}
\begin{equation}
P_{\rm s}(k,\mu) = b^2 (1+\beta \mu^2)^2 F(k,\mu)P(k,\mu)\,.
\end{equation}
In our analysis, we use the streaming model for FoG, $F(k,\mu) = (1+k^2\mu^2\Sigma_s^2)^{-2}$, with $\Sigma_{s} = 3 \,h^{-1}\,{\rm Mpc}$.
Subsequently, we calculate the correlation function by taking the Fourier transform of the redshift-space power spectrum.
\begin{figure}[t]
\centering
\includegraphics[width=1\textwidth,angle=0]{plots/dilation_vs_d_50to150.png}
\caption{
The effect of small-displacement interlopers on dilation parameters when interlopers are not included in the model, shown for $2$, $5$, and $10\%$ fraction of interlopers. As can be seen, for small-displacements, the estimation of $\alpha$ (upper panel) and $\epsilon$ (lower panel) are notably affected by interlopers, and therefore, small-displacement interloper corrections need to be implemented in the fitting model to avoid systematic bias.
For large-displacements, the effect of the cross-correlation on the recovered parameters rapidly decreases.
}
\label{fig:dilation_vs_deltad}
\end{figure}
\section{Results}
\label{sec:results}
In this Section, we present the results obtained by fitting the mean of the multipole moments of the correlation functions from $1,000$ mocks in redshift space.
To find the best fit parameters, we carried out a Markov-chain Monte Carlo (hereafter, MCMC) analysis (emcee package \cite{foreman_mackey_emcee_2013}) to minimize the posterior function introduced by Percival et al. \cite{Percival_posterior2021}:
\begin{equation}
\label{eq:posterior}
f(\xi^{\rm model} |\xi^{\rm data},C) \propto \left[ 1+\frac{\chi^2}{n_s-1} \right]^{-\frac{m}{2}}\,,
\end{equation}
where $m$ is given by Eq.~54 of \cite{Percival_posterior2021}.
We also assumed the following uniform priors on our parameters:
\begin{equation}
f(\alpha, \epsilon, f_i) =
\begin{cases}
1 & \text{if $0.6<\alpha<1.4$}\,,\\
& \text{\& $-1<\epsilon<1$}\,,\\
& \text{\& $0<f_i<0.5$}\\
0 & \text{otherwise}
\end{cases}
\end{equation}
The best fit values and the statistical errors presented in this Section (i.e., Figure~\ref{fig:dilation_vs_deltad}, \ref{fig:results}, Table~\ref{tab:results_all}, \ref{tab:results_mono_only}) are respectively calculated from the mean and standard deviation calculated using a Monte Carlo Markov Chain used to explore the posterior with 3000 steps.
To calculate $\chi^2$, we estimated the covariance matrix of $n_s = 1000$ independent contaminated mocks using
\begin{equation}
\label{eq:covariance}
C_{ij}[\xi(r_i)\xi(r_j)] = \frac{1}{(n_s-1)}\sum_{n=1}^{n_s}
[\xi_n(r_i) - \bar{\xi}(r_i)]
[\xi_n(r_j) - \bar{\xi}(r_j)]\,.
\end{equation}
To fit the model to the mean of $n_s$ mocks, Eq.~\ref{eq:covariance}, which gives the covariance matrix used for an individual mock, is divided by $n_s$. For the purpose of this paper, we only need to fit the data around the BAO peak. Therefore, in our analyses, we fit our model to data on scales between $r_{\rm min} = 50 \,h^{-1}\,{\rm Mpc}$ and $r_{\rm max} = 150 \,h^{-1}\,{\rm Mpc}$.
Let us begin with highlighting the necessity of correcting for the small-displacement interloper effect, and also roughly quantifying what we mean by "small" displacements. For this purpose, we repeated the process of making contaminated catalogues described in Section~\ref{sec:simulations}, but for a wider range of displacements, from $45$ to $200 \,h^{-1}\,{\rm Mpc}$. Without correcting for the interloper effect, we found the best fit estimations on $\alpha$ and $\epsilon$. In Figure~\ref{fig:dilation_vs_deltad}, the difference between the estimation and true is plotted against displacement, and from this figure we find that:
\begin{itemize}
\item How much the estimated dilation parameters deviate from their true value depends on the level of contamination of the catalogue (different colors). In the small-displacement regime, estimations are worse for higher fractions of of interlopers. For $10\%$ interlopers, the MCMC code does not even converge for displacements smaller than $75 \,h^{-1}\,{\rm Mpc}$ (hence why not shown in this plot).
\item
For $\Delta d \approx 100 \,h^{-1}\,{\rm Mpc}$, the best fit does not differ from 1 significantly, since the two peaks ($\Delta d$ and BAO) overlap, and there is no need to change the peak location in the model by varying $\alpha$. As we make $\Delta d$ slightly larger (smaller) than the BAO scale, until around $120 \,h^{-1}\,{\rm Mpc}$ ($80 \,h^{-1}\,{\rm Mpc}$), $\alpha$ becomes smaller (larger) than 1 to allow the model without the interloper signal to compromise between fitting the BAO and the interloper cross-correlation term.
\item
There is a smooth transition around $\Delta d \approx 140-160 \,h^{-1}\,{\rm Mpc}$ such that, at smaller displacements, the estimated dilation parameters deviate from the truth, while on larger scales they become less affected by the presence of interlopers.
This is because as we increase the displacement, the cross-correlation peak phases out of the maximum fitting range of $150 \,h^{-1}\,{\rm Mpc}$, and the tail that remains within the fitting range gets corrected by polynomials, rather than the dilation parameters.
\item
The extrema reflect a jump in the best-fit position in the posterior between two isolated solutions with different $\Delta d$: when the cross-correlation becomes too far from the position of the BAO, the best fit solution jumps from being a compromise between fitting both BAO and spike to a solution that fits the BAO (only) better, ignoring the cross-correlation spike.
\end{itemize}
We should emphasize that large-displacements can still bias the BAO measurement, even though the cross-correlation would be negligible, as mentioned in Section~\ref{sec:intro}.
\begin{figure}[t]
\centering
\includegraphics[width=1\textwidth,angle=0]{plots/rsd_alp_eps_fi_1000sims_all.JPEG}
\caption{
The difference between estimation and true for $\alpha$ (top row), $\epsilon$ (middle row), and $f_i$ (bottom row), as plotted against the true fraction of interlopers. For comparison, the green lines show the fit without taking into account interlopers and the orange shows the fit with interlopers (our analysis).
}
\label{fig:results}
\end{figure}
Now, we take into account the effect of interlopers in the fitting model by using the formalism described in Section~\ref{sec:Modelling_the_contaminated_correlation_function}. Figure~\ref{fig:results} demonstrates how much the estimation of our pipeline for $\alpha$ (top row), $\epsilon$ (middle row), and $f_i$ (lower row) deviates from the corresponding true values when considering catalogs contaminated by different fractions of interlopers (x-axis), and three different values for the displacements $\Delta d$: $85\,h^{-1}\,{\rm Mpc}$ (left column), $90\,h^{-1}\,{\rm Mpc}$ (middle column), and $97 \,h^{-1}\,{\rm Mpc}$ (right column). The orange symbols corresponds to the case where we account for the presence of interlopers in our analysis using Eq.~\ref{eq:xi_final}.
We also consider green symbols which show the results when ignoring interlopers (as if $f_i$ was set to zero in Eq.~\ref{eq:xi_final}, and as assumed in Figure~\ref{fig:dilation_vs_deltad}). We considered fits to the monopole alone (shown in triangle) and to the monopole and quadrupole simultaneous (rectangles). From Figure~\ref{fig:results}, we deduce that:
\begin{itemize}
\item
\emph{Monopole only with no correction (green triangles):}
At a fixed value for the displacement $\Delta d$, the estimated isotropic dilation parameter $\alpha_{\rm est}$ increasingly deviates from 1 when considering larger interloper fractions. This means that the bias in $\alpha$ introduced by the interlopers increases with the amount of contamination, as previously noted in \cite{massara2020line}. How much the bias increases with $f_i$ depends on the value of the displacement: the smaller the displacement, the faster the bias increases with the amount of contamination. Indeed, when the displacement is close to the BAO position, e.g. $\Delta d = 97\,h^{-1}\,{\rm Mpc}$, the peak in the galaxy-interloper cross-correlation is very close to the BAO location and it causes a mild distortion and shift of the BAO peak. On the other hand, smaller displacements present a peak in the cross-correlation that is further away from the BAO peak location, causing a much larger distortion of the BAO feature (see also Figure~\ref{fig:mono_general}).
\item
\emph{Monopole+Quadrupole with no correction (green rectangles):}
Compared to the previous case (monopole only with no correction) and for each combination of $\Delta d$ and $f_i^{\rm true}$, $\alpha_{\rm est}$ is closer to $1$ when including the quadrupole in the fit. This happens because the contamination is anisotropic: By including the quadrupole, the parameter $\epsilon$ can mimic and absorb the anisotropy, leaving $\alpha$ less biased. However, the $\epsilon$ measurements are strongly biased, with deviations from the fiducial value increasing with the amount of contamination and decreasing with increasing $\Delta d$.
\item
\emph{Monopole only with interloper correction (orange triangles):}
Compared to the no correction case, the values of $\alpha_{\rm est}$ are now much closer to the true value and they exhibit a systematic bias smaller than $3.9 \times 10^{-3}$ for all considered level of contamination and displacements. This model allows us to obtain an estimation for the fraction of interlopers $f_i^{\rm est}$ with a bias that is below $8 \times 10^{-3}$ for all levels of contamination and displacements. (More details in Table~\ref{tab:results_mono_only})
\item
\emph{Monopole+Quadrupole with interloper correction (orange rectangles):} The values of $\alpha_{\rm est}$ and $\epsilon_{\rm est}$ are much closer to the truth than in the no correction case. The residual systematic biases are smaller than $5 \times 10^{-3}$ for $\alpha$ and $0.01$ for $\epsilon$. The percentage of contaminants can be estimated with better precision in this case than when fitting for the monopole alone: the statistical errors decrease by a factor of two on average. The accuracy on $f_i$ is also generally improved when the displacement is small, while no significant amelioration appears if $\Delta d = 97\,h^{-1}\,{\rm Mpc}$.
\end{itemize}
A more detailed description of our results of the monopole+quadrupole and monopole only analyses are given in Table~\ref{tab:results_all} and Table~\ref{tab:results_mono_only} respectively, where the systematic and statistical errors on $\alpha$, $\epsilon$ and $f_i$ are reported for all the displacements and fractions considered. We find that the systematic error on $\alpha$ with interlopers using our model is $4.4 \times 10^{-3}$ which is consistent with pre-reconstruction BAO analyses on SDSS data without interlopers (\cite{bautista_completed_2020}, \cite{magana_sdss-iii_2013}, \cite{alam_clustering_2017}). This shows that slitless spectroscopy does not degrade BAO measurements.
\begin{center}
\input{Table_result_all}
\end{center}
\begin{center}
\input{Table_results_monopole}
\end{center}
\section{Discussions}\label{sec:discussions}
We have investigated how to extract the BAO scale from a contaminated galaxy catalogue containing a sub-sample of interlopers that are correlated with the main target galaxies. This case is particularly important for the future Roman Space Telescope, where a population of [O{\sc\,iii}]\ emitting galaxies extracted thanks to the observation of a single line is likely to be contaminated by $\rm H\beta$ interlopers.
In this study, we assumed that the $\rm H \beta$ line is misidentified as the primary line at $500.7 \, \rm nm$ of the [O{\sc\,iii}] \ doublet.
Depending on how the doublet is fitted, one might need to take into account the weighted average of the primary and the secondary lines of [O{\sc\,iii}]. Nonetheless, this will only impact the value of $\Delta d$ as a function of redshift in Eq.~\ref{eq:displacement}, and will not affect our conclusions.
In the following, we highlight some of the main outcomes of our analysis.
Generally, the contaminated correlation function has three terms (Eq.\ref{eq:xi_tt_general}): The galaxy-galaxy term, the interloper-interloper term, and the galaxy-interloper term. We have split interlopers into three categories: random, large-displacement and small-displacement interlopers. We have argued that random contaminants, uncorrelated with either themselves or the target galaxies are trivial to account for.
For both small-displacement and large-displacement interlopers, we need to model their auto-correlation, but this is generally easy to do. In addition, the effect of the interloper-interloper auto-correlation on the contaminated signal is a simple linear super-position of two correlation functions: in the extremes, either the change in $H(z)$ between true redshifts of the targets and interlopers is small and so the BAO position is unaffected, or it is large and the superimposed BAO peak is far from that of the targets. In either case and in between, to measure the BAO position, we only need a model on large scales where linear theory is sufficient to describe the BAO feature and we know the position through the relative line positions. If interlopers and galaxies have different bias, an extra (bias) parameter can be added to the second term of Eq.~\ref{eq:xi_tt_general} and can be fitted for along other parameters introduced in our analysis. It also may be possible to measure this using calibration data, by comparing the interloper-interloper and target-target correlation functions. In contrast, for small-displacement interlopers, it is critical to have a good estimate of the interloper-target cross-correlation.
For our simulations, the interlopers are drawn from the same underlying distribution as the target galaxies. This allowed us to simplify the equations, merging the galaxy-galaxy and interloper-interloper terms together. In practice, the populations of [O{\sc\,iii}]\ and $\rm H \beta$ emission line galaxies are different, and this can change our equations requiring us to keep these terms separate. We do not expect this to significantly alter our conclusions, or to present a difficulty for modeling.
For small-displacement interlopers, the interloper shift can be considered as being constant, and is dependent on the fiducial cosmology used to convert redshifts to distances. Thus its position does not provide cosmological information. We made the assumption of a constant shift because the displacement only changes by less than one percent across the bin considered, similar to the case for the BAO scale being fixed within an bin of observational data. Hence, the BAO signal in $\xi_{\rm ii}(\Vec{r})$ will not be shifted as a correlation function is not modified by a translation, while it is modified in $\xi_{\rm gi}(\Vec{r})$.
Thus, the key to understanding and modelling the contaminated correlation function is through accurately modelling the cross-correlation between the main galaxy targets and the interlopers that are misidentified as main targets, which is described in Eq.~\ref{eq:cross_equation} and required a mapping between true and wrongly inferred positions for the interlopers.
Again, with calibration data, it is possible to apply this mapping (refer to \ref{eq:polar_coordinates}) to the measured cross-correlation at the true distance.
We leave it to future work to determine the size of calibration data required to determine this function with sufficient accuracy that it does not impact on BAO measurements. The good news is that we need to know this function on small scales, where small volumes of the Universe will be sufficient.
For our study, we used the same assumptions used to create the simulations (the [O{\sc\,iii}]\ and $\rm H \beta$ emitting galaxies are drawn from the same population) to estimate the cross-correlation with the auto-correlation function of the main galaxy targets. We considered models based on CAMB and HALOFIT to derive an estimation of the cross-correlation. We found that they both fail in precisely modelling the peak of the monopole and quadrupole of the cross-correlation. These models also need to account for the difference between fiducial and true cosmologies via the AP parameters, that would need to be added in the cross-correlation function of Eq.~\ref{eq:xi_model_0}, and Eq.~\ref{eq:xi_model_2} as well.
Therefore, we tested another possibility that is more powerful on smaller scales: using the measured contaminated correlation functions to estimate cross-correlation. We argued that using the latter is the most precise for the Quijote mocks used in this paper, since it matches the way that the mocks were constructed (interlopers and target galaxies have the same galaxy bias). Indeed, the resulting cross-correlation is able to better model the peak of the monopole of the cross-correlation, and both peak and the shape of the quadrupole of the cross-correlation (refer to Figure~\ref{fig:cross_correlation}).
With all of the mentioned assumptions made, we finally presented a model for monopole and quadrupole of the contaminated correlation function. We tested our model by fitting it to the mean of 1,000 mocks for five different interloper fractions and three different displacements. We found that for all these cases, our method is successful at making an unbiased prediction of the dilation parameters (recovering the BAO peak), with a systematic error of less than $4.4 \times 10^{-3}$, which is consistent with previous pre-reconstruction analyses. It also enables us to have a good estimation of the fraction of interlopers. Whereas if we did not take interlopers into account in the model, for $15\%$ interlopers we would have an order of magnitude higher systematic errors -- as high as $-0.020$ and $0.044$ for $\alpha$ and $\epsilon$.
Moreover, the signal in the correlation function from interlopers is sufficiently clear that it allows us to measure the fraction of interlopers with a systematic error of at most few percent. This will be useful for cosmological analysis other than measuring BAO where the fraction of interlopers can change the amplitude of the signal, and knowing the fraction is important.
Since our analysis is specifically useful for future Roman [O{\sc\,iii}]\ targets that are contaminated by $\rm H \beta$ interlopers, we predict the statistical error that our method would have for Roman telescope data. To do so, we scaled the statistical errors on the mean calculations from 1,000 mocks at redshift 1 ($\Delta d = 85 \,h^{-1}\,{\rm Mpc}$). These 1,000 mocks have a cumulative volume of $\rm 1,000 (h^{-1} Gpc)^3$. The approximate volume of Roman survey between redshift 1.8 and 2.8 (where [O{\sc\,iii}]\ galaxies live) is $V \approx \rm 10 (h^{-1} Gpc)^3$. Thus, the statistical error is scaled by $\sqrt{1000/10} = 100$.
For the range of $0-15\%$ interlopers, we find that the statistical error on $\alpha$ and $\epsilon$ is $0.01$ which is larger than their corresponding systematic errors.
Small-displacement interlopers give rise to a strong signal in the correlation function on similar scales to the cosmological signal. Even so, this signal has key differences from the cosmological signal such that a joint model fitting both the cosmological signal and the effect of interlopers can be constructed and we do not see a degradation in the cosmological measurements in the tests we have performed. Furthermore, the interloper signal itself allows excellent diagnostics on the contamination, allowing a simultaneous fit to the contaminant fraction. For random interlopers however, it is more difficult to measure the contaminant fraction, which has to be done before the amplitude of clustering can be measured. Thus, while small-displacement interlopers are definitely the most pernicious in terms of the correlation function, for fits to the data this is not the case.
\acknowledgments
We thank Yun Wang for useful comments and discussions. Research at Perimeter Institute is supported in part by the Government of Canada through the Department of Innovation, Science and Economic Development Canada and by the Province of Ontario through the Ministry of Colleges and Universities. This research was enabled in part by support provided by Compute Ontario (computeontario.ca) and the Digital Research Alliance of Canada (alliancecan.ca).
\printbibliography
\end{document}
|
\section{An Efficient Adoption to Software Verification}
\label{sect:approach}
In this section, we describe the proposed approach for adopting IMC to software verification.
The presentation will be split into two parts.
The first part will introduce the main idea towards an efficient adoption.
In essence, we take advantage of \textit{large-block encoding}~\cite{LBE} to draw an analogy between a program and the state-transition system discussed in~\cref{sect:background}.
The idea is not only helpful for this paper, but might also shed light on the efficient adoptions of other algorithms.
The second part will focus on the implementation, discussing the data structure and the algorithmic procedures.
We implemented the proposed adoption in the verification framework \cpachecker~\cite{CPACHECKER},
leveraging its flexibility provided by configurable program analysis~\cite{CPA}.
\subsection{Obtaining Transition Relation with Large-Block Encoding}
\label{sect:approach-LBE}
As discussed in~\cref{sect:background-imc}, obtaining a transition relation from a CFA is not a straightforward process.
Unlike sequential Boolean-logic circuits, which usually have only one feedback loop, a CFA could have arbitrarily many loops.
To simplify the problem, we start by considering single-loop programs.
As a program with multiple loops can be converted into a single-loop program by a standard transformation~\cite{DragonBook,kIndForDMARaces},
this simplification will not hurt the generality of the proposed approach.
To obtain the transition relation of a single-loop program, we take advantage of \textit{large-block encoding} (LBE)~\cite{LBE}.
Given a CFA, LBE repeatedly rewrites the original CFA in order to summarize it.
In the summarized CFA, each loop-free subgraph of the original CFA is represented by a single control-flow edge.
The edge is annotated with a formula that encodes the program behavior of the represented subgraph of the original CFA.
A summarized single-loop CFA is given in~\cref{fig:summarized-cfa}.
It has an initial location~$l_0$, a loop-head location~$l_H$, a loop-body location~$l_B$, a loop-tail location~$l_T$, and an error location~$l_E$.
These locations correspond to the program locations in the original single-loop CFA before summarization.
The edges of the summarized single-loop CFA are labeled with the following formulas:
Formula~$\varphi_0$ summarizes the subgraph from~$l_0$ to~$l_H$,
formula~$\varphi_L$ summarizes the subgraph from~$l_B$ back to~$l_H$, and
formulas~$\varphi_E, \varphi_E'$ summarize the subgraphs from~$l_T, l_B$ to~$l_E$, respectively.
\newsavebox{\summarizedCFA}
\begin{lrbox}{\summarizedCFA}
\begin{minipage}[b]{0.4\textwidth}
\centering
\scalebox{0.8}{\input{figures/single-loop-cfa.tex}}
\end{minipage}
\end{lrbox}
\newsavebox{\summarizedARG}
\begin{lrbox}{\summarizedARG}
\begin{minipage}[b]{0.4\textwidth}
\centering
\scalebox{0.8}{\input{figures/single-loop-arg.tex}}
\end{minipage}
\end{lrbox}
\begin{figure*}[t]
\centering
\subfloat[Summarized single-loop CFA]{\usebox{\summarizedCFA}\label{fig:summarized-cfa}}
\subfloat[$k$-Unrolling]{\usebox{\summarizedARG}\label{fig:summarized-arg}}
\caption{%
A summarized single-loop CFA (a) and its $k$-unrolling (b)
}
\label{fig:summarized}
\end{figure*}
We notice that,
the summarized single-loop CFA has a natural analogy to those predicates used in~\cref{sect:background-imc}:
The initial-state predicate $I(s)$ is analogous to $\varphi_0$,
the transition relation $T(s,s')$ is analogous to $C\land\varphi_L$, and
the negated safety property $\lnot P(s)$ is analogous to $(\lnot C\land\varphi_E)\lor (C\land\varphi_E')$.
Using LBE, we successfully obtain the required predicates without destroying the control flow of the CFA.
Furthermore, in order to perform IMC, we have to unroll the summarized single-loop CFA and construct the BMC query~\cref{eq:BMC}.
In~\cref{fig:summarized-arg},
we unroll the CFA by drawing all possible paths starting from $l_0$,
iterating the loop $k$ times ($k+1$ visits to $l_H$),
and finally reaching $l_E$.
A node in~\cref{fig:summarized-arg} consists of a program location which the control flow is currently at and a formula $\blkf$ to encode all possible paths starting from the program location of the preceding node.
Note that $\blkf$ is indexed with the unrolling counter $i$ to distinguish between different iterations.
To discover the similarity between~\cref{eq:BMC} and~\cref{fig:summarized-arg}, we additionally label a node in~\cref{fig:summarized-arg} with the subformula in~\cref{eq:BMC} that $\blkf$ corresponds to.
From those labels, we observe that the formulas in the unrolled CFA nicely match the subformulas in~\cref{eq:BMC}.
We name the formula matching $I(s_0)$ \textit{prefix formula},
the formula matching $T(s_0,s_1)$ \textit{loop formula}, and
the formula matching $T(s_1,s_2) \land \ldots \land T(s_{k-1},s_k) \land \lnot P(s_k)$ \textit{suffix formula}.
In the following we will describe an implementation in~\cpachecker to construct these formulas and perform IMC.
\subsection{Implementation in \cpachecker}
\label{sect:approach-impl}
Before we delve into implementation details, we would like to emphasize that the idea to extract a transition relation with LBE is general and independent of the underlying framework.
We chose to implement the proposed adoption in~\cpachecker because it provides (1) the necessary components for the adoption, which are highly configurable, and (2) the implementations of various state-of-the-art software-verification algorithms, which is convenient for the evaluation.
To implement the proposed adoption in~\cpachecker, we use the \emph{Predicate~CPA}~\cite{ABE} as the core data structure to store formulas that encode program behavior, and build the algorithmic procedures of IMC on top of the \cpapa algorithm~\cite{AlgorithmComparison-JAR}.
\subsubsection{Predicate CPA}
The \emph{Predicate~CPA}~$\predcpa$ for \emph{adjustable-block encoding} (ABE)~\cite{ABE} uses a triple~$(\as,\labs{}{},\pf)$ of an abstraction formula~$\as$, an abstraction location~$\labs{}{}$, and a path formula~$\pf$ as an abstract state.
The abstraction formula~$\as$ stores the abstraction of the program state computed at the program location~$\labs{}{}$.
The path formula~$\pf$ syntactically encodes the program behavior from the abstraction location~$\labs{}{}$ to the current program location.
Abstract states where the path formula~$\pf$ is $\true$ are called \emph{abstraction states}; other abstract states are \emph{intermediate states}.
In this paper, we augment an abstract state of the Predicate CPA by a fourth element: a \emph{block formula} $\blkf$.
The block formula encodes all possible paths from the previous abstraction location to the current abstraction location, and is used to compute the abstraction formula.
In the implementation of~\cpachecker, the block formula is already stored in the data structure for the abstraction formula.
We append it to an abstract state of the Predicate CPA in order to make the subsequent discussion more understandable.
With the help of ABE, we can achieve the effect of LBE via using the block-adjustment operator~$\blk^{l}$~\cite{ABE}.
The operator~$\blk^{l}$ will make the Predicate CPA convert an intermediate state to an abstraction state if the current program location is at the loop head or the error location.
Under this configuration, the unrolled ARG, if projected to abstraction states, will have a similar structure to~\cref{fig:summarized-arg}.
Therefore, we can easily obtain the required formulas by collecting and combining the block formulas from the corresponding abstraction states in the ARG.
It is worth noting that
here we take advantage of the flexibility of the Predicate~CPA:
By choosing an appropriate implementation for the block-adjustment operator,
we can configure the Predicate~CPA to be suitable for IMC
(together with the algorithms described in the following)
without further changes to its definition.
Other choices for its operators would allow it to be combined with different algorithms
and implement different approaches
like \impact, predicate abstraction, and \kinduction~\cite{AlgorithmComparison-JAR}.
Using the Predicate~CPA as common framework
not only highlights conceptual differences and similarities between the approaches,
but also allows for comparing them experimentally
with the set of confounding variables kept to a minimum.
\subsubsection{Algorithmic Procedures}
We present an algorithm for the adoption of IMC to software verification in~\cref{imc-algo-main},
based on the \cpapa algorithm~\cite{AlgorithmComparison-JAR}.
The algorithm assumes single-loop programs as input.
We apply single-loop transformation~\cite{DragonBook,kIndForDMARaces} to input programs with multiple loops as a preprocessing.
The algorithm takes as input an upper limit $k_{max}$ for a counter $k$ that tracks the number of loop iterations on a program path and a composite CPA consisting of the Location CPA, the Predicate CPA described above, and the Loop-Bound CPA.
Note that the \cpapa algorithm~\cite{AlgorithmComparison-JAR} will stop exploring a program path
if the number of loop iterations on this path has reached the given unrolling upper bound.
While the algorithm is asked to unroll the program $k$~times,
it uses the last unrolling only for encoding the predicate~$P(s)$
and thus only $k-1$~copies of $T(s,s')$ appear in its BMC query.
This is done for consistency with other algorithms
expressed on top of the same unifying framework~\cite{AlgorithmComparison-JAR}.
For analyzing up to $k$~unrollings of a program with IMC,
we can pass $k+1$ as parameter to \cref{imc-algo-main}.
\begin{algorithm}[t]
\caption{IMC: main procedure}
\label{imc-algo-main}
\begin{algorithmic}[1]
\newcommand\lc{i
\REQUIRE
an upper limit $k_{max}$ for the loop unrolling bound $k$,\\ \hspace{4.5mm}
a composite CPA $\cpa$ with components: the Location CPA $\loccpa$,\\ \hspace{4.5mm}
the Predicate CPA $\predcpa$, and the Loop-Bound CPA $\boundscpa$\\
\ENSURE
\FALSE{} if an error path to $l_E$ is found,\\ \hspace{6.5mm}
\TRUE{} if a fixed point is obtained,\\ \hspace{6.5mm}
\textbf{unknown} otherwise
\STATE $k := 1$
\STATE $\initial\astate := (\initial\pc,(\true,\initial\pc,\true,\true),\{l_H\mapsto-1\})$\hfill\COMMENT{Create initial abstract state at $\pci$}
\STATE $\mathsf{abort}^\mathit{never} := \{\cdot\mapsto\false\}$ \hfill\COMMENT{$\mathsf{abort}^\mathit{never}$ always returns $\false$}
\STATE $\reached := \wait := \{\initial\astate\}$
\WHILE{$k \leq k_{max}$}
\STATE $(\reached, \wait) :=$\\
\hspace{5mm}$\textsf{\cpapa}(\cpa, \reached, \wait, \mathsf{abort}^\mathit{never}, k)$\label{imc-unroll-cfa}
\STATE $\blkf_p := \blkf \mid (l_H,(\cdot,\cdot,\cdot,\blkf),\{l_H\mapsto0\}) \in \reached$\label{imc-collect-formula-start}
\STATE $\blkf_l := \true$
\IF{$k>1$}
\STATE $\blkf_l := \blkf \mid (l_H,(\cdot,\cdot,\cdot,\blkf),\{l_H\mapsto1\}) \in \reached$
\ENDIF
\STATE $\blkf_s := \bigwedge_{i=2}^{k-1} \blkf \mid (l_H,(\cdot,\cdot,\cdot,\blkf),\{l_H\mapsto\lc\}) \in \reached ~~\land$\\
\hspace{7.2mm}$\bigvee \left\{ \blkf \mid (l_E,(\cdot,\cdot,\cdot,\blkf),\{l_H\mapsto(k-1)\}) \in \reached\right\}$\label{imc-collect-formula-end}
\IF{\textsf{sat}($\blkf_p\wedge \blkf_l\wedge \blkf_s$)}\label{imc-bmc-step}
\RETURN \FALSE{} \hfill\COMMENT{Found an error path via BMC query}
\ENDIF
\IF{$k>1$ \textbf{and} \textsf{reach\_fixed\_point($\blkf_p$,$\blkf_l$,$\blkf_s$)}}
\RETURN \TRUE{} \hfill\COMMENT{Obtained a fixed point via interpolation}
\ENDIF
\STATE $k:=k+1$
\ENDWHILE
\RETURN \textbf{unknown}
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[t]
\caption{IMC: \textsf{reach\_fixed\_point($\blkf_p$,$\blkf_l$,$\blkf_s$)}}
\label{imc-algo-fixed}
\begin{algorithmic}[1]
\REQUIRE
prefix formula $\blkf_p$, loop formula $\blkf_l$, and suffix formula $\blkf_s$
\ENSURE \TRUE{} if a fixed point is reached, \FALSE{} otherwise
\STATE \textsf{image} $:=$ \textsf{start} $:= \blkf_p$ \hfill \COMMENT{Set current reachable and starting states to initial states}
\WHILE{$\lnot\mathsf{sat}$($\mathsf{start}\land\blkf_l\land\blkf_s$)}\label{imc-inner-while-loop}
\STATE $\itp:= \mathsf{get\_interpolant}(\mathsf{start}\land\blkf_l,\blkf_s)$ \hfill \COMMENT{formula $A$: $\mathsf{start}\land\blkf_l$; formula $B$: $\blkf_s$}
\STATE $\itp:= \mathsf{shift\_variable\_index}(\itp,\blkf_p)$
\IF{$\lnot\mathsf{sat}$($\itp\land\lnot\mathsf{image}$)}
\RETURN \TRUE{} \hfill \COMMENT{Interpolant implies image: fixed point}
\ENDIF
\STATE $\mathsf{image}:=\mathsf{image}\lor\itp$ \hfill \COMMENT{Find new states: enlarge image}
\STATE $\mathsf{start}:=\itp$ \hfill \COMMENT{Start next iteration from new states}
\ENDWHILE
\RETURN \FALSE{} \hfill \COMMENT{Reach error: might be wrong alarm}
\end{algorithmic}
\end{algorithm}
An abstract state of the composite CPA is $(\pc, (\as,\labs{}{},\pf,\blkf), \{l_H\mapsto i\})$,
where the first element is an abstract state of the Location CPA representing the current program location $l$,
the second element is an abstract state of the Predicate CPA as explained above,
and the third element is an abstract state of the Loop-Bound CPA recording that the loop body starting from $l_H$ has been
completely traversed $i$~times already.
We also use the ARG CPA~$\argcpa$ to store the predecessor-successor relationship between abstract states.
To increase readability we simply use abstract states as elements in the ARG,
and explicitly give the unrolling upper bound~$k$ as a parameter to the \cpapa algorithm
(instead of passing it via the precision of the initial abstract state as in the literature~\cite{AlgorithmComparison-JAR}).
After unrolling the CFA with the \cpapa algorithm (\cref{imc-unroll-cfa}), we have to collect prefix, loop, and suffix formulas to pose a BMC query and perform the fixed-point computation via interpolation.
The formula collection is described in \crefrange{imc-collect-formula-start}{imc-collect-formula-end}, where we write $\blkf | {(\pc, (\as,\labs{}{},\pf,\blkf), \{l_H\mapsto i\})}$ to denote the block formula $\blkf$ of the abstract state ${(\pc, (\as,\labs{}{},\pf,\blkf), \{l_H\mapsto i\})}$.
The prefix formula $\blkf_p$ is the block formula of the abstract state ${(l_H, (\cdot,\cdot,\cdot,\blkf), \{l_H\mapsto 0\})}$;
if the loop head is visited more than once, i.e., $k>1$,
the loop formula $\blkf_l$ is the block formula of the abstract state ${(l_H, (\cdot,\cdot,\cdot,\blkf), \{l_H\mapsto 1\})}$,
otherwise, it is set to $\true$;
the suffix formula $\blkf_s$ is the conjunction of the following two formulas:
the conjunction of block formulas of the abstract state ${(l_H, (\cdot,\cdot,\cdot,\blkf), \{l_H\mapsto i\})}$ for $i=2,\ldots,(k-1)$, and
the disjunction of the block formulas of the abstract states ${(l_E, (\cdot,\cdot,\cdot,\blkf), \{l_H\mapsto (k-1)\})}$.
Note that the above formula collection at abstract states whose locations equal $l_H$ is unambiguous, meaning that there is a unique abstract state satisfying the conditions imposed by the Location~CPA~$\loccpa$ and the Loop-Bound CPA~$\boundscpa$.
This is because we assume single-loop programs and use LBE to summarize all paths between two adjacent abstraction states.
After collecting these formulas, the BMC query is simply the conjunction of the prefix, loop, and suffix formulas.
If the BMC query is unsatisfiable, we try to compute a fixed point using~\cref{imc-algo-fixed},
which implements the procedure described in~\cref{sect:background-imc} to iteratively derive interpolants from unsatisfiable BMC queries and grow a fixed point as their union.
We first initialize both \textsf{image}, which stores an overapproximation of the reachable states,
and \textsf{start}, which stores the starting states of BMC queries, to be the prefix formula.
Using $\mathsf{start}\land\blkf_l$ as formula $A$ and $\blkf_s$ as formula $B$, we derive an interpolant $\itp$.
As discussed in~\cref{sect:background-imc},
the $i^{th}$~interpolant is an overapproximation of the reachable states after $i$~loop unrollings.
We change the variables used in the interpolant to those in the prefix formula
and check whether the interpolant implies \textsf{image}.
If yes, a fixed point has been reached, and we conclude the property is true;
otherwise, we enlarge \textsf{image} by adding the states contained in the interpolant to it
and pose another BMC query starting from the interpolant.
If any BMC query during the iteration is satisfiable,
we return back to~\cref{imc-algo-main} and increase the loop-unrolling counter $k$ to check whether the violation is a wrong alarm.
\subsection{Example}
\label{sect:approach-example}
\begin{figure}[t]
\centering
\scalebox{0.9}{\input{figures/even-arg.tex}}
\caption{ARG constructed by the \cpapa algorithm~\cite{AlgorithmComparison-JAR} for the CFA in \cref{fig:even-cfa} ($k=2$)}
\label{fig:example-imc}
\end{figure}
We demonstrate step-by-step how to apply~\cref{imc-algo-main} and~\cref{imc-algo-fixed} to verify the CFA in~\cref{fig:even-cfa}.
The ARG constructed by the \cpapa algorithm when $k=2$ is shown in~\cref{fig:example-imc}.
In this figure, each abstract state is a tuple ${(\pc, (\as,\labs{}{},\pf,\blkf), \{\pc_4\mapsto i\})}$
of the abstract states of $\loccpa$, $\predcpa$, and $\boundscpa$.
Note that every abstract state in the ARG has an abstraction formula $\as$ (the first element in an abstract state of Predicate CPA) equal to $\true$, because IMC does not compute an abstraction formula.
Instead, it relies on interpolants for the abstraction of program states.
Abstract states whose predicate abstract state is an abstraction state are highlighted in gray.
We use~$r$ to denote the return value of function~\texttt{nondet}.
The prefix formula~$\blkf_p$ is the block formula $x_0=0$ of abstract state~$e_1$,
the loop formula~$\blkf_l$ is the block formula $\lnot(r_0=0)\land x_1=x_0+2$ of abstract state~$e_6$,
and the suffix formula~$\blkf_s$ is the block formula $r_1=0\land\lnot(x_1\%2=0)$ of abstract state~$e_8$
(note that the block formula of abstract state~$e_3$, which also has location~$l_8$,
is not selected because $l_4 \mapsto 0$ does not match in \cref{imc-collect-formula-end} of \cref{imc-algo-main}).
As the BMC query $x_0=0 \land \lnot(r_0=0)\land x_1=x_0+2 \land r_1=0 \land \lnot(x_1\%2=0)$ is unsatisfiable, we try to compute a fixed point using~\cref{imc-algo-fixed}.
Variables \textsf{image} and \textsf{start} are initialized to $x_0=0$.
Using $x_0=0 \land \lnot(r_0=0)\land x_1=x_0+2$ as formula $A$ and $r_1=0\land\lnot(x_1\%2=0)$ as formula $B$, we can derive an interpolant $\itp$ from the unsatisfiable BMC query.
Assume that $\itp$ is $x_1\%2=0$, referring to the common variable $x_1$ between formulas $A$ and $B$.
After shifting the variable to the one used in $\blkf_p$, we obtain $x_0\%2=0$.
As the interpolant does not imply \textsf{image}, we enlarge the current image by disjoining it with the interpolant.
The computation is then repeated again, with \textsf{start} equal to $x_0\%2=0$ this time.
The BMC query in the second iteration becomes $x_0\%2=0 \land \lnot(r_0=0)\land x_1=x_0+2 \land r_1=0 \land\lnot(x_1\%2=0)$, which is still unsatisfiable.
Assume the interpolant is again $x_1\%2=0$.
Obviously, we have reached a fixed point, as the newly derived interpolant implies \textsf{image}.
Therefore, we conclude that the property holds.
\subsection{Correctness}
It is straightforward to see that~\cref{imc-algo-main} is \emph{precise}, i.e.,
does not produce wrong alarms,
because if it returns \textbf{false},
then the BMC query for all paths from~$l_0$ to~$l_E$ at~\cref{imc-bmc-step} is satisfiable,
which implies that the CFA has a feasible path to~$l_E$.
More interesting is the soundness of~\cref{imc-algo-main},
i.e., whether it may produce wrong proofs,
which we discuss in the following.
Its soundness follows from that of large-block encoding~\cite{LBE} and the original IMC algorithm~\cite{McMillanCraig}.
We state the soundness of~\cref{imc-algo-main} when it is applied to a single-loop CFA in~\cref{thm:imc-soundness}.
For CFAs with multiple loops,
the soundness will also depend on that of the single-loop transformation~\cite{DragonBook,kIndForDMARaces}.
\begin{theorem}\label{thm:imc-soundness}
Given a single-loop CFA~$A$ and its corresponding composite CPA~$\cpa$ as input,
if \cref{imc-algo-main} returns \textbf{true} upon~$\cpa$,
then $A$ does not have a feasible path to~$l_E$.
\end{theorem}
\begin{proof}
We prove the statement by contradiction.
Suppose~\cref{imc-algo-main} returns \textbf{true} when the value of the loop-unrolling counter $k$ equals~$\hat{k}$,
but the single-loop CFA has a feasible path to $l_E$.
We split into two cases based on the number $\hat{h}$ of the visits to $l_H$ on the error path.
First, assume $\hat{h}\leq\hat{k}$.
Thanks to the sound summarization of LBE,
the formula of the error path must imply $\blkf_p\land\blkf_l\land\blkf_s$ when $k=\hat{h}$.
Therefore, \cref{imc-algo-main} should have returned \textbf{false} as $k=\hat{h}$,
because the BMC query at~\cref{imc-bmc-step} of~\cref{imc-algo-main} is satisfiable.
This result contradicts to the assumption that~\cref{imc-algo-main} returns \textbf{true}.
Second, assume $\hat{h} > \hat{k}$.
Such error path indicates the existence of a state $\hat{s}$
that is reachable from $l_0$ by traversing the loop $\hat{h}-\hat{k}$ times
and will reach $l_E$ after further traversing the loop $\hat{k}-1$ times.
We will show that~\cref{imc-algo-fixed} will return \textbf{false} after discovering $\hat{s}$ via interpolation.
Note that~\cref{imc-algo-fixed} cannot return \textbf{true} before finding $\hat{s}$,
because the state must be contained in the computed fixed point.
According to the property of the original IMC algorithm described in~\cref{sect:background-imc},
the interpolant derived in the $i^{th}$ while-loop iteration of~\cref{imc-algo-fixed} is an overapproximation of the set of states reachable from $l_0$ by traversing the loop $i$ times.
Therefore, $\hat{s}$ must belong to the interpolant $\itp$
derived in the $(\hat{h}-\hat{k})^{th}$ while-loop iteration of~\cref{imc-algo-fixed},
which will be used as new starting states in the next iteration.
Moreover, because of the soundness of LBE,
the formula from the $(\hat{h}-\hat{k}+1)^{th}$ $l_H$ to $l_E$ (involving $\hat{k}$ visits to $l_H$) on the error path
must imply $\blkf_l\land\blkf_s$ when we enter~\cref{imc-algo-fixed} with $k=\hat{k}$.
As a result, in the beginning of the next iteration,
the satisfiability query at~\cref{imc-inner-while-loop} of~\cref{imc-algo-fixed} must be satisfiable,
which makes~\cref{imc-algo-fixed} return \textbf{false}.
This in turn prevents~\cref{imc-algo-main} from returning \textbf{true} when $k=\hat{k}$,
contradicting to our assumption.
Having analyzed the above two possibilities,
we conclude that such a feasible error path does not exist,
and hence~\cref{imc-algo-main} is sound.
\end{proof}
\subsection{Backward Derivation of Interpolants}
Notice that in the example of~\cref{sect:approach-example}, the ``quality'' of interpolants heavily affects the convergence of the fixed-point computation.
For example, instead of $x_1\%2=0$, which is actually the loop invariant, suppose the interpolant derived by the solver is $x_1=2$.
Starting from this interpolant, we might be trapped in a sequence of interpolants $x_1=4$, $x_1=6$, $x_1=8, \ldots$, and never reach a fixed point.
While in general it is difficult to control the interpolation process of the solver, there is a trick to mitigate this problem.
First, we switch the labels of the two formulas, i.e., we label the original formula $B$ as the new formula $A$ and the original formula $A$ as the new formula $B$.
Second, we ask the solver to derive an interpolant for the new formulas and then negate it.
The negated interpolant is a valid interpolant for the original formulas $A$ and $B$.
In other words, instead of $\mathsf{get\_interpolant}(A, B)$, we use $\lnot \mathsf{get\_interpolant}(B, A)$.
Using this trick in IMC, we are actually deriving interpolants backwards, from the safety property.
Therefore, we call it \textit{backward derivation} of interpolants.
With the backward derivation, we can in practice often avoid the bad interpolant $x_1=2$ and obtain the good one $x_1\%2=0$ for fast convergence of the example program in~\cref{sect:approach-example}.
Empirically, we found that the backward derivation performs slightly better than the forward derivation.
This phenomenon might be attributed to the fact that deriving the interpolants backward from the safety property side is likely to yield interpolants summarizing information more relevant to proving the property.
As a result, we use it as default in our implementation.
\section{Background}
\label{sect:background}
In the following, Boolean connectives $\lnot, \lor, \land, \rightarrow, \equiv$ are used in their conventional semantics.
A first-order logical formula is also interpreted as a set of (program) states that satisfy the formula, and we use the two terms interchangeably when it is clear from the context.
\subsection{{Interpolation-Based Model Checking}\xspace}
\label{sect:background-imc}
{Interpolation-based model checking}\xspace (IMC)~\cite{McMillanCraig} is an algorithm for unbounded model checking to verify safety properties of state-transition systems.
It can be considered as an extension of BMC, which is well-known for bug hunting.
In order to describe IMC, we first define the notation to formalize a state-transition system.
Second, we review Craig's interpolation theorem~\cite{Craig57}, which is the core concept to extend BMC to unbounded model checking.
\subsubsection{State-Transition System}
Let $s$ and $s'$ be two arbitrary states in the state space of a state-transition system.
We formalize the state-transition system by three predicates over states.
Predicate $I(s)$ evaluates to $\true$ if state $s$ is an initial state of the system.
Predicate $T(s,s')$ evaluates to $\true$ if the system can transit from state $s$ to state $s'$.
It is also called the \textit{transition relation} of the system.
Predicate $P(s)$ evaluates to $\true$ if state $s$ satisfies the safety property to be verified.
In the above formulation of a state-transition system, we do not assume the state space to be finite or infinite.
The working of IMC is similar in both cases, provided that the underlying constraint solver (SAT/SMT solver) supports the reasoning over the corresponding logical formulas.
\subsubsection{Craig's Interpolation Theorem}
Given two logical formulas $A$ and $B$, if $A \land B$ is unsatisfiable,
Craig's interpolation theorem~\cite{Craig57} guarantees the existence of a logical formula $C$,
called an interpolant of $A$ and $B$, which only refers to the common variables of $A$ and $B$ and satisfies:
(1)~$A \rightarrow C$ and (2)~$C \land B$ is unsatisfiable.
\subsubsection{Algorithm Description}
The overall procedure of IMC~\cite{McMillanCraig} can be decomposed into two phases.
The first phase poses a BMC query by unrolling the transition relation $k$~times
and constructing a formula representing all possible execution paths from an initial state to a bad state
(a~state that violates the safety property) with $k$~transitions.
We use variable $s_i$ to denote the state after the $i^{th}$ transition.
Furthermore, to facilitate Craig interpolation in the second phase, the BMC query is partitioned into two formulas $A$ and $B$:
\begin{align}
\underbrace{I(s_0)T(s_0,s_1)}_{\text{$A(s_0,s_1)$}}\underbrace{T(s_1,s_2)\ldots T(s_{k-1},s_k)\neg P(s_k)}_{\text{$B(s_1,s_2,\ldots,s_k)$}}
\label{eq:BMC}
\end{align}
If this formula is satisfiable, a violation is found and we conclude that the system does not fulfill the safety property.
Otherwise, instead of simply increasing the unrolling upper bound, IMC tries to prove the safety property from the unsatisfiable BMC query as its second phase.
According to Craig's interpolation theorem, there exists an interpolant $C(s_1)$ referring to the common variable $s_1$, such that the following two conditions hold:
\begin{align*}
& I(s_0)T(s_0,s_1) \rightarrow C(s_1)\text{ ~and} \\
& C(s_1)T(s_1,s_2)\ldots T(s_{k-1},s_k)\neg P(s_k)\text{ is unsatisfiable.}
\end{align*}
The above two conditions indicate that $C(s_1)$ is an overapproximation of the set of states reachable from the initial states with one transition, and that states in $C(s_1)$ will not violate the safety property after $(k-1)$ transitions.
An overapproximation of the set of reachable states can be built by iteratively computing these interpolants.
Suppose the interpolant contains some noninitial states.
Changing the variable used in the interpolant from $s_1$ to $s_0$, another BMC query starting from the interpolant can be posed with $I(s_0)$ replaced by $C(s_0)$:
\begin{align*}
\underbrace{C(s_0)T(s_0,s_1)}_{\text{$A'(s_0,s_1)$}}\underbrace{T(s_1,s_2)\ldots T(s_{k-1},s_k)\neg P(s_k)}_{\text{$B'(s_1,s_2,\ldots,s_k)$}}
\end{align*}
If the formula is again unsatisfiable, another interpolant $C'(s_1)$ exists, which is an overapproximation of the set of states reachable from the initial states with two transitions.
Such computation is repeated until the newly derived interpolant is contained in the union of the initial states and all previous interpolants.
In other words, the procedure stops when the union of the initial states and all previous interpolants grows to a \textit{fixed point}, i.e., an inductive set of states with respect to the transition relation, which contains all reachable states.
From the second condition of Craig's interpolation theorem, it is guaranteed that this fixed point implies the safety property, and hence the safety property is proved.
If any BMC query is satisfiable during the iteration in the second phase, we cannot conclude that the property is violated.
The violation could be a wrong alarm, as some starting states in the interpolants might not be reachable.
Therefore, we have to return back to the first phase, increase the unrolling upper bound, and precisely check the existence of a violation starting from the initial states.
\subsubsection{Towards an Efficient Adoption}
While IMC is described in terms of logical formulas in the above discussion, the adoption of this algorithm to a concrete state-transition system, such as a sequential Boolean-logic circuit (hardware) or a program (software), requires a conversion from the system under verification to the three predicates $I(s)$, $T(s,s')$, and $P(s)$.
The conversion is simple for sequential Boolean-logic circuits which IMC originally focused on, as the input wires to the registers of the circuit encode the function to compute the next state (i.e., the state after transition) via the downstream circuitry in terms of the output wires of the registers (i.e., the current state).
This state-transition function can be naturally expressed as a transition relation.
It is less straightforward, by contrast, to extract a transition relation from a program.
Although a brute-force conversion is available, representing a program via a transition relation with the program counter often destroys the control-flow structure, and in turn decreases the computational efficiency of verification.
The main challenge towards an efficient adoption to software verification thus lies in obtaining all required predicates without blowing up the program's structure.
\subsection{Program Representation}
To facilitate the subsequent discussion of program analysis, here we provide some fundamental definitions for program representation from the literature~\cite{BLAST,HBMC-dataflow}.
We consider an imperative programming language whose variables are all integers.
The operations are either variable assignment or Boolean-expression evaluation.
A \textit{control-flow automaton} (CFA) is a representation of such a program.
It is a directed graph with nodes being program locations and edges annotated with program operations.
A \textit{reachability-safety verification task} consists of a CFA and an error location of the CFA.
The task is to either prove that the error location is unreachable from the initial location, or find a feasible error path to the error location otherwise.
For instance, the CFA of the example C program in~\cref{fig:even-code} is shown in~\cref{fig:even-cfa}.
The initial location of this CFA is $l_3$, and the error location is $l_8$.
\subsection{Configurable Program Analysis}
\label{sec:cpa}
A configurable program analysis (CPA)~\cite{CPA,CPAplus} defines the abstract domain used for a program analysis.
As we implemented the proposed adoption of IMC in the framework \cpachecker, which utilizes CPA as the core concept, we provide necessary background knowledge about CPA as follows.
For the rigorous definition of a CPA as well as further details,
please refer to the literature~\cite{AlgorithmComparison-JAR,HBMC-dataflow}.
\subsubsection{Fundamental CPAs}
Several fundamental CPAs are used in this paper:
The \emph{Location CPA}~$\loccpa$~\cite{CPAplus} uses a flat lattice over all program locations to track the program counter explicitly;
the \emph{Loop-Bound CPA}~$\boundscpa$~\cite{kInduction,AlgorithmComparison-JAR} tracks in its abstract states for every loop of the program how often the loop body has been traversed on the current program path;
the \emph{ARG CPA}~$\argcpa$ stores the predecessor-successor relationship between abstract states to track the abstract reachability graph~(ARG).
Another important CPA, namely the \emph{Predicate~CPA}~$\predcpa$~\cite{AlgorithmComparison-JAR}, will be used in~\cref{sect:approach-impl}.
\subsubsection{Composite CPA}
Several CPAs can be combined using a \emph{Composite CPA}~\cite{CPA} to achieve synergy.
The abstract states of the Composite CPA are tuples of one abstract state from each component CPA
and the operators of the Composite CPA delegate to the component CPAs' operators accordingly.
\subsubsection{CPA Algorithm}
CPAs can be used by the CPA~algorithm~\cite{AlgorithmComparison-JAR,HBMC-dataflow,CPA} for reachability analysis, which gets as input a CPA and an initial abstract state.
The algorithm performs a classic fixed-point iteration by looping until all abstract states have been completely processed and returns the set of reachable abstract states.
\section{Conclusion}
\label{sect:conclusion}
Software verification is a hard problem, and it is imperative to leverage
as much knowledge of the verification community as possible.
{Interpolation-based model checking}\xspace (McMillan, 2003 \cite{McMillanCraig}) is a successful hardware-verification algorithm,
but in contrast to many other interpolation-based verification approaches
this algorithm was not yet adopted to software verification
and the characteristics of the algorithm when applied to software systems were unknown.
This paper presents the first theoretical adoption
and practical implementation of the algorithm for software verification,
and provides a base-line for other researchers to build on.
Surprisingly, it has taken 19~years to close this significant gap of knowledge
by investigating the applicability to software verification.
We present the novel idea of utilizing
the well established technique large-block encoding
to extract transition relations
from programs while preserving the control-flow structures.
The proposed adoption was implemented in the open-source software-verification framework \cpachecker and
evaluated against other state-of-the-art software-verification algorithms
on a large benchmark set of C~verification tasks for reachability properties.
Among the competing approaches, our implementation achieved a comparable performance, evaluated in terms of both effectiveness (the number of correctly solved tasks) and efficiency (the CPU time to solve tasks).
Our IMC implementation was the most effective and efficient interpolation-based approach in the evaluation.
Furthermore,
the new approach was able to solve 6~programs for which all other approaches ran out of resources (\SI{15}{min} CPU time or \SI{15}{GB} memory usage),
which shows that the new approach \emph{improves} the state of the art and \emph{complements} the other approaches.
We hope that our promising results stimulate other researchers to further
improve the approach for software verification,
and that our open-source implementation in the flexible framework \cpachecker helps
other researchers to understand the details of the algorithm.
\section{Evaluation}
\label{sect:evaluation}
In order to answer the research questions posed in~\cref{sect:intro-research-question}, namely,
\begin{itemize}
\item \textbf{RQ1}: How to efficiently adopt IMC to software verification?
\item \textbf{RQ2}: How does the software adoption of IMC compare with the state of the art?
\end{itemize}
we evaluated the proposed adoption of IMC against several state-of-the-art algorithms,
including BMC~\cite{BMC},
\kinduction~\cite{kInduction},
predicate abstraction~\cite{GrafSaidi97},
\impact~\cite{IMPACT},
and PDR~\cite{CTIGAR},
over the largest publicly available benchmark suite of C safety-verification tasks~\cite{SVCOMP22}.
All of the compared approaches are implemented in \cpachecker.
The implementations of BMC, \kinduction, predicate abstraction, and \impact are built on top of the \cpapa algorithm in a unified manner~\cite{AlgorithmComparison-JAR}.
The implementation of PDR in \cpachecker follows a software-verification adaption named \ctigar~\cite{CTIGAR},
which was compared against other PDR-related approaches recently in the literature~\cite{PDR}.
We did not include other state-of-the-art verifiers in this paper for space reasons
(and because the results are available already from SV-COMP\,'22~\cite{SVCOMP22})
and to limit confounding variables to a minimum
(same parser, same libraries, same SMT solver, etc.).
We chose \cpachecker as platform because it is a flexible framework
that performed well in the competitions.
\subsection{Benchmark Set}
As the benchmark set we used the verification tasks from the 2022 Competition on Software Verification (SV-COMP\,'22)~\cite{SVCOMP22-SVBENCHMARKS-artifact}.
We used only verification tasks where the safety property is the reachability of a program location.
From those,
we further excluded verification tasks that are not supported by at least one of the compared approaches,
e.g., those from the categories
\textit{ReachSafety-Recursive} and
\textit{ConcurrencySafety-Main}.
\iffalse
Furthermore, because our implementation of IMC
relies on the single-loop transformation of \cpachecker,
we removed all
tasks for which this transformation fails.
\fi
The resulting benchmark set consists of a total of~\num{6027}~verification tasks from
the subcategories
\textit{AWS-C-Common-ReachSafety},
\textit{DeviceDriversLinux64-ReachSafety},
\textit{DeviceDriversLinux64Large-ReachSafety}, and
\textit{uthash-ReachSafety}
of the category \textit{SoftwareSystems}
and from the following subcategories of the category \textit{ReachSafety}:
\textit{Arrays},
\textit{Bitvectors},
\textit{ControlFlow},
\textit{ECA},
\textit{Floats},
\textit{Heap},
\textit{Loops},
\textit{ProductLines},
\textit{Sequentialized},
\textit{XCSP}, and
\textit{Combinations}.
A total of~\num{1793}~tasks in the benchmark set contain a known specification violation,
while the other~\num{4234}~tasks are assumed to be correct.
\subsection{Experimental Setup}
Our experiments were performed on machines with
one 3.4\,GHz CPU (Intel Xeon E3-1230~v5) with 8~processing units and 33\,GB of RAM each.
The operating system was Ubuntu~20.04 (64~bit),
using Linux~5.4 and OpenJDK~11.0.8.
Each verification task was limited to
two CPU cores,
a CPU time of \SI{15}{min},
and a memory usage of \SI{15}{GB}.
We used \benchexec\footnote{\url{\benchexecurl}}~\cite{Benchmarking-STTT} to achieve reliable benchmarking
and revision \num{40806} of branch \texttt{cfa-single-loop-transformation}\xspace of \cpachecker for evaluation.
We configured \cpachecker to use the SMT theories
of equality with uninterpreted functions, bit vectors, floats, and arrays.
All SMT queries were handled by \mathsat~\cite{MATHSAT5}.
\input{\plotpath/data-commands}
\begin{table*}[t]
\centering
\caption{Summary of the results for~\num{6027}~reachability-safety verification tasks}
\label{tab:results}
\newcommand\stoh[1]{\fpeval{#1/3600}}
\newcommand\precnum[1]{\tablenum[round-mode=off,table-format=4]{#1}}
\newcommand\rndnum[1]{\tablenum[round-mode=figures,round-precision=2,table-format=4.1]{#1}}
\begin{tabular}{lccc@{\hspace{3.5mm}}c@{\hspace{3.5mm}}c@{\hspace{3.5mm}}c}
\toprule
Algorithm & {IMC} & {PDR} & {BMC} & {\kInduction} & {Predicate Abstraction} & {\impact} \\
\midrule
Correct results & \precnum{\JarFullInterpolationModelCheckingImcCorrectCount}
& \precnum{\JarFullInterpolationModelCheckingPdrCorrectCount}
& \precnum{\JarFullInterpolationModelCheckingBmcCorrectCount}
& \precnum{\JarFullInterpolationModelCheckingKInductionCorrectCount}
& \precnum{\JarFullInterpolationModelCheckingPredicateAbstractionCorrectCount}
& \precnum{\JarFullInterpolationModelCheckingImpactCorrectCount} \\
\quad proofs & \precnum{\JarFullInterpolationModelCheckingImcCorrectTrueCount}
& \precnum{\JarFullInterpolationModelCheckingPdrCorrectTrueCount}
& \precnum{\JarFullInterpolationModelCheckingBmcCorrectTrueCount}
& \precnum{\JarFullInterpolationModelCheckingKInductionCorrectTrueCount}
& \precnum{\JarFullInterpolationModelCheckingPredicateAbstractionCorrectTrueCount}
& \precnum{\JarFullInterpolationModelCheckingImpactCorrectTrueCount} \\
\quad alarms & \precnum{\JarFullInterpolationModelCheckingImcCorrectFalseCount}
& \precnum{\JarFullInterpolationModelCheckingPdrCorrectFalseCount}
& \precnum{\JarFullInterpolationModelCheckingBmcCorrectFalseCount}
& \precnum{\JarFullInterpolationModelCheckingKInductionCorrectFalseCount}
& \precnum{\JarFullInterpolationModelCheckingPredicateAbstractionCorrectFalseCount}
& \precnum{\JarFullInterpolationModelCheckingImpactCorrectFalseCount} \\
Wrong proofs & \precnum{\JarFullInterpolationModelCheckingImcWrongTrueCount}
& \precnum{\JarFullInterpolationModelCheckingPdrWrongTrueCount}
& \precnum{\JarFullInterpolationModelCheckingBmcWrongTrueCount}
& \precnum{\JarFullInterpolationModelCheckingKInductionWrongTrueCount}
& \precnum{\JarFullInterpolationModelCheckingPredicateAbstractionWrongTrueCount}
& \precnum{\JarFullInterpolationModelCheckingImpactWrongTrueCount} \\
Wrong alarms & \precnum{\JarFullInterpolationModelCheckingImcWrongFalseCount}
& \precnum{\JarFullInterpolationModelCheckingPdrWrongFalseCount}
& \precnum{\JarFullInterpolationModelCheckingBmcWrongFalseCount}
& \precnum{\JarFullInterpolationModelCheckingKInductionWrongFalseCount}
& \precnum{\JarFullInterpolationModelCheckingPredicateAbstractionWrongFalseCount}
& \precnum{\JarFullInterpolationModelCheckingImpactWrongFalseCount} \\
Timeouts & \precnum{\JarFullInterpolationModelCheckingImcErrorTimeoutCount}
& \precnum{\JarFullInterpolationModelCheckingPdrErrorTimeoutCount}
& \precnum{\JarFullInterpolationModelCheckingBmcErrorTimeoutCount}
& \precnum{\JarFullInterpolationModelCheckingKInductionErrorTimeoutCount}
& \precnum{\JarFullInterpolationModelCheckingPredicateAbstractionErrorTimeoutCount}
& \precnum{\JarFullInterpolationModelCheckingImpactErrorTimeoutCount} \\
Out of memory & \precnum{\JarFullInterpolationModelCheckingImcErrorOutOfMemoryCount}
& \precnum{\JarFullInterpolationModelCheckingPdrErrorOutOfMemoryCount}
& \precnum{\JarFullInterpolationModelCheckingBmcErrorOutOfMemoryCount}
& \precnum{\JarFullInterpolationModelCheckingKInductionErrorOutOfMemoryCount}
& \precnum{\JarFullInterpolationModelCheckingPredicateAbstractionErrorOutOfMemoryCount}
& \precnum{\JarFullInterpolationModelCheckingImpactErrorOutOfMemoryCount} \\
Other inconclusive & \precnum{\JarFullInterpolationModelCheckingImcErrorOtherInconclusiveCount}
& \precnum{\JarFullInterpolationModelCheckingPdrErrorOtherInconclusiveCount}
& \precnum{\JarFullInterpolationModelCheckingBmcErrorOtherInconclusiveCount}
& \precnum{\JarFullInterpolationModelCheckingKInductionErrorOtherInconclusiveCount}
& \precnum{\JarFullInterpolationModelCheckingPredicateAbstractionErrorOtherInconclusiveCount}
& \precnum{\JarFullInterpolationModelCheckingImpactErrorOtherInconclusiveCount} \\
\bottomrule
\end{tabular}
\end{table*}
\subsection{Results}
The experimental results of all compared approaches are summarized in~\cref{tab:results}.
Observe that IMC produced the most correct results for both proofs and alarms
among the interpolation-based approaches (IMC, PDR, predicate abstraction, and \impact),
and was second only to \kinduction in the evaluation.
Furthermore, it uniquely solved 6~programs for which all other approaches ran out of resources.
In comparison to the most-related approach \impact,
IMC proved the safety of 320~more programs and found 29~more bugs (an increase of 21\,\% and 3.3\,\%, respectively).
Meanwhile, BMC generated the most correct alarms as expected,
and \kinduction correctly solved the most tasks, with the most correct proofs and the second-most correct alarms.
Moreover, although IMC is a new addition to \cpachecker,
it did not produce any wrong proof in the evaluation,
identical to the other long-established approaches in the software-verification framework.
We consider the 3~wrong alarms of IMC not caused by our implementation.
They are related to the program encoding of~\cpachecker,
and other approaches, such as predicate abstraction, also failed to solve these tasks correctly.
The quantile plots for the correct proofs and alarms of the compared approaches are shown in~\cref{fig:evaluation:quantile-true} and~\cref{fig:evaluation:quantile-false}, respectively.
A data point $(x,y)$ in the plots indicates that there are $x$ tasks correctly solved by the respective algorithm within a CPU time of $y$~seconds each.
Note that IMC is not only effective in producing proofs and alarms but is also efficient.
From~\cref{fig:evaluation:quantile-cputime},
we see that it is the most efficient and effective interpolation-based approach in the evaluation.
\input{\plotpath/plot-defs}
\newsavebox{\quantileTrue}
\begin{lrbox}{\quantileTrue}
\begin{minipage}[b]{0.5\textwidth}
\scalebox{0.85}{\input{\plotpath/quantile-true}}
\end{minipage}
\end{lrbox}
\newsavebox{\quantileFalse}
\begin{lrbox}{\quantileFalse}
\begin{minipage}[b]{0.5\textwidth}
\scalebox{0.85}{\input{\plotpath/quantile-false}}
\end{minipage}
\end{lrbox}
\begin{figure*}[t]
\centering
\subfloat[Proofs]{\usebox{\quantileTrue}\label{fig:evaluation:quantile-true}}
\subfloat[Alarms]{\usebox{\quantileFalse}\label{fig:evaluation:quantile-false}}\\
\bigskip
\begin{minipage}[c]{.55\linewidth}
\pgfplotslegendfromname{legend:quantile-true}
\end{minipage}
\hfill
\begin{minipage}[c]{.37\linewidth}
$x$-Axis: n-th fastest correct result\\
$y$-Axis: CPU time (\second)
\end{minipage}
\caption{Quantile plots for all correct proofs and alarms}
\label{fig:evaluation:quantile-cputime}
\end{figure*}
\newsavebox{\scatterPdr}
\begin{lrbox}{\scatterPdr}
\begin{minipage}[b]{0.5\textwidth}
\scalebox{0.8}{\input{\plotpath/scatter-pdr}}
\end{minipage}
\end{lrbox}
\newsavebox{\scatterKidf}
\begin{lrbox}{\scatterKidf}
\begin{minipage}[b]{0.5\textwidth}
\scalebox{0.8}{\input{\plotpath/scatter-ki-df}}
\end{minipage}
\end{lrbox}
\newsavebox{\scatterPredicate}
\begin{lrbox}{\scatterPredicate}
\begin{minipage}[b]{0.5\textwidth}
\scalebox{0.8}{\input{\plotpath/scatter-predicate}}
\end{minipage}
\end{lrbox}
\newsavebox{\scatterImpact}
\begin{lrbox}{\scatterImpact}
\begin{minipage}[b]{0.5\textwidth}
\scalebox{0.8}{\input{\plotpath/scatter-predicate-impact}}
\end{minipage}
\end{lrbox}
\begin{figure*}[t]
\centering
\subfloat[PDR]{\usebox{\scatterPdr}\label{fig:evaluation:scatter-pdr}}
\subfloat[\kInduction]{\usebox{\scatterKidf}\label{fig:evaluation:scatter-ki}}\\
\subfloat[Predicate Abstraction]{\usebox{\scatterPredicate}\label{fig:evaluation:scatter-predicate}}
\subfloat[\impact]{\usebox{\scatterImpact}\label{fig:evaluation:scatter-impact}}
\caption{Scatter plots of CPU time in seconds for all correct results with IMC in $y$-axis and compared approaches in $x$-axis}
\label{fig:evaluation:scatter-cputime}
\end{figure*}
The scatter plots for the correctly solved tasks (including both proofs and alarms) of the compared approaches are shown in~\cref{fig:evaluation:scatter-cputime}.
We omitted the scatter plot for BMC as it is mainly inclined to bug hunting, while other approaches have more balanced behavior.
A data point $(x,y)$ in the plots indicates that there is a task correctly solved by both IMC and a compared approach,
while IMC took a CPU time of $y$~seconds and the other approach took a CPU time of $x$~seconds.
Observe that IMC is often more efficient than a compared approach.
For example, while it solved less tasks compared to \kinduction, its time efficiency is often better than \kinduction on the tasks which can be solved by both algorithms.
This phenomenon could be explained by the fact that, unlike \kinduction,
which relies on an external procedure to generate auxiliary invariants,
IMC generates interpolants from BMC queries and uses them to construct fixed points purely internally.
Moreover, in our evaluation, there were \num{122} programs that IMC solved, but \kinduction ran out of time or memory.
These programs often have a loop for which \textit{helpful} loop invariants, i.e., those that help to prove the property, require complex formulas.
IMC is better in such cases.
The default configuration of \kinduction in \cpachecker uses an interval-based data-flow analysis to generate candidate invariants whose expressiveness is limited.
By contrast, IMC constructs a candidate fixed point (also a loop invariant) as a union of previously derived interpolants, which in principle can express any combination of reachable states.
The best subcategory for
IMC was category \textit{ReachSafety-ECA}.
These event-condition-action (ECA) programs have a loop to receive external inputs,
generate outputs, and update internal variables based on the ECA rules.
Inside the loop is a complex control flow to implement the ECA rules.
Overall, the working of these programs is similar to that of sequential Boolean-logic circuits.
IMC naturally performs well on them because it originated from hardware verification.
Out of a total \num{1265} ECA programs, IMC solved a second most \num{561} programs,
while predicate abstraction, \impact, and \kinduction solved \num{476}, \num{555}, and \num{578} tasks, respectively.
\inlineheadingbf{Answers to the Research Questions}
Regarding RQ1, we conclude that the proposed approach with large-block encoding is an efficient adoption of IMC to software verification.
The conclusion is well supported by the experimental results,
because our IMC implementation not only solved the second most verification tasks (\cref{tab:results}) in the evaluation,
but was also efficient compared to other state-of-the-art approaches (\cref{fig:evaluation:scatter-cputime}).
In our experiments, it was the most efficient and effective interpolation-based approach (\cref{fig:evaluation:quantile-cputime}).
Note that the efficiency and effectiveness might not be achieved without large-block encoding,
because the control-flow structure would be destroyed
if one simply converts a program into a transition relation with the program counter to perform IMC.
The empirical results demonstrate the unique value of using large-block encoding to adopt the IMC algorithm.
As for RQ2, we conclude that IMC,
the first interpolation-based formal-verification approach ever invented,
is competitive against the other state-of-the-art algorithms,
which have been investigated much more by the research community.
It has great potential to improve the state of the art of software verification and thus deserves more attention.
Our experimental results show that IMC is not only comparable but also complements other approaches,
as it uniquely solved 6~verification tasks for which all other algorithms ran out of resources.
\subsection{Threats to Validity}
Here we discuss some threats that may affect the validity of our conclusions and how we reduced them.
To enhance internal validity, all the compared algorithms are implemented in~\cpachecker.
This practice minimizes the confounding variables (front ends and utilities) and rules out differences unrelated to the algorithms.
We also use \benchexec~\cite{Benchmarking-STTT} to ensure best possible measurement accuracy.
To reduce the external threat resulting from the selection bias of verification tasks,
we conduct the experiments using the largest publicly available benchmark set of C safety-verification tasks.
Other external threats arise from the selection of the compared approaches and underlying framework.
It is clear from the literature~\cite{AlgorithmComparison-JAR} that the compared approaches in this paper indeed represent the state of the art of software verification;
the only missing main related state-of-the-art approach is trace abstraction~\cite{UAUTOMIZER2013},
for which the implementation in \cpachecker is not yet mature enough.
Moreover, the chosen platform~\cpachecker is a well-maintained software project that performs well in the competitions, and the relative performance between~\cpachecker and other verifiers is available from SV-COMP\,'22~\cite{SVCOMP22}.
\section{Introduction}
Automatic software verification~\cite{SoftwareModelChecking} is an active research field in which automated solutions of the following problem are studied:
Given a program and a specification, decide whether the program satisfies the specification or not.
In this paper, we focus on the verification of reachability-safety properties, asserting that some error location in the program should never be reached by the control flow.
Other specifications, including termination, memory safety, concurrency safety, and overflows, are also investigated in the literature.
Although the problem of software verification is in general undecidable,
many important concepts, including various predicate-abstraction techniques~\cite{GrafSaidi97,FlanaganQadeer02,LazyAbstraction,AutomaticPredicateAbstraction2001},
counterexample-guided abstraction refinement~(CEGAR)~\cite{ClarkeCEGAR},
large-block encoding~\cite{LBE,ABE},
interpolation~\cite{AbstractionsFromProofs,IMPACT},
together with the advances in SMT solving~\cite{HBMC-SMT} and combinations with data-flow analysis~\cite{HBMC-dataflow},
make it feasible to verify industry-scale software such as operating-systems code~\cite{LDV,LDV12,SLAM,SLAMtransfer,INFER,AWS}.
To illustrate the reachability-safety verification of a program,
consider the C~program in~\cref{fig:even-code}.
The program first initializes variable~\texttt{x} to $0$, and keeps incrementing \texttt{x} by $2$ while the nondeterministic value returned from function~\texttt{nondet()} is nonzero.
Once the nondeterministic value equals zero, the control flow exits the loop and tests whether variable~\texttt{x} is odd.
If variable~\texttt{x} is odd, the control flow reaches the error location at line~\texttt{8}; otherwise the program terminates without errors.
The goal of the reachability-safety verification is to either prove that the error location is unreachable by the control flow, or find an execution path of the program reaching the error location.
\newsavebox{\exampleCode}
\begin{lrbox}{\exampleCode}
\begin{minipage}[b]{0.48\textwidth}
\centering
\lstinputlisting[
style=C,
basicstyle=\ttfamilywithbold,
numberstyle=\scriptsize,
lineskip=2pt,
aboveskip=0pt,
belowskip=0pt,
]{figures/even.c}
\end{minipage}
\end{lrbox}
\newsavebox{\exampleCFA}
\begin{lrbox}{\exampleCFA}
\begin{minipage}[b]{0.48\textwidth}
\centering
\scalebox{0.9}{\input{figures/even-cfa.tex}}
\end{minipage}
\end{lrbox}
\begin{figure*}[t]
\centering
\subfloat[C program]{\usebox{\exampleCode}\label{fig:even-code}}
\subfloat[Control-flow automaton]{\usebox{\exampleCFA}\label{fig:even-cfa}}
\caption{
An example C program (a) and its CFA (b)
(adopted from \href{https://gitlab.com/sosy-lab/benchmarking/sv-benchmarks/-/blob/svcomp22/c/loop-invariants/even.c}{\texttt{loop-invariants/even.c}} of the benchmark set of the 2022 Competition on Software Verification (SV-COMP\,'22)~\cite{SVCOMP22})
}
\label{fig:example}
\end{figure*}
As the verification of finite-state transition systems and infinite-state transition systems share much similarity,
some classic model-checking algorithms for software (infinite-state systems), such as bounded model checking (BMC)~\cite{BMC,CBMC} or \kinduction~\cite{kInduction,k-Induction,PKind},
were originally developed for hardware (finite-state systems).
A well-known example of such technology transfer is property-directed reachability (\pdr)~\cite{IC3}.
After it obtained huge success in hardware model checking,
many research efforts have been invested for its software-verification adoption~\cite{SoftwareIC3,IC3-CFA,CTIGAR,PDR-kInduction,PDR}.
\subsection{Interpolation-Based Verification Approaches}
\begin{figure*}[t]
\centering
\scalebox{0.8}{\input{figures/overview-tree}}
\caption{Classification of different abstractions using Craig interpolation}
\label{fig:classification}
\end{figure*}
McMillan's algorithm~\cite{McMillanCraig} from 2003 is another state-of-the-art approach for hardware model checking, prior to the invention of \pdr.
It utilizes Craig interpolation~\cite{Craig57} to derive interpolants from unsatisfiable \bmc queries, and computes an overapproximation of the set of reachable states as the union of the interpolants.
Its idea of abstracting objects with interpolants has been extended beyond state sets and underpinned various interpolation-based verification approaches and tools.
Abstractions of
transition relations~\cite{TRApproximation},
traces~\cite{TraceAbstraction},
predicates over program variables~\cite{AbstractionsFromProofs},
and function calls~\cite{SeryFunFrog11}
have been studied in the literature.
We classify in~\cref{fig:classification} different usages
of Craig interpolation and highlight several important algorithms regarding state-set abstraction.
An overview of several representative interpolation-based formal-verification approaches is provided in~\cref{sect:related-work}.
Despite its success in hardware model checking and profound theoretical impact on program analysis,
McMillan's algorithm~\cite{McMillanCraig} from 2003 has not been investigated for software verification.
We emphasize that
\emph{McMillan's interpolation-based algorithm for model checking from 2003 should not be mistaken for other interpolation-based verification approaches}.
In the following, we refer to the algorithm proposed by McMillan from 2003 as \emph{{interpolation-based model checking}\xspace} and abbreviate it as IMC.
One potential concern to apply IMC to software,
raised by its inventor McMillan in his later paper~\cite{IMPACT} presenting the algorithm \impact,
was the scalability of the underlying decision procedure to handle the entire unrolled program.
Compared to IMC, \impact derives interpolants only for individual execution paths, reducing the workload of the solver.
Fortunately, due to the advancements in SMT solving,
delegating formulas encoding the entire unrolled program to the solvers has become feasible.
Therefore, it is time to revisit IMC and evaluate its performance against the state of the art.
Other SMT-based approaches have been thoroughly compared already in the literature~\cite{AlgorithmComparison-JAR}.
\subsection{Our Research Questions and Contributions}
\label{sect:intro-research-question}
In this paper, we explore the applicability of the IMC algorithm to software verification.
Specifically, we answer the following two research questions.
First, we investigate \emph{how to efficiently adopt IMC to software verification}.
As mentioned earlier, IMC was originally invented to verify sequential Boolean-logic circuits (hardware),
whose transition relations, required to perform IMC, are easy to derive:
The downstream circuitry of the memory elements (i.e., registers) encodes the next-state function of the system,
which can then be naturally expressed as a transition relation between system states.
It is less straightforward, by contrast, to extract a transition relation from a program (software).
Although representing a program as a transition relation with the program counter is possible,
such conversion often blows up the control-flow structure,
and in turn influences the computational efficiency of verification.
To address this research question,
we propose an efficient software adoption of IMC via large-block encoding,
exploring the analogy between the execution paths of a sequential Boolean-logic circuit and a program.
We also present the first implementation of the IMC algorithm for software verification,
and make it available in the open-source framework \cpachecker.
The details of the proposed adoption and implementation will be discussed in~\cref{sect:approach}.
Our second research question focuses on \emph{evaluating the performance of the IMC adoption against the state of the art}.
To address this research question,
we compare the proposed implementation against other state-of-the-art software-verification algorithms,
including \pdr, BMC, \kinduction, predicate abstraction, and \impact,
over the largest benchmark suite of C safety-verification tasks in~\cref{sect:evaluation}.
Our experimental results show that IMC is competitive in terms of both effectiveness (the number of solved tasks)
and efficiency (the elapsed CPU time).
\inlineheadingbf{Novelty}
(1)~This paper closes the 19~years old gap of knowledge by investigating the applicability of IMC to software verification.
We analyze the characteristics of IMC in the context of software verification,
and our empirical evaluation indicates its competitiveness against the state-of-the-art approaches.
(2)~Our replication of the IMC algorithm as open-source implementation broadens
the spectrum of available software-verification techniques,
which might have important implications in practice,
because researchers, developers, and tool users now have more choices at their disposal.
(3)~While the application of large-block encoding to program analysis has a long history,
to the best of our knowledge, using large-block encoding to represent an algorithm originated from a different research community for software is a new idea, which may shed light on the efficient adoptions of other algorithms.
\inlineheadingbf{Significance}
IMC is an important verification algorithm in hardware verification.
It is a risk to leave the potential of it unleveraged for the verification of software.
Therefore, we believe that the knowledge about the algorithm's adoption to software is
a significant improvement of the state of the art and has the potential to inspire other
works in the area of software verification.
\inlineheadingbf{Correctness}
We show the correctness of our algorithms in~\cref{thm:imc-soundness}.
Our implementation is based on components from the \cpachecker framework,
which is a well-maintained software project with lots of evidence that the components work well.
Large-block encoding is a sound component from the literature.
The effectiveness and efficiency of our implementation is empirically evaluated with
experiments on a large benchmark set.
We discuss possible threats to validity that might affect the soundness of our conclusions from the experimental results.
\section*{Data-Availability Statement}
To enhance the verifiability and transparency of the results reported in this paper,
all used software, input programs, and raw experimental results are
available in a supplemental reproduction package~\cite{IMC-artifact-JAR-submission}.
For convenient browsing through the results,
interactive tables are available at \url{https://www.sosy-lab.org/research/cpa-imc}.
Current versions of \cpachecker are also available at \url{https://cpachecker.sosy-lab.org}.
\interlinepenalty=10000
\section{Related Work}
\label{sect:related-work}
IMC has popularized the idea of using interpolation for verification,
and although IMC itself has not been applied to software so far,
there are many approaches for software verification that make use of interpolation.
Based on the classification in~\cref{fig:classification},
we will discuss several representative interpolation-based approaches and tools,
as summarized in~\cref{tab:related-work}.
Interested readers are referred to the chapter~\cite{HBMC-interpolation} by McMillan
in the Handbook of Model Checking for a broader survey.
Of course, there exist many techniques for computing interpolants.
We do not discuss them here as interpolant computation is typically orthogonal
to the used verification algorithm.
In our implementation, we use an off-the-shelf SMT solver for interpolation (\mathsat~\cite{MATHSAT5}).
\begin{table*}[t]
\centering
\scriptsize
\caption{Important interpolation-based formal-verification approaches and tools}
\label{tab:related-work}
\vspace{-2mm}
\newcommand\newapproach{\\[0.1cm]}
\begin{tabular}{l|ccl}
\toprule
Approach & Year & Publication & Contribution \\
\midrule
IMC & 2003 & \cite{McMillanCraig} & First interpolation-based model-checking algorithm
\newapproach
\multirow{2}{*}{Predicate abstraction} & \multirow{2}{*}{2004} & \multirow{2}{*}{\cite{AbstractionsFromProofs}} & Discovering relevant predicates \\
& & & from interpolants to refute false alarms
\newapproach
\multirow{2}{*}{TR approximation} & \multirow{2}{*}{2005} & \multirow{2}{*}{\cite{TRApproximation}} & Refining an abstract TR with interpolants \\
& & & to avoid exact image computation
\newapproach
{\impact} & 2006 & \cite{IMPACT} & Combining lazy abstraction with interpolants
\newapproach
Slicing abstraction & 2007 & \cite{SlicingAbstractions} & Splitting abstract states with interpolants
\newapproach
\multirow{2}{*}{ISB} & \multirow{2}{*}{2009} & \multirow{2}{*}{\cite{VizelFMCAD09}} & Imitating BDD-based model checking by \\
& & & abstracting states with interpolants
\newapproach
\multirow{2}{*}{Trace abstraction} & \multirow{2}{*}{2009} & \multirow{2}{*}{\cite{TraceAbstraction}} & Refining an overapproximation of \\
& & & possible traces with interpolant automata
\newapproach
\multirow{2}{*}{Lazy annotation} & \multirow{2}{*}{2010} & \multirow{2}{*}{\cite{LazyAnnotation}} & Annotating a program with interpolants \\
& & & derived from Hoare triples
\newapproach
\multirow{2}{*}{Function summaries} & \multirow{2}{*}{2011} & \multirow{2}{*}{\cite{SeryFunFrog11}} & Summarizing function calls with interpolants \\
& & & to reduce future analysis effort
\newapproach
\multirow{2}{*}{Software PDR} & \multirow{2}{*}{2012} & \multirow{2}{*}{\cite{SoftwareIC3}} & Combining {\impact}-like proof-based interpolants \\
& & & and PDR clause generation
\newapproach
\multirow{2}{*}{\ctigar} & \multirow{2}{*}{2014} & \multirow{2}{*}{\cite{CTIGAR}} & Refining abstraction failures \\
& & & relative to single steps with interpolants \\
\midrule
{\blast} & 2004 & \cite{AbstractionsFromProofs} & First software model checker using interpolation
\newapproach
{\cpachecker} & 2004 & \cite{CPACHECKER} & Large-block encoding and interpolation
\newapproach
{\wolverine} & 2011 & \cite{Wolverine} & First public implementation of {\impact}
\newapproach
{\ufo} & 2012 & \cite{UFO} & Combining predicate and interpolation methods
\newapproach
{\duality} & 2013 & \cite{DUALITY} & Solving constrained Horn clauses with interpolation
\newapproach
{\spacer} & 2013 & \cite{SPACER} & Combining proof-based approaches and CEGAR
\newapproach
{\safari} & 2014 & \cite{LazyAbstractionWithArrays-JOURNAL} & Backward {\impact}-like analysis with arrays \\
\bottomrule
\end{tabular}
\vspace{-2mm}
\end{table*}
\subsection{State Sets}
The most closely related algorithm is \impact~\cite{IMPACT} from the same author,
which is also based on the idea of computing a fixed point from interpolants,
but applies interpolation to formulas of single program paths instead of the whole program,
generating one interpolant for each step of a spurious counterexample.
It also computes fixed points of reachable states per program location
instead of globally.
One adaption of property-directed reachability~(PDR) to software~\cite{SoftwareIC3}
computes sequences of sets of clauses for refuting spurious counterexamples,
and these sequences also form a valid sequence of interpolants.
Under this view, the approach is similar to \impact, only differing in how the interpolants are computed.
A hybrid approach with a combination of proof-based interpolation (as in \impact)
and PDR-based clause generation has also been suggested.
\ctigar~\cite{CTIGAR} is another attempt to extend PDR to software.
It combines Cartesian predicate abstraction with PDR and considers an abstract state as a conjunction of the predicates satisfied by the corresponding concrete state.
Different from other adaptions of PDR,
\ctigar avoids expensive pre-image computation by focusing on refinement relative to single steps of the transition relation.
A related approach for hardware model checking
is \emph{interpolation-sequence based model checking} (ISB)~\cite{VizelFMCAD09}.
In contrast to IMC, which computes only one interpolant at a time
that overapproximates states reachable within a certain number of steps,
ISB derives a sequence of interpolants from an unsatisfiable BMC query,
and each interpolant is an overapproximation of the states reachable within
an increasing number of steps.
This is similar to \impact,
just with ISB computing interpolant sequences for an unrolling of the whole transition relation
instead of single program paths like \impact.
In ISB, the fixed point is found if the interpolant derived at the last unrolled loop head
implies the disjunction of all previous interpolants.
The approach of \emph{lazy annotation}~\cite{LazyAnnotation} combines symbolic execution and interpolation to generate Hoare-style annotations for a program in a similar way as a conflict-driven clause-learning SAT solver.
An annotation on a program edge is a condition that will block any future execution from this edge to an error location.
The method symbolically executes the input program along some chosen path to search for an error location.
If the execution is blocked by an edge,
it backtracks and produces an annotation by interpolation,
which is a valid precondition of the edge's Hoare triple.
This method is also applicable to program testing, because it explores only feasible traces.
\subsection{Predicates, Traces, and Functions}
Another popular use of interpolation for software verification
is to derive predicates from interpolants for predicate abstraction~\cite{AbstractionsFromProofs}
in the refinement step of CEGAR,
typically by breaking up the interpolants into atomic predicates.
In contrast to IMC and \impact,
which both create the final abstract model of the program
(the overapproximation of the set of reachable states)
directly from interpolants,
predicate abstraction uses Boolean or Cartesian abstraction
over the set of derived predicates
and may generalize better.
Interpolation has also been used to avoid the expensive exact image computation in predicate abstraction~\cite{TRApproximation}, refining an abstract transition relation to guarantee convergence given adequate predicates.
Slicing abstraction~\cite{SlicingAbstractions} is another technique related to predicate abstraction.
It splits abstract states using predicates obtained from Craig interpolants to refine the abstraction.
Trace abstraction~\cite{TraceAbstraction,UAUTOMIZER2013} extends the concept of abstracting information by Craig interpolation to representing program paths with interpolants.
Given an unsatisfiable BMC query,
it derives a sequence of interpolants and constructs an interpolant automaton out of them.
This interpolant automaton excludes spurious traces that share the same reason of infeasibility with the current one.
A novel counterexample-guided abstraction refinement scheme is proposed for trace abstraction to prove the correctness of a program.
Interpolants are also applied to summarize function calls in a program~\cite{SeryFunFrog11}.
This approach replaces function calls with interpolants obtained in a previous analysis to reduce the subsequent verification effort.
Given an unsatisfiable BMC query involving a function call,
a summary of the function is computed as an interpolant between the function's corresponding formula and the rest of the BMC formula.
\subsection{Tools Based on Craig Interpolation}
Several software-verification tools are developed on top of Craig interpolation.
The tool \blast~\cite{AbstractionsFromProofs,BLAST} provides the first implementation
of a software-verification tool that uses interpolants for computing abstractions.
The verification framework \cpachecker~\cite{CPACHECKER} applies Craig interpolation to large-block encodings of program code.
The tool \wolverine~\cite{Wolverine} provides the first publicly available implementation of \impact,
featuring a built-in interpolation procedure and some support for bit-vector operations.
The framework \ufo~\cite{UFO} is parameterized by definable components of abstract post, refinement, and expansion,
allowing various verification techniques based on over- and under-approximation.
Craig interpolation has also been applied to solve constrained Horn clauses (CHC's).
The tool \duality~\cite{DUALITY} generalizes \impact to gradually unroll a program and solves the corresponding CHC's with interpolation until it yields valid inductive invariants.
The tool \spacer~\cite{SPACER} combines proof-based techniques with CEGAR,
maintaining both an overapproximation and an underapproximation of the input program.
The tool \safari~\cite{LazyAbstractionWithArrays-JOURNAL} implements a backward reachability analysis with lazy abstraction based on the MCMT framework~\cite{MCMT}, which can be understood as a backward variant of \impact, to support reasoning of arrays with unknown length.
|
\section{Introduction}
A precise control of the motion in narrow tracks of stable localized excitations is desirable to realize different devices of current interest in applied physics research.
Paradigmatic examples are magnetic domain walls ~\cite{Parkin2008,Emori2013,Blasing2020} or skyrmions \cite{Kiselev2011,Schulz2012,Fert2013,Reichhardt2021} moving in an array of magnetic nanowires.
Furthermore, such kind of devices may be also of interest for other driven objects embedded in different materials, such as ferroelectric domain walls~\cite{Catalan2012,Paruch2013}, vortices in superconductors~\cite{Kwok2016}, colloids~\cite{Lowen2008,Volpe2014,Bechinger2016}, general brownian particles~\cite{hanggi2009}, as well as kinks and solitons~\cite{Martinez2008,Laliena2020,Osorio2021}.
Although each one may have a different internal structure we will generically refer to all these driven objects as ``particles''.
One of the interesting applications proposed for the above mentioned devices is the ``racetrack memory''~\cite{Parkin2008}, a novel type of non-volatile memories where particles are used as carriers of information in an array of tracks. This proposal assumes the ability
to write, read and move the particles in the tracks. To achieve a competitive capacity and efficiency compared with existing memories, these tracks have been reduced to the nanoscale, making the consideration of (the usually unavoidable) quenched spatial heterogeneities in the host materials of prime importance.
Indeed, spatial inhomogeneities may have disruptive effects, mainly metastability, a depinning transition, and a shaking effect on the sliding particles. As a consequence, the motion of the driven particles becomes more difficult to predict, specially in the ubiquitous case of quenched disorder. Moreover, in the case of track arrays, each track has a different realization of the quenched disorder, so different tracks may produce a different particle trajectory under an identical protocol. This is also a very important issue if we aim to synchronize the motion of many particles in many independent tracks, such as in a ``memory bus'' and, for instance, perform computation with the whole bunch of data.
Although the effects of quenched disorder or quenched periodic potentials on driven particles have been extensively discussed in the literature, particularly focusing in the collective transport in two dimensions or higher
~\cite{Reichhardt2016,Reichhardt2021}, a detailed statistical study of the effect of quenched disorder on independent particles driven in an array of quasi one-dimensional racetracks is lacking.
In this paper we address the problem of independent particles driven in an array of disordered tracks with the aim to identify robust statistical features, independent of specific model details. To be concrete we focus in the specific ``obstacle race'' problem illustrated schematically in Fig.\ref{fig:dispersion}. At $t=0$ many particles start at $x=0$ in a track array, one per track, and are driven by the same constant force $f$. As tracks are independent realizations of an otherwise statistically identical quenched disorder with some assumed short-range correlations, at any given time $t>0$ particles disperse, even at zero temperature, provided the force exceeds the depinning threshold (assumed to be finite). We will be mainly interested in the mean velocity $v$ and in the dispersion constant $D$ around the mean forward motion of all the particles, as a function of $f$. Such race problem is similar, except some subtle differences we discuss, to the well studied problem of dispersion of tracers in porous media. However, the results turn out to be radically different: while in the latter the dispersion constant $D$ in general increases with increasing the average flow~\cite{BouchaudGeorges1990}, in the problem we address here it decreases with increasing the average flow $v$. We find, in particular, that $D$ has rather robust properties at large $f$, in the limit when $v\to f$, though quite sensitive to the nature of disorder. For the two physically important cases that we study here, the so-called Random-Bond (RB) and the Random-Field (RF) type, we find that $D\sim f^{-3}$ and $D\sim f^{-1}$ respectively.
These results are obtained from different toy models where the full velocity-force characteristic and dispersion can be analytically obtained.
More complicated models, motivated in different physically relevant situations are solved numerically and shown to display the same dispersion beyond a velocity crossover.
In section \ref{sec:mechanicalmodel} we introduce the main properties of interest in relation to a simple mechanical model. Then, in section \ref{sec:exactresults} we introduce several models that can be solved analytically for the properties of interest, $v$ and $D$. These examples illustrate various effects and in particular suggest a universal behaviour of the dispersion constant $D$. In section \ref{sec:trapmodel} we show how these results can be connected to those of a stochastic trap model in the context of flow in porous media.
In order to test the robustness of the different analytical results, in section \ref{sec:numerics} we compare them with the numerical solution of various physically relevant models for which an exact solution is difficult to obtain. Finally, in section \ref{sec:discussion} we discuss the results and summarize our conclusions.
\section{A mechanical toy model}
\label{sec:mechanicalmodel}
To be concrete let us consider a damped particle of mass $m$ driven by a constant uniform force $f$ in a one-dimensional space presenting quenched random forces at {\it zero temperature} (see section \ref{sec:langevin} for a discussion about temperature effects),
\begin{equation}
m \ddot x = {F}({x})+ {f} - \eta \dot {x},
\end{equation}
where $\eta$ is a friction constant and ${F}(x)$ is a short-range correlated quenched random force field such that
\begin{eqnarray}
\overline{{F}({x})}&=&0,\\
\overline{{F}({x}){F}({x'})}&=& f_0^2 g(|{x}-{x'}|/d_0).
\label{eq:dampedmechmodel}
\end{eqnarray}
Here $d_0$ is a characteristic length,
$f_0$ a characteristic force amplitude, and $g(u)$ a rapidly decaying function of unit range and $g(0)=1$. We will be particularly interested in the cases where $\int_x g(x)=0$, corresponding to the so-called ``Random-Bond'' case (RB), and the case where $\int_x g(x)>0$, corresponding to the so-called ``Random-Field'' (RF) case. If we write $F(x)=-U'(x)$ (which is always possible to do in $d=1$ but not necessarily in $d>1$), and the potential $U(x)$ is bounded, then $F(x)$ is RB type. If otherwise the potential diffuses as $\langle [U(x)-U(x')]^2 \rangle \sim |x-x'|$ for long $|x-x'|$, then it is of the RF type. The great physical relevance of these two particular cases is better appreciated by noting that the RB type can be generated by forces derived from a bounded short-range correlated random potential while the RF type can be generated by a short-range correlated random force-field.
In both cases we will also assume that $F(x)$ is bounded, so a finite critical depinning force $f_c = \max_x[-F(x)]$ exists,
such that only for $f>f_c$ a steady-state motion is generated.
\footnote{See Ref.\cite{LeDoussal2009}for a thorough and general study of the one particle depinning in different universality classes corresponding to the three extreme value statistics, Gumbel, Weibull, and Fréchet.}
Without any loss of generality
we can nondimensionalize the equation of motion Eq.(\ref{eq:dampedmechmodel}) by measuring distances in units of $d_0$, forces in units of $f_0$ and time in units of $\tau_0=\eta d_0 /f_0$, and mass in units of $m_0=\eta^2 d_0/f_0$. The derived velocity unit is therefore $v_0=d_0/\tau_0=f_0/\tau_0 \eta$.
Then, overriding notation for all nondimensionalized quantities we get
\begin{equation}
m \ddot x = {F}({x})+ {f} - \dot {x}.
\label{eq:equationwithinertia}
\end{equation}
For a given initial condition and disorder realization $F(x)$ the solution $x(t)$ of this equation are only parameterized by the nondimensional mass $m$.
We will be mainly interested in the statistics of the time-dependent fluctuations induced by quenched disorder in these solutions.
Given the position $x(t)$ of a particle as a function of time $t$ for many particular realizations of $F(x)$, we will be mainly interested in the average velocity $v$ and the dispersion constant $D$ in the commoving frame at long times, as a function of the driving force $f$ and in the steady-state regime. Assuming $x(t=0)=0$, they are defined as
\begin{eqnarray}
v &=& \lim_{t\to \infty} \langle x(t)\rangle/t
\label{eq:vdef} \\
D &=& \lim_{t\to \infty} \langle [x(t)-v t]^2 \rangle/t
\label{eq:Ddef}
\end{eqnarray}
where the $\langle \dots \rangle$ denotes average over disorder realizations
\footnote{Note that we are thus interested here in the so-called ``annealead'' dispersion constant, different than the ``quenched'' one that describes the spreading of a packet in a single channel due to thermal noise or chaotic behaviour at zero temperature~\cite{LeDoussal1989}.}.
\begin{figure}[htp]
\centering
\includegraphics[width=0.5\textwidth]{dispersion}
\caption{Schematics of the particle (filled circles) deterministic dispersion problem at finite force $f$ above the depinning threshold, induced by quenched disorder in an array of narrow disordered racetracks (grey bars). Tracks are different disorder realizations of statistically {\it identical} media (light-blue lines illustrate two force-fields $F(x)$ and $F'(x)$). Particles disperse around the average motion $vt$ in the the right direction, with $v$ the disorder-averaged or track-averaged velocity. The ``obstacle race'' starts at $t=0$, and for increasing times (indicated with different colors) the particle distribution is shown below. Dispersion is measured by the dispersion constant $D$. The dependence of $v$ and $D$ with $f$ contains useful and sensitive information about the type of disorder in the media. }
\label{fig:dispersion}
\end{figure}
To give these definitions a concrete physical interpretation we can think of particles running on parallel racetracks, each one with its own realization $F(x)$ but identical statistical properties, with one particle per track (see Fig.\ref{fig:dispersion}). Then the average velocity $v$ is then the center of mass velocity and $D t$ is the variance of their distribution around the center of mass, for a large collection of particles.
We can think the channel array is a memory bus and the particles as carriers of information. In such imaginary application $D$ and $v$ are relevant quantities to perform efficient parallel computation. If particles are actually localized objects of a certain finite size, such as magnetic domain walls, magnetic skyrmions, solitons or colloidal particles, the zero temperature approximation may be well justified.
In addition to $v$ and $D$ We will be also interested in the steady-state force dependent differential mobility, defined as
\begin{eqnarray}
\mu = \frac{dv}{df}
\end{eqnarray}
as it allows us to define an effective temperature from a generalized Einstein relation
as
\begin{eqnarray}
T_{\rm eff} = \frac{D}{2\mu}.
\label{eq:Teffdefinition}
\end{eqnarray}
$T_{\rm eff}$ will in general depend on the driving force $f$, or the velocity $v$.
\section{Exact results}
\label{sec:exactresults}
{\it Exact results for any $f$} can be readily obtained in the $m\to 0$ overdamped case
\begin{equation}
\dot x = F(x) + f.
\label{eq:monomereq}
\end{equation}
when $F(x)$ is a piece-wise short-range correlated random function, such that the random-force $F(x)$ is constant in successive regular intervals (i.e. a random-steps force-field). If we then chose a bounded distribution for $F(x)$, a finite depinning threshold $f_c = \max[-F(x)]$, such that $v=0$ for $f<f_c$, is guaranteed to exist. Within this family of models we find (see appendix \ref{sec:appendixformulaD}),
\begin{eqnarray}
v &=& 1/\langle \Delta t \rangle \label{eq:vformula}
\\
D &=& \frac{\langle \Delta t^2 \rangle-\langle \Delta t \rangle^2}{\langle \Delta t \rangle^3}
\label{eq:Dformula}
\end{eqnarray}
where $\Delta t$ the time a particle spends on a given cell of size $\Delta x = 1$.
Therefore, the calculation of $v$ and $D$ reduces to single cell
averages in the steady-state.
Below we derive results for two particular cases, a Random-Field RF case and a Random-Bond RB case and later explore their
robustness.
\subsection{Box distribution of random forces}
\label{sec:boxdist}
Let us start with an instructive toy model with
uniformly distributed bounded forces at each cell.
\begin{figure}[htp]
\centering
\includegraphics[width=0.5\textwidth]{fig0}
\caption{A piece-wise RF force field (a) and its potential (c). A piece-wise RB force field (b) and its potential (d). In both cases the depinning threshold is $f_c \to 1$ for large systems. Exact calculations show that both cases have the same $v$ but different $D$.
}
\label{fig:RFandRB}
\end{figure}
\subsubsection{RF case}
To construct a RF disorder such that $\int_y g(y)>0$ we take
\begin{eqnarray}
F(x)=R_{[x]}
\end{eqnarray}
where $[\dots]$ denotes the integer part. Constant force intervals are hence of size $\Delta x=1$ and labeled with an integer $n$.
The $R_n$ are uniformly distributed
random numbers in the interval $[-1,1]$
such that
\begin{eqnarray}
\overline{R_n R_m} &=&\frac{\delta_{n,m}}{3} .
\end{eqnarray}
A sample of this RF force field at short distances is shown in Fig.\ref{fig:RFandRB}(a), and it corresponding potential
$U(x)=-\int_x F(x)$ in in Fig.\ref{fig:RFandRB}(c).
From Eq.(\ref{eq:monomereq})
the time spent in the interval $n$
is
\begin{eqnarray}
\Delta t_n = \frac{1}{R_n+f} ,
\end{eqnarray}
and hence
\begin{eqnarray}
\langle \Delta t\rangle &=&
\frac{1}{2} \int_{-1}^{1}
\frac{1}{R+f} \; dR
= \frac{1}{2}\ln \left(\frac{f+1}{f-1}\right), \\
\langle \Delta t^2 \rangle &=&
\frac{1}{2} \int_{-1}^{1}
\frac{1}{(R+f)^2} \; dR
= \frac{1}{f^2-1}.\qquad
\end{eqnarray}
The {\it exact} mean velocity for $f \geq 1$ is
\begin{eqnarray}
v =
\frac{2}{\ln \left(\frac{f+1}{f-1}\right)},
\label{eq:exactvuniform}
\end{eqnarray}
displaying a depinning transition at $f=1$.
Approaching the threshold, $f\to 1+$, the velocity vanishes continuously as $v \sim -2/\ln (f-1)$. This model is thus peculiar from the standard critical phenomena point of view where we expect a power-law $v \sim (f-1)^\beta$ for an order parameter $f$ and a control parameter $f$
(see however section \ref{sec:sintheta}).
On the other hand, for $f\gg 1$, $v\simeq f$ as expected from the washing out of the pinning force at large velocities and thus free flow is recovered. The velocity force characteristics can be also inverted, yielding
\begin{eqnarray}
f = \frac{e^{2/v}+1}{e^{2/v}-1}.
\end{eqnarray}
The exact (differential) mobility for $f \geq 1$ is
\begin{eqnarray}
\mu =
\frac{4}{
(f^2-1) \ln^2 \left(\frac{f+1}{f-1}\right)},
\end{eqnarray}
such that $\mu \to \infty$ when $f\to 1$ (i.e. right at the depinning transition), and $\mu\to 1$ when $f\to \infty$ as expected.
Using Eq.(\ref{eq:Dformula}) the exact dispersion constant for $f \geq 1$ is
\begin{eqnarray}
D =
\frac{8 \left(\frac{1}{f^2-1}-\frac{1}{4} \ln ^2\left(\frac{f+1}{f-1}\right)\right)}{\ln ^3\left(\frac{f+1}{f-1}\right)}
\label{eq:exactDRF}
\end{eqnarray}
and therefore, using the generalized Einstein relation we get the exact effective temperature
\begin{eqnarray}
T_{\rm eff}=
\frac{2 \left(1-\frac{f^2-1}{4} \ln ^2\left(\frac{f+1}{f-1}\right)\right)}{\ln \left(\frac{f+1}{f-1}\right)
}
\end{eqnarray}
We are interested in the $f\gg 1$ fast-flow behaviour of $D$ and $T_{\tt eff}$. Expanding in powers of $1/f$ we first get
\begin{eqnarray}
v &\simeq& f - \frac{1}{3f} - \frac{4}{45 f^3} + \mathcal{O}(f^{-5})
\label{eq:largevexpansion}
\\
\mu &\simeq& 1 + \frac{1}{3f^2}+\mathcal{O}(f^{-4}),
\end{eqnarray}
and in particular for the dispersion and associated effective temperature we get
\begin{eqnarray}
D &\simeq& \frac{1}{3f} + \frac{7}{45 f^3 } + \mathcal{O}(f^{-5}) \nonumber \\
&\simeq&
\frac{1}{3v} + \frac{2}{45 v^3} + \mathcal{O}(v^{-4}),\\
T_{\tt eff}&\simeq& \frac{1}{3f}+\frac{2}{45f^3}
+\mathcal{O}(f^{-5}).
\end{eqnarray}
We will particularly focus our attention on the large-velocity dominant term,
which in this case is
\begin{eqnarray}
D\sim 1/f \sim 1/v,\;\;\; \text{RF}.
\label{eq:RFresult}
\end{eqnarray}
The above RF scaling can be obtained by first order perturbation theory in $f^{-1}$. At zero-th order
Eq.(\ref{eq:monomereq}) implies $x \approx f t$,
and then at first order $\dot{x}=f+F(f t)$. The random fluctuating force $F(ft)$ mimics a colored noise because $\langle F(ft) \rangle=0$, $\langle F(ft)F(ft') \rangle = g(f(t-t'))$, where $g(x)$ is short-ranged. Most importantly, because we are considering RF disorder, $\int g(x)>0$ and thus we can define a positive effective temperature $T_{\rm eff} \propto \int_{-\infty}^{\infty} g(ft)dt \sim f^{-1}$. Hence, since $\mu\to 1$, $D\sim 1/f$.
\subsubsection{RB case}
\label{sec:RBcaseUniform}
To model RB disorder
with $\int_y g(y)=0$,
we define the random forces in Eq.(\ref{eq:monomereq}) as
\begin{eqnarray}
F(x) = R_{[x]} \text{sign}(x-[x]-1/2),
\end{eqnarray}
where as above the $R_n$ are iid random variables, uniformly distributed in $[-1,1]$.
This choice splits the unit interval of each cell in two parts of size $1/2$, such that the constant force in the second half is minus the force of the first half.
A sample of this RF force field at short distances is shown in Fig.\ref{fig:RFandRB}(b), and it corresponding potential
$U(x)=-\int_x F(x)$ in in Fig.\ref{fig:RFandRB}(d).
Evidently this is a particular way to obtain the auto-correlation needed for a RB type of disorder. As we show below, the choice also has the advantage of leaving $v$ and $\mu$ invariant from $RF$ to $RB$.
By repeating the procedure of the previous section we obtain $\langle \Delta t \rangle$ as for the RF case,
and thus {\it identical} $v$ and $\mu$ as a function of $f$.
However, $\langle \Delta t^2 \rangle$ is different,
\begin{eqnarray}
\langle \Delta t^2 \rangle &=&
\frac{1}{2} \int_{-1}^{1}
\left[ \frac1{ 2(f+R)} + \frac1{2 (f-R)} \right] ^2\; dR \\
&=&
\frac{1}{4 f}
\left[\frac{2f}{f^2-1}+\ln \left(\frac{f+1}{f-1}\right)\right].
\end{eqnarray}
The {\it exact} dispersion constant is
\begin{eqnarray}
D=
\frac{2 \left(\frac{2}{f^2-1}-\log ^2\left(\frac{f+1}{f-1}\right)+\frac{\log \left(\frac{f+1}{f-1}\right)}{f}\right)}{\log ^3\left(\frac{f+1}{f-1}\right)}.
\label{eq:exactDRB}
\end{eqnarray}
$D$ thus diverges at the depinning threshold
and at large driving forces vanishes as
\begin{eqnarray}
D \simeq T_{\tt eff} \simeq \frac{4}{45 f^3} + \mathcal{O}(f^{-5})
\simeq \frac{4}{45 v^3} + \mathcal{O}(v^{-5})
\end{eqnarray}
which is clearly different from the RF case.
Then, the dominant term in this case is,
\begin{eqnarray}
D\sim 1/f^3 \sim 1/v^3,\;\;\; \text{RB}
\label{eq:RBresult}
\end{eqnarray}
faster than the RF case.
It is worth noting that the $f^{-3}$ RB scaling, at variance with the $f^{-1}$ RF, thus can not be obtained from first order perturbation in $f^{-1}$. Indeed, dispersion is (incorrectly) zero at first order, because for RB we have
$\int g(x)=0$.
\subsubsection{Summary}
In Fig.\ref{fig:boxdist} we summarize the exact results for this model.
We show that {\it identical} $v$ vs $f$ characteristics can be accompanied by different $D$ vs $f$ characteristics.
We show in particular that $D$ vanish differently at large $f$ in the RF and RB cases, thus becoming a sensitive tool to characterize the random-medium. It is worth noting that the decay of $D$ with $f$ depends crucially on the fact that the random-medium produce finite random forces whose effect becomes weaker at large velocities. A simple disordered model where this is not the case and $D$ increases with $f$ is discussed in the appendix \ref{sec:randomfriction}.
\begin{figure}[htp]
\centering
\includegraphics[width=0.5\textwidth]{fig1.pdf}
\caption{$v$ and $D$ vs $f$ for RF and RB force fields with a box-distribution. In this model, $v$ is invariant with respect to the RF or RB nature of the disorder.
At $f\to 1$ (depinning threshold) $v$ vanishes logarithmically, while $D$ diverges. At large forces, as indicated, $v \sim f$, while $D \sim 1/f$ and $D \sim 1/f^3$ for the RF and RB cases respectively, as indicated with dotted-lines.}
\label{fig:boxdist}
\end{figure}
\subsection{``RB + $\epsilon$ RF'' case}
\label{sec:rb+rf}
We now discuss the case where disorder is not pure RB or RF but where a RB disorder is perturbed by a RF disorder. To achieve this we split the $n$-esime unit interval in two equal parts, and put a uniformly distributed random force $R_n$ in the first half, followed by a random force $-R_n+\delta_n$ in the second half. We will consider that $\delta_n$ is a small random variable with $\langle \delta_n \rangle=0$ and $\langle \delta_n^2 \rangle=\epsilon^2 \ll 1$.
We will also consider that $\langle R_m \delta_n\rangle=0$ so to ensure that $v$ remains invariant under the perturbation.
The $n$-esime residence time is
\begin{eqnarray}
\Delta t_n &=& \frac{1}{2} \left( \frac{1}{f+R_n}+\frac{1}{f-R_n+\delta_n} \right) \nonumber \\
&\approx& \Delta t_{RB} + \Delta t_{RF}
+ \mathcal{O}(\epsilon^2)
\end{eqnarray}
where
\begin{eqnarray}
\Delta t_{RB} &=& \frac{f}{f^2-R_n^2}, \\
\Delta t_{RF} &=& -\frac{\delta_n}{2(f-R_n)^2},
\end{eqnarray}
are the pure RB and RF random contributions.
Since $\langle R_m \delta_n\rangle=0$
we have
$\langle \Delta t_{RF} \rangle \propto \langle \delta_n \rangle =0$.
Therefore the RF perturbation does not alter the mean velocity, as $v^{-1}=\langle \Delta t_n \rangle=\langle \Delta t_{RB} \rangle$.
Nevertheless,
\begin{equation}
\langle \Delta t^2 \rangle
= \langle \Delta t_{RB}^2 \rangle +
\langle \Delta t_{RF}^2 \rangle
\end{equation}
where in the last equality we have used $\langle \Delta t_{RB} \Delta t_{RF}\rangle \propto \langle \delta_n \rangle =0$. Since we already know
$\langle \Delta t^2_{RB} \rangle$ from Sec.\ref{sec:RBcaseUniform} we have only to calculate $\langle \Delta t^2_{RF} \rangle$, which for uniformly distributed forces is simply
\begin{eqnarray}
\langle \Delta t^2_{RF} \rangle&=&
\langle \delta_n^2 \rangle \frac{1}{2}\int_{-1}^1 d R \left(\frac{1}{2(f-R)^2}\right)^2
\nonumber \\
&=&
\frac{\epsilon^2 \left(3 f^2+1\right)}{12 \left(f^2-1\right)^3}
\end{eqnarray}
Since
\begin{eqnarray}
D= \frac{
\langle \Delta t_{RB}^2 \rangle- \langle \Delta t^2 \rangle
+\langle \Delta t_{RF}^2 \rangle}{\langle \Delta t \rangle^3}
\end{eqnarray}
we obtain
\begin{eqnarray}
D= D_{RB} + \frac{\langle \Delta t_{RF}^2\rangle}{\langle \Delta t\rangle^3} = D_{RB} + v^3 \langle \Delta t_{RF}^2\rangle.
\end{eqnarray}
To the leading term $\langle \Delta t_{RF}^2\rangle \approx 1/(4 f^4)$, $D_{RB}\approx 4/(45f^3)$ and $v \sim f$, we have
\begin{eqnarray}
D \approx \frac{4}{45f^3} + \frac{\epsilon^2}{4f}.
\end{eqnarray}
Therefore, the RF perturbation dominates at large enough velocities and $D \sim 1/f$. The crossover in this model roughly occurs at
$f^* \sim \frac{4}{3\sqrt{5}\epsilon}$ if $\epsilon \ll 1$.
In Fig.\ref{fig:RBperturbRF} we summarize the exact results for this toy model.
\begin{figure}[htp]
\centering
\includegraphics[width=0.5\textwidth]{fig2}
\caption{
Crossover behaviour in a RB force field perturbed with a RF force field (both with a box-distribution of forces), for perturbing parameter $\epsilon=0.01$. While $v$ remains invariant under the small perturbation $D$ crossovers from $RB$ to $RF$ behaviour at large enough forces, as indicated with dotted-lines.
}
\label{fig:RBperturbRF}
\end{figure}
The crossover just described may be important for instance for flat domain walls in media with RB disorder such as non-magnetic impurities, crystalline defects or rough borders, contaminated with a few magnetic impurities which produce RF pinning. On the other hand, it may be important for particles described by two or more coupled degrees of freedom, since in that case the force field seen by the coordinate describing its position in the track is effective.
More subtly, the crossover is important for the case of extended systems such as (non-flat) elastic interfaces but with pure RB microscopic disorder, since in that case the renormalized pinning force is not pure RB anymore.
At interface depinning indeed, the renormalized disorder becomes of a RF type and hence a merging of RF and RB into a unique RF depinning universality class emerges. However, as the velocity increases, such renormalization is expected to be weaker because the correlation length $\mathcal{l}$ along the interface becomes smaller (near $f_c$ as $\mathcal{l} \sim (f-f_c)^{-\nu}$), and then the renormalized pinning force tend to flow down to the RB microscopic disorder again. A toy model that beautifully illustrates this physics near the depinning threshold is the the particle quasistatically dragged in a random force field by a parabolic potential of curvature $m^2$~\cite{LeDoussal2009}.
Then the effect of increasing $v$ can be related to the effect of increasing $m \sim 1/\mathcal{l}$, and one can fully appreciate the flow from RB to RF by approaching the depinning threshold.
The results for the present toy model hence suggests that even the small RF part that remains from renormalization at large drives (or large $m$) should be still detectable beyond a putative crossover of $D$ as a function of $f$.
\subsection{A non-uniform distribution of random forces}
\label{sec:sintheta}
In order to show the robustness of the previous results we now consider a simple model variant with the same scaling properties for $D$ at large velocities in the RF and RB case, but quite different critical behaviour near the depinning threshold. The particular variant proposed also display much simpler exact analytical expressions for all quantities and could thus be used as a convenient toy model for other studies.
In order to achieve this we change the probability distribution of random-forces in each cell such that
\begin{eqnarray}
R_n = \sin \theta_n
\end{eqnarray}
with $\theta_n$ uniformly distributed in the $[-\pi,\pi)$ interval. This disorder may physically correspond to a particle coupled to the $x$ component of a random unit vector force in a plane.
\subsubsection{RF case}
As before, we compute the mean residence time in a cell
\begin{eqnarray}
\langle \Delta t \rangle =
\frac{1}{2\pi} \int_{-\pi}^{\pi}
\frac{d\theta}{\sin(\theta)+f} \; = \frac{1}{\sqrt{f^2-1}},
\end{eqnarray}
and hence the exact velocity and differential mobility are respectively
\begin{eqnarray}
v = \sqrt{f^2-1}\\
\mu = \frac{f}{\sqrt{f^2-1}}.
\end{eqnarray}
Interestingly, this is exactly the mean velocity of the
$\dot x = f + \sin(x)$ model and
it has a finite and trivial critical exponent $v\sim (f-1)^{\beta}$, $\beta=1/2$, different than the previous logarithmic depinning case. Nevertheless, at variance with the particle in the periodic potential, where there is no dispersion at zero temperature, here we have a disorder-induced dispersion with respect to the disorder-averaged displacement.
To see this we compute as before
\begin{eqnarray}
\langle \Delta t^2 \rangle =
\frac{1}{2\pi} \int_{-\pi}^{\pi}
\frac{d\theta }{(\sin(\theta)+f)^2} \;
= \frac{f}{(f^2-1)^{3/2}}
\label{eq:nonUniformDt2}
\end{eqnarray}
Then, using Eq.(\ref{eq:Dformula}), we get the {\it exact} dispersion constant
\begin{eqnarray}
D = f-\sqrt{f^2-1}
= -v + \sqrt{1+v^2}
\end{eqnarray}
which {\it remains finite} as $f\to 1$ ($D\to 1$) and
at large forces behaves as \begin{eqnarray}
D
\approx
\frac{1}{2f}+\frac{1}{8 f^3} + \mathcal{O}(f^{-5})
\approx
\frac{1}{2v}-\frac{1}{8v^3}+\mathcal{O}(v^{-5})
\end{eqnarray}
These results show that the RF dependence $D\sim 1/f$ at large velocities (Eq.(\ref{eq:RFresult})) is robust under the change of the step distribution we have made, in spite that the critical depinning behaviour is clearly different.
\subsubsection{RB case}
As before, we use the trick of dividing the unit interval in two correlated parts,
\begin{eqnarray}
F(x) = \sin \theta_{[x]} \text{sign}(x-[x]-1/2).
\end{eqnarray}
where $\theta_n$ is uniformely distributed in $[\pi,\pi)$ as before. The convenience of this choice is, as before, that $v$ remains invariant.
The residence time in a single cell then is split in two contributions
\begin{eqnarray}
\Delta t = \frac{1/2}{f+\sin(\theta)}+\frac{1/2}{f-\sin(\theta)}
\end{eqnarray}
Therefore,
\begin{eqnarray}
\langle \Delta t \rangle &=&
\frac{1}{4\pi} \int_{-\pi}^{\pi}
\frac{d\theta}{\sin(\theta)+f}
+\frac{1}{4\pi} \int_{-\pi}^{\pi}
\frac{d\theta}{-\sin(\theta)+f}
\nonumber \\
&=& \frac{1}{\sqrt{f^2-1}},
\end{eqnarray}
{\it identical} to the RF case of the previous subsection, and hence $v$ and $\mu$ are also identical.
Nevertheless
\begin{eqnarray}
\langle \Delta t^2 \rangle =
\frac{2f^2-1}{2f(f^2-1)^{3/2}}
\end{eqnarray}
is different than the RF case, Eq.(\ref{eq:nonUniformDt2}).
The {\it exact} dispersion constant is
\begin{eqnarray}
D=-\sqrt{f^2-1}+f-\frac{1}{2 f},
\end{eqnarray}
which {\it remains finite} as $f \to 1$ ($D\to 1/2$) and its large velocity expansion yields
\begin{eqnarray}
D &=& \frac{1}{8f^3 }+\mathcal{O}(f^{-5})
\end{eqnarray}
These results show now that the RB dependence $D\sim 1/f^3 \sim 1/v^3$ at large velocities (Eq.(\ref{eq:RBresult})) is also robust under a change of the step distribution.
In Fig.\ref{fig:sintheta} we summarize the exact results for this model.
We note first that when $f\to 1+$, $v\sim (f-1)^{1/2}$, in contrast with the logarithmic behaviour (Eq.(\ref{eq:exactvuniform})) obtained for the box distribution in the previous sections. Moreover, we find that $D \to 1$ (RF) and $D \to 0.5$ (RB), also in sharp contrast with the divergent dispersion constants in the same cases for the box distribution, Eqs.(\ref{eq:exactDRF}) and (\ref{eq:exactDRB})
respectively.
Nevertheless, the (RF) $1/f$ and (RB) $1/f^3$ asymptotics of $D$ at large $f$ remains a robust feature.
\begin{figure}[htp]
\centering
\includegraphics[width=0.5\textwidth]{fig3.pdf}
\caption{$v$ and $D$ vs $f$ for the RF and RB cases of the model with a non-uniform force distribution.
At the depinning threshold $v\sim (f-1)^{1/2}$ and $D\to 1$ and $D\to 1/2$ (indicated with Red dashes lines) for RF and RB respectively, in sharp contrast with the box-distribution model.
At large forces, we recover the robust $D\sim 1/f$ and $D\sim 1/f^3$ behaviours of the RF and RB cases respectively (indicated with dotted-lines).
}
\label{fig:sintheta}
\end{figure}
\subsection{Conexion with a trap model for dispersion}
\label{sec:trapmodel}
Dispersion under an average flow is an old problem, relevant for the physics of porous media where the dispersion constant $D$ measures the spread of an injected packet of tracer particles.
In this context a pedagogical toy model proposed by Bouchaud and Georges~\cite{BouchaudGeorges1990} allows to understand the basic physics and estimate $D$. In that model the medium is idealized as one dimensional and made of a ``backbone'' along which the particle is convected with velocity $V$ at regularly spaced positions (separated by a distance $\xi$) and with probability $p$ can leave the backbone during a random waiting time $\tau$. Importantly, the particle can not go backwards.
\begin{figure}[htp]
\centering
\includegraphics[width=0.5\textwidth]{bouchaud}
\caption{Adapting the Bouchaud and Georges trap model (a) to the quenched-force field toy model above the depinning threshold, at zero temperature (c). To convert it in (b) we consider that that free flow is $V=f$, and that the effective trapping occurs with probability one with waiting times that are random but quenched, each one given by the ``excess residence time'' in a cell $\tau_n \equiv \frac{\xi}{F_n+f}-\frac{\xi}{f}$, where $n\equiv [x]$ is the cell number. }
\label{fig:bouchaud}
\end{figure}
Bouchaud and Georges showed that
\begin{eqnarray}
U^{-1}&=&V^{-1}+p\langle \tau \rangle/\xi
\label{eq:bouchaudgeorgesU}
\\
D_\parallel &=& (pU^3/2\xi)(\langle \tau^2 \rangle - p \langle \tau \rangle^2).
\label{eq:bouchaudgeorgesD}
\end{eqnarray}
where $U$ is the average velocity of the particle in the medium and $D_\parallel$ denotes the longitudinal dispersion constant. These results show that for a given drive $V$, $D_\parallel$ {\it grows} with $V$, in good qualitative agreement with some porous media experiments. It is interesting to note that at large velocities this model predicts $D_\parallel \sim V^3$, in sharp contrast with our results for uncorrelated random media, $D\sim 1/f^3 \sim 1/v^3$ (RB) and $D\sim 1/f \sim 1/v$ (RF).
In the following we explain such difference and show how to reconcile the results of Eqs.(\ref{eq:bouchaudgeorgesU}) and (\ref{eq:bouchaudgeorgesD}) with the predictions of our dispersion model.
To explain the differences between the Bouchaud-Georges model predictions and the ones we presented in the previous sections we start by noting that the average in Eq.(\ref{eq:bouchaudgeorgesD}) is over different stochastic trajectories in one track, while in
Eq.(\ref{eq:Dformula}) the average is over disorder realizations (see Fig.\ref{fig:bouchaud}). The former is the so-called quenched dispersion constant while the later is the so-called annealed dispersion constant. We show below that this does not prevent however to find a concrete connection between the two mathematical approaches.
We also note that in the problem we are interested in, the delay $\langle \Delta t \rangle$ in each cell is not given by a fixed distribution, but by a drive dependent distribution. In particular, the larger $f$, the closest is $\langle \Delta t \rangle$ to $1/f$.
Hence, to make a concrete connection with the overdamped mechanical model we replace the trap in the $n$-esime cell of size $\xi=1$ by a quenched random force field
and redefine $\tau$ as the {\it excess time} with respect to the free motion time in the cell $=1/f$.
Note that replacing $\tau$ by quenched random values is not problematic because the particle never moves backward. That is,
\begin{eqnarray}
\tau \equiv \Delta t - \frac{1}{f},
\end{eqnarray}
and we also make the identification $V \equiv f$ and $U \equiv v = 1/\langle \Delta t \rangle$. We can also take $p=1$ since the non-trapping event can be absorbed in the distribution of $\tau$, corresponding to $F_n=0$. With such analogy, noting that $2 D_\parallel \equiv D$ (because of the convention we have used to define $D$, see Eq.(\ref{eq:Ddef})), we finally get
\begin{eqnarray}
v^{-1}&=& V^{-1}+ \langle \tau \rangle = {\langle \Delta t\rangle} \\
D &=& U^3 (\langle \tau^2 \rangle - \langle \tau \rangle^2)
= \frac{\langle \Delta t^2 \rangle - \langle \Delta t \rangle^2}{\langle \Delta t \rangle^3}
\end{eqnarray}
which is identical to our expressions, see Eq.(\ref{eq:Dformula}).
The later adaptation and reinterpretation of the different terms shows that our results for the mechanical models ultimately reduce to compute a few statistical properties (first and second moments) of the excess waiting time in a trap equivalent model.
Interestingly, the mechanical model yields always a {\it decreasing} $D$ with increasing $v$, and the microscopic nature of the short-correlated disorder is important to determine how exactly $D$ vanishes at large velocities.
\section{Numerical Results for different models}
\label{sec:numerics}
In order to assess the universality of some of the analytical predictions we have made in the previous sections for simple one-dimensional toy models, in this section we compare them with more complicated physically relevant disordered models, for which an exact solution is not available, but precise numerical simulations can be performed.
We consider overdamped and damped models, the effect of temperature and the effect of transverse fluctuations in wires with a finite or infinite width.
We also consider the dispersion of a non-rigid particle and a micromagnetic model of a magnetic domain wall driven by an external electrical current and/or applied magnetic field.
Solving all the proposed models reduce to integrate a system of ordinary non-linear differential equations of at most two variables, with disorder. In all cases we solve them by standard integration techniques. We use between $8192$ to $32768$ independent disorder realizations (or tracks) to average the properties of interest, and run long-time simulations so to capture the diffusive behaviour beyond any finite time crossover.
\subsection{Inertia: massive damped particles}
Embedded soft matter systems, such as magnetic domain walls, magnetic skyrmions, or superconducting vortices behave effectively as overdamped driven objects and inertia can be ignored. Inertia may be important however for other experimental realizations which could be modeled by a driven massive damped particle or soliton in a quenched random field. Inertia opens, on the other hand, a new phenomenology because the state space is doubled compared to the overdamped case. In our particular one-dimensional case of Eq.(\ref{eq:equationwithinertia}), the state of the damped particle must be described by two variables instead of one, and it is a priori not evident whether the universal results we obtain for $D$ in the RF and RB cases remain valid in this case. Since to the best of our knowledge no analytical solutions for the dispersion properties are known for Eq.(\ref{eq:equationwithinertia}) for finite $m$, here we solve it numerically.
For the simulations we solve the nondimensional Eq.(\ref{eq:equationwithinertia}) using the same piece-wise $F(x)$ considered in sections \ref{sec:boxdist} for the RB and RF cases. From Eq.(\ref{eq:equationwithinertia})
we note that $v \approx f$ in the large $f$ limit, because the average acceleration must be zero in presence of damping. We will be particularly interest in the fluctuations when $f-v \ll f$.
\subsubsection{RF case}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figMSDVarP.pdf}
\caption{Quadratic mean displacement (a) and momentum variance (b) in the moving frame vs $t$ for the RF disorder case of Eq.(\ref{eq:equationwithinertia}), using $m=0.1$, and indicated driving forces. The dashed-lines indicates normal {dispersion}.
}
\label{fig:msdRF}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figDvsfvsMRF.pdf}
\caption{Numerical results for the inertial RF disorder case of Eq.(\ref{eq:equationwithinertia}). (a) Rescaled velocity deficit as a function of the drive $f$, for different masses $m$ (indicated in (b)), in the fast flow regime $f-v \ll v$. Dash and dashed-dotted lines indicate approximate asymptotic behaviours.
(b) Dispersion constant vs $f$ for different masses. The dotted line shows a $\sim 1/f$ decay.
(c) Rescaled momentum variance vs $f$. The dahed-dotted and dashed lines shows asymptotic predictions (see text).
}
\label{fig:figinertiaRF}
\end{figure}
We start describing the results for the RF field.
In Fig.\ref{fig:msdRF}(a) we show the quadratic mean displacement in the moving frame $\langle [x-v t]^2\rangle$ vs $t$, where $v=\langle \dot x \rangle$ is the steady-state mean velocity. We see that dispersion is normal and thus a driving force dependent dispersion constant $D$ can be fitted. In Fig.\ref{fig:msdRF}(b) we see that the variance of the momentum in the moving frame $\langle p-m v \rangle$, shown for the same mass and driving forces, has a well defined drive dependent steady-state value (temporal fluctuations are controlled by the number of tracks considered in the average).
In Fig.\ref{fig:figinertiaRF}(a) we show that in the fast-flow regime
$f-v \approx m G_v(fm)$, with $G_v(x)$ displaying approximately a crossover between two power laws, $G_v(x)\sim x^{-1}$ for $x\ll 1$ and $G_v(x)\sim x^{-3}$ for $x\gg 1$. For a fixed $f$ and a small $m$ the result is consistent with the expansion obtained for the over-damped toy model, Eq.(\ref{eq:largevexpansion}). Interestingly, there is a crossover to a different regime for a fixed value of $f m$, so the smaller $m$ the larger the crossover force $f_m\sim m^{-1}$.
Therefore, the expansion of Eq.(\ref{eq:largevexpansion}) is not robust.
In Fig.\ref{fig:figinertiaRF}(b) we show nevertheless that the dispersion constant $D$ is practically independent of $m$ and decays as $D \approx f^{-1}/3$, with the same power-law predicted analytically with the toy model in the RF case, Eq.(\ref{eq:RFresult}).
In Fig.\ref{fig:figinertiaRF}(c)
we show that the steady-state variance of the momentum obeys also a scaling $\langle p^2 \rangle-\langle p \rangle^2 \sim m^2 G_p(f m)$ with a crossover, at $f m \sim 1$, from a constant to a power law decay, $G_p(x)\approx x^{-1}/12$. Interestingly, the exact fits to $D$ and $\langle p^2 \rangle-\langle p \rangle^2$ for large $f m$ are consistent with the relation
\begin{equation}
\frac{\langle [p-\langle p \rangle]^2 \rangle}{2m}
\approx \frac{D}{4} \approx \frac{T_{\tt eff}(f)}{2}
\label{eq:equipartitionRF}
\end{equation}
where in the last equality we have used the definition of the effective temperature from the generalized Einstein relation of Eq.(\ref{eq:Teffdefinition}), with $\mu\approx 1$ for $f-v \ll f$.
In other words, the mean kinetic energy calculated in the moving frame with velocity $v=\langle p/m \rangle$ appear to satisfy the equilibrium equipartition law in spite of being a system far from equilibrium.
We can understand this result at the lowest order in a high velocity expansion. If we write that the random force is $F(x)\approx F(vt)$, and that $\dot x \approx v \approx f$, from Eq.(\ref{eq:equationwithinertia})
\begin{equation}
m \ddot u \approx {F}(vt)-\dot u.
\label{eq:langevinRF}
\end{equation}
for the position coordinate $u=x-vt$ in the moving frame.
Then, using that $\overline{F(x)F(x')}=g(|x-x'|)$ with $g(x)$ a rapidly decaying function with $\int_x g(x)>0$ for RF,
we can compute the time autocorrelation function of the random force for $v \gg |t-t'|^{-1}$, $\overline{F(vt)F(vt')}=g(v|t-t'|)\sim \delta(t-t')/v$.
Hence, in such large velocity limit, the random force behaves effectively as a Langevin noise $\xi_v(t)\equiv F(vt)$ \footnote{The ensemble average of the Langevin noise coincides in this limit with the average over tracks.} at an effective temperature $T_{\tt eff}\sim v^{-1}$ (since the mobility is constant and $\mu=1$). The system thus reaches thermal equilibrium at the $T_{\tt eff}$, and hence the mean kinetic energy is $\langle \frac{m}{2}\dot u^2 \rangle = T_{\tt eff}/2$,
in agreement with the asymptotic result of Eq.(\ref{eq:equipartitionRF}).
On the other hand, for $fm<1$ the particle enters a regime where the friction force $\dot x$ dominates over the inertial term $m\ddot x$
and then, using the same arguments as above we get
\begin{equation}
\dot u \approx F(vt)
\label{eq:smallfm}
\end{equation}
which implies that $\langle \dot u^2 \rangle = \langle F(vt)^2 \rangle=1/3$. Therefore,
$\langle [p-mv]^2 \rangle/2 m^2 = 1/6$ in agreement with the scaled data shown in Fig.\ref{fig:figinertiaRF}(c) for $fm\ll 1$.
\subsubsection{RB case}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figMSDVarPRB.pdf}
\caption{Quadratic mean displacement (a) and momentum variance (b) in the moving frame vs $t$, using $m=0.1$ and indicated driving forces, for the RB disorder case of Eq.(\ref{eq:equationwithinertia}). The dashed-line in (a) indicates normal {dispersion}.
}
\label{fig:msdRB}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figDvsfvsMRB.pdf}
\caption{Numerical results for the inertial RB case of Eq.(\ref{eq:equationwithinertia}). (a) Rescaled velocity deficit as a function of the drive $f$, for different masses $m$ (indicated in (b)), in the fast flow regime $f-v \ll v$. Dash and dashed-dotted lines indicate approximate asymptotic behaviours, as indicated.
(b) Dispersion constant vs $f$ for different masses. The dotted line highlights the $\sim f^{-3}$ decay.
(c) Rescaled momentum variance vs $f$. The dashed-dotted and dashed lines indicate approximate asymptotic behaviour. Grey symbols in (a) and (c) correspond to the results of the RF case.
}
\label{fig:figinertiaRB}
\end{figure}
We now describe the results for the RB field.
In Fig.\ref{fig:msdRB}(a) we show the quadratic mean displacement in the moving frame $\langle [x-v t]^2\rangle$ vs $t$, where $v=\langle \dot x \rangle$ is the steady-state mean velocity. We see that {dispersion} is normal and thus a driving force dependent dispersion constant $D$ can be fitted. In Fig.\ref{fig:msdRB}(b) we see that the variance of the momentum in the moving frame $\langle (p-m v)^2 \rangle$, shown for the same mass and driving forces, has a well defined drive-dependent steady-state value.
In Fig.\ref{fig:figinertiaRB}(a) we show, for various masses (indicated in Fig.\ref{fig:figinertiaRB}(b)), that the deficit of velocity with respect to the free velocity obeys approximately the same scaling $(f-v)\sim m G_f(f m)$ found for the RF case, and also presents a crossover between two regimes. The first regime displays the correction $G_f(x)\sim x^{-1}$ expected from the large velocity expansion of the overdamped toy-model Eq.(\ref{eq:largevexpansion}). The second regime $G_f(x)\sim x^{-2}$ for $f m \gg 1$, shows that inertia can have an important effect, and thus the scaling of Eq.(\ref{eq:largevexpansion}) is not universal.
In Fig.\ref{fig:figinertiaRB}(b) we show nevertheless that the dispersion coefficient $D$ follows closely the $f^{-3}$ predicted by the toy model in the RB case, Eq.(\ref{eq:RBresult}).
It is nevertheless worth noting that while in the RF case $D$ is almost independent of $m$, in the RB case there is a strong dependence, as evidenced by the rescaling of curves into a master curve.
We also note that $D$, for the same range of $m$ and $f$, is in general much smaller in the RB case than in the RF case, and more difficult to sample precisely.
Finally, in Fig.\ref{fig:figinertiaRB}(c)
we show that the momentum variance follows the same scaling $\langle p^2\rangle-\langle p\rangle^2 = m^2 G_p(fm)$ than in the RF case, and a similar crossover from small to large $fm$. Interestingly, the, small $fm$ regime saturates to an identical value than for the RB case. This is consistent with the effective equation of motion Eq.(\ref{eq:smallfm}), with the saturation value controlled only by the variance of the disorder.
The large $fm$ regime however, where empirically we find $G_p(x)\sim x^{-1.75}$, differs appreciably from the RB case. The faster decay of
$G_p(x)$ is however consistent with the faster decay of $G_f(x)$
shown in Fig.\ref{fig:figinertiaRB}(a).
At variance with the RF case, The large $fm$ regime can not be rationalized using an effective equipartition theorem because $D \sim T_{\tt eff}$ does not decay in the same way in this limit.
The interpretation of this remains open.
\subsubsection{Summary}
In summary, we find that damped particles display:
\begin{itemize}
\item $D\sim 1/f$ and $D\sim 1/f^3$ as predicted for the overdamped toy model for the RF and RB case respectively, Eqs.(\ref{eq:RFresult}) and (\ref{eq:RBresult}). The result thus appears to be robust under inertia.
\item A non trivial dependence with the mass $m$ both in the $f-v$ and in $\langle p^2 \rangle-\langle p \rangle^2$ reveals a crossover between two regimes varying the parameter $f m$. The first regime, identical for RB and RF is well described by the overdamped limit with an effective drive-dependent noise. The second regime is different for RF and RB. In the RF case the second regime is compatible with a generalized equipartition law at the disorder-induced drive dependent effective temperature $T_{\tt eff}=D/2$. For RB the later equipartition law {\it fails}.
\end{itemize}
\subsection{Quasi one- and two-dimensional tracks}
\label{sec:transverse}
When the size of the particles is smaller than the track width transverse fluctuations of a transversely confined particle may modify the predicted longitudinal dispersion properties. This kind of situation may arise for instance in the cases of driven vortices in narrow and thin superconducting strips, driven colloids, or current driven magnetic skyrmions. Compared to the simple one-dimensional overdamped case, now the state-space doubles (as in the one-dimensional damped case) and the effective one-dimensional coupling to the underlying disorder changes.
We study the consequences of these new properties in the velocity and dispersion of particles.
\subsubsection{RB case}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figconfined1.pdf}
\caption{
RB results for the two dimensional model of Eq.(\ref{eq:2d}) with transverse confinement in a two-dimensional random bounded potential.
(a) Quadratic mean displacement with respect to the mean motion in the driving direction vs time, for different driving forces $f$ for a confinement constant $\kappa=1$. The dashed line indicates normal dispersion.
(b) Fitted dispersion constant in the direction of the drive $D_x$ vs $f$ for different confinement constants $\kappa$. The dashed lines indicates the $1/f^3$ behaviour. For comparison $1/f^2$ and $1/f^4$ behavior are drawn in dotted lines.
}
\label{fig:figconfined1}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figconfined2.pdf}
\caption{
RB results for the two dimensional model of Eq.(\ref{eq:2d}) with transverse confinement ($\kappa>0$) for the two-dimensional uncorrelated random potential of Eq.(\ref{eq:bilinear}).
(a) Mean squared displacement in the transverse direction as a function of time, for $\kappa=1$.
(b) The variance of transverse fluctuations approximately satisfies $\kappa \langle y^2 \rangle \sim f^{-1}$ at large forces for different confinement constants $\kappa$. Dotted line indicate the $1/f^2$ behavior for comparison.}
\label{fig:figconfined2}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figconfined3.pdf}
\caption{
RB results for the two dimensional model (Eq.(\ref{eq:2d})) \textit{without} confinement ($\kappa=0$), for the two-dimensional uncorrelated random potential of Eq.(\ref{eq:bilinear}).
Longitudinal (a) and transverse (b) quadratic mean displacement vs time for different driving forces $f$, with respect to the mean motion.
(c) Longitudinal and transverse dispersion constants $D_x$ and $D_y$ as a function of $f$. The dashed-line shows the $\sim f^{-1}$ and the dashed-dotted the $\sim f^{-3}$ behaviours.
}
\label{fig:figconfined3}
\end{figure}
In order to study the effect of transverse fluctuations we will consider here only the relevant RB case of an overdamped particle in a narrow channel transversely confined by a parabolic potential
\begin{eqnarray}
\dot x &=& f - \partial_x U(x,y),
\nonumber \\
\dot y &=& -\kappa y -\partial_y U(x,y),
\label{eq:2d}
\end{eqnarray}
with $\kappa \geq 0$ and $U(x,y)$ a random potential in the plane, obtained by bilinear interpolation,
such that for $x \in [n,n+1]$ and $y \in [m,m+1]$
we have
\begin{eqnarray}
U(x,y)&=& U_{n,m}(n+1-x)(m+1-y) \nonumber\\
&+& U_{n+1,m}(x-n)(m+1-y) \nonumber\\
&+& U_{n,m+1}(y-m)(n+1-x) \nonumber\\
&+& U_{n+1,m+1}(y-m)(x-n),
\label{eq:bilinear}
\end{eqnarray}
where $U_{n,m}$ are quenched independent random values, $\langle U_{n,m} \rangle=0$ and $\langle U_{n,m}U_{n',m'} \rangle=\delta_{n,n'}\delta_{m,m'}/3$, drawn from a uniform distribution.
In Fig.\ref{fig:figconfined1} we show results for $\kappa=1$. The quadratic mean displacement with respect to the mean motion at velocity $v$ in the direction of the drive (in the other direction the mean velocity vanished by symmetry) grows linearly with time, as indicated in Fig.\ref{fig:figconfined1}(a), and dispersion is normal with different $f$-dependent dispersion constants $D_x$, which are fitted as $\langle [x-vt]^2 \rangle \sim D_x t$ at long times. In Fig.\ref{fig:figconfined1}(b) we show that $D_x \sim 1/f^3$ at large $f$, being almost independent of the confinement in one decade of variation of the parameter $\kappa$. This dependence with $f$ is consistent with the one predicted for the one-dimensional system in a RB force field. In Fig.\ref{fig:figconfined2}(a) we show that the mean quadratic transverse displacement $\langle y^2 \rangle$ vs time $t$ saturates at a well defined $f$-dependent value. Interestingly, we find that $\kappa \langle y^2 \rangle \sim f^{-1}$, as indicated in Fig.\ref{fig:figconfined2}(b).
The result of Fig.\ref{fig:figconfined2}(b) can be interpreted as a sort of equipartition law in the transverse direction, with an effective transverse temperature vanishing as $T^y_{\tt eff}\sim f^{-1}$, at variance with the longitudinal fluctuations that vanish as $D_x \sim T^x_{\tt eff} \sim f^{-3}$.
To further test this picture in Fig.\ref{fig:figconfined3} we show the unconfined $\kappa=0$ case, for the longitudinal (a) and transverse (b) fluctuations. As can be observed, in this case transverse fluctuation lead to normal transverse dispersion. The dependencies $D_x\sim 1/f^3$ and $D_y\sim 1/f$ are consistent with the picture of two different temperatures, one for each direction. These temperatures satisfy a generalized fluctuation relation since the differential mobility in each direction becomes drive independent at large $v$. This kind of behaviour was already observed in a two-dimensional toy models with randomly located parabolic wells~\cite{Kolton2006}. A simple (but not general) argument starts by noting that in the particular bilinearly interpolated potential transverse force in a given square cell derived from Eq.(\ref{eq:bilinear}) is independent of $y$, i.e. $-\partial_y U(x,y)\equiv F_y(x)$. In the next cell the same holds but the force $F_y(x)$ changes, because the random potential at the vertex of the unit squares $V_{n,m}$ are uncorrelated quenched random numbers.
Since $\langle \partial_y U(x,y)\partial_y U(x',y')\rangle=\Delta_{yy}(x-x',y-y')$,
at large $v$ we can write $\langle F_y(x(t))F_y(x(t')) \rangle \approx \Delta_{yy}(x(t)-x(t'),y=0) \approx \Delta_{yy}(v(t-t'),y=0) \sim \delta(t-t')/v \equiv 2T^y_{\tt eff}(v)\delta(t-t')$, where in the first approximate equality we used that the particle moves mostly forward in a distance equal to one correlation length of the disorder along $x$, and in the fourth term that $\Delta_{yy}$ is short-ranged and satisfies $\int_u \Delta_{yy}(u) \neq 0$.
Therefore, we see that the transverse force mimics a thermal-noise at an effective transverse temperature vanishing as $T^y_{\tt eff} \sim 1/f \sim 1/v$.
A more general argument to understand the $1/f$ transverse fluctuations in the RB case at large velocities is given in Ref.\cite{Elias2022}, in the context of a vortex line dynamics.
\subsubsection{RF case}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figconfined1RF.pdf}
\caption{
RF results for the two dimensional model with transverse confinement, Eq.(\ref{eq:2d}).
(a) Quadratic mean displacement vs time with respect to the mean motion in the driving direction, for different driving forces $f$ for a confinement constant $\kappa=1$. The dashed line indicates normal dispersion.
(b) Fitted dispersion constant in the direction of the drive $D_x$ vs $f$ for different confinement constants $\kappa$. The dashed lines indicates the $1/f$ decay and dotted line indicate the $1/f^2$ behavior for comparison.
}
\label{fig:figconfined1RF}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figconfined2RF.pdf}
\caption{
RF results for the two dimensional model with transverse confinement $\kappa>0$, Eq.(\ref{eq:2d}).
(a) Mean squared displacement in the transverse direction as a function of time, for $\kappa=1$.
(b) The variance of transverse fluctuations approximately satisfies $\kappa \langle y^2 \rangle \sim f^{-1}$ at large forces for different confinement constants $\kappa$. Dotted line indicate the $1/f^2$ behavior for comparison.
}
\label{fig:figconfined2RF}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figconfined3RF.pdf}
\caption{
RF results for the two dimensional model \textit{without} confinement ($\kappa=0$), from Eq.(\ref{eq:2dRF}).
Longitudinal (a) and transverse (b) quadratic mean displacement vs time for different driving forces $f$, with respect to the mean motion. Dashed-lines in (a) and (b) highlight normal dispersion.
(c) Longitudinal and transverse dispersion constants $D_x$ and $D_y$ as a function of $f$. The dashed-line shows the $\sim f^{-1}$ decay.
}
\label{fig:figconfined3RF}
\end{figure}
In two dimensions the RF case contemplates various possible cases. The random force field can be compressible or incompressible, rotational or irrotational or a mixture of components with different properties. Here we consider a simple one, which is to take
$\overline{F_x(x,y)F_y(x',y')}=0$
and $\overline{F_x(x,y)F_x(x',y')}=\overline{F_y(x,y)F_y(x',y')}=\Delta(|x-x'|)\Delta(|y-y'|)$, with $\Delta(z)$ a positive short-ranged function.
Note that, unlike the $d=1$ RF case, this force can not be derived from a potential i.e. ${\bf F}({\bf x})\neq -\nabla U({\bf x})$.
In practice, for coordinates
$x \in [n,n+1]$ and $y\in (m,m+1)$
in cell $(n,m)$ we generate two uniformly distributed independent random numbers in the $(-1/2,1/2)$ interval, one for $F_x(x,y)$ and the other for $F_y(x,y)$~\footnote{Different choices, such as $F_x=F_0 \cos(\alpha_{n,m})$, $F_y=F_0 \sin(\alpha_{n,m})$, with $\alpha_{n,m}\in (0,2\pi)$ a random angle, yield equivalent results.}.
The equations of motion then read
\begin{eqnarray}
\dot x &=& f + F_x(x,y),
\nonumber \\
\dot y &=& -\kappa y +F_y(x,y).
\label{eq:2dRF}
\end{eqnarray}
In Fig.\ref{fig:figconfined1RF}(a) we show the quadratic mean displacements in the moving frame vs time. Dispersion in the longitudinal direction is normal and drive-dependent. Comparing with Fig.\ref{fig:figconfined1} for RB we can see a larger dispersion for the RF case. This is reflected in a larger dispersion coefficient $D_x$ in Fig.\ref{fig:figconfined1RF}(b), where we can also observe a $D_x\sim f^{-1}$ decay at large $f$, at variance with the RB case, but consistent with the prediction of the one dimensional toy model, Eq.(\ref{eq:RFresult}). We can also observe that this result is almost independent of the confinement constant $\kappa$ in Eq.(\ref{eq:2dRF}), as was also observed in the RB case.
In Fig.\ref{fig:figconfined2RF}(a) we show the behaviour of the transverse fluctuation vs time, for $\kappa>0$.
As expected, it has a well defined temporal mean value, which is drive dependent. In Fig.\ref{fig:figconfined2RF}(b) we show, for three values of $\kappa$, that $\kappa \langle y^2 \rangle \sim f^{-1}$, as for the RB case.
Interestingly, for the unconfined two dimensional case, $\kappa=0$, the dispersion is {\it isotropic} in the commoving frame, as can be appreciated in Fig.\ref{fig:figconfined3RF}(c), which is obtained by fitting the quadratic mean displacements of Figs. \ref{fig:figconfined3RF}(a) and (b).
\subsubsection{Summary}
In summary in the two-dimensional case (either confined or unconfined)
we find that:
\begin{itemize}
\item $D\sim 1/f^3 \sim 1/v^3$ for RB and
$D\sim 1/f \sim 1/v$ for RF,
at large $v$.
This shows that the results for the toy model, Eqs.(\ref{eq:RFresult}) and (\ref{eq:RBresult}) respectively, are robust. This is somehow expected for $\kappa \gg 1$, as the system becomes quasi one-dimensional. Nevertheless, we have also shown that it even holds for the truly two-dimensional $\kappa=0$ case.
\item In the $\kappa=0$ case there is a normal transverse dispersion with dispersion constant decaying $D_y\sim f^{-1}$, both for RF and RB. Nevertheless dispersion in the moving frame is highly anisotropic in the RB case while it is isotropic in the RF case. This striking difference can be used as a fingerprint of the underlying disorder.
\item Transverse fluctuations in the confined case $\kappa>0$, which decay as $\kappa \langle y^2 \rangle \sim f^{-1}$, are consistent with the effective temperature $T^{x}_{\tt eff}=D_x/2$ obtained from the unconfined $\kappa=0$ case.
\end{itemize}
\subsection{Soft particles: elastic dimer model}
Strictly, skyrmions, vortices, walls or solitons in a one-dimensional disordered track are not rigid objects, and hence quenched disorder affects not only the longitudinal dispersion of their centers of mass but also induce shape fluctuations which in turn affect the way the deformable object couples to the underlying quenched disorder.
To model this effect, and check the robustness of dispersion properties, we will consider a model of a soft localized object composed by two particles coupled with a spring of natural size $l$ in a one dimensional force field $F(x)$,
\begin{eqnarray}
\dot{x_1}&=&f - \kappa (x_1-x_2+l) + F(x_1), \nonumber \\ \dot{x_2}&=&f + \kappa (x_1-x_2+l) + F(x_2),
\label{eq:coupled}
\end{eqnarray}
where $\kappa>0$ a coupling constant controlling the (longitudinal) shape fluctuations, allowing us to go from the rigid object for $\kappa \gg 1$ to the soft limit for $\kappa \ll 1$.
For the simulations we use the same piece-wise $F(x)$ considered in sections \ref{sec:boxdist} for the RB and RF cases. We will be interested in the dispersion of the center of mass of the two particles.
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figcoupled.pdf}
\caption{
Quadratic mean displacement vs time of two coupled particles in a one dimensional track, Eq.(\ref{eq:coupled}), in the RF (a) and RB (b) disorder cases, for different driving forces. The dashed-line shows that dispersion is normal in both cases. (c) The corresponding dispersion constants, for different spring constants $\kappa$, vanish as $D_{RF}\sim 1/f \sim 1/v$ (dot-dashed line)
and $D_{RB}\sim 1/f^3 \sim 1/v^3$ (dashed- line) at large $v$.
}
\label{fig:figcoupled}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figbond.pdf}
\caption{
Bond fluctuations vs $f$ for two coupled particles, from Eq.(\ref{eq:coupled}), in the RF and RB cases for spring constant $\kappa=1$. As indicated by the dashed and dashed-dotted lines both follow an empirical $\approx 1/f^{7/4}$ decay (of different amplitude), contrasting with the different decays $D_{RF}\sim 1/f$ and $D_{RB} \sim 1/f^3$ (RB) observed in the dispersion constant of individual particles (see Fig.\ref{fig:figcoupled}(c)).
}
\label{fig:figbond}
\end{figure}
In Fig.\ref{fig:figcoupled} (a)-(b) we show the quadratic mean displacement in the RF and RB pinning fields respectively. In both cases dispersion is normal and the dispersion constant decays with increasing $f$. In Fig.\ref{fig:figcoupled} (c) we see that $D_{RF}\sim 1/f$ and $D_{RB}\sim 1/f^3$, as obtained for the toy particle model (Eq.(\ref{eq:RFresult}) and (\ref{eq:RBresult}) respectively), for various values of the spring constant $\kappa$. This shows again the robustness of these predictions.
Interestingly, as shown in Fig.\ref{fig:figbond}, we find that the bond fluctuations, measuring the mean Hook energy of the pair, follows the same $\sim 1/f^{7/4}$ decay at large $f$, though with a different prefactor. This shows on one hand that the effective equipartition law that approximately holds for the transverse fluctuations of transversely confined particles (see section \ref{sec:transverse}) does not work in this case, in the sense that bond fluctuations are clearly not controlled by the same effective temperature $T_{\tt eff}\approx D/\mu$ describing the center of mass dispersion.
This may be attributed to the fact that the effective disorder-induced noise is correlated between particles, because they follow each other and thus visit exactly the same quenched disorder after a mean characteristic time $\sim l/v$. Hence, this noise can not mimic thermal-like noise $\xi_n(t)$, which must satisfy $\langle \xi_n(t)\xi_m(t') \rangle \propto \delta_{n,m}\delta(t-t')$ for particles $n$ and $m$ in a coarse-grained level. It is nevertheless quite remarkable that the RB and RF cases seem to follow a unique decay law for bond fluctuations, in sharp contrast with $D_{RF}$ and $D_{RB}$ for the center of mass. It would be interesting to investigate if this system can be still be described by multiple effective temperatures, depending on the typical time-scale of the observable fluctuations.
\subsubsection{Summary}
In summary, we have shown that a simple model of a simple deformable object (an elastic dimer) driven in a disordered one-dimensional track, presents at large velocity a dispersion described by Eqs.(\ref{eq:RFresult}) and (\ref{eq:RBresult}) made for the one overdamped toy model. These predictions thus appear robust under shape fluctuations of soft but yet localized objects.
\subsection{Effect of temperature: driven Brownian particle}
\label{sec:langevin}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figtemp.pdf}
\caption{
Effect of temperature on the dispersion in the one dimensional overdamped particle model.
Quadratic mean displacement vs time for different temperatures in the RF (a) and RB (b) random force fields, for a given driving force $f=3$ (dashed-lines indicate normal dispersion).
(c) At large $f$, thermal noise becomes just an additive effect in the dispersion, and $D(f,T)\approx D(f,T=0)+2T$ (as indicated by the asymptotic dotted-line) independently of the RF or RB pinning nature.
}
\label{fig:temperature}
\end{figure}
So far we have discussed deterministic dispersion of different objects in narrow tracks which can be described as particles with one or more degrees of freedom.
When the dispersing objects are small enough thermal fluctuations may become important however and may contribute to the dispersion in the racetrack array. To study its effect on particles we can use the nondimensionalized Langevin equation
\begin{equation}
\dot x = F(x)+ f + \xi(t).
\label{eq:langevin}
\end{equation}
where $F(x)$ is the same random force field as in Eq.(\ref{eq:monomereq}) and $\xi(t)$ is a standard Langevin noise with the properties $\overline{\xi(t)}=0$, and $\overline{\xi(t)\xi(t')}=2T\delta(t-t')$ averaging over the ensemble of possible noise realizations. Here $T$ the nondimensionalized temperature (to get the physical temperature we multiply by the unit of temperature $f_0 d_0/k_B$). For simplicity and to compare with that case, we will consider $F(x)$ to be of the same types discussed in section \ref{sec:boxdist}.
At variance with the other deterministic cases analyzed before, before discussing the dispersion properties of Eq.(\ref{eq:langevin}) it is important to remark here the distinction between the so-called ``annealed'' and ``quenched'' dispersion constants~\cite{BouchaudGeorges1990}. When we describe the normal spreading of a packet of particles in a single track due to finite temperature we are dealing with the ``quenched'' dispersion constant. From Fig.\ref{fig:dispersion} it is clear that we are instead interested in the so-called ``annealed'' dispersion constant $D$. In the absence of disorder, $v=f$ and $D=2T$, since it is standard driven Brownian motion, and then the two dispersion constant coincide. In the presence of disorder however the ``annealed'' is in general larger than the ``quenched'' dispersion constant~\cite{LeDoussal1989}.
In the presence of disorder and thermal fluctuations the problem becomes in principle quite complex. On one hand there is dispersion even below the depinning threshold, as the particles are able, in each track, to jump their energy barriers separating the metastable states by thermal activation. The stochastic dynamics is non trivial and can lead in general to anomalous dispersion, specially at low dimensions and for correlated random forces. We refer the reader to Ref.\cite{BouchaudGeorges1990} for a detailed review on this problem.
Above the putative depinning threshold however, advection dominates and destroy
any long-range time correlations. Also, the residence time in each correlated cell has a finite mean and variance for any $T$. Hence, we expect the Central Limit Theorem to hold and normal dispersion to have a well defined $D$. Our aim here is to describe such regime and show how $D$ behaves vs $f$ and $T$, particularly well above the depinning threshold where the robust power-law scaling emerges as a function of $f$ or $v$.
In Fig.\ref{fig:temperature}(a)-(b) we show
the effect of temperature on the quadratic mean displacement vs time for different temperatures in the RF and RB cases respectively. As can be observed, dispersion is normal and temperature monotonically increases dispersion. To better appreciate its effect
in Fig.\ref{fig:temperature}(c) we show that the effect of temperature is approximately additive at large $f$: $D(f,T)\approx D(f,T=0)+ 2T$, regardless of the type of disorder.
\subsubsection{Summary}
The effect of temperature for the driven overdamped particle in a RF or RB disorder is just to add the trivial thermal dispersion constant $D_0=2T$, describing dispersion in the absence of disorder. This result is expected as the disorder-induced fluctuations become small compared with the (fixed strength) thermal bath-induced fluctuations.
Therefore, the scaling (Eqs.(\ref{eq:RFresult}) and (\ref{eq:RBresult})) remain robust in the sense that now $D(f,T)-D_0 \approx f^{-1} \approx v^{-1}$ for RF and $D(f,T)-D_0 \approx f^{-3} \approx v^{-3}$ for RB.
\subsection{A model for magnetic domain walls in wires}
The dynamics of magnetic domain walls in narrow tracks have attracted a great interest after the proposal of using it to devise a non-volatile racetrack memory~\cite{Parkin2008}.
As a starting point for modelling a magnetic domain wall (DW) driven by either an applied magnetic field and/or a current we will follow Ref.\cite{Thiaville2005} and consider the modified Landau-Lifshitz-Gilbert equation for the local magnetization vector $\vec{m}$,
\begin{eqnarray}
\dot{\vec{m}} &=&
\gamma_0 \vec{H} \times \vec{m} +
\alpha \vec{m} \times \dot{\vec{m}} -
(\vec{\mathcal{U}}.\vec{\nabla})\vec{m}
\nonumber \\
&+& \beta \vec{m} \times
[(\vec{\mathcal{U}}.\vec{\nabla})\vec{m}],
\label{eq:micromagnetics}
\end{eqnarray}
where $\vec{m}$ is a unit vector, $\gamma_0$ the gyromagnetic constant,
$\vec{H}$ the micromagnetic effective field and $\alpha$ the Gilbert damping constant.
The velocity $\vec{\mathcal{U}}$ is a vector directed along the direction of electron motion, with an amplitude $\mathcal{U}=JPg\mu_B/(2eM_s)$, where $M_s$ is the spontaneous magnetization, $J$ is the current density and $P$ its polarization rate. The non-dimensional adiabatic parameter $\beta$ is expected to be comparable to $\alpha$. Thermal fluctuations are neglected.
Solving Eq.(\ref{eq:micromagnetics}) for the dynamics of a driven domain wall is computationally expensive, due to the large number of degrees of freedom. Indeed, even in one dimension, where $\vec{m}\equiv \vec{m}(x,t)$, the system is extended. An usual approximation to leverage this difficulty is the collective coordinate approach which assumes a flat DW profile characterized by only two degrees of freedom: the DW position $q$ along the wire longitudinal $x$ direction, and the internal degree of freedom $\phi$ indicating the orientation of the magnetization at the DW center. For a transverse domain wall Thiaville \textit{et al.} \cite{Thiaville2005}
have obtained, for an homogeneous material,
\begin{eqnarray}
\frac{\alpha}{\Delta}\dot q + \dot \phi &=& \gamma_0 (H_a+ H_p(q)) + \frac{\beta}{\Delta}\mathcal{U} \\
\alpha \dot \phi - \frac{\dot q}{\Delta} &=& -\frac{\gamma_0 H_K}{2} \sin 2\phi - \frac{\mathcal{U}}{\Delta}
\end{eqnarray}
where $\Delta=[A/(K_0 + K \sin^2 \phi)]^{1/2}$, with $K_0$ and $K$ the effective longitudinal and transverse anisotropy respectively, and $H_K=2K/\mu_0 M_s$. For simplicity we will make a rigid wall approximation or assume that $K \ll K_0$, so to approximate $\Delta$ with a constant. We will also assume that quenched disorder (arising from intrinsic magnetic or non-magnetic defects in the material, thickness modulations or from the wire borders roughness) give place to an extra random field coupled only to $q$, $H_p(q)$, as in Lecomte \textit{et al.} \cite{Lecomte2009}.
By measuring $H_a$ and $H_p(q)$ in units of $H_K/2$, $u$ in units of $v_W=\Delta \gamma_0 H_K/2$, time in units of $(1+\alpha^2)\Delta/v_W$, and $q$ in units of $\Delta$ we get (overriding notation for nondimensionalized $q$, $\phi$ and $t$), the nondimensional dynamical system
\begin{eqnarray}
\dot q &=& \alpha [h + F(q)] + \sin(2\phi)+ u (1+\alpha \beta)
\label{eq:microm1}
\\
\dot \phi &=& h+ F(q) + (\beta-\alpha) u -\alpha \sin(2\phi),
\label{eq:microm2}
\end{eqnarray}
controlled by the dimensionless parameters $\alpha$ and $\beta$, and drivings $h=2H_a/H_K$ and $u=\mathcal{U}/v_w$, with $F(q)=2H_p(q)/H_K$ the dimensionless pinning field.
It is worth noting that if we freeze the internal degree of freedom $\phi$, of the DW the model reduces to the overdamped mechanical model, essentially identical to Eq.(\ref{eq:monomereq}).
We also note that in the absence of disorder ($F(q)=0$) at long times $\dot \phi=0$ if $h+(\beta-\alpha)u < \alpha$. This threshold corresponds to the model approximation of the Walker breakdown, separating the so-called stationary regime, where $\alpha \dot q = h + \beta u$ and $\dot \phi=0$, from the precessional regime where $\langle \dot \phi \rangle \neq 0$. With pinning, there is a fixed point $\dot q = \dot \phi =0$ or pinned state for small enough yet finite drives.
To solve Eqs.(\ref{eq:microm1}) and (\ref{eq:microm2}) we will consider again the piece-wise random-force field $F(q)$ used in section \ref{sec:boxdist}, both in the RF and RB cases. We will consider both the case of pure magnetic field driving ($u=0$) and the pure current driving $(h=0)$, and the cases $\langle \dot \phi \rangle=0$ and $\langle \dot \phi \rangle \neq 0$,
and will focus on their large velocity regimes, $v\equiv \langle \dot q \rangle \sim u(1+\alpha \beta)$
and
$v\equiv \langle \dot q \rangle \sim \alpha h$, respectively.
\subsubsection{RF case}
In Fig.\ref{fig:DWRF}(a)-(b) we show the results for the RF case for the current driven ($h=0$, $u>0$) and field driven ($h>0$, $u=0$) cases respectively, for drivings well above the corresponding depinning thresholds.
In Fig.\ref{fig:DWRF}(a) we show the rescaled dispersion constant $D$ for the current driven case for $u$ values such that $v\equiv \langle \dot q \rangle \approx u(1+\alpha\beta)$. We empirically find that $D=|\alpha-\beta|\alpha^3 G_u(u|\alpha-\beta|\alpha)$, with $G_u(x)$ a scaling function, describes well the data for different pairs of values $(\alpha,\beta)$ which are sound for real materials. At large $u|\alpha-\beta|\alpha \gg 1$ we find $D\sim u^{-1}$ in agreement with the toy model with a RF disorder, Eq.(\ref{eq:RFresult}). Interestingly however, for smaller $u|\alpha-\beta|\alpha \ll 1$ a crossover towards a $D\sim u^{-3}$ is observed, coinciding with the decay predicted from the toy model for a RB disorder.
In Fig.\ref{fig:DWRF}(b), for the field driven case, we show that $D \approx h^{-1}$ at large $h$, with a power-law decay in $h$ in agreement with the RF case of the toy model, Eq.\ref{eq:RFresult}. Interestingly, no crossover is observed in this case.
A qualitative explanation of the results of Fig.\ref{fig:DWRF} can be given.
The crossover in the current driven case (Fig.\ref{fig:DWRF}(a)) occurs at $u^*\approx |\alpha-\beta|^{-1}\alpha^{-1}$. For $u>u^*$, we have $D\approx \alpha^2/u$, which is consistent with an RF disorder of amplitude $\alpha$ in the one degree of freedom toy model, Eq.(\ref{eq:RFresult}). Therefore, the $\sin(2\phi)$ term in Eq.(\ref{eq:microm1}) appears to have a negligible influence in the dispersion compared with the one expected from the term $\alpha F(q)$. To explain this we note, on one hand, that the larger $\alpha$ (the smaller $u^*$), the more important $\alpha F(q)$ is compared with the average effect of $\sin(2\phi)$ on the dispersion.
One the other hand, when $u<u^*$ dispersion is dominated by the $\sin(2\phi)$. Subtly, its effect on the dispersion of $q$ is similar to the one of a RB disorder, yielding $D\sim 1/u^3$. Nevertheless, this case can not be described by a simple one degree of freedom toy model such as Eq.(\ref{eq:monomereq}) for an effective RB force field $F(q)$ (see appendix \ref{sec:reducido}).
The above arguments also explain qualitatively the field driven case of Fig.\ref{fig:DWRF}(b), since in this case the characteristic frequency of fluctuations of the $\sin(2\phi)$ term is always of order $h$ and can not be reduced by changing $\alpha$ and/or $\beta$. Since $v\sim h$, the average effect of $\sin(2\phi)$ can be neglected and then $D\sim \alpha^2 h^{-1}$ as expected for the toy model in a RF disorder of amplitude $\alpha$, Eq.(\ref{eq:RFresult}).
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{domain_wall_RF.png}
\caption{Scaled dispersion constant of a domain wall with RF disorder, from the numerical solution of Eq.(\ref{eq:microm2}). In (a) we show the current driven case ($h = 0,\; u > 0$), for $\alpha=\{0.02, 0.2, 0.25, 0.35 0.5, 0.75, 1.0\}$ and $\beta = \{0.0,0.02, 0.04, 0.08, 0.1, 0.2\}$ (all $\{\alpha, \beta\}$ pairs). The data is in agreement with a scaling of the form $D=|\alpha-\beta|\alpha^3 G_u(u|\alpha-\beta|\alpha)$ where $G_u(x) \sim x^{-1}$ for $x \gg 1$ and $G_u(x) \sim x^{-3}$ for $x \ll 1$. In (b) we show the pure field driven case ($u = 0,\; h > 0$) for $\alpha=\{0.01, 0.02, 0.2, 0.35 0.5, 1.0\}$. No crossover is observed and the data follows approximately the relation $D/\alpha \propto h^{-1}$.
\label{fig:DWRF}}
\end{figure}
\subsubsection{RB case}
In Fig.\ref{fig:DWRB}(a)-(b) we show the results for the RB case for the current driven ($h=0$, $u>0$) and field driven ($h>0$, $u=0$) cases respectively, from the numrical solution of Eq.(\ref{eq:microm2}). We observe $D\sim u^{-3}$ and $D\sim h^{-3}$, in agreement with the toy model prediction Eq.(\ref{eq:RBresult}) for RB disorder.
Interestingly, no crossover is observed in the current-driven case of Fig.\ref{fig:DWRF}(a). Using the arguments above, this may be explained by the fact that in this case both, $\alpha F(q)$ and (in a subtly effective way) $\sin(2\phi)$, are of a RB type, and thus $D\sim u^{-3}$ is in agreement the toy model, Eq.(\ref{eq:RBresult}) for RB disorder.
Similarly, the field driven case can be explained by the fact that $\sin(2\phi)$ in Eq.(\ref{eq:microm1}) oscillates at a frequency $\sim 2h$ larger than $v$, which is the characteristic frequency induced by $\alpha F(q)$. Therefore, its average effect on dispersion vanishes, and the system effectively behaves as the toy model in the RB case, (Eq.\ref{eq:RBresult}).
Power-law prefactors in Figs.\ref{fig:DWRB}(a)-(b) are different and less simple to predict in the RB case as compared to the RF case.
The scaling collapse in Figs.\ref{fig:DWRB}(a)-(b)
is purely empiric.
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{domain_wall_RB.png}
\caption{Scaled dispersion constant of a domain wall with RB disorder for $\alpha=\{0.3, 0.4, 0.5 0.6, 0.8\}$. In (a) in shown the pure current driven case ($h = 0,\; u > 0$) for $\beta = \{0.0 , 0.02, 0.04\}$ showing the expected RB behavior as $D \cdot \alpha \propto 1/u^3$. In (b) in shown the pure field driven case ($u = 0,\; h > 0$) showing again the expected RB behavior but $D \cdot (1-\alpha)^{-1} \propto 1/h^3$. In both cases, no crossover is observed, at variance with the current driven RF case (see Fig.\ref{fig:DWRF}(a)).}
\label{fig:DWRB}
\end{figure}
\subsubsection{Summary}
In summary, we have shown that a zero-dimensional domain-wall derived from micromagnetic considerations, presents the following features:
\begin{itemize}
\item At large drives, either from an applied current or magnetic field, $D$ decays as expected from the toy model predictions Eqs.(\ref{eq:RFresult})-(\ref{eq:RBresult}) for the RF and RB cases respectively.
\item In all cases $D$ scales differently with respect to the Gilbert damping parameter $\alpha$ and the adiabatic parameter $\beta$.
\item A crossover between different power-law decays is observed in $D$ only in the RF case, for the range of $\alpha$ and $\beta$ analyzed.
\end{itemize}
\section{Discussion}
\label{sec:discussion}
Our results show that the way the dispersion constant of particles $D$ vanishes at large velocities is independent of many model details. It is nevertheless particularly sensitive to the nature of disorder. One may consider other types of random-force fields, such as long-ranged or periodically correlated ones, and the way $D$ vanishes may change. The RB and RF cases are nevertheless rather important physical cases to start with. For zero-dimensional magnetic domain walls for instance (i.e. when can neglect its transverse fluctuations and consider it as a flat interface), the RB type is generated by non magnetic-impurities, rough borders, modulated thickness or by a space-dependent magnetic anisotropy, while the RF is realized in the presence of magnetic-impurities inducing a random magnetic field. The situation is similar for zero-dimensional ferroelectric domain walls, where the two types of quenched disorders can be in principle realized. On the other hand, magnetic skyrmions, colloids or vortices typically couple to a random bounded potential and hence realize a RB disorder.
Extended systems may be also considered, such as one- or two-dimensional domain walls anchored by a narrow channel of a finite width. Skyrmions chains in one dimensional tracks also fall in the extended system category. In these systems with many elastically coupled degrees of freedom the random-force acting on the center of mass may be renormalized. A prominent example are elastic walls in a short-ranged correlated random-force field, for which a RB microscopic disorder is renormalized to a cuspy RF-type near depinning. The force-force correlator then gets rounded at larger velocities as the dynamical correlation length reduces~\cite{chauve2000}. In these cases we expect our results for the dispersion to hold for the center of mass respecting the shape of the renormalized disorder instead of the microscopic one. It is plausible that, in such case, a scenario similar to the one described in section \ref{sec:rb+rf} takes place at large velocities. Indeed, for a driven elastic vortex line in three dimensions, with RB disorder, an effective temperature vanishing as $\sim v^{-3}$ was found for a wide range of velocity in spite that the depinning transition is better described by a renormalized RF-type random force field~\cite{Elias2022}.
Bringing back the dimensions, our results show that at large velocities, we can roughly expect
\begin{eqnarray}
v&\approx& (v_0/f_0)f, \\
D&\approx& v_0 d_0 (f_0/f)^n,
\end{eqnarray}
where $f_0$ is the pinning force amplitude, $d_0$ the disorder correlation length $v_0=\eta f_0$ is the characteristic velocity, $\eta$ the friction constant, and $n=1$ or $n=3$ for RF and RB disorder respectively. To compare with experiments we may use that the experimentally accessible depinning force is $f_c \sim f_0$ (neglecting interaction effects between particles) and hence $v_c \equiv \eta f_c \sim v_0$. The disorder correlation length $d_0$ is usually unknown or difficult to estimate directly, but one can use that $d_0=\max(R,r_0)$, where $R$ is the size of the particle and $r_0$ the correlation length-scale of the inhomogeneity. We can thus write a more practical and simple prediction
\begin{eqnarray}
D\approx v d_0 (f_c/f)^{n+1} \approx v d_0 (v_c/v)^{n+1}.
\label{eq:Dwithunits}
\end{eqnarray}
From an applied viewpoint, to realize a memory bus analog, it would be desirable to have particles in different tracks concurrently and precisely controlled. If we need for instance to transport coherently and at the largest possible velocity a few bytes from one check point to the next by moving particles in different parallel tracks, the less possible dispersion would be ideal. To make a concrete estimation let us define the maximum time $t_{\rm max}$ and distance $x_{\rm max}$ we can move the particles at given velocity $v$ with a given tolerance $\delta$ for the dispersed packet width, i.e $D t_{\rm max}= \delta^2$. It makes sense to ask for a maximum tolerated error to be of the order of one particle (one bit) size,
i.e. $\delta \sim R$.
We then get
\begin{eqnarray}
x_{\rm max} \equiv
v t_{\rm max} \approx
\frac{R^2}{d_0}\left(\frac{f}{f_c}\right)^{n+1} \approx
\frac{R^2}{d_0}\left(\frac{v}{v_c}\right)^{n+1}.
\label{eq:predictions}
\end{eqnarray}
For skyrmions for instance, if we assume point disorder, we have $d_0 \approx R$. Then, using a skyrmion size $R\approx 10 \text{ nm}$ moving at $v\approx 0.34 \text{ mm/s}$ for $f/f_c \equiv j/j_c \approx 5$~\cite{Fert2013} we obtain $x_{\rm max} \approx 6 \text{ }\mu\text{m}$ and $t_{\rm max}\approx 0.025\text{ s}$ for a coherent parallel transport. If the disorder were otherwise of a pure RF type ($n=1$) then $x_{\rm max}$ and $t_{\rm max}$ would be more than one order of magnitude smaller. On the other hand, if the microscopic correlation length of the inhomogeneity is $r_0>R$ the previous estimates for $x_{\rm max}$ and $t_{\rm max}$ enhance by a factor $R/r_0$.
Rough estimates similar to the above can be made for domain walls, superconducting vortices or colloids, provided we know a few parameters such as the size of the particles
and the velocity-force characteristics under the assumption of a short-ranged random force correlator. Besides specific applications however, Eq.(\ref{eq:Dwithunits}) is a rather robust prediction which might be useful to characterize the disorder in the host media, regardless of its microscopic origin, by measuring the dispersion $D$ of independent racing particles as a function of the driving force.
\begin{acknowledgments}
We thank Kay Wiese for illuminating discussions. We also acknowledge support from grant PICT-2019-01991.
\end{acknowledgments}
\section{Introduction}
A precise control of the motion in narrow tracks of stable localized excitations is desirable to realize different devices of current interest in applied physics research.
Paradigmatic examples are magnetic domain walls ~\cite{Parkin2008,Emori2013,Blasing2020} or skyrmions \cite{Kiselev2011,Schulz2012,Fert2013,Reichhardt2021} moving in an array of magnetic nanowires.
Furthermore, such kind of devices may be also of interest for other driven objects embedded in different materials, such as ferroelectric domain walls~\cite{Catalan2012,Paruch2013}, vortices in superconductors~\cite{Kwok2016}, colloids~\cite{Lowen2008,Volpe2014,Bechinger2016}, general brownian particles~\cite{hanggi2009}, as well as kinks and solitons~\cite{Martinez2008,Laliena2020,Osorio2021}.
Although each one may have a different internal structure we will generically refer to all these driven objects as ``particles''.
One of the interesting applications proposed for the above mentioned devices is the ``racetrack memory''~\cite{Parkin2008}, a novel type of non-volatile memories where particles are used as carriers of information in an array of tracks. This proposal assumes the ability
to write, read and move the particles in the tracks. To achieve a competitive capacity and efficiency compared with existing memories, these tracks have been reduced to the nanoscale, making the consideration of (the usually unavoidable) quenched spatial heterogeneities in the host materials of prime importance.
Indeed, spatial inhomogeneities may have disruptive effects, mainly metastability, a depinning transition, and a shaking effect on the sliding particles. As a consequence, the motion of the driven particles becomes more difficult to predict, specially in the ubiquitous case of quenched disorder. Moreover, in the case of track arrays, each track has a different realization of the quenched disorder, so different tracks may produce a different particle trajectory under an identical protocol. This is also a very important issue if we aim to synchronize the motion of many particles in many independent tracks, such as in a ``memory bus'' and, for instance, perform computation with the whole bunch of data.
Although the effects of quenched disorder or quenched periodic potentials on driven particles have been extensively discussed in the literature, particularly focusing in the collective transport in two dimensions or higher
~\cite{Reichhardt2016,Reichhardt2021}, a detailed statistical study of the effect of quenched disorder on independent particles driven in an array of quasi one-dimensional racetracks is lacking.
In this paper we address the problem of independent particles driven in an array of disordered tracks with the aim to identify robust statistical features, independent of specific model details. To be concrete we focus in the specific ``obstacle race'' problem illustrated schematically in Fig.\ref{fig:dispersion}. At $t=0$ many particles start at $x=0$ in a track array, one per track, and are driven by the same constant force $f$. As tracks are independent realizations of an otherwise statistically identical quenched disorder with some assumed short-range correlations, at any given time $t>0$ particles disperse, even at zero temperature, provided the force exceeds the depinning threshold (assumed to be finite). We will be mainly interested in the mean velocity $v$ and in the dispersion constant $D$ around the mean forward motion of all the particles, as a function of $f$. Such race problem is similar, except some subtle differences we discuss, to the well studied problem of dispersion of tracers in porous media. However, the results turn out to be radically different: while in the latter the dispersion constant $D$ in general increases with increasing the average flow~\cite{BouchaudGeorges1990}, in the problem we address here it decreases with increasing the average flow $v$. We find, in particular, that $D$ has rather robust properties at large $f$, in the limit when $v\to f$, though quite sensitive to the nature of disorder. For the two physically important cases that we study here, the so-called Random-Bond (RB) and the Random-Field (RF) type, we find that $D\sim f^{-3}$ and $D\sim f^{-1}$ respectively.
These results are obtained from different toy models where the full velocity-force characteristic and dispersion can be analytically obtained.
More complicated models, motivated in different physically relevant situations are solved numerically and shown to display the same dispersion beyond a velocity crossover.
In section \ref{sec:mechanicalmodel} we introduce the main properties of interest in relation to a simple mechanical model. Then, in section \ref{sec:exactresults} we introduce several models that can be solved analytically for the properties of interest, $v$ and $D$. These examples illustrate various effects and in particular suggest a universal behaviour of the dispersion constant $D$. In section \ref{sec:trapmodel} we show how these results can be connected to those of a stochastic trap model in the context of flow in porous media.
In order to test the robustness of the different analytical results, in section \ref{sec:numerics} we compare them with the numerical solution of various physically relevant models for which an exact solution is difficult to obtain. Finally, in section \ref{sec:discussion} we discuss the results and summarize our conclusions.
\section{A mechanical toy model}
\label{sec:mechanicalmodel}
To be concrete let us consider a damped particle of mass $m$ driven by a constant uniform force $f$ in a one-dimensional space presenting quenched random forces at {\it zero temperature} (see section \ref{sec:langevin} for a discussion about temperature effects),
\begin{equation}
m \ddot x = {F}({x})+ {f} - \eta \dot {x},
\end{equation}
where $\eta$ is a friction constant and ${F}(x)$ is a short-range correlated quenched random force field such that
\begin{eqnarray}
\overline{{F}({x})}&=&0,\\
\overline{{F}({x}){F}({x'})}&=& f_0^2 g(|{x}-{x'}|/d_0).
\label{eq:dampedmechmodel}
\end{eqnarray}
Here $d_0$ is a characteristic length,
$f_0$ a characteristic force amplitude, and $g(u)$ a rapidly decaying function of unit range and $g(0)=1$. We will be particularly interested in the cases where $\int_x g(x)=0$, corresponding to the so-called ``Random-Bond'' case (RB), and the case where $\int_x g(x)>0$, corresponding to the so-called ``Random-Field'' (RF) case. If we write $F(x)=-U'(x)$ (which is always possible to do in $d=1$ but not necessarily in $d>1$), and the potential $U(x)$ is bounded, then $F(x)$ is RB type. If otherwise the potential diffuses as $\langle [U(x)-U(x')]^2 \rangle \sim |x-x'|$ for long $|x-x'|$, then it is of the RF type. The great physical relevance of these two particular cases is better appreciated by noting that the RB type can be generated by forces derived from a bounded short-range correlated random potential while the RF type can be generated by a short-range correlated random force-field.
In both cases we will also assume that $F(x)$ is bounded, so a finite critical depinning force $f_c = \max_x[-F(x)]$ exists,
such that only for $f>f_c$ a steady-state motion is generated.
\footnote{See Ref.\cite{LeDoussal2009}for a thorough and general study of the one particle depinning in different universality classes corresponding to the three extreme value statistics, Gumbel, Weibull, and Fréchet.}
Without any loss of generality
we can nondimensionalize the equation of motion Eq.(\ref{eq:dampedmechmodel}) by measuring distances in units of $d_0$, forces in units of $f_0$ and time in units of $\tau_0=\eta d_0 /f_0$, and mass in units of $m_0=\eta^2 d_0/f_0$. The derived velocity unit is therefore $v_0=d_0/\tau_0=f_0/\tau_0 \eta$.
Then, overriding notation for all nondimensionalized quantities we get
\begin{equation}
m \ddot x = {F}({x})+ {f} - \dot {x}.
\label{eq:equationwithinertia}
\end{equation}
For a given initial condition and disorder realization $F(x)$ the solution $x(t)$ of this equation are only parameterized by the nondimensional mass $m$.
We will be mainly interested in the statistics of the time-dependent fluctuations induced by quenched disorder in these solutions.
Given the position $x(t)$ of a particle as a function of time $t$ for many particular realizations of $F(x)$, we will be mainly interested in the average velocity $v$ and the dispersion constant $D$ in the commoving frame at long times, as a function of the driving force $f$ and in the steady-state regime. Assuming $x(t=0)=0$, they are defined as
\begin{eqnarray}
v &=& \lim_{t\to \infty} \langle x(t)\rangle/t
\label{eq:vdef} \\
D &=& \lim_{t\to \infty} \langle [x(t)-v t]^2 \rangle/t
\label{eq:Ddef}
\end{eqnarray}
where the $\langle \dots \rangle$ denotes average over disorder realizations
\footnote{Note that we are thus interested here in the so-called ``annealead'' dispersion constant, different than the ``quenched'' one that describes the spreading of a packet in a single channel due to thermal noise or chaotic behaviour at zero temperature~\cite{LeDoussal1989}.}.
\begin{figure}[htp]
\centering
\includegraphics[width=0.5\textwidth]{dispersion}
\caption{Schematics of the particle (filled circles) deterministic dispersion problem at finite force $f$ above the depinning threshold, induced by quenched disorder in an array of narrow disordered racetracks (grey bars). Tracks are different disorder realizations of statistically {\it identical} media (light-blue lines illustrate two force-fields $F(x)$ and $F'(x)$). Particles disperse around the average motion $vt$ in the the right direction, with $v$ the disorder-averaged or track-averaged velocity. The ``obstacle race'' starts at $t=0$, and for increasing times (indicated with different colors) the particle distribution is shown below. Dispersion is measured by the dispersion constant $D$. The dependence of $v$ and $D$ with $f$ contains useful and sensitive information about the type of disorder in the media. }
\label{fig:dispersion}
\end{figure}
To give these definitions a concrete physical interpretation we can think of particles running on parallel racetracks, each one with its own realization $F(x)$ but identical statistical properties, with one particle per track (see Fig.\ref{fig:dispersion}). Then the average velocity $v$ is then the center of mass velocity and $D t$ is the variance of their distribution around the center of mass, for a large collection of particles.
We can think the channel array is a memory bus and the particles as carriers of information. In such imaginary application $D$ and $v$ are relevant quantities to perform efficient parallel computation. If particles are actually localized objects of a certain finite size, such as magnetic domain walls, magnetic skyrmions, solitons or colloidal particles, the zero temperature approximation may be well justified.
In addition to $v$ and $D$ We will be also interested in the steady-state force dependent differential mobility, defined as
\begin{eqnarray}
\mu = \frac{dv}{df}
\end{eqnarray}
as it allows us to define an effective temperature from a generalized Einstein relation
as
\begin{eqnarray}
T_{\rm eff} = \frac{D}{2\mu}.
\label{eq:Teffdefinition}
\end{eqnarray}
$T_{\rm eff}$ will in general depend on the driving force $f$, or the velocity $v$.
\section{Exact results}
\label{sec:exactresults}
{\it Exact results for any $f$} can be readily obtained in the $m\to 0$ overdamped case
\begin{equation}
\dot x = F(x) + f.
\label{eq:monomereq}
\end{equation}
when $F(x)$ is a piece-wise short-range correlated random function, such that the random-force $F(x)$ is constant in successive regular intervals (i.e. a random-steps force-field). If we then chose a bounded distribution for $F(x)$, a finite depinning threshold $f_c = \max[-F(x)]$, such that $v=0$ for $f<f_c$, is guaranteed to exist. Within this family of models we find (see appendix \ref{sec:appendixformulaD}),
\begin{eqnarray}
v &=& 1/\langle \Delta t \rangle \label{eq:vformula}
\\
D &=& \frac{\langle \Delta t^2 \rangle-\langle \Delta t \rangle^2}{\langle \Delta t \rangle^3}
\label{eq:Dformula}
\end{eqnarray}
where $\Delta t$ the time a particle spends on a given cell of size $\Delta x = 1$.
Therefore, the calculation of $v$ and $D$ reduces to single cell
averages in the steady-state.
Below we derive results for two particular cases, a Random-Field RF case and a Random-Bond RB case and later explore their
robustness.
\subsection{Box distribution of random forces}
\label{sec:boxdist}
Let us start with an instructive toy model with
uniformly distributed bounded forces at each cell.
\begin{figure}[htp]
\centering
\includegraphics[width=0.5\textwidth]{fig0}
\caption{A piece-wise RF force field (a) and its potential (c). A piece-wise RB force field (b) and its potential (d). In both cases the depinning threshold is $f_c \to 1$ for large systems. Exact calculations show that both cases have the same $v$ but different $D$.
}
\label{fig:RFandRB}
\end{figure}
\subsubsection{RF case}
To construct a RF disorder such that $\int_y g(y)>0$ we take
\begin{eqnarray}
F(x)=R_{[x]}
\end{eqnarray}
where $[\dots]$ denotes the integer part. Constant force intervals are hence of size $\Delta x=1$ and labeled with an integer $n$.
The $R_n$ are uniformly distributed
random numbers in the interval $[-1,1]$
such that
\begin{eqnarray}
\overline{R_n R_m} &=&\frac{\delta_{n,m}}{3} .
\end{eqnarray}
A sample of this RF force field at short distances is shown in Fig.\ref{fig:RFandRB}(a), and it corresponding potential
$U(x)=-\int_x F(x)$ in in Fig.\ref{fig:RFandRB}(c).
From Eq.(\ref{eq:monomereq})
the time spent in the interval $n$
is
\begin{eqnarray}
\Delta t_n = \frac{1}{R_n+f} ,
\end{eqnarray}
and hence
\begin{eqnarray}
\langle \Delta t\rangle &=&
\frac{1}{2} \int_{-1}^{1}
\frac{1}{R+f} \; dR
= \frac{1}{2}\ln \left(\frac{f+1}{f-1}\right), \\
\langle \Delta t^2 \rangle &=&
\frac{1}{2} \int_{-1}^{1}
\frac{1}{(R+f)^2} \; dR
= \frac{1}{f^2-1}.\qquad
\end{eqnarray}
The {\it exact} mean velocity for $f \geq 1$ is
\begin{eqnarray}
v =
\frac{2}{\ln \left(\frac{f+1}{f-1}\right)},
\label{eq:exactvuniform}
\end{eqnarray}
displaying a depinning transition at $f=1$.
Approaching the threshold, $f\to 1+$, the velocity vanishes continuously as $v \sim -2/\ln (f-1)$. This model is thus peculiar from the standard critical phenomena point of view where we expect a power-law $v \sim (f-1)^\beta$ for an order parameter $f$ and a control parameter $f$
(see however section \ref{sec:sintheta}).
On the other hand, for $f\gg 1$, $v\simeq f$ as expected from the washing out of the pinning force at large velocities and thus free flow is recovered. The velocity force characteristics can be also inverted, yielding
\begin{eqnarray}
f = \frac{e^{2/v}+1}{e^{2/v}-1}.
\end{eqnarray}
The exact (differential) mobility for $f \geq 1$ is
\begin{eqnarray}
\mu =
\frac{4}{
(f^2-1) \ln^2 \left(\frac{f+1}{f-1}\right)},
\end{eqnarray}
such that $\mu \to \infty$ when $f\to 1$ (i.e. right at the depinning transition), and $\mu\to 1$ when $f\to \infty$ as expected.
Using Eq.(\ref{eq:Dformula}) the exact dispersion constant for $f \geq 1$ is
\begin{eqnarray}
D =
\frac{8 \left(\frac{1}{f^2-1}-\frac{1}{4} \ln ^2\left(\frac{f+1}{f-1}\right)\right)}{\ln ^3\left(\frac{f+1}{f-1}\right)}
\label{eq:exactDRF}
\end{eqnarray}
and therefore, using the generalized Einstein relation we get the exact effective temperature
\begin{eqnarray}
T_{\rm eff}=
\frac{2 \left(1-\frac{f^2-1}{4} \ln ^2\left(\frac{f+1}{f-1}\right)\right)}{\ln \left(\frac{f+1}{f-1}\right)
}
\end{eqnarray}
We are interested in the $f\gg 1$ fast-flow behaviour of $D$ and $T_{\tt eff}$. Expanding in powers of $1/f$ we first get
\begin{eqnarray}
v &\simeq& f - \frac{1}{3f} - \frac{4}{45 f^3} + \mathcal{O}(f^{-5})
\label{eq:largevexpansion}
\\
\mu &\simeq& 1 + \frac{1}{3f^2}+\mathcal{O}(f^{-4}),
\end{eqnarray}
and in particular for the dispersion and associated effective temperature we get
\begin{eqnarray}
D &\simeq& \frac{1}{3f} + \frac{7}{45 f^3 } + \mathcal{O}(f^{-5}) \nonumber \\
&\simeq&
\frac{1}{3v} + \frac{2}{45 v^3} + \mathcal{O}(v^{-4}),\\
T_{\tt eff}&\simeq& \frac{1}{3f}+\frac{2}{45f^3}
+\mathcal{O}(f^{-5}).
\end{eqnarray}
We will particularly focus our attention on the large-velocity dominant term,
which in this case is
\begin{eqnarray}
D\sim 1/f \sim 1/v,\;\;\; \text{RF}.
\label{eq:RFresult}
\end{eqnarray}
The above RF scaling can be obtained by first order perturbation theory in $f^{-1}$. At zero-th order
Eq.(\ref{eq:monomereq}) implies $x \approx f t$,
and then at first order $\dot{x}=f+F(f t)$. The random fluctuating force $F(ft)$ mimics a colored noise because $\langle F(ft) \rangle=0$, $\langle F(ft)F(ft') \rangle = g(f(t-t'))$, where $g(x)$ is short-ranged. Most importantly, because we are considering RF disorder, $\int g(x)>0$ and thus we can define a positive effective temperature $T_{\rm eff} \propto \int_{-\infty}^{\infty} g(ft)dt \sim f^{-1}$. Hence, since $\mu\to 1$, $D\sim 1/f$.
\subsubsection{RB case}
\label{sec:RBcaseUniform}
To model RB disorder
with $\int_y g(y)=0$,
we define the random forces in Eq.(\ref{eq:monomereq}) as
\begin{eqnarray}
F(x) = R_{[x]} \text{sign}(x-[x]-1/2),
\end{eqnarray}
where as above the $R_n$ are iid random variables, uniformly distributed in $[-1,1]$.
This choice splits the unit interval of each cell in two parts of size $1/2$, such that the constant force in the second half is minus the force of the first half.
A sample of this RF force field at short distances is shown in Fig.\ref{fig:RFandRB}(b), and it corresponding potential
$U(x)=-\int_x F(x)$ in in Fig.\ref{fig:RFandRB}(d).
Evidently this is a particular way to obtain the auto-correlation needed for a RB type of disorder. As we show below, the choice also has the advantage of leaving $v$ and $\mu$ invariant from $RF$ to $RB$.
By repeating the procedure of the previous section we obtain $\langle \Delta t \rangle$ as for the RF case,
and thus {\it identical} $v$ and $\mu$ as a function of $f$.
However, $\langle \Delta t^2 \rangle$ is different,
\begin{eqnarray}
\langle \Delta t^2 \rangle &=&
\frac{1}{2} \int_{-1}^{1}
\left[ \frac1{ 2(f+R)} + \frac1{2 (f-R)} \right] ^2\; dR \\
&=&
\frac{1}{4 f}
\left[\frac{2f}{f^2-1}+\ln \left(\frac{f+1}{f-1}\right)\right].
\end{eqnarray}
The {\it exact} dispersion constant is
\begin{eqnarray}
D=
\frac{2 \left(\frac{2}{f^2-1}-\log ^2\left(\frac{f+1}{f-1}\right)+\frac{\log \left(\frac{f+1}{f-1}\right)}{f}\right)}{\log ^3\left(\frac{f+1}{f-1}\right)}.
\label{eq:exactDRB}
\end{eqnarray}
$D$ thus diverges at the depinning threshold
and at large driving forces vanishes as
\begin{eqnarray}
D \simeq T_{\tt eff} \simeq \frac{4}{45 f^3} + \mathcal{O}(f^{-5})
\simeq \frac{4}{45 v^3} + \mathcal{O}(v^{-5})
\end{eqnarray}
which is clearly different from the RF case.
Then, the dominant term in this case is,
\begin{eqnarray}
D\sim 1/f^3 \sim 1/v^3,\;\;\; \text{RB}
\label{eq:RBresult}
\end{eqnarray}
faster than the RF case.
It is worth noting that the $f^{-3}$ RB scaling, at variance with the $f^{-1}$ RF, thus can not be obtained from first order perturbation in $f^{-1}$. Indeed, dispersion is (incorrectly) zero at first order, because for RB we have
$\int g(x)=0$.
\subsubsection{Summary}
In Fig.\ref{fig:boxdist} we summarize the exact results for this model.
We show that {\it identical} $v$ vs $f$ characteristics can be accompanied by different $D$ vs $f$ characteristics.
We show in particular that $D$ vanish differently at large $f$ in the RF and RB cases, thus becoming a sensitive tool to characterize the random-medium. It is worth noting that the decay of $D$ with $f$ depends crucially on the fact that the random-medium produce finite random forces whose effect becomes weaker at large velocities. A simple disordered model where this is not the case and $D$ increases with $f$ is discussed in the appendix \ref{sec:randomfriction}.
\begin{figure}[htp]
\centering
\includegraphics[width=0.5\textwidth]{fig1.pdf}
\caption{$v$ and $D$ vs $f$ for RF and RB force fields with a box-distribution. In this model, $v$ is invariant with respect to the RF or RB nature of the disorder.
At $f\to 1$ (depinning threshold) $v$ vanishes logarithmically, while $D$ diverges. At large forces, as indicated, $v \sim f$, while $D \sim 1/f$ and $D \sim 1/f^3$ for the RF and RB cases respectively, as indicated with dotted-lines.}
\label{fig:boxdist}
\end{figure}
\subsection{``RB + $\epsilon$ RF'' case}
\label{sec:rb+rf}
We now discuss the case where disorder is not pure RB or RF but where a RB disorder is perturbed by a RF disorder. To achieve this we split the $n$-esime unit interval in two equal parts, and put a uniformly distributed random force $R_n$ in the first half, followed by a random force $-R_n+\delta_n$ in the second half. We will consider that $\delta_n$ is a small random variable with $\langle \delta_n \rangle=0$ and $\langle \delta_n^2 \rangle=\epsilon^2 \ll 1$.
We will also consider that $\langle R_m \delta_n\rangle=0$ so to ensure that $v$ remains invariant under the perturbation.
The $n$-esime residence time is
\begin{eqnarray}
\Delta t_n &=& \frac{1}{2} \left( \frac{1}{f+R_n}+\frac{1}{f-R_n+\delta_n} \right) \nonumber \\
&\approx& \Delta t_{RB} + \Delta t_{RF}
+ \mathcal{O}(\epsilon^2)
\end{eqnarray}
where
\begin{eqnarray}
\Delta t_{RB} &=& \frac{f}{f^2-R_n^2}, \\
\Delta t_{RF} &=& -\frac{\delta_n}{2(f-R_n)^2},
\end{eqnarray}
are the pure RB and RF random contributions.
Since $\langle R_m \delta_n\rangle=0$
we have
$\langle \Delta t_{RF} \rangle \propto \langle \delta_n \rangle =0$.
Therefore the RF perturbation does not alter the mean velocity, as $v^{-1}=\langle \Delta t_n \rangle=\langle \Delta t_{RB} \rangle$.
Nevertheless,
\begin{equation}
\langle \Delta t^2 \rangle
= \langle \Delta t_{RB}^2 \rangle +
\langle \Delta t_{RF}^2 \rangle
\end{equation}
where in the last equality we have used $\langle \Delta t_{RB} \Delta t_{RF}\rangle \propto \langle \delta_n \rangle =0$. Since we already know
$\langle \Delta t^2_{RB} \rangle$ from Sec.\ref{sec:RBcaseUniform} we have only to calculate $\langle \Delta t^2_{RF} \rangle$, which for uniformly distributed forces is simply
\begin{eqnarray}
\langle \Delta t^2_{RF} \rangle&=&
\langle \delta_n^2 \rangle \frac{1}{2}\int_{-1}^1 d R \left(\frac{1}{2(f-R)^2}\right)^2
\nonumber \\
&=&
\frac{\epsilon^2 \left(3 f^2+1\right)}{12 \left(f^2-1\right)^3}
\end{eqnarray}
Since
\begin{eqnarray}
D= \frac{
\langle \Delta t_{RB}^2 \rangle- \langle \Delta t^2 \rangle
+\langle \Delta t_{RF}^2 \rangle}{\langle \Delta t \rangle^3}
\end{eqnarray}
we obtain
\begin{eqnarray}
D= D_{RB} + \frac{\langle \Delta t_{RF}^2\rangle}{\langle \Delta t\rangle^3} = D_{RB} + v^3 \langle \Delta t_{RF}^2\rangle.
\end{eqnarray}
To the leading term $\langle \Delta t_{RF}^2\rangle \approx 1/(4 f^4)$, $D_{RB}\approx 4/(45f^3)$ and $v \sim f$, we have
\begin{eqnarray}
D \approx \frac{4}{45f^3} + \frac{\epsilon^2}{4f}.
\end{eqnarray}
Therefore, the RF perturbation dominates at large enough velocities and $D \sim 1/f$. The crossover in this model roughly occurs at
$f^* \sim \frac{4}{3\sqrt{5}\epsilon}$ if $\epsilon \ll 1$.
In Fig.\ref{fig:RBperturbRF} we summarize the exact results for this toy model.
\begin{figure}[htp]
\centering
\includegraphics[width=0.5\textwidth]{fig2}
\caption{
Crossover behaviour in a RB force field perturbed with a RF force field (both with a box-distribution of forces), for perturbing parameter $\epsilon=0.01$. While $v$ remains invariant under the small perturbation $D$ crossovers from $RB$ to $RF$ behaviour at large enough forces, as indicated with dotted-lines.
}
\label{fig:RBperturbRF}
\end{figure}
The crossover just described may be important for instance for flat domain walls in media with RB disorder such as non-magnetic impurities, crystalline defects or rough borders, contaminated with a few magnetic impurities which produce RF pinning. On the other hand, it may be important for particles described by two or more coupled degrees of freedom, since in that case the force field seen by the coordinate describing its position in the track is effective.
More subtly, the crossover is important for the case of extended systems such as (non-flat) elastic interfaces but with pure RB microscopic disorder, since in that case the renormalized pinning force is not pure RB anymore.
At interface depinning indeed, the renormalized disorder becomes of a RF type and hence a merging of RF and RB into a unique RF depinning universality class emerges. However, as the velocity increases, such renormalization is expected to be weaker because the correlation length ${l}$ along the interface becomes smaller (near $f_c$ as $l \sim (f-f_c)^{-\nu}$), and then the renormalized pinning force tend to flow down to the RB microscopic disorder again. A toy model that beautifully illustrates this physics near the depinning threshold is the the particle quasistatically dragged in a random force field by a parabolic potential of curvature $m^2$~\cite{LeDoussal2009}.
Then the effect of increasing $v$ can be related to the effect of increasing $m \sim 1/l$, and one can fully appreciate the flow from RB to RF by approaching the depinning threshold.
The results for the present toy model hence suggests that even the small RF part that remains from renormalization at large drives (or large $m$) should be still detectable beyond a putative crossover of $D$ as a function of $f$.
\subsection{A non-uniform distribution of random forces}
\label{sec:sintheta}
In order to show the robustness of the previous results we now consider a simple model variant with the same scaling properties for $D$ at large velocities in the RF and RB case, but quite different critical behaviour near the depinning threshold. The particular variant proposed also display much simpler exact analytical expressions for all quantities and could thus be used as a convenient toy model for other studies.
In order to achieve this we change the probability distribution of random-forces in each cell such that
\begin{eqnarray}
R_n = \sin \theta_n
\end{eqnarray}
with $\theta_n$ uniformly distributed in the $[-\pi,\pi)$ interval. This disorder may physically correspond to a particle coupled to the $x$ component of a random unit vector force in a plane.
\subsubsection{RF case}
As before, we compute the mean residence time in a cell
\begin{eqnarray}
\langle \Delta t \rangle =
\frac{1}{2\pi} \int_{-\pi}^{\pi}
\frac{d\theta}{\sin(\theta)+f} \; = \frac{1}{\sqrt{f^2-1}},
\end{eqnarray}
and hence the exact velocity and differential mobility are respectively
\begin{eqnarray}
v = \sqrt{f^2-1}\\
\mu = \frac{f}{\sqrt{f^2-1}}.
\end{eqnarray}
Interestingly, this is exactly the mean velocity of the
$\dot x = f + \sin(x)$ model and
it has a finite and trivial critical exponent $v\sim (f-1)^{\beta}$, $\beta=1/2$, different than the previous logarithmic depinning case. Nevertheless, at variance with the particle in the periodic potential, where there is no dispersion at zero temperature, here we have a disorder-induced dispersion with respect to the disorder-averaged displacement.
To see this we compute as before
\begin{eqnarray}
\langle \Delta t^2 \rangle =
\frac{1}{2\pi} \int_{-\pi}^{\pi}
\frac{d\theta }{(\sin(\theta)+f)^2} \;
= \frac{f}{(f^2-1)^{3/2}}
\label{eq:nonUniformDt2}
\end{eqnarray}
Then, using Eq.(\ref{eq:Dformula}), we get the {\it exact} dispersion constant
\begin{eqnarray}
D = f-\sqrt{f^2-1}
= -v + \sqrt{1+v^2}
\end{eqnarray}
which {\it remains finite} as $f\to 1$ ($D\to 1$) and
at large forces behaves as \begin{eqnarray}
D
\approx
\frac{1}{2f}+\frac{1}{8 f^3} + \mathcal{O}(f^{-5})
\approx
\frac{1}{2v}-\frac{1}{8v^3}+\mathcal{O}(v^{-5})
\end{eqnarray}
These results show that the RF dependence $D\sim 1/f$ at large velocities (Eq.(\ref{eq:RFresult})) is robust under the change of the step distribution we have made, in spite that the critical depinning behaviour is clearly different.
\subsubsection{RB case}
As before, we use the trick of dividing the unit interval in two correlated parts,
\begin{eqnarray}
F(x) = \sin \theta_{[x]} \text{sign}(x-[x]-1/2).
\end{eqnarray}
where $\theta_n$ is uniformely distributed in $[\pi,\pi)$ as before. The convenience of this choice is, as before, that $v$ remains invariant.
The residence time in a single cell then is split in two contributions
\begin{eqnarray}
\Delta t = \frac{1/2}{f+\sin(\theta)}+\frac{1/2}{f-\sin(\theta)}
\end{eqnarray}
Therefore,
\begin{eqnarray}
\langle \Delta t \rangle &=&
\frac{1}{4\pi} \int_{-\pi}^{\pi}
\frac{d\theta}{\sin(\theta)+f}
+\frac{1}{4\pi} \int_{-\pi}^{\pi}
\frac{d\theta}{-\sin(\theta)+f}
\nonumber \\
&=& \frac{1}{\sqrt{f^2-1}},
\end{eqnarray}
{\it identical} to the RF case of the previous subsection, and hence $v$ and $\mu$ are also identical.
Nevertheless
\begin{eqnarray}
\langle \Delta t^2 \rangle =
\frac{2f^2-1}{2f(f^2-1)^{3/2}}
\end{eqnarray}
is different than the RF case, Eq.(\ref{eq:nonUniformDt2}).
The {\it exact} dispersion constant is
\begin{eqnarray}
D=-\sqrt{f^2-1}+f-\frac{1}{2 f},
\end{eqnarray}
which {\it remains finite} as $f \to 1$ ($D\to 1/2$) and its large velocity expansion yields
\begin{eqnarray}
D &=& \frac{1}{8f^3 }+\mathcal{O}(f^{-5})
\end{eqnarray}
These results show now that the RB dependence $D\sim 1/f^3 \sim 1/v^3$ at large velocities (Eq.(\ref{eq:RBresult})) is also robust under a change of the step distribution.
In Fig.\ref{fig:sintheta} we summarize the exact results for this model.
We note first that when $f\to 1+$, $v\sim (f-1)^{1/2}$, in contrast with the logarithmic behaviour (Eq.(\ref{eq:exactvuniform})) obtained for the box distribution in the previous sections. Moreover, we find that $D \to 1$ (RF) and $D \to 0.5$ (RB), also in sharp contrast with the divergent dispersion constants in the same cases for the box distribution, Eqs.(\ref{eq:exactDRF}) and (\ref{eq:exactDRB})
respectively.
Nevertheless, the (RF) $1/f$ and (RB) $1/f^3$ asymptotics of $D$ at large $f$ remains a robust feature.
\begin{figure}[htp]
\centering
\includegraphics[width=0.5\textwidth]{fig3.pdf}
\caption{$v$ and $D$ vs $f$ for the RF and RB cases of the model with a non-uniform force distribution.
At the depinning threshold $v\sim (f-1)^{1/2}$ and $D\to 1$ and $D\to 1/2$ (indicated with Red dashes lines) for RF and RB respectively, in sharp contrast with the box-distribution model.
At large forces, we recover the robust $D\sim 1/f$ and $D\sim 1/f^3$ behaviours of the RF and RB cases respectively (indicated with dotted-lines).
}
\label{fig:sintheta}
\end{figure}
\subsection{Conexion with a trap model for dispersion}
\label{sec:trapmodel}
Dispersion under an average flow is an old problem, relevant for the physics of porous media where the dispersion constant $D$ measures the spread of an injected packet of tracer particles.
In this context a pedagogical toy model proposed by Bouchaud and Georges~\cite{BouchaudGeorges1990} allows to understand the basic physics and estimate $D$. In that model the medium is idealized as one dimensional and made of a ``backbone'' along which the particle is convected with velocity $V$ at regularly spaced positions (separated by a distance $\xi$) and with probability $p$ can leave the backbone during a random waiting time $\tau$. Importantly, the particle can not go backwards.
\begin{figure}[htp]
\centering
\includegraphics[width=0.5\textwidth]{bouchaud}
\caption{Adapting the Bouchaud and Georges trap model (a) to the quenched-force field toy model above the depinning threshold, at zero temperature (c). To convert it in (b) we consider that that free flow is $V=f$, and that the effective trapping occurs with probability one with waiting times that are random but quenched, each one given by the ``excess residence time'' in a cell $\tau_n \equiv \frac{\xi}{F_n+f}-\frac{\xi}{f}$, where $n\equiv [x]$ is the cell number. }
\label{fig:bouchaud}
\end{figure}
Bouchaud and Georges showed that
\begin{eqnarray}
U^{-1}&=&V^{-1}+p\langle \tau \rangle/\xi
\label{eq:bouchaudgeorgesU}
\\
D_\parallel &=& (pU^3/2\xi)(\langle \tau^2 \rangle - p \langle \tau \rangle^2).
\label{eq:bouchaudgeorgesD}
\end{eqnarray}
where $U$ is the average velocity of the particle in the medium and $D_\parallel$ denotes the longitudinal dispersion constant. These results show that for a given drive $V$, $D_\parallel$ {\it grows} with $V$, in good qualitative agreement with some porous media experiments. It is interesting to note that at large velocities this model predicts $D_\parallel \sim V^3$, in sharp contrast with our results for uncorrelated random media, $D\sim 1/f^3 \sim 1/v^3$ (RB) and $D\sim 1/f \sim 1/v$ (RF).
In the following we explain such difference and show how to reconcile the results of Eqs.(\ref{eq:bouchaudgeorgesU}) and (\ref{eq:bouchaudgeorgesD}) with the predictions of our dispersion model.
To explain the differences between the Bouchaud-Georges model predictions and the ones we presented in the previous sections we start by noting that the average in Eq.(\ref{eq:bouchaudgeorgesD}) is over different stochastic trajectories in one track, while in
Eq.(\ref{eq:Dformula}) the average is over disorder realizations (see Fig.\ref{fig:bouchaud}). The former is the so-called quenched dispersion constant while the later is the so-called annealed dispersion constant. We show below that this does not prevent however to find a concrete connection between the two mathematical approaches.
We also note that in the problem we are interested in, the delay $\langle \Delta t \rangle$ in each cell is not given by a fixed distribution, but by a drive dependent distribution. In particular, the larger $f$, the closest is $\langle \Delta t \rangle$ to $1/f$.
Hence, to make a concrete connection with the overdamped mechanical model we replace the trap in the $n$-esime cell of size $\xi=1$ by a quenched random force field
and redefine $\tau$ as the {\it excess time} with respect to the free motion time in the cell $=1/f$.
Note that replacing $\tau$ by quenched random values is not problematic because the particle never moves backward. That is,
\begin{eqnarray}
\tau \equiv \Delta t - \frac{1}{f},
\end{eqnarray}
and we also make the identification $V \equiv f$ and $U \equiv v = 1/\langle \Delta t \rangle$. We can also take $p=1$ since the non-trapping event can be absorbed in the distribution of $\tau$, corresponding to $F_n=0$. With such analogy, noting that $2 D_\parallel \equiv D$ (because of the convention we have used to define $D$, see Eq.(\ref{eq:Ddef})), we finally get
\begin{eqnarray}
v^{-1}&=& V^{-1}+ \langle \tau \rangle = {\langle \Delta t\rangle} \\
D &=& U^3 (\langle \tau^2 \rangle - \langle \tau \rangle^2)
= \frac{\langle \Delta t^2 \rangle - \langle \Delta t \rangle^2}{\langle \Delta t \rangle^3}
\end{eqnarray}
which is identical to our expressions, see Eq.(\ref{eq:Dformula}).
The later adaptation and reinterpretation of the different terms shows that our results for the mechanical models ultimately reduce to compute a few statistical properties (first and second moments) of the excess waiting time in a trap equivalent model.
Interestingly, the mechanical model yields always a {\it decreasing} $D$ with increasing $v$, and the microscopic nature of the short-correlated disorder is important to determine how exactly $D$ vanishes at large velocities.
\section{Numerical Results for different models}
\label{sec:numerics}
In order to assess the universality of some of the analytical predictions we have made in the previous sections for simple one-dimensional toy models, in this section we compare them with more complicated physically relevant disordered models, for which an exact solution is not available, but precise numerical simulations can be performed.
We consider overdamped and damped models, the effect of temperature and the effect of transverse fluctuations in wires with a finite or infinite width.
We also consider the dispersion of a non-rigid particle and a micromagnetic model of a magnetic domain wall driven by an external electrical current and/or applied magnetic field.
Solving all the proposed models reduce to integrate a system of ordinary non-linear differential equations of at most two variables, with disorder. In all cases we solve them by standard integration techniques. We use between $8192$ to $32768$ independent disorder realizations (or tracks) to average the properties of interest, and run long-time simulations so to capture the diffusive behaviour beyond any finite time crossover.
\subsection{Inertia: massive damped particles}
Embedded soft matter systems, such as magnetic domain walls, magnetic skyrmions, or superconducting vortices behave effectively as overdamped driven objects and inertia can be ignored. Inertia may be important however for other experimental realizations which could be modeled by a driven massive damped particle or soliton in a quenched random field. Inertia opens, on the other hand, a new phenomenology because the state space is doubled compared to the overdamped case. In our particular one-dimensional case of Eq.(\ref{eq:equationwithinertia}), the state of the damped particle must be described by two variables instead of one, and it is a priori not evident whether the universal results we obtain for $D$ in the RF and RB cases remain valid in this case. Since to the best of our knowledge no analytical solutions for the dispersion properties are known for Eq.(\ref{eq:equationwithinertia}) for finite $m$, here we solve it numerically.
For the simulations we solve the nondimensional Eq.(\ref{eq:equationwithinertia}) using the same piece-wise $F(x)$ considered in sections \ref{sec:boxdist} for the RB and RF cases. From Eq.(\ref{eq:equationwithinertia})
we note that $v \approx f$ in the large $f$ limit, because the average acceleration must be zero in presence of damping. We will be particularly interest in the fluctuations when $f-v \ll f$.
\subsubsection{RF case}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figMSDVarP.pdf}
\caption{Quadratic mean displacement (a) and momentum variance (b) in the moving frame vs $t$ for the RF disorder case of Eq.(\ref{eq:equationwithinertia}), using $m=0.1$, and indicated driving forces. The dashed-lines indicates normal {dispersion}.
}
\label{fig:msdRF}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figDvsfvsMRF.pdf}
\caption{Numerical results for the inertial RF disorder case of Eq.(\ref{eq:equationwithinertia}). (a) Rescaled velocity deficit as a function of the drive $f$, for different masses $m$ (indicated in (b)), in the fast flow regime $f-v \ll v$. Dash and dashed-dotted lines indicate approximate asymptotic behaviours.
(b) Dispersion constant vs $f$ for different masses. The dotted line shows a $\sim 1/f$ decay.
(c) Rescaled momentum variance vs $f$. The dahed-dotted and dashed lines shows asymptotic predictions (see text).
}
\label{fig:figinertiaRF}
\end{figure}
We start describing the results for the RF field.
In Fig.\ref{fig:msdRF}(a) we show the quadratic mean displacement in the moving frame $\langle [x-v t]^2\rangle$ vs $t$, where $v=\langle \dot x \rangle$ is the steady-state mean velocity. We see that dispersion is normal and thus a driving force dependent dispersion constant $D$ can be fitted. In Fig.\ref{fig:msdRF}(b) we see that the variance of the momentum in the moving frame $\langle p-m v \rangle$, shown for the same mass and driving forces, has a well defined drive dependent steady-state value (temporal fluctuations are controlled by the number of tracks considered in the average).
In Fig.\ref{fig:figinertiaRF}(a) we show that in the fast-flow regime
$f-v \approx m G_v(fm)$, with $G_v(x)$ displaying approximately a crossover between two power laws, $G_v(x)\sim x^{-1}$ for $x\ll 1$ and $G_v(x)\sim x^{-3}$ for $x\gg 1$. For a fixed $f$ and a small $m$ the result is consistent with the expansion obtained for the over-damped toy model, Eq.(\ref{eq:largevexpansion}). Interestingly, there is a crossover to a different regime for a fixed value of $f m$, so the smaller $m$ the larger the crossover force $f_m\sim m^{-1}$.
Therefore, the expansion of Eq.(\ref{eq:largevexpansion}) is not robust.
In Fig.\ref{fig:figinertiaRF}(b) we show nevertheless that the dispersion constant $D$ is practically independent of $m$ and decays as $D \approx f^{-1}/3$, with the same power-law predicted analytically with the toy model in the RF case, Eq.(\ref{eq:RFresult}).
In Fig.\ref{fig:figinertiaRF}(c)
we show that the steady-state variance of the momentum obeys also a scaling $\langle p^2 \rangle-\langle p \rangle^2 \sim m^2 G_p(f m)$ with a crossover, at $f m \sim 1$, from a constant to a power law decay, $G_p(x)\approx x^{-1}/12$. Interestingly, the exact fits to $D$ and $\langle p^2 \rangle-\langle p \rangle^2$ for large $f m$ are consistent with the relation
\begin{equation}
\frac{\langle [p-\langle p \rangle]^2 \rangle}{2m}
\approx \frac{D}{4} \approx \frac{T_{\tt eff}(f)}{2}
\label{eq:equipartitionRF}
\end{equation}
where in the last equality we have used the definition of the effective temperature from the generalized Einstein relation of Eq.(\ref{eq:Teffdefinition}), with $\mu\approx 1$ for $f-v \ll f$.
In other words, the mean kinetic energy calculated in the moving frame with velocity $v=\langle p/m \rangle$ appear to satisfy the equilibrium equipartition law in spite of being a system far from equilibrium.
We can understand this result at the lowest order in a high velocity expansion. If we write that the random force is $F(x)\approx F(vt)$, and that $\dot x \approx v \approx f$, from Eq.(\ref{eq:equationwithinertia})
\begin{equation}
m \ddot u \approx {F}(vt)-\dot u.
\label{eq:langevinRF}
\end{equation}
for the position coordinate $u=x-vt$ in the moving frame.
Then, using that $\overline{F(x)F(x')}=g(|x-x'|)$ with $g(x)$ a rapidly decaying function with $\int_x g(x)>0$ for RF,
we can compute the time autocorrelation function of the random force for $v \gg |t-t'|^{-1}$, $\overline{F(vt)F(vt')}=g(v|t-t'|)\sim \delta(t-t')/v$.
Hence, in such large velocity limit, the random force behaves effectively as a Langevin noise $\xi_v(t)\equiv F(vt)$ \footnote{The ensemble average of the Langevin noise coincides in this limit with the average over tracks.} at an effective temperature $T_{\tt eff}\sim v^{-1}$ (since the mobility is constant and $\mu=1$). The system thus reaches thermal equilibrium at the $T_{\tt eff}$, and hence the mean kinetic energy is $\langle \frac{m}{2}\dot u^2 \rangle = T_{\tt eff}/2$,
in agreement with the asymptotic result of Eq.(\ref{eq:equipartitionRF}).
On the other hand, for $fm<1$ the particle enters a regime where the friction force $\dot x$ dominates over the inertial term $m\ddot x$
and then, using the same arguments as above we get
\begin{equation}
\dot u \approx F(vt)
\label{eq:smallfm}
\end{equation}
which implies that $\langle \dot u^2 \rangle = \langle F(vt)^2 \rangle=1/3$. Therefore,
$\langle [p-mv]^2 \rangle/2 m^2 = 1/6$ in agreement with the scaled data shown in Fig.\ref{fig:figinertiaRF}(c) for $fm\ll 1$.
\subsubsection{RB case}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figMSDVarPRB.pdf}
\caption{Quadratic mean displacement (a) and momentum variance (b) in the moving frame vs $t$, using $m=0.1$ and indicated driving forces, for the RB disorder case of Eq.(\ref{eq:equationwithinertia}). The dashed-line in (a) indicates normal {dispersion}.
}
\label{fig:msdRB}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figDvsfvsMRB.pdf}
\caption{Numerical results for the inertial RB case of Eq.(\ref{eq:equationwithinertia}). (a) Rescaled velocity deficit as a function of the drive $f$, for different masses $m$ (indicated in (b)), in the fast flow regime $f-v \ll v$. Dash and dashed-dotted lines indicate approximate asymptotic behaviours, as indicated.
(b) Dispersion constant vs $f$ for different masses. The dotted line highlights the $\sim f^{-3}$ decay.
(c) Rescaled momentum variance vs $f$. The dashed-dotted and dashed lines indicate approximate asymptotic behaviour. Grey symbols in (a) and (c) correspond to the results of the RF case.
}
\label{fig:figinertiaRB}
\end{figure}
We now describe the results for the RB field.
In Fig.\ref{fig:msdRB}(a) we show the quadratic mean displacement in the moving frame $\langle [x-v t]^2\rangle$ vs $t$, where $v=\langle \dot x \rangle$ is the steady-state mean velocity. We see that {dispersion} is normal and thus a driving force dependent dispersion constant $D$ can be fitted. In Fig.\ref{fig:msdRB}(b) we see that the variance of the momentum in the moving frame $\langle (p-m v)^2 \rangle$, shown for the same mass and driving forces, has a well defined drive-dependent steady-state value.
In Fig.\ref{fig:figinertiaRB}(a) we show, for various masses (indicated in Fig.\ref{fig:figinertiaRB}(b)), that the deficit of velocity with respect to the free velocity obeys approximately the same scaling $(f-v)\sim m G_f(f m)$ found for the RF case, and also presents a crossover between two regimes. The first regime displays the correction $G_f(x)\sim x^{-1}$ expected from the large velocity expansion of the overdamped toy-model Eq.(\ref{eq:largevexpansion}). The second regime $G_f(x)\sim x^{-2}$ for $f m \gg 1$, shows that inertia can have an important effect, and thus the scaling of Eq.(\ref{eq:largevexpansion}) is not universal.
In Fig.\ref{fig:figinertiaRB}(b) we show nevertheless that the dispersion coefficient $D$ follows closely the $f^{-3}$ predicted by the toy model in the RB case, Eq.(\ref{eq:RBresult}).
It is nevertheless worth noting that while in the RF case $D$ is almost independent of $m$, in the RB case there is a strong dependence, as evidenced by the rescaling of curves into a master curve.
We also note that $D$, for the same range of $m$ and $f$, is in general much smaller in the RB case than in the RF case, and more difficult to sample precisely.
Finally, in Fig.\ref{fig:figinertiaRB}(c)
we show that the momentum variance follows the same scaling $\langle p^2\rangle-\langle p\rangle^2 = m^2 G_p(fm)$ than in the RF case, and a similar crossover from small to large $fm$. Interestingly, the, small $fm$ regime saturates to an identical value than for the RB case. This is consistent with the effective equation of motion Eq.(\ref{eq:smallfm}), with the saturation value controlled only by the variance of the disorder.
The large $fm$ regime however, where empirically we find $G_p(x)\sim x^{-1.75}$, differs appreciably from the RB case. The faster decay of
$G_p(x)$ is however consistent with the faster decay of $G_f(x)$
shown in Fig.\ref{fig:figinertiaRB}(a).
At variance with the RF case, The large $fm$ regime can not be rationalized using an effective equipartition theorem because $D \sim T_{\tt eff}$ does not decay in the same way in this limit.
The interpretation of this remains open.
\subsubsection{Summary}
In summary, we find that damped particles display:
\begin{itemize}
\item $D\sim 1/f$ and $D\sim 1/f^3$ as predicted for the overdamped toy model for the RF and RB case respectively, Eqs.(\ref{eq:RFresult}) and (\ref{eq:RBresult}). The result thus appears to be robust under inertia.
\item A non trivial dependence with the mass $m$ both in the $f-v$ and in $\langle p^2 \rangle-\langle p \rangle^2$ reveals a crossover between two regimes varying the parameter $f m$. The first regime, identical for RB and RF is well described by the overdamped limit with an effective drive-dependent noise. The second regime is different for RF and RB. In the RF case the second regime is compatible with a generalized equipartition law at the disorder-induced drive dependent effective temperature $T_{\tt eff}=D/2$. For RB the later equipartition law {\it fails}.
\end{itemize}
\subsection{Quasi one- and two-dimensional tracks}
\label{sec:transverse}
When the size of the particles is smaller than the track width transverse fluctuations of a transversely confined particle may modify the predicted longitudinal dispersion properties. This kind of situation may arise for instance in the cases of driven vortices in narrow and thin superconducting strips, driven colloids, or current driven magnetic skyrmions. Compared to the simple one-dimensional overdamped case, now the state-space doubles (as in the one-dimensional damped case) and the effective one-dimensional coupling to the underlying disorder changes.
We study the consequences of these new properties in the velocity and dispersion of particles.
\subsubsection{RB case}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figconfined1.pdf}
\caption{
RB results for the two dimensional model of Eq.(\ref{eq:2d}) with transverse confinement in a two-dimensional random bounded potential.
(a) Quadratic mean displacement with respect to the mean motion in the driving direction vs time, for different driving forces $f$ for a confinement constant $\kappa=1$. The dashed line indicates normal dispersion.
(b) Fitted dispersion constant in the direction of the drive $D_x$ vs $f$ for different confinement constants $\kappa$. The dashed lines indicates the $1/f^3$ behaviour. For comparison $1/f^2$ and $1/f^4$ behavior are drawn in dotted lines.
}
\label{fig:figconfined1}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figconfined2.pdf}
\caption{
RB results for the two dimensional model of Eq.(\ref{eq:2d}) with transverse confinement ($\kappa>0$) for the two-dimensional uncorrelated random potential of Eq.(\ref{eq:bilinear}).
(a) Mean squared displacement in the transverse direction as a function of time, for $\kappa=1$.
(b) The variance of transverse fluctuations approximately satisfies $\kappa \langle y^2 \rangle \sim f^{-1}$ at large forces for different confinement constants $\kappa$. Dotted line indicate the $1/f^2$ behavior for comparison.}
\label{fig:figconfined2}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figconfined3.pdf}
\caption{
RB results for the two dimensional model (Eq.(\ref{eq:2d})) \textit{without} confinement ($\kappa=0$), for the two-dimensional uncorrelated random potential of Eq.(\ref{eq:bilinear}).
Longitudinal (a) and transverse (b) quadratic mean displacement vs time for different driving forces $f$, with respect to the mean motion.
(c) Longitudinal and transverse dispersion constants $D_x$ and $D_y$ as a function of $f$. The dashed-line shows the $\sim f^{-1}$ and the dashed-dotted the $\sim f^{-3}$ behaviours.
}
\label{fig:figconfined3}
\end{figure}
In order to study the effect of transverse fluctuations we will consider here only the relevant RB case of an overdamped particle in a narrow channel transversely confined by a parabolic potential
\begin{eqnarray}
\dot x &=& f - \partial_x U(x,y),
\nonumber \\
\dot y &=& -\kappa y -\partial_y U(x,y),
\label{eq:2d}
\end{eqnarray}
with $\kappa \geq 0$ and $U(x,y)$ a random potential in the plane, obtained by bilinear interpolation,
such that for $x \in [n,n+1]$ and $y \in [m,m+1]$
we have
\begin{eqnarray}
U(x,y)&=& U_{n,m}(n+1-x)(m+1-y) \nonumber\\
&+& U_{n+1,m}(x-n)(m+1-y) \nonumber\\
&+& U_{n,m+1}(y-m)(n+1-x) \nonumber\\
&+& U_{n+1,m+1}(y-m)(x-n),
\label{eq:bilinear}
\end{eqnarray}
where $U_{n,m}$ are quenched independent random values, $\langle U_{n,m} \rangle=0$ and $\langle U_{n,m}U_{n',m'} \rangle=\delta_{n,n'}\delta_{m,m'}/3$, drawn from a uniform distribution.
In Fig.\ref{fig:figconfined1} we show results for $\kappa=1$. The quadratic mean displacement with respect to the mean motion at velocity $v$ in the direction of the drive (in the other direction the mean velocity vanished by symmetry) grows linearly with time, as indicated in Fig.\ref{fig:figconfined1}(a), and dispersion is normal with different $f$-dependent dispersion constants $D_x$, which are fitted as $\langle [x-vt]^2 \rangle \sim D_x t$ at long times. In Fig.\ref{fig:figconfined1}(b) we show that $D_x \sim 1/f^3$ at large $f$, being almost independent of the confinement in one decade of variation of the parameter $\kappa$. This dependence with $f$ is consistent with the one predicted for the one-dimensional system in a RB force field. In Fig.\ref{fig:figconfined2}(a) we show that the mean quadratic transverse displacement $\langle y^2 \rangle$ vs time $t$ saturates at a well defined $f$-dependent value. Interestingly, we find that $\kappa \langle y^2 \rangle \sim f^{-1}$, as indicated in Fig.\ref{fig:figconfined2}(b).
The result of Fig.\ref{fig:figconfined2}(b) can be interpreted as a sort of equipartition law in the transverse direction, with an effective transverse temperature vanishing as $T^y_{\tt eff}\sim f^{-1}$, at variance with the longitudinal fluctuations that vanish as $D_x \sim T^x_{\tt eff} \sim f^{-3}$.
To further test this picture in Fig.\ref{fig:figconfined3} we show the unconfined $\kappa=0$ case, for the longitudinal (a) and transverse (b) fluctuations. As can be observed, in this case transverse fluctuation lead to normal transverse dispersion. The dependencies $D_x\sim 1/f^3$ and $D_y\sim 1/f$ are consistent with the picture of two different temperatures, one for each direction. These temperatures satisfy a generalized fluctuation relation since the differential mobility in each direction becomes drive independent at large $v$. This kind of behaviour was already observed in a two-dimensional toy models with randomly located parabolic wells~\cite{Kolton2006}. A simple (but not general) argument starts by noting that in the particular bilinearly interpolated potential transverse force in a given square cell derived from Eq.(\ref{eq:bilinear}) is independent of $y$, i.e. $-\partial_y U(x,y)\equiv F_y(x)$. In the next cell the same holds but the force $F_y(x)$ changes, because the random potential at the vertex of the unit squares $V_{n,m}$ are uncorrelated quenched random numbers.
Since $\langle \partial_y U(x,y)\partial_y U(x',y')\rangle=\Delta_{yy}(x-x',y-y')$,
at large $v$ we can write $\langle F_y(x(t))F_y(x(t')) \rangle \approx \Delta_{yy}(x(t)-x(t'),y=0) \approx \Delta_{yy}(v(t-t'),y=0) \sim \delta(t-t')/v \equiv 2T^y_{\tt eff}(v)\delta(t-t')$, where in the first approximate equality we used that the particle moves mostly forward in a distance equal to one correlation length of the disorder along $x$, and in the fourth term that $\Delta_{yy}$ is short-ranged and satisfies $\int_u \Delta_{yy}(u) \neq 0$.
Therefore, we see that the transverse force mimics a thermal-noise at an effective transverse temperature vanishing as $T^y_{\tt eff} \sim 1/f \sim 1/v$.
A more general argument to understand the $1/f$ transverse fluctuations in the RB case at large velocities is given in Ref.\cite{Elias2022}, in the context of a vortex line dynamics.
\subsubsection{RF case}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figconfined1RF.pdf}
\caption{
RF results for the two dimensional model with transverse confinement, Eq.(\ref{eq:2d}).
(a) Quadratic mean displacement vs time with respect to the mean motion in the driving direction, for different driving forces $f$ for a confinement constant $\kappa=1$. The dashed line indicates normal dispersion.
(b) Fitted dispersion constant in the direction of the drive $D_x$ vs $f$ for different confinement constants $\kappa$. The dashed lines indicates the $1/f$ decay and dotted line indicate the $1/f^2$ behavior for comparison.
}
\label{fig:figconfined1RF}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figconfined2RF.pdf}
\caption{
RF results for the two dimensional model with transverse confinement $\kappa>0$, Eq.(\ref{eq:2d}).
(a) Mean squared displacement in the transverse direction as a function of time, for $\kappa=1$.
(b) The variance of transverse fluctuations approximately satisfies $\kappa \langle y^2 \rangle \sim f^{-1}$ at large forces for different confinement constants $\kappa$. Dotted line indicate the $1/f^2$ behavior for comparison.
}
\label{fig:figconfined2RF}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figconfined3RF.pdf}
\caption{
RF results for the two dimensional model \textit{without} confinement ($\kappa=0$), from Eq.(\ref{eq:2dRF}).
Longitudinal (a) and transverse (b) quadratic mean displacement vs time for different driving forces $f$, with respect to the mean motion. Dashed-lines in (a) and (b) highlight normal dispersion.
(c) Longitudinal and transverse dispersion constants $D_x$ and $D_y$ as a function of $f$. The dashed-line shows the $\sim f^{-1}$ decay.
}
\label{fig:figconfined3RF}
\end{figure}
In two dimensions the RF case contemplates various possible cases. The random force field can be compressible or incompressible, rotational or irrotational or a mixture of components with different properties. Here we consider a simple one, which is to take
$\overline{F_x(x,y)F_y(x',y')}=0$
and $\overline{F_x(x,y)F_x(x',y')}=\overline{F_y(x,y)F_y(x',y')}=\Delta(|x-x'|)\Delta(|y-y'|)$, with $\Delta(z)$ a positive short-ranged function.
Note that, unlike the $d=1$ RF case, this force can not be derived from a potential i.e. ${\bf F}({\bf x})\neq -\nabla U({\bf x})$.
In practice, for coordinates
$x \in [n,n+1]$ and $y\in (m,m+1)$
in cell $(n,m)$ we generate two uniformly distributed independent random numbers in the $(-1/2,1/2)$ interval, one for $F_x(x,y)$ and the other for $F_y(x,y)$~\footnote{Different choices, such as $F_x=F_0 \cos(\alpha_{n,m})$, $F_y=F_0 \sin(\alpha_{n,m})$, with $\alpha_{n,m}\in (0,2\pi)$ a random angle, yield equivalent results.}.
The equations of motion then read
\begin{eqnarray}
\dot x &=& f + F_x(x,y),
\nonumber \\
\dot y &=& -\kappa y +F_y(x,y).
\label{eq:2dRF}
\end{eqnarray}
In Fig.\ref{fig:figconfined1RF}(a) we show the quadratic mean displacements in the moving frame vs time. Dispersion in the longitudinal direction is normal and drive-dependent. Comparing with Fig.\ref{fig:figconfined1} for RB we can see a larger dispersion for the RF case. This is reflected in a larger dispersion coefficient $D_x$ in Fig.\ref{fig:figconfined1RF}(b), where we can also observe a $D_x\sim f^{-1}$ decay at large $f$, at variance with the RB case, but consistent with the prediction of the one dimensional toy model, Eq.(\ref{eq:RFresult}). We can also observe that this result is almost independent of the confinement constant $\kappa$ in Eq.(\ref{eq:2dRF}), as was also observed in the RB case.
In Fig.\ref{fig:figconfined2RF}(a) we show the behaviour of the transverse fluctuation vs time, for $\kappa>0$.
As expected, it has a well defined temporal mean value, which is drive dependent. In Fig.\ref{fig:figconfined2RF}(b) we show, for three values of $\kappa$, that $\kappa \langle y^2 \rangle \sim f^{-1}$, as for the RB case.
Interestingly, for the unconfined two dimensional case, $\kappa=0$, the dispersion is {\it isotropic} in the commoving frame, as can be appreciated in Fig.\ref{fig:figconfined3RF}(c), which is obtained by fitting the quadratic mean displacements of Figs. \ref{fig:figconfined3RF}(a) and (b).
\subsubsection{Summary}
In summary in the two-dimensional case (either confined or unconfined)
we find that:
\begin{itemize}
\item $D\sim 1/f^3 \sim 1/v^3$ for RB and
$D\sim 1/f \sim 1/v$ for RF,
at large $v$.
This shows that the results for the toy model, Eqs.(\ref{eq:RFresult}) and (\ref{eq:RBresult}) respectively, are robust. This is somehow expected for $\kappa \gg 1$, as the system becomes quasi one-dimensional. Nevertheless, we have also shown that it even holds for the truly two-dimensional $\kappa=0$ case.
\item In the $\kappa=0$ case there is a normal transverse dispersion with dispersion constant decaying $D_y\sim f^{-1}$, both for RF and RB. Nevertheless dispersion in the moving frame is highly anisotropic in the RB case while it is isotropic in the RF case. This striking difference can be used as a fingerprint of the underlying disorder.
\item Transverse fluctuations in the confined case $\kappa>0$, which decay as $\kappa \langle y^2 \rangle \sim f^{-1}$, are consistent with the effective temperature $T^{x}_{\tt eff}=D_x/2$ obtained from the unconfined $\kappa=0$ case.
\end{itemize}
\subsection{Soft particles: elastic dimer model}
Strictly, skyrmions, vortices, walls or solitons in a one-dimensional disordered track are not rigid objects, and hence quenched disorder affects not only the longitudinal dispersion of their centers of mass but also induce shape fluctuations which in turn affect the way the deformable object couples to the underlying quenched disorder.
To model this effect, and check the robustness of dispersion properties, we will consider a model of a soft localized object composed by two particles coupled with a spring of natural size $l$ in a one dimensional force field $F(x)$,
\begin{eqnarray}
\dot{x_1}&=&f - \kappa (x_1-x_2+l_0) + F(x_1), \nonumber \\ \dot{x_2}&=&f + \kappa (x_1-x_2+l_0) + F(x_2),
\label{eq:coupled}
\end{eqnarray}
where $\kappa>0$ a coupling constant controlling the (longitudinal) shape fluctuations, allowing us to go from the rigid object for $\kappa \gg 1$ to the soft limit for $\kappa \ll 1$.
For the simulations we use the same piece-wise $F(x)$ considered in sections \ref{sec:boxdist} for the RB and RF cases. We will be interested in the dispersion of the center of mass of the two particles.
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figcoupled.pdf}
\caption{
Quadratic mean displacement vs time of two coupled particles in a one dimensional track, Eq.(\ref{eq:coupled}), in the RF (a) and RB (b) disorder cases, for different driving forces. The dashed-line shows that dispersion is normal in both cases. (c) The corresponding dispersion constants, for different spring constants $\kappa$, vanish as $D_{RF}\sim 1/f \sim 1/v$ (dot-dashed line)
and $D_{RB}\sim 1/f^3 \sim 1/v^3$ (dashed- line) at large $v$.
}
\label{fig:figcoupled}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figbond.pdf}
\caption{
Bond fluctuations vs $f$ for two coupled particles, from Eq.(\ref{eq:coupled}), in the RF and RB cases for spring constant $\kappa=1$. As indicated by the dashed and dashed-dotted lines both follow an empirical $\approx 1/f^{7/4}$ decay (of different amplitude), contrasting with the different decays $D_{RF}\sim 1/f$ and $D_{RB} \sim 1/f^3$ (RB) observed in the dispersion constant of individual particles (see Fig.\ref{fig:figcoupled}(c)).
}
\label{fig:figbond}
\end{figure}
In Fig.\ref{fig:figcoupled} (a)-(b) we show the quadratic mean displacement in the RF and RB pinning fields respectively. In both cases dispersion is normal and the dispersion constant decays with increasing $f$. In Fig.\ref{fig:figcoupled} (c) we see that $D_{RF}\sim 1/f$ and $D_{RB}\sim 1/f^3$, as obtained for the toy particle model (Eq.(\ref{eq:RFresult}) and (\ref{eq:RBresult}) respectively), for various values of the spring constant $\kappa$. This shows again the robustness of these predictions.
Interestingly, as shown in Fig.\ref{fig:figbond}, we find that the bond fluctuations, measuring the mean Hook energy of the pair, follows the same $\sim 1/f^{7/4}$ decay at large $f$, though with a different prefactor. This shows on one hand that the effective equipartition law that approximately holds for the transverse fluctuations of transversely confined particles (see section \ref{sec:transverse}) does not work in this case, in the sense that bond fluctuations are clearly not controlled by the same effective temperature $T_{\tt eff}\approx D/\mu$ describing the center of mass dispersion.
This may be attributed to the fact that the effective disorder-induced noise is correlated between particles, because they follow each other and thus visit exactly the same quenched disorder after a mean characteristic time $\sim l_0/v$. Hence, this noise can not mimic thermal-like noise $\xi_n(t)$, which must satisfy $\langle \xi_n(t)\xi_m(t') \rangle \propto \delta_{n,m}\delta(t-t')$ for particles $n$ and $m$ in a coarse-grained level. It is nevertheless quite remarkable that the RB and RF cases seem to follow a unique decay law for bond fluctuations, in sharp contrast with $D_{RF}$ and $D_{RB}$ for the center of mass. It would be interesting to investigate if this system can be still be described by multiple effective temperatures, depending on the typical time-scale of the observable fluctuations.
\subsubsection{Summary}
In summary, we have shown that a simple model of a simple deformable object (an elastic dimer) driven in a disordered one-dimensional track, presents at large velocity a dispersion described by Eqs.(\ref{eq:RFresult}) and (\ref{eq:RBresult}) made for the one overdamped toy model. These predictions thus appear robust under shape fluctuations of soft but yet localized objects.
\subsection{Effect of temperature: driven Brownian particle}
\label{sec:langevin}
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{figtemp.pdf}
\caption{
Effect of temperature on the dispersion in the one dimensional overdamped particle model.
Quadratic mean displacement vs time for different temperatures in the RF (a) and RB (b) random force fields, for a given driving force $f=3$ (dashed-lines indicate normal dispersion).
(c) At large $f$, thermal noise becomes just an additive effect in the dispersion, and $D(f,T)\approx D(f,T=0)+2T$ (as indicated by the asymptotic dotted-line) independently of the RF or RB pinning nature.
}
\label{fig:temperature}
\end{figure}
So far we have discussed deterministic dispersion of different objects in narrow tracks which can be described as particles with one or more degrees of freedom.
When the dispersing objects are small enough thermal fluctuations may become important however and may contribute to the dispersion in the racetrack array. To study its effect on particles we can use the nondimensionalized Langevin equation
\begin{equation}
\dot x = F(x)+ f + \xi(t).
\label{eq:langevin}
\end{equation}
where $F(x)$ is the same random force field as in Eq.(\ref{eq:monomereq}) and $\xi(t)$ is a standard Langevin noise with the properties $\overline{\xi(t)}=0$, and $\overline{\xi(t)\xi(t')}=2T\delta(t-t')$ averaging over the ensemble of possible noise realizations. Here $T$ the nondimensionalized temperature (to get the physical temperature we multiply by the unit of temperature $f_0 d_0/k_B$). For simplicity and to compare with that case, we will consider $F(x)$ to be of the same types discussed in section \ref{sec:boxdist}.
At variance with the other deterministic cases analyzed before, before discussing the dispersion properties of Eq.(\ref{eq:langevin}) it is important to remark here the distinction between the so-called ``annealed'' and ``quenched'' dispersion constants~\cite{BouchaudGeorges1990}. When we describe the normal spreading of a packet of particles in a single track due to finite temperature we are dealing with the ``quenched'' dispersion constant. From Fig.\ref{fig:dispersion} it is clear that we are instead interested in the so-called ``annealed'' dispersion constant $D$. In the absence of disorder, $v=f$ and $D=2T$, since it is standard driven Brownian motion, and then the two dispersion constant coincide. In the presence of disorder however the ``annealed'' is in general larger than the ``quenched'' dispersion constant~\cite{LeDoussal1989}.
In the presence of disorder and thermal fluctuations the problem becomes in principle quite complex. On one hand there is dispersion even below the depinning threshold, as the particles are able, in each track, to jump their energy barriers separating the metastable states by thermal activation. The stochastic dynamics is non trivial and can lead in general to anomalous dispersion, specially at low dimensions and for correlated random forces. We refer the reader to Ref.\cite{BouchaudGeorges1990} for a detailed review on this problem.
Above the putative depinning threshold however, advection dominates and destroy
any long-range time correlations. Also, the residence time in each correlated cell has a finite mean and variance for any $T$. Hence, we expect the Central Limit Theorem to hold and normal dispersion to have a well defined $D$. Our aim here is to describe such regime and show how $D$ behaves vs $f$ and $T$, particularly well above the depinning threshold where the robust power-law scaling emerges as a function of $f$ or $v$.
In Fig.\ref{fig:temperature}(a)-(b) we show
the effect of temperature on the quadratic mean displacement vs time for different temperatures in the RF and RB cases respectively. As can be observed, dispersion is normal and temperature monotonically increases dispersion. To better appreciate its effect
in Fig.\ref{fig:temperature}(c) we show that the effect of temperature is approximately additive at large $f$: $D(f,T)\approx D(f,T=0)+ 2T$, regardless of the type of disorder.
\subsubsection{Summary}
The effect of temperature for the driven overdamped particle in a RF or RB disorder is just to add the trivial thermal dispersion constant $D_0=2T$, describing dispersion in the absence of disorder. This result is expected as the disorder-induced fluctuations become small compared with the (fixed strength) thermal bath-induced fluctuations.
Therefore, the scaling (Eqs.(\ref{eq:RFresult}) and (\ref{eq:RBresult})) remain robust in the sense that now $D(f,T)-D_0 \approx f^{-1} \approx v^{-1}$ for RF and $D(f,T)-D_0 \approx f^{-3} \approx v^{-3}$ for RB.
\subsection{A model for magnetic domain walls in wires}
The dynamics of magnetic domain walls in narrow tracks have attracted a great interest after the proposal of using it to devise a non-volatile racetrack memory~\cite{Parkin2008}.
As a starting point for modelling a magnetic domain wall (DW) driven by either an applied magnetic field and/or a current we will follow Ref.\cite{Thiaville2005} and consider the modified Landau-Lifshitz-Gilbert equation for the local magnetization vector $\vec{m}$,
\begin{eqnarray}
\dot{\vec{m}} &=&
\gamma_0 \vec{H} \times \vec{m} +
\alpha \vec{m} \times \dot{\vec{m}} -
(\vec{\mathcal{U}}.\vec{\nabla})\vec{m}
\nonumber \\
&+& \beta \vec{m} \times
[(\vec{\mathcal{U}}.\vec{\nabla})\vec{m}],
\label{eq:micromagnetics}
\end{eqnarray}
where $\vec{m}$ is a unit vector, $\gamma_0$ the gyromagnetic constant,
$\vec{H}$ the micromagnetic effective field and $\alpha$ the Gilbert damping constant.
The velocity $\vec{\mathcal{U}}$ is a vector directed along the direction of electron motion, with an amplitude $\mathcal{U}=JPg\mu_B/(2eM_s)$, where $M_s$ is the spontaneous magnetization, $J$ is the current density and $P$ its polarization rate. The non-dimensional adiabatic parameter $\beta$ is expected to be comparable to $\alpha$. Thermal fluctuations are neglected.
Solving Eq.(\ref{eq:micromagnetics}) for the dynamics of a driven domain wall is computationally expensive, due to the large number of degrees of freedom. Indeed, even in one dimension, where $\vec{m}\equiv \vec{m}(x,t)$, the system is extended. An usual approximation to leverage this difficulty is the collective coordinate approach which assumes a flat DW profile characterized by only two degrees of freedom: the DW position $q$ along the wire longitudinal $x$ direction, and the internal degree of freedom $\phi$ indicating the orientation of the magnetization at the DW center. For a transverse domain wall Thiaville \textit{et al.} \cite{Thiaville2005}
have obtained, for an homogeneous material,
\begin{eqnarray}
\frac{\alpha}{\Delta}\dot q + \dot \phi &=& \gamma_0 (H_a+ H_p(q)) + \frac{\beta}{\Delta}\mathcal{U} \\
\alpha \dot \phi - \frac{\dot q}{\Delta} &=& -\frac{\gamma_0 H_K}{2} \sin 2\phi - \frac{\mathcal{U}}{\Delta}
\end{eqnarray}
where $\Delta=[A/(K_0 + K \sin^2 \phi)]^{1/2}$, with $K_0$ and $K$ the effective longitudinal and transverse anisotropy respectively, and $H_K=2K/\mu_0 M_s$. For simplicity we will make a rigid wall approximation or assume that $K \ll K_0$, so to approximate $\Delta$ with a constant. We will also assume that quenched disorder (arising from intrinsic magnetic or non-magnetic defects in the material, thickness modulations or from the wire borders roughness) give place to an extra random field coupled only to $q$, $H_p(q)$, as in Lecomte \textit{et al.} \cite{Lecomte2009}.
By measuring $H_a$ and $H_p(q)$ in units of $H_K/2$, $u$ in units of $v_W=\Delta \gamma_0 H_K/2$, time in units of $(1+\alpha^2)\Delta/v_W$, and $q$ in units of $\Delta$ we get (overriding notation for nondimensionalized $q$, $\phi$ and $t$), the nondimensional dynamical system
\begin{eqnarray}
\dot q &=& \alpha [h + F(q)] + \sin(2\phi)+ u (1+\alpha \beta)
\label{eq:microm1}
\\
\dot \phi &=& h+ F(q) + (\beta-\alpha) u -\alpha \sin(2\phi),
\label{eq:microm2}
\end{eqnarray}
controlled by the dimensionless parameters $\alpha$ and $\beta$, and drivings $h=2H_a/H_K$ and $u=\mathcal{U}/v_w$, with $F(q)=2H_p(q)/H_K$ the dimensionless pinning field.
It is worth noting that if we freeze the internal degree of freedom $\phi$, of the DW the model reduces to the overdamped mechanical model, essentially identical to Eq.(\ref{eq:monomereq}).
We also note that in the absence of disorder ($F(q)=0$) at long times $\dot \phi=0$ if $h+(\beta-\alpha)u < \alpha$. This threshold corresponds to the model approximation of the Walker breakdown, separating the so-called stationary regime, where $\alpha \dot q = h + \beta u$ and $\dot \phi=0$, from the precessional regime where $\langle \dot \phi \rangle \neq 0$. With pinning, there is a fixed point $\dot q = \dot \phi =0$ or pinned state for small enough yet finite drives.
To solve Eqs.(\ref{eq:microm1}) and (\ref{eq:microm2}) we will consider again the piece-wise random-force field $F(q)$ used in section \ref{sec:boxdist}, both in the RF and RB cases. We will consider both the case of pure magnetic field driving ($u=0$) and the pure current driving $(h=0)$, and the cases $\langle \dot \phi \rangle=0$ and $\langle \dot \phi \rangle \neq 0$,
and will focus on their large velocity regimes, $v\equiv \langle \dot q \rangle \sim u(1+\alpha \beta)$
and
$v\equiv \langle \dot q \rangle \sim \alpha h$, respectively.
\subsubsection{RF case}
In Fig.\ref{fig:DWRF}(a)-(b) we show the results for the RF case for the current driven ($h=0$, $u>0$) and field driven ($h>0$, $u=0$) cases respectively, for drivings well above the corresponding depinning thresholds.
In Fig.\ref{fig:DWRF}(a) we show the rescaled dispersion constant $D$ for the current driven case for $u$ values such that $v\equiv \langle \dot q \rangle \approx u(1+\alpha\beta)$. We empirically find that $D=|\alpha-\beta|\alpha^3 G_u(u|\alpha-\beta|\alpha)$, with $G_u(x)$ a scaling function, describes well the data for different pairs of values $(\alpha,\beta)$ which are sound for real materials. At large $u|\alpha-\beta|\alpha \gg 1$ we find $D\sim u^{-1}$ in agreement with the toy model with a RF disorder, Eq.(\ref{eq:RFresult}). Interestingly however, for smaller $u|\alpha-\beta|\alpha \ll 1$ a crossover towards a $D\sim u^{-3}$ is observed, coinciding with the decay predicted from the toy model for a RB disorder.
In Fig.\ref{fig:DWRF}(b), for the field driven case, we show that $D \approx h^{-1}$ at large $h$, with a power-law decay in $h$ in agreement with the RF case of the toy model, Eq.\ref{eq:RFresult}. Interestingly, no crossover is observed in this case.
A qualitative explanation of the results of Fig.\ref{fig:DWRF} can be given.
The crossover in the current driven case (Fig.\ref{fig:DWRF}(a)) occurs at $u^*\approx |\alpha-\beta|^{-1}\alpha^{-1}$. For $u>u^*$, we have $D\approx \alpha^2/u$, which is consistent with an RF disorder of amplitude $\alpha$ in the one degree of freedom toy model, Eq.(\ref{eq:RFresult}). Therefore, the $\sin(2\phi)$ term in Eq.(\ref{eq:microm1}) appears to have a negligible influence in the dispersion compared with the one expected from the term $\alpha F(q)$. To explain this we note, on one hand, that the larger $\alpha$ (the smaller $u^*$), the more important $\alpha F(q)$ is compared with the average effect of $\sin(2\phi)$ on the dispersion.
One the other hand, when $u<u^*$ dispersion is dominated by the $\sin(2\phi)$. Subtly, its effect on the dispersion of $q$ is similar to the one of a RB disorder, yielding $D\sim 1/u^3$. Nevertheless, this case can not be described by a simple one degree of freedom toy model such as Eq.(\ref{eq:monomereq}) for an effective RB force field $F(q)$ (see appendix \ref{sec:reducido}).
The above arguments also explain qualitatively the field driven case of Fig.\ref{fig:DWRF}(b), since in this case the characteristic frequency of fluctuations of the $\sin(2\phi)$ term is always of order $h$ and can not be reduced by changing $\alpha$ and/or $\beta$. Since $v\sim h$, the average effect of $\sin(2\phi)$ can be neglected and then $D\sim \alpha^2 h^{-1}$ as expected for the toy model in a RF disorder of amplitude $\alpha$, Eq.(\ref{eq:RFresult}).
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{domain_wall_RF.png}
\caption{Scaled dispersion constant of a domain wall with RF disorder, from the numerical solution of Eq.(\ref{eq:microm2}). In (a) we show the current driven case ($h = 0,\; u > 0$), for $\alpha=\{0.02, 0.2, 0.25, 0.35 0.5, 0.75, 1.0\}$ and $\beta = \{0.0,0.02, 0.04, 0.08, 0.1, 0.2\}$ (all $\{\alpha, \beta\}$ pairs). The data is in agreement with a scaling of the form $D=|\alpha-\beta|\alpha^3 G_u(u|\alpha-\beta|\alpha)$ where $G_u(x) \sim x^{-1}$ for $x \gg 1$ and $G_u(x) \sim x^{-3}$ for $x \ll 1$. In (b) we show the pure field driven case ($u = 0,\; h > 0$) for $\alpha=\{0.01, 0.02, 0.2, 0.35 0.5, 1.0\}$. No crossover is observed and the data follows approximately the relation $D/\alpha \propto h^{-1}$.
\label{fig:DWRF}}
\end{figure}
\subsubsection{RB case}
In Fig.\ref{fig:DWRB}(a)-(b) we show the results for the RB case for the current driven ($h=0$, $u>0$) and field driven ($h>0$, $u=0$) cases respectively, from the numrical solution of Eq.(\ref{eq:microm2}). We observe $D\sim u^{-3}$ and $D\sim h^{-3}$, in agreement with the toy model prediction Eq.(\ref{eq:RBresult}) for RB disorder.
Interestingly, no crossover is observed in the current-driven case of Fig.\ref{fig:DWRF}(a). Using the arguments above, this may be explained by the fact that in this case both, $\alpha F(q)$ and (in a subtly effective way) $\sin(2\phi)$, are of a RB type, and thus $D\sim u^{-3}$ is in agreement the toy model, Eq.(\ref{eq:RBresult}) for RB disorder.
Similarly, the field driven case can be explained by the fact that $\sin(2\phi)$ in Eq.(\ref{eq:microm1}) oscillates at a frequency $\sim 2h$ larger than $v$, which is the characteristic frequency induced by $\alpha F(q)$. Therefore, its average effect on dispersion vanishes, and the system effectively behaves as the toy model in the RB case, (Eq.\ref{eq:RBresult}).
Power-law prefactors in Figs.\ref{fig:DWRB}(a)-(b) are different and less simple to predict in the RB case as compared to the RF case.
The scaling collapse in Figs.\ref{fig:DWRB}(a)-(b)
is purely empiric.
\begin{figure}[htp]
\centering
\includegraphics[width=\columnwidth]{domain_wall_RB.png}
\caption{Scaled dispersion constant of a domain wall with RB disorder for $\alpha=\{0.3, 0.4, 0.5 0.6, 0.8\}$. In (a) in shown the pure current driven case ($h = 0,\; u > 0$) for $\beta = \{0.0 , 0.02, 0.04\}$ showing the expected RB behavior as $D \cdot \alpha \propto 1/u^3$. In (b) in shown the pure field driven case ($u = 0,\; h > 0$) showing again the expected RB behavior but $D \cdot (1-\alpha)^{-1} \propto 1/h^3$. In both cases, no crossover is observed, at variance with the current driven RF case (see Fig.\ref{fig:DWRF}(a)).}
\label{fig:DWRB}
\end{figure}
\subsubsection{Summary}
In summary, we have shown that a zero-dimensional domain-wall derived from micromagnetic considerations, presents the following features:
\begin{itemize}
\item At large drives, either from an applied current or magnetic field, $D$ decays as expected from the toy model predictions Eqs.(\ref{eq:RFresult})-(\ref{eq:RBresult}) for the RF and RB cases respectively.
\item In all cases $D$ scales differently with respect to the Gilbert damping parameter $\alpha$ and the adiabatic parameter $\beta$.
\item A crossover between different power-law decays is observed in $D$ only in the RF case, for the range of $\alpha$ and $\beta$ analyzed.
\end{itemize}
\section{Discussion}
\label{sec:discussion}
Our results show that the way the dispersion constant of particles $D$ vanishes at large velocities is independent of many model details. It is nevertheless particularly sensitive to the nature of disorder. One may consider other types of random-force fields, such as long-ranged or periodically correlated ones, and the way $D$ vanishes may change. The RB and RF cases are nevertheless rather important physical cases to start with. For zero-dimensional magnetic domain walls for instance (i.e. when can neglect its transverse fluctuations and consider it as a flat interface), the RB type is generated by non magnetic-impurities, rough borders, modulated thickness or by a space-dependent magnetic anisotropy, while the RF is realized in the presence of magnetic-impurities inducing a random magnetic field. The situation is similar for zero-dimensional ferroelectric domain walls, where the two types of quenched disorders can be in principle realized. On the other hand, magnetic skyrmions, colloids or vortices typically couple to a random bounded potential and hence realize a RB disorder.
Extended systems may be also considered, such as one- or two-dimensional domain walls anchored by a narrow channel of a finite width. Skyrmions chains in one dimensional tracks also fall in the extended system category. In these systems with many elastically coupled degrees of freedom the random-force acting on the center of mass may be renormalized. A prominent example are elastic walls in a short-ranged correlated random-force field, for which a RB microscopic disorder is renormalized to a cuspy RF-type near depinning. The force-force correlator then gets rounded at larger velocities as the dynamical correlation length reduces~\cite{chauve2000}. In these cases we expect our results for the dispersion to hold for the center of mass respecting the shape of the renormalized disorder instead of the microscopic one. It is plausible that, in such case, a scenario similar to the one described in section \ref{sec:rb+rf} takes place at large velocities. Indeed, for a driven elastic vortex line in three dimensions, with RB disorder, an effective temperature vanishing as $\sim v^{-3}$ was found for a wide range of velocity in spite that the depinning transition is better described by a renormalized RF-type random force field~\cite{Elias2022}.
Bringing back the dimensions, our results show that at large velocities, we can roughly expect
\begin{eqnarray}
v&\approx& (v_0/f_0)f, \\
D&\approx& v_0 d_0 (f_0/f)^n,
\end{eqnarray}
where $f_0$ is the pinning force amplitude, $d_0$ the disorder correlation length $v_0=\eta f_0$ is the characteristic velocity, $\eta$ the friction constant, and $n=1$ or $n=3$ for RF and RB disorder respectively. To compare with experiments we may use that the experimentally accessible depinning force is $f_c \sim f_0$ (neglecting interaction effects between particles) and hence $v_c \equiv \eta f_c \sim v_0$. The disorder correlation length $d_0$ is usually unknown or difficult to estimate directly, but one can use that $d_0=\max(R,r_0)$, where $R$ is the size of the particle and $r_0$ the correlation length-scale of the inhomogeneity. We can thus write a more practical and simple prediction
\begin{eqnarray}
D\approx v d_0 (f_c/f)^{n+1} \approx v d_0 (v_c/v)^{n+1}.
\label{eq:Dwithunits}
\end{eqnarray}
From an applied viewpoint, to realize a memory bus analog, it would be desirable to have particles in different tracks concurrently and precisely controlled. If we need for instance to transport coherently and at the largest possible velocity a few bytes from one check point to the next by moving particles in different parallel tracks, the less possible dispersion would be ideal. To make a concrete estimation let us define the maximum time $t_{\rm max}$ and distance $x_{\rm max}$ we can move the particles at given velocity $v$ with a given tolerance $\delta$ for the dispersed packet width, i.e $D t_{\rm max}= \delta^2$. It makes sense to ask for a maximum tolerated error to be of the order of one particle (one bit) size,
i.e. $\delta \sim R$.
We then get
\begin{eqnarray}
x_{\rm max} \equiv
v t_{\rm max} \approx
\frac{R^2}{d_0}\left(\frac{f}{f_c}\right)^{n+1} \approx
\frac{R^2}{d_0}\left(\frac{v}{v_c}\right)^{n+1}.
\label{eq:predictions}
\end{eqnarray}
For skyrmions for instance, if we assume point disorder, we have $d_0 \approx R$. Then, using a skyrmion size $R\approx 10 \text{ nm}$ moving at $v\approx 0.34 \text{ mm/s}$ for $f/f_c \equiv j/j_c \approx 5$~\cite{Fert2013} we obtain $x_{\rm max} \approx 6 \text{ }\mu\text{m}$ and $t_{\rm max}\approx 0.025\text{ s}$ for a coherent parallel transport. If the disorder were otherwise of a pure RF type ($n=1$) then $x_{\rm max}$ and $t_{\rm max}$ would be more than one order of magnitude smaller. On the other hand, if the microscopic correlation length of the inhomogeneity is $r_0>R$ the previous estimates for $x_{\rm max}$ and $t_{\rm max}$ enhance by a factor $R/r_0$.
Rough estimates similar to the above can be made for domain walls, superconducting vortices or colloids, provided we know a few parameters such as the size of the particles
and the velocity-force characteristics under the assumption of a short-ranged random force correlator. Besides specific applications however, Eq.(\ref{eq:Dwithunits}) is a rather robust prediction which might be useful to characterize the disorder in the host media, regardless of its microscopic origin, by measuring the dispersion $D$ of independent racing particles as a function of the driving force.
\begin{acknowledgments}
We thank Kay Wiese for illuminating discussions. We also acknowledge support from grant PICT-2019-01991.
\end{acknowledgments}
|
\chapter*{``Concerning a new kind of L-Series'': Introduction}
By E. Artin in Hamburg.
\medskip
{\color{blue}
In what follows the black paragraphs give my very free translation of Artin's original paper. The blue paragraph gives my notes and comments.
A similar convention will hold for footnotes.
We start with the with the introduction, which consists of a short paragraph:}
\medskip
For investigating non-Abelian algebraic number fields one needs a new kind of $L$-series
that generalizes the usual $L$-series for Abelian algebraic number fields.
These analytic functions are formed with Frobenius style group characters.
This article is dedicated to the investigation of such functions.
\chapter{Frobenius Style Group Characters: Review}
For the convenience of the reader, I will begin by briefly giving the formulas and notation that we will need from
the theory of group characters.\footnote{See J. Schur 1905, \emph{Neue Begr\"undung der Theorie der Gruppencharaktere} (New foundation for the theory of group characters),
Sitzungsberichte (conference reports), Berlin, and Speiser~\cite{SpeiserGroups} Chapters~10-12.}
Let $G$ be a finite group of order~$n$. Decompose $G$ into $x$ conjugacy classes~$C_1, \ldots, C_x$, and let
$h_i$ be the number of elements of $C_i$.
Let $\Gamma$ be a representation of the group $G$ as nonsingular matrices. Given $\Gamma$ we get a \emph{character}~$\chi$
which is a function $G\to {\mathbb{C}}$ that assigns to $\sigma \in G$ the trace of the associated matrix.
There are~$x$ irreducible representations $\Gamma_1, \ldots, \Gamma_x$, and let $\chi^1, \ldots, \chi^x$ be their associated characters.
These characters are called \emph{simple characters}. Every character $\chi$ is in fact the linear combination of simple characters:
\begin{equation}\label{E1}
\chi(\sigma) = \sum_{i=1}^x r_i \chi^i(\sigma)
\end{equation}
where $r_i$ are nonnegative integers associated with the decomposition of $\Gamma$ into irreducible representations.
The simple characters satisfy the following formulas
\begin{equation}\label{E2}
\sum_{\sigma} \chi^i (\sigma) \chi^k(\sigma^{-1}) = n \delta_{i k}
\end{equation}
and
\begin{equation} \label{E3}
\sum_{i=1}^x \chi^i(\sigma) \chi^i (\tau^{-1})
=
\begin{cases} 0 \text{ if $\sigma$ and $\tau$ are in different classes}, \\ \frac{n}{h_r} \text{ if $s$ and $\tau$ are both in the class $C_r$}. \end{cases}
\end{equation}
Furthermore, suppose $H$ is a subgroup of $G$ and that
\begin{equation}\label{E4}
G = \sum_{i=1}^s H S_i
\end{equation}
is the decomposition into cosets (here $S_i \in G$).
Let $\Delta$ be a representation of the subgroup $H$ of degree $\delta$, and let $A_\sigma$ be the matrix associated to $\sigma \in H$.
If $\sigma \in G$ is not in $H$ we take $A_\sigma$ to be the zero matrix. We build the matrix $B_\sigma$ out of blocks in the following way:
\begin{equation}\label{E5}
B_\sigma = \left(
A_{S_i \sigma S_k^{-1} }\right).
\end{equation}
As stated this is an $s$ by $s$ square matrix with entries equal to $\delta$ by $\delta$ square matrices, where $s$ is the index of $H$ in~$G$.
But we regard $B_\sigma$ as defining a square~$s\delta$ by $s \delta$ matrix, and it turns out that this gives a representation of $G$
called the representative of $G$ induced by the representation of $H$.\footnote{See Speiser~\cite{SpeiserGroups} \S 52
from which formula (44) can easily be easily derived.
See also an 1898 report by Frobenius called \emph{\"Uber Relationen zwischen den
Charakteren einer Gruppe und denen ihrer Untergruppen} (Concerning the connection between the characters of a group
and those of its subgroups).}
{\color{blue}
\begin{remark}
In the above $A_{S_i \sigma S_k^{-1}}$ designates the $(i, k)$ block (using the $i$th row partition and $k$th column partition). There are $s^2$ such blocks total, and each block is a~$\delta$ by $\delta$ matrix.
So the induced representation is given concretely in terms of~$s \delta$ by $s \delta$ matrices associated to each $\sigma \in G$.
Artin is viewing the group acting on the right of vectors. If we act on the left, which is common today,
we end up with the $(i, k)$ block looking like $A_{S_i^{-1} \sigma S_k}$.
\end{remark}
}
If $\psi$ is the character of
the representation $\Delta$ then the character $\chi_\psi$ associated to the representation~(\ref{E5}) is called the \emph{character of $G$ induced by the
character $\psi$
of~$H$}.
Let $\psi_1, \ldots, \psi_\lambda$ be the simple characters of the subgroup~$H$.
Then we can express the restriction of $\chi^i$ to $H$ as a nonnegative integral linear combination of $\psi_1, \ldots, \psi_\lambda$
with nonnegative integer coefficients $r_{1 i}, \ldots, r_{\lambda i}$:
\begin{equation} \label{E6}
\chi^i(\tau) = \sum_{\nu = 1}^\lambda r_{ \nu i} \psi_\nu (\tau)\qquad (i=1, \dots, x)
\end{equation}
for all $\tau \in H$.
Similarly, we can express the induced character $\chi_{\psi_i}$
as a nonnegative integral linear combination of the simple characters $\chi_1, \ldots, \chi_x$ of $G$,
and in fact the nonnegative coefficients are just the coefficients that arise in (\ref{E6}):
\begin{equation} \label{E7}
\chi_{\psi_i} (\tau) = \sum_{\nu =1}^x r_{i \nu} \chi^{\nu} (\tau) \qquad (i=1, \dots, \lambda).
\end{equation}
for all $\tau \in G$.
{\color {blue}
\begin{remark}
The above is an expression of the Frobenius reciprocity law.
The version of Serre~\cite{Serre1977} Section 7.1 can be written
$$
\left< \psi, \mathrm{Res}\, \chi \right>_H
=
\left<\mathrm{Ind}\,\psi, \chi \right>_G.
$$
The above statement can be derived from this.
\end{remark}
}
\chapter{Construction of the $L$-Series}
From now on let $k$ be an algebraic number field, let $K$ be a Galois extension of $k$, and let~$G$ be the Galois group of $K / k$.
Let $\mf p$ be a prime ideal in the ring of integers of $k$ not dividing the relative discriminant of $K/k$. Let $\mf P$ be a prime
ideal of $\mc O_K$ dividing $\mf p \mc O_K$.
We chose an element $\sigma \in G$ such that for all algebraic integers $A$ in $K$ we have
\begin{equation}\label{E8}
\sigma A \equiv A^{N \mf p} \pmod {\frak P}
\end{equation}
where $N\frak p$ is the norm of $\frak p$ in $k$.
For the existence of such a $\sigma$ see Weber's Algebra~\cite{Weber2v1899}, \S 178 (volume 2).
This congruence determines $\sigma$ uniquely given a choice of $\frak P$, since if $\sigma_1$ satisfies the same congruence then, for all algebraic integers $A$ in $K$,
$$
\sigma^{-1} \sigma_1 A \equiv A \pmod {\frak P},
$$
and so $\sigma^{-1} \sigma_1$ belongs to the inertia group (Tr\"agheitsgruppe) of $\frak P$. By our assumption ($\frak p$ not dividing the relative discriminant)
the inertia group is trivial.
Next suppose one chooses $\frak P'$ instead of $\frak P$
as a designated prime divisor of $\mf p \mc O_K$. Since~$G$ acts transitively on primes above $\frak p$, we have
$
\tau \frak P = \frak P'
$
for some $\tau \in G$. It is easy to check that one gets~$\tau \sigma \tau^{-1}$ as the corresponding element of $G$ (where $\sigma$ is the
corresponding element for $\frak P$).
So we have a way to associate to $\frak p$ a well-defined conjugacy class $C$ of $G$.
It is well-known that each element of $C$ generates the decomposition group
for some~$\frak P$ above $\frak p$ but this property does not in general completely determine the class $C$ (in fact certain powers
of this class $C$ with have this property).\footnote{\label{FrobFoot}This assignment of conjugacy classes to prime ideals was
already carried out by Frobenius. See the 1896 Berlin report called \emph{\"Uber Beziehungen zwischen Primidealen eines algebraischen K\"orpers und den Substitutionen
seiner Gruppe} (concerning the relationships between prime ideals of an algebraic field and the elements of its Galois group).}
We will say that the prime ideal $\frak p$ \emph{belongs to the class $C$} and we will write this class as $C_{\frak p}$.
{\color {blue}
\begin{remark}
We call each element of $C_{\frak p}$ a \emph{Frobenius element}, and the class as a whole the \emph{Frobenius class},
in honor of Frobenius who, as Artin points out in the footnote, developed this idea earlier.
Artin does not really use these terms in the German original of this paper, but I will use them in the translation below for the convenience of the modern reader.
\end{remark}
}
From now on let $\Gamma$ be a linear representation of $G$. For $\frak p$ as above let $A_{\frak p}$ be a matrix associated to an element of $C_{\frak p}$ via $\Gamma$.
Since the elements of $C_{\frak p}$ are conjugate, the characteristic polynomial
$$
\left| E - t A_{\frak p} \right|
$$ of $A_{\frak p}$ does not depend on the choice of $A_{\frak p}$. Here $E$ is the identity matrix and, as usual, the absolute values indicates determinant.
Note that $A_p$ will change by a conjugate if~$\Gamma$ is replaced by an equivalent representation, so the characteristic polynomial only depends
on the representation $\Gamma$ up to equivalence.
We define the associated $L$-series by the formula
\begin{equation}\label{E9}
L(s, \chi; k) = \prod_{\mf p} \frac{1}{| E - \left(N \mf p\right)^{-s} A_{\mf p} |}
\end{equation}
where $s$ is a complex variable and $\chi$ denotes the character associated with the representation $\Gamma$.
Here, the product varies only for the set of prime ideals $\mf p$ of $\mc O_k$ that do no divide the relative discriminant of $K/k$.
{\color {blue}
\begin{remark}
In a later paper, Artin gives an explicit formula for terms associated to primes that do divide the relative discriminant of $K/k$.
Note that the above $L$ series is expressed using $\chi$ instead of $\Gamma$, since $\chi$ determines $\Gamma$ up to equivalence and
so determines the expression on the right-hand side of (\ref{E9}).
\end{remark}
}
The function $L(s, \chi; k)$ converges absolutely and uniformly on any closed and bounded region in the half plane $\frak R (s) > 1$. To see this observe
that every root of the characteristic polynomial $\left| E - t A_{\frak p} \right|$ is a root of unity. Thus $L(s, \chi; k)$ is a product of terms of the form
$$
\frac{1}{1 - \left(N \frak p\right)^{-s} \varepsilon }
$$
where $\varepsilon$ is a root of unity.
{\color {blue}
\begin{remark}
Since $A_\mf p$ has finite order it is diagonalizable with eigenvalues all equal to roots of unity.
So its characteristic polynomial factors as described by Artin.
Some of the convergence issues can be handled with the following well-known criterion: if an infinite series $\sum |a_i|$ converges then the corresponding infinite product~$\prod (1+a_i)$ converges, and the terms $1+a_i$ can be reordered freely with convergence to the same result. Furthermore, if each term $1+ a_i$ is nonzero then
the limit is nonzero. (See, for example, \cite{Stein2}, Chapter 5, Proposition 3.1 for some justification.)
On the other hand,
it might be convenient to wait on convergence issues until we have the formula for the logarithm given by Artin below.
\end{remark}
}
One can now expand (\ref{E9}) in a Dirichlet series and express the coefficients in terms of the character $\chi$.
The resulting formulas are not very clear (``Die Formeln werden aber wenig übersichtlich''). On the other hand, we arrive at a simple formula for the
logarithm of~(\ref{E9}).
First we associate a conjugacy class $C_{\mf p^\nu}$ to any power $\mf p^\nu$ of a prime ideal $\mf p$. We simply take the class consisting of $A_{\frak p}^\nu$
where $A_{\frak p} \in C_{\frak p}$.
It is easy to see that this forms a conjugacy class of~$G$. We write
\begin{equation}\label{E10}
\chi \left( \frak p^\nu \right) = \chi (\sigma)
\end{equation}
where $\sigma$ is any member of $C_{\frak p^\nu}$.
Now let $\varepsilon_1, \varepsilon_2, \ldots, \varepsilon_f$ be the roots of the equation $\left| Et - A_{\frak p} \right| = 0$. Then
\begin{equation}\label{E11}
\chi \left( \frak p^\nu \right) = \varepsilon_1^\nu + \varepsilon_2^\nu + \ldots + \varepsilon_f^\nu.
\end{equation}
So we get for $|t| < 1$
$$
-\log|E - t A_{\frak p}| = - \sum_{i=1}^f \log (1 - t \varepsilon_i) = \sum_{i=1}^f \sum_{\nu=1}^\infty \frac{\varepsilon_i^\nu}{\nu} t^\nu
=
\sum_{\nu=1}^\infty \frac{\chi(\frak p^\nu)}{\nu} t^\nu
$$
{\color {blue}
\begin{remark}
Here we understand $\log$ as a multivalued functions, or equivalently we regard some of our equations as being valid modulo $(2\pi i ) {\mathbb{Z}}$.
So for example, the first equation above can be regarded as a congruence modulo~$(2\pi i ){\mathbb{Z}}$.
\end{remark}
}
This leads to the desired formula:
\begin{equation}\label{E12}
+ \log L(s, \chi; k) = \sum_{\frak p^\nu} \frac{\chi(\frak p^\nu)}{\nu \left(N \frak p^\nu \right)^{s}},
\end{equation}
where the sum varies over all powers of prime ideals of $k$ not dividing the relative discriminant of $K/k$.
{\color {blue}
\begin{remark}
Associated convergence issues can be justified by the observation that
$$
\sum_{\mf p^\nu}
\left|
\frac{\chi(\frak p^\nu)}
{\nu \left( N \mf p^\nu \right)^s }
\right|
\; < \;
\sum_{M=1}^\infty
m \frac{f}
{ M ^{\sigma_0}}
= mf \zeta(\sigma_0) < \infty
$$
assuming that $\Re(s) \ge s_0 > 1$. Here $\zeta(s)$ is the classical Zeta function,
the left sum is taken over all ideals of the form $\mf p^\nu$, in any order, where $\mf p$ is a prime ideal of~$\mc O_k$ relatively prime to the relative
discriminant of $K/k$, and $\nu$ is a positive integer. Also
$f$ is the degree of the representation $\Gamma$ and~$m$ is the degree~$[k\colon {\mathbb{Q}}]$, so that at most $m$ ideals of the form $\mf p^\nu$ can share the same norm.
In particular we have the absolute convergence of
$$
\sum_{\mf p^\nu}
\frac{\chi(\frak p^\nu)}
{\nu \left( N \mf p^\nu \right)^s }
$$
which justifies the manipulations above. We also get uniform convergence on the set $\Re(s) \ge s_0$ for each $s_0 > 1$, and so the sum gives
a homomorphic function on the set defined by~$\Re(s) > 1$.
By exponentiation we get the desired convergence properties for our Euler product expansion of $L$ as well,
including the invariance under reordering of terms with a product that defines a holomophic function in $s$
with no zeros on the set defined by $\Re(s) > 1$.
\end{remark}
}
Either from (\ref{E9})
or even better from (\ref{E12}) one sees that
\begin{equation}\label{E13}
L(s, \chi + \chi') = L(s, \chi) L(s, \chi')
\end{equation}
for any two characters $\chi$ and $\chi'$.
If $\chi$ is a simple character then we will call the associated $L$-series a \emph{primitive $L$-series}.
If $\chi$ is a general character expressed in terms of simple characters, as in~(\ref{E1}) then (\ref{E13}) gives us
\begin{equation}\label{E14}
L(s, \chi) = \prod_{i=1}^x \left( L(s, \chi^i) \right)^{r_i}.
\end{equation}
A brief remark about the dependence on the field $K$: suppose $\Omega$ is an extension of $K$ that is Galois over $k$.
Then $G = \mathrm{Gal}(K/k)$ is isomorphic to
the quotient group~$\mathrm{Gal}(\Omega/k) / \mathrm{Gal}(\Omega/K)$. If $\sigma \in \mathrm{Gal}(\Omega/k)$ is
such that (\ref{E8}) is valid for all algebraic integers in $\Omega$ then it will of course be valid for all algebraic integers in~$K$.
Furthermore, (\ref{E8}) will be valid for algebraic integers $A$ in $K$ if we replace $\sigma$ with any element of the coset~$\sigma \, \mathrm{Gal}(\Omega/K)$.
Next observe that every character of~$\mathrm{Gal}(\Omega/k) / \mathrm{Gal}(\Omega/K)$
is a character of $\mathrm{Gal}(\Omega/k)$, and every simple character of~$\mathrm{Gal}(\Omega/k) / \mathrm{Gal}(\Omega/K)$
is a simple character of $\mathrm{Gal}(\Omega/k)$. In particular every $L$-series using $K$ as the extension will essentially be an $L$-series using $\Omega$
as the extension, and if the $L$ series is primitive using $K$ then it will be primitive using $\Omega$.
However, the relative discriminant of~$\Omega/k$ may exclude a finite number of prime factors in the $L$-series that occur using the relative discriminant of~$K/k$.
But we will consider $L$-series that differ from each other by only a finite number of factors as being essentially the same.
By the way, we will be able to normalize the $L$-series later to be truly invariant of $K$.
\color{blue}
\begin{remark}
The above uses a fundamental compatibility principle for of the Frobenius element associated with two extensions $\Omega/k$ and $K/k$ of a common base field $k$.
This principle is needed in several places in this paper, so I will go ahead and codify it as a lemma.
I will switch the roles of $K$ and $\Omega$ here since in what follows $\Omega$ is often used to denote an intermediate field.
\end{remark}
\begin{lemma} \label{triple_lemma}
Suppose $K/k$ is a Galois extension of number fields with Galois group~$G$ and let $\Omega$ be an intermediate field such that $\Omega/k$ is also Galois.
Let $\mf p$ be a prime ideal of $\mc O_k$ not dividing the relative discriminant of $K/k$, let $\mf q$ be a prime ideal of~$\mc O_\Omega$ above $\mf p$, and
let $\mf P$ be a prime ideal of $\mc O_K$ above $\mf q$. In other words we have a triple extension $K/\Omega/k$ with corresponding prime ideals $\mf P, \mf q, \mf p$.
Then if $\sigma \in G$ is the Frobenius element associated to $\mf P$, then the restriction~$\sigma'$ of $\sigma$ to $\Omega$ is the Frobenius element
of $\mf q$ in the Galois group of~$\Omega/k$. When we identify the Galois group of $\Omega/k$ with $G/H$
where $H$ is the Galois group of $K/\Omega$, then this Frobenius element $\sigma'$
is the coset $\sigma H \in G/H$.
\end{lemma}
\begin{proof}
By (\ref{E8}) we have that $\sigma A - A^{N \mf p} \in \mf P$ for all $A \in \mc O_K$.
So
$$
\sigma' A - A^{N \mf p} \in \mf P \cap \mc O_\Omega = \mf q
$$
for all $A \in \mc O_\Omega$. Hence
\begin{equation*}
\sigma' A \equiv A^{N \mf p} \pmod {\frak q}
\end{equation*}
for all $A \in \mc O_\Omega$, and so $\sigma'$ is the desired Frobenius element.
By basic Galois theory,~$\sigma'$ corresponds to the coset $\sigma H \in G/H$.
\end{proof}
\color{black}
\chapter{The Theorem on Induced Representations}
Let $H$ be a subgroup of $G$, let $\Omega$ be the subfield of $K$ fixed by $H$, so~$H$ is the Galois group of $K/\Omega$.
The first main theorem (Satz 1) concerns the following situation:
\begin{itemize}
\item $\Delta$ is a representation of $H$.
\item $\Gamma_\Delta$ is the induced representation of $G$.
\item $\psi$ is the character of $\Delta$, and $\chi_\psi$ is the character of $\Gamma_\Delta$.
\item Exclude as factors of $L(s, \psi; \Omega)$ any prime dividing the relative discriminant of $K/k$ (considered as an ideal of the
ring of integers of $\Omega$).
\end{itemize}
\begin{satz}
In the situation discussed above
\begin{equation}\label{E15}
L(s, \psi; \Omega) = L(s, \chi_\psi ; k).
\end{equation}
\end{satz}
\begin{proof}
We set up the following notation:
\begin{itemize}
\item
Let $\mf p$ be a prime ideal of $\cO_k$ not dividing the relative discriminant of $K/k$.
\item
Let $\mf q_1, \mf q_2, \ldots, \mf q_r$ be the prime ideals of $\cO_\Omega$ dividing $\fp \cO_\Omega$:
$$
\fp \cO_\Omega = \mathfrak q_1 \mathfrak q_2 \cdots \mathfrak q_r.
$$
\item
Let~$l_i$ be the relative degree of $\mathfrak q_i$ over $\fp$. In other words, $(N \fp)^{l_i}$ is the size~$N \mathfrak q_i$ of
the residue field~$\cO_\Omega/ \mathfrak q_i$.
\item
For each $\mf q_i$ choose a prime ideal $\fP_i$ of $\cO_K$ dividing $\mf q_i \cO_K$.
\item
For each such $\fP_i$ let $\tau_i \in G$ be chosen so that $\fP_i = \tau_i \fP_1$. (Recall that the Galois group $G$ acts transitively on
the primes of $\cO_K$ dividing $\mf p \cO_K$).
\item
Let $\sigma \in G$ be the Frobenius element associated with $\fP_1$ over $k$. In other words,
$$
\sigma A \equiv A^{N\frak p} \pmod {\fP_1}
$$
for all $A \in \cO_K$.
\item
Let $\sigma_i \stackrel{\mathrm{def}}{\, = \,} \tau_i \sigma \tau^{-1}_i$. Observe that
\begin{equation}\label{E16}
\sigma_i A \equiv A^{N\frak p} \pmod {\fP_i}
\end{equation}
for all $A \in \cO_K$
so $\sigma_i$ is the Frobenius element associated with~$\fP_i$.
Thus $\sigma_i$ generates the decomposition group of $\fP_i$.
\end{itemize}
Claim: \emph{The Frobenius element associated with $\fP_i$ over $\Omega$ is equal to $\sigma_i^{l_i}$}.
To see this first observe that from (\ref{E16})
\begin{equation}\label{E17}
\sigma^{l_i}_i A \equiv A^{(N\frak p)^{l_i}} \equiv A^{N \frak q_i} \pmod {\fP_i},
\end{equation}
So to establish that $\sigma_i^{l_i}$ is the Frobenius element we just need to show that $\sigma_i^{l_i} \in H$.
In the special case where $A = \alpha \in \cO_\Omega$ we have from (\ref{E17}) and Fermat's little theorem that
$$
\sigma^{l_i}_i \alpha \equiv \alpha^{N \frak q_i} \equiv \alpha \pmod {\fP_i}.
$$
Since $\mf p$ does not divide the relative discriminant of~$K/k$, this means
that~$\sigma^{l_i}_i \alpha = \alpha$ for all $\alpha \in \cO_\Omega$ and hence for all $\alpha \in \Omega$.
So $\sigma^{l_i}_i \in H$ as desired.
\color{blue}
\begin{remark}
Note that $\sigma_i$ is in the decomposition group of $\mf P_i$, and so $\sigma_i^{l_i}$ is, of course, in this decomposition group.
Since $\mf P_i$ is unramifield over $\mf q_i$, the canonical map from the decomposition group of $\mf P_i$ to the Galois group of $\cO_K/\mf P_i$
is injective.
\end{remark}
\color{black}
Next we observe that $l_i$ is the smallest positive power $\nu$ of $\sigma_i$ such that $\sigma_i^\nu \in H$.
To see this observe that if $\sigma_i^\nu \in H$ then by (\ref{E16})
$$
\sigma^{\nu}_i \alpha = \alpha \equiv \alpha^{(N\mathfrak p)^{\nu}} \pmod {\frak P_i}
$$
for all $\alpha \in \cO_\Omega$. Thus $(N\mathfrak p)^{\nu} \ge (N \fp)^{l_i} = N \mf q_i$ and so $\nu \ge l_i$.
\color{blue}
\begin{remark}
The last step becomes clear when we observe that every element of the residue field has been shown to be a root of $X^{(N\mathfrak p)^{\nu}} - X$ which
is a polynomial in~$X$ of
degree $(N\mathfrak p)^{\nu}$. But the residue field has $(N \fp)^{l_i} = N \mf q_i$ elements.
\end{remark}
\color{black}
Claim: \emph{Consider cosets $H \sigma_{\nu}^a \tau_\nu$ and~$H \sigma_{\mu}^b \tau_\mu$. These cosets are equal if and only if
$\nu = \mu$ and~$a \equiv b \pmod {l_\nu}$.
}
One direction of this claim is straightforward since $\sigma_\nu^{l_\nu} \in H$, so if $a \equiv b$ modulo~$l_\nu$ then
$H \sigma_{\nu}^a =H \sigma_{\nu}^b$ and so $H \sigma_{\nu}^a \tau_\nu = H \sigma_{\nu}^b \tau_\nu$.
For the other direction, assume that~$H \sigma_{\nu}^a \tau_\nu = H \sigma_{\mu}^b \tau_\mu$, and so $$\sigma_{\nu}^a\; \tau_\nu = \tau_0 \; \sigma_{\mu}^b \; \tau_\mu$$
with $\tau_0 \in H$. So by the definition of $\sigma_\nu$ and $\sigma_\mu$
$$
\tau_0 = \sigma_{\nu}^a\; \tau_\nu \; \tau_\mu^{-1}\; \sigma_{\mu}^{-b} = \tau_\nu \; \sigma^{a-b} \; \tau_\mu^{-1}
$$
and, since $\sigma$ is in the decomposition group of $\mf P_1$,
$$
\tau_0 \mf P_\mu = \tau_\nu \; \sigma^{a-b} \; \tau_\mu^{-1} \mf P_\mu = \tau_\nu \; \sigma^{a-b} \mf P_1 = \tau_\nu \mf P_1 = \mf P_\nu.
$$
Since $\tau_0 \in H$, it is the identity map on $\mf P_\mu \cap \mc O_\Omega = \mf q_\mu$, but the image of $\mf P_\mu \cap \mc O_\Omega$
is $\mf P_\nu \cap \mc O_\Omega = \mf q_\nu$. So $\mf q_\mu = \mf q_\nu$. Thus $\mu = \nu$. We
then have $\sigma_{\nu}^a\; \tau_\nu = \tau_0 \; \sigma_{\nu}^b \; \tau_\nu$ so that~$\sigma_\nu^{a-b} \in H$
which implies that $a \equiv b \pmod {l_\nu}$.
So we have identified $l_1 + \ldots + l_r$ distinct cosets of $H$.
But we know that $$l_1 + \ldots + l_r = [\Omega \colon k] = [G: H].$$
So we have identified all the right cosets of $H$.
Note that $H \sigma_{\nu}^a \tau_\nu = H \tau_\nu \sigma^a$, and so
have coset representations, as in (\ref{E4}) with $S_i$ varying in the sequence
$$
\tau_1, \tau_1 \sigma, \ldots, \tau_1 \sigma^{l_1-1}, \tau_2, \tau_2 \sigma, \ldots \ldots, \tau_r, \tau_r\sigma, \ldots, \tau_r \sigma^{l_r-1}.
$$
In other words, each $S_i$ is of the form $\tau_\nu \sigma^a$ with $0\le \nu \le r$ and $0\le a < l_\nu$.
According to (\ref{E5}), in the induced representation $\Gamma_\Delta$ of $G$, the element $\sigma \in G$ is represented by the matrix
described in terms of blocks as follows:
$$
B_\sigma = \left(
A_{S_i \sigma S_k^{-1} }\right)
=
\left(
A_{\tau_\nu \sigma^{a-b + 1} \tau_{\mu}^{-1} }
\right)
$$
where, as above, $A_{\tau_\nu \sigma^{a-b + 1} \tau_{\mu}^{-1} }$ is the zero block if $\tau_\nu \sigma^{a-b + 1} \tau_{\mu}^{-1}$ is not in $H$.
\color{blue}
\begin{remark}
Here the row blocks are indexed by $(\nu, a)$ and the column blocks are indexed by $(\mu, b)$.
\end{remark}
\color{black}
Note that $\tau_\nu \sigma^{a-b + 1} \tau_{\mu}^{-1} \in H$ if and only if $\tau_\nu \sigma^{a-b + 1} \in H \tau_\mu$.
But since $\tau_\nu \sigma^{a-b + 1}$ is in the coset $H \sigma_\nu^{a-b+1}\tau_\nu$ we conclude that the block is zero unless
both~$\mu = \nu$ and~$a-b+1 \equiv 0 \pmod {l_\nu}$.
For a fixed $\nu$ we can consider the square matrix $C_\nu$
which is described as a block matrix whose $(a, b)$ block is the $\delta$ by $\delta$ square matrix
$\left( A_{\tau_\nu \sigma^{a-b + 1} \tau_{\nu}^{-1} }\right)$.
In particular the $(a, b)$ block is zero unless $a-b+1 \equiv 0 \pmod {l_\nu}$. Note that $C_\nu$ is an $\l_\nu \delta$ by $\l_\nu \delta$ square matrix.
Then (for a suitable ordering of a basis) one can write $B_\sigma$ in terms of blocks as follows:
$$
B_\sigma = \begin{pmatrix}
C_1 & 0 & \cdots & 0 \\
0 &C_2 & \cdots & 0 \\
\vdots & \vdots & & \vdots \\
0 & 0 & \cdots & C_r
\end{pmatrix}.
$$
If $a = 0, 1, \ldots, l_\nu -2$ then the $(a, b)$ block of $C_\nu$ is zero unless $b=a+1$, and when~$b=a+1$ the
block is $A_{\tau_\nu \sigma^{a-b + 1} \tau_{\nu}^{-1} }$ which is the $\delta$ by $\delta$ identity matrix $E$.
If~$a = l_\nu - 1$ then the $(a, b)$ block of $C_\nu$ is zero unless $b = 0$ and the $(l_\nu-1, 0)$ block is
given by~$A_{\tau_\nu \sigma^{l_\nu} \tau_{\nu}^{-1} } = A_{ \sigma_\nu^{l_\nu} } = A_{\sigma_\nu}^{l_\nu}$.
So $C_\nu$ decomposes into blocks as follows:
$$
C_\nu = \begin{pmatrix}
0 & E & 0& \cdots & 0 \\
0 &0 & E & \cdots & 0 \\
\vdots & \vdots & \vdots & & \vdots \\
0 & 0 & 0 & \cdots & E\\
A_{\sigma_\nu^{l_\nu}} & 0 & 0 & \cdots & 0
\end{pmatrix}.
$$
The characteristic polynomial in $t$ is then
$$
|E -t B_\sigma| = \prod_{\nu = 1}^r \left| E - t C_\nu \right|
= \prod_{\nu = 1}^r
\begin{vmatrix}
E & - t E & 0& \cdots & 0 \\
0 &E &-t E & \cdots & 0 \\
\vdots & \vdots & \vdots & & \vdots \\
0 & 0 & 0 & \cdots & - t E\\
- t A_{\sigma_\nu^{l_\nu}} & 0 & 0 & \cdots &E
\end{vmatrix}.
$$
Adding $t$ times the first column to the second, then $t$ times the (new) second to the third, and so on, one gets
$$
|E -t B_\sigma| = \prod_{\nu = 1}^r
\begin{vmatrix}
E & 0 & 0& \cdots & 0 \\
0 &E &0 & \cdots & 0 \\
\vdots & \vdots & \vdots & & \vdots \\
0 & 0 & 0 & \cdots & 0\\
- t A_{\sigma_\nu^{l_\nu}} & - t^2 A_{\sigma_\nu^{l_\nu}} & - t^3 A_{\sigma_\nu^{l_\nu}} & \cdots &E - t^{l_\nu} A_{\sigma_\nu^{l_\nu}}
\end{vmatrix}.
$$
Thus
$$
|E - t B_\sigma|
= \prod_{\nu = 1}^r
\left| E - t^{l_\nu} A_{\sigma_\nu^{l_\nu}} \right|.
$$
Note this last formula does not depend on the choice of $\sigma$ since different choices of Frobenius elements gives the same characteristic polynomials.
The contribution of $\mf p$ to $L(s, \chi_\psi; k)$ is
$$
\frac{1}{|E - (N\mf p)^{-s} B_\sigma|}
=\prod_{\nu = 1}^r
\frac{1}{| E - (N\mf p)^{-l_\nu s} A_{\sigma_\nu^{l_\nu}} |}
=\prod_{\nu = 1}^r
\frac{1}{| E - (N\mf q_\nu)^{-s} A_{\sigma_\nu^{l_\nu}} |}.
$$
We have already seen that the Frobenius element associated with $\fP_i$ over $\Omega$ is equal to $\sigma_i^{l_i}$.
So the right hand side of the above formula gives the product of the~$\mf q_\nu$ contributions to $L(s, \psi; \Omega)$.
Thus Satz 1 is proved.
\end{proof}
\chapter{Factorization of Zeta Functions}
Satz 1 gives us, for starters, a factorization of zeta functions of intermediate fields $\Omega$ in terms of primitive $L$-series associated to $K/k$
When we consider the trivial representation and the trivial character $\chi = 1$ (der Hauptcharakter $\chi_1$) we get
$$
L(s, \chi_1; k) = \prod_{\frak p} \frac{1}{| E - \left(N \frak p\right)^{-s} A_{\frak p} |} = \prod_{\frak p} \frac{1}{1 - \left(N \frak p\right)^{-s} }
$$
which is, up to a finite number of factors, just the zeta function $\zeta_k(s)$ of the base field.
More generally if $\Omega$ is an intermediate field between $k$ and $K$, and if $H$ is the Galois group of $K/\Omega$ with trivial character (Hauptcharakter) $\psi_1$
then
$$L(s, \psi_1; \Omega) = \zeta_\Omega (s),$$
at least up to a finite number of factors. Let $\Pi_\Omega$ be the induced representation
associated with the trivial representation of $H$. Note that $\Pi_\Omega$ is simply the representation associated with the permutation of cosets of $H$ in $G$ (so if $\Omega$
is itself Galois over~$k$, it corresponds to the regular representation of the Galois group of $\Omega$ over $k$).
Thus the associated character $\chi_\Omega$ has the property that, for any $\sigma \in G$, the value~$\chi_\Omega(\sigma)$ is the number of cosets fixed by $\sigma$ under this action,
so is determined in a most simple manner. If we decompose~$\chi_\Omega$ in terms of primitive characters
$$
\chi_\Omega (\sigma) = \sum_{i = 1}^x g_i \chi^i(\sigma)
$$
then $g_i$ is obtained using (\ref{E2}):
\begin{equation}\label{E18}
g_i = \frac{1}{n} \sum_{\sigma} \chi_\Omega (\sigma) \chi^i(\sigma^{-1})
\end{equation}
($n$ is the order of $G$ and so is $n = [K:k]$).
So Satz 1 in combination with (\ref{E14}) implies
\begin{equation}\label{E19}
\zeta_{\Omega} (s) = \prod_{i=1}^x \left( L(s, \chi^i) \right)^{g_i}
\end{equation}
which is the desired factorization (up to a finite number of factors).
{\color{blue}
\begin{remark}
From what Artin has said up to this point it is apparent that he regards~$G$ as acting on the left for its natural action on $K$, but regards~$G$ as
acting on the right for linear representation. Under this convention
$\Pi_\Omega$ is the permutation representation of the right action of $G$ on the collection $H\backslash G$ of right cosets. However, the associated character
$\chi_\Omega(\sigma)$ is the same whether we use left actions or right actions here (in other words, the number of left cosets fixed by $\sigma \in G$ is
the same as the number of right cosets fixed by $\sigma$).
\end{remark}
}
In the special case of $K = \Omega$ the induced representation is the regular representation and we get the simple formula
\begin{equation}\label{E20}
\zeta_{K} (s) = \prod_{i=1}^x \left( L(s, \chi^i) \right)^{f_i}.
\end{equation}
{\color{blue}
\begin{remark}
Here $f_i$ is the degree of character $\chi_i$. So if $K/k$ is Abelian, we have $f_i = 1$.
In general, all the primitive $L$-series for $K/k$ occur in the factorization.
\end{remark}
}
Formula (\ref{E19}) gives all the relations between the zeta functions of intermediate fields. To get such a relation, one uses (19) for various $\Omega$ and eliminates the $L$-series factors $L(s, \chi^i)$.
One is left with relations between zeta functions.
So the equations~(\ref{E19}) can be regarded as parameterizing relations. We will show later (Section~8) that this is the only way to get relations between zeta functions (when we reduce to the case~$k={\mathbb{Q}}$).\footnote{See E.~Artin, \"Uber
die Zetafunktionen gewisser algebraisher Zahlk\"orper
(Concerning the zeta functions of certain algebraic number fields),
Math. Ann Bd. 89, where the relations in special cases are obtained.}
This essentially solves the problem of relations between zeta functions.
There is another way to formulate our results. Observe that the factorization~(\ref{E19}) of $\zeta_\Omega(s)$
runs parallel to the factorization into irreducible polynomials of the group determinant (Gruppendeterminante) associated to the permutation representation $\Pi_\Omega$.
So one can say the following:
One gets all the relations between zeta functions of intermediate fields by finding the relations between
the group determinants associated with transitive permutation actions of $G$, and replacing
the group determinants with the corresponding zeta functions.
{\color{blue}
\begin{remark}
The ``group determinants'' that Artin mentions above are certain homogeneous polynomials associated to groups and their representations.
They are not as familiar today as they were when Artin wrote this paper, so I will give some details.
They are called ``determinants'' since
they arise as determinants of matrices with entries
that are homogeneous linear polynomials. These
polynomials were studied by Dedekind and Frobenius, and their study led Frobenius to his theory of characters of non-Abelian groups in 1896
that is in fact the basis of the current paper (see~\cite{Hawkins1971history}).
They are easy enough to define:
consider the polynomial ring~${\mathbb{C}} [X_{g_1}, \ldots, X_{g_n}]$ associated to a given finite group $G = \{g_1, \ldots, g_n\}$ where the $X_{g_i}$
are independent variables. If $g \mapsto A_{g}$ is a representation of $G$ by complex matrices, then the determinant associated to the representation is
simply the determinant of the following matrix:
$$
A_G \stackrel{\mathrm{def}}{\, = \,} \sum_{g \in G} X_g A_g.
$$
The matrix $A_G$ has a particularly nice description if the representation is a permutation representation, and even more so for the regular
representation (it is a good exercise to work these out). The determinant associated to the regular representation is called the ``group determinant'' of $G$ and can be thought of as
a fundamental algebraic invariant of $G$.
The determinant associated to a representation is an irreducible polynomial if and only if the representation is an irreducible representation,
and the decomposition of a representation is reflected in the factorization of its associated determinant. Observe also that the degree of such a determinant polynomial is equal to
the degree of the representation. Note that, historically speaking, the problem of factoring the group determinant proceeds, and in fact motives, the problem of decomposing a representation into irreducible factors that is the starting point of modern representation theory (See \cite{Hawkins1971history}).
For a simple example, the group determinant of a two-element group~$G=\{1, \sigma\}$ is just~$X_1^2 - X_\sigma^2$ which factors
as $(X_1 + X_\sigma) (X_1 - X_\sigma)$, reflecting the fact that the regular representation of $G$ decomposes into two
irreducible representations, each of degree 1.
\end{remark}
}
For now these relations are only valid up to a finite number of factors. Because of the existence of functional equations for zeta functions, we can use the well-known methods of
Herrn Hecke to show the relations are exactly valid.\footnote{E.~Hecke: \"Uber eine neue Anwendung der Zetafunktion auf die Arithmetik
der Zahlk\"orper
(concerning a new application of zeta functions to the arithmetic of number fields)\color{black}. G\"ottinger Nachrichten 1917.}
{\color{blue}
\begin{remark}
These methods of Hecke allow us to use functional equations of zeta functions to conclude that if a relation between zeta functions is valid up to a finite number
of Euler factors, then the relation holds exactly.
(See Lemma~\ref{fixingL_lemma} below for an illustration of this phenomenon.)
\end{remark}
}
Of course, similar considerations apply for relations between $L$-Series of intermediate fields.
{\color{blue}
\begin{remark}
We can use Artin's results to get an even more dramatic conclusion.
Suppose $\zeta_\Omega$ is a zeta function with base number field $\Omega$, or more generally consider $L$ functions with base field $\Omega$.
Then by result alluded to at the and of Section~2, we can take $K$ to be an extension of $\Omega$ that is Galois over ${\mathbb{Q}}$. So the above
considerations allow us to express $\zeta_\Omega$ (or more general $L$-functions) in terms of primitive~$L$-functions over ${\mathbb{Q}}$.
Artin, in Section 8 below, will show that this decomposition is unique.
\end{remark}
}
\chapter{The Abelian Case}
We now consider the case where $G$ is Abelian. We investigate whether the primitive~$L$-series defined in this document correspond to the
usual $L$-series.
{\color{blue}
\begin{remark}
These earlier $L$-series were defined by Weber and generalize those defined by Dirichlet.
They are defined in terms of characters of class groups (where characters are understood here in the traditional Dirichlet-Dedekind sense
as a homomorphism from a finite Abelian group into ${\mathbb{C}}^\times$).
\end{remark}
}
When $G$ is Abelian, each conjugacy class has a single element. So for each prime~$\mf p$ of~$k$ not dividing the relative discriminant of~$K/k$ there is exactly one Frobenius
element $\sigma \in G$, and (\ref{E8}) holds for all primes $\mf P$ in $K$ above $\mf p$.
One can replace (\ref{E8}) with the congruence
\begin{equation}\label{E21}
\sigma A \equiv A^{N \frak p} \pmod {\mf p}.
\end{equation}
Further, the irreducible representations of $G$ are all of degree 1, and they
correspond to the ordinary Abelian characters $\chi^i(\sigma)$ of $G$. Hence
\begin{equation}\label{E22}
L(s, \chi^i) = \prod_{\mf p} \frac{1}{1-\frac{\chi^i (\sigma)}{N \mf p^s}}
\end{equation}
where $\sigma$ denotes the Frobenius element associated to $\mf p$.
{\color{blue}
\begin{remark}
Equation (\ref{E21}) follows from the Chinese remainder theory. In (\ref{E22}) the Frobenius element $\sigma$ depends on $\mf p$. Artin makes this implicit, but a notation such as $\sigma(\mf p)$ could be used here if we wanted to make this explicit.
\end{remark}
}
Now in this situation $K$ is the class field of a certain
class group $\{ C_1, \ldots, C_n \}$ for a certain modulus $\mf m$ (a certain ideal of $\mc O_k$)
with the property that a prime ideal~$\mf p$ of $\mc O_k$ prime to $\mf m$ splits into prime ideals of the first degree in $\mc O_K$ if and only if $\mf p$ is in~$C_1$ where~$C_1$ is the identity class (Hauptklasse).\footnote{See
Teiji Takagi: \emph{\"Uber eine Theorie des relativ Abelschen Zahlk\"orpers}
{(concerning a theory of relative Abelian number fields)}, Journal of the College of Science, Tokyo 1920 \cite{Takagi1920}.
Further reference to Takagi will generally be from this paper.}
{\color{blue}
\begin{remark}
We can think of $\{ C_1, \ldots, C_n \}$ as a certain quotient group of the multiplicative group of fractional ideals whose prime factors are prime to $\mf m$.
In other words, each $C_i$ is a class of fractional ideals prime to $\mf m$.
There is a minimal ideal~$\mf m$ that we can use called the conductor, but we get well-defined version of the class group when we use multiplies of this minimal modulus.
Replacing a modulus by a multiple gives a class group that is naturally isomorphic to the first, so we can
often say ``the ideal class group'' associated to $K/k$ is we are not concerned about the exact modulus. However replacing a modulus with a multiple can
reduce the set of prime ideals of $\mc O_k$ prime to the modulus, but only by a finite number.
\end{remark}
}
The identity between our new $L$-series and the usual $L$-series will be shown once we are able to prove the following:
\begin{satz}
\
\emph{a)} The Frobenius element $\sigma$ of $\mf p$ depends only on the ideal class $C_i$ containing~$\mf p$, (so we can
assign a Frobenius element to each ideal class $C_i$ by choosing any prime ideal in that class as a representative).
\emph{b)} This Frobenius map gives an isomorphism between
the ideal class group and the Galois group~$G$.
\end{satz}
{\color{blue}
\begin{remark}
Observe that if Satz 2 holds for a certain modulus $\mf m$ then it automatically holds for any multiple of $\mf m$.
So there are really two versions of Satz 2, the strong version and the weak version. The strong version asserts the result where the class group
is taken with any valid modulus $\mf m$, or equivalently with the conductor as the modulus. The weak version asserts the result for some modulus $\mf m$, or equivalently
asserts (a) for ``almost all'' prime ideals, i.e. all prime ideals of $\mc O_k$ outside a certain finite subset (and where we can then let $\mf m$ be any valid modulus).
When we know that almost all prime ideals of a given ideal class $C_i$ must have the same Frobenius element, we can conclude that all ideal classes containing
infinitely many prime ideals can be assigned a well-defined Frobenius element.
But note that every ideal class $C_i$ contains an infinite number of prime ideals $\mf p$ of $\mc O_k$ by a suitable generalization of Dirichlet's theorem concerning
primes in arithmetic progressions. Thus we can assign a Frobenius element to any class. This is the content of the first part of Satz 2.
\end{remark}
}
This result implies that every character of the Galois group $G$ is then a character of the ideal class group and conversely.
So any $L$-series in our sense is then a $L$-series in the usual sense.
Conversely, if an ordinary $L$-series is given for an ideal class group then it will be an $L$-series for the character of the Galois group
of the associated class field.
So Satz 2 implies that our new definition is indeed a generalization of the old definition, agreeing with the old definition in the case where $K/k$
is Abelian.
{\color{blue}
\begin{remark}
Satz 2 is called ``Artin reciprocity''. It is the culmination of classical class field theory, and will be proved
by Artin in an article~\cite{Artin1927} appearing a few years later in 1927.
When Artin wrote the current article in 1923, Teiji Takagi had already developed class field theory to a very high degree, and Artin builds on this here.
Takagi's results give the following.
If $K/k$ is an Abelian extension of degree~$n$ then~$K$ is the class field of a class group~$\mathcal C = \{ C_1, \ldots, C_n \}$ defined with respect to a modulus $\mf m$
for some ideal~$\mf m$ of $\cO_k$. What this means is that $\{ C_1, \ldots, C_n \}$ partitions the collection of ideals, and even fractional ideals,
of~$\cO_k$ prime to~$\mf m$.
Furthermore, the set $\mathcal C = \{ C_1, \ldots, C_n \}$ of these classes is a group where $C_i C_j$ is defined as the class containing $I_i I_j$
for any choice $I_i \in C_i$ and $I_j \in C_j$.
The modulus $\mf m$ is such that all prime ideals $\mf p \in C_i$ prime to $\mf m$ are unramified in $\cO_K$ in the sense that~$\mf p \cO_K$
factors into distinct prime ideals. Furthermore, for such $\mf p$ prime to $\mf m$,
we have that~$\mf p$ is in the identity class $C_1$ if and only if $\mf p$ splits in $\cO_K$ (in the sense that~$\mf p \cO_K$ factors into $n$ distinct primes of relative degree 1).
Another very important result of Takagi is that $\mathcal C = \{ C_1, \ldots, C_n \}$ is isomorphic to the Galois group $G$ of $K/k$.
Interestingly, Takagi showed the isomorphism abstractly and did not supply a particular isomorphism. What Artin reciprocity does is gives a explicit
canonical isomorphism $\mathcal C \to G$.
\end{remark}
}
Satz 2 is also of interest in itself. It gives an explicit description of the isomorphism between the Galois group $G$ and the ideal class group.
In the case where $G$ is cyclic, Satz 2 is completely identical with the general reciprocity law, assuming the base field $k$ has the associated roots
of unity. And indeed the agreement is so obvious that Satz 2 has to be interpreted as as the general reciprocity law (even when~$k$ does not have the associated roots of unity) even if the formulation seems a bit strange (fremdartig) at first as a reciprocity law.
{\color{blue}
\begin{remark}
The general reciprocity referred here, and in the next paragraph, seems to be a version developed by Takagi mentioned in special case 5.~below. This law is less familiar today than other reciprocity laws, but
the important take-away is that Takagi's law generalizes the classical reciprocity laws. Since Artin reciprocity generalizes Takagi's reciprocity law
it automatically generalizes all the more familiar classical reciprocity laws.
\end{remark}
}
The situation is, however,
that our provisional proof of Satz 2 only really succeeds in the cases where the general reciprocity law is accessible to us, that is for~$K$ of prime degree
over $k$ or composite fields of such extensions. For general fields we must, for the time being, just postulate Satz 2. We will do so in
future sections which will allow us to regard all purely Abelian matters as being settled.
In this section we will prove Satz 2 in the cases accessible to us. We will proceed in stepwise fashion where we give the most general results possible in
in order to make the relationships stand out more clearly.
\bigskip
1. \emph{A prime ideal $\mf p$ is in the identity class $C_1$ (the ``Hauptklasse'') if and only if the corresponding Frobenius
element~$\sigma$ is the identity in $G$.}
{\color{blue}
\begin{remark}
Of course here we are only interested in prime ideals $\mf p$ of $\mc O_k$ prime to the modulus $\mf m$. As we will see in the proof, this result holds for any valid modulus.
\end{remark}
}
\begin{proof}
If the Frobenius of $\mf p$ is the identity element of $G$ then $A \equiv A^{N \mf p} \pmod {\mf P}$ holds for all $A \in \cO_K$
and all primes~$\mf P$ of $\cO_K$ dividing $\mf p \cO_K$. This implies that the residue field $\cO_K/\mf P$
has $N \mf p$ elements and so the
degree $[\cO_K/\mf P : \cO_k/\mf p]$ is $1$. Thus~$\mf p \cO_K$ factors into primes of relative degree 1, which means that~$\mf p \in C_1$ by Takagi Satz 31.
Conversely, if $\mf p$ is in the identity class $C_1$ then $\mf p \cO_K$
factors into primes of relative degree~1. Thus $A \equiv A^{N \mf p} \pmod {\mf P}$ holds for all prime ideals~$\mf P$ dividing~$\mf p \cO_K$
and all $A \in \cO_K$.
This means that $\sigma = 1$ works as the Frobenius element.
\end{proof}
{\color{blue}
\begin{remark}
Note that $A \equiv A^{N \mf p} \pmod {\mf P}$ holds for all $A \in \cO_K$ if and only if every element of the residue field $\cO_K/\mf P$ is a root of $x^{Np} - x$.
Lagrange's theorem on the number of roots of a polynomial of a given degree and by Fermat's little theorem, this holds in turn if and only if $\cO_K/\mf P$ is equal to its subfield $\cO_k/\mf p$.
\end{remark}
\begin{remark}
The above, when combined with Takagi's class field theory, allows us to jump from homomorphisms to isomorphisms.
Suppose in fact that we have a homomorphism $\mathcal C \to G$ from the class group $\mathcal C$ associated to $K/k$ to the Galois group $G$ of $K/k$.
Suppose also that the class of any prime ideal $\mf p$ maps to the associated Frobenius element (perhaps even with a finite number of exceptions). Assume $C \in \mathcal C$ is a class in the kernel. Then 1.~implies that $C$ is the identity class
(using a density result via
Weber $L$-functions). Thus $\mathcal C \to G$ is injective. From Takagi's class field theory,
$\mathcal C$ and $G$ have the same size (in fact Takagi showed they are isomorphic), thus $\mathcal C \to G$ is surjective as well.
\end{remark}
\begin{remark}
The following result is one where we have to be careful about the distinction between the strong and weak versions of~Satz~2. The proof seems
to give the following: any modulus for which Satz 2 holds for $K/k$ will also yield Satz 2 for~$\Omega / k$ where~$\Omega$ is an intermediate field.
\end{remark}
}
\medskip
2. \emph{If Satz 2 is valid for an Abelian extension $K/k$ then it is valid for $\Omega/k$ for any intermediate field~$\Omega$.}
\begin{proof}
Let $G_\Omega \subseteq G$ be the Galois group of $K/\Omega$. Let $r$ be the order of $G_\Omega$ and let~$s$ be the index
of $G_\Omega$ in $G$. As usual the quotient $G/G_\Omega$ will be identified with the Galois group of $\Omega/K$.
We assume Satz 2 for the extension $K/k$ so there is a class group $\{C_1, \ldots, C_n\}$ relative to some modulus~$\mf m$,
and a Frobenius isomorphism~$\mc C \to G$ sending $C_i \in \mc C$ to the Frobenius element~$\sigma \in G$ associated to any prime ideal in~$C_i$.
Since $\Omega$ is an intermediate field, by Takagi's results $\Omega$ is the class field for a class group $\mc H = \{ H_1, \ldots, H_s\}$, and moreover $\mc H$ can
be chosen to come from a quotient group of~$\mc C$. In other words, we can use the same modulus $\mf m$ for $\mc H$ as for~$\mc C$, and we can
write the identity class (Hauptklasse) of $\mc H$ as the union of classes of~$\mc C$
$$
H_1 = C_1 \cup C_2 \cup \dots \cup C_r
$$
(where we reindex the elements of $\mc C$ as necessary).\footnote{\label{fn9}\color{blue}
Let $C_1$ be the identity class (die Hauptklasse) of $\mc C$
and let $\mc I_{\mf m}$ be the full group of fractional ideals of $\mc O_k$ relatively prime to the modulus $\mf m$.
Then
there is a principle of class field theory similar to what we find in Galois theory: the intermediate fields of $K/k$ are in bijective correspondence
with subgroups of $\mc I_\mf m$ containing $C_1$. This correspondence reverses inclusion.
Given such a subgroup $H_1$ of fractional ideals, the Galois group of the corresponding intermediate extension~$\Omega/k$ is isomorphic to $\mc I_m / H_1$, and the cosets
of $H_1$ in $\mc I_m$ give the class group associated to $\Omega / k$.
Note also that the prime ideals of $H_1$ are exactly the prime ideals in $\mc I_\mf m$ that split in $\Omega$.
}
By 1.~(above) if $\mf p$ is a prime ideal of $\mc O_k$ not dividing $\mf m$ then the Frobenius element of $\mf p$ in $G/G_\Omega$ (relative to the extension $\Omega/k$)
is equal to the identity coset $G_\Omega \in G / G_\Omega$ if and only if $\mf p \in H_1$.
So by the compatibility of the Frobenius for $K/k$ compared to $\Omega/k$ we have that the Frobenius element of $C_i$ in $G$ is in $G_\Omega$
if and only if $C_i \subseteq H_i$.\footnote{\color{blue} See Lemma~\ref{triple_lemma}.}
We now show that all primes in a given class $H_i$ have the same Frobenius element in $G/G_\Omega$. Since $\mc H$ comes from a quotient group of $\mc C$, we can
write $H_i$ as $C'_i H_1$ for some~$C'_i \in \mc C$. So if $\mf p \in H_i$ is a prime ideal we have $\mf p \in C'_i C_j$ for some $1\le j \le r$ (by the decomposition of $H_1$).
By Satz 1 for $K/k$ we have that $\mf p$ has Frobenius element~$\sigma_i \tau_j \in G$ where $\sigma_i \in G$ is the Frobenius element of the class $C'_i$
and $\tau_j \in G_\Omega$ is the Frobenius element of $C_j$ (recall $C_j \subseteq H_1$ so $\tau_j \in G_\Omega$). Observe that $\sigma_i \tau_j$ is in the coset $\sigma_i G_\Omega$,
and so by the compatibility of the Frobenius elements for $K/k$ compared to $\Omega/k$ we have that the Frobenius element of $\mf p$ in $G/G_\Omega$ is $\sigma_i G_\Omega$.
Thus all primes $\mf p$ in $H_i$ have the same Frobenius element in $G/G_\Omega$. This proves the first part of Satz 1 for $\Omega/k$.
Now we have a well-defined Frobenius function $\mc H \to G/G_\Omega$, and we must show it is a homomorphism. This follows from the fact
that the following commutes, where the horizontal maps are the Frobenius maps and the vertical maps are the natural quotient maps:
$$
\begin{tikzcd} [column sep = normal, row sep = large]
\mc C \arrow[rr] \arrow[d]
& & G \arrow[d]
\\
\mc H
\arrow[rr]
& & G/G_\Omega
\end{tikzcd}
$$
Since the vertical map $\mc C \to \mc H$ is surjective, and since the top three maps are homomorphisms, the bottom map must also be a homomorphism.
This Frobenius map $\mc H \to G/G_\Omega$ is surjective since if $\sigma G_\Omega$ is in $G/G_\Omega$,
then~$\sigma$ is the Frobenius in~$G$ for some $\mf p$, which means~$\sigma G_\Omega$ is the corresponding Frobenius in~$G/G_\Omega$.
Since~$\mc H$ and $G/G_\Omega$ have the same order, the map is in fact an isomorphism.
\end{proof}
{\color{blue}
\begin{remark}
Artin's proof original is a bit terse, so I expanded it a bit in my translation (and even snuck in a commutative diagram not in the original).
I will add extra explanatory details to other proofs as we proceed.
One
thing Artin did not need to do, however, is to argue that the map $\mc H \to G/G_\Omega$ is surjective since as pointed in a remark after result 1.~above, we know such a Frobenius map
must be an isomorphism once we know it is a homomorphism. Alternatively, we can see surjectivity right away from the commutative diagram and the fact that the top and right maps are obviously surjective.
\end{remark}
\begin{remark}
In several places in the above proof we used the the compatibility of the Frobenius for $K/k$ compared to $\Omega/k$.
This was addressed at the end of Section~2 (and is summarized in Lemma~\ref{triple_lemma} in the commentary).
Note that this compatibility is what justifies the commutative diagram that I inserted into the above proof.
\end{remark}
\begin{remark}
The next result can also be regarded as a justification for either the strong or the weak versions of Satz 2. In other words, if the strong version of Satz 2
holds for $K_1$ and $K_2$ then the proof yields the strong version for $K_1 K_2$. If, however, only the weak version of Satz 2 holds for $K_1$ and $K_2$
then the proof can be regarded as a proof for the weak version of $K_1 K_2$. This is based on the observation
that any modulus valid for an Abelian extension is valid for any subextension.
\end{remark}
}
3. \emph{Suppose Satz 2 holds for two Abelian extension $K_1$ and $K_2$ of $k$ whose intersection is $k$, then it holds for the composite field $K = K_1 K_2$.
.}
\begin{proof}
Let $\mf m$ be common modulus such that Satz 2 holds for $K_1$ and $K_2$ with modulus~$\mf m$,
and let $C_1, \dots, C_n; D_1, \dots, D_m$ be classes taken for the modulus $\mf m$ where the $C_i$ form the class group for $K_1$
and $D_i$ form the class group for $K_2$.
Let~$G_1$ be the Galois group of $K_1/k$ and $G_2$ be the Galois group of~$K_2/k$.
Suppose that $C_i$ has Frobenius~$\sigma_i \in G_1$ and $D_j$ has Frobenius~$\tau_j \in G_2$.
As we know from Galois theory, the Galois group of $K_1 K_2 / k$ can be identified with $G_1 \times G_2$.
Note that $K_1 K_2$ is the class field associated to the class group
described by the partition of ideals prime to $\mf m$ given by the intersections $C_r \cap D_s$.\footnote{\color{blue}
This is not too difficult to show. In fact we
can appeal the the principle of footnote~\ref{fn9}. Let $\mc I_\mf m$ be the group of fractional ideals of $\mc O_k$ prime to $\mf m$, and let
$\mc P_\mf m$ be the subgroup of principal ideals with totally positive generators congruent to 1 modulo $\mf m$.
Then $\mc P_\mf m$ corresponds to the ray class
field $L_\mf m$ that clearly contains $K_1 K_2$ since it contains both $K_1$ and $K_2$.
Under the correspondence between subfields of $L_\mf m$ containing $k$ and subgroups of $\mc I_\mf m$ containing $\mc P_\mf m$,
the group $C_1$ corresponds to $K_1$ and $D_1$ corresponds to $K_2$.
So $C_1 \cap D_1$ corresponds to the
smallest subfield of $L_\mf m$ containing both $K_1$ and $K_2$, which is just $K_1 K_2$. The cosets of $C_1 \cap D_1$ in $\mc I_\mf m$ can be seen to be
the sets $C_r \cap D_s$
as desired. To see this consider the injective homomorphism~$\mc I_\mf m/(C_1 \cap D_1) \to \mc I_\mf m / C_1 \times\mc I_\mf m/ D_1$ which must be an isomorphism since
$[K_1 K_2 : k] = [K_1 : k ] [K_2 : k]$ (or equivalently, since $C_1 D_1$ corresponds to $k$, the smallest common subfield of $K_1$ and $K_2$, and so must be all of $\mc I_\mf m$).
}
The product of classes for this class group
is described by the following equation:
$$
(C_r \cap D_s) (C_u \cap D_v) = C_r C_u \cap D_s D_v.
$$
Now let $A_1 \in \mc O_{K_1}$ and $A_2 \in \mc O_{K_2}$ be generators fo $K_1/k$ and $K_2/k$ respectively. Let $A = \varphi(A_1, A_2)$
be in $\mc O_{K_1 K_2}$. Let $\mf p$ be in $C_r \cap D_s$. Then
$$
A^{N\mf p} \equiv \varphi( A_1^{N\mf p}, A_2^{N\mf p}) \equiv \varphi( \sigma_r A_1, \tau_s A_2) \equiv (\sigma_r, \tau_s) A \pmod \mf p.
$$
Suppose $A$ is an integral element of $K_1 K_2$ of the form $A_1 A_2$ with $A_1 \in \cO_{K_1}$ and~$A_2 \in \cO_{K_2}$.
If $\mf p$ is a prime ideal in $C_r \cap D_s$ then
$$
A^{N\mf p} = A_1^{N\mf p} A_2^{N\mf p} \equiv (\sigma_r A_1)(\tau_s A_2) = (\sigma_r, \tau_s) A \pmod{\mf p}
$$
Thus the Frobenius of any prime ideal in $C_r \cap D_s$ is $(\sigma_r, \tau_s)$, which is independent of the choice of $\mf p$. So the first part of Satz 2 holds.
The second part follows as well based on what we have shown.
\end{proof}
{\color{blue}
\begin{remark}
In the above proof, Artin does not describe explicity what $\varphi(x, y)$ is, but from context it seems to be a polynomial in $k[x, y]$. Furthermore,
to support the congruences, the coefficients should
be expressible as fractions of integral elements with denominators not in $\mf p$. Artin does not address the existence of such a polynomial. Fortunately, there
is straightforward way to prove the result that does not rely such a polynomial $\varphi(x, y)$:
As in the above proof, let $\mf p$ be a prime ideal of $C_r \cap D_s$ with Frobenius element~$(\sigma, \tau) \in G_1 \times G_2$.
By Lemma~\ref{triple_lemma}, and thinking of the Galois group of $K_1/k$ as the quotient~$G_1 \times G_2 / G_2$ (with $G_2$ embedded in $G_1 \times G_2$
in the usual way) then the Frobenius element of $\mf p$ for the extension $K_1/k$ is the coset
$$
(\sigma, \tau) G_2 = (\sigma, 1) G_2.
$$
Under the identification of $G_1 \times G_2 / G_2$ with $G_1$, which identifies the two descriptions of the Galois group of $K_1/k$,
this element is $\sigma$. Thus $\sigma = \sigma_r$ since $\sigma_r$ is the Frobenius element of $\mf p$ for $K_1/k$.
Similarly, $\tau = \tau_s$. Thus the Frobenius of $\mf p$ is~$(\sigma_r, \tau_t)$ as claimed.
\end{remark}
}
Note that because of 3.~(and the structure theorem of finite Abelian groups) we can reduce the proof of Satz 2 to cyclic extensions of
prime power degree. However, in this paper we will only fully succeed in proving Satz 2 in the case of cyclic extensions of prime degree.
\bigskip
4. \emph{Satz 2 holds for $K = k(\zeta)$ where $\zeta = e^{\frac{2\pi i}{m}}$ is an $m$th root of unity.}\footnote{An analogous proof
can be produced for class fields of complex multiplication. This shows how the reciprocity laws can be obtained through transcendental
generators of the class fields.}
\begin{proof}
Let $\mc C = \{C_1, \ldots, C_n\}$ be a class group associated to the field extension~$K/k$ where, as usual, $C_1$ is the identity class (die Hauptklasse).
For now we allow any modulus $\mf m$ for $\mc C$, valid for $K/k$, that at least satisfies the following condition: every prime ideal dividing $m\mc O_k$ also divides $\mf m$ (we will later show that $\mf m = m \mc O_k$ is in fact valid).
The first step is to identify the prime ideals in $C_1$ by determining a splitting law. In other words, we wish to describe which prime ideals $\mf p$ of $\mc O_k$
prime to $\mf m$ have
the property that $\mf p \mc O_K$ factors into distinct primes of relative degree one.
Given such a prime ideal $\mf p$, we know that $\mf p$ is unramified in $K/k$ and that the distinct $m$th-roots of unity in $\cO_K$ map to distinct
$m$th roots of unity in the residue field~$\cO_K/\mf P$
for any prime $\mf P$ above~$\mf p$. So $\cO_K/\mf P$ contains all the $m$th roots of unity.
Since $\mf p$ splits in~$\cO_K$, the residue field $\cO_k/\mf p$
is isomorphic to~$\cO_K/\mf P$ and so itself contains all the
$m$th root of unity. In this case the order~$N \mf p - 1$ of the multiplicative group $(\cO_k/\mf p)^\times$ is divisible by $m$. In other words,~$N \mf p \equiv 1 \pmod m$.
Conversely, suppose $N \mf p\equiv 1 \pmod m$ where $\mf p$ is a prime ideal of $\cO_k$ not dividing~$\mf m$.
Then for each algebraic integer
$A = \alpha_0 + \alpha_1 \zeta + \dots $ in $\cO_K$ (with $\alpha_i \in \cO_k$)
$$
A^{N \mf p} \equiv A \pmod{\mf p}.
$$
So the residue field~$\cO_K/\mf P$ has size bounded by $N\mf p$, and so equal to $N \mf p$, for all
primes $\mf P$ above~$\mf p$. Thus $\mf p$ splits in~$\cO_K$.
We have now established our desired splitting law: for prime ideals $\mf p$ of $\cO_k$ prime to $\mf m$, then $\mf p$ splits if and only if $N\mf p \equiv 1 \pmod m$.
So by a fundamental result of class field theory, for prime ideals $\mf p$ of $\cO_k$ prime to $\mf m$, we have $\mf p \in C_1$ if and only if $N\mf p \equiv 1 \pmod m$.
We wish to extend this to showing that $C_1$ consists the of the fractional ideals~$\mf a$ prime to $\mf m$ such that $N \mf a \equiv 1 \pmod m$,
and in fact that all fractional ideals in a given class~$C_i$ have the same norm modulo $m$. It turns out that we can do this by showing that $K$ is contained
in the ray class field of $k$ for modulus $m\mc O_k$, which will allow us to choose $\mf m$ to be~$m\mc O_k$.
So let $\mc C_m$ be the ray class group of $k$ modulo~$m$.\footnote{\color{blue}
The ray class group modulo $\mf m$ can be defined as $\mc I_\mf m / \mathcal P_\mf m$ where $\mc I_\mf m$ is the group of fractional ideals prime to $\mf m$
and $\mc P_\mf m$ is the subgroup of principal ideals generated by elements~$\alpha \in k$ such that~$\alpha \equiv 1 \pmod{\mf m}$
and such that $\alpha$ is positive in all real embeddings of~$k$. It is a basic result that every class of the ray class group contains integral ideals, and in fact prime ideals (by a generalization of Dirichlet's theorem).
The condition
$\alpha \equiv 1 \pmod{\mf m}$ can be interpreted as saying that $\alpha$ is the quotient $\beta/\gamma$ of algebraic integers such
that $\beta$ and $\gamma$ are prime to $\mf m$ and such that $\beta \equiv \gamma \pmod{\mf m}$.
In the current proof we are concerned with the ideal $\mf m = m \mc O_k$, and so we have $\sigma \beta \equiv \sigma \gamma \pmod{m}$
for all $\sigma$ in the Galois group of $K/k$. In particular,
$N \beta \equiv N \gamma \pmod {m}$, which we can express as saying that~$N\alpha \equiv 1 \pmod {m}$.
This is the norm of $\alpha$
as an element of ${\mathbb{Q}}$; the norm of the associated principal fractional ideal is the absolute value of the norm of its generator $\alpha$.
Since we assume that~$\alpha$ is positive in all real embeddings of $k$ in~${\mathbb{R}}$, its norm is positive, and so we get that~$N(\alpha\mc O_k) \equiv 1 \pmod m$
where here we mean the norm of the associated principal fractional ideal.}
Suppose $\mf a$ and $\mf b$ are ideals of~$\cO_k$ in the same class in~$\mathcal C_m$.
Then~$\mf a = \alpha \mf b$ for some~$\alpha \in k$ positive in all embeddings of $k$ into~${\mathbb{R}}$ and
such that~$\alpha\equiv 1 \pmod m$. For such $\alpha$ we have
$$
N (\alpha \mc O_k)
= |N\alpha | = N \alpha \equiv 1 \pmod m,
$$
so
$$
N \mf a \equiv N \alpha N \mf b \equiv N \mf b \pmod m.
$$
So all the ideals in a given class of $\mc C_m$ have the same norm, and we have a homomorphism~$\mathcal C_m \to ({\mathbb{Z}}/m{\mathbb{Z}})^\times$.
Combining classes of norm 1 yields a subgroup $\mc K_m$ of~$\mc C_m$ (the kernel of this norm homomorphism), and
the quotient $\mathcal C_m / \mc K_m$ determines a class group with the property that two fractional ideals (prime to $m$) are in the same class if and only if they
have the same norm.
In particular, $\mc C$ and $\mc C_m / \mc K_m$ both have the property that (with at most finitely many exceptions) a prime ideal $\mf p$ is in the identity class if
and only if $N \mf p= 1$. According to class field theory this means that the class fields of $\mc C$ and $\mathcal C_m / \mc K_m$ have the same primes that split (with a finite number of possible exceptions),
and so must be equal.
So $K$ is the class field of the class group $\mc C_m/\mc K_m$, where this class group is taken to have modulus $m \mc O_k$. We can now fix $\mf m$ to be $m \mc O_k$,
and identify~$\mc C$ with~$\mc C_m/\mc K_m$.
In particular all fractional ideals of a given class $C_i$ have the same norm modulo $m$.
Since $K = k(\zeta)$, we can view the Galois group $G$ of $K/k$ to be a subgroup of~$({\mathbb{Z}}/m{\mathbb{Z}})^\times$ where $\sigma$ is identified with the integer $t$ modulo $m$
for which $\sigma \zeta = \zeta^t$.
Let $C_i$ be a class of $\mc C$, and assume that the fractional ideals of $C_i$ have norm congruent to $n_i$ modulo~$m$. Let $\sigma \in G$ be the Frobenius element
of some prime $\mf p$ of $C_i$.
Since~$\sigma A \equiv A^{N\mf p} \pmod {\mf P}$
for all $A\in \cO_K$ and all primes $\mf P$ in $\mc O_K$ above $\mf p$, we have in
particular that
$$\sigma \zeta \equiv \zeta^{N\mf p} \equiv \zeta^{n_i} \pmod {\mf P}.$$
However $\sigma \zeta = \zeta^t$ for some integer~$t$.
So~$\zeta^{n_i} \equiv \zeta^t \pmod {\mf P}$. As mentioned above, distinct $m$-roots of unity in $\cO_K$ map to distinct $m$th roots of unity in the residue field~$\cO_K/\mf P$.
We conclude that $\zeta^{n_i} = \zeta^t$, and so, identifying $G$ with a subgroup of~$({\mathbb{Z}}/m{\mathbb{Z}})^\times$ ,
we see that the Frobenius element is just $n_i \in ({\mathbb{Z}}/m{\mathbb{Z}})^\times$.
In particular all primes of $C_i$ share the same Frobenius element, proving the first part of Satz 2.
By the multiplicativity of the norm map, the Frobenius map is a homomorphism. Observe that the Frobenius map has kernel consisting of the class $C_1$ alone
since only ideals in $C_1$ have norm conguent to 1 modulo $m$. Since $\mathcal C$ and $G$ have the same number of elements (according to Takagi's theory), the induced map $\mathcal C \to G$ is an isomorphism.
\end{proof}
{\color{blue}
\begin{remark}
As mentioned above, it is not really necessary to prove the map is an isomorphism since it being a homomorphism is enough. (See remark after claim~1.).
\end{remark}
\begin{remark}
This gives Satz 2 specifically for modulus $m \mc O_k$.
\end{remark}
\begin{remark}
At this point we know that at least a weak form of Satz 2 holds when~$k= {\mathbb{Q}}$ (using 2.,~4.~and the Kronecker-Weber theorem that every finite Abelian extension of~${\mathbb{Q}}$
is a subfield of~${\mathbb{Q}}(\zeta)$ for suitable $\zeta$).
\end{remark}
\begin{remark}
As with other proofs in this translation, the above proof is much expanded and somewhat modified from Artin's original proof in order to make
the argument more accessible to the modern reader. Here we provide more commentary for the proof.
Let $\mf p$ be prime to $\mf m$.
We can use the factorization of the polynomial $x^m - 1$ in $\cO_K [x]$ into linear polynomials and its reduction modulo $\mf p$ to get a factorization into linear
polynomials
$(\cO_K/\mf P) [x]$. Since the deriviative $m x^{m-1}$ is relatively prime to $x^m-1$, the roots in $(\cO_K/\mf P) [x]$ must be distinct
(we know that $m$ is not in $\mf P$ by our assumption on $\mf m$).
This explains why distinct
$m$th roots of unity map to distinct roots of unity in the residue field~$\cO_K/\mf P$.
We also used the fact that $(A_1 + A_2)^{N(\mf p)} \equiv A_1^{N(\mf p)} + A_2^{N(\mf p) }\pmod {\mf p}$ for all~$A_1, A_2 \in \mc O_K$. This follows from the fact that $N(\mf p)$
is a power of the characteristic $p$ of $\mc O_k/\mf p$.
\end{remark}
}
\bigskip
{\color{blue} The next result gives Satz 2 for a class of Kummer extensions:}
\medskip
5. \emph{Suppose $k$ contains the root of unity $\zeta = e^{\frac{2\pi i}{m}}$ where $m = l^n$ is a power of a prime $l$. Then Satz 2 holds for all
cyclic extensions $K$ of $k$ of degree $m = l^n$.}
\begin{proof}
It is a standard result of Galois theory\footnote{\color{blue}
See for instance Aluffi~\cite{Aluffi2009}, Chapter VII, Proposition 6.19. In fact, this result is so central to Galois theory that
it was essentially stated by Galois himself in the case that $m$ is prime, but
Galois' argument has a gap (essentially he fails to show $\mu \ne 0$).
See Edwards~\cite{Edwards1984}, \S 46, Page 63 for a discussion of the gap in Galois' manuscript and a simple way to fix it in a manner that would have been
accessible to Galois himself.}
that any such extension $K$ is of the form~$k\left( \mu^{1/m} \right)$ for some $\mu \in k$ and some fixed choice $\mu^{1/m}$ of $m$th root.
Observe also that the Galois group $G$ can be identified with the group of $m$th roots of unity: the action of $\sigma \in G$ is determined
by image of $\mu^{1/m}$ which must be of the form $c(\sigma) \mu^{1/m}$ for some $m$th root of unity~$c(\sigma)$.
The map~$\sigma \mapsto c(\sigma)$
is our desired isomorphism of~$G$ with the group of $m$th roots of unity.
For convenience we can take $\mu$ to be in~$\mc O_k$ so that $\mu^{1/m} \in \mc O_K$.
Suppose $\mf p$ is a prime ideal of $\cO_k$ prime to $l$. Since $\zeta \in \cO_k$, the
residue field~$\cO_k/\mf p$ has a primitive $m$th root of unity.
In other words $m$ divides~$N \mf p -1$, the order of the multiplicative group of the residue field~$\cO_k/\mf p$.
In other words, $N \mf p \equiv 1 \pmod {m}$.
Therefore,
$$
\left( \mu^{{1}/{m}} \right)^{N\mf p}
\equiv
\mu^{{(N\mf p - 1)}/{m}} \mu^{{1}/{m}}
\equiv
\left( \frac{\mu}{\mf p}\right) \mu^{{1}/{m}} \pmod {\mf p}
$$
where $\displaystyle \left( \frac{\mu}{\mf p}\right)$ is the $m$th power character, whose values are $m$th roots of unity.\footnote{\color{blue}
See for instance Lemmermeyer~\cite{Lemmermeyer2000}, Section 4.1, or Ireland and Rosen~\cite{IrelandR1990}, Section 14.2.}
In particular, the Frobenius element for $\mf p$ is the element of $G$ identified with the $m$th root of unity $\displaystyle \left( \frac{\mu}{\mf p}\right)$.
The essential statement of the general reciprocity law, as given by Takagi, is exactly that
$\displaystyle \left( \frac{\mu}{\mf p}\right)$ only depends on the class containing~$\mf p$
(in fact, this holds for any ideal $\mf a$ prime to~$\mu$).\footnote{Takagi~\cite{Takagi1922}.
}
So let $\mc C$ be a class group for $K/k$ with modulus $\mf m$ (containing~$\mu$ and $l$, say) for which we are certain that
$\displaystyle \left( \frac{\mu}{\mf a}\right)$ depends only on the class of $\mf a$ in~$\mc C$ for all integral ideals relatively prime to~$\mf m$.\footnote{\color{blue}
Artin's original proof does not specify what modulus $\mf m$ will work here. Perhaps it is $l \mu \mc O_k$ or $m \mu \mc O_k$. In any case, it should be clear by looking at Takagi's paper \cite{Takagi1922}.
Until I have the opportunity to consult Takagi's paper, I will just use any modulus that gets the job done here.
In fact, Artin does not mention the modulus at all in the proof. He also does not specify what~$l$ is, but it is pretty clear from context that $l$ must
at least be a prime. It could be that $l$ is restricted to odd primes. Again, it might require digging into Takagi's paper.
}
It follows now that the first part of Satz 2 holds for such a modulus $\mf m$.
The multiplicativity of the power character implies that the Frobenius map is a homomorphism. From this we get the rest of Satz 2.\footnote{\color{blue}
We get at least the weak form of Satz 2. Artin also mentions that $\displaystyle \left( \frac{\mu}{\mf p}\right)$ can take on any $m$th root of unity as a value and uses
this to justify surjectivity of the Frobenius map, but as mentioned after result 1. above this follows already from what we have done.
}
\end{proof}
{\color{blue}
\begin{remark}
The above proof is perhaps the most challenging for the modern reader to verify since it relies on results of Takagi that Artin does not
spell out in detail (nor do the modern sources I have consulted). The power character is well-known though and is easy to define.
Following Section 4.1 of \cite{Lemmermeyer2000},
let $k$ be a number field containing all the $m$th roots of unity where $m$ is a positive integer.
Recall that the reduction mod $\mf p$ map
sends distinct $m$th roots of unity to distinct $m$th roots of unity when $\mf p$ is prime to $m$. By Fermat's little theorem we have
$$
\alpha^{N\mf p - 1} \equiv 1 \pmod {\mf p}
$$
for all $\alpha \in \mc O_k$ outside of $\mf p$, and so
$
\alpha^{(N\mf p - 1)/m}
$
reduces to an $m$th root of unity in the residue field $\mc O_k/ \mf p$.
The power character $\left( \frac{\alpha}{\mf p} \right)_m$ for such an $\alpha$ and $\mf p$ is defined to be the unique $m$ root of unity such that
$$
\left( \frac{\alpha}{\mf p} \right)_m \equiv \alpha^{(N\mf p -1)/m} \pmod {\mf p}.
$$
This can be extended from $\mf p$ to other ideals
prime to $m$ by defining the symbol to be multiplicative with respect to ideal multiplication.
We can even define $\left( \frac{\alpha}{\beta} \right)_m$
for relatively prime elements nonzero elements $\alpha, \beta \in \mc O_k$, with $\beta$ prime to $m$, by
considering
the ideal generated by $\beta$.
The power character is central to the study of various reciprocity laws. For example,
the Eisenstein reciprocity law (\cite{IrelandR1990}, Section 14.2, or \cite{Lemmermeyer2000} Section 11.2) can be elegantly expressed using the power character for the field $k = {\mathbb{Q}}(\zeta_l)$:
\begin{thm} [Eisenstein reciprocity]
Suppose $l$ is an odd prime, suppose $\zeta_l$ is a primitive $l$th root of unity, and
suppose $a \in {\mathbb{Z}}$ is not divisible by $l$.
If $\alpha \in {\mathbb{Z}}[\zeta_l]$ is relatively prime to $a$, and if $\alpha$ is a primary element (meaning that $\alpha$ is not a unit, is prime to $l$, and is congruent to an element of ${\mathbb{Z}}$ modulo~$(1-\zeta_l)^2$)
then
$$
\left( \frac{\alpha}{a} \right)_l = \left( \frac{a}{\alpha} \right)_l .
$$
\end{thm}
According to~\cite{Lemmermeyer2000} (Section 11.4),
Takagi~\cite{Takagi1922} generalized this result from~${\mathbb{Q}} (\zeta_l)$ to any number field containing~$\zeta_l$.
Apparently Takagi connected such reciprocity laws with his class field theory, which then opened the door to the above result of Artin and to Artin's general reciprocity law.
\end{remark}
\begin{remark}
Because of the close connection between the Frobenius element and the power character illustrated in the above proof, it is common to introduce reciprocity-like symbols
for the Frobenius. The expression
$$
\left( \frac{K/k} {\mf p} \right)
$$
denotes the Frobenius element associated to $\mf p$. As usual, here $K/k$ is an Abelian extension of number fields, and $\mf p$ is a prime ideal of $\mc O_k$
unramified in $K/k$. This symbol is called the \emph{Artin symbol} in honor of the ideas introduced in this paper. When $K/k$ is Galois but not Abelian,
the Frobenius element depends on the choice of prime above $\mf p$, and this leads to the \emph{Frobenius symbol} (introduced by Hasse)
$$
\left[ \frac{K/k} {\mf P} \right]
$$
where $\mf P$ is a prime ideal of $\mc O_K$ unramifield in $K/k$
(See Section 3.2 of \cite{Lemmermeyer2000}).
\end{remark}
}
6. \emph{Suppose $K = k(\alpha)$ is cyclic of degree $r = l^n$ over $k$ where $l$ is a prime, and suppose $\Omega = k(\zeta)$ is an extension of $k$ of degree $m$
where $\zeta = e^{{2\pi i}/{l}}$. So $m$ divides~$l-1$ and is necessarily prime to $l$.
If Satz 2 holds for $K^* = \Omega(\alpha)$ over~$\Omega$ then Satz 2 holds for $K$ over $k$.}
{\color{blue}
\begin{remark}
This claim can be generalized, with essentially the same proof, to the following:
\emph{Suppose $K/k$ is an an Abelian extension of degree $r$, and suppose $\Omega/k$ is an Abelian extension of degree
$m$ where $m$ and $r$ are relatively prime. Let $K^* = K \Omega$ be the composite field. If Satz 2
holds for $K^*/\Omega$ then Satz~2 holds for $K/k$ as well.}
In the following translation, Artin's original argument has been adapted to support this more general statement.
\end{remark}
}
\begin{proof}
We write our Galois groups as $G(K/k), G(K^*/K), G(\Omega/k), G(K^*/\Omega),$ and~$G(K^*/k)$, where $K^*$ is the composite field $K \Omega$.
Since $r$ and $l$ are relatively prime, the intersection of $K$ and $\Omega$ is $k$, and $G(K^*/k)$ can be identified with
$$
G(K/k) \times G(\Omega/k)
$$
using the usual isomorphisms from Galois theory.
This identification also allows us to identify $G(K/k)$ with $G(K^*/\Omega)$, and $G(\Omega/k)$ with~$G(K^*/K)$. We will write $G$ for both
$G(K/k)$ and $G(K^*/\Omega)$, and we will write $H$ for both $G(\Omega/k)$ with~$G(K^*/K)$. Thus, for example, if $\sigma \in G$
then $\sigma$ can be regarded as an automorphism of $K$ fixing $k$, or as the unique extension of this automorphism to
an automorphism $K^*$ that fixes $\Omega$.
Fix an ideal $\mf m$ of $\cO_k$ that gives a valid modulus for the class group of~$K^*/k$.
So~$\mf m$ is also a valid modulus for the subextensions $K/k$ and $\Omega/k$. Let~$\cC(K^*/k)$,~$\cC (K/k)$, and~$\cC(\Omega/k)$ be the respective class
groups all using modulus~$\mf m$. By replacing $\mf m$ by a multiple if necessary we can also choose~$\mf m$ so that Satz 2 holds for $K/\Omega$ with modulus $\mf m \mc O_\Omega$
\textbf{Step 1.}
The first step of the proof is to construct a class group $\cC(K^*/\Omega)$ with modulus~$\mf m \cO_\Omega$ together with
an explicit isomorphism $\cC(K^*/\Omega) \to \cC(K/k)$.
We begin by considering the relative norm map $\mathcal I_\Omega \to \mathcal I_k$ where $\mathcal I_\Omega$ is the group of fractional ideals
of $\Omega$ prime to $\mf m \cO_\Omega$
and where $\mathcal I_k$ is the group of fractional ideals of $k$ prime to $\mf m$.
Note that $\cC(K/k)$ can be described as a quotient group of $\mathcal I_k$ and so the composition
$$
\mathcal I_\Omega \to \mathcal I_k \to \cC(K/k)
$$
is a homomorphism. Let $\mf C_0$ be the kernel of this composition.
Observe that \text{if~$\beta \in \Omega$} is prime to $\mf m\mc O_\Omega$ and satisfies $\beta \equiv 1 \pmod {\mf m\mc O_\Omega}$
then the relative norm $N\beta$ in~$k$ must satisfy the congruence~$N \beta \equiv 1 \pmod{\mf m}$ (since $\mf m$ is the intersection
of $\mf m \cO_\Omega$ with $k$).
Furthermore, if $\beta$ is also positive in all real embeddings of $\Omega$ into~${\mathbb{R}}$ then the relative norm
$N(\beta) \in k$ is totally positive as well.
In particular, the principal ideal generated by such $\beta$ must be in the kernel~$\mf C_0$. This means that the quotient
group $\mathcal I_\Omega/\mf C_0$ yields a class group for modulus~$\mf m \cO_\Omega$.
Let $\mf q$ be a prime ideal of $\cO_\Omega$ prime to $\mf m \cO_\Omega$, and let $\mf p$ be the intersection of $\mf q$
with the subfield $k$. In particular the relative norm $N\mf q \subseteq \cO_k$ is $\mf p^f$ where $f$ divides the relative degree~$m$.
Observe that $\mf q$ splits in $K^*$ if and only if $\mf p$ splits in $K$, since $f$ does not divide $r$.
But $\mf p$ splits in $K$ if and only it is in the identity class of $\cC(K/k)$. Since $f$ is prime to the order of $\cC(K/k)$
this occurs if and only if $\mf p^f$ in in the identity class of~$\cC(K/k)$. In other words, $\mf q$ splits in $K^*$
if and only if $\mf q \in \mf C_0$. By Takagi's results this means that $K^*$ is the class field extension of $\Omega$
corresponding to $\mathcal I_\Omega/\mf C_0$. So we write $\cC(K^*/\Omega)$ for $\mathcal I_\Omega/\mf C_0$.
Furthermore, the homomorphism $\mathcal I_\Omega \to \cC(K/k)$ induces an injective homomorphism
$\cC(K^*/\Omega) \to \cC(K/k)$. Since both groups have order $r$, this map $\cC(K^*/\Omega) \to \cC(K/k)$,
induced by the relative norm map, is an isomorphism.
\textbf{Step 2.} The second step is to use the isomorphism of step 1, and the assumption of Satz 2 for $K^*/\Omega$,
to define a Frobenius isomorphism on the class group~$\cC(K/k)$. For the isomorphism we will try the composition
$$
\cC(K/k)
\to
\cC(K^*/\Omega)
\to
G(K^*/ \Omega)
\to
G(K/k)
$$
where the first map is the inverse of the isomorphism of step 1, the second is the Frobenius isomorphism that exists
by assumption of Satz 2 for $K^*/\Omega$, and the third map is the natural isomorphism given by restrictions of automorphisms.
This composition is an isomorphism, so to prove Satz 2 for $K/k$ and modulus $\mf m$ we just need to show that this maps the class of a prime ideal $\mf p$
to its corresponding Frobenius element in $G(K/k)$.
So fix a prime ideal~$\mf p$ of $\cO_k$ prime to~$\mf m$ and
let~$\mf q$ be a prime ideal in $\cO_\Omega$ above~$\mf p$.
Let~$(\sigma, \tau)$ be the Frobenius element in $G(K^*/k) = G\times H$ corresponding to $\mf p$.
Identifying $G$ and~$H$ with subgroups of $G\times H$, we can write this element as~$\sigma \tau$ and
$$
\sigma \tau A \equiv A^{N\mf p} \pmod {\mf p}
$$
for all $A \in \cO_{K^*}$.
So $\sigma A \equiv A^{N\mf p} \pmod {\mf p}$ for~$A \in \cO_K$,
and $\tau A \equiv A^{N\mf p} \pmod {\mf p}$ for $A \in \cO_\Omega$
(where here $G$ is identified with $G(K^*/\Omega)$ and $H$ is identified with~$G(K^*/K)$).
Thus $\sigma$ is the Frobenius element of $\mf p$ in $G=G(K/k)$, and $\tau$ is the Frobenius
element of $\mf p$ in $H=G(\Omega/k)$.
Note that the relative norm $N \mf q$ is~$\mf p^f$ where $f$ is the order of $\tau$ in $H$. So
$$
\sigma^f A \equiv
\sigma^f \tau^f A \equiv (\sigma \tau)^f A \equiv A^{{N\mf p}^f}
\equiv A^{{N\mf q}} \pmod {\mf q}
$$
for all $A \in \cO_{K^*}$ (where here $N \mf q$ is the absolute norm).
Thus $\sigma^f$ is the Frobenius element associated with $\mf q$.
Note that $f$ divides $m = [\Omega: k]$, so $f$ is relatively prime to $r = [K: k]$. This means that
$u f \equiv 1 \pmod r$ for some $u$, and $(\sigma^f)^u = \sigma$.
The isomorphism~$\cC(K^*/\Omega) \to \cC(K/k)$ from step 1
sends the class of $\mf q^u$ to the class of its relative norm~$(\mf p^f)^u$. The class of $\mf p^{f u}$ is
the class of $\mf p$ since $f u \equiv 1$ modulo~$r$. Thus the inverse isomorphism $\cC(K/k)
\to
\cC(K^*/\Omega)$
maps the class of $\mf p$ to the class of $\mf q^u$. Since the class of $\mf q$ maps to its Frobenius $\sigma^f$
under the next map~$\cC(K^*/\Omega) \to G(K^*/\Omega)$, the class of $\mf q^u$ maps to
$$(\sigma^f)^{u} = \sigma^{fu} = \sigma.$$
Finally, $\sigma$ maps to $\sigma$ under the map $G(K^*/ \Omega)
\to
G(K/k)$ (here we are identifying~$G = G(K/k)$ with $G(K^*/ \Omega)$).
In conclusion the above composition $\cC(K/k) \to G(K/k)$ sends the class of a prime ideal to its Frobenius element.
\end{proof}
\color{blue}
\begin{remark}
This shows that the weak version of Satz 2 for $K^*/\Omega$ implies the weak version of Satz 2 for $K/k$.
\end{remark}
\color{black}
\begin{remark}
We now see that Satz 2 holds for any Abelian extension of degree equal to the product of distinct primes.
To see this first assume that $K/k$ has prime degree~$l$. Using 5.~we have Satz 2 for $K(\zeta)/k(\zeta)$ where $\zeta$ is a primitive
$l$th root of unity. By 6.~we have Satz 2 for $K/k$ as well. Finally 3.~extends Satz 2 to $K/k$ when~$[K:k]$ the product of distinct primes, or
more generally when the Galois group is the product of cyclic groups of prime order.
\end{remark}
\color{blue}
\begin{remark}
So Artin has proved the following:
\end{remark}
\begin{thm}
Suppose $K/k$ is an Abelian extension of number fields with Galois group~$G$. If $G$ can be factored into cyclic groups of prime order
then the weak form of Satz 2 holds for $K/k$.
\end{thm}
\color{black}
\chapter{Continuation of $L(s, \chi)$ to $\Re(s) \le 1$}
We return to the general case, assuming Satz 2 holds for the Abelian case.
We write~$m(\sigma)$ for the order of an element $\sigma \in G$ of the Galois group of $K/k$.
For each~$\sigma \in G$, let $\mf g^\sigma$ be the subgroup generated by $\sigma$,
and let $\Omega_\sigma$ be the subfield of~$K$ of elements fixed by $\sigma$.
So $\mf g^\sigma$ is the Galois group of $K/\Omega_\sigma$.
Let $\psi_i^{(\sigma)}$ for $i=1, \ldots, m(\sigma)$ be the irreducible characters of the Abelian group~$\mf g^\sigma$
where $\psi_1^{(\sigma)}$ is the trivial character (the ``Hauptcharakter" or the ``principal character'').
If we denote by $\chi_{\psi^{(\sigma)}_i}$ the induced character of $G$ then, by
Satz 1, equation~(\ref{E15}), we have the following which is valid up to a finite
number of factors in the Euler products:
$$
L\left(s, \psi_i^{(\sigma)} \; ; \; \Omega_\sigma \right) = L\left(s, \chi_{\psi_i^{(\sigma)}} \; ; \; k \right).
$$
As in (\ref{E7}), we decompose each $\chi_{\psi^{(\sigma)}_i}$ and obtain the factorizations
\begin{equation} \label{E23}
L\left(s, \psi_i^{(\sigma)} \right)
=
\prod_{\nu = 1}^x \left( L (s, \chi^\nu) \right)^{r^{(\sigma)}_{i\nu}}
\qquad
(i = 1, 2, \dots, m(\sigma))
\end{equation}
where each $r_{i\nu}^{(\sigma)}$ is a nonnegative integer, and again with validity up to a finite number of factors in the Euler product.
By Satz 2, the left-hand side of (\ref{E23}) corresponds to a traditional $L$-series whose extension to ${\mathbb{C}}$ and functional equation was established by Hecke.
We can use the equations (\ref{E23}) to solve for $L(s, \chi^\nu)$ in order to prove the continuation of each~$L(s, \chi^\nu)$.
We can focus on the case $\nu > 1$ since~$L(s, \chi^1) = \zeta_k(s)$
is a Dedekind zeta function whose meromorphic continuation is known.\footnote{\color{blue}Actually for any
$\chi^i$ of degree 1 we have the continuation since that is the case that Hecke considered (assuming Satz 2).}
For $\nu > 1$ we will show that $L(s, \chi^\nu)$ can be expressed in terms of a product of rational powers of the $L\bigl(s, \psi_i^{(\sigma)}\bigr)$
where $\sigma$ varies in $G$ and $i$ varies in $\{2, \ldots, m(\sigma)\}$, avoiding the trivial character $\psi_1^{(\sigma)}$.
Because of (\ref{E23}) it suffices to show that the system of $x$ linear equations
\begin{equation} \label{E24}
\sum_{\sigma \ne 1}
\sum_{i=2}^{m(\sigma)}
r^{(\sigma)}_{i\nu} x_i^\sigma
= \delta_{k \nu} \qquad \nu = 1, 2, \dots, x
\end{equation}
can be solved for each given $k$ in the sequence $2, \ldots, x$.\footnote{\color{blue}
Here Artin is using $k$ as an index. Once we show (\ref{E24}) can be solved, $k$ will return
to its role as denoting the base field.}
{\color{blue}
\begin{remark}
Suppose $x^\sigma_i \in {\mathbb{Q}}$ is a solution to the above system of linear equations (for a fixed $k$), then
\begin{eqnarray*}
\prod_{\sigma\ne 1} \prod_{i=2}^{m(\sigma)} L\left(s, \psi_i^{(\sigma)} \right)^{x_i^{\sigma}}
&=&
\prod_{\sigma\ne 1} \prod_{i=2}^{m(\sigma)} \prod_{\nu = 1}^x L (s, \chi^\nu)^{r^{(\sigma)}_{i\nu} x^\sigma_i}
\\
&=&
\prod_{\nu = 1}^x \prod_{\sigma\ne 1} \prod_{i=2}^{m(\sigma)} L \left(s, \chi^\nu \right)^{r^{(\sigma)}_{i\nu} x^\sigma_i}\\
&=&
\prod_{\nu = 1}^x L (s, \chi^\nu)^{\delta_{k\nu}}\\
& =&
L (s, \chi^k).
\end{eqnarray*}
There is a subtlety here:
the $x_i^\sigma$ are allowed to be rational and so the above quantities are dependent on how the various rational powers
are chosen. Depending on the choices it is possible that the calculation is
valid only up to a $d$th root of unity where $d$ is a common denominator for the~$x_i^\sigma$.
So we should think of this equality as holding up to a $d$th root of unity, and as usual up to a finite number of Euler factors.
What we can safely say is that $L(s, \chi^k)^d$ can be expressed in terms of a product of integral powers of
the $L\bigl(s, \psi_i^{(\sigma)} \bigr)$ (ignoring a finite number of Euler factors), and so $L(s, \chi^k)^d$ has a meromorphic
continuation to ${\mathbb{C}}$. Another way to say this is that there is a meromorphic continuation of $L(s, \chi^k)$
on a Riemann surfaces $\mc L$ mapping onto~${\mathbb{C}}$ with fibers of size bounded by $d$. Or we can take the old point of view that $L(s, \chi^k)$
is a ``multivalued function'' that has an analytic continuation outside a discrete set of singularities.
As we will see, Artin suspects this continuation is single valued (that is, $\mc L$ can be taken to be ${\mathbb{C}}$). In other words,
Artin hoped that $L(s, \chi^k)$ itself, and not a power, has a meromorphic continuation.
This was first proved by R.~Brauer~\cite{Brauer1947a} in~1947. Artin's deeper conjecture that this continuation is actually analytic when~$\chi^k \ne 1$ is
still open.
\end{remark}
}
Now $r_{i1} = 0$ for each $i>1$, so equation~(\ref{E24}) with $\nu =1$ automatically holds.\footnote{
\color{blue}
This follows from (\ref{E6}).}
Thus we only need
to consider $\nu \ge 2$. So in order for (\ref{E24}) to be solvable, it is sufficient that the matrix
$$
\left( r_{i\nu}^{(\sigma)} \right) \qquad
\sigma \ne 1,
\qquad
i=2, \dots, m(\sigma);
\qquad
\nu = 2, \dots, x
$$
has rank $x-1$. Here we regard the columns as being indexed by $(\sigma, i)$ and rows as being indexed by~$\nu$.
For this matrix to have rank $x-1$ it is necessary and sufficient that the $x-1$ rows of this matrix be linearly independent.
So we just need to to show that the only solution to the system of linear equations
\begin{equation} \label{E25}
\sum_{\nu=2}^x r_{i \nu}^{(\sigma)} y_\nu = 0
\end{equation}
is the zero solution with $y_\nu = 0$
(where the system contains an equation for each~$(\sigma, i)$ where~$\sigma \ne 1$
and $2 \le i \le m(\sigma)$.)
So assume $y_2, \ldots, y_x$ is a solution to the system.
Fix $\sigma$ and $\tau \in \mf g^\sigma$ (where $\tau = 1$ is allowed), and for each $i$ from $2$ to~$m(\sigma)$
multiply~(\ref{E25}) by $\psi_i^{(\sigma)} (\tau)$. Now sum the resulting equations as $i$ varies:
$$
\sum_{i=2}^{m(\sigma)} \sum_{\nu=2}^x r_{i \nu}^{(\sigma)} \psi_i^{(\sigma)} (\tau) \, y_\nu = 0.
$$
Using (\ref{E6}) we can simplify this equation, giving
the following equation for each choice of $\sigma \in G$ and~$\tau \in \mf g^\sigma$:
$$
\sum_{\nu=2}^x \bigl( \chi^\nu(\tau) - r_{1 \nu}^{(\sigma)} \bigr) y_\nu
= 0
$$
or
$$
\sum_{\nu=2}^x \chi^\nu(\tau) y_\nu
=
\sum_{\nu=2}^x r_{1 \nu}^{(\sigma)} y_\nu.
$$
The right hand side does not depend on $\tau$, so the left hand side has the same value for all $\tau\in \mf g^\sigma$. In particular,
$$
\sum_{\nu=2}^x \chi^\nu(\tau) y_\nu = \sum_{\nu=2}^x \chi^\nu(1) y_\nu
$$
for all $\tau \in \mf g^\sigma$. Note the right hand side of this equation does not depend on $\sigma$, and so the left hand side has the same
value for all $\tau \in G$. Call this value $-y_1$, so
$$
\sum_{\nu=1}^x \chi^\nu(\tau) y_\nu = 0
$$
for all $\tau \in G$. Using (\ref{E2}), we see that for each $i \in \{1, \ldots, x\}$
\begin{eqnarray*}
0 = 0 \cdot \sum_{\tau \in G} \chi^{i} (\tau^{-1}) &=&\left(\sum_{\nu=1}^x \chi^\nu(\tau) y_\nu\right) \sum_{\tau \in G} \chi^{i} (\tau^{-1}) \\
&=&
\sum_{\nu=1}^x \left( \sum_{\tau \in G} \chi^\nu(\tau) \chi^{i} (\tau^{-1}) \right) y_\nu \\
&=&
\sum_{\nu=1}^x n \delta_{\nu i} y_\nu = n y_i.
\end{eqnarray*}
So $y_i = 0$ for all $i \in \{1, \ldots, x\}$, establishing the linear independence claim, and so the solvability of (\ref{E24}).
We can now express each $L(s, \chi^\nu)$ in terms of Abelian $L$-series, which gives us a way to extend $L(s, \chi^\nu)$
with properties similar to those of traditional $L$-series. For example, if $\chi^\nu$ is not the identity character ($\nu >1$) the expression
only involves~$L\bigr(s, \psi^{(\sigma)}_i\bigr)$ with~$i \ne 1$, so $L(s, \chi^j)$ is regular and nonvanishing at $s=1$.
Now we change our initial definition of $L$-functions.
A solution to (\ref{E24}) expresses~$L(s, \chi^\nu)$ in terms of a product of rational powers of traditional
$L$-series but only up to a finite number of factors. We can modify the definition of $L(s, \chi^\nu)$ so that this expression
is an exact equality, and then use (\ref{E14}) to define $L(s, \chi)$ for general characters.
This modified definition changes $L(s, \chi)$ up to a finite number of factors, so all our results that are valid up to a finite
number of factors will continue to hold with the modified definition.
The resulting $L(s, \chi)$ will analytically continue as a multivalued function on the whole plane ${\mathbb{C}}$ minus possibly a discrete set of branch points,
and going around a branch point will only change the value by a root of unity.
The functional equation of Hecke holds for the Abelian $L$-series,
so will yield a functional equation for our new $L$-series.
This functional equation can be used to show that the definition of our $L$-series is independent of the solution to~(\ref{E24})
used to build our new $L$-series.
{\color{blue}
\begin{remark}
Let $x_i^{\sigma} \in {\mathbb{Q}}$ be the numbers occurring in a solution to (\ref{E24}) (where we change~$k$ to $j$ in what follows), then Artin proposes to
use the resulting relation, originally valid only up to a finite number of Euler factors, as a new, modified definition:
$$
L (s, \chi^j) \; \stackrel{\mathrm{def}}{\, = \,} \; \prod_{\sigma\ne 1} \prod_{i=2}^{m(\sigma)} L\left(s, \psi_i^{(\sigma)} \right)^{x_i^{\sigma}}.
$$
This makes $L(s, \chi^j)$ a multivalued function on ${\mathbb{C}}$ minus a discrete set of branch points, that is to say it is a
meromorphic function on a Riemann surface covering~${\mathbb{C}}$.
If $d$ is the common denominator of the $x_i^{\sigma}$ then
$$
L (s, \chi^j)^d = \prod_{\sigma\ne 1} \prod_{i=2}^{m(\sigma)} L\left(s, \psi_i^{(\sigma)} \right)^{d x_i^{\sigma}}
$$
gives an exact equation between meromorphic functions, where the functions on the right
satisfy nice functional equations established by Hecke.
From this we see that Artin's definition actually gives $L(s, \chi^j)$ as a meromorphic function on a Riemann surface $\mc L$ which covers
${\mathbb{C}}$ with degree bounded by $d$.
If we want to derive a functional equation for this meromorphic function~$L (s, \chi^j)^d$ we need to observe that
we can use the same solution to (\ref{E24}) for writing $L(s, \overline {\chi}^j)^d$ in terms of Abelian $L$-series:
$$
L \left(s, \overline \chi^j\right)^d = \prod_{\sigma\ne 1} \prod_{i=2}^{m(\sigma)} L\left(s, \overline \psi_i^{(\sigma)} \, \right)^{d x_i^{\sigma}}
$$
where $\overline \chi^j$ denotes the complex conjugate of $\chi^j$, and ${\overline \psi_i^{(\sigma)}}$ denotes the complex conjugate of ${\psi_i^{(\sigma)}}$.
The validity of this can be
seen by oberving that~(\ref{E6}) and~(\ref{E7}) are well-behaved under complex conjugation, and the induced character of $\overline \psi_i^{(\sigma)}$ satisfies
$$
\chi_{\overline \psi_i^{(\sigma)}} \; = \; \overline{\chi_ {\psi_i^{(\sigma)}}}.
$$
This gives us a version of~(\ref{E23}) for conjugate characters using the same integers~${r^{(\sigma)}_{i\nu}}$ as the original~(\ref{E23}), and so a solution to (\ref{E24}) will work for both $L (s, \chi^j)^d$
and~$L \bigl(s, \overline \chi^j\bigr)^d$.
As we will see, the functional equation for Abelian $L$ series is of a form that is closed under
products, so gives a nice functional equation for $L(s, \chi^j)^d$. Artin further observes that the functional equation
forces $L(s, \chi^j)$,
or better $L(s, \chi^j)^d$,
to be independent of the solution to (\ref{E24}). In other words, there can be only one definition for $L(s, \chi^j)^d$
that satisfies such a functional equation and agrees with the earlier definition up to a finite number of Euler factors.
\end{remark}
}
The functional equations for the Abelian $L$-series, and hence our new $L$-series, has the following form:\footnote{E.~Landau, \"Uber Ideale und Primideale in Idealklassen \color{blue} (concerning ideals and prime ideals in ideal classes)\color{black}. Math. Zeitschrift Bd. 2, Seite 104, Satz LXVI.}
$$
{L (1-s, \overline \chi^i)}
=
a_i A^s
\left( \Gamma(s)\right)^{l_i^{(1)}}
\left( \cos \frac{s\pi}{2}\right)^{l_i^{(2)}}
\left( \sin \frac{s\pi}{2}\right)^{l_i^{(3)}}
{L(s, \chi^i)}.
$$
Here $l_i^{(1)}, l_i^{(2)}, l_i^{(3)}$ are rational, and $A$ is a positive real number. Note that $a_i$ depends on
a choice of branch, and $a_i$ may change by a root of unity
if we change the branch.\footnote{\color{blue}We can take $a_i$ to be a true constant and we can take $l_i^{(j)}$ to be integers if we raise both sides of the equation to an appropriate integral power.}
{\color{blue}
\begin{remark}
In verifying these claims it might be best to work with a power $L(s, \chi^j)^d$ that is meromorphic.
As mentioned above, the transformation from $\chi^j$ and $\overline \chi^j$ is well-behaved and we
can use the same solution to (\ref{E24}) for both $\chi^j$ and $\overline \chi^j$ to get compatible
decompositions. So since the above functional equation has a form that is closed under powers and products, we get
a functional equation for $L(s, \chi^j)^d$, and so for $L(s, \chi^j)$ for a choice of branch.
\end{remark}
\begin{remark}
The form of the functional equation for Abelian $L$-series used here by Artin is a bit different than the form it is usually given today, so it is worth a few comments. (I have not consulted Hecke's original paper, nor the paper of Landau cited by Artin. Instead I consulted
Tate's thesis. Tate was a student of Artin in the 1940s who showed how to replace Hecke's approach with an approach
using harmonic analysis on the id\`eles.)
Suppose $\chi$ is an Abelian character with conductor~$\mf f$. Then Tate's thesis gives a form of the functional equation
(see \cite{CasselsF1967} pages 342--346) that leads naturally to the version used by Artin.
To describe this, let $S$ be a finite set of places of $k$ including all divisors of the conductor $\mf f$ and all Archimedean places.
Tate shows that
$$
L(1-s, \chi^{-1}) = \left( \prod_{\mf p \in S} \rho_{\mf p} (s) \prod_{\mf p \not\in S} (N {\mf d}_{\mf p})^{s-1/2}
\; \chi^{-1}( {\mf d}_{\mf p}) \right) L(s, \chi)
$$
where $\rho_{\mf p}(s)$ denotes certain meromorphic functions related to $\chi$ and the place~$\mf p$, which are
explicitly
calculated in Tate's thesis (\cite{CasselsF1967}, Pages 317, 319, 322). Here~${\mf d}_{\mf p}$ denotes the local different ideal.
Recall that the absolute norm of the \text{product~$\mf d = \prod \mf d_{\mf p}$} of these ideals gives the absolute discriminant $|\Delta_k|$ of the field~$k$
(where $\mf p$ includes the non-Archimedean primes in $S$).
The functions~$\rho_{\mf p}(s)$ are as follows:
\begin{itemize}
\item
Suppose $\mf p$ is a real place.
Consider a principal ideals generated by $\alpha \in k^\times$ such that
(1) $\alpha \equiv 1 \pmod{\mf f}$, (2) $\alpha < 0$ at $\mf p$, and (3) $\alpha > 0$ for all real places not equal to~$\mf p$.
(Weak approximations assures such an $\alpha$ exists).
Then if $\chi$ has value $1$ on such a principal ideal $\alpha \cO_k$ then
$$
\rho_{\mf p}(s) = \left( \frac{2}{(2\pi)^s} \right) \cos \left( \frac{\pi s}{2}\right) \Gamma(s),
$$
but if $\chi$ has value $-1$ on $\alpha \cO_k$ then
$$
\rho_{\mf p}(s) = -i \left( \frac{2}{(2\pi)^s} \right) \sin \left( \frac{\pi s}{2}\right) \Gamma(s).
$$
\item
If $\mf p$ is a complex place then
\begin{eqnarray*}
\rho_{\mf p}(s) &=& (2\pi)^{1-2s}
\frac{\Gamma(s)}
{\Gamma(1-s)}\\
&=& 2 (2 \pi)^{-2s} \sin(\pi s) \Gamma(s)^2 \\
& = & \left( \frac{2}{(2\pi)^s} \right)^2\sin \left( \frac{\pi s}{2}\right)\cos \left( \frac{\pi s}{2}\right) \Gamma(s)^2
\end{eqnarray*}
(Note this is just $i$ times the product of the two formulas for real $\rho$).
The first equation is essentially the formula from Tate's thesis. The other equations are justified by
the identity
$$\Gamma(s) \Gamma(1-s) = \frac{\pi}{\sin \pi s}$$
and the double angle identity for the sine function.
\item
If $\mf p$ is a ramified non-Archimedean place with conductor component $\mf f_{\mf p}$ then
$$
\rho_{\mf p}(s) = N( \mf d_{\mf p} \mf f_{\mf p})^{s-1/2} W_{\mf p}(\chi)
$$
where $W_{\mf p}(\chi)$ is a certain root of unity called the root number.
\end{itemize}
When we substitute these formulas for $\rho_{\mf p}(s)$ and simplify we obtain the formula
$$
L(1-s, \chi^{-1}) = w
\left( \frac{2}{(2\pi)^s} \right)^{n}
(N(\mf f) |\Delta_k|)^{s-1/2} \sin \left( \frac{\pi s}{2}\right)^{n_1} \cos \left( \frac{\pi s}{2}\right)^{n_2}
\Gamma(s)^n L(s, \chi)
$$
where $n = [k: {\mathbb{Q}}]$, where $n_1$ and $n_2$ are two nonnegative integers with $n_1 + n_2 = n$, where $N(\mf f)$ is the norm of the conductor of $\chi$,
and where $w$ is a root of unity.
\end{remark}
{
\color{blue}
\begin{remark}
Artin makes the observation that the functional equation fixes the $L$-series exactly, not just up to a finite number of factors.
In other words, the functional equation picks out a canonical representation of a class of $L$-series up to ``finite Euler factor equivalence''.
Undoubtably this principal was well-known when Artin wrote this article, but it might be helpful to supply the details.
The following Lemma helps make this clear and can be proved just by considering the location of zeros and poles.
Note this lemma can be generalized to a much broader class of admissible functional equations, but we will stick to the concrete form given
in the paper.
\end{remark}
\begin{lemma}\label{fixingL_lemma}
Suppose $L_1(s), \widetilde{L_1}(s)$, $L_2(s), \widetilde{L_2}(s)$ are nonzero meromorphic functions on ${\mathbb{C}}$ and assume
the following two conditions. (i)
$$
\frac{L_i (s)}{ \widetilde{L_i}(1-s)} = a_i \; A_i^s \;
\Gamma(s)^{n_i^{(1)}}
\left( \cos \frac{s\pi}{2}\right)^{n_i^{(2)}}
\left( \sin \frac{s\pi}{2}\right)^{n_i^{(3)}}
$$
where $a_i \in {\mathbb{C}}^\times$, where $n_i^{(j)} \in {\mathbb{Z}}$, and where $A_i$ is a positive real constant. And (ii)
$$
L_2 (s) = P(s) L_1 (s), \qquad \widetilde{L_2}(s) = \widetilde{P}(s) \, \widetilde{L_1}(s)
$$
where
$$
P(s) = \prod_{i=1}^e (1 - \varepsilon_i p_i^{-s})^{u_i}, \qquad \widetilde{P}(s) = \prod_{i=1}^e (1 - \overline\varepsilon_i p_i^{-s})^{u_i}
$$
where the product is over distinct pairs $(\varepsilon_i, p_i)$ composed of a prime $p_i \in {\mathbb{Z}}$
and a root of unity~$\varepsilon_i \in {\mathbb{C}}^\times$, and where $u_i \in {\mathbb{Z}}$.
Then necessarily
$$
L_1 (s) = L_2(s)
$$
as meromorphic functions.
\end{lemma}
\begin{proof}
Consider the function
$$
R(s) = \frac{L_2 (s)}{ \widetilde{L_2}(1-s)} \cdot \frac { \widetilde{L_1}(1-s)} {L_1 (s)} = \frac{P(s)}{ \widetilde{P}(1-s)}
$$
which, according to the functional equations of assumption (i), has the form
$$
R(s) = a \; A^s \;
\Gamma(s)^{n^{(1)}}
\left( \cos \frac{s\pi}{2}\right)^{n^{(2)}}
\left( \sin \frac{s\pi}{2}\right)^{n^{(3)}}
$$
for some $a \in {\mathbb{C}}^\times$, $n^{(j)} \in {\mathbb{Z}}$, and $A$ a positive real constant.
Note that all the zeros and poles of $P(s)$ occur on the line $\Re(s) = 0$ (because $p_i^s$ can equal $\varepsilon_i$ only
on this line),
and all the zeros and poles of $\widetilde P(1-s)$ occur on the line $\Re(s) = 1$.
So the only possible real zeros and poles of $R(s)$ occur when~$s=0$ or~$s=1$; in particular the number of real zeros and poles of $R(s)$ is finite. Since~$\Gamma(s)$ has no zeros and poles for real~$s > 0$,
this forces~$n^{(2)} = n^{(3)} = 0$ in order to avoid an infinite number of real zeros or poles for~$R(s)$.
Since $\Gamma(s)$ has an infinite number of real poles (at nonnegative integers), we can conclude that $n^{(1)} = 0$ as well.
So $R(s) = a A^s$ has no zeros or poles.
As mentioned above, $P(s)$ and $\widetilde P(1-s)$ have disjoint sets of zeros and poles. Since the quotient $R(s)$ has no zeros or poles, this forces both $P(s)$ and $\widetilde P(1-s)$ to have no zeros or poles.
Consider
$$
P(s) = \prod_{j=1}^e (1 - \varepsilon_j p_j^{-s})^{u_j}
$$
and the zero sets of the factors $1 - \varepsilon_j p_j^{-s}$. We see that $s$ is in the zero set of the~$j$th factor
if and only if $p_j^s = \varepsilon_j$. If $\varepsilon_j = \exp(2\pi r_j i)$ with $r_j \in {\mathbb{Q}}$, then
$s$ is in the zero set if and only if $s =0 + t i$ with
$$
t = 2\pi \; \frac{r_j + k}{\log p_j}
$$
for some $k\in {\mathbb{Z}}$.
Observe that if $p_j = p_l$ but $\varepsilon_j \ne \varepsilon_l$, then there can be no common root
to $1 - \varepsilon_j p_j^{-s}$ and $1 - \varepsilon_l p_l^{-s}$ simply because $p_j^s = p_l^s$ cannot
be equal to both~$\varepsilon_j$ and~$\varepsilon_l$. So consider the case where $p_i \ne p_l$.
A common zero of $1 - \varepsilon_j p_j^{-s}$ and~$1 - \varepsilon_l p_l^{-s}$ would yield a real $t$ with
$$
t = 2\pi \; \frac{r_j + k}{\log p_j} = 2\pi \; \frac{r_l + k'}{\log p_l}
$$
where $k, k'\in {\mathbb{Z}}$. If, in addition, $t$ is no zero, then
we would be able to find two nonzero integers $a, b \in {\mathbb{Z}}$ where
$$
\frac{\log p_l}{\log p_j} = \frac{a} {b}
$$
and so $\log p_l^b = \log p_j^a$, or more simply $p_l^b = p_j^a$, a contradiction.
So the only possible common zero of $1 - \varepsilon_j p_j^{-s}$ and~$1 - \varepsilon_l p_l^{-s}$
is $s=0$ (and that occurs only if $\varepsilon_j = \varepsilon_l = 1$).
Thus each factor $1 - \varepsilon_j p_j^{-s}$ of $P(s)$ has a zero that is not a zero of any other factor.
Since $P(s)$ has no zeros or poles this implies that each $u_j = 0$.
So $P(s) = 1$ as desired.
\end{proof}
\begin{corollary}
Aside from a possible multiplication by a root of unity,
the definition of $L(\chi^j, s)$ is independent of the solution to
(\ref{E24}).
\end{corollary}
\begin{proof}
Let $\left\{ x_i^{\sigma}\right\}$ and $\left\{ \tilde x_i^{\sigma} \right\}$ be two solutions to (\ref{E24}), where we use $j$ for $k$ in (\ref{E24}).
Fix a positive integer $d$ such that each~$ x_i^\sigma d$ and $ \tilde x_i^\sigma d$ is in~${\mathbb{Z}}$ and so $L(\chi^j, s)^d$ is meromorphic
whether we use the $x_i^{\sigma}$ or the $\tilde x_i^{\sigma}$ to define $L(\chi^j, s)$.
Consider $L_1(s), L_2(s)$ be equal to $L(\chi^j, s)^d$ according to the two expressions
given by $x_i^{\sigma}$ and $\tilde x_i^{\sigma}$ respectively.
Note that $L_1(s)$ and $L_2(s)$ agree up to a finite number of Euler factor, and the Euler factors
where they differ are powers of terms of the form
$$
\frac{1}{1 - \varepsilon N(\mf p)^{-s}}
$$
for some prime ideal $\mf p$ in some number field and some root of unity $\varepsilon$.
So $N(\mf p) = p^l$ for some prime $p\in {\mathbb{Z}}$. By factoring the polynomial $1 - \varepsilon X^l$ into linear factors,
we get the following:
$$
\frac{1}{1 - \varepsilon N(\mf p)^s} = \prod_{\mu =1}^l \frac{1}{1 - \varepsilon_\mu p^{-s}}
$$
where each $\varepsilon_\mu$ is a root of unity.
With these ideas we can verify that $L_1$ and $L_2$ satisfy the requirements of the above lemma. So $L_1 = L_2$.
This implies that the two definitions of $L(\chi^j, s)$ differ only by a $d$th root of unity factor.
\end{proof}
\begin{remark}
Note that the above lemma also implies that (\ref{E23}) is an exact equation (up to root of unity), a fact that Artin uses in the calculation of $l_i^{(1)}$.
\end{remark}
}
}
To determine $l_i^{(1)}$ explicitly in the functional equation we use (\ref{E23}) and examine the exponent of $\Gamma(s)$
appearing in the functional equation. This is carried out in the following lemma. In the Abelian case
the exponent of $\Gamma$ is $[k: {\mathbb{Q}}]$, and the following lemma shows how to generalize this to the non-Abelian case.
\begin{lemma}
The exponent of $\Gamma(s)$ appearing in the functional equation of $L(s, \chi^i)$ is equal to $f_i [k:{\mathbb{Q}}]$ where $f_i$ is the degree
of the representation associated to $\chi^i$. In other words $f_i = \chi^i(1)$.
\end{lemma}
\begin{proof}
By equation (\ref{E23}) we see that the exponent of $\Gamma(s)$ appearing in
the functional equation of $L\bigl(s, \psi_i^{(\sigma)} \bigr)$ is equal to
$$
\sum_{\nu = 1}^x r_{i \nu}^{(\sigma)} l_\nu^{(1)}
$$
but from the functional equation for Abelian $L$-series we know that this exponent should be the degree $[\Omega_\sigma: {\mathbb{Q}}]$. Thus,
for each $\sigma \in G$ and $i = 1, \ldots, m(\sigma)$,
$$
\sum_{\nu = 1}^x r_{i \nu}^{(\sigma)} l_\nu^{(1)} = [\Omega_\sigma: {\mathbb{Q}}] = [k: {\mathbb{Q}}] \frac{|G|}{m(\sigma)}.
$$
Multiply by $\psi_i^{(\sigma)} (\tau)$ with $\tau \in \mf g^\sigma$, and sum over $i$:
$$
\sum_{i=1}^{m(\sigma)} \sum_{\nu = 1}^x l_\nu^{(1)} r_{i \nu}^{(\sigma)} \psi_i^{(\sigma)} (\tau)
= [k: {\mathbb{Q}}] \frac{|G|}{m(\sigma)} \sum_{i=1}^{m(\sigma)} \psi_i^{(\sigma)} (\tau).
$$
Using (\ref{E6}) on the left and (\ref{E3}) on the right, this equation simplifies as
$$
\sum_{\nu = 1}^x l_\nu^{(1)}\chi^{\nu} (\tau)
= [k: {\mathbb{Q}}] |G| \, \varepsilon_{\tau}
$$
where $\varepsilon_\tau$ is $1$ or $0$ depending on whether $\tau = 1$ or $\tau \ne 1$.
The above equation is independent of $\sigma$, and so applies to all $\tau \in G$.
Now multiply by $\chi^i(\tau^{-1})$ and sum over $\tau \in G$:
$$
\sum_{\tau \in G} \sum_{\nu = 1}^x l_\nu^{(1)} \chi^{\nu} (\tau) \chi^i(\tau^{-1})
\! = \sum_{\tau \in G} [k: {\mathbb{Q}}] |G| \, \varepsilon_{\tau}
\chi^i(\tau^{-1}) = [k: {\mathbb{Q}}] |G| \, \chi^i(1) = [k: {\mathbb{Q}}] |G| f_i
$$
but the left-hand simplifies by (\ref{E2}) to give $|G| \, l_i^{(1)}$. So $l_i^{(1)}|G| = f_i [k: {\mathbb{Q}}] |G|$.
In other words, $l_i^{(1)} = f_i [k: {\mathbb{Q}}]$.
\end{proof}
We can determine some of the other constants in the functional equation in a similar manner.
{\color{blue}
\begin{remark}
In the above proof Artin regards equation (\ref{E23}) not as an equation valid up to a finite number of Euler factors but as an exact
equation (or at least up to multiplication by a root of unity). This is justified based because both
sides of (\ref{E23}) satisfy the right type of functional equations (see Lemma~\ref{fixingL_lemma}).
Here is another proof of the above lemma that might be of interest; it does not use the strong version of (\ref{E23}).
We begin with a special case of~(\ref{E7}):
\begin{equation*}
\chi_{\psi^{(\sigma)}_i} (1) = \sum_{\nu =1}^x r_{i \nu}^{(\sigma)} \chi^{\nu} (1) = \sum_{\nu =1}^x r^{(\sigma)}_{i \nu} f_\nu
\end{equation*}
for each $\sigma$ and each $i = 1, \ldots, m(\sigma)$.
But the degree of the induced representation is just the index $[G: \mf g^{\sigma}]$ so
$$
\sum_{\nu =1}^x r^{(\sigma)}_{i \nu} f_\nu = \chi_{\psi^{(\sigma)}_i} (1) = [G: \mf g^{\sigma}] = \frac{|G|}{m(\sigma)}.
$$
In what follows let $x_i^{\sigma}$ a solution to (\ref{E24}), (where we use $j$ for $k$ in that equation). Then by
the previous equation and (\ref{E24})
\begin{eqnarray*}
\sum_{\sigma\ne 1} \sum_{i=2}^{m(\sigma)} x_i^\sigma \frac{|G|}{m(\sigma)}
&=&
\sum_{\sigma\ne 1} \sum_{i=2}^{m(\sigma)} \sum_{\nu =1}^x x_i^\sigma r^{(\sigma)}_{i \nu} f_\nu\\
&=&
\sum_{\nu =1}^x \left( \sum_{\sigma\ne 1} \sum_{i=2}^{m(\sigma)} x_i^\sigma r^{(\sigma)}_{i \nu} \right) f_\nu\\
&=& \sum_{\nu =1}^x \delta_{j\nu} f_\nu\\
&=& f_j.
\end{eqnarray*}
With this identity we can easily calculate the exponent of the expression
$$ \left( \frac{2}{(2\pi)^s} \right) \Gamma(s)$$
in the functional equation for $L(s, \chi^j)$.\footnote{\color{blue} As usual, if it makes matters clearer take a power $L(s, \chi_i)^d$
that is meromorphic on ${\mathbb{C}}$ instead of dealing with branches. It is clear how to adapt this argument to such a power.}
From Hecke's functional equation for Abelian $L$-series we have that the contribution from each $L\bigl(s, \psi_i^{(\sigma)} \bigr)$ is equal to
$$
[\Omega_{\sigma}: {\mathbb{Q}}] = [k: {\mathbb{Q}}] \; \frac{|G|}{m(\sigma)},
$$
and so the total contribution for $L(s, \chi^j)$ is
$$
\sum_{\sigma\ne 1} \sum_{i=2}^{m(\sigma)} {x_i^\sigma } [\Omega_{\sigma}: {\mathbb{Q}}]
=[k:{\mathbb{Q}}] \sum_{\sigma\ne 1} \sum_{i=2}^{m(\sigma)} x_i^\sigma \frac{|G|}{m(\sigma)} = [k:{\mathbb{Q}}] f_j.
$$
We can argue similarly for the part of the function equation of $L(s, \chi^j)$ coming from factors of the type~$(N(\mf f) |\Delta_k|)^{s-1/2}$
from the Abelian $L$-series factors.
For each~$L\big(s, \psi_i^{(\sigma)} \big)$ we can write this factor as $(B_{i}^\sigma |\Delta_\sigma|)^{s-1/2}$
where $|\Delta_{\sigma}|$ is the absolute discriminant of the field $\Omega_\sigma$ and $B_{i}^\sigma$ is a positive integer. But
$$
|\Delta_\sigma| = N_\sigma |\Delta_k|^{[\Omega_\sigma: k]}$$
where $\Delta_k$ is the discriminant of $k$ and $N_\sigma$ is some positive integer.\footnote{\color{blue}
This is a standard result in algebraic number theory. See, for instance, Neukirch~\cite{Neukirch1999}, Corollary 2.10, page 202. }
So we can write
$$
(B_{i}^\sigma |\Delta_\sigma|)^{s-\frac{1}{2}}
= (B_{i}^\sigma N_\sigma)^{s-\frac{1}{2}} \left(|\Delta_k|^{s-\frac{1}{2}}\right)^{[\Omega_\sigma: k]}.
$$
In the functional equation of $L(s, \chi^j)$ the terms $(B_{i}^\sigma N_\sigma)^{s-1/2}$ combine to give
$$
\left( \prod_{\sigma \ne 1} \prod_{i=2}^{m(\sigma)} \left( B_{i, \sigma} N_\sigma \right)^{x_i^\sigma}\right)^{s-\frac{1}{2}}
$$
which in Artin's notation is $\alpha_j^{s-\frac{1}{2}}$. Observe that $\alpha_i$ is the product of rational powers of positive integers.
The exponent of $|\Delta_k|^{s-1/2}$ in the functional equation of $L(s, \chi^j)$ will be given by
$$
\sum_{\sigma\ne 1} \sum_{i=2}^{m(\sigma)} {x_i^\sigma} [\Omega_{\sigma}: k]
= \sum_{\sigma\ne 1} \sum_{i=2}^{m(\sigma)} x_i^\sigma \frac{|G|}{m(\sigma)} = f_j
$$
where we have used the formula for $f_j$ established above. Thus we get the discriminant factor~$\left(|\Delta_k|^{f_j}\right)^{s-\frac{1}{2}}$.
\end{remark}
}
\medskip
All in all, we get the following:
\begin{satz}
The primitive $L$-series $L(s, \chi^i)$ can be analytically continued to the whole plane aside from a possibly discrete
set of branch points. The orders of the branch points are (unformly) bounded.\footnote{\color{blue} In other words, $L(s, \chi^i)$ can be meromorphically continued on some Riemann surface covering~${\mathbb{C}}$ of finite degree $d$. In fact, $L(s, \chi^i)^d$ can be meromorphically continued on ${\mathbb{C}}$ itself for some positive power $d$.} For $i>1$
the continuation of (each branch of)~$L(s, \chi^i)$ is holomorphic and nonzero in a neighborhood of $s=1$.
There are zero-free neighborhoods of the line~\text{$\Re (s) = 1$}, including a region on the plane
defined by $\sigma \ge 1 - c/\log t$ for some constant $c>0$ (here we write a complex number as $s = \sigma + i t$ with $\sigma, t \in {\mathbb{R}}$).
These \text{$L$-functions} satisfy a functional equation of the form:
\begin{equation} \label{E26}
\frac{L (1-s, \overline \chi^i)}{L(s, \chi^i)}
=
\varepsilon_i
\left( \frac{2}{(2\pi)^s}\right)^{mf_i}
\left( \alpha_i |\Delta_k|^{f_i}\right)^{s-\frac{1}{2}}
\left( \cos \frac{s\pi}{2}\right)^{l_i^{(2)}}
\left( \sin \frac{s\pi}{2}\right)^{l_i^{(3)}}
\left( \Gamma(s)\right)^{m f_i}.
\end{equation}
where $\Delta_k$ is the discriminant of $k$, $\alpha_i$ is a product of rational powers of (rational) positive integers, $\varepsilon_i$
are algebraic integers that depends only on the branch under consideration with $|\varepsilon_i| = 1$, $m = [k:{\mathbb{Q}}]$, and $f_i = \chi^i(1)$ is the
degree of the representation associated to the character $\chi^i$.
Furthermore, $l_i^{(2)}$ and $l_i^{(3)}$ are rational numbers.
\end{satz}
With these types of methods (``Auf demselben Wege'') it should also be possible to establish the single-valuedness of our functions,
of which one can easily convince oneself in special cases. At least one can prove that the branching orders are divisible only by primes dividing $|G|$.
Completely new methods will probably be needed to show that our $L$-Series are analytic on all of ${\mathbb{C}}$ (aside from the $L$-series associated
to the trivial character (Hauptcharakter)).
{\color{blue}
\begin{remark}
As mentioned above, the methods of this paper show that $L(s, \chi^i)^d$ is entire for some positive integer $d$. In other words, $L(s, \chi^i)$
can be regarded as a $d$-valued function. Artin mentions here that it should be possible to prove that~$d = 1$ is possible (``die Eindeutigkeit unseren Funktionen''), in other words that $L(s, \chi^i)$
is meromorphic on the whole plane.
His next sentence means that we should at least be able to find a $d$ such that the only primes dividing $d$ are divisors of $|G|$.
The former claim would have to wait until 1947 when it was proved by R. Brauer~\cite{Brauer1947a},
but later claim is, as Artin says, fairly easy to show: see the following remark.
It is still an open problem however on whether the Artin $L$-series is analytic in general. It has been shown in some cases by Langlands and Tunnell, and
these cases were used by Wiles in his proof of Fermat's Last Theorem.
\end{remark}
\begin{remark}
We now outline an argument for Artin's claim on branching orders for primes $p$ not dividing the order~$|G|$.
Recall that as part of the proof of the solvability of~(\ref{E24}), Artin shows that
the matrix~$(r_{i \nu}^{(\sigma)})$ has linearly independent rows. We can reduce this matrix modulo $p$,
and by working in a suitable extension ${\mathbb{F}}_q$ of ${\mathbb{F}}_p$ (containing roots of unity of order $|G|$) we can mimic the proof
given above for ${\mathbb{Q}}$ and show that it also works over ${\mathbb{F}}_p$ as long as $|G|$ is
not zero modulo $p$.
Once we know that the matrix $(r_{i \nu}^{(\sigma)})$ has linearly independent rows modulo~$p$, we can find an $x-1$ by $x-1$ submatrix whose mod $p$ reduction is
nonsingular. In other words, we can find a $x-1$ by $x-1$ submatrix whose determinant is an integer not divisible by $p$.
We can then
find a solution to (\ref{E23}) in terms of rational numbers whose denominators are not divisible by $p$. This gives a $d_p$ sheeted cover of ${\mathbb{C}}$
such that $L(s, \chi^i)$ is meromorphic on the cover.
In particular, if one goes around a branch point of $L(s, \chi^i)$ then the value will change value by a multiplicative factor that is a $d_p$-root of unity. In other words, the order of the branch is relatively prime to $p$.
This applies to all primes not dividing~$|G|$ as one goes around a branch point. Let $d$ be the GCD
of all the~$d_p$. Going around any branch point changes the value by a $d$-th root of unity, so
$L(s, \psi^i)^d$ descends to a meromorphic function on ${\mathbb{C}}$, and at the same time the only primes dividing $d$
are primes dividing $|G|$.
\end{remark}
}
\chapter{ Conjecture of Frobenius (now called the Chebotaryov Density Theorem)}
With the the result just derived one can easily confirm a conjecture of Frobenius using Formula (\ref{E12}).\footnote{See
\S 5, Formulas (16) and (18)
of the 1896 work of Frobenius cited in footnote~\ref{FrobFoot}. }
\color{blue}
\begin{remark}
This density conjecture of Frobenius that Artin proves here is what we today call the \emph{Chebotaryov (or Chebotarev) density theorem}.\footnote{\color{blue}
Nikolai Chebotaryov (1894--1947) was a mathematician from Ukraine and Russia. The spelling ``Chebotaryov'' is a transliteration of the Ukrainian version of his name, while
``Chebotarev'' is a transliteration of the Russian version. He was born in Ukraine and was educated at Kyiv University. He later became a professor at Kazan University in Russia in 1928
where he spent the remainder of his career.}
Unbeknownst to Artin, Nikolai Chebotaryov had already proved this result about a year
earlier in~1922
without using these new $L$-series. Artin gives a proof here, but it is requires Artin's reciprocity (Satz~2) in order
to be assured that Satz~3 holds. Satz 2 was not fully proved until 1927 when Artin proved his reciprocity law. It is interesting to note that Artin's 1927 proof
of his reciprocity law was inspired by the 1925 German versions of Chebotaryov's proof of this density theorem that Artin read only after he completed the current paper.
\end{remark}
\color{black}
Not only can you derive the conjecture results, but you can also sharpen them without effort.
From formula (\ref{E12})
for $\log L(s, \chi)$
it follows from known methods that
\begin{equation}\label{E27}
\sum_{N \mf p \le x}
\chi^i (\mf p)
=
\delta_{1 i} \mathrm{Li}(x) + O\left( x e^{-a\sqrt{\log x}}\right),
\end{equation}
where $\delta_{11} = 1$, but otherwise $\delta_{1i}=0$.
{
\color{blue}
\begin{remark}
By ``known methods'', Artin is presumably referring to a combination of methods used to prove Dirichlet's theorem
together with those needed for the the prime number theorem generalized to number fields.
As usual, the error term can be greatly improved
if one assumes the generalized Riemann hypothesis.
For example, a classical form of the prime number theorem is that
\begin{equation*}
\pi(x)
=
\mathrm{Li}(x) + O\left( x e^{-a\sqrt{\log x}}\right),
\end{equation*}
for some $a>0$. (See for instance Theorem 6.9, page 179 of \cite{ MV2007}).
Here $\pi(x)$ is the number of primes in ${\mathbb{Z}}$ less than $x$
and
$$
\mathrm{Li}(x) = \int_{2}^x \frac{1}{\log t} dt.
$$
The proof of the prime number theorem uses a zero-free region for $\zeta(s)$ similar to that described in Satz~3 for Artin $L$-functions.
(See Theorem~6.6, page 172 of \cite{ MV2007} for the classical zero-free region).
\end{remark}
}
For a real number $x$ and a conjugacy class $C$ of $G$,
let~$\pi(x, C)$ be the number of prime ideals $\mf p$ of $k$ with $N\mf p \le x$
whose Frobenius class is $C$.
We multiply (\ref{E27}) by $\chi^i(\sigma^{-1})$ where $\sigma \in C$, and sum over $i$. From (\ref{E3})
we get
$$
\frac{|G|}{|C_r|}\, \pi(x, C_r) = \sum_{i=1}^x \sum_{N \mf p \le x} \chi^i (\sigma^{-1})
\chi^i (\mf p) = \mathrm{Li}(x) + O\left( x e^{-a\sqrt{\log x}}\right).
$$
\begin{satz}
For a real number $x$ and a conjugacy class $C$ of $G$,
let~$\pi(x, C)$ be the number of prime ideals $\mf p$ of $k$ with $N\mf p \le x$
whose Frobenius class is $C$. Then
\begin{equation} \label{E28}
\pi(x, C) = \frac{|C|}{|G|} \mathrm{Li}(x) + O\left( x e^{-a\sqrt{\log x}}\right).
\end{equation}
So the density of prime ideals in the class $C$ is equal to the density of $C$ in $G$.
In particular, in each class $C$ there is an infinite number of prime ideals whose Frobenius class is $C$.
\end{satz}
This theorem is a generalization of Dirichlet's theorem concerning primes in an arithmetic progression,
which (with the help of our general reciprocity law) can be seen to be a special case.\footnote{\color{blue}Although
Satz 2 is not fully proved
in this paper, it is proved in special cases including that of ${\mathbb{Q}}(\zeta) / {\mathbb{Q}}$. When we work out the class field theory for ${\mathbb{Q}}(\zeta)/{\mathbb{Q}}$, we find that Artin's reciporocity gives a correspondence between
the set of primes of ${\mathbb{Q}}$ with a fixed Frobenius in the Galois group of~${\mathbb{Q}}(\zeta)/{\mathbb{Q}}$ and the set of primes of ${\mathbb{Q}}$
in a certain arithmetic progression. So Satz 5 applied to the fields ${\mathbb{Q}}(\zeta) / {\mathbb{Q}}$ is really just Dirichlet's theorem.
}
Its true meaning has yet to be clarified (``Seine wahre Bedeutung harrt noch der Aufkl\"arung'').
\chapter{Multiplicative Relations Between $L$-Series}
\begin{satz}
If the base field $k$ is ${\mathbb{Q}}$ then there are no multiplicative relations between the primitive $L$-Series.
\end{satz}
\begin{proof}
Suppose $x_i$ are integers such that
$$\prod_{i = 1}^x \left( L(s, \chi^i) \right)^{x_i} = 1.$$
Then by (\ref{E12}), with $k={\mathbb{Q}}$,
\begin{equation*}
\log L(s, \chi^i) = \sum_{p^\nu} \frac{\chi^i(p^\nu)}{\nu p^{\nu s} }
\end{equation*}
where the sum is over all prime powers $p^\nu > 1$.
So when we sum over the $\chi^i$ we get
$$
\sum_{p^\nu} \left( \sum_{i=1}^x x_i {\chi^i(p^\nu)}\right) \frac{1}{\nu {p^{\nu s}} } = 0.
$$
{\color{blue}
\begin{remark}
One can tentavely think of the above equality as holding modulo~$2\pi i {\mathbb{Z}}$. But in any case the right hand side is a constant
on the connected set~$\Re(s)>1$. Since the left hand side is a Dirichlet series with constant term $0$ this forces the right hand side to be $0$ as asserted
(by the uniqueness of coefficients of a Dirichlet series).
\end{remark}
}
By the uniqueness of the coefficients of a Dirichlet series we have
$$
\sum_{i=1}^x x_i {\chi^i(p)} = 0
$$
for all primes $p$ (and in fact, the prime power $p^\nu$ coefficients vanish as well).
Recall from (\ref{E10}) that~$\chi^i(p)$ denotes to the value of~$\chi^i$ at the Frobenius of $p$.
By Satz 4, each conjugacy class $C$ of $G$ is the Frobenius class for an infinite number of primes $p$. So
$$
\sum_{i=1}^x x_i {\chi^i(\tau)} = 0
$$
for all $\tau \in G$. This implies, in the usual way, that $x_i = 0$ for all $i$.
\end{proof}
{\color{blue}\begin{remark}
The last step is just due to the linear independence of characters. This can be shown using (\ref{E2}).
\end{remark}
\begin{remark}
We assumed $x_i$ were integers in the above proof since that is the main case under consideration, but we can let $x_i$ be complex and use the above argument
to show that the functions~$\log L(s, \chi^1), \ldots, \log L(s, \chi^x), 1$ are linearly independent over ${\mathbb{C}}$.
\end{remark}
}
Satz 5 is not valid for general algebraic number fields $k$ since conjugate prime ideals can undermine the result.
In fact, one can easily construct examples (even with~$[k: {\mathbb{Q}}]=2$) in which conjugate characters give rise to the same $L$-series.
(In fact, we will see some examples in Section~\ref{S9} where different characters of a given Galois group can give rise to the same $L$-series).
Based on Satz 5
we see how to find all the relationships between any finite collection of $\zeta$-functions or $L$-series.
Find a Galois extension
$E$ of ${\mathbb{Q}}$ that contains all the field extensions $K/k$ used to define the zeta and $L$-series that you are interested in.
We can consider all of our given functions as being defined using characters for~$E/{\mathbb{Q}}$,
and all of these can be expressed in terms of primitive $L$-series for $E/{\mathbb{Q}}$, which are independent by Satz 5.
We can use elimination to find all the relations between our functions because any additional relations are ruled out by Satz 5.
The remark at the end of Section~2 shows we do not necessarily have to transition to a common $E$
to get our decompositions since extending the field
does not change the decomposition. (The common field was mainly used to prove uniqueness).
So we have reached a conclusion to the problem of determining multiplicative relations.
\color{blue}
\begin{remark}
Here Artin to the independence of the common Galois extension $E$. If you wish to break the dependency on a common
Galois $E/{\mathbb{Q}}$, you would want a way to identify when two primitive $L$-series
for $E_1/{\mathbb{Q}}$ and $E_2/{\mathbb{Q}}$ are equal. One way is to agree to classify each primitive $L$-series by the minimal Galois extension $E/{\mathbb{Q}}$
for which it arises. In other words, consider only irreducible \emph{faithful} representations of Galois groups with base field~${\mathbb{Q}}$.
Note that we have independence for the infinite collection of such primitive $L$-series over ${\mathbb{Q}}$ (and the ${\mathbb{C}}$-linear independence
for their repective logarithmic functions). When we combine this section with the results of Section 2 we have the following result:
\end{remark}
\begin{corollary}
Every Artin $L$-series factors uniquely as the product of primitive~$L$-series defined over ${\mathbb{Q}}$.
\end{corollary}
\begin{remark}
Above Artin mentions using elimination to find relations. This essentially means using commonplace matrix manipulations on integral matrices.
We describe this in more detail.
Suppose $\ell_1, \ldots, \ell_m$ gives a collection of $L$-series, and $L_1, \ldots, L_t$ are all the primitive~$L$-series
defined over ${\mathbb{Q}}$ that arises in the decompositions of $\ell_1, \ldots, \ell_m$. Then we can identify each $\ell_i$
with an element of ${\mathbb{Z}}^t$ through the exponents of its decomposition. Consider the ${\mathbb{Z}}$-module homomorphism
$$
\Phi\colon {\mathbb{Z}}^m \to {\mathbb{Z}}^t
$$
sending $(c_1, \ldots, c_m)$ to the element of ${\mathbb{Z}}^t$ associated to $\ell^{c_1}_1 \cdots \ell_m^{c_m}$.
Then the kernel of $\Phi$ is a free ${\mathbb{Z}}$-module of rank bounded by $m$. The elements in this kernel
give us our relations between $\ell_1, \ldots, \ell_m$, and Satz 5 assures us that these are all the relations.
We can concretely calculate a basis for the kernel, i.e. identify all fundamental relations for $\ell_1, \ldots, \ell_m$, by using
row and column operations on the matrix representing~$\Phi$ to identify the kernel.
The matrix representing $\Phi$ can be written down as soon as we have decomposed each $\ell_i$: its $j$th column is the exponents
occurring in the decomposition of $\ell_i$
(assuming we multiply the matrix on the left). For example, using (\ref{E29}) below, the fundamental relations among $\zeta, \zeta_5, \zeta_6, \zeta_{10}, \zeta_{12}, \zeta_{15}, \zeta_{20}, \zeta_{30}, \zeta_{60}$ discussed there can be calculated from calculating the kernel of the following:
$$
\begin{bmatrix}
1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 \\
0 & 0 & 0 & 0 & 1 & 0 & 1 & 1 & 3 \\
0 & 0 & 0 & 0 & 1 & 0 & 1 & 1 & 3 \\
0 & 1 & 0 & 1 & 0 & 1 & 2 & 2 & 4 \\
0 & 0 & 1 & 1 & 1 & 2 & 1 & 3 & 5
\end{bmatrix}.
$$
For example, the column vector $(2, 0, -2, 0, 0, 0, -1, 1, 0)$ is in the kernel and
so gives the relation $\zeta^2 \zeta_6^{-2} \zeta_{20}^{-1} \zeta_{30} = 1$.
\end{remark}
\color{black}
\chapter{Applications to Icosahedral Fields} \label{S9}
Finally we apply these results to icosahedral extensions, the simplest extensions that cannot be obtained through a series of Abelian extensions.
Let $K/k$ be a Galois extension of number fields with Galois group $G$ isomorphic to the icosahedral group.
Observe that Satz 2 holds for intermediate Abelian extensions $K'/k'$. To see this
observe that Abelian groups of the form $H_1/H_2$, where $H_1$ is a subgroup of $G$ and~$H_2$ is a normal subgroup of $H_1$,
have order dividing $60=2^2 \cdot 3 \cdot 5$.
The $p$-power part of such a group is a cyclic group of order dividing $p$ for $p=3, 5$, and
the~2-power part of such a group is either cyclic of order dividing $2$, or is the
(Klein) four groups (``Vierergruppe") since $G$ has no elements of order 4.
(Satz 2 has been proved for Abelian Galois groups that are products of cyclic groups of prime order.)
{\color{blue}
\begin{remark}
The group $G$ of symmetries of the icosahedron is isomorphic to the alternating group $A_5$, which is a simple group of order 60.
The subgroups of $A_5$ include cyclic subgroups of the following orders: 1, 2, 3, 5.
There are also Klein four groups, and dihedral subgroups of order~6 and 10. Finally there are subgroups isomorphic to~$A_4$, and of course $A_5$ itself.
So there are intermediate fields of degree~$1, 5, 6, 10, 12, 15, 20, 30$ and $60$ over $k$.
Note that two subgroups $A_5$ of the same order are actually conjugate, and so are isomorphic.
This implies that two intermediate subfields of $K/k$ of the same degree over $k$ must be isomorphic, and so have equal zeta functions.
Artin uses the notation $\Omega_n$ for a field of degree $n$ over~$k$, and~$\zeta_n$ for its zeta function. We let $\zeta$ be the zeta function of
the base field~$k$, so~$\zeta = \zeta_1$.
\end{remark}}
In $G$ we have 5 conjugacy classes $C_1, C_2, C_3, C_4, C_5$ with $1, 15, 20, 12, 12$ elements respectively. The densities of prime ideals in these classes must be
$$
\frac{1}{60}, \frac{1}{4}, \frac{1}{3}, \frac{1}{5}, \frac{1}{5}.
$$
Furthermore, by the theory of characters developed by Frobenius, we have five simple characters of $G$ and their degrees are $1, 3, 3, 4, 5$. We call the associated
primitive $L$-series~$\zeta$, $L_3^{(1)}, L_3^{(2)}, L_4, L_5$.
We easily get the following factorizations using our methods (where the index refers to the degree of the field over $k$):
\begin{eqnarray} \label{E29}
\zeta_5 & = & \zeta \; L_4 \\
\zeta_6 & = & \zeta\; L_5 \nonumber\\
\zeta_{10 }& = & \zeta\; L_4 L_5 \nonumber\\
\zeta_{12} & = & \zeta \;L_3^{(1)} L_3^{(2)} L_5 \nonumber\\
\zeta_{15} & = & \zeta\; L_4 \, (L_5)^2 \nonumber\\
\zeta_{20} & = & \zeta \; L_3^{(1)} L_3^{(2)} \left( L_4\right)^2 L_5\nonumber\\
\zeta_{30} & = & \zeta \; L_3^{(1)} L_3^{(2)} \left( L_4\right)^2 \left( L_5 \right)^3 \nonumber\\
\zeta_{60} & = & \zeta \; ( L_3^{(1)} L_3^{(2)} )^3 \left( L_4\right)^4 \left( L_5 \right)^5 \nonumber
\end{eqnarray}
\bigskip
{\color{blue}
\begin{remark}
Verifying (\ref{E29}) is an exercise. One way to verify it is to do the following:
\begin{itemize}
\item
Identify all subgroups of $A_5$, and the size of the intersections with each of the conjugacy classes $C_1, \ldots, C_5$.
\item
Derive explicit formulas for the five simple characters of $A_5$.
\item
Use Frobenius reciprocity to calculate the induced characters of trivial characters in terms of the irreducible characters of $A_5$.
\item
Use (\ref{E14}) and Satz 1.
\end{itemize}
\end{remark}}
\begin{proposition} \label{prop_4}
If $G = \mathrm{Gal}(K/k)$ is the icosahedral group, then all the $L$-series associated to representations of $G$ are meromorphic. In other words,
they are single valued (outside of poles) when extended to~${\mathbb{C}}$.
\end{proposition}
\begin{proof}
It is enough to verify this for irreducible representatives. The function $\zeta$ and each $\zeta_n$ are already known to be meromorphic (Hecke).
Note that the first two equations of~(\ref{E29}) show that $L_4$ and $L_5$ are meromorphic.
Observe that $K$ is cyclic of degree 5 over an intermediate field $\Omega_{12}$ of degree 12 over $k$. There are four
primitive $L$-series (in addition to $\zeta_{12}$) associated to $K/\Omega_{12}$ and these can all be expressed in terms of our primitive $L$-series.
They are also known to be entire (Hecke). Note that $\zeta_{60}$ factors as $\zeta_{12}$ times the product of these four $L$-series.
Comparing the expressions for $\zeta_{12}$ and $\zeta_{60}$
in (\ref{E29}), we see that the product of these four $L$-series for $K/\Omega_{12}$ is
$$
(L_3^{(1)} L_3^{(2)} )^2 \left( L_4\right)^4 \left( L_5 \right)^4.
$$
Each of these $L$-series is based on a one-dimensional representative of $\mathrm{Gal}(K/\Omega_{12})$, and so can be
expressed in terms of a degree $12$ induced representation of~$G$.
So each decomposition gives 12 as the sum in terms of the integers 3, 4, 5.
We conclude that $12$ is decomposed as $3 + 4 + 5$ for each (note that $5$ cannot occur twice in any one of the sums, so $5$ must occur exactly once
in each sum).
We conclude that two of these $L$-series factor as $L_3^{(1)} L_4 L_5$ and the other two as $L_3^{(2)} L_4 L_5$. (By the way, this gives an example of a field with identical primitive $L$-series where conjugate characters give the same $L$-series)\footnote{\color{blue}Note that any five cycle and its inverse
are in the same conjugacy class of $A_4$, which by Frobenius reciprocity implies that $L$-series for conjugate characters for $K/\Omega_{12}$ will
have the same decomposition.}. So $L_3^{(1)}$ and $L_3^{(2)}$ are meromorphic.
\end{proof}
The above proof also shows the following:
\begin{proposition}
The functions $L_3^{(1)} L_4 L_5$ and $L_3^{(2)} L_4 L_5$ are entire.
\end{proposition}
We can identify other entire function:
\begin{proposition}
The function $L_5$ is entire.
\end{proposition}
\begin{proof}
We have an intermediate extension $\Omega_5/k$ of degree $5$, and an intermediate extension $\Omega_{15}/k$ of degree 15
such that $\Omega_{15}$ is a cyclic cubic extension of $\Omega_5$. This gives us two nontrivial degree 1 characters of $\mathrm{Gal}(K/\Omega_{5})$
of order 3,
whose induced characters are degree 5 characters of $G$. The associated $L$-series are entire since they are Abelian $L$-series with nontrivial characters.
The product of these series is~$\zeta_{15}/\zeta_5 = (L_5)^2$.
Since the associated induced characters of $G$ are of degree 5, both $L$-series must be equal to $L_5$. So $L_5$ is entire.
(And this gives another example where distinct representations gives the same $L$-series).\end{proof}
\begin{proposition}
The products $L_3^{(1)} L_3^{(2)}$ and $L_3^{(1)} L_3^{(2)} L_4$ are entire.
\end{proposition}
\begin{proof}
The proof is similar to the last proof. The first comes from using the quadratic extension~$\Omega_{12}/\Omega_6$ where one produces an entire
$L$-series equal to $\zeta_{12}/\zeta_{6}$.
The second comes from using the quadratic extension~$\Omega_{20}/\Omega_{10}$ where one produces an entire
$L$-series equal to $\zeta_{20}/\zeta_{10}$.
\end{proof}
\begin{remark}
We do not get anything essentially new from the other Abelian extensions.
So $L_5$ and the combinations $L_3^{(1)} L_4 L_5, L_3^{(2)} L_4 L_5, L_3^{(1)} L_3^{(3)}$, and $L_3^{(1)} L_3^{(2)} L_4$
(and their products) are the only $L$-series we can prove are entire.
\end{remark}
{\color{blue}
\begin{remark}
Let's look at the other Abelian subextensions in addition to those treated in the above two Propositions:
\begin{itemize}
\item
$\Omega_{30}/\Omega_{15}$ gives the entire function $\zeta_{30} / \zeta_{15} = L_3^{(1)} L_3^{(2)} L_4 L_5$.
\item
$K/\Omega_{30}$ gives the entire function $\zeta_{60} / \zeta_{30} = \left( L_3^{(1)} L_3^{(2)} L_4\right)^2 \left( L_5 \right)^2$.
\item
$K/\Omega_{20}$ gives the entire function $\zeta_{60} / \zeta_{20} = \left( L_3^{(1)} L_3^{(2)} L_4\right)^2 \left( L_5 \right)^4$
which must factor into two Abelian (and so entire) $L$-functions
corresponding to the two nontrivial characters $\psi$ and $\psi^{-1}$ of the corresponding cyclic Galois group $H_3$ of order~$3$.
It turns out that the $L$-series associated to $\psi$ and $\psi^{-1}$ are equal and so are both $L_3^{(1)} L_3^{(2)} L_4 \left( L_5 \right)^2$.
To see this, note that the decomposition depends on the the multiplicities of the simple characters $\chi^i$ in the corresponding induced representations, and
these can be calculated using Frobenius reciprocity: the multiplicities are respectively
$$
\left< \psi, \mathrm{res}\, \chi^i\right>_{H_3}, \quad\text{and}\quad
\left< \psi^{-1}, \mathrm{res}\, \chi^i\right>_{H_3}
$$
where in these inner products we restrict $\chi^i$ to $H_3$. However, a three cycle in~$A_5$ and its inverse are in the same conjugate class of $A_5$ and so have the same value under $\chi^i$, which means that these two inner products
are actually given by the same sum. This shows the multiplicities are the same.
\item
$K/\Omega_{15}$ gives the entire function $\zeta_{60} / \zeta_{15} = ( L_3^{(1)} L_3^{(2)} )^3 \left( L_4\right)^3 \left( L_5 \right)^3$
that factors into three entire functions coming from $\Omega_{30}/\Omega_{15}$ extensions.
Looking at the earlier case $\Omega_{30}/\Omega_{15}$, we see these three functions are each~$L_3^{(1)} L_3^{(2)} L_4 L_5$.
\item $K/\Omega_{12}$ was treated above (Proposition~\ref{prop_4}).
\end{itemize}
\end{remark}
Observe that these entire functions are all just products of the entire products already considered; nothing new.
}
\begin{remark}
Of the zeta functions from (\ref{E29}), we see that the following are divisible by~$\zeta$ with entire quotient: $\zeta_{6}, \zeta_{12}, \zeta_{30}, \zeta_{60}$.
(This leaves the other half in question, namely $\zeta_{5}, \zeta_{10}, \zeta_{15}, \zeta_{20}$).
On can also verify immediately the relations between zeta functions from my earlier article~\cite{Artin1923}.
\end{remark}
{\color{blue}
\begin{remark}
This shows Artin's interest in the following question: if $K/k$ then is $\zeta_K/\zeta_k$ entire?
This helps motivate Artin's conjecture that primitive $L$-series that are not zeta functions are entire.
Artin's earlier article~\cite{Artin1923}, published in 1923, has some interesting relationships between these zeta functions
in the current case of $G$ isomorphic to $A_5$, the icosahedral group. Some of these include
$$\zeta_{20}\, \zeta^2 = \zeta_5^2 \, \zeta_{12}, \qquad\text{and} \quad
\zeta_{30}\, \zeta^2 = \zeta_{6}^2 \, \zeta_{20}.
$$
These are immediate given (\ref{E29}) above.
\end{remark}
}
\textbf{Hamburg, Mathematics Seminar, July 1923}
\bibliographystyle{plain}
|
\section{Introduction}
The strong CP problem~\cite{Peccei:1977hh,Peccei:1977ur,Weinberg:1977ma,Wilczek:1977pj} in the Standard Model (SM) can be solved by introducing a global $U(1)_{\text{PQ}}$ symmetry which was spontaneously broken down by a dynamical CP-conserving axion field.
The corresponding pseudo-Nambu-Goldstone-boson of the broken symmetry is called the QCD axion, which in addition serves as a dark matter candidate~\cite{Preskill:1982cy,Abbott:1982af,Dine:1982ah}.
The breaking scale of the new symmetry should be high: $f_a\gtrsim 10^9$ GeV~\cite{Feng:1997tn}, demanding tiny masses of the QCD axion and their couplings to the SM particles, since the latter two are inversely proportional to $f_a$.
This results in a very long lifetime of the QCD axions.
A closely-related hypothetical particle is known as axion-like particle (ALP), which, like the QCD axion, is also a pseudoscalar boson.
Unlike the QCD axion, the ALP mass is not linearly proportional to the couplings to the SM particles, and the ALP hence does not necessarily fix the strong CP problem.
However, the ALP remains one of the possible dark matter candidates, and its mass could possibly range across more than 20 orders of magnitude~\cite{Kim:2015yna,DeMartino:2017qsa,Rubakov:1997vp}.
Further, such ALPs appear in various theoretical models beyond the SM~\cite{Svrcek:2006yi,Gelmini:1980re,Wilczek:1982rv}.
In general, the ALPs can couple to photons, leptons, quarks, as well as gauge bosons at either tree level or loop level.
The phenomenology with only ALP-photon interaction $g_{a\gamma\gamma}$ has been vastly investigated (see Ref.~\cite{Agrawal:2021dbo} and the references therein).
In particular, for sub-GeV ALPs, PRIMEX~\cite{Aloni:2019ruo} and Belle II~\cite{Belle-II:2020jti} provide the most stringent upper bounds on $|g_{a\gamma\gamma}|$, and ALPs of mass $m_a\lesssim 30$ MeV~\cite{Agrawal:2021dbo,Buen-Abad:2021fwq} are disfavored by beam-dump experiments.
However, for ALP-muon interactions, only BaBar~\cite{BaBar:2016sci} gives constraints, for ALPs heavier than twice the muon mass~\cite{Buen-Abad:2021fwq}.
As far as we know, muonphilic ALPs lighter than twice the muon mass have not been directly constrained.
Therefore, we choose to focus on this scenario in the present work.
When cosmic rays reach the Earth's atmosphere, large atmospheric air showers are produced including copious production of pseudoscalar mesons.
Such mesons can decay to light long-lived particles (LLPs) (see Refs.~\cite{Curtin:2018mvb,Lee:2018pag,Alimena:2019zri} for reviews on LLPs), which travel macroscopic distances before decaying potentially in the large-volume neutrino experiments at the Earth surface.
This allows to probe various models predicting such LLPs including heavy neutral leptons~\cite{Kusenko:2004qc,Asaka:2012hc,Masip:2014xna,Arguelles:2019ziu,Coloma:2019htx,Meighen-Berger:2020eun}, the lightest neutralinos in the R-parity-violating supersymmetry~\cite{Candia:2021bsl}, the hadrophilic light dark matter~\cite{Arguelles:2022fqq}, and milli-charged particles~\cite{ArguellesDelgado:2021lek}.
Similarly, the muonphilic ALPs can be abundantly produced via
charged-meson decays from the atmosphere air showers.
Such ALPs should be long-lived, because both they are very light and their decay channels are radiatively suppressed if their mass is below twice the muon mass.
After traveling tens of kilometers across the atmosphere, these ALPs may subsequently decay into two photons inside the detectors of neutrino experiments, such as Super-Kamiokande (SK).
With the tool MCEq~\cite{Fedynitch:2015zma}, we numerically compute the ALPs' flux from the air showers including the propagation through dense medium.
We then estimate the signal event rates at the SK detector, which is sensitive to events of energy below $\mathcal{O}(100)~{\rm GeV}$~\cite{Super-Kamiokande:2017yvm}.
After discussing the background events, we obtain SK bounds on both physical observables and model parameters.
This article is organized as follows.
In Sec.~\ref{sec:model}, we introduce the theoretical scenario we investigate in this work.
The estimation of the ALP flux from the air showers is detailed in Sec.~\ref{sec:alp_flux}, followed by Sec.~\ref{sec:alp_detection} and Sec.~\ref{sec:SuperK} explaining the ALP detection on the Earth and introducing the SK experiment, respectively.
The final numerical results are presented and discussed in Sec.~\ref{sec:results}.
At the end, Sec.~\ref{sec:conclusions} provides a summary and outlook of this work.
\section{ALP-muon interaction}\label{sec:model}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.48\textwidth]{feynman_diagram_production.pdf}\hfill
\includegraphics[width=0.48\textwidth]{feynman_diagram_decay.pdf}
\end{center}
\caption{\textit{Left panel}: Feynman diagram for production of the muonphilic ALP from the charged-pion decay.
The decay vertex factor $g_{\pi \mu\nu}$ denotes the effective
coupling for the charged pion decay $\pi^+ \to \mu^+ \nu_\mu$.
Its conjugated diagram with $\pi^-$ decays is not shown here.
\textit{Right panel}: Feynman diagram for the decay of the muonphilic ALP into a pair of photons, via a triangular muon loop.
}
\label{fig:feynman}
\end{figure*}
In this work, we consider the interaction between the ALP and muon with an effective Lagrangian expressed as
\begin{align}
\mathcal{L} \supset -i g_{a\mu\mu} a \bar{\mu} \gamma_5 \mu\,,
\end{align}
where $g_{a\mu\mu}$ is a dimensionless coupling constant.
For ALP mass $m_a$ larger than twice the muon mass $m_\mu$, the ALP can decay into a pair of muons, while for a lighter ALP only decays into a SM photon pair induced by a triangular muon loop.
The loop-induced interaction between ALP and photons can be described by
\begin{align}
\mathcal{L}_{\rm loop} \supset -\dfrac{1}{4} g_{a\gamma\gamma}^{\rm eff} a F^{\mu\nu} \tilde{F}_{\mu\nu}\,,
\end{align}
with the effective coupling being~\cite{Bauer:2017ris}
\begin{align}
g_{a\gamma\gamma}^{\rm eff} = \dfrac{g_{a\mu\mu} \alpha}{m_\mu \pi } \left[1- \dfrac{4m_\mu^2}{m_a^2} \arcsin^2 \left(\dfrac{m_a}{2m_\mu} \right)\right]\,,
\end{align}
which is valid for $m_a \le 2 m_\mu$.
The lifetime of ALP with $m_a < 2m_\mu$ then reads
\begin{align}
\label{eq:ALP_lifetime}
\tau_a = \Gamma_{a\rightarrow \gamma \gamma}^{-1} = \dfrac{64\pi}{(g_{a\gamma\gamma}^{\rm eff})^{2} m_a^{3}}\,.
\end{align}
With the ALP-muon coupling, ALPs can be produced from charged-meson decays in air shower, dominated by the decay of charged pions $\pi^\pm \rightarrow \mu^\pm \nu a$; therefore, the kinematically allowed ALP mass range is $0 \leq m_a \leq m_\pi - m_\mu$ assuming massless SM neutrinos.
In Fig.~\ref{fig:feynman}, we present two Feynman diagrams for the production and decay of the ALPs, respectively.
The ALP-muon and ALP-photon couplings both contribute to the muon magnetic dipole moment, $a_\mu=(g-2)_\mu/2$~\cite{Buen-Abad:2021fwq}.
The one-loop result of $g_{a\mu\mu}$ leads to negative contributions to $a_\mu$.
However, if we also include the ALP-photon coupling, $g_{a\gamma\gamma}$, the two couplings will induce two-loop light-by-light and Barr-Zee diagrams~\cite{Chun:2015xfx,Marciano:2016yhf}, which, in combination, provide positive contributions to $a_\mu$~\cite{Keung:2021rps}.
On the experimental side, the updated combined results of Fermilab~\cite{Muong-2:2021ojo} and BNL~\cite{Muong-2:2006rrc} measurements indicate a 4.25$\sigma$ positive deviation from the SM theoretical prediction:
\begin{eqnarray}
\Delta a^{\rm BNL}_\mu &=& a^{\rm BNL}_\mu-a^{\rm SM}_\mu = (251\pm 59) \times 10^{-11}\,.
\end{eqnarray}
However, theoretical uncertainties arising from hadronic vacuum polarization may alleviate the tension between these measurements and the SM~\cite{Borsanyi:2020mff}.
Given the large uncertainties within the SM computation, we do not take into account $(g-2)_\mu$ in our analysis.
\section{ALP flux from air shower}\label{sec:alp_flux}
We utilize the numerical code \texttt{MCEq}~\cite{Fedynitch:2015zma} to compute the ALP flux at the Earth surface.
\texttt{MCEq} numerically solves cascade equations of particles propagating in a dense medium; in this work, we use it to study the ALP production throughout the cascade of secondary cosmic rays.
We adopt the parameterization of the cosmic ray flux at the top of atmosphere provided in~\cite{Gaisser:2011klf} and take the hadronic interaction model in~\cite{Riehn:2017mfm}.
The atmosphere is modeled by the CORSIKA parameterizations~\cite{Heck:1998vt}.
To implement the process $\pi^\pm \rightarrow \mu^\pm \nu a$ in \texttt{MCEq}, we compute the corresponding decay matrix
\begin{align}
D^{ij}_{\pi^\pm \rightarrow a} = \Delta T_{\pi^\pm}^i \dfrac{dN_a}{dT_a} (T_{\pi^\pm}^i, T_a^j)\,,
\end{align}
where $T_{\pi^\pm}$ and $T_a$ are the kinetic energy of the pion and
the ALP in the lab frame, with $i,j$ and $\Delta T_{\pi^\pm}^i$ being the kinetic energy bin indices and width, respectively.
The ALP energy spectrum $dN_a/dT_a$ in the lab frame is obtained by applying a Lorentz boost to the energy spectrum in the pion rest frame
\begin{align}
\dfrac{dN_a}{dT_a} = \int \dfrac{d\Omega}{4\pi} \dfrac{dN_a}{dE_a^*} \left| \dfrac{\partial E_a^*}{\partial T_a}\right|\,,
\end{align}
where $E_a^*$ is the energy of the ALP in the pion rest frame and $|\partial E_a^* /\partial T_a |$ is the Jacobian between $E_a^*$ and $T_a$.
After the decay matrix is tabulated, we augment the decay channels of $\pi^\pm$ with $\pi^\pm \rightarrow \mu \nu a$.
We first consider the case that production and the subsequent decay of the ALP into two photons are uncorrelated for simplicity, \textit{i.e.}, the ALP flux at production is proportional to $g_{a\mu\mu}^2$ while the decay is determined by the decay length $c\tau_a$
in the ALP rest frame.
The results are model-independent in the sense that they can be mapped onto another model with similar decay topologies.
Then, we study the case that both production and decay depend on the coupling constant $g_{a\mu\mu}$.
\section{ALP detection on the Earth}\label{sec:alp_detection}
After arriving at the Earth, the ALP can decay into two photons through a muon loop, with a lifetime given in Eq.~\eqref{eq:ALP_lifetime}.
The photons so-produced can then be detected by the Cherenkov detector in neutrino experiments.
Given zenith angle $\theta$, detector geometry, and data-taking time $\Delta t$, the event rate can be calculated by
\begin{align}
\dfrac{d N_{\rm event}}{dT_a} = \epsilon \Delta t A_{\rm eff} \dfrac{d \Phi_a}{dT_a}\,,
\end{align}
where $\epsilon$ is the detection efficiency and we use the output of $\texttt{MCEq}$ for the differential flux $d \Phi_a/dT_a$.
The effective decay area $A_{\rm eff}$, depending on $T_a$, $\theta$ and $c\tau_a$, is given in the appendix of Ref.~\cite{Arguelles:2019ziu}.
The main SM background of such a two-photon signal from the ALP decay
consists of neutral pions decaying into two photons, and neutrino-induced electron-like events that create multiple Cherenkov rings in the electromagnetic shower.
The number of these events have been studied in Ref.~\cite{Super-Kamiokande:2017yvm}, with the best-fit values being 1727 and 797, respectively.
These background events will be taken into account in Sec.~\ref{sec:results} when we estimate the sensitivity reach of SK.
In addition to the signals from ALP decaying into a $\gamma$-pair, ALP can interact with atoms in the detector to create mono-$\gamma$ signal with an energy similar to the energy of ALP, the so-called inverse-Primakoff process.
The cross section of inverse-Primakoff process was studied in details in \cite{Abe:2020mcs}, which can be expressed as
\begin{align}
\sigma_{\rm IP} \simeq \left(\dfrac{g_{a\gamma\gamma}^{\rm eff}}{1\,{\rm GeV}^{-1}}\right)^2 \times 2\,{\rm GeV}^{-2}.
\end{align}
However, since $A_{\rm eff}$ of decay is larger by orders of magnitude than the effective cross section of inverse-Primakoff process $N_T \sigma_{\rm IP}$ with $N_T$ being the total number of target atoms inside the fiducial volume of the detector, we can infer that the event rate from the ALP decay dominates over that from the inverse-Primakoff process; therefore, we will not consider this possibility further in this work.
\section{Super-Kamiokande}
\label{sec:SuperK}
We note that only when the charged pions have a kinetic energy below $\mathcal{O}({\rm TeV})$ and hence a small Lorentz boost, do they essentially all decay well before reaching the Earth surface; therefore, ALP flux at $T_a > \mathcal{O}({\rm TeV})$ is strongly suppressed.
In order to maximize the sensitivity, we focus on the water-based Cherenkov detector of Super-Kamiokande (SK), which has good energy resolution in the sub- and multi-GeV ranges~\cite{Super-Kamiokande:2017yvm}.
Following the analysis in Ref.~\cite{Super-Kamiokande:2017yvm}, the geometry of the SK detector is assumed to be a cylinder with radius $R_{\rm SK} = 20\,{\rm m}$ and height $H_{\rm SK} = 40\,{\rm m}$. The lifetime of SK is taken to be 5326 days with a flat detection efficiency of $0.75$.
Fully-contained events in SK can be grouped into different categories according to the energy and configuration of observed Cherenkov rings.
Since the signal from ALP decay constitutes two electron-like Cherenkov rings, we consider data of $\pi^0$-like two-ring events in 5 energy-bins for sub-GeV $T_a$ and electron-like multi-ring events in 5 $\cos\theta$-bins for multi-GeV $T_a$ provided in Ref.~\cite{Super-Kamiokande:2017yvm}.
\section{Constraints on the parameter space}\label{sec:results}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.48\textwidth]{gamumu_vs_ctau.pdf}\hfill
\includegraphics[width=0.48\textwidth]{mALP_vs_gamumu}
\end{center}
\caption{\textit{Left panel}: sensitivity reach of SK to the muonphilic ALPs for uncorrelated $g_{a\mu\mu}$ and $c\tau_a$ (solid curves), and $c\tau_a$ as a function of $g_{a\mu\mu}$ according to Eq.~\eqref{eq:ALP_lifetime} (dashed lines) in the ($c\tau_a$, $g_{a\mu\mu}$) plane, for three benchmark values of $m_a$.
\textit{Right panel}: constraints on ($m_a$, $g_{a\mu\mu}$) assuming $c\tau_a$ is correlated with $g_{a\mu\mu}$. For comparison, we also include the constraint from BaBar which holds only for larger $m_a$~\cite{BaBar:2016sci}, and the bounds from SN1987A which cover $g_{a\mu\mu} \sim [10^{-10}, 5\times 10^{-3}]$ for sub-GeV $m_a$~\cite{Croon:2020lrf}.
}
\label{fig:SK_reach}
\end{figure*}
We perform $\chi^2$-fit to the SK data mentioned in Sec.~\ref{sec:SuperK} using \cite{Arguelles:2019ziu}
\begin{align}
\chi^2_i = 2 \left\lbrace N_{\rm sig}^i + N_{\rm bkg}^i - N_{\rm obs}^i \left[ 1- \log\left( \dfrac{N_{\rm obs}^i}{N_{\rm sig}^i + N_{\rm bkg}^i}\right) \right] \right\rbrace\,,
\end{align}
where $N_{\rm sig}^i$, $N_{\rm bkg}^i$, and $N_{\rm obs}^i$ are numbers of expected signal events of ALP, background events, and observed events in each data bin, respectively.
The background and observed events are extracted from Ref.~\cite{Super-Kamiokande:2017yvm}.
For a total of 10 bins considered in this work, we derive the 90\% C.L. constraint by requiring $\Delta \chi^2 \equiv \chi^2 - \chi^2_0 \leq 4.865$ with $\chi^2 = \sum_i \chi_i^2$ and $\chi^2_0$ being the case without ALP contribution.
We show the sensitivity reach of SK in Fig.~\ref{fig:SK_reach} for $g_{a\mu\mu}$-$c\tau_a$ uncorrelated (solid lines in the left panel) and correlated (blue area in the right panel) cases.
In the left panel, we observe that for $m_a = 1\,{\rm MeV}$, the best sensitivity of SK lies at $c\tau_a \sim 5\times 10^{-2}\,{\rm km}$, while for larger $m_a$, the best sensitivity of SK is reached at larger $c\tau_a$.
In the same plot we also overlap the uncorrelated sensitivity curves with dashed lines depicting $c\tau_a$ as a function of $g_{a\mu\mu}$ using Eq.~\eqref{eq:ALP_lifetime}.
This allows us to crosscheck with the right panel where we assume both production and decay are mediated by $g_{a\mu\mu}$; the intersection of the solid and dashed lines in the left plot for each fixed mass should coincide with the corresponding parameter point on the outer edge of the blue area in the $g_{a\mu\mu}$ vs.~$m_a$ plane in the right panel.
We find in the right plot that for $m_a = [0.1, 33]$~MeV, SK can exclude $g_{a\mu\mu} = [5\times 10^{-3}, 50]$, comparable to the BaBar exclusion limits, which, however, only hold for larger $m_a$~\cite{BaBar:2016sci}.
Below $g_{a\mu\mu} \sim 5\times 10^{-3}$ and for sub-GeV $m_a$, we note the parameter space is covered by the SN1987A constraint~\cite{Bollig:2020xdr,Croon:2020lrf}.
For each mass value within the sensitive range, the SK exclusion limits are bounded from both top and bottom.
This is because when $g_{a\mu\mu}$ is too small, the production rate of the ALPs is insufficient and the decay length is too long.
On the other hand, with a too large $g_{a\mu\mu}$, despite the enhanced production rate, the decay length is so short that the ALPs decay before reaching the SK detector.
We note that future muon beam-dump experiments can improve the sensitivity at the ALP mass range of interest in this work down to around $g_{a\mu\mu}\sim 6\times 10^{-6}$; see Ref.~\cite{Chen:2017awl} for details.
The results presented here are only based on charged pion flux.
One can in principle extend the constraint to $m_a \geq m_{\pi^\pm} - m_\mu$ by further including heavier mesons such as kaons that have a similar decay channel.
However, once $m_a \geq 2 m_\mu$ is fulfilled, the ALPs can decay into two muons, resulting in further complication during the air shower; we reserve this possibility for future work.
In addition to ALP-muon interaction, we have checked the case that the ALP has a direct coupling to the SM photons.
In this case, the main ALP production channel in air shower is $\pi^0$ decay, with a smaller branching ratio compared to charged pion decay as a result of the much shorter lifetime of $\pi^0$.
Therefore, we infer that given a fixed decay rate of $\pi \rightarrow a$, for $T_a < \mathcal{O}({\rm TeV})$ the ALP flux from $\pi^0$ is less intense than that from $\pi^\pm$, but for higher masses the ALP flux from $\pi^0$ dominates, as $\pi^\pm$'s will not all decay well before reaching the Earth surface.
We find that this excludes $g_{a \gamma \gamma} \gtrsim 10^{-2}\,{\rm GeV}^{-1}$ for $m_a \sim \mathcal{O}(1\text{--}100\,{\rm MeV})$, assuming the decay of ALP is independently determined by $c\tau_a$.
Since $g_{a \gamma \gamma} \gtrsim 10^{-2}\,{\rm GeV}^{-1}$ is already excluded by accelerator and collider constraints (see Ref.~\cite{dEnterria:2021ljz} for a recent summary), we do not demonstrate the result in this work.
In addition, we note that in this scenario the case that the production and the decay are correlated cannot be probed by SK, since tree-level decay results in a much shorter $c \tau_a$ for the given couplings; see also other scenarios with tree-level decays discussed in Ref.~\cite{Arguelles:2019ziu}.
\section{Conclusions}\label{sec:conclusions}
Similar to QCD axions, axion-like-particles (ALPs), denoted as $a$ in this work, are also among the most plausible candidates of dark matter.
With its mass and couplings to the SM particles uncorrelated, such exotic particles are being searched for at various experimental facilities across a wide range of masses.
While the ALPs can in theory couple to various types of particles, here, we have chosen to focus on the case that the ALPs are dominantly or solely interacting with the SM muons at tree level.
In addition, we have restricted ourselves to ALP mass below the muon-pair threshold so that the ALPs only decay radiatively into a pair of photons.
Such ALPs can be produced from charged pion decays, $\pi^\pm\to \mu^\pm \nu a$ via the ALP-muon coupling $g_{a\mu\mu}$.
Large numbers of mesons including charged pions are produced in the atmospheric air showers resulting from cosmic rays.
Once ALPs are produced from these charged-pion decays, if long-lived, they can travel tens of kilometers downwards to the Earth surface thanks to the large Lorentz boost, and decay in large-volume neutrino experiments such as Super-Kamiokande (SK), leading to Cherenkov signal events.
We make use of the numerical tool \texttt{MCEq} in order to estimate the ALP flux at the Earth surface stemming from cosmic-ray-induced atmospheric air showers of the charged pions.
We further compute the signal event rates, taking into account the differential ALP flux, detector efficiency, data-taking time, etc., at SK.
In addition, there are background events mainly stemming from neutral pion decays into two photons as well as neutrino-induced electron-like events that lead to multiple Cherenkov rings in the electromagnetic shower.
We have extracted the level of these background events from Ref.~\cite{Super-Kamiokande:2017yvm}, and evaluated the sensitivity reach of the SK experiments to such muonphilic ALPs.
Results are presented for both production-decay correlated and uncorrelated cases, shown in Fig.~\ref{fig:SK_reach}.
In particular, we find that if both production and decay are mediated by the coupling $g_{a\mu\mu}$, SK can probe $g_{a\mu\mu}$ down to $5\times 10^{-3}$ at $m_a\sim 20$ MeV, complementary to the exclusion limits obtained at BaBar which is sensitive to larger masses, as well as SN1987A constraint which covers $g_{a\mu\mu} \lesssim 5\times 10^{-3}$.
Additionally, we have commented on further possibilities such as the inverse-Primakoff process, and the case that the ALP is coupled to the SM photons at tree level.
While the former is expected to be dominated over by the main process considered in this work, the latter is checked to give only rather weak limits that have already been excluded by past experiments.
We have also briefly discussed the implication of the anomalous muon magnetic moment measurements on our model.
Based on the current sensitivity of Fermilab $(g-2)_\mu$ measurement and once theoretical uncertainties are clarified, {$|g_{a\mu\mu}|\gtrsim 5\times 10^{-4}$ can be probed for correlated $g_{a\gamma\gamma}$-$g_{a\gamma\gamma}$ and $m_a\le 2 m_\mu$}.
However, with this coupling correlation, the ALPs always contribute negatively to $a_\mu$.
Therefore, our model is unable to alleviate the present tension between the SM prediction and observation results.
Before closing, we comment on the sensitivities of other present and future neutrino telescopes to the scenario considered in this work.
For the future Hyper-Kamiokande experiment (HK)~\cite{Abe:2011ts,Hyper-KamiokandeWorkingGroup:2013hcb}, its fiducial volume is increased by a factor 25 compared to SK.
Therefore, we naively estimate that it can improve the sensitivity reach to $g_{a\mu\mu}$ by a factor $\sqrt{5}$ ($5^{1/4}$) for the uncorrelated (correlated) case.
We note that the precise sensitivity reach ultimately depends on the detector configuration of HK. Moreover, if the signal discrimination rate is also enhanced, we expect the sensitivity reach could potentially enclose a large portion of the parameter space currently covered by the SN1987A constraint.
Finally, we note that in principle IceCube~\cite{IceCube:2016zyt} is also capable of probing the atmospheric ALP flux considered in this work. However, since IceCube focuses on the ultrahigh energy range, the best sensitivity lies at $c\tau_a \sim 5\times 10^{-5}\,{\rm km}$ for $m_a \sim 10\,{\rm MeV}$~\cite{Arguelles:2019ziu}; therefore, with a smaller $c\tau_a$ needed, we can expect that the constraint on $g_{a\mu\mu}$ will be much weaker, despite a much larger fiducial volume at IceCube.
\section*{Acknowledgment}
We thank Giovanna Cottin, Felix Kling, and V\'ictor Mu\~{n}oz for useful discussions.
Z.S.W. is supported by the Ministry of Science and Technology (MoST) of Taiwan with grant number MoST-110-2811-M-007-542-MY3. J.L.K. is supported by the U.S. National Science Foundation (NSF) Theoretical Physics Program, Grant PHY-1915005.
P.Y.T. and K.C. are supported in part by MoST with grant numbers MoST-111-2112-M-007-012-MY3 and MoST-110-2112-M-007-017-MY3, respectively.
\newpage
|
\section{Introduction}
\label{S:saruma-intro}
Let $\{Y_{t}\}$ be a time series whose evolution can be described by the equation
\begin{equation}
\label{eq:ARUMA}
U(B) \phi(B) Y_t = \theta(B) \varepsilon_t
.
\end{equation}
\citet{TiaoTsay1983} refer to this model as a nonstationary ARMA model. \citet{HuangAnh1990}
call this model autoregressive unit root moving average (ARUMA), see also
\citet{WoodwardGrayElliott2017}. Here $\{\varepsilon_{t}\}$ is white noise, $B$ is the
backward shift operator and all roots of the polynomials $\phi(z)$ and $\theta(z)$ are
outside the unit circle. The nonstationary part is specified by the polynomial
$U(z) = 1 - U_1 z - U_2 z^2 - \dots - U_d z^d$ whose all roots have moduli~1 (i.e., lie on
the unit circle). Traditionally the polynomial $U(z)$ does not have coefficients to be
estimated. This is the case, for example, for the familiar ARIMA and seasonal ARIMA (SARIMA)
models obtained when $U(z) = (1-B)^{d}$ and $U(z) = (1-B)^{d}(1-B^s)^{d_{s}}$, respectively.
\citet{TiaoTsay1983} and \citet{TsayTiao1984} study (iterative) ordinary least squares
procedures for estimation of such models and, in particular, show how the unit roots can be
estimated consistently.
For time series data it is typical to consider whether seasonal trends appear. This behaviour
is easily captured by the existing models by allowing further polynomials to appear in the
model with the appropriate power transformation of $B$ to account for the
seasonality. Standard example is the \textit{SARIMA} class of models, mentioned above. The
operator $(1-B^{s})^{d_{s}}$ however is sometimes too crude and may be inpractical when the
number of seasons, $s$, is large or in the case of multiple seasons. A more flexible class
of models is obtained by replacing it with a operator containg only some harmonics of $1/s$.
With a seasonal extension, we refer to this class of models as \textit{SARUMA}. Here is a
symbolic representation:
\begin{equation}
\label{eq:SARUMA}
U_s(B^s) U(B) \phi_s(B^s) \phi(B) Y_t = \theta_s(B^s) \theta(B) \varepsilon_t,
\end{equation}
where $U_s(z)$ is a seasonal polynomial of degree $d_s$ where all roots are unit, $\phi_s(z)$
is a seasonal autoregressive polynomial of degree $p_s$, $\theta_s(z)$ is a seasonal moving
average polynomial of degree $q_s$ such that all roots of $\phi_s(z)$ and $\theta_s(z)$ lie
outside the unit circle. The remaining terms are as in Equation~\ref{eq:ARUMA}.
We also require that there are no common roots between the $\phi_s(z^s)\phi(z)$ and
$\theta_s(z^s)\theta(z)$ components of the model. In practice, it is sometimes useful to
factor $U(z)$ and $U_{s}(z)$ into further factors in order to obtain more meaningful and/or
manageable models.
In principle the SARUMA model can be written in the form of model~\eqref{eq:ARUMA} by
expanding $U_{s}(B^{s})U(B)$ and estimate it using the OLS method of \citet{TsayTiao1984} but
this looses any parsimony that might be achievable otherwise.
Parameterisations of stationary models through partial autocorrelations are widely used in
the stationary case but for unit root models partial autocorrelations are not
defined. Nevertheless, we show that partial autocorrelations equal to $\pm1$ naturally
describe multiplicative ARUMA models and neatly fit with the standard practice of fitting
ARIMA models. We continue to call them partial autocorrelations though they do not have the
usual statistical interpretation and are purely a parameterisation of the polynomial on the
left-hand side of Equation~\eqref{eq:SARUMA}.
The transformation from partial autocorrelations to polynomial coefficients is unique, so
residuals and sums of squares are easily available and estimation is posible.
In this paper we obtain the algebraic properties of the partial autocorrelations in the
context of unit roots. The main result is that if a partial autocorrelation sequence contains
some values equal to $1$ or $-1$, then it can be split at these values into sequences
each of which represents the partial autocorrelations of a factor of the overall polynomial
on the left-hand side of the model.
A separate paper will discuss the details of the estimation procedure
and its properties. An implementation is provided by \citet[][function
\texttt{sarima}]{Rsarima}.
\section{Levinson-Durbin algorithm and its inverse}
The use of partial autocorrelations as a parameterisation of autoregressive (AR) stationary
models and stable filters is well established. For stationary AR models there is a
one-to-one map between the autoregressive parameters and the partial autocorrelations. The
partial autocorrelations have a clear statistical meaning in this case. The one-to-one map
allows to think of the partial autocorrelations also as an alternative way to parameterise
the coefficients of the associated autoregressive polynomial.
For a stationary process $\{X_{t}\}$, let $\parcoef{1}{n},\dots,\parcoef{n}{n}$ be the
partial prediction coefficients for the best linear predictor,
$\parcoef{1}{n}Y_{t} + \dots + \parcoef{n}{n}Y_{t-n+1}$, of $Y_{t+1}$ based on the latest
available $n$ observations. Let $\parcor{1}, \parcor{2},\ldots$ be the partial
autocorrelations. It is convenient to define $\parcor{0} = 1$. Consider also the polynomial
\begin{align*}
1 - \parcoef{1}{n}z - \dots - \parcoef{n}{n}z^{n}
.
\end{align*}
The statistical meaning of the partial autocorrelations and partial prediction coefficients
is not really needed for the exposition below but gives context.
The Levinson-Durbin recursions \citep{BrockwellDavis1991} can be used to compute the partial
prediction coefficients from the partial autocorrelations, as follows:
\begin{align}
\label{eq:pacf2ar1}
\parcoef{n}{n} &= \parcor{n}
\\
\label{eq:pacf2ar2}
\parcoef{k}{n} &= \parcoef{k}{n - 1} - \parcor{n} \parcoef{n - k}{n - 1}
\qquad{}
\text{$k = 1, \dots, n - 1$}
\\ & \text{(for $n = 1, 2, \ldots $)} \nonumber
.
\end{align}
It is evident that the transformation from partial autocorrelations to partial coefficients is uniquely
defined without the need to put restrictions on
$\parcor{1}, \dots, \parcor{n}$. Note that, strictly speaking, the Levinson-Durbin algorithm
contains an additional step at each $n$ for computing the partial autocorrelation from
autocorrelations, which we don't need since we start with partial autocorrelations.
The recursions can be arranged in reverse order to compute the partial autocorrelations
from the partial coefficients $\parcoef{1}{m}, \dots \parcoef{m}{m}$:
\begin{align}
\label{eq:ar2pacf1}
\parcor{n} &= \parcoef{n}{n}
\\
\label{eq:ar2pacf2}
\parcoef{k}{n - 1} &= (\parcoef{k}{n} + \parcor{n} \parcoef{n - k}{n} )
/ (1 - \parcor{n}^{2})
\qquad{}
\text{$k = 1, \dots, n - 1$}
\\ & \text{(for $n = m, m - 1, \ldots, 1$)} \nonumber
.
\end{align}
At the end we have $\parcor{1},\dots,\parcor{n}$. Detailed discussion of several variants of
the Levinson-Durbin algorithm is given by \citet{porat2008digital}.
Of course, the inverse recursion will work only if $|\parcor{k}| \neq 1$ for
$k = 1, \dots, n$. In that case the relationship between the two sets of coefficients is
one-to-one. The case $\abs{\parcor{k}} > 1$ is not of interest to us here. Our aim is to show
that allowing some of the partial autocorrelations to be equal to one provides a very natural
parameterisation for models with arbitrary unit roots, including seasonal ARIMA models.
Since partial autocorrelations uniquely determine the filter coefficients, this means that
residuals can be computed and so a non-linear least squares estimation of the unit root
filter can be performed.
Some further insight can be obtained by noticing that the equations are paired for $k$ and
$n-k$:
\begin{align*}
\parcoef{k}{n} &= \parcoef{k}{n - 1} - \parcor{n} \parcoef{n - k}{n - 1} \\
\parcoef{n-k}{n} &= \parcoef{n-k}{n - 1} - \parcor{n} \parcoef{k}{n - 1}
\qquad{}
\text{$k = 1, \dots, [n/2]$}
.
\end{align*}
If $n$ is even and $k = n/2 = n-k$ the two equations can be reduced to
\begin{equation*}
\parcoef{n/2}{n}
= \parcoef{n/2}{n - 1} - \parcor{n} \parcoef{n/2}{n - 1}
= \parcoef{n/2}{n - 1}(1 - \parcor{n})
.
\end{equation*}
In particular, if $\parcor{n} = 1$ then $\parcoef{n/2}{n} = 0$ and if $\parcor{n} = -1$
then $\parcoef{n/2}{n} = 2\parcoef{n/2}{n - 1}$. It is also obvious that when $k \neq n/2$
that $\parcoef{k}{n} = -\parcoef{n-k}{n}$ when $\parcor{n} = 1$ and
$\parcoef{k}{n} = \parcoef{n-k}{n}$ when $\parcor{n} = -1$. For example, when $n=2$,
the above gives $\parcoef{1}{2} = 0$ if $\parcor{2} = 1$ and the polynomial must be
$1 - z^{2}$. When $\parcor{2} = -1$ then $\parcoef{1}{2} = 2\parcoef{1}{1} = 2\parcor{1}$
and the polynomial is $1 - 2\parcor{1}z + z^{2}$, which generates a pair of complex roots.
In what follows we show how polynomials can be separated after the occurrence of
a partial autocorrelation value of unit magnitude and show that sequence of partial
autocorrelations ending with a unit value produces a polynomial that contains only
roots on the unit circle. This methodology can be used to define each polynomial in
Equation~\eqref{eq:SARUMA}.
\section{Parameterisation using partial autocorrelations}
Let $\parcor{k}$, $k = 1, 2, \ldots$, be a sequence of partial autocorrellations.
Define polynomials $\Pn(z)$ by
\begin{equation} \label{eq:Pn}
\Pn(z) = \sum_{k = 1}^n \parcoef{k}{n} z^k
,\quad \text{for $n = 1, 2, \ldots$,}
\qquad \Pn[0](z) = 0
,
\end{equation}
where $\parcoef{k}{n}$ are the partial coefficients obtained from
$\parcor{1},\dots,\parcor{n}$, using
Equations~\eqref{eq:pacf2ar1}--\eqref{eq:pacf2ar2}.
Our main interest is in the positions of the zeroes of the polynomials
\begin{equation*}
\Phin(z) = 1 - \Pn(z)
= 1 - \sum_{k = 1}^{n} \parcoef{k}{n} z^k
,\quad \text{for $n = 0, 1, 2, \ldots$}
.
\end{equation*}
It is well known that if the coefficients of the polynomial $\Phin{(z)}$ are obtained from
partial autocorrelations $\parcor{1}, \dots, \parcor{n}$, such that $\abs{\parcor{i}} < 1$
for $i=1,\dots,n$, then all zeroes of the polynomial $\Phin{(z)}$ are outside the unit circle
(i.e., have moduli greater than~1). In particular, their product has modulus larger than~1.
What happens if $\abs{\parcor{i}} < 1$ for $i=1,\dots,n-1$, but $\parcor{n} = \pm1$? We
formulate the result as a lemma. It is hardly new but not easily available.
\begin{lemma} \label{le:unitRoots}
If $\abs{\parcor{i}} < 1$ for $i=1,\dots,n-1$, $\parcor{n} = \pm1$, then all zeroes,
$z_1, \dots, z_{n}$, of the polynomial $\Phin{(z)} = 1 - P_{n}(z)$ are on the unit circle
(i.e., $\abs{z_i} = 1$ for $i = 1, \dots, n$).
\end{lemma}
One way to show this is to notice that in that case the Vietta formulas imply that the
product of the zeroes of $\Phin(z)$ is $\pm{1}$. Then let $\parcor{n}^{(i)} \to \parcor{n}$,
$\abs{\parcor{n}^{(i)}} < 1$ for $i=1,2,\ldots$ and consider the sequence of polynomials
$\Phi_{n}^{(i)}(z)$, $i=1,2,\ldots$. Since the zeroes of polynomials are continuous
functions of their coefficients, and hence the partial autocorrelations, the zeroes of
$\Phi_{n}^{(i)}(z)$ converge to the zeroes of $\Phin(z)$. But all zeroes of
$\Phi_{n}^{(i)}(z)$ are strictly outside the unit circle, so their limits (the zeroes of
$\Phin(z)$) are outside or on the unit circle. This means that their product can be equal
to~1 only if all of them have modulus~1.
The following relation between the polynomials $\Pn(z)$ can be
obtained from the Levinson-Durbin recursions. Let $n \ge 2$. For
general $z$, multiply Equation~\eqref{eq:pacf2ar2} by $z^k$ for
$k = 1, \dots, n-1$, and sum to obtain
\begin{equation*}
\sum_{k=1}^{n-1} \parcoef{k}{n} z^k
=
\sum_{k=1}^{n-1} \phi_k^{(n-1)} z^k - \parcor{n} \sum_{k=1}^{n-1} \phi_{n-k}^{(n-1)} z^k
.
\end{equation*}
Using the definition of the polynomial $\Pn(z)$ and $\parcor{n} = \parcoef{n}{n}$,
this can be written as
\begin{equation*}
\Pn(z) - \parcor{n} z^n = P_{n-1}(z) - \parcor{n} z^n P_{n-1}(z^{-1})
,
\end{equation*}
which after rearranging becomes
\begin{equation}
\label{eq:PD1n}
(1 - \Pn(z)) = (1 - P_{n-1}(z)) - \parcor{n} z^n \left( 1 - P_{n-1}(z^{-1}) \right)
.
\end{equation}
The above equation was derived for $n \ge 2$ but it holds also, trivially, for $n = 1$.
Note that the coefficients of the polynomial $z^n \left( 1 - P_{n-1}(z^{-1})\right)$ are
those of $(1 - \Pn(z))$ in reverse order.
In general, the polynomials $1 - P_{n}(z)$, $n = 1, 2,\ldots$, do not have common zeroes. A
remarkable exception, particularly important for unit root models, is given by the following
lemma. It shows that if $z_{0}$ is such that it and $\zzero^{-1}$ are both zeroes of the
polynomial $1 - P_{m}(z)$, then they are also zeroes of the polynomials $1 - P_{n}(z)$ for
all $n\ge m$.
\begin{lemma}
Let $z_{0}$ be such that $1 - \Pn[m](z_{0}) = 0$ and $1 - \Pn[m](\zzero^{-1}) = 0$ for
some $m \in \mathbb{Z}^+$. Then $1 - \Pn(z_{0}) = 0$ and
$1 - \Pn(\zzero^{-1}) = 0$ for any $n \geq m$.
\end{lemma}
\begin{proof}
Setting $n = m + 1$ in Equation~\eqref{eq:PD1n} gives
\begin{equation}
(1 - \Pn[m+1](z))
= (1 - \Pn[m](z)) - \parcor{m+1} z^{m+1} \left( 1 - \Pn[m](z^{-1}) \right).
\end{equation}
If $z = z_{0}$ or $\zzero^{-1}$, then both terms on the right-hand side of the last equation
are zero, by the assumptions of the lemma. Hence, the left-hand side is also zero, i.e.
$1 - P_{m+1}(z_0) = 0$ and $1 - P_{m+1}(z_0^{-1}) = 0$. So, the claim of the lemma holds for
$n = m + 1$. But Equation~\eqref{eq:PD1n} holds also for $n > m + 1$, so the proof can be
completed by induction.
\end{proof}
The following corollary concerning roots on the unit circle is of primary interest for our
purposes. Indeed, complex roots of polynomials with real coefficients come in complex
conjugate pairs. Moreover, if $\abs{z_{0}} = 1$ then $z_{0}^{-1} = \bar{z}_{0}$. So, in this
case $1 - P_{m}(z_0) = 0$ implies $1 - P_{m}(z_0^{-1}) = 0$ and we have:
\begin{lemma} \label{le:z}
If $\abs{z_0} = 1$ and $1 - P_{m}(z_0) = 0$
then $1 - \Pn(z_{0}) = 0$ and $1 - \Pn(\zzero^{-1}) = 0$ for any $n \geq m$.
\end{lemma}
A useful consequence of Lemma~\ref{le:z} is the following result.
\begin{lemma} \label{le:zall} If all roots, $z_1, \dots, z_{m}$, of the polynomial
$1 - P_{m}(z)$ are on the unit circle (i.e., $\abs{z_i} = 1$ for $i = 1, \dots, m$), then
$1 - P_{m}(z)$ is a factor of $1 - \Pn(z)$ for any $n \geq m$.
\end{lemma}
\begin{proof}
Since the roots have moduli equal to 1 and $1 - P_{m-1}(z)$ has real
coefficients, it follows from Lemma~\ref{le:z} that
$z_{1},\dots,z_{m}$ are roots of $1 - \Pn(z)$ for all $n\ge m$,
hence the result.
\end{proof}
Lemma~\ref{le:zall} shows that if $\Pn(z)$ is the polynomial generated from the
partial autocorrelation sequence
$\parcor{1},\dots,\parcor{m},\parcor{m+1},\dots,\parcor{n}$, where
$\parcor{m}=\pm 1$ and $|\parcor{m+i}|<1$ for $i=1,\dots,n-m$, then $1 - \Pn(z)
= (1 - \Pn[m](z)(1 - T(z))$, where $T(z)$ is some polynomial. It turns out that
$\parcor{m+1},\dots,\parcor{n}$ are, up to possible sign changes, the partial
autocorrelations generating the polynomial $T(z)$. Our main result in this
section states the complete result.
\begin{theorem}[Main result] \label{thm:polyDecomp}
Let
$\parcor{1},\dots,\parcor{m},\parcor{m+1},\dots$, be partial autocorrelations, such that
$|\parcor{i}|\le 1$, for $i = 1,\dots, m - 1$, $\parcor{m}=\pm 1$, and
$|\parcor{m+i}| \le 1$ for $i \ge 1$.
Let $\Pn(z)$ be the polynomials defined by Equation~\eqref{eq:Pn}.
Let also $\pmbeta_{i} = (-1)^{d_{+}}\parcor{m+i}$, $i \ge 1$, where $d_{+}$ is
the number of zeroes of $1 - P_{m}(z)$ equal to $+1$.
Then, for each $n \ge m + 1$, $(1 - \Pn(z)) = (1 - \Pn[m](z))(1 - \Qn(z))$,
where the polynomial $\Qn(z)$ is generated from the partial autocorrelations
$\pmbeta_{1},\dots,\pmbeta_{n - m}$.
\end{theorem}
\begin{proof}
Changing $n$ to $l$ in Equation~\eqref{eq:PD1n} and summing from $m+1$ to $n$ we
obtain
\begin{align*}
\sum_{l=m+1}^n (1-P_l(z))
&= \sum_{l=m+1}^{n} (1-P_{l-1}(z)) - \sum_{l=m+1}^n \parcor{l} z^l \left( 1 - P_{l-1}(z^{-1})\right).
\\ &= \sum_{l=m}^{n-1} (1-P_{l}(z)) - \sum_{l=m+1}^n \parcor{l} z^l \left( 1 - P_{l-1}(z^{-1})\right).
\end{align*}
After cancelling the common terms in the two sides of the equation and
rearranging, we get
\begin{align}
\label{eq:PD2}
(1 - \Pn(z))
&= (1 - P_{m}(z)) - \sum_{l=m+1}^n \parcor{l} z^l \left( 1 - P_{l-1}(z^{-1})\right).
\nonumber
\end{align}
In particular, for $m = 0$ we have
\begin{align}
(1 - \Pn(z))
&= (1 - P_{0}(z)) - \sum_{l=1}^n \parcor{l} z^l \left( 1 - P_{l-1}(z^{-1})\right).
\nonumber
\\ &= (1 - \sum_{l=1}^n \parcor{l} z^l \left( 1 - P_{l-1}(z^{-1})\right).
\end{align}
By Lemma~\ref{le:unitRoots} all roots of the polynomial $1 - P_{m}(z)$ are on the unit
circle. Let $d_{+}$ and $d_{-}$ be the number of roots equal to $+1$ and $-1$,
respectively. The remaining $2r$ roots are complex conjugate pairs,
$\alpha_{i}, \alpha_{i}^{-1}$, $i=1,\dots,r$, where $\alpha_{i}^{-1}$ is the complex
conjugate of $\alpha_{i}$ since $|\alpha_{i}|=1$. Obviously, $m = d_{+} + d_{-} + 2r$.
We have
\begin{align*}
1 - \Pn[m](z)
&= \left( 1 - z \right)^{d_+} \left( 1 + z \right)^{d_-}
\prod_{i=1}^r \left(1 - \frac{z}{\alpha_i} \right) \left(1 - \alpha_i z
\right)
\\ &= \left( 1 - z \right)^{d_+} \left( 1 + z \right)^{d_-}
\prod_{i=1}^r \left(1 - (\frac{1}{\alpha_i} + \alpha_i) z + z^{2} \right)
.
\end{align*}
From this we get
\begin{align}
1 - \Pn[m](z^{-1})
&= \left( 1 - z^{-1} \right)^{d_+} \left( 1 + z^{-1} \right)^{d_-}
\prod_{i=1}^r \left(1 - \frac{z^{-1}}{\alpha_i} \right) \left(1 - \alpha_i z^{-1}
\right), \nonumber \\
&= z^{-d_{+}}(z - 1)^{d_{+}}
z^{-d_{-}}(z + 1)^{d_{-}}
\prod_{i=1}^r z^{-2} \left(z - \frac{1}{\alpha_i} \right) \left(z - \alpha_i \right)
\nonumber \\
&= z^{-d_{+}}(z - 1)^{d_{+}}
z^{-d_{-}}(z + 1)^{d_{-}}
z^{-2r} \prod_{i=1}^r \left(z^{2} - (\frac{1}{\alpha_i} + \alpha_{i} ) z + 1 \right)
\nonumber \\
&= z^{-m} (-1)^{d_+} \left( 1 - z \right)^{d_+} \left( 1 + z \right)^{d_-}
\prod_{i=1}^r \left(1 - \frac{z}{\alpha_i} \right) \left(1 - \alpha_i z \right)
\nonumber \\
\label{eq:PD4}
&= z^{-m} (-1)^{d_+} \left( 1 - P_{m}(z) \right)
\end{align}
Together with Equation~\eqref{eq:PD2} (with $n = m + 1$) this gives
\begin{align*}
1 - \Pn[m+1](z) &= 1 - \Pn[m](z) -
\parcor{m+1} z^{m+1} \left( z^{-m} (-1)^{d_+} \left( 1 - P_{m}(z) \right) \right),
\\ &= \left( 1 - P_{m}(z) \right) \left( 1 - (-1)^{d_+} \parcor{m+1} z \right).
\\ &= \left( 1 - P_{m}(z) \right) \left( 1 - \pmbeta_{1} z \right)
.
\end{align*}
Therefore, when $n = m + 1$, $1 - P_{m}(z)$ is a factor of $1 - P_{m+1}(z)$ and,
moreover, we have the explicit factorisation with $Q_{1}(z) = \pmbeta_{1} z$.
For the general case, let $n > m + 1$ and assume that the claim is true for all $l < n$.
Concentrate on the case $l = n$ and let $1 - Q_{l-m}(z)$ represent the polynomial remaining
after division of $1 - P_{l}(z)$ by $1 - P_{m}(z)$ for $l>m$, so that
\begin{equation}
\label{eq:PD5}
1 - P_{n}(z) = \left( 1 - P_{m}(z) \right) \left( 1 - Q_{n-m}(z) \right)
\end{equation}
Starting from Equation~\eqref{eq:PD1n}, and with the help of
Equation~\eqref{eq:PD4},
\begin{align}
1 - P_{n}(z)
&= 1 - P_{m}(z) - \sum_{l=m+1}^n \parcor{l} z^l \left( 1 - P_{l-1}(z^{-1}) \right)
\nonumber \\
&= 1 - P_{m}(z) - \sum_{l=m+1}^n \parcor{l} z^l \left( 1 - P_{m}(z^{-1}) \right)
\left( 1 - Q_{l-1-m}(z^{-1}) \right), \nonumber \\
&= 1 - P_{m}(z) - \sum_{l=m+1}^n \parcor{l} z^{l-m} (-1)^{d_+} \left( 1 - P_{m}(z) \right)
\left( 1 - Q_{l-1-m}(z^{-1}) \right), \nonumber \\
&= (1 - P_{m}(z))\left(1 - \sum_{l=m+1}^n (-1)^{d_+} \parcor{l} z^{l-m}
\left( 1 - Q_{l-1-m}(z^{-1}) \right) \right), \nonumber \\
&= (1 - P_{m}(z))\left(1 - \sum_{l=m+1}^n \pmbeta_{l-m} z^{l-m}
\left( 1 - Q_{l-1-m}(z^{-1}) \right) \right), \nonumber \\
&= (1 - P_{m}(z))\left(1 - \sum_{k=1}^{n-m} \pmbeta_{k} z^{k}
\left( 1 - Q_{k-1}(z^{-1}) \right) \right),
\label{eq:PD6}
\end{align}
Equation~\eqref{eq:PD6} shows that $1 - P_{m}(z)$ is a factor of $1 - P_{n}(z)$ for some
$n>m$ and moreover, by comparing it with Equation~\eqref{eq:PD5} we can see that
\begin{equation*}
1 - Q_{n-m}(z) = 1 - \sum_{l=1}^{n-m} \gamma_l z^l \left(1 - Q_{l-1}(z^{-1}) \right),
\end{equation*}
where $\gamma_l = (-1)^{d_+} \parcor{l}$. Notice the similarities between this equation
and Equation~\eqref{eq:PD2}. $1 - Q_{n-m}(z)$ is of the same form as the original
polynomial $1-\Pn(z)$ except that the original partial autocorrelation coefficients
$\parcor{k}$ have been replaced by $\gamma_k$.
By induction,
the claim of the theorem is proved.
\end{proof}
If there are more partial autocorrelations with modulus~1,
Theorem~\ref{thm:polyDecomp} can be applied recursively to get a factorisation
of the unit root polynomials.
\begin{theorem}
Let $m_{1} < m_{2} < \dots < m_{r}$, be positive integers such that
$\abs{\parcor{m_{i}}} = 1$, $i=1,\dots,r$.
Then for each $n \ge m_{r} + 1$
\begin{align*}
(1 - \Pn(z))
&= (1 - \Pn[m_{1}](z))(1 - \Pn[m_{2} - m_{1}](z))
\cdots
(1 - \Pn[m_{r} - m_{r-1}](z))
(1 - \Qn[n-m_{r}](z))
,
\end{align*}
where the polynomials $(1 - \Pn[m_{i}](z))$ are obtained from the partial
autocorrelations $\parcor{i}$, $i=m_{i-1}+1,\dots,m_{i}$ with adjusted signs
as given by Theorem~\ref{thm:polyDecomp} (applied recursively)
and the polynomial $\Qn(z)$ is generated from the partial autocorrelations
$\pmbeta_{1},\dots,\pmbeta_{n - m_{r}}$.
\end{theorem}
There are a number of ways to use Theorem~\ref{thm:polyDecomp} in
modelling. The most transparent and useful is given by the following result.
\begin{theorem}[ARUMA result] \label{thm:polyDecomp2}
Let $n > m$ and $\parcor{1},\dots,\parcor{m},\parcor{m+1},\dots,\parcor{n}$,
be partial autocorrelations, such that $|\parcor{i}|\le 1$, for $i = 1,\dots,
m - 1$, $\parcor{m}=\pm 1$, and $|\parcor{m+i}| < 1$ for $i = 1, \dots, n -
m$. Let $\Pn(z)$ be the polynomials defined by Equation~\eqref{eq:Pn}.
Then $(1 - \Pn(z)) = (1 - \Pn[m](z))(1 - \Qn(z))$, where all zeroes of $(1 -
\Pn[m](z))$ are on the unit circle and all zeroes of $(1 - \Qn(z))$ are
outside the unit circle. Further, $1 - \Pn[m](z)$ is generated by
$\parcor{1},\dots,\parcor{m}$ and $1 - \Qn(z))$ by $\pmbeta_{1},\dots,\pmbeta_{n
- m}$, where $\pmbeta_{i} = (-1)^{d_{+}}\parcor{m+i}$, $i = 1,\dots, n - m$
and $d_{+}$ is the number of zeroes of $1 - P_{m}(z)$ equal to $+1$.
\end{theorem}
\begin{proof}
The factorisation $(1 - \Pn(z)) = (1 - \Pn[m](z))(1 - \Qn(z))$ follows from
Theorem~\ref{thm:polyDecomp}. By Lemma~\ref{le:unitRoots} all zeroes of
$(1 - \Pn[m](z))$ are on the unit circle. Further, $(1 - \Qn(z))$ since by
Theorem~\ref{thm:polyDecomp} they are generated by partial autocorrelations
$|\pmbeta_{i}| < 1$, $i = 1,\dots, n - m$, which have the same moduli as
$\parcor{m+1},\dots, \parcor{n}$.
\end{proof}
Theorems~\ref{thm:polyDecomp} and~\ref{thm:polyDecomp2} fit nicely with the
standard practice of applying unit root and/or seasonal unit root filters
(represented here by the polynomial $1 - \Pn[m](z)$) to make a time series
stationary and then fitting a stationary model to the filtered time series. The
unit root filters are typically chosen in advance. Our results allow for
estimating the unit root filter. In the simplest case, $\parcor{m}$ (where $m$
is as in Theorem~\ref{thm:polyDecomp2}) is fixed to $\pm1$ and the remaining
partial autocorrelations are estimated using non-linear optimisation in the unit
cube.
Recall that for the SARUMA model $\Phi(z) = 1 - P_n(z)$. From the results above, we know
that $\Phi(z)$ decomposes into $\left( 1 - P_m(z) \right) \left( 1 - Q_{n-m}(z) \right)$
if all roots of $\left( 1 - P_m(z) \right)$ are on the unit circle. We express
$\left( 1 - P_m(z) \right)$ as $U(z)$, the unit root polynomial. If no unit partial
autocorrelation values remain in $\left( 1 - Q_{n-m}(z) \right)$ then this corresponds to
the stationary $\phi(z)$. Otherwise, the unit root polynomials can be iteratively
separated and stored as a product in $U(z)$. When $U(z)$ contains all nonstationary
aspects of the model, the Levinson-Durbin recursion can be used to generate the
coefficients of $U(z)$ by fixing the final coefficient to $\pm 1$. For example, say that
$U(z)$ is of degree $d$. The remaining partial autocorrelations can be used to estimate
the coefficients in $\phi(z)$, starting from $\beta_{d+1}$ and after multiplication with
$(-1)^{d_+}$.
Firstly assume, without loss of generality, that all seasonal polynomials
can be dropped ($U_s(z) = \phi_s(z) = \theta_s (z) \equiv 1$). Then the resulting ARUMA
model can be written
\begin{equation}
\label{eq:nsARMA}
\Phi(B) Y_t = \theta(B) \varepsilon_t.
\end{equation}
Furthermore, define the polynomial $P_n(z)$ as
\begin{equation*}
P_n(z) = \sum_{k = 1}^n \phi_k^{(n)} z^k
,\quad \text{for $n = 1, 2, \ldots$,}
\qquad \Pn[0](z) = 0
,
\end{equation*}
so that $\Phi(z) = 1 - P_n(z)$ with $n = p + d$.
We will discuss the details and the properties of an estimation procedure for
ARUMA models based on the results here in a separate paper. An implementation
can be found in package `sarima' \citep{Rsarima}.
\bibliographystyle{abbrvnat}
|
\section{Introduction}\label{sec-1}
Let $\mathbb{N}$ be the set of positive integers, $\mathbb{N}_0=\mathbb{N} \cup \{ 0\}$,
$\mathbb{Z}$ the set of integers,
$\mathbb{P}$ the set of prime numbers,
$\mathbb{R}$ the field of real numbers, $\mathbb{C}$ the field of complex numbers and $i:=\sqrt {-1}$.
In our previous paper \cite{MNT2021}, we considered general double Dirichlet series
of the form
\begin{align}\label{def:mDS}
\Phi _2 &(s_1,s_2;\alpha _1,\alpha _2)=\sum _{m_1=1}^\infty \sum _{m_2=1}^\infty \frac {\alpha _1 (m_1)\alpha _2 (m_2)}{m_1^{s_1}(m_1+m_2)^{s_2}}
\end{align}
where $\alpha_1,\,\alpha_2$ are arithmetic functions satisfying certain conditions.
If $\alpha_k$ ($k=1,2$) are periodic, then the study of
$\Phi _2 (s_1,s_2;\alpha _1,\alpha _2)$ essentially reduces to
the case of double Hurwitz zeta function defined by
\begin{align*}
& \zeta _r (s_1,s_2;d_1,d_2)=\sum _{m_1=0}^\infty \sum _{m_2=0}^\infty \frac{1}{(m_1+d_1)^{s_1}(m_1+m_2+d_1+d_2)^{s_2}}
\end{align*}
for positive rational numbers $d_1$ and $d_2$, which has been investigated classically.
A more general situation was studied by Tanigawa and the first-named author
\cite{MatsumotoTanigawa}. They considered a certain class of arithmetic functions $\{ \alpha\,:\mathbb{N} \to \mathbb{C}\}$, the Dirichlet series associated with $\alpha$:
\begin{equation}\label{def:DS}
\Phi (s;\alpha )=\sum _{n=1}^\infty \frac {\alpha (n)}{n^s}
\end{equation}
and its multiple version as follows.
Let $\mathcal{A}$ be the set of arithmetic functions satisfying the following three
conditions: If $\alpha\in\mathcal{A}$, then
\begin{itemize}
\item[(I)] $\Phi (s;\alpha)$ is absolutely convergent for $\Re s>\delta=\delta(\alpha)(>0)$;
\item[(II)] $\Phi (s;\alpha)$ can be continued meromorphically to the whole plane $\mathbb{C}$, holomorphic except for a possible pole (of order at most 1) at $s=\delta$;
\item[(III)] in any fixed strip $\sigma _1\leq \sigma \leq \sigma _2$, $\Phi (\sigma +it;\alpha)=O(|t|^A)$ holds as $|t|\to \infty$, where $A$ is a non-negative constant.
\end{itemize}
Then they showed that
for arithmetic functions $\alpha_1,\ldots,\alpha_r$ belonging to $\mathcal{A}$,
\begin{align*}
&\Phi _r (s_1,\dots ,s_r;\alpha _1,\dots ,\alpha _r)\\
&=\sum _{m_1=1}^\infty \cdots\sum _{m_r=1}^\infty \frac {\alpha _1 (m_1)
\alpha_2(m_2)\cdots
\alpha _r (m_r)}{m_1^{s_1}(m_1+m_2)^{s_2}\cdots(m_1+\cdots+m_r)^{s_r}}
\end{align*}
can be continued meromorphically to the whole space $\mathbb{C}^r$, and location of its possible singularities can be described explicitly. In particular, if all $\Phi (s;\alpha_k)$'s are entire, then $\Phi _r (s_1,\dots ,s_r;\alpha _1,\dots ,\alpha _r)$ is also entire
(see \cite[Theorem 1]{MatsumotoTanigawa}).
Further it is easy to extend the results in \cite{MatsumotoTanigawa} to the case when
the condition (II) is replaced by
(II)': $\Phi(s;\alpha)$ is continued to $\mathbb{C}$ and holomorphic except for finitely many poles.
As another example, Egami and the first-named author \cite{MatsumotoEgami} considered the double series associated with the von Mangoldt function. Let $\zeta(s)$ be the Riemann zeta-function and denote by
$\{\rho_{n}\}_{n\geq 1}$ the non-trivial zeros of $\zeta(s)$ numbered by the size of absolute values of their imaginary parts. (As for the pair of complex conjugates, we first count the zero whose imaginary part is positive, and then count negative.)
Let
$\Lambda$ be the von Mangoldt function
defined by
\begin{eqnarray*}
\Lambda (n)=\left\{ \begin{array}{ll}
\log p & (n=p^m\ \textrm{for}\ p\in \mathbb{P},\ m\in \mathbb{N}) \\
0 & ({\rm otherwise}) \\
\end{array} \right. .
\end{eqnarray*}
Then it is well-known that
\begin{equation*}
\Phi (s;\Lambda )=\sum _{n=1}^\infty \frac {\Lambda (n)}{n^s}=-\frac {\zeta '(s)}{\zeta (s)}
\end{equation*}
(see \cite[$\S$ 1.1]{Titch}). We denote this function by $M(s)$.
We see that $M(s)$ has poles at $s=1$, $s=-2m\ (m\in \mathbb{N})$ and $s=\rho_l$ $(l\in \mathbb{N})$, hence does not satisfy the assumption (II), or even (II)'. That is,
$\Lambda\notin\mathcal{A}$.
As a double version of $M(s)$, Egami and the first-named author \cite{MatsumotoEgami}
introduced
\begin{eqnarray*}
{\mathcal M}_2(s)&=&\sum _{m_1=1}^\infty \sum _{m_2=1}^\infty \frac {\Lambda (m_1)\Lambda (m_2)}{(m_1+m_2)^s}\ (=\Phi _2(0,s;\Lambda ,\Lambda ))
\end{eqnarray*}
which can be written as $\sum _{n=1}^\infty G_2(n)n^{-s}$,
where
\begin{equation*}
G_2(n)=\sum _{m_1+m_2=n}\Lambda (m_1)\Lambda (m_2).
\end{equation*}
We emphasize that ${\mathcal M}_2(s)$ is closely connected to the Goldbach conjecture, that is, $G_2(n)>0$ for all even $n\geq 4$ (cf. \cite{BHMS}).
Since $\Lambda\notin\mathcal{A}$, we cannot apply the result in \cite{MatsumotoTanigawa}. In fact, the line $\Re s=1$ can be shown to be the natural boundary of ${\mathcal M}_2(s)$ under some plausible assumptions, hence ${\mathcal M}_2(s)$ would not be continued meromorphically to the whole complex plane $\mathbb{C}$ (see \cite{MatsumotoEgami} and \cite{BSP11}).
In \cite{MNT2021}, we considered a certain family of the double series defined by \eqref{def:mDS} with $\alpha_1\equiv 1$, $\alpha_2\notin\mathcal{A}$, but
can be continued meromorphically to the whole space.
The first example is
\begin{equation}\label{def:phi(1,l)}
\Phi _2 (s_1,s_2;1,\Lambda )=\sum _{m_1=1}^\infty \sum _{m_2=1}^\infty \frac {\Lambda (m_2)}{m_1^{s_1} (m_1+m_2)^{s_2}}.
\end{equation}
The right-hand side of (\ref{def:phi(1,l)}) is absolutely convergent for $\Re s_2>1,\Re (s_1+s_2)>2$. Furthermore we can prove that $\Phi _2(s_1,s_2;1,\Lambda )$ can be continued meromorphically to the whole space $\mathbb{C}^2$ (see Theorem \ref{state1}). We also determine
the location of singularities of $\Phi _2(s_1,s_2;1,\Lambda )$ (see
Corollary \ref{C-2-1}).
The second example is, for any $\Phi (s;\alpha)$ with $\alpha\in\mathcal{A}$,
\begin{equation}\label{def_double_tilde}
\Phi _2 (s_1,s_2;1,\widetilde{\alpha} )=\sum _{m_1=1}^\infty \sum _{m_2=1}^\infty \frac {\widetilde{\alpha} (m_2)}{m_1^{s_1} (m_1+m_2)^{s_2}},
\end{equation}
where $\widetilde{\alpha}:\,\mathbb{N} \to \mathbb{C}$ is defined by
\begin{equation}\label{def_tilde_alpha}
\frac{\Phi (s;\alpha)}{\zeta(s)}=\sum _{n=1}^\infty \frac {\widetilde{\alpha}(n)}{n^s}\left(=:\Phi(s;\widetilde{\alpha})\right).
\end{equation}
In \cite{MNT2021} it was shown that $\Phi _2 (s_1,s_2;1,\widetilde{\alpha} )$ is absolutely convergent in the region
\begin{equation}
\left\{(s_1,s_2)\in \mathbb{C}^2 \mid \Re s_2>\max\{1,\delta\},\ \Re (s_1+s_2)>\max\{ 2,1+\delta\}\right\}\label{conv-region-intro}
\end{equation}
(see Lemma \ref{Prop-region}),
and $\Phi _2(s_1,s_2;1,\widetilde{\alpha})$ can be continued meromorphically to the whole space $\mathbb{C}^2$ (see Theorem \ref{Kocchiga-Th-4-2}), under the assumption of a certain quantitative version of the simplicity conjecture of $\zeta(s)$ (see Assumption \ref{Ass-2}).
Note that we can apply this result to the cases when $\widetilde{\alpha}$ is the M\"obius function $\mu$, the Euler totient function $\phi$, and so on.
The points $(u_1,u_2)$ on singularity sets are usually points of indeterminacy,
whose ``values''
can be understood only as limit values which depend on how to choose a limiting process.
Two typical limit processes are ``regular values'' and ``reverse values'', defined by
\begin{align}
& \lim_{s_1\to u_1}\lim_{s_2\to u_2}, \label{reg-vals} \\
& \lim_{s_2\to u_2}\lim_{s_1\to u_1}, \label{rev-vals}
\end{align}
respectively. These values were first introduced and studied by
Akiyama, Egami and Tanigawa \cite{AET}, and Akiyama and Tanigawa \cite{AT} in the case of multiple zeta-functions of Euler-Zagier type. It should be noted that Onozuka \cite{Onozuka2013} showed that limit values of Euler-Zagier multiple zeta-functions at non-positive integer points are always convergent, under any choice of the limiting process.
In contrast, the situation is different when we consider the reverse values of
\eqref{def:phi(1,l)} and \eqref{def_double_tilde}.
In \cite{MNT2021}, we explicitly computed the reverse values of \eqref{def:phi(1,l)} (see Proposition \ref{spv1}) and \eqref{def_double_tilde}
with $\widetilde{\alpha}=\mu$,
when $(s_1,s_2)=(-m,-n)$ where $m,n\in \mathbb{N}_{ 0}$ with $m \equiv n \mod 2$,
and $m\geq 1$ when $n\geq 2$.
In these cases, the reverse values are still finite definite values.
In the present paper, we observe the behavior of \eqref{def:phi(1,l)} and \eqref{def_double_tilde} with respect to the limiting process for the case \eqref{rev-vals} at $(-m,-n)$ with $m\in \mathbb{Z}$ and $n\in \mathbb{N}_{0}$ such that $m \not\equiv n \mod 2$. In particular, even if $(m,n)\in \mathbb{N}_{ 0}^2$, we find that the reverse values are sometimes not convergent in this case (see Theorems \ref{Theorem-3-1} and \ref{Theorem-5-1}). First we recall their analytic behavior (see Section \ref{sec-2}). From these results, we give asymptotic formulas for reverse values of \eqref{def:mDS} (see Section \ref{sec-3}). Moreover
we prove a certain reciprocity law of their residues which comes from the known reciprocity law for Bernoulli numbers (see Section \ref{sec-4}). Similarly, we observe the analytic behavior of \eqref{def_double_tilde} (see Section \ref{sec-5}).
Finally we give corrigendum and addendum to the previous paper \cite{MNT2021} (see Section \ref{sec-6}).
\
\section{Known results}\label{sec-2}
In this section, we recall several known results for $\Phi _2 (s_1,s_2;1,\Lambda )$ and
$\Phi _2(s_1,s_2;1,\widetilde{\alpha})$ given in \cite{MNT2021}.
In order to prove those results, we use the
Mellin-Barnes integral formula (see, for example, \cite[Section 14.51, p.289, Corollary]{WW}):
\begin{equation}\label{MB}
(1+\lambda )^{-s}=\frac 1{2\pi i}\int _{(c)}\frac {\Gamma (s+z)\Gamma (-z)}{\Gamma (s)}\lambda ^{z}dz,
\end{equation}
where $s,\lambda \in \mathbb{C}$ with $\lambda \neq 0,|{{\rm arg}\lambda}|<\pi,\ \Re s>0$,
$c\in \mathbb{R}$ with $-\Re s<c<0$, and the path of integration is the vertical line form $c-i\infty$ to $c+i \infty$.
We prepare some notation as follows. Let $a_k$ and $b_l$ be
constant terms of Laurent series of $M(s)=-\zeta'(s)/\zeta(s)$ at $s=-k$
($k\in\mathbb{N}$, $k:\textrm{even}$), and of $\Gamma(s)$ at $s=-l$
($l\in\mathbb{N}_0$),
respectively. Then we have
\begin{align}
&a_k=\lim_{s\to -k}\ \frac{d}{ds}(s+k)M(s)=-\frac{\zeta''(-k)}{2\zeta'(-k)},\label{def-ak}\\
&b_l=\lim_{s\to -l}\ \frac{d}{ds}(s+l)\Gamma(s)=\frac{(-1)^l}{l!}\left(\sum_{j=1}^{l}\frac{1}{j}-\gamma\right), \label{def-bk}
\end{align}
where $\gamma$ is the Euler constant.
Expressing $\Phi _2 (s_1,s_2;1,\Lambda )$ by using \eqref{MB} and shifting
the path of integration, we obtained
\begin{thm}[\cite{MNT2021}\ Theorem 2.1] \label{state1} $\Phi _2(s_1,s_2;1,\Lambda )$ can be continued meromorphically to the
whole space $\mathbb{C}^2$ by the following expression:
\begin{align}\label{mainthm}
&\Phi _2 (s_1,s_2;1,\Lambda ) \\
&=\frac {\zeta (s_1 +s_2 -1)}{s_2 -1}-(\log 2\pi ) \zeta (s_1+s_2)\nonumber\\
&+\sum ^{N-1}_{k=1 \atop k:{\rm odd}} \binom {-s_2}k M(-k)\zeta (s_1+s_2+k)\nonumber \\
&-\sum ^{N-1}_{k=1 \atop k:{\rm even}} \bigg[ \binom {-s_2}k\left\{ \left( -{a_k}+k!b_k\right)\zeta (s_1+s_2+k)-\zeta'(s_1+s_2+k)\right\} \nonumber \\
& \qquad\qquad -\frac {1}{k!}\frac{\Gamma'(s_2+k)}{\Gamma(s_2)}\zeta(s_1+s_2+k)\bigg]\nonumber\\
&-\frac{1}{\Gamma (s_2)}\sum _{n=1}^\infty {\rm ord}(\rho _n)\Gamma (s_2-\rho _n)\Gamma (\rho _n)\zeta (s_1+s_2-\rho _n)\nonumber \\
&+\frac{1}{2\pi i\Gamma (s_2)}\int_{(N-\eta )}\Gamma (s_2+z)\Gamma (-z)M(-z)\zeta (s_1+s_2+z)dz,\nonumber
\end{align}
where $N\in \mathbb N$,
$\eta$ is a small positive number, and ${\rm ord}(\rho _n)$ is the order of $\rho _n$ as a zero of $\zeta(s)$.
\end{thm}
\begin{cor}[\cite{MNT2021}\ Theorem 2.2] \label{C-2-1}\
The list of singularities of $\Phi _2 (s_1,s_2;1,\Lambda )$ are given as follows:
\begin{align}
& s_2=1, \label{sing-1}\\
& s_2=-l\quad (l\in \mathbb{N},\ l\geq 2),\label{sing-2}\\
& s_1+s_2=2-l\quad (l\in \mathbb{N}_0),\label{sing-3}\\
& s_2=-l+\rho_n\quad (n\in \mathbb{N},\ l\in \mathbb{N}_0),\label{sing-4}\\
&s_1+ s_2=1+\rho_n\quad (n\in \mathbb{N}).\label{sing-5}
\end{align}
\end{cor}
Next we consider a general class of double series $\Phi _2 (s_1,s_2;1,\widetilde{\alpha} )$ defined by
\eqref{def_double_tilde}.
Since
\begin{equation}
\frac{1}{\zeta(s)}=\sum _{n=1}^\infty \frac {\mu(n)}{n^s}\quad (\Re s>1),\label{mu-Dir}
\end{equation}
where $\mu$ is
the M{\"o}bius function
(see \cite[$\S$ 1.1]{Titch}),
from \eqref{def_tilde_alpha} we have
\begin{equation}
\widetilde{\alpha}(n)=\sum_{1\leq d \leq n \atop d\mid n}\alpha\left(\frac{n}{d}\right)\mu(d). \label{Rev-form}
\end{equation}
\begin{lem}[\cite{MNT2021}\ Proposition 3.1] \label{Prop-region}
If $\alpha$ satisfies the condition (I) of the class $\mathcal{A}$, we see that
$\Phi _2 (s_1,s_2;1,\widetilde{\alpha} )$ is absolutely convergent in the region
\begin{equation}
\left\{(s_1,s_2)\in \mathbb{C}^2 \mid \Re s_2>\max\{1,\delta\},\ \Re (s_1+s_2)>\max\{ 2,1+\delta\}\right\}.\label{conv-region}
\end{equation}
\end{lem}
Here we introduce the following assumption, which is plausible in view of
the Gonek-Hejhal conjecture (Gonek \cite{Gonek}, Hejhal \cite{Hejhal}).
\begin{assumption}\label{Ass-2}\ All non-trivial zeros of $\zeta(s)$ are simple
and
\begin{equation}\label{order:1/zeta'(rho)}
\frac 1{\zeta '(\rho _n)}=O(|{\rho _n}|^B)\quad (n\to \infty)
\end{equation}
with some constant $B> 0$.
\end{assumption}
Let $c_k(\alpha)$ be
the constant term of the Laurent series of
$\Phi(s;\alpha)/\zeta (s)$ at $s=-k$ ($k\in\mathbb{N}$, $k:\mathrm{even}$).
Define the Bernoulli numbers $\{B_n\}_{n\geq 0}$ by
\begin{equation*}
\frac {te^t}{e^t-1}=\sum _{n=0}^\infty B_n\frac {t^n}{n!}.
\end{equation*}
Then we obtained
\begin{thm}[\cite{MNT2021}\ (4.4)] \label{Kocchiga-Th-4-2}
Let $\alpha\in\mathcal{A}$.
Under Assumption \ref{Ass-2}, $\Phi _2(s_1,s_2;1,\widetilde{\alpha} )$ can be continued meromorphically to the whole space $\mathbb{C}^2$ by {\rm (\ref{mainthm2})} by the following expression:
\begin{align}\label{mainthm2}
&\Phi _2(s_1,s_2;1,\widetilde{\alpha} )\\
&=\frac{\Gamma(s_2-\delta)\Gamma(\delta)}{\Gamma(s_2)} \Res_{s=\delta}\ \left(\frac{\Phi(s;\alpha)}{\zeta(s)}\right)\zeta(s_1+s_2-\delta) \notag\\
& \ -2\Phi(0;\alpha)\zeta (s_1+s_2)-\sum _{k=1 \atop k:{\rm odd}}^{N-1}\binom {-s_2}k \frac {(k+1)\Phi(-k;\alpha)}{B_{k+1}}\zeta (s_1+s_2+k)\nonumber \\
&\ +\sum _{k=1\atop k:{\rm even}}^{N-1} \bigg[ \binom {-s_2}k\bigg\{ \frac {(-1)^{k/2}2(2\pi )^{k}\Phi(-k;\alpha)}{\zeta (1+k)}\notag \\
& \ \ \times \left(b_k \zeta(s_1+s_2+k)-\frac{1}{k!}\zeta'(s_1+s_2+k)\right)+c_k(\alpha)\zeta(s_1+s_2+k)\bigg\}\nonumber \\
& \qquad -\frac{\Gamma'(s_2+k)}{\Gamma(s_2)}\frac {(-1)^{k/2}2(2\pi )^{k}\Phi(-k;\alpha)}{(k!)^2\zeta (1+k)}\zeta(s_1+s_2+k)\bigg]\notag\\
&\ +\frac 1 {\Gamma (s_2)}\sum _{n=1}^\infty \Gamma (s_2-\rho _n)\Gamma (\rho _n)\frac {\Phi(\rho_n;\alpha)}{\zeta'(\rho _n)}\zeta (s_1+s_2-\rho _n)\nonumber \\
&\ +\frac 1 {2\pi i\Gamma (s_2)}\int _{(N-\eta)}\Gamma (s_2+z)\Gamma (-z)\Phi(-z;\widetilde{\alpha})\zeta (s_1+s_2+z)dz\nonumber
\end{align}
for $N\in \mathbb N$, where $\eta$ is a small positive number, and $b_k$ is defined by \eqref{def-bk}.
\end{thm}
\begin{cor}[\cite{MNT2021}\ Theorem 4.3] \label{Th-4-2}
The list of singularities of $\Phi _2 (s_1,s_2;1,\widetilde{\alpha} )$ is given by:
\begin{equation}
\begin{split}
& s_1+s_2=1-k\quad (k\in \mathbb{N}_0),\\
& s_2=-k\quad (k\in \mathbb{N},\ k\geq 2),\\
& s_2=-l+\rho_n\quad (l\in \mathbb{N}_0,\ n\in \mathbb{N}),\\
& s_1+ s_2=1+\rho_n\quad (n\in \mathbb{N}),\\
& s_2=-l+\delta\quad (l\in \mathbb{N}_0),\\
& s_1+ s_2=1+\delta,
\end{split}
\label{coro-2}
\end{equation}
where the last two equations are omitted when $\Phi(s;\alpha)$ has no pole or $\delta=1$.
\end{cor}
We define the reverse value of $\Phi _2 (s_1,s_2;\alpha_1,\alpha_2 )$ at $(u_1,u_2)$ on each singular set by
\begin{align*}
& \Phi _2^{\rm Rev} (u_1,u_2;\alpha_1,\alpha_2 )=\lim_{s_2\to u_2}\lim_{s_1\to u_1}\Phi _2 (s_1,s_2;\alpha_1,\alpha_2 ).
\end{align*}
Let $(s)_k:=s(s+1)(s+2)\cdots (s+k-1)$. Then we obtained
\begin{prop}[\cite{MNT2021}\ Proposition 5.1] \label{spv1} Let $m,n\in \mathbb{N}_0$ with $2\mid (m+n)$ and assume $m\geq 1$ when $n\geq 2$. Then
\begin{align*}
\Phi _2^{\rm Rev} (-m,-n;1,\Lambda )
&\ =\frac {B_{m+n+2}}{(n+1)(m+n+2)}+(\log 2\pi ) \frac{B_{m+n+1}}{m+n+1}\\
&\ \ -\sum _{k=1 \atop k:{\rm odd}}^n \binom n k \frac {k+1}{B_{k+1}}\zeta '(-k)\frac {B_{m+n-k+1}}{m+n-k+1}\\
&\ \ +\sum _{k=1\atop k:{\rm even}}^n \binom n k \frac {B_{m+n-k+1}}{m+n-k+1}\left( -a_k+k!b_k\right) \\
& \qquad\quad -\frac{(-1)^m m!n!}{(m+n+1)!} M(-m-n-1).
\end{align*}
\end{prop}
\begin{remark}\label{exceptionalcase}
In the above proposition, the case $m=0$, $n\geq 2$ is excluded. In this case
we can see that $\Phi _2^{\rm Rev} (-m,-n;1,\Lambda )$ is not convergent.
See Section \ref{sec-6}.
\end{remark}
Next we consider the case $\Phi(s;\alpha)=1$, namely $\widetilde{\alpha}=\mu$.
In this case, $c_k=c_k(\alpha)$ is defined by
\begin{align}
&c_k=\lim_{s\to -k}\ \frac{d}{ds}\frac{s+k}{\zeta(s)}=-\frac{\zeta''(-k)}{2(\zeta'(-k))^2}\quad (k:\textrm{even}). \label{def-ck-2}
\end{align}
From \eqref{mainthm2},
it follows that
\begin{align}
&\Phi _2(s_1,s_2;1,\mu )\label{mainthm2-2} \\
&=-2\zeta (s_1+s_2)-\sum _{k=1\atop k:{\rm odd}}^{N-1}\binom {-s_2}k \frac {(k+1)\zeta (s_1+s_2+k)}{B_{k+1}}\nonumber \\
&+\sum _{k=1\atop k:{\rm even}}^{N-1}\bigg[\binom {-s_2}k\bigg\{ \frac {(-1)^{k/2}2(2\pi )^{k}}{\zeta(1+k)}\left(b_k \zeta(s_1+s_2+k)-\frac{\zeta'(s_1+s_2+k)}{k!}\right)\notag\\
& \quad +c_k\zeta(s_1+s_2+k)\notag\bigg\} -\frac{\Gamma'(s_2+k)}{\Gamma(s_2)}\frac {(-1)^{k/2}2(2\pi )^{k}}{(k!)^2 \zeta (1+k)}\zeta (s_1+s_2+k)\bigg]\nonumber \\
&+\frac 1 {\Gamma (s_2)}\sum _{n=1}^\infty \Gamma (s_2-\rho _n)\Gamma (\rho _n)\frac {\zeta (s_1+s_2-\rho _n)}{\zeta '(\rho _n)}\nonumber \\
&+\frac 1 {2\pi i\Gamma (s_2)}\int _{(N-\eta)}\Gamma (s_2+z)\Gamma (-z)\frac {\zeta (s_1+s_2+z)}{\zeta (-z)}dz \notag
\end{align}
for $N\in \mathbb N$, where $\eta$ is a small positive number, $b_k$ and $c_k$ are defined by \eqref{def-bk} and \eqref{def-ck-2} (see \cite[(6.2)]{MNT2021}).
Since $\Phi(s;\alpha)=1$ has no pole, \eqref{coro-2} in this case implies
\begin{equation*}
\begin{split}
& s_1+s_2=1-k\quad (k\in \mathbb{N}_0),\\
& s_2=-k\quad (k\in \mathbb{N},\ k\geq 2),\\
& s_2=-l+\rho_n\quad (l\in \mathbb{N}_0,\ n\in \mathbb{N}),\\
& s_1+ s_2=1+\rho_n\quad (n\in \mathbb{N}).
\end{split}
\end{equation*}
We can calculate, similar to Proposition \ref{spv1}, the reverse value of
$\Phi _2(s_1,s_2;1,\mu )$ at $(s_1,s_2)=(-m,-n)$ for $m,n\in \mathbb{N}_0$,
$2\mid (m+n)$ and the additional condition $m\geq 1$ when $n\geq 2$.
Also, similar to Remark \ref{zzz}, $\Phi_2(s_1,s_2;1,\mu )$ is not
convergent if $m=0$, $n\geq 2$.
\section{Analytic behavior of $\Phi _2 (s_1,s_2;1,\Lambda )$ at nonpositive integer points}\label{sec-3}
In this section, we observe the analytic behavior of
$\Phi _2 (s_1,s_2;1,\Lambda )$
at $(s_1,s_2)=(-m,-n)$ for $m\in \mathbb{Z}$ and $n \in \mathbb{N}_{0}$ with
$m \not\equiv n \mod 2$. In fact, we aim to confirm whether the reverse values $\Phi _2^{\rm Rev} (-m,-n;1,\Lambda )$
at these points are convergent or not. For this aim, it is necessary to observe their analytic behavior at these points.
We prove the following.
\begin{thm}\label{Theorem-3-1}\ Let $m\in \mathbb{Z}$, $n\in \mathbb{N}_{0}$ and $\ell \in \mathbb{N}$ with $m+n=2\ell-1$. \\
If $m\geq 0$, then
\begin{align}\label{3-1}
&\Phi_2(-m,-n+\varepsilon;1,\Lambda)
= \frac{1}{\varepsilon}\bigg\{
(-1)^n\frac{m!n!}{(m+n+1)!}
\\
& \qquad +\sum_{k=1 \atop k:\,\text{\rm even}}^{n}\binom{n}{k}\zeta(-m-n+k)\bigg\}+O(1)\quad (\varepsilon \to 0). \notag
\end{align}
If $m \leq -1$, then
\begin{align}
\Phi_2&(-m,-n+\varepsilon;1,\Lambda) \label{Kocchiga-3-2}\\
& = \frac{2}{\varepsilon^2}\binom{n}{2\ell}+\frac{1}{\varepsilon}
\bigg[ \binom{n}{2\ell}\left\{(-a_{2\ell}+(2\ell)!b_{2\ell})-\sum_{j=0}^{2\ell-1}\frac{1}{n-j}\right\}\notag\\
& \quad +\sum_{k=1 \atop k:\,\text{\rm even}}^{n}\binom{n}{k}\zeta(-m-n+k)+\binom{n}{2\ell}\gamma-\frac{C(2\ell,n)}{(2\ell)!}\bigg] \notag\\
& \quad +O(1)\quad (\varepsilon \to 0), \notag
\end{align}
where $\gamma$ is the Euler constant and $C(2\ell,n)$ is the constant defined in \eqref{3-8} determined by the gamma function (see below).
\end{thm}
\begin{proof}
First we set $(s_1,s_2)=(-m,-n+\varepsilon)$ for a small $\varepsilon>0$ in \eqref{mainthm}. Then
\begin{align} \label{3-3}
&\Phi _2 (-m,-n+\varepsilon;1,\Lambda ) \\
&=\frac {\zeta (-m-n-1+\varepsilon)}{-n-1+\varepsilon}-(\log 2\pi ) \zeta (-m-n+\varepsilon)\notag\\
&\ +\sum ^{N-1}_{k=1 \atop k:{\rm odd}} \binom {n-\varepsilon}k M(-k)\zeta (-m-n+k+\varepsilon)\nonumber \\
&\ -\sum ^{N-1}_{k=1 \atop k:{\rm even}} \bigg[ \binom {n-\varepsilon}k\left\{ \left( -{a_k}+k!b_k\right)\zeta (-m-n+k+\varepsilon)-\zeta'(-m-n+k+\varepsilon)\right\} \nonumber \\
& \qquad\qquad -\frac {1}{k!}\frac{\Gamma'(-n+k+\varepsilon)}{\Gamma(-n+\varepsilon)}\zeta(-m-n+k+\varepsilon)\bigg]\nonumber\\
&\ -\frac{1}{\Gamma (-n+\varepsilon)}\sum _{h=1}^\infty {\rm ord}(\rho _h)\Gamma (-n-\rho _h+\varepsilon)\Gamma (\rho _h)\zeta (-m-n-\rho_h+\varepsilon)\nonumber \\
&\ +\frac{1}{2\pi i\Gamma (-n+\varepsilon)}\int_{(N-\eta)}\Gamma (-n+z+\varepsilon)\Gamma (-z)M(-z)\zeta (-m-n+z+\varepsilon)dz\nonumber \\
& = A_1-A_2+A_3-A_4-A_5+A_6, \notag
\end{align}
say. Since $1/\Gamma(-n+\varepsilon) \to 0$ as $\varepsilon \to 0$, we see that $A_5$ and $A_6$ tend to $0$. Also, letting $\varepsilon \to 0$, we have
\begin{align*}
&A_1 \to \frac {\zeta (-m-n-1)}{-n-1}=\frac {\zeta (-2\ell)}{-n-1}=0,\\
&A_2 \to (\log 2\pi)\zeta(1-2\ell),\\
&A_3 \to \sum ^{N-1}_{k=1 \atop k:{\rm odd}} \binom {n}{k} M(-k)\zeta (1-2\ell+k),
\end{align*}
which are convergent. The only remaining part is
\begin{align*}
A_4&=\sum ^{N-1}_{k=1 \atop k:{\rm even}} \binom {n-\varepsilon}k \left( -{a_k}+k!b_k\right)\zeta (-m-n+k+\varepsilon)\\
& \quad -\sum ^{N-1}_{k=1 \atop k:{\rm even}}\binom{n-\varepsilon}{k} \zeta'(-m-n+k+\varepsilon) \nonumber \\
& \quad -\sum ^{N-1}_{k=1 \atop k:{\rm even}} \frac {1}{k!}\frac{\Gamma'(-n+k+\varepsilon)}{\Gamma(-n+\varepsilon)}\zeta(-m-n+k+\varepsilon)\nonumber\\
& =A_{41}-A_{42}-A_{43},
\end{align*}
say, where we choose a sufficiently large $N$ such as $2\ell\leq N-1$.
As for $A_{41}$, since $m+n=2\ell-1$, we can see that
\begin{align}
A_{41}&=\binom {n-\varepsilon}{2\ell} \left( -a_{2\ell}+(2\ell)!b_{2\ell}\right)\zeta (1+\varepsilon)+O(1) \label{3-4}\\
&=\binom {n-\varepsilon}{2\ell} \left( -a_{2\ell}+(2\ell)!b_{2\ell}\right)\left(\frac{1}{\varepsilon}+\gamma+O(\varepsilon)\right)+O(1)\notag\\
&=\frac{1}{\varepsilon}\binom {n}{2\ell} \left( -a_{2\ell}+(2\ell)!b_{2\ell}\right)+O(1)\quad (\varepsilon \to 0).\notag
\end{align}
We remark that,
if $m\geq 0$ (namely $n\leq 2\ell-1$), then $\binom{n}{2\ell}=0$ and hence
$A_{41}=O(1)$.
As for $A_{42}$, we can similarly see that
\begin{align*}
A_{42}&=\binom {n-\varepsilon}{2\ell}\zeta'(1+\varepsilon)+O(1)\\
& =\frac{(n-\varepsilon)(n-1-\varepsilon)\cdots(n-2\ell+1-\varepsilon)}{(2\ell)!}\left(-\frac{1}{\varepsilon^2}+O(1)\right)+O(1).\notag
\end{align*}
For the case $n\geq 2\ell$, namely $m\leq -1$, we can see that
\begin{align} \label{3-5}
A_{42}&=-\frac{1}{\varepsilon^2}\binom {n}{2\ell} +\frac{1}{\varepsilon}\binom {n}{2\ell}\left(\sum_{j=0}^{2\ell-1}\frac{1}{n-j}\right)+O(1)\quad (\varepsilon \to 0).
\end{align}
For the case $n\leq 2\ell-1$, namely $m \geq 0$, we have
\begin{align*}
\binom{n-\varepsilon}{2\ell}&=\frac{(n-\varepsilon)\cdots(1-\varepsilon)(-\varepsilon)(-1-\varepsilon)\cdots(n-2\ell+1-\varepsilon)}{(2\ell)!}\\
& =(-\varepsilon)\frac{n!(-1)^{2\ell-n-1}(2\ell-n-1)!}{(2\ell)!}(1+O(\varepsilon))\\
& =\varepsilon(-1)^n\frac{n!(2\ell-n-1)!}{(2\ell)!}(1+O(\varepsilon)).
\end{align*}
Therefore we have
\begin{align}
A_{42}&=-\frac{(-1)^n}{\varepsilon}\frac{n!(2\ell-n-1)!}{(2\ell)!}+O(1)\quad (\varepsilon \to 0). \label{3-6}
\end{align}
Finally we consider
$$A_{43}=\sum ^{N-1}_{k=1 \atop k:{\rm even}} \frac {1}{k!}\frac{\Gamma'(-n+k+\varepsilon)}{\Gamma(-n+\varepsilon)}\zeta(-m-n+k+\varepsilon).$$
When $k\geq n+1$, we have $\Gamma'(-n+k+\varepsilon)=O(1)$ $(\varepsilon \to 0)$. Hence, by $1/ \Gamma(-n+\varepsilon) \to 0$, we have
\begin{align}\label{333}
\frac{\Gamma'(-n+k+\varepsilon)}{\Gamma(-n+\varepsilon)}\zeta(-m-n+k+\varepsilon)=O(1)\quad (\varepsilon \to 0).
\end{align}
Note that this includes the case $k=m+n+1$ (because the pole of the zeta
factor is cancelled by the gamma factor).
When $k\leq n$, we quote a result given in \cite[(2.18)]{MNT2021}:
\begin{equation}\label{3-7}
\frac{\Gamma'(s+k)}{\Gamma(s)}=\frac{(-1)^{k-1}(k+h)!}{h!}\frac{1}{s+k+h}+O(1) \quad (s\to -k-h)
\end{equation}
for even $k\geq 2$ and $h \in \mathbb{N}_{0}$. Letting $s=-n+\varepsilon$ and $h=n-k\geq 0$ in \eqref{3-7}, we can write
\begin{equation}\label{3-8}
\frac{\Gamma'(-n+k+\varepsilon)}{\Gamma(-n+\varepsilon)}=\frac{1}{\varepsilon}\frac{(-1)^{k-1}n!}{(n-k)!}+C(k,n)+O(\varepsilon) \quad (\varepsilon \to 0)
\end{equation}
with a constant $C(k,n)$.
The zeta factor $\zeta(-m-n+k+\varepsilon)$ remains finite unless
$k=2\ell$, while for $k=2\ell$ (which occurs when $2\ell\leq n$), we have
\begin{equation}\label{3-9}
\zeta(-m-n+k+\varepsilon)=\zeta(1+\varepsilon)=\frac{1}{\varepsilon}+\gamma+O(\varepsilon)\quad (\varepsilon \to 0).
\end{equation}
.
Summarizing the above results, we obtain the following.\\
For the case $n\leq 2\ell-1$, namely $m\geq 0$, we obtain from
\eqref{333} and \eqref{3-8} that
\begin{align}\label{3-10}
A_{43}& = \sum ^{n}_{k=1 \atop k:{\rm even}} \frac {1}{k!}\left(\frac{1}{\varepsilon}\frac{(-1)^{k-1}n!}{(n-k)!}\right)\zeta(1-2\ell+k+\varepsilon)+O(1)\\
& =\frac{1}{\varepsilon} \sum^{n}_{k=1 \atop k:{\rm even}} (-1)^{k-1}\binom{n}{k}\zeta(1-2\ell+k)+O(1)\quad (\varepsilon \to 0). \notag
\end{align}
For the case $n\geq 2\ell$, namely $m\leq -1$, we obtain from \eqref{333}, \eqref{3-8} and \eqref{3-9} that
\begin{align}\label{3-11}
A_{43}& = \sum ^{n}_{k=1 \atop {k:{\rm even} \atop k\neq 2\ell}} \frac {1}{k!}\left(\frac{1}{\varepsilon}\frac{(-1)^{k-1}n!}{(n-k)!}\right)\zeta(1-2\ell+k+\varepsilon)\\
&\ \ +\frac{1}{(2\ell)!}\left(\frac{1}{\varepsilon}\frac{(-1)^{2\ell-1}n!}{(n-2\ell)!}+C(2\ell,n)+O(\varepsilon)\right)\notag\\
& \qquad \times \left(\frac{1}{\varepsilon}+\gamma+O(\varepsilon)\right)+O(1) \notag\\
& = -\frac{1}{\varepsilon^2}\binom{n}{2\ell} \notag\\
&\ \ -\frac{1}{\varepsilon}\bigg(\sum ^{n}_{k=1 \atop {k:{\rm even} \atop k\neq 2\ell}} \binom{n}{k}\zeta(1-2\ell+k)+\binom{n}{2\ell}\gamma-\frac{C(2\ell,n)}{(2\ell)!}\bigg)\notag\\
&\ \ +O(1)\quad (\varepsilon \to 0). \notag
\end{align}
Combining \eqref{3-4}-\eqref{3-6}, \eqref{3-10} and \eqref{3-11}, we obtain the proof of Theorem \ref{Theorem-3-1}.
\end{proof}
\begin{remark}\label{Remark-3-2}\
As mentioned before, the condition $n \geq 2\ell$ implies $m \leq -1$. Let $h=-m\in \mathbb{N}$.
Then
\eqref{Kocchiga-3-2} in Theorem \ref{Theorem-3-1} shows that for $h,\,n\in \mathbb{N}$ such that $n-h$ is odd positive, $\Phi_2^{\rm rev}(h,-n;1,\Lambda)$ is clearly not convergent.
On the other hand, the condition $n \leq 2\ell-1$ implies that $m \geq 0$.
Hence
\eqref{3-1} in Theorem \ref{Theorem-3-1} shows that for $m,\,n\in \mathbb{Z}_{\geq 0}$ such that $m+n$ is odd positive, $\Phi_2^{\rm rev}(-m,-n;1,\Lambda)$ is convergent if and only if the residue
\begin{align}\label{3-12}
R(-m,-n):=(-1)^n\frac{m! n!}{(m+n+1)!} +\sum_{k=1 \atop k:\,\text{\rm even}}^{n}\binom{n}{k}\zeta(k-m-n)
\end{align}
equals to zero. We will give further results on the values of
$R(-m,-n)$ in the next section.
\end{remark}
\section{Reciprocity relations for residues}\label{sec-4}
In this section, we will evaluate the values of $R(-m,-n)$ in several cases.
In particular, we will prove certain reciprocity relation between
$R(-m,-n)$ and $R(-n,-m)$
for $m,n \in \mathbb{N}$ with $2\nmid (m+n)$.
Since
\begin{align}\label{zzz}
\zeta(1-r)=(-1)^{r-1}B_{r}/r \qquad(r\in \mathbb{N})
\end{align}
(see \cite[Chapter II]{Titch}), $R(-m,-n)$ defined by \eqref{3-12} can also be written as
\begin{align}\label{4-1}
R(-m,-n
=(-1)^n\frac{m!n!}{(m+n+1)!} -\sum_{j=1}^{[n/2]}\binom{n}{2j}\frac{B_{m+n+1-2j}}{m+n+1-2j}.
\end{align}
We first prove the following.
\begin{thm}\label{Theorem-4-1}\
For $N\in \mathbb{N}_{0}$,
\begin{align}\label{4-2}
& R(-1,-2N)=
\begin{cases}
0 & (N\geq 1),\\
\frac{1}{2} & (N=0).
\end{cases}
\end{align}
Hence $\Phi_2^{\rm Rev}(-1,-2N;1,\Lambda)$ is convergent for $N\geq 1$ and $\Phi_2^{\rm Rev}(-1,0;1,\Lambda)$ is not convergent.
\end{thm}
In order to prove this result, we recall the following reciprocity relation among Bernoulli numbers. This result was first proved by Saalsch\"utz \cite{Saal1892}, and was recovered by Gelfand \cite{Gelfand1967} (see also Agoh-Dilcher \cite{AgohD2008} for the details).
\begin{lem}\label{Lemma-4-2}\
For $p,q\in \mathbb{N}_{ 0}$,
\begin{align}\label{4-3}
& (-1)^{p+1}\sum_{l=0}^{q}\binom{q}{l}\frac{B_{p+1+l}}{p+1+l}+(-1)^{q+1}\sum_{l=0}^{p}\binom{p}{l}\frac{B_{q+1+l}}{q+1+l}\\
&\quad =\frac{p!q!}{(p+q+1)!}. \notag
\end{align}
\end{lem}
Using this relation, we will prove Theorem \ref{Theorem-4-1}.
\begin{proof}[Proof of Theorem \ref{Theorem-4-1}]
By \eqref{4-1} with $(m,n)=(1,2\ell-2)$, we have
\begin{align*
R(-1,-2\ell+2)&=\frac{(2\ell-2)!}{(2\ell)!} -\sum_{j=1}^{\ell-1}\binom{2\ell-2}{2j}\frac{B_{2\ell-2j}}{2\ell-2j}. \notag
\end{align*}
Setting $N=\ell-1\geq 0$ and $h=\ell-j-1$, we have
\begin{align}\label{rrr}
& R(-1,-2N)=\frac{(2N)!}{(2N+2)!}-\sum_{h=0}^{N-1}\binom{2N}{2h}\frac{B_{2h+2}}{2h+2}.
\end{align}
When $N=0$, \eqref{rrr} implies $R(-1,0)=1/2$. When $N\geq 1$,
setting $(p,q)=(2N,1)$ in Lemma \ref{Lemma-4-2}, we have
$$-\frac{B_{2N+2}}{2N+2}+\sum_{h=0}^{N}\binom{2N}{2h}\frac{B_{2h+2}}{2h+2}=\frac{(2N)!}{(2N+2)!},$$
because $B_{2r+1}=0$ $(r \geq 1)$. Combining this with \eqref{rrr},
we complete the proof.
\end{proof}
Next we aim to prove the following reciprocity relation among $R(-m,-n)$.
\begin{thm}\label{Theorem-4-3}\
For $m,n \in \mathbb{N}$ with $2\nmid (m+n)$,
\begin{align}\label{4-4}
& (-1)^n R(-m,-n)+(-1)^m R(-n,-m)=\frac{m!n!}{(m+n+1)!}.
\end{align}
\end{thm}
\begin{proof}
Using \eqref{zzz} we see that the left-hand side of \eqref{4-3}
(with $(p,q)=(m,n)\in\mathbb{N}^2$) equals to
\begin{align*}
& (-1)^{m+1}\sum_{l=0}^{n}\binom{n}{l}(-1)^{m+l}\zeta(-m-l)+(-1)^{n+1}\sum_{l=0}^{m}\binom{m}{l}(-1)^{n+l}\zeta(-n-l).
\end{align*}
By setting $j=n-l$ and $j=m-l$ in the first and the second terms, respectively,
this can be written as
\begin{align*}
& \sum_{j=0}^{n}\binom{n}{j}(-1)^{n-j+1}\zeta(-m-n+j)+\sum_{j=0}^{m}\binom{m}{j}(-1)^{m-j+1}\zeta(-m-n+j).
\end{align*}
Now assume $2\nmid (m+n)$. Then the two terms corresponding to $j=0$ are
cancelled with each other, and hence,
\eqref{4-3} implies that
\begin{align}\label{4-5}
& (-1)^{n-1}\sum_{j=1}^{n}\binom{n}{j}(-1)^{j}\zeta(-m-n+j) \\
& \quad +(-1)^{m-1}\sum_{j=1}^{m}\binom{m}{j}(-1)^{j}\zeta(-m-n+j)=\frac{m!n!}{(m+n+1)!}. \notag
\end{align}
From \eqref{3-12}, we have
\begin{align}\label{4-6}
R(-m,-n)&=(-1)^n\frac{m!n!}{(m+n+1)!} +\sum_{k=1 \atop k:\,\text{\rm even}}^{n}\binom{n}{k}\zeta(-m-n+k)\\
&=(-1)^n\frac{m!n!}{(m+n+1)!} +\sum_{k=1}^{n}\binom{n}{k}(-1)^{k}\zeta(-m-n+k), \notag
\end{align}
because $\zeta(-m-n+k)=0$ for any odd $k$ with $k\leq n$ and $m\geq 1$. In fact, $-m-n+k$ is even and negative in this case. From \eqref{4-5} and \eqref{4-6}, we have
\begin{align*}
& (-1)^{n-1}\left\{R(-m,-n)-(-1)^n\frac{m!n!}{(m+n+1)!}\right\}\\
& +(-1)^{m-1}\left\{R(-n,-m)-(-1)^m\frac{m!n!}{(m+n+1)!}\right\}=\frac{m!n!}{(m+n+1)!}.
\end{align*}
Thus we complete the proof.
\end{proof}
Combining Theorems \ref{Theorem-4-1} and \ref{Theorem-4-3}, we obtain the following.
\begin{cor} \label{Cor-4-4}\ For $N \in \mathbb{N}$,
$$ R(-2N,-1)=-\frac{1}{(2N+1)(2N+2)}.$$
\end{cor}
Finally we prove the following.
\begin{prop} \label{Prop-4-5}\ For $N \in \mathbb{N}$,
$$R(0,-2N-1)=-\frac{1}{2}.$$
\end{prop}
\begin{proof}
By \eqref{4-1} with $(m,n)=(0,2N-1)$, we have
\begin{align*}
R(0,1-2N)&=-\frac{1}{2N} -\sum_{j=1}^{N-1}\binom{2N-1}{2j}\frac{B_{2N-2j}}{2N-2j}.
\end{align*}
We obtain from \eqref{4-3} with $(p,q)=(0,2N-1)$ that
\begin{align*}
\frac{1}{2N}&=-B_1-\sum_{\mu=0}^{N-1}\binom{2N-1}{2\mu+1}\frac{B_{2\mu+2}}{2\mu+2}+\frac{B_{2N}}{2N}\\
& = -B_1-\sum_{\mu=0}^{N-2}\binom{2N-1}{2\mu+1}\frac{B_{2\mu+2}}{2\mu+2}\\
& =\frac{1}{2}-\sum_{j=1}^{N-1}\binom{2N-1}{2j}\frac{B_{2N-2j}}{2N-2j},
\end{align*}
by setting $j=N-1-\mu$. Combining these results, we complete the proof.
\end{proof}
\section{Residues of $\Phi_2(s_1,s_2;1,\widetilde{\alpha})$} \label{sec-5}
In this section, similar to Section \ref{sec-3}, we observe the analytic behavior of
$\Phi _2 (s_1,s_2;1,\widetilde{\alpha} )$
at $(s_1,s_2)=(-m,-n)$ for $m\in \mathbb{Z}$ and $n \in \mathbb{N}_{0}$ with
$m \not\equiv n \mod 2$, under Assumption \ref{Ass-2}.
Let $\alpha\in \mathcal{A}$ satisfying three conditions (I)-(III) in Section \ref{sec-1}. Let $\delta=\delta(\alpha)>0$ be defined in (I).
Furthermore, for even $k\in \mathbb{N}$, let
\begin{equation*}
D_k=D_k^{(\alpha)}=\frac{\Phi(-k;\alpha)}{\zeta'(-k)}=\frac{(-1)^{k/2}2(2\pi)^k\Phi(-k;\alpha)}{k! \zeta(k+1)}.
\end{equation*}
We prove the following.
\begin{thm}\label{Theorem-5-1}\ Let $m\in \mathbb{Z}$, $n\in \mathbb{N}_{0}$ and $\ell \in \mathbb{N}$ with $m+n=2\ell-1$.\\
Under Assumption \ref{Ass-2}, we have
\begin{align}
\Phi_2&(-m,-n+\varepsilon;1,\widetilde{\alpha}) = \frac{1}{\varepsilon}\bigg\{ (-1)^nD_{2\ell}\frac{m!n!}{(m+n+1)!} \label{5-1}\\
& \qquad +\sum_{k=1 \atop k:\,\text{\rm even}}^{n}(-1)^k\binom{n}{k}D_k\zeta(-m-n+k)\bigg\}+O(1)\quad (\varepsilon \to 0) \notag
\end{align}
if $m\geq 0$, and
\begin{align}\label{3-2}
\Phi_2&(-m,-n+\varepsilon;1,\widetilde{\alpha}) = \frac{2D_{2\ell}}{\varepsilon^2}\binom{n}{2\ell}+O(1/\varepsilon)\quad (\varepsilon \to 0
\end{align}
if $m \leq -1$.
\end{thm}
Note that the reverse value \eqref{5-1} is convergent if and only if the residue on the right-hand side vanishes.
Also, if $D_{2\ell}\neq 0$ (namely $\Phi(-2\ell;\alpha)\neq 0$), the reverse value \eqref{3-2} is not convergent.
\begin{proof}[Proof of Theorem \ref{Theorem-5-1}]\
Since the proof is quite similar to the case of $\Phi _2 (s_1,s_2;1,\Lambda)$, we will give a sketch of the proof.
Set $(s_1,s_2)=(-m,-n+\varepsilon)$ for a small $\varepsilon>0$ in
\eqref{mainthm2}. Then
\begin{align*}
&\Phi _2 (-m,-n+\varepsilon;1,\widetilde{\alpha} ) \\
&=\frac{\Gamma(-n+\varepsilon-\delta)\Gamma(\delta)}{\Gamma(-n+\varepsilon)} \Res_{s=\delta}\ \left(\frac{\Phi(s;\alpha)}{\zeta(s)}\right)\zeta(-m-n+\varepsilon-\delta) \notag\\
& \ -2\Phi(0;\alpha)\zeta (-m-n+\varepsilon)-\sum _{k=1 \atop k:{\rm odd}}^{N-1}\binom {n-\varepsilon}k \frac {(k+1)\Phi(-k;\alpha)}{B_{k+1}}\zeta (-m-n+\varepsilon+k)\nonumber \\
&\ +\sum _{k=1\atop k:{\rm even}}^{N-1} \bigg[ \binom {n-\varepsilon}k\bigg\{ \frac {(-1)^{k/2}2(2\pi )^{k}\Phi(-k;\alpha)}{\zeta (1+k)}\notag \\
& \ \ \times \left(b_k \zeta(-m-n+\varepsilon+k)-\frac{1}{k!}\zeta'(-m-n+\varepsilon+k)\right)+c_k(\alpha)\zeta(-m-n+\varepsilon+k)\bigg\}\nonumber \\
& \qquad -\frac{\Gamma'(-n+\varepsilon+k)}{\Gamma(-n+\varepsilon)}\frac {(-1)^{k/2}2(2\pi )^{k}\Phi(-k;\alpha)}{(k!)^2\zeta (1+k)}\zeta(-m-n+\varepsilon+k)\bigg]\notag\\
&\ +\frac 1 {\Gamma (-n+\varepsilon)}\sum _{n=1}^\infty \Gamma (-n+\varepsilon-\rho _n)\Gamma (\rho _n)\frac {\Phi(\rho_n;\alpha)}{\zeta'(\rho _n)}\zeta (-m-n+\varepsilon-\rho _n)\nonumber \\
&\ +\frac 1 {2\pi i\Gamma (-n+\varepsilon)}\int _{(N-\eta )}\Gamma (-n+\varepsilon+z)\Gamma (-z)\Phi(-z;\widetilde{\alpha})\zeta (-m-n+\varepsilon+z)dz\nonumber\\
& = A_1-A_2-A_3+A_4+A_5+A_6, \notag
\end{align*}
say. By just the same way as in the case of $\Phi _2 (-m,-n+\varepsilon;1,\Lambda)$, we can check that $A_1,\,A_2,\,A_3,\,A_5$ and $A_6$ are $O(1)$ as $\varepsilon \to 0$. By the definition of $D_k$, we have
\begin{align*}
A_4&=\sum _{k=1\atop k:{\rm even}}^{N-1} \binom {n-\varepsilon}{k} k!D_{k}b_k \zeta(-m-n+\varepsilon+k)\\
& -\sum _{k=1\atop k:{\rm even}}^{N-1} \binom {n-\varepsilon}{k} D_{k}\zeta'(-m-n+\varepsilon+k)\\
& +\sum _{k=1\atop k:{\rm even}}^{N-1} \binom {n-\varepsilon}{k}c_k(\alpha)\zeta(-m-n+\varepsilon+k)\nonumber \\
& -\sum _{k=1\atop k:{\rm even}}^{N-1}\frac{\Gamma'(-n+\varepsilon+k)}{\Gamma(-n+\varepsilon)}\frac {D_k}{k!}\zeta(-m-n+\varepsilon+k)\notag\\
& =A_{41}-A_{42}+A_{43}-A_{44},
\end{align*}
say, where we choose a sufficiently large $N$. Let $m+n=2\ell-1$ $(\ell \in \mathbb{N})$. Similarly as in the proof of Theorem \ref{Theorem-3-1},
we obtain
\begin{equation*}
A_{41}=
\begin{cases}
\frac{1}{\varepsilon}\binom{n}{2\ell}(2\ell)!D_{2\ell}b_{2\ell}+O(1) & (m\leq -1),\\
O(1) & (m\geq 0),
\end{cases}
\end{equation*}
\begin{equation*}
A_{42}=
\begin{cases}
-\frac{1}{\varepsilon^2}\binom{n}{2\ell}D_{2\ell}+O(1/\varepsilon) & (m\leq -1),\\
\frac{1}{\varepsilon}(-1)^{n-1}D_{2\ell}\frac{m!n!}{(m+n+1)!}+O(1) & (m\geq 0),
\end{cases}
\end{equation*}
\begin{equation*}
A_{43}=
\begin{cases}
\frac{1}{\varepsilon}\binom{n}{2\ell}c_{2\ell}(\alpha)+O(1) & (m\leq -1),\\
O(1) & (m\geq 0),
\end{cases}
\end{equation*}
\begin{equation*}
A_{44}=
\begin{cases}
-\frac{1}{\varepsilon^2}\binom{n}{2\ell}D_{2\ell}+O(1/\varepsilon) & (m\leq -1),\\
\frac{1}{\varepsilon}\sum_{1\leq k \leq n \atop k:{\rm even}}(-1)^{k-1}\binom{n}{k}D_k\zeta(-m-n+k)+O(1) & (m\geq 0).
\end{cases}
\end{equation*}
Thus we complete the proof.
\end{proof}
\section{Corrigendum and addendum to \cite{MNT2021}} \label{sec-6}
\noindent
Corrigendum to \cite{MNT2021}:
\begin{enumerate}[$\bullet$]
\item In \cite[Remark 5.2]{MNT2021}, we mentioned that if $m+n$ is odd with $n\geq 2$, then $\Phi_2(s_1,s_2;1,\Lambda)$ is not convergent as ${s_1\to -m}$ and ${s_2\to -n}$. However, as stated in Sections \ref{sec-3} and \ref{sec-4}
(especially Theorem \ref{Theorem-4-1}), it sometimes happens that $\Phi_2^{\rm Rev}(-m,-n;1,\Lambda)$ is convergent.
\item In \cite[Example 5.3]{MNT2021}, we listed
$$\Phi _2^{\rm Rev} (-1,0;1,\Lambda )=\frac 1 {12}\log 2\pi -\frac{3}{4}-\frac{\zeta''(-2)}{4\zeta'(-2)}+\frac{\gamma}{2},$$
which is not correct. It follows from Theorem \ref{Theorem-4-1} that
$\Phi _2^{\rm Rev} (-1,0;1,\Lambda )$ is not convergent.
\item In \cite[Example 6.1]{MNT2021}, we listed
$$\Phi _2^{\rm Rev}(-1,0;1,\mu )=\frac 1 6 -\frac {2\pi ^2}{\zeta (3)}\left(\frac{3}{2}-\gamma\right)-\frac{\zeta''(-2)}{4\zeta'(-2)^2},$$
which is not correct. In fact, as noted in \eqref{mu-Dir}, we see that $\widetilde{\alpha}=\mu$ implies $\Phi(s;\alpha)=1$, hence
\begin{equation*}
D_k=D_k^{(\alpha)}=\frac{1}{\zeta'(-k)}=\frac{(-1)^{k/2}2(2\pi)^k}{k! \zeta(k+1)}\neq 0
\end{equation*}
for even $k\geq 0$. Therefore it follows from Theorem \ref{Theorem-5-1} with $\widetilde{\alpha}=\mu$ that
\begin{align}
\Phi_2&(-1,\varepsilon;1,\mu) = \frac{1}{\varepsilon}\frac{D_{2}}{2} +O(1)\qquad (\varepsilon \to 0),
\end{align}
which implies that $\Phi _2^{\rm Rev}(-1,0;1,\mu )$ is not convergent.
\end{enumerate}
\
\noindent
Addendum to \cite{MNT2021}:
In \cite[Proposition 5.1]{MNT2021} (= Proposition \ref{spv1}),
we considered the situation $m,n\in\mathbb{N}_0$ with $2|(m+n)$, but
the case
$m=0$, $n\geq 2$ was excluded. As mentioned in Remark
\ref{exceptionalcase}, in this excluded case,
$\Phi _2^{\rm Rev} (-m,-n;1,\Lambda )$
is indeed not convergent.
In fact, in the proof of \cite[Proposition 5.1]{MNT2021}, we need to
compute
$$
-\frac{1}{k!}\frac{\Gamma'(-n+k+\varepsilon)}{\Gamma(-n+\varepsilon)}
\zeta(-m-n+k+\varepsilon)
$$
for even $k\leq n$. Using \eqref{3-8} we find that the above is
\begin{align}\label{add}
&=(-1)^k\binom{n}{k}\zeta(-m-n+k)\frac{1}{\varepsilon}
-\frac{C(k,n)}{k!}\zeta(-m-n+k)\\
&\quad +(-1)^k\binom{n}{k}\zeta'(-m-n+k)
+O(\varepsilon).\notag
\end{align}
If $m\geq 1$, then $-m-n+k$ is an even negative integer (because
$k$ and $m+n$ are even, $k\leq n$), and so $\zeta(-m-n+k)=0$.
Therefore \eqref{add} tends to
$\binom{n}{k}\zeta'(-m-n+k)$ as $\varepsilon \to 0$, which is the
formula on the last line of
\cite[p.452]{MNT2021}.
However if $m=0$, then $n$ is even, and
$\zeta(-m-n+k)=\zeta(0)=-1/2$ for $k=n$.
Therefore in this case \eqref{add} is $-(2\varepsilon)^{-1}+O(1)$,
which is not convergent.
Similar phenomenon also happens for $\Phi_2(s_1,s_2;1,\mu)$.
\
\section*{Acknowledgments.}
The authors are sincerely grateful to Professor Masatoshi Suzuki who first
pointed out a mistake in \cite{MNT2021}.
|
\section{Introduction}
Dynamical systems are mathematical equations written to describe the interactions of quantities that change in time in many science and engineering applications. Traditionally, mechanistic models describing these systems were obtained from the iterative process of deriving equations from first principles and testing the models with physical experimentation. The power of mechanistic models comes from the ability to directly explain the system at hand with known physical principles such as the thermodynamics of the system, heat and mass transfer processes, chemical kinetics, and the system's forces. Reliable predictions can be made about the system behavior in different regimes from mechanistic models due to the scientist's knowledge of the assumptions governing the various first principles and physical laws. Mechanistic models are preferred by scientists and engineers; however, it can take years to develop a complete and accurate description of a system and we do not know the underlying first principles describing all systems. \cite{SINDY} sped up the identification of mechanistic models for nonlinear dynamical systems with the method Sparse Identification of Nonlinear Dynamics (SINDy), which is a sparse regression problem of time derivatives obtained by finite difference methods with a library of candidate terms appearing in the dynamical system. SINDy has had great success with system identification for various applications such as plasma physics \citep{plasma_SINDy}, nonlinear optical communication \citep{nonlinear_optics_SINDy}, biological chemical reaction networks \citep{reaction_networks_SINDy, reactive_SINDy}, and fluid dynamics \citep{PDE_SINDy}.
With the emergence of increased computational power from GPUs and CPUs, along with the exponential growth of the amount of data gathered such as via automated experimentation in chemistry and biology \citep{trends_highthroughput_screening, szymanski2011adaptation}, sensor data from factories \citep{sensor_advances}, satellite and in situ earth observations \citep{satellite}, and large-scale simulations such as computational fluid dynamics (CFD) simulations \citep{calzolari2021deep} and climate models \citep{Randall, esd-12-401-2021}, data-driven models such as deep learning have emerged as a way to process and understand this large amount of data quickly.
Neural ordinary differential equations (ODEs) are a recent approach to data-driven modeling of time-series data and dynamical systems in which a neural network is used to learn an approximation to an equation governing the dynamics of the system. Neural ODEs were first introduced in \cite{NeuralODEPaper}'s seminal NeurIPS best paper. In the last few years, several other types of neural ODEs have emerged such as latent ODEs \citep{latent_ODEs}, Bayesian neural ODEs \citep{bayesianneuralode}, and neural stochastic ODEs \citep{stochastic_neural_ode}. Models describing the system can be created solely from observed data without the need for expert domain knowledge, which \cite{universal_diffeq} term as the universal differential equation. These models can then be used to make predictions of what the system will do for unobserved conditions. Traditional data-driven neural ODE models can be obtained in a few hours as opposed to the many years required to develop complete and accurate mechanistic models for dynamical systems. Neural ODEs integrate the neural ODE in time to obtain predictions for the observed data, whereas SINDy uses finite difference methods with the data to obtain numerical approximations for the time derivatives, which gives neural ODEs the advantage of having less stringent requirements on the frequency of the observations \citep{SINDy-sampling}. However, neural ODEs have the well-known major problems that they are not directly interpretable, and they do not make reliable predictions outside of the domain of their training region. \citep{SINDY-neural} were able to make the neural ODE interpretable by successfully recovering symbolic equations from conventional neural ODEs by using SINDy with predictions for the time derivatives obtained from the trained neural ODE.
Our work is motivated by the need to make neural ODEs directly interpretable without using additional methods like SINDy after training. We address this problem by making the case to use directly interpretable neural network architectures inside of the neural ODE framework rather than the conventional neural network based on the standard multilayer perceptron (MLP) \citep{Goodfellow-et-al-2016}. Nonlinear dynamical systems can be complex symbolic expressions. We chose to address the class of dynamical systems described by polynomials first before tackling more advanced expressions. Dynamical systems involving polynomials arise in a number of physical systems such as gene regulatory networks \citep{sanguinetti2018gene} and cell signaling networks \citep{gutkind2000signaling} in systems biology, population models in ecology \citep{royle2008hierarchical} and epidemiology \citep{singh2018mathematical}, and atmospheric chemical kinetics \citep{brasseur2017modeling}. In the process of this work, we developed a few deep polynomial neural network architectures, but had the most success with \citep{PiNetPaper}'s $\pi$-net. We are the first to use any deep polynomial neural network for the purpose of direct symbolic regression, which we demonstrate on a fourth order univariate polynomial. We are also the first to put a deep polynomial neural network into the neural ODE framework, which we term the polynomial neural ODE, and use the polynomial neural ODE to perform direct symbolic regression on nonstiff dynamical systems described by a polynomials such as the Lotka-Volterra model, Damped Oscillatory System, and Van der Pol model. In addition, we test the polynomial neural ODE on a model that does not involve polynomials, and demonstrate its effectiveness as another form of a deep polynomial function approximator. This work, along with future work, will allow the neural ODE to serve as a tool complementary to SINDy for system identification.
\FloatBarrier
\section{Methods}
\FloatBarrier
\subsection{Neural ODEs}
Neural ordinary differential equations (ODEs) learn an approximate ODE, given data for the solution, y(t) \citep{NeuralODEPaper}. The ODE that we seek to approximate is given by
\begin{equation}
\frac{dy(t)}{dt} = f\left(t, y(t), \theta \right),
\end{equation}
\noindent where t is the time, $y(t)$ is the vector of state variables, $\theta$ is the vector of parameters, and $f$ is the ODE model. For many scientific problems, it can take years to discover the functional form of the ODE described by $f$. Neural ODEs solve this problem by learning an approximation to the dynamics described by $f$ without learning the exact functional equation. The neural ode, which we denote by $NN$, is a neural network that approximates the model $f$:
\begin{equation}
\frac{dy(t)}{dt} \approx NN\left(t, y(t), \theta \right).
\end{equation}
\noindent Once the neural ODE has been trained, it is treated exactly the same way as an ODE. To obtain predictions for $y(t)$, the neural ODE is integrated as an initial value problem (IVP) with an ODE solver. Traditionally, neural ODEs have used the same architecture as a standard multilayer perceptron (MLP)
\begin{equation}
NN(x) = (L_1 \circ L_2 \dots \circ L_{l-1} \circ L_l)(x),
\end{equation}
\noindent that is the composition of several neural network layers, $L_i$
\begin{equation}
L_i(x) = \sigma(x*w_i+b_i),
\end{equation}
\noindent with nonlinear activation function $\sigma$, weights $w_i$, and bias $b_i$ \citep{Goodfellow-et-al-2016}.
\FloatBarrier
\subsection{Polynomial Neural ODEs}
Mathematical models in numerous application areas including gene regulatory networks \citep{sanguinetti2018gene} and cell signaling networks \citep{gutkind2000signaling} in systems biology, population models in ecology \citep{royle2008hierarchical} and epidemiology \citep{singh2018mathematical}, and atmospheric chemical kinetics \citep{brasseur2017modeling} are expressed as differential equations where the right hand side functions $f$ are polynomials. For this class of problems we present the polynomial neural ODE. While \cite{Learning-Polynomials} have theoretically and experimentally shown that conventional feedforward MLPs work as universal approximators for polynomials, we will make the case for using polynomial neural ODEs instead of conventional neural ODEs for this application space.
Polynomial neural networks are function approximators in which the output layer is expressed as a polynomial of the input layer.
There are several types of polynomial neural networks. Designing polynomial neural networks that can be trained easily without an explosion of parameters is still an active area of research \citep{PiNetPaper, FAN2020383, du2018power, liang2016deep}. In this paper, we use the $\pi$-net architecture from \citet{PiNetPaper}. $\pi$-nets were specifically designed to use skip connections to avoid the combinatorial explosion in the number of parameters of polynomial activation functions, which make the network harder to train. Essentially, these networks learn tensor decompositions of the polynomials. For our work we use $\pi$-net V1, for which the architecture is shown in Fig. \ref{fig:PiNetV1_arch}. The architecture is centered around Hadamard Products \citep{horn1994topics} of linear layers without activation functions to form higher order polynomials.
\begin{figure}[!ht]
\centering
\includegraphics[width=\textwidth]{fig/PiNetV1fig/NN_visual.pdf}
\caption{Example polynomial neural ODE for a chemical reaction system with molecules A, X, and Y. The neural network outputs a polynomial transformation of the input, which are the concentrations of the chemical species.}
\label{fig:NN_visual}
\end{figure}
Conventional neural networks are not interpretable, due to the complex arrangement of nonlinear activation functions. \cite{SINDY-neural} have had success using \cite{SINDY}'s Sparse Identification of Nonlinear Dynamics (SINDy) to recover symbolic forms of neural ODEs following training. However, the polynomial neural ODE architecture has a unique advantage. Since the output layer is a direct mapping of the input in terms of tensor and Hadamard products, symbolic tensor math can be used to obtain a direct polynomial form of the polynomial neural network without using additional tools such as SINDy. We use SymPy, the Python library for symbolic mathematics, to obtain a symbolic form of the polynomial neural ODE following training \citep{SymPy}. Since we expect to see a plethora of new neural network architectures for symbolic regression beyond polynomials, we anticipate the need for compatibility of symbolic computation with software such as PyTorch, TensorFlow, and JAX, which would avoid the need to extract the network architecture, weights, and biases into SymPy following training to obtain a symbolic equation.
\begin{figure}[h!]
\centering
\includegraphics[width=\textwidth]{fig/PiNetV1fig/PiNetArch.pdf}
\caption{The network architecture of \cite{PiNetPaper}'s $\pi$-net V1 is shown (left) alongside a worked example of a one layer input with variable name $x$ (right). The boxes labeled L represent standard linear layers, while the circles with the $*$ symbol represent a layer that is the Hadamard product of the layer's inputs. There are no standard activation functions such as tanh or ReLU in this network architecture, which makes the architecture directly interpretable.}
\label{fig:PiNetV1_arch}
\end{figure}
\FloatBarrier
\subsection{Training Neural ODEs}
Prior to training the neural ODE, the architecture is defined and the parameters in the network are initialized. For all of our neural ODEs, we initialized our weights and biases with a normal distribution with mean of 0. We used a standard deviation of 0.00005 and 0.01 for the conventional and $\pi$-net V1 neural ODEs respectively.
When training a neural ODE, the goal is to fit the neural ODE to observed data for the state variables, $y_{obs}$, as a function of time. The neural ODE is integrated with an ODE solver to obtain predictions for $y_{obs}$, which we will call $y_{pred}$. We used gradient descent to minimize the normalized MSE loss $L(y_{obs}, y_{pred})$ between $y_{obs}$ and $y_{pred}$:
\begin{equation}
L(y_{obs}, y_{pred}) = \frac{1}{N} \sum_i^N \left (\frac{y_{pred, i} - y_{obs, i}}{y_{scale}}\right)^2,
\end{equation}
\noindent where $y_{scale}=\mid y_{max} - y_{min} \mid$ is the scale factor used to normalize the values in the loss function. We batched our data into $N-1$ samples consisting of IVPs between 2 adjacent known data points. For each iteration of gradient descent, we simultaneously solve the $N-1$ initial value problems forward in time to the next observed data point explicitly using the fourth order explicit Runge–Kutta–Fehlberg method \citep{fehlberg1968classical}. Since we are only working with nonstiff ODEs, we are able to use an explicit discretization method to solve the neural ODEs. The advantage of using this method to solve the neural ODE is efficient direct backpropagation through the explicit ODE scheme, which the popular continuous time sensitivity adjoint method from \cite{NeuralODEPaper} lacks.
\FloatBarrier
\section{Results}
\FloatBarrier
\subsection{Univariate 4th Order Polynomial}
Prior to looking at any dynamical systems with neural ODEs, the ability of $\pi$-net V1 to learn basis polynomials was first tested. For the test case, we used the following fourth order univariate function:
\begin{equation}
f(x) = 3 x^4 + 16 x^3 + 5 x^2
\end{equation}
\noindent The training data for the x-values consisted of 20 uniformly spaced data points in the range -5.3 to 2.2. The values of $f(x)$ corresponding to the values of $x$ were obtained by directly substituting the $x$-values into the function.
Two neural networks were trained using the training data: (1) a conventional neural network with 5 layers consisting of 1x100x100x100x1 neurons in each layer with tanh activation functions and (2) a $\pi$-net V1 polynomial neural network that outputs fourth order polynomials. Each neural network was trained a total of 5 times and the best network was chosen based on the normalized MSE test loss, $L(y_{obs}, y_{pred})$, as well as the visual fit of the neural network's predictions against the known data.
Results from the neural networks with the best fit are shown in Fig \ref{fig:1D-example}. While not shown, conventional neural networks with ReLU activation functions produced similar results. The fourth order $\pi$-net V1 polynomial neural network reliably reproduced the same model, due to the polynomial constraint defined by its architecture, whereas there was some variability in the final conventional neural network model. The conventional neural network exhibits Gibbs phenomenon, an oscillatory behavior around the observed data usually caused by discontinuities, which is typical for approximation functions such as Fourier series, orthogonal polynomials, splines, and wavelets \citep{GibbsPhenom, jerri2013gibbs}. Since a neural network is another type of approximation function, this behavior is not surprising. Gibbs phenomenon was found every time the conventional neural network was fit to the data. The model in Fig \ref{fig:1D-example} was chosen because it exhibited the least amount of Gibbs phenomenon. Additionally, we observed that ReLU activation functions produce Gibbs phenomenon to a lesser extent than tanh activation functions. We suspect that this is because tanh functions bound the output between -1 and 1, whereas ReLU functions bound the output between 0 and $\infty$, which leads to fewer discontinuities.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{fig/PiNetV1fig/1D-example.pdf}
\caption{Predictions inside and outside of the training region for a conventional neural network with tanh activation functions, and a fourth order polynomial neural network after learning the fourth order polynomial. The recovered symbolic equation from the polynomial neural network is shown below the figures. The red terms indicate additional terms.}
\label{fig:1D-example}
\end{figure}
Neural networks are known to be poor at making predictions outside of their training region, which Fig \ref{fig:1D-example} also shows. When the prediction range is slightly extended, the conventional neural network is completely unable to make an accurate prediction, whereas the polynomial neural network makes accurate predictions in this range because it learned the functional form of the data.
The main advantage of polynomial neural networks is the ability to directly obtain a symbolic representation of the network. As shown in Fig. \ref{fig:1D-example}, coefficients of the polynomial accurate to 7 significant figures were directly predicted from 20 observations. After performing equation recovery, the extra terms can be systematically removed to simplify the model. In this case, the red terms are small enough to be dropped after further analysis by the modeler.
\FloatBarrier
\subsection{Lotka-Volterra Deterministic Oscillator}
Our first demonstration of a polynomial neural ODE is on the deterministic Lotka-Volterra ODE model, which describes the predator-prey population dynamics of a biological system \citep{Lotka1925, volterra1926variazioni}. When written as a set of first order nonlinear ODEs, the model is given by
\begin{align} \label{lotka-volterra-equations}
\frac{dx}{dt} = 1.5 x - x y, \\
\frac{dy}{dt} = -3 y + x y,
\end{align}
\noindent with initial conditions $x=1$ and $y=1$.
Since the problem is nonstiff, we generate our training data by integrating the IVP with SciPy and torchdiffeq using DOPRI5, a fourth order embedded method in the Runge–Kutta family of ODE solvers, with the default settings at 200 points uniformly spaced in time between 0 and 10 \citep{2020SciPy-NMeth, Chen_torchdiffeq_2021, dopriref}. As discussed in the methods section, we batch our data into 199 training samples consisting of IVPs between 2 adjacent known data points, and simultaneously solve the 199 IVPs during each epoch using our own code for the fourth order explicit Runge–Kutta–Fehlberg method, which allows us to directly perform backpropagation through the ODE discretization scheme \citep{fehlberg1968classical}.
Four neural networks were trained using the training data: (1) a conventional neural network with 5 layers consisting of 2x50x50x50x2 neurons in each layer with tanh activation functions and (2) three separate $\pi$-net V1 polynomial neural networks of degree two, three, and four.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{fig/PiNetV1fig/Lotka-Volterra/Lotka_Volterra-Full.pdf}
\caption{The conventional neural ODE and the $\pi$-net V1 polynomial neural ODE were integrated from the initial conditions to a time of 40, to test the performance of the neural ODEs. The vertical green line indicates where the training data ends and the neural ODE begins making a prediction. The symbolic form of the $\pi$-net V1 polynomial neural network was obtained for each of the degrees shown. The correct equation is shown above the recovered equations. The red terms indicate terms that can be dropped.}
\label{fig:PiNet-Lotka-Voltera-Full}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{fig/PiNetV1fig/Lotka-Volterra/Grid_VectorField.pdf}
\caption{The vector fields of the time derivatives of variables x and y are plotted for the conventional neural ODE with tanh activations (left), the true ODE equation (middle), and the second order $\pi$-net V1 polynomial neural ode (right). The top row shows the training region, while the remaining rows show the region outside of the training region. The blue circles mark the training data.}
\label{fig:PiNet-Lotka-Volterra-VectorField}
\end{figure}
Following training, the performance of the neural networks was tested by integrating the neural ODE from the initial conditions up to a time value of 40. As shown in Fig \ref{fig:PiNet-Lotka-Voltera-Full}, both the conventional neural ODE and the fourth order polynomial neural ODE are able to accurately predict the trajectory of the dynamical system beyond the training region for the same initial values as the training data.
Fig \ref{fig:PiNet-Lotka-Voltera-Full} doesn't offer much insight about how the conventional neural ODE differs from the polynomial neural ODE, so the vector fields of the time derivatives of x and y were also plotted. As shown in Fig \ref{fig:PiNet-Lotka-Volterra-VectorField}, the second order $\pi$-net V1 neural ODE accurately learns the true vector field for the training region, as well as outside of the training region. The conventional neural ODE learns a close approximation to the training region's vector field, which is why it was able to accurately predict the trajectory of the dynamical system beyond the training region for the same initial values as the training data, as shown in Fig \ref{fig:PiNet-Lotka-Voltera-Full}. However, it doesn't learn an accurate enough approximation to make predictions for observations outside of its training region. This is clear just by looking at the difference in vector fields for the region of space outside of the training region. On the other hand, the polynomial neural ODE has no problem with the same task.
To show that the $\pi$-net V1 polynomial neural ODE learns the same equation regardless of the degree of the polynomial output, three separate polynomial neural ODEs were trained with degrees of two, three, and four. Following training, the symbolic form of each of the polynomial neural ODEs was obtained, as shown in Fig \ref{fig:PiNet-Lotka-Voltera-Full}. The second, third, and fourth order polynomial neural ODEs were able to recover 6, 2, and 3 significant digits for the coefficients respectively. The values of the coefficients belonging to the terms not found in the original ODE differ each time, which serves as a clue to the modeler that these terms can be dropped from the final equation.
\FloatBarrier
\subsection{Damped Oscillatory System}
Our next demonstration of a polynomial neural ODE is on the deterministic damped oscillatory system, a popular toy problem for Neural ODEs \citep{NeuralODEPaper, CollocationNeuralODE}. Damped oscillations are common in engineering, physics, and biology \citep{doi:10.1080/00107514.2011.644441, karnopp1990system}. For our work, we refer to the following model as the damped oscillator:
\begin{eqnarray} \label{eq:damped-oscillator}
\frac{dx}{dt} &=& -0.1 x^3 - 2 y^3 \nonumber \\
\frac{dy}{dt} &=& 2 x^3 - 0.1 y^3,
\end{eqnarray}
\noindent with initial conditions $x=1$ and $y=1$. Since the problem is nonstiff, we generate our training data by integrating the IVP with SciPy and torchdiffeq using DOPRI5, a fourth order embedded method in the Runge–Kutta family of ODE solvers, with the default settings at 100 points uniformly spaced in time between 0 and 25 \citep{2020SciPy-NMeth, Chen_torchdiffeq_2021, dopriref}. As discussed in the methods section, we batch our data into 99 training samples consisting of IVPs between 2 adjacent known data points, and simultaneously solve the 99 IVPs during each epoch using our own code for the fourth order explicit Runge–Kutta–Fehlberg method, which allows us to directly perform backpropagation through the ODE discretization scheme \citep{fehlberg1968classical}.
Three neural networks were trained using the training data: (1) a conventional neural network with 5 layers consisting of 2x50x50x50x2 neurons in each layer, with tanh activation functions, and (2) two separate $\pi$-net V1 polynomial neural networks of degree three and four. Previous work has demonstrated that this damped oscillatory system is challenging for traditional neural ODEs to learn \citep{NeuralODEPaper, CollocationNeuralODE}. Several authors have made the model easier to learn with neural ODEs by making the first layer of the neural network a cubic function; however, this approach requires a priori knowledge that the functional form of the equation is cubic \citep{Chen_torchdiffeq_2021, CollocationNeuralODE}. Since the traditional neural network is supposed to learn the dynamics on its own, we do not use this approach for any of our neural ODEs.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{fig/PiNetV1fig/Damped-Oscillator/Damped_Oscillator-Full.pdf}
\caption{The conventional neural ODE and the $\pi$-net V1 polynomial neural ODE were integrated from the initial conditions to a time of 70, to test the performance of the neural ODEs. The vertical green line indicates where the training data ends and the neural ODE begins making a prediction. The symbolic forms of the polynomial neural ODEs were also obtained for each of the degrees shown. The red terms indicate additional terms that can be dropped.}
\label{fig:PiNet-Damped-Oscillator-Full}
\end{figure}
Following training, the performance of the neural networks was tested by integrating the neural ODE from the initial conditions up to a time value of 70. As shown in Fig \ref{fig:PiNet-Damped-Oscillator-Full}, the third order polynomial neural ODE is able to accurately predict the trajectory of the dynamical system beyond the training region, for the same initial values as the training data. The identical match of vector fields between the third order polynomial and that of the true ODE model, as shown in Fig \ref{fig:PiNet-Damped-Oscillator-VectorField}, indicates why the polynomial neural network is able to accurately predict the trajectory.
In contrast, the conventional neural ODE is able to make an accurate prediction only up to about 15 time units past the training region. The conventional neural network's predictions do not preserve the general shape of the solution to the ODE, whereas the polynomial constraint on the polynomial neural ODE ensures that the functional shape is preserved for the prediction task. Looking at the vector field for the conventional neural ode between -0.5 and 0.5, which is the zoomed in portion of the training region, since the training data does not have many values in this range, the neural ode was unable to learn the center of the vector field's spiral well. We observe and hypothesize that this is the reason why the conventional neural ODE is known to struggle with learning the damped oscillator model: the errors from the center of the spiral increase over the trajectory as it dampens. Additionally, the zoomed out vector field in the range -1000 to 1000 shows that the conventional neural ODE's vector field undergoes major distortion outside of its training region. This effect is not observed in the polynomial neural network.
Fig \ref{fig:PiNet-Damped-Oscillator-Full} shows the symbolic equations recovered from the third and fourth order polynomial neural ODEs. Five significant figures were recovered from both of the neural odes, and the remaining coefficients are small enough to be dropped.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{fig/PiNetV1fig/Damped-Oscillator/Grid_VectorField.pdf}
\caption{The vector fields of the time derivatives of variables x and y are plotted for the conventional neural ODE with tanh activations (left), the true ODE equation (middle), and the third order $\pi$-net V1 polynomial neural ode (right). The top row shows the entire training region, the middle row shows a smaller subset of the training region where oscillations dampen, and the bottom row shows the region outside of the training region. The blue circles mark the training data.}
\label{fig:PiNet-Damped-Oscillator-VectorField}
\end{figure}
\FloatBarrier
\subsection{Van der Pol Deterministic Oscillator}
Our final demonstration of the performance of polynomial neural ODEs is on the Van der Pol oscillator, a nonconservative oscillator with nonlinear damping \citep{VanderPolModel}. The Van der Pol equation describes several processes of relaxation-oscillations in the physical and biological sciences. For example, it has been used to model action potentials of neurons, tectonic plates in a geological fault, and oscillations of the left and right vocal cords during speech \citep{FITZHUGH1961445, 4066548, doi:10.1142/S0218127499001620, doi:10.1121/1.4798467}. The Van der Pol oscillator is described by the following second order ordinary differential equation:
\begin{equation}
\frac{d^2x}{dt^2}-\mu(1-x^2) \frac{dx}{dt}+x= 0
\end{equation}
\noindent and can be rewritten as a system of first order ODEs:
\begin{eqnarray}
\frac{dx}{dt} &=& y \nonumber \\
\frac{dy}{dt} &=& \mu y - \mu x^2 y - x.
\end{eqnarray}
We chose to assign $\mu=5$ and use initial conditions $x=2$ and $\frac{dx}{dt}=y=0$. Since the problem is nonstiff, we generate our training data by integrating the IVP with SciPy and torchdiffeq using DOPRI5, a fourth order embedded method in the Runge–Kutta family of ODE solvers, with the default settings at 200 points uniformly spaced in time between 0 and 25 \citep{2020SciPy-NMeth, Chen_torchdiffeq_2021, dopriref}. As discussed in the methods section, we batch our data into 199 training samples consisting of IVPs between 2 adjacent known data points, and simultaneously solve the 199 IVPs during each epoch using our own code for the fourth order explicit Runge–Kutta–Fehlberg method, which allows us to directly perform backpropagation through the ODE discretization scheme \citep{fehlberg1968classical}. Three neural networks were trained using the training data: (1) a conventional neural network with 5 layers consisting of 2x50x50x50x2 neurons in each layer with tanh activation functions, and (2) two separate $\pi$-net V1 polynomial neural networks of degree three and four.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{fig/PiNetV1fig/Van-der-Pol/Van_der_Pol-Full.pdf}
\caption{The conventional neural ODE and the $\pi$-net V1 polynomial neural ODE were integrated from the initial conditions to a time of 80 to test the performance of the neural ODEs. The vertical green line indicates where the training data ends and the neural ODE begins making a prediction. The symbolic forms of the polynomial neural ODEs were also obtained for each of the degrees shown. The red terms indicate additional terms that can be dropped.}
\label{fig:PiNet-Van-der-Pol-Full}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{fig/PiNetV1fig/Van-der-Pol/Limit_Cycle_Test_VdPol2.pdf}
\caption{To test the ability of the neural ODEs to learn the limit cycle of the Van der Pol oscillator, the various models were integrated forward in time from initial conditions $x=4$ and $y=\frac{dx}{dt}=4$. The conventional neural ODE with tanh activation functions did not learn the limit cycle from the data, whereas the third order polynomial neural ODE successfully did.}
\label{fig:PiNet-Van-der-Pol-LimitCycle}
\end{figure}
Following training, the performance of the neural networks were tested by: (1) integrating the neural ODEs from the training data's initial conditions up to a time value of 80 and (2) plotting the vector fields of the time derivatives of the variables x and y. As shown in Fig \ref{fig:PiNet-Van-der-Pol-Full} and \ref{fig:PiNet-Van-der-Pol-VectorField}, the third order $\pi$-net polynomial neural ODE was able to successfully learn the dynamics of the system. When the equations were recovered from the polynomial neural ODEs, as shown in Fig \ref{fig:PiNet-Van-der-Pol-Full}, the coefficients of the original Van der Pol ODE model were successfully recovered to 7 significant digits.
The conventional neural ODE with tanh activation functions was able to correctly predict the dynamical system's trajectory starting at the training data's initial conditions, as shown in Fig \ref{fig:PiNet-Van-der-Pol-Full}; however, the vector fields in Fig \ref{fig:PiNet-Van-der-Pol-VectorField} demonstrate that it learned slightly different dynamics from the original Van der Pol ODE the data was generated from. The vector field of the traditional neural ODE does not show the limit cycle found in the Van der Pol ODE. We can show this by integrating the neural ODEs and Van der Pol ODE forward in time starting at a different set of initial conditions. Fig \ref{fig:PiNet-Van-der-Pol-LimitCycle} integrates the ODEs forward in time starting at $x=4$ and $y=\frac{dx}{dt}=4$. It can be seen that the conventional neural ODE fails to learn the limit cycle, whereas the polynomial neural ODE does not have the same issue.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{fig/PiNetV1fig/Van-der-Pol/Grid_VectorField.pdf}
\caption{The vector fields of the time derivatives of variables x and y are plotted for the conventional neural ODE with tanh activations (left), the true ODE equation (middle), and the third order $\pi$-net V1 polynomial neural ode (right). The top row shows the entire training region, while the bottom row shows the region outside of the training region. The blue circles mark the training data.}
\label{fig:PiNet-Van-der-Pol-VectorField}
\end{figure}
\FloatBarrier
\subsection{Learning Other Types of Equations with Polynomial Neural ODEs}
All of our previous demonstrations were on ODEs with equations defined by polynomials, so our last demonstration is on a dynamical system which is not a polynomial:
\begin{eqnarray} \label{eq:sincos}
\frac{dx}{dt} &=& \cos(y) \nonumber \\
\frac{dy}{dt} &=& - \sin(x).
\end{eqnarray}
\noindent with initial conditions $x=0.5$ and $y=1$. Since the problem is nonstiff, we generate our training data by integrating the IVP with SciPy and torchdiffeq using DOPRI5, a fourth order embedded method in the Runge–Kutta family of ODE solvers, with the default settings at 200 points uniformly spaced in time between 0 and 40 \citep{2020SciPy-NMeth, Chen_torchdiffeq_2021, dopriref}. As discussed in the methods section, we batch our data into 199 training samples consisting of IVPs between 2 adjacent known data points, and simultaneously solve the 199 IVPs during each epoch using the fourth order explicit Runge–Kutta–Fehlberg method \citep{fehlberg1968classical}. The following neural networks were trained using the training data: (1) a conventional neural network with 5 layers consisting of 2x50x50x50x2 neurons in each layer with tanh activation functions, and (2) separate $\pi$-net V1 polynomial neural networks of degrees 4, 5, 6, and 15.
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{fig/PiNetV1fig/sincos/Grid_VectorField.pdf}
\caption{The vector fields of the time derivatives of variables x and y are plotted for the conventional neural ODE with hidden layers 2x50x50x50x2 and tanh activations (left), the true ODE equation (middle), and the sixth order $\pi$-net V1 polynomial neural ode (right). The blue circles mark the training data.}
\label{fig:PiNet-sincos-VectorField}
\end{figure}
Following training, the performance of the neural networks were tested by plotting the vector fields of the time derivatives of the variables x and y. As shown in Fig \ref{fig:PiNet-sincos-VectorField}, the polynomial neural ODE was able to learn a better approximation to the true vector field than the conventional neural ODE. While the polynomial neural ODE does not learn a perfect approximation to the field, it is better able to predict a rough approximation of the structure of the true vector field, including where additional spirals in the vector field would be, whereas the conventional neural ODE only learned one spiral.
\FloatBarrier
\section{Discussion/Conclusion}
This work introduced (1) symbolic regression with deep polynomial neural networks and (2) the polynomial neural ODE, which is a deep polynomial neural network implemented into the neural ODE framework. Deep polynomial neural networks directly output a polynomial transformation of the input, which makes them directly interpretable. We are the first to show that symbolic computing can be used to directly recover a symbolic representation of deep polynomial neural networks. We demonstrated successful symbolic regression with a deep polynomial neural network on data generated from a univariate fourth order polynomial. We also successfully demonstrated symbolic regression of dynamical systems governed by ODEs with the polynomial neural ODE on data from the Lotka-Volterra deterministic oscillator, damped oscillatory system, and Van der Pol deterministic oscillator.
We understand that real world experimental data will be noisy, but we have chosen to devote an analysis on experimental noise to a future study. We have a follow up paper on the data requirements for training conventional and polynomial neural ODEs such as noise, sampling frequency, and training size. Additionally, most chemical kinetics systems are stiff ODEs \citep{doi:10.1137/1021001}, which arise from reaction rates that differ by many orders of magnitude, so neural ODEs need to be able to handle stiffness for this application space. This work only shows examples for nonstiff ODEs. Stiff ODEs require special treatment \citep{stiff_neural_ode}, which we will show in a follow up paper.
Traditional deep learning approaches from the field of computer science can be used on scientific problems, but this paper makes the case for developing deep learning techniques specifically tailored for scientific applications. Rather than using "black box" data-driven approaches to describe physical phenomenon, we should be creating a suite of mechanistic data-driven approaches. The polynomial neural ODE is one such approach; however, mathematical models usually have additional types of functions such as trigonometric functions and exponential functions. The scientific machine learning community will need to work on developing more interpretable neural network architectures including more complicated functions such as these.
\section{Acknowledgements}
The authors acknowledge research funding from National Institutes of Health (NIH) NIBIB Award No. 2-R01-EB014877-04A1. Use was made of computational facilities purchased with funds from the National Science Foundation (CNS-1725797) and administered by the Center for Scientific Computing (CSC). The CSC is supported by the California NanoSystems Institute and the Materials Research Science and Engineering Center (MRSEC; NSF DMR 1720256) at UC Santa Barbara.
This work was supported in part by NSF awards CNS-1730158, ACI-1540112, ACI-1541349, OAC-1826967, OAC-2112167, CNS-2120019, the University of California Office of the President, and the University of California San Diego's California Institute for Telecommunications and Information Technology/Qualcomm Institute. Thanks to CENIC for the 100Gbps networks. The content of the information does not necessarily reflect the position or the policy of the funding agencies, and no official endorsement should be inferred.
\FloatBarrier
|
\section{Introduction}
Quantum coherence plays a significant role in quantum theory, which has a number of applications in quantum optics, quantum information processing, nanoscale thermodynamics and biological systems \cite{MMTC,SMP,FGM,FGSL,MJJ,YYZW,MBS,DZLX,KDWZ,ZTYH,DGGG,TMG,CLYG,KPF,MKT}. Recently, it has been recognized that coherence can be treated as one kind of quantum resources. Therefore, detecting coherence is crucial in quantum physics.
In Ref. \cite{TGM}, quantitative investigations of quantum coherence have been launched, and several coherence measures for standard coherence, with respect to von Neumann measurements, have been proposed. For standard coherence, consider a $d$-dimensional Hilbert space $\mathcal{H}$, a state $\delta$ is incoherent under a chosen reference basis $\{|i\rangle\}_{i=1}^{d}$ if and only if $\delta$ is diagonal under the reference basis \cite{TGM}, i.e.
\begin{equation}\label{}
\delta =\sum_{i=1}^{d} p_i|i\rangle\langle i|,
\end{equation}
with probabilities $\{p_i\}$. One can define the standard dephasing operation $\Delta$ as
\begin{equation}\label{de}
\Delta(\rho):=\sum_{i=1}^{d}|i\rangle\langle i|\rho|i\rangle\langle i|.
\end{equation}
Thus, a state $\delta$ is incoherent under a chosen reference basis $\{|i\rangle\}_{i=1}^{d}$ if and only if the following condition holds,
\begin{equation}\label{}
\delta=\Delta(\delta),
\end{equation}
where $\Delta$ is defined in Eq. (\ref{de}).
In Ref. \cite{Aberg}, {\AA}berg proposed a framework which actually defined block coherence with respect to projective measurements. The standard coherence can be viewed as a special case of block coherence. Moreover, Bischof \textit{et al.} generalized block coherence to coherence with respect to the most general quantum measurements \cite{FHD}, i.e., the positive-operator-valued measure (POVM)-based coherence. Therefore, the resource theory of coherence was generalized by extending the standard coherence to block coherence and even POVM-based coherence. It is worth noticing that POVMs describe the most general type of quantum measurements, and they may be more advantageous compared to von Neumann measurements.
However, unlike the standard coherence, there are only a few results reported on quantifying block coherence and POVM-based coherence \cite{FHD,CMWR,QCWRT}. The methods to detect whether a state has nonzero block coherence and POVM-based coherence are missing. For the standard coherence, Ref. \cite{CTRM} first introduced the standard coherence witness $W$. Similar to entanglement witnesses, a Hermitian operator $W$ is a standard coherence witness if $\tr(\delta W)\geq0$ holds for all incoherent states $\delta$. If one finds $\tr(\rho W)<0$ for a state $\rho$, then the state $\rho$ must be a standard coherent state. Compared with coherence measures, which usually need full information of the state by using quantum state tomography, coherence witnesses can be measured with much less measurements since quantum state tomography requires exponentially growing measurements with the number of qubits.
Thus, it is necessary to construct block coherence witnesses and POVM-based coherence witnesses to detect them without quantum state tomography, especially for experimentally unknown states.
The purpose of this work is two-fold. On the one hand, we present witnesses for block coherence and POVM-based coherence, and obtain the necessary and sufficient conditions for arbitrary block coherence witnesses and POVM-based coherence witnesses. Moreover, we discuss possible realizations of some block coherence and POVM-based coherence witnesses in real experiments, and present examples to detect block coherence coherence by measuring block coherence witnesses. On the other hand, we provide an application of block coherence witnesses in a quantum parameter estimation task with a degenerate Hamiltonian, and one can estimate the unknown parameter by measuring our block coherence witnesses if the input state is block coherent. We also prove that the quantum Fisher information of any block incoherent state is equal to zero, which coincides with the result from measuring block coherence witnesses.
\section{Detecting block coherence based on block coherence witnesses}
Before embarking on our main results, let us first review the definition of block incoherent states.
In Refs. \cite{FHD,CMWR,QCWRT,CCBP,Aberg,WJD}, block incoherent state has been defined as follows.
Given a $d$-dimensional Hilbert space $\mathcal{H}$, which has been divided into $n$ ($n\leq d$) subspaces, and subspace projectors are $\mathbf{P}:=\{P_s\}_{s=1}^{n}$ with $\sum_{s=1}^n P_s=\idol$ (where $\idol$ is the identity operator). A state $\widetilde{\delta}$ is block incoherent under the reference subspace projectors $\mathbf{P}$, if and only if $\widetilde{\delta}$ is block diagonal under the reference $\mathbf{P}$, i.e,
\begin{equation}\label{}
\widetilde{\delta}=\sum_{s=1}^{n}P_{s}\widetilde{\delta} P_{s}:=\widetilde{\Delta}(\widetilde{\delta}),
\end{equation}
where we define the modified dephasing operation as
\begin{equation}
\widetilde{\Delta}(\rho) := \sum_{s} P_s \rho P_s.
\end{equation}
Similar to the witnesses for standard coherence \cite{CTRM,ZZY}, we can construct block coherence witnesses as follows
\textbf{Theorem 1.} (a) For any Hermitian operator $A$, we can construct a block coherence witness
\begin{equation}\label{10}
\widetilde{W}_A=\widetilde{\Delta}(A)- A.
\end{equation}
(b) An arbitrary Hermitian operator $\widetilde{W}$ is a block coherence witness if and only if $\widetilde{\Delta}(\widetilde{W})\geq 0$.
\textbf{Proof.---} (a) We first prove that $\widetilde{W}_A$ is a block coherence witness. Since $A$ is a Hermitian operator, $\widetilde{W}_A$ must be Hermitian as well. Thus, for an arbitrary block incoherent state $\widetilde{\delta}=\sum_{s}P_{s}\widetilde{\delta} P_{s}$, we can obtain
\begin{eqnarray}
\mathrm{Tr}(\widetilde{\delta} \widetilde{W}_A) &=& \mathrm{Tr}[\widetilde{\delta}\widetilde{\Delta}(A)]-\tr[\widetilde{\delta}A] \nonumber\\
&=& \mathrm{Tr}[\widetilde{\delta}\sum_{s} P_{s}A P_{s}] -\tr[\widetilde{\delta}A ] \nonumber\\
&=& \mathrm{Tr}[\sum_{s}P_{s}\widetilde{\delta}P_{s}A] - \tr[\widetilde{\delta}A] \nonumber\\
&=& 0,
\end{eqnarray}
which means $\widetilde{W}_A$ is a block coherence witness.
(b) It is worth noticing that a Hermitian operator $W$ is a coherence witness for standard coherence if and only if $\Delta(W) \geq 0$ \cite{CTRM}. Similarly, we can prove that a Hermitian operator $\widetilde{W}$ is a block coherence witness if and only if $\widetilde{\Delta}(\widetilde{W})\geq 0$.
Firstly, if $\widetilde\Delta(\widetilde{W})\geq 0$ holds, then for any block incoherent state $\widetilde{\delta}$ we can obtain that
\begin{eqnarray}
\tr[\widetilde{\delta} \widetilde{W}] &=&\tr[\widetilde\Delta(\widetilde{\delta}) \widetilde{W}]\nonumber \\
&=&\tr[\widetilde\Delta(\widetilde{W}) \widetilde{\delta}]\nonumber \\
& \geq& 0,
\end{eqnarray}
i.e., $\widetilde{W}$ is a block coherence witness.
Conversely, we prove that if for any block incoherent state $\widetilde{\delta}$ $\tr[\widetilde{\delta} \widetilde{W}]\geq 0$ holds, then $\widetilde\Delta(\widetilde{W})\geq 0$. For any quantum state $\rho$, one can obtain
\begin{eqnarray}
\tr[\rho\widetilde{\Delta}( \widetilde{W})]
&=&\tr[\widetilde\Delta(\rho) \widetilde{W}] \nonumber \\
&=&\tr[\widetilde{\delta}_{\rho}\widetilde{W}] \nonumber \\
& \geq& 0,
\end{eqnarray}
where $\widetilde{\delta}_{\rho}:=\widetilde\Delta(\rho)$ is a block incoherent state.
Thus, $ \widetilde{\Delta}(\widetilde{W})$ is positive-semidefinite, i.e., $\widetilde\Delta(\widetilde{W})\geq 0$.
Therefore, it can be given that $\widetilde{W}$ is block coherence witness if and only if $\widetilde{\Delta}(\widetilde{W})\geq 0$. \hfill $\square$
\textbf{Remark.---} Based on Theorem 1(a), one can construct a block coherence witness $\widetilde{W}_{\sigma}$ by using any density matrix $\sigma$ as the Hermitian operator $A$ in Eq. (\ref{10}),
\begin{equation}\label{15}
\widetilde{W}_{\sigma} = \widetilde{\Delta}(\sigma) - \sigma.
\end{equation}
Moreover, if $\sigma$ is a pure state $|\phi\rangle$, we can obtain that
\begin{equation}\label{pureW1}
\widetilde{W}_{\phi} = \widetilde{\Delta}(|\phi\rangle\langle\phi|) -|\phi\rangle\langle\phi|,
\end{equation}
and
\begin{eqnarray}
\mathrm{Tr}[\rho \widetilde{W}_{\phi}]&=&\mathrm{Tr}[\rho(\widetilde{\Delta}(|\phi\rangle\langle\phi|) -|\phi\rangle\langle\phi|)]\nonumber \\
&=&\langle \phi|\widetilde\Delta(\rho)|\phi\rangle-\langle \phi|\rho|\phi\rangle \nonumber\\
&=&F(\widetilde\Delta(\rho),|\phi\rangle)-F(\rho,|\phi\rangle),
\end{eqnarray}
where $F(\rho,|\phi\rangle):=\langle \phi|\rho|\phi\rangle$ is the fidelity between the state $\rho$ and the pure state $|\phi\rangle$. Therefore, the expect value of block coherence witness $ \widetilde{W}_{\phi}$ is related to the two fidelities.
\section{Coherence witness with respect to general measurements}
Recently, Bishof $et \ al.$ introduced a generalization to a resource theory of coherence with respect to the general quantum measurement. The POVM-based coherence is defined as follows \cite{FHD,CMWR,QCWRT}.
Let $\mathbf{E}$ be a collection of $n$ positive operators $\mathbf{E}:=\{E_{i}\}_{i=1}^{n}$ with $\sum_{i=1}^{n} E_{i}=\idol$. The corresponding measurement operator of each $E_i$ is denoted by $A_i$, such that $E_{i}=A_{i}^{\dagger}A_{i}$ holds. Thus, a state $\bar{\delta}$ is called incoherent state with respect to the general measurement $\mathbf{E}$ if and only if
\begin{equation}
E_{i}\bar{\delta} E_{i'}=0, \ \forall i\neq i'.
\end{equation}
Note that it is equivalent to \cite{FHD}
\begin{equation}
A_{i}\bar{\delta} A_{i'}^{\dagger}=0, \ \forall i\neq i'.
\end{equation}
Therefore, any POVM-based incoherent state $\bar{\delta}$ should satisfy
\begin{equation}\label{bardelta}
\bar{\delta}= \sum_{i}E_{i}\bar{\delta} E_{i}:=\bar{\Delta}(\bar{\delta}),
\end{equation}
where $\bar{\Delta}$ is defined as
\begin{equation}
\bar{\Delta}(\rho) := \sum_{i} E_i \rho E_i.
\end{equation}
It is worth noticing that Eq. (\ref{bardelta}) can be easily proved from the definition of POVM-based incoherent state, since for any POVM-based incoherent state $\bar{\delta}$, one can obtain
\begin{eqnarray}
\bar{\delta} &=&(\sum_{i}E_{i})\bar{\delta} (\sum_{j}E_{j}) \nonumber \\
&=&\sum_{i} E_{i}\bar{\delta} E_{i} +\sum_{i\neq j} E_{i}\bar{\delta} E_{j}\nonumber \\
&=&\sum_{i}E_{i}\bar{\delta} E_{i},
\end{eqnarray}
where we have used $E_{i}\bar{\delta} E_{j}=0, \forall i\neq j$.
\textbf{Theorem 2.} (a) For any Hermitian operator $A$, we can construct a POVM-based coherence witness $\bar{W}_A$ as follows,
\begin{equation}\label{}
\bar{W}_A=\bar{\Delta} (A)- A.
\end{equation}
(b) An arbitrary Hermitian operator $\bar{W}$ is a POVM-based coherence witness if and only if $\bar{\Delta}(\bar{W})\geq 0$.
\textbf{Proof.---} (a) We provide that $\bar{W}_A$ is a POVM-based coherence witness. As $A$ is a Hermitian operator, $\bar{W}_A$ must be
Hermitian. For any incoherent state $\bar{\delta}$ with respect to $\{E_{i}\}$, $\bar{\delta}=\sum_{i}E_{i}\bar{\delta} E_{i}$ holds and thus
\begin{eqnarray}
\mathrm{Tr}(\bar{\delta} \bar{W}_A) &=& \mathrm{Tr}[\bar{\delta}\bar{\Delta} (A)]-\tr[\bar{\delta} A] \nonumber \\
&=& \mathrm{Tr}[\bar{\delta}\sum_{i} E_{i}A E_{i}] - \tr[\bar{\delta} A ] \nonumber \\
&=& \mathrm{Tr}[\sum_{i}E_{i}\bar{\delta} E_{i}A] - \tr[\bar{\delta} A] \nonumber \\
&=& 0,
\end{eqnarray}
which means $\bar{W}_A$ is a POVM-based coherence witness.
(b) Firstly, if $\bar{\Delta}(\bar{W})\geq 0$ holds, then for any POVM-based incoherent state $\bar{\delta}$ we can obtain that
\begin{eqnarray}
\tr[\bar{\delta} \bar{W}] &=&\tr[\bar{\Delta} (\bar{\delta}) \bar{W}] \nonumber \\
&=&\tr[\bar{\Delta}(\bar{W}) \bar{\delta}] \nonumber \\
& \geq& 0.
\end{eqnarray}
Thus, $\bar{W}$ is a POVM-based coherence witness.
Conversely, we prove that if $\tr[\bar{\delta} \bar{W}]\geq 0$, then $\bar{\Delta}(\bar{W})\geq 0$. For any quantum state $\rho$, one can obtain that
\begin{eqnarray}
\tr[\rho\bar{\Delta}(\bar{ W})]
&=&\tr[\bar{\Delta}(\rho) \bar{W}] \nonumber \\
&=&\tr[\bar{\delta}_{\rho}\bar{W}] \nonumber \\
& \geq& 0,
\end{eqnarray}
where $\bar{\delta}_{\rho}:=\bar{\Delta}(\rho)$ is a POVM-based incoherent state. Thus, $\bar{\Delta}(\bar{W})$ is positive-semidefinite, i.e., $\bar{\Delta}(\bar{W})\geq 0$.
Therefore, we proved that $\bar{W}$ is a POVM-based coherence witness if and only if $\bar{\Delta}(\bar{W})\geq 0$. \hfill $\square$
\textbf{Remark.---} Based on Theorem 2(a), we can also construct a POVM-based coherence witness $\bar{W}_{\sigma}$ by choosing any density matrix $\sigma$ as the Hermitian operator $A$,
\begin{equation}\label{43}
\bar{W}_{\sigma} = \bar{\Delta}({\sigma}) - \sigma.
\end{equation}
Moreover, if $\sigma$ is a pure state $|\phi\rangle$, one can obtain that
\begin{equation}\label{pureW2}
\bar{W}_{\phi} = \bar{\Delta}(|\phi\rangle\langle\phi|) - |\phi\rangle\langle\phi|,
\end{equation}
and
\begin{eqnarray}
\mathrm{Tr}[\rho \bar{W}_{\phi}]&=&\mathrm{Tr}[\rho(\bar{\Delta}(|\phi\rangle\langle\phi|) - |\phi\rangle\langle\phi|)] \nonumber\\
&=&\tr[\rho \bar{\Delta}(|\phi\rangle\langle\phi|)] -\tr[\rho |\phi\rangle\langle\phi|] \nonumber\\
&=&F( \bar{\Delta}(\rho), |\phi\rangle)-F(\rho,|\phi\rangle),
\end{eqnarray}
where it demonstrates the relationship of the POVM-based coherence witness $\bar{W}_{\phi}$ and fidelities between the state $\rho$ (or $\bar\Delta(\rho)$) and the pure state $|\phi\rangle$.
\section{Possible experimental realization for witnesses and examples}
Many experiments have measured fidelities $F=\langle\phi|\rho_{exp}|\phi\rangle$ between the experimental state $\rho_{exp}$ and a target pure state $|\phi\rangle$ \cite{ion,Nph1,Nph2,Nph3}. For bipartite and multipartite systems, one possible way to measure fidelities is to decompose the operator $|\phi\rangle\langle \phi|$ as sum of tensor products of local observables \cite{Nph1,Nph2,Nph3}. Therefore, one can measure our witnesses (\ref{pureW1}) and (\ref{pureW2}) in the same manner.
In the following, we will present examples of $N$-qubit W states $|W_N\rangle$ from real experimental data, where
\begin{equation}\label{W_N}
|W_N\rangle=(|0\cdots 001\rangle+|0\cdots 010\rangle+ \cdots +|10\cdots0\rangle)/\sqrt{N}.
\end{equation}
We use the block coherence witness (\ref{pureW1}) to detect block coherence of W states. In Ref. \cite{ion}, $N$-qubit W states ($4\leq N\leq8$) have be experimentally generated by trapped ions, with fidelities between experimental states and perfect W states being $F_4=0.846$, $F_5=0.759$, $F_6=0.788$, $F_7=0.763$, $F_8=0.722$ for the $4$-, $5$-, $6$-, $7$- and $8$-ion W states, respectively. Moreover, numerical values of the density matrices of experiment states with $4\leq N\leq8$ have been presented in Ref. \cite{ion}. It is worth noticing that the experimental states have local phases, and one can find the local phases by maximizing the fidelities $F=\langle \tilde{W}_N|\rho_{exp}|\tilde{W}_N\rangle$ with $|\tilde{W}_N\rangle$ being W states containing local phases shown in \cite{ion}. After choosing local unitary transformations based on local phases, we can transform $\rho_{exp}$ to $\rho'_{exp}$ such that $F=\langle \tilde{W}_N|\rho_{exp}|\tilde{W}_N\rangle=\langle W_N|\rho'_{exp}|W_N\rangle$. We consider the following reference subspace projectors $\mathrm{\mathbf{P}}$ with
\begin{eqnarray}
P_0&=&|\phi^-\rangle\langle \phi^-|,\nonumber\\
P_1&=&|\phi^+\rangle\langle \phi^+|,\nonumber\\
P_2&=&|0\cdots 010\rangle\langle 0\cdots010|,\nonumber\\
&&\cdots,\nonumber\\
P_{N-1}&=&|010\cdots0\rangle\langle 010\cdots0|,\nonumber\\
P_N&=&\idol-\sum_{i=0}^{N-1}P_{i},\nonumber
\end{eqnarray}
where $|\phi^{\pm}\rangle:=(|00\cdots 01\rangle\pm |10\cdots0\rangle)/\sqrt{2}$. Thus, our block coherence witness is as follows,
\begin{eqnarray}
\widetilde{W}&=&\widetilde{\Delta}(|W_N\rangle\langle W_N|)- |W_N\rangle\langle W_N|\nonumber\\
&=&\sum_{i=0}^N P_i |W_N\rangle\langle W_N| P_i -|W_N\rangle\langle W_N|.
\end{eqnarray}
From Table \ref{lab1}, we can see that $-\tr(\rho'_{exp} \widetilde{W})$ is always greater than zero, which means all the experimental states in Ref. \cite{ion} contain the block coherence under the above reference subspace projectors $\mathrm{\mathbf{P}}$.
\begin{table}
\begin{tabular}{c|ccccc}
\hline
\hline
$|W_N\rangle$ & $4$ & $5$ & $6$ & $7$& $8$ \\
\hline
$F_{N}$ & $ 0.846 $ \ \ & $ 0.759 $ \ \ & $ 0.788 $\ \ & $ 0.763 $\ \ & $ 0.722 $ \\
\hline
$\langle \widetilde{\Delta}(|W_N\rangle\langle W_N|) \rangle$ & $0.321$ \ \ & $0.207$\ \ & $0.173$\ \ & $0.141$\ \ & $0.119$ \\
\hline
$-\tr(\rho'_{exp} \widetilde{W}) $ &0.525 \ \ &0.552 \ \ &0.615\ \ &0.622\ \ &0.603 \\
\hline
\hline
\end{tabular}
\caption{The fidelity $F_{N}$ is the result of overlap $\langle W_N |\rho'_{exp}|W_N\rangle$ from Ref. \cite{ion}. The values $\langle \widetilde{\Delta}(|W_N\rangle\langle W_N|) \rangle$ and $-\tr(\rho'_{exp} \widetilde{W})$ are obtained from the density matrices of experiment states in Ref. \cite{ion}. One can see that $-\tr(\rho'_{exp} \widetilde{W})$ is always greater than zero. }\label{lab1}
\end{table}
\section{Quantum parameter estimation task with degenerate Hamiltionians}
For quantum metrology, one of the main tasks is parameter estimation in a quantum channel to improve accuracy that is different from the standard quantum limits \cite{VGS,JPD}. Quantum coherence plays a fundamental role in quantum parameter estimation \cite{GTI}. For the case of unitary evolution with a degenerate Hamiltonian, we will propose a simple application of quantum block coherence, and find that the quantum Fisher information is strongly related to the block coherence.
\subsection{Quantum parameter estimation by using block coherent states}
Now we consider a $d$-dimensional Hilbert space, suppose that $H$ is a degenerate Hamiltonian,
\begin{equation}\label{1}
H=\sum_{s=1}^{n}\sum_{g=1}^{k_s}E_s|s,g \rangle\langle s,g|,
\end{equation}
where $H$ has $n$ different eigenvalues $\{E_s\}_{s=1}^n$, and each eigenvalue $E_s$ has $k_s$ degenerate eigenstates $\{|s,g\rangle\}_{g = 1}^{k_s}$. Here the index $s$ is for the $s$-th different eigenvalue, and the index $g$ is denoted as the $g$-th $E_s$ (the total number of the same eigenvalue $E_s$ is $k_s$). $|s,g\rangle$ is the eigenstate corresponding to the eigenvalue, i.e., the $g$-th $E_s$. According to the definition of block coherence, one can naturally choose the degenerate subspaces of $H$ in Eq. (\ref{1}) as the reference subspaces. Therefore,
\begin{equation}\label{Ps}
P_s = \sum_{g = 1}^{k_s} |s,g\rangle\langle s,g|,
\end{equation}
is the $s$-th subspace projector, where every pure state in this subspace is an eigenstate of $H$ with the eigenvalue being $E_s$.
\textbf{Proposition 3.} With a degenerate Hamiltonian $H$ in Eq. (\ref{1}), we choose $\{P_s\}_{s=1}^n$ in (\ref{Ps}) as the reference subspace projectors $\mathrm{\mathbf{P}}$. The output state $\rho_{out}=U_{\varphi}\rho_{in}U_{\varphi}^{\dag}:=\rho_{\varphi}$ can be use to estimate the unknown parameter $\varphi$ in the black box in Fig. \ref{Fig1}, if and only if $\rho_{in}$ and $\rho_{out}$ have nonzero block coherence under the reference subspaces $\mathrm{\mathbf{P}}$.
\textbf{Proof.---} For an arbitrary input state $\rho_{in}$, it can be expressed under the eigenstates of $H$ (\ref{1}), i.e.,
\begin{equation}
\rho_{in} =\sum_{s,s'} \sum_{g,g'}\rho_{(s,g),(s',g')}|s,g\rangle\langle s',g'|,
\end{equation}
where we used $\rho_{(s,g),(s',g')} := \langle s,g|\rho_{in}|s',g'\rangle$ with $\sum_{s,g}\rho_{(s,g),(s,g)} = 1$. Thus,
the corresponding output state $\rho_{out} = U_{\varphi}\rho_{in} U_{\varphi}^{\dag}:=\rho_{\varphi}$ can be expressed as follows,
\begin{eqnarray}
\rho_{out}&=& U_{\varphi}\rho_{in} U_{\varphi}^{\dag}\nonumber\\
&=&\sum_{\substack{s,s'\\s\neq s'}}\sum_{g,g'}\rho_{(s,g),(s',g')}e^{-i(E_{s} - E_{s'})\varphi}|s,g\rangle\langle s',g'|\nonumber\\
&& + \sum_{s}\sum_{g,g'}\rho_{(s,g),(s,g')}|s,g\rangle\langle s,g'|.
\end{eqnarray}
We can see that $\rho_{out}$ is dependent on $\varphi$ if and only if there exists nonzero $\rho_{(s,g),(s',g')}$ with $s\neq s'$, i.e., $\rho_{out}$ and $\rho_{in}$ have nonzero block coherence under the reference subspaces. \hfill $\square$
\begin{figure}
\includegraphics[scale=1.3]{Fig1.eps}
\caption{The black box implements an unitary evolution $U_{\varphi}=e^{-iH\varphi}$ on the input state $\rho_{in}$, and the unknown parameter $\varphi$ of the black box must be estimated by measuring $\rho_{out}=U_{\varphi}\rho_{in}U_{\varphi}^{\dag}:=\rho_{\varphi}$. Suppose that the Hamiltonian $H$ is degenerate and already known, the quantum parameter estimation task is estimating the unknown parameter $\varphi$ from the output state $\rho_{out}$.}\label{Fig1}
\end{figure}
\textbf{Remark.---} It is worth noticing that witnesses for standard coherence may be useless to estimate the unknown parameter when the Hamiltonian $H$ is degenerate, although the input state has standard coherence. Consider a special case, if we use an input state which is block incoherent, but its density matrix contains off-diagonal nonzero elements in some degenerate subspaces under a chosen eigenvectors of $H$ as reference basis, i.e., this input state contains standard coherence but no block coherence. The expect values of standard coherence witnesses has no information of the parameter, and one cannot estimate it by measuring witnesses of standard coherence, although the input state contains standard coherence.
\subsection{Quantum Fisher information of block incoherent states}
For an arbitrary output state $\rho_{\varphi}$, the quantum Fisher information $F_q$ can be obtained \cite{SCG,SCG2,MGA,VGSLM}
\begin{eqnarray}
F_q&=&\tr[\rho_{\varphi}L_{\varphi}^{2}]\nonumber\\
&=&\sum_{m,n}4c_{m}(\frac{c_{n}-c_{m}}{c_{n}+c_{m}})^2 |\langle m|H|n\rangle|^2,
\end{eqnarray}
where $U_{\varphi}=e^{-iH\varphi}$ is an unitary operator and $H$ is the corresponding Hermitian Hamiltonian, and the output state $\rho_{\varphi}=U_{\varphi}\rho_{in}U_{\varphi}^{\dagger}=U_{\varphi}(\sum_{n}c_{n}|n\rangle\langle n|)U_{\varphi}^{\dagger}$ with $\rho_{in}=\sum_{n}c_{n}|n\rangle\langle n|$. $\{c_{n}\}$ and $\{|n\rangle\}$ are the eigenvalues and eigenvectors of $\rho_{in}$, respectively. The symmetric logarithmic derivative operator is $L_{\varphi}=U_{\varphi}(-2i\sum_{m,n}\frac{\langle m|[H,\rho_{in}]|n\rangle}{c_{n}+c_{m}}|m\rangle\langle n|)U_{\varphi}^{\dagger}$ \cite{SCG,SCG2,MGA,VGSLM}.
Furthermore, we will discuss the quantum Fisher information with a block incoherent state.
\textbf{Proposition 4.} Consider a degenerate Hamiltonian $H$ in Eq. (\ref{1}) and the quantum parameter estimation task in Fig. \ref{Fig1}, one can choose the degenerate subspaces of $H$ as the reference subspaces. If the input state $\rho_{in}$ is a block incoherent state, then the quantum Fisher information for the output state is $F_q=0$.
\textbf{Proof.---} Consider an arbitrary block incoherent state as the input state $\rho_{in}$,
\begin{eqnarray}
\rho_{in}&=&\sum_{i}\sum_{g,g'}\rho_{(i,g),(i,g')}|i,g\rangle\langle i,g'|\nonumber\\
&=&\sum_{i}\sum_{\tilde{g}}A_{\tilde{g}}^{(i)}|i,\tilde{g}\rangle\langle i,\tilde{g}|,
\end{eqnarray}
where the second equation holds since we diagonalize $\rho_{in}$ in each subspace. Thus, the eigenvalues of $\rho_{in}$ are $\{A_{\tilde{g}}^{(i)}\}$ with corresponding eigenvectors $\{|i,\tilde{g}\rangle\}$. The output state $\rho_{\varphi}$ can be expressed as
\begin{eqnarray}
\rho_{\varphi}&=&U_{\varphi}\rho_{in}U_{\varphi}^{\dagger}\nonumber\\
&=&U_{\varphi}\sum_{i}\sum_{\tilde{g}}A_{\tilde{g}}^{(i)}|i,\tilde{g}\rangle\langle i,\tilde{g}|U_{\varphi} ^{\dagger},
\end{eqnarray}
where $U_{\varphi}=e^{-i H\varphi}$ is an unitary operator. Furthermore, the symmetric logarithmic derivative operator $L_{\varphi}$ can be expressed as \cite{SCG,SCG2,MGA,VGSLM},
\begin{eqnarray}
L_{\varphi}
&=&U_{\varphi}(-2i\sum_{i,j}\sum_{\tilde{g},\tilde{h}}\frac{\langle i,\tilde{g}|[H,\rho_{in}]|j,\tilde{h}\rangle}{A_{\tilde{g}}^{i}+A_{\tilde{h}}^{(j)}}|
i,\tilde{g}\rangle\langle j,\tilde{h}|)U_{\varphi}^{\dagger}\nonumber\\
&=&U_{\varphi}\Bigg(-2i\sum_{i,j}\sum_{\tilde{g},\tilde{h}}\frac{A_{\tilde{h}}^{(j)}-
A_{\tilde{g}}^{(i)}}{A_{\tilde{g}}^{(i)}+A_{\tilde{h}}^{(j)}}\langle i,\tilde{g}|H|j,\tilde{h}\rangle |i,\tilde{g}\rangle\langle j,\tilde{h}|\Bigg)U_{\varphi}^{\dagger}.\nonumber
\end{eqnarray}
Finally, we obtain the quantum Fisher information $F_q$ as
\begin{eqnarray}
F_q&=&\tr(\rho_{\varphi}L_{\varphi}^{2})\nonumber\\
&=&4\sum_{i,j}\sum_{\tilde{g},\tilde{h}}A_{\tilde{g}}^{(i)}\big|\langle i,\tilde{g}|H|j,\tilde{h}\rangle\big|^{2}\Bigg(\frac{A_{\tilde{h}}^{(j)}-A_{\tilde{g}}^{(i)}}
{A_{\tilde{g}}^{(i)}+A_{\tilde{h}}^{(j)}}\Bigg)^2\nonumber\\
&=&4\sum_{\substack{i,j \\i\neq j}}\sum_{\tilde{g},\tilde{h}}A_{\tilde{g}}^{(i)}\big|\langle i,\tilde{g}|H|j,\tilde{h}\rangle\big|^{2}\Bigg(\frac{A_{\tilde{h}}^{(j)}-A_{\tilde{g}}^{(i)}}
{A_{\tilde{g}}^{(i)}+A_{\tilde{h}}^{(j)}}\Bigg)^2 \nonumber\\
&&+4\sum_{i}\sum_{\tilde{g},\tilde{h}}A_{\tilde{g}}^{(i)}\big|\langle i,\tilde{g}|H|i,\tilde{h}\rangle\big|^{2}\Bigg(\frac{A_{\tilde{h}}^{(i)}-A_{\tilde{g}}^{(i)}}
{A_{\tilde{g}}^{(i)}+A_{\tilde{h}}^{(i)}}\Bigg)^2 \nonumber\\
&=&0,\label{35}
\end{eqnarray}
where we have used $\langle i,\tilde{g}|H|j,\tilde{h}\rangle= 0$ with $i \neq j$, and $\langle i,\tilde{g}|H|i,\tilde{h}\rangle=E_i \delta_{\tilde{g}\tilde{h}}$. \hfill $\square$
\textbf{Remark.---} When the input state is a block incoherent state, the quantum Fisher information of the output state is equal to zero, which means one cannot estimate the parameter from the output state, i.e., the output state contains no information of the parameter. That coincides with the result from Proposition 3.
\section{Discussions and conclusions}
Besides the Fisher information estimation mentioned in the above example, when performing quantum coherence operations on any degenerate quantum system, we also need to take into account the measure of coherence. In the practical systems, degeneracy is a very common situation and has been widely reported in the microscopic field, e.g., orbital degenerate states and degenerate spin states in $\lambda$-type and the chainlike structures atomic system \cite{EXP1}, near-degenerate states in few-electron ions system \cite{EXP2}, etc. Moreover, degeneracy also exists in macroscopic systems, e.g., macroscopically degenerate in intrinsic quasi-crystals \cite{EXP3}, and been regarded to be appeared during the conversion or transition of black holes \cite{EXP4}. In addition, degeneracy plays an important role in both quantum computing and quantum networks construction \cite{EXP5,EXP6}. It is often used as a protected qubit in fault-tolerant quantum operation \cite{EXP5}. And it can also be used as a control device to control the coherence between quantum systems \cite{EXP7}. In the discussion of topological structure and phase transition, the study of degeneracy and near-degeneracy also plays an important role \cite{EXP8,EXP9,EXP10}. These studies related to the issue of how to correctly measure the coherence of degenerate quantum systems. Therefore, our coherence detection scheme has potential value in the measurement and application of degenerate states.
In conclusion, we have discussed witnesses for block coherence and POVM-based coherence. The necessary and sufficient conditions for arbitrary block coherence and POVM-based coherence witnesses have been obtained. Furthermore, we have shown that block coherent states can be used in a quantum parameter estimation task with a degenerate Hamiltonian if the input state is block coherent. The quantum Fisher information of block incoherent states is equal to zero, which coincides with the result from the block coherence witness.
\section*{ACKNOWLEDGMENTS}
This work is supported by the National Natural Science Foundation of China (Grants No. 11734015, No. 11704205, and No. 12074206), the open funding program from State Key Laboratory of Precision Spectroscopy (East China Normal University), and K.C. Wong Magna Fund in Ningbo University.
|
\section{Introduction}
Ad hoc teamwork (AHT) is a step towards general multi-agent intelligence where decision-making agents (human and computational) are trained to collaborate with other previously unseen agents. That is, AHT assumes that each agent will learn to collaborate without prior coordination or direct control of the other agents \cite{mirsky2022survey}. Although the challenge of AHT was first formalized by \citet{stone2010ad}, there has been a recent resurgence of AHT research that incorporates advances in deep reinforcement learning (RL).
Established works have focused on multi-agent consensus among a network of agents typically with centralized training and decentralized execution, where a fixed set of agents are known to all \cite{lowe2017multi, rashid2018qmix}. For example, this may be useful in real-world infrastructure protection settings where a team of collaborative defense agents might be tasked with securing an asset of interest. This might not be practical, however, in dynamic, adversarial, and contested environments. In these settings, new agents may need to enter the environment and start competing or collaborating without access to a centralized policy. Even if we assume such a policy, it would likely lead to brittle agents that are tethered to their training partners for optimal performance \cite{li2019robust}. This leads the group to be unable to perform optimally in the presence of adversaries. One appeal of AHT is the potential for multiple people or organizations to train agents that collaborate with each other without the need to retrain with previously unseen agents. Outside of \citet{shafipour2021task}, AHT in the presence of adversaries has not been explored due to the assumption of full cooperation between agents. Adversarial machine learning (ML) researchers have an opportunity to chart the next path of truly robust multi-agent learning. In this paper, we contribute (1) a summary of past research in AHT, (2) why it could be a necessary next step in high-stakes, real-world multi-agent RL, (3) a list of some challenges in AHT in the presence of adversaries, and (4) a proposal for some new directions for researchers in both adversarial ML and AHT.
\section{Ad Hoc Teamwork}
\subsection{Background}
AHT is a subfield of cooperative AI, which is defined as AI methods that help individuals, humans and machines, to find ways that improve joint welfare \cite{dafoe2020open}. As mentioned previously, AHT methods train agents to collaborate without prior coordination or direct control of other agents. In this paper, AHT will mostly be synonymous with RL-based methods in AHT, called zero-shot coordination (ZSC). The main reason is because ZSC has enjoyed some recent success. These successes include a general and differentiable information theoretic objective for training a diverse population of optimal policies \cite{lupu2021trajectory}, training methods that prevent agents from exchanging information through arbitrary conventions \cite{hu2021off}, and graph neural networks that learn agent models and joint-action value models under varying team compositions \cite{rahman2021towards}. Some approaches even outperform self-play and behavioral cloning when collaborating with humans in video games \cite{strouse2021collaborating}. ZSC is an evolving field and there is currently some disagreement over how to precisely formulate it \cite{treutlein2021new}. We recommend \citet{mirsky2022survey} as an up-to-date survey of AHT.
\subsection{Potential Applications}
While AHT has shown some success in robot soccer \cite{genter2017three}, there are still many potential applications yet to be explored. Ideally, AHT can perform more robustly than centralized multi-agent RL methods, so any multi-agent environment can potentially train an AHT application. In the context of secure cyber-physical infrastructure operations (such as transportation and energy distribution), multi-agent learning and control is essential for achieving system resilience goals \cite{januario2019distributed, hou2021reinforcement, phan2021resilient}. Some future AHT applications worth mentioning include important automated infrastructure, like energy and power distribution \cite{biagioni2021powergridworld, pigott2021gridlearn, wang2021multi}. AHT would be particularly useful here because of the possibility that different buildings or power plants might need to coordinate with previously unseen agents from other allied power plants. Another application would be human-automated vehicle interaction. It is still difficult for automated cars to learn and appropriately react to humans while driving \cite{dommes2021young}.
\section{Potential Challenges with Adversaries in Ad Hoc Teamwork}
We now briefly describe some challenges that agents trained using AHT methods could face unless it is framed as an adversarial ML problem.
\subsection{Past Research in Adversarial RL}
It is possible that some attacks described in past research on adversarial RL \cite{huang2017adversarial, gleave2019adversarial, lin2020robustness, sun2020stealthy, figura2021adversarial, fujimoto2021reward} will have similar negative effects in AHT. Given that there are currently no universally effective defenses against adversarial attacks in supervised learning \cite{athalye2018obfuscated, short2019defending}, the same is likely true for RL.
\subsection{Balancing the Presence of Allies and Adversaries}
Even if we assume adversaries might exist in the environment, it is not clear how to train the agents to collaborate. It might be that the adversary is secretly controlling the policy of one of the agents. We would need new methods that assume some of the training agent's partners might not be aligned with the objectives of the group.
\subsection{Methods of Evaluation}
One aspect of AHT that might be overlooked is the evaluation of the agents at test time. An AHT agent might have to act differently in the presence of an adversary instead of a low-skill partner. In general, we would need AHT methods that are more risk-aware and define what success looks like when there is a high probability that some of the other agents will try to subvert the group's goal.
\section{Future Research Directions}
AHT was established with robust multi-agent behavior as the main goal. Some past works in AHT have made steps in that direction. Type-based methods, like in \citet{rahman2021towards}, attempt to infer the agent's behavior from its actions. Agents can also use communication to help each other quickly adapt to new scenarios \cite{macke2021expected}, which as been successful in defending against attacks in multi-agent deep learning \cite{tu2021adversarial}. Extending past work in AHT can show us which methods have the most potential in being robust against adversaries. Past works in adversarial training in RL \cite{pinto2017robust, pan2019risk} show that robustness can be achieved if the agent is trained with adversarial examples.
Extending past methods alone, however, will likely not be enough. Introducing new agents to other previously trained agents exacerbate the brittleness of current RL methods \cite{gleave2019adversarial}. The success of adversarial AHT, and AHT in general, will depend on a foundation of risk-awareness for multi-agent RL in the presence of adversaries. This foundation needs to account for the ``openness" of the environment, which models how new agents enter the group. The presence of adversaries compels us to carefully understand how openness impacts the training of an AHT agent. Openness that includes adversarial agents will require researchers to assume additional formal descriptions of the environment, which include (1) the distribution of new adversarial agents to introduce to the group, (2) the diversity of the adversaries, and (3) the reward function that accounts for the number and types of adversaries. As multi-agent RL is starting to be applied to more real-world tasks \cite{wang2021multi, bae2022scientific}, there is now more motivation to train agents that are robust and risk-aware. This creates an incentive to train AHT agents in high-stakes environments. An adversarial ML perspective can make these endeavors more successful.
\section{Conclusion}
As RL continues to be pursued to accomplish complex, real-world tasks, we will inevitably need to train RL agents that robustly cooperate with others. Since AHT seems to be a promising path toward that direction, we propose accounting for the possible presence of adversaries as a necessary component of that field. We put forward that adversarial ML researchers will be critical to ensuring AHT will advance safe collaboration between both humans and RL agents.
|
\section{Introduction}
Dynamical systems are pervasive in engineering and science applications. They are typically time-dependent systems of ordinary differential equations (ODEs) or partial differential equations (PDEs). The latter is not different from the former from the method of lines viewpoint in which a PDE reduces to a system of ODEs after a spatial discretization. For practical settings, simulating a dynamical system could be challenging due to a large number of degrees of freedom, and hence the number of ODEs, interdependent on each other in a highly nonlinear manner. For multi-scale or stiff systems of ODEs, explicit time discretization schemes, though straightforward, are not efficient to due time stepsize limitation to ensure stability. Implicit schemes, on the other hand, are stable but computationally expensive as a large linear system of equations needs to be solved at each time step. Though currently infeasible, real-time accurate approximate solutions for the practical complex dynamical system are highly desirable for control, optimization, uncertainty quantification, and decision-making.
Towards achieving real-time solutions for dynamical systems, various pure data-driven deep learning attempts have been made.
Autoencoder architecture has been explored to simulate fluid flows \cite{kim2019deep}. Autoencoder with physics-informed regularization to improve accuracy has been proposed to predict the future sea surface temperature given past series of measurements \cite{de2019deep}.
{In \cite{sanchez2018graph}, a graph network-based model is trained to approximate the forward map and inference model, and then used to speed up control algorithms.}
As an effort to combine traditional and machine learning approaches, {the authors in} \cite{morton2018deep} introduce a deep Koopman model\textemdash an auto-encoder architecture of convolution neural network\textemdash to predict the dynamics of airflow over a cylinder. Comprehensive overviews of machine learning methods for forecasting dynamical systems can be found in \cite{lim2021time} and \cite{benidis2018deep}. The work in \cite{duraisamy2021perspectives} presents a review and aspects of using machine learning techniques to simulate turbulent flows.
Instead of replacing traditional computational approaches with pure data-driven machine learning models, which is debatable and an active research direction, one can use machine learning methods to speed up only computationally demanding modules. This could maintain desirable physics constraints as in traditional approaches while gaining computational time. Indeed, a convolution neural network (CNN) can be trained to learn the numerical error between high-resolution and low-resolution simulations \cite{pathak2020using}. Combining the CNN prediction with low-resolution simulations can then achieve similar high-resolution accuracy while being faster and at that the same time not compromising the physics. {In a different effort, neural networks are trained to replace components/terms severely affected by a low-resolution grid \cite{kochkov2021machine}}. The predictions from neural networks are then unrolled over multiple time steps to improve long-time inference performance. A recurrent neural network can also be used to enhance the effectiveness of geometric multigrid methods for simulating Navier-Stokes equations \cite{margenberg2022neural}.
Completely replacing traditional methods while respecting governing equations, we argue, is highly desirable for machine learning methods because fast but nonphysical solutions are undesirable. A popular deep-learning approach aiming to accomplish this goal is the physics-informed neural
network (PINN) \cite{raissi2017physics1}. Similar to least squares finite element
methods, PINN trains deep learning solution
constrained by the PDE residual through a regularization
\cite{raissi2017physics1, RaissiEtAl2019, RaissiKarniadakis2018,
RaissiEtAl2017, YangPerdikaris2019, TripathyBilionis2018}).
PINN can learn solutions that attempt to make the
PDE residual small. However, the PINN approach directly approximates
the PDE solution in infinite dimensional spaces. While universal
approximation results (see, e.g.,
\cite{Cybenko1989,hornik1989multilayer,Zhou17,johnson2018deep}) could
ensure any desired accuracy with a sufficiently large number of neurons,
practical network architectures are moderate in both depth and width,
and hence the number of weights and biases,
the
accuracy of PINN can be limited. Moreover, PINN requires a retrain for new scenarios such as new boundary conditions, or new initial conditions, or new values of the underlying parameters. A physics-informed recurrent neural network has also been studied in \cite{jia2019physics}. In order to produce physically consistent and better prediction results, energy flow and density-depth constraint laws are integrated into the loss function.
Instead of learning the infinite-dimensional solution as in PINN, learning discretized solutions of dynamical systems is equally popular.
{The work in}
\cite{zhuang2021model} uses a neural network to approximate the derivative of the system state in reduced projected subspace. The neural network is then combined with forward Euler and Runge-Kutta time discretization schemes to achieve high-accuracy solutions. Alternatively, {a feed forward neural network can be used to directly learn the map from the solution at the current time step to the solution in the next time step \cite{pan2018long}}. The stability and accuracy of long-time prediction are reinforced by introducing a Jacobian regularization into the loss function. Realizing several drawbacks of the direct learning approach,
{the authors in \cite{Wang1998RungeKuttaNN} propose to learn the tangent slope with Runge-Kutta schemes}. Once trained, the learned tangent slope can be used with any time discretization schemes and any time step size.
In
\cite{um2020solver}, the authors propose to learn a correction neural network that lifts low-resolution solutions to high-resolution accuracy, and the training procedure includes low-resolution differentiable codes. Similarly, differential molecular dynamics simulations \cite{jaxmd2020} have been implemented in \texttt{Jax}
\cite{jax2018github}. Alternatively,
{the authors in }\cite{hu2019difftaichi} develop a differentiable simulations package that wraps a numerical simulator as a gradient kernel for end-to-end back-propagation used in optimization algorithms. Similar to \cite{jaxmd2020}, {a differentiable physic simulations package equipped with the adjoint method for backpropagation is developed in \cite{holl2020phiflow}, which enables the embedding of physical forward model into the training process.}
In this paper, aiming at simulating dynamical systems in real-time, we propose a model-constrained tangent slope deep learning (\texttt{mcTangent}) approach that has several appealing features over existing methods. First, it operates on finite dimensional systems and is thus in principle easier to train. However, it is spatial discretization-dependent for systems governed by PDEs. Second, it learns the underlying tangent slope and thus is semi-discrete in nature. Once trained, it can be deployed with any time discretization schemes with any time step size. The next three features are the main advances
beyond the work in \cite{Wang1998RungeKuttaNN}.
Third, it aims to fulfill the governing equations by constraining a fully discrete system in the loss function during training. Fourth, it is equipped with sequential learning strategies and thus promotes stability and accuracy in simulating the underlying dynamical systems far beyond the training time horizon. Fifth, our approach imposes regularizations on the smoothness of the neural network tangent and its derivatives implicitly via data randomization. This provides extra stability and accuracy for \texttt{mcTangent} solutions.
The paper is organized as follows. Section \secref{mcTangent} introduces an abstract dynamical system and a model-constrained tangent slope learning (\texttt{mcTangent}) approach. Both sequential machine learning and sequential model-constrained strategies will be discussed in detail in \cref{sect:mc} and \cref{sect:sMC}. Data randomization approach then follows with an in-depth semi-heuristic argument to reveal its implicit regularization nature in \cref{sect:noise_data_sec}. In particular, data randomization induces smoothness regularization for the underlying neural network via the standard machine learning loss. The beauty of the model-constrained loss term is that it not only enforces the likeliness of the neural network and the truth tangent {slopes} but also implicitly constrains their likeliness up to second-order derivatives via data randomization. \cref{sect:error} provides a rigorous estimation for prediction error using \texttt{mcTangent} approach. Several numerical results using the proposed \texttt{mcTangent} approach for transport equation, viscous Burger's equation, and Navier-Stokes equation are presented in \cref{sect:numerics}.
{We also provide detailed information on parameter tuning, randomness setting, and the cost for both training and testing.}
Section \secref{conclusions} concludes the paper with future work.
\section{Model-constrained tangent {slope} deep learning solutions for dynamical systems}
\seclab{mcTangent}
\subsection{Motivation}
For the concreteness and simplicity of the exposition, let us consider an abstract dynamical system governed by the following time-dependent scalar PDE equation of the form
\begin{equation}
\eqnlab{eq_base}
\pp{u}{t} = \mc{G}\LRp{u, \Grad u, \hdots} \quad \text{ in } \Omega \subset {\bs{\mathbb{R}}}^\d,
\end{equation}
where $t \in \LRs{0,T}$, $u\LRp{\bs{x}} \in {\bs{\mathbb{R}}}$ for any $\bs{x} \in {\bs{\mathbb{R}}}^d$, and $\d \in \LRc{1,2,3}$. We also assume \eqnref{eq_base} is equipped with appropriate initial conditions and boundary conditions to ensure its well-posedness.
In this paper, we are interested in {\em parametrized PDEs}. For downstream tasks such as design, control, optimization, inference, and
uncertainty quantification, these PDEs need to be solved many times. As such,
we wish to approximate solutions of \eqnref{eq_base} in real time for
any parameters (e.g. initial conditions or boundary conditions, or some parameter).
Training a PINN
together with parameters (either by themselves or their neural
networks weights and biases as another set of optimization variables)
\cite{Chen20, RAISSI2019686, DeepXDE21,lu2021physicsinformed} may not
be efficient as a new solution (corresponding to new
parameters) requires a retrain. We note that attempts using pure
data-driven deep learning to learn the parameter-to-solution map have
been explored (see, e.g., \cite{Kojima17, WHITE20191118,
Pestourie2020, Tahersima2019, Peurifoyeaar4206, Kojima17,
DNNInverseNanoPhotonnics20, Jiang2020}). On the other hand,
standard numerical methods such as finite difference, finite volume,
and finite elements \cite{smith1985numerical, leveque2002finite, johnson2012numerical} discretize \eqnref{eq_base} both in time
and space. One of the most popular approaches is perhaps the method of lines (see, e.g., \cite{schiesser2012numerical})
in which one performs spatial discretization first to obtain a system
of (possibly nonlinear) ordinary differential equations of the form
\begin{equation}
\eqnlab{MoLines}
\pp{\pb}{t} = \bs{G} \LRp{\pb},
\end{equation}
where $\pb$ and $\bs{G}$ are vector representations of finite dimensional
approximations of $\u$ and $\mc{G}$, respectively. Now, either an
explicit or implicit (or their combination) can be deployed to
discretize the temporal derivative. For the former, the most expensive
operation is the evaluation of tangent {slope\footnote{We call the right hand side $\bs{G} \LRp{\pb}$ as the tangent slope as it is a generalization of the tangent slope field in scalar ordinary differential equation.}} $\bs{G} \LRp{\pb}$. For
the latter, evaluating both $\bs{G} \LRp{\pb}$ and its (possibly
approximate) Jacobian for each time step play a vital
role. Implementing the Jacobian, even with the adjoint method \cite{tromp2005seismic}, is
a significant part of the programming effort. Automatic
differentiation can mitigate this programming burden at the expense
of more memory bandwidth. In summary, computing $\bs{G}\LRp{\pb}$ and its
Jacobian is a major part, both in implementation and computational time, of existing numerical methods.
To overcome the time burden of estimating the tangent slope and its
Jacobian, we present a model-constrained tangent slope deep learning approach (\texttt{mcTangent}) inspired by the semi-discrete nature of the method of
lines. In particular, we first learn the tangent slope $\bs{G}
\LRp{\pb}$ using neural network and then use a time discretization to
solve for approximations of $\pb$. Our approach thus aims to approximate only the spatial discretization and leaves the temporal discretization for traditional time integrators. At the heart of our approach is the incorporation of the governing equations into the
neural network tangent by constraining the learning task to respect a temporal
discretization of \eqnref{MoLines}. Again, unlike PINN and its siblings which learn the
infinite-dimensional solution $u$, our approach learns the tangent
slope of the finite-dimensional approximation $\pb$. Furthermore,
we constrain the physics on the discrete level. Clearly, our approach
is discretization-dependent while PINN requires neither spatial discretization nor temporal discretization.
\subsection{Model-constrained neural network approach with sequential data learning}
\seclab{mc}
In this section, we construct a model-constrained neural network
$\NN{\pb}$ to learn $\bs{G}\LRp{\pb}$. This is done in tandem with a
time discretization of \eqnref{MoLines}. For clarity, we limit our presentation to forward Euler method
\begin{equation}
\eqnlab{ubFE}
\ui{k+1} = \ui{k} + \Delta t \, \bs{G} \LRp{\ui{k}},
\end{equation}
as it is straightforward
to extend the approach to any time discretization scheme, and we provide a brief discussion at the end of the section. The task at hand is to train $\NN{\pb}$ on a certain spatial mesh $\mc{T}$ corresponding to a spatial discretization. To begin, let us
denote the numerical solutions of \eqnref{ubFE} at $N_t + 1$ time
steps on a finer mesh $\mc{T}^f$ as
\[
\LRc{\ui{0}, \ui{1}, \hdots, \ui{N_t}}.
\]
which are then down-sampled on $\mc{T}$ for training $\NN{\pb}$. Doing so
has proved to yield more accurate predictions than training directly
on the solutions on $\mc{T}$ \cite{pathak2020using, kochkov2021machine,
zhuang2021learned}. This is not surprising as the down-sampled
training data on $\mc{T}$ is more accurate than the solution on
$\mc{T}$.
The next idea that we like to incorporate into our approach is sequential training. The key is to feed the predictions back to the neural network model to enable a better long-time predictive capability. Using this idea \cite{wu2022learning} deploys a mixture of graph neural network and 3D-U-Net neural network to model fluid flows. Similarly, in
\cite{zhuang2021learned}
sequential learning is used to train a network to obtain the optimal finite difference coefficients from the high-resolution training data.
In the context of atmosphere modeling, \cite{brenowitz2018prognostic} introduces a stable and highly accurate long-time prediction loss function with sequential training.
Following \cite{pathak2020using, kochkov2021machine,wu2022learning,
zhuang2021learned}, we partition the training data in $N_t -S$ overlapping subsets
\[
\mc{U} := \LRc{\LRp{\ui{0}, \ui{1}, \hdots, \ui{S+1}}, \LRp{\ui{1}, \ui{2}, \hdots, \ui{S+2}}, \hdots, \LRp{\ui{{N_t-S-1}}, \ui{{N_t-S}}, \hdots, \ui{N_t}}}.
\]
For convenience in the exposition, we enumerate these $N_t -S$ subsets as
\begin{multline*}
\mc{U} := \big \{ \LRp{\ui{{0,0}}, \ui{{0,1}}, \hdots, \ui{{0,S+1}}}, \LRp{\ui{{1,0}}, \ui{{1,1}}, \hdots, \ui{{1, S+1}}}, \\
\hdots, \LRp{\ui{{N_t-S-1, 0}}, \ui{{N_t-S-1, 1}}, \hdots, \ui{{N_t-S-1, S+1}}} \big \},
\end{multline*}
where the second superscript denotes the local index in each subset. To distinguish from the sequential model-constrained learning in \cref{sect:sMC}, let us call the machine learning approach based on these overlapping subsets as {\em sequential data learning}.
We next discuss how we use each subset in our model-constrained
approach. Consider the $k$th subset $\LRp{\ui{{k,0}}, \ui{{k,1}},
\hdots, \ui{{k,S+1}}}$, for $k = 0, \hdots, N_t-S-1$. Starting from
$\ut{k,0} = \ui{k,0}$, we can write the sequence of approximate
solutions $\LRc{\ut{k,i}}_{i=1}^{S+1}$ for \eqnref{MoLines} using
forward Euler time discretization with the neural network tangent
$\NN{\pb}$ as
\begin{equation}
\eqnlab{ml_pred}
\ut{k,i+1} = \ut{k,i} + \Delta t \NN{\ut{k,i}}, \quad i = 0, \hdots, S.
\end{equation}
On the other hand, if we feed $\ut{k,i}$ through the forward Euler discretization \eqnref{ubFE} we obtain
\begin{equation}
\eqnlab{u_mc_pred}
\ubar{k,i+1} = \ut{k,i} + \Delta t \, \bs{G} \LRp{\ut{k,i}}, \quad i = 0, \hdots, S.
\end{equation}
As can be seen $\ut{k,i+1} \ne \ubar{k,i+1}$, though we wish they are the same.
{\em If they were, the approximate solutions using neural network tangent would respect the governing discretized equation exactly}.
Obviously, this is not feasible in general.
Thus, we resort to requiring $\ut{k,i+1} $ as close as possible to $ \ubar{k,i+1}$.
One way to accomplish this is to consider the following loss function
for the $k$th batch:
\begin{equation}
\eqnlab{loss_1}
\mc{J}_k := \quad \frac{1}{S+1} \sum_{i=1}^{S+1} \LRp{ \MSE{\ui{k,i} - \ut{k,i}} + \alpha \MSE{ \ubar{k,i} - \ut{k,i}}},
\end{equation}
where $\alpha$ is a model-constrained penalty (or regularization) parameter{, which controls the magnitude of the model-constrained loss (relative to the machine learning loss and). Parameter tuning in \cref{sect:Comp_cost} shows that a single value $\alpha = 10^5$ works well for all numerical examples in \cref{sect:numerics}}. The first term of the loss \eqnref{loss_1}\textemdash the ML Loss in \cref{fig:NN_architecture}\textemdash ensures the data
consistency, while the second term\textemdash the MC Loss in \cref{fig:NN_architecture}\textemdash is to force approximate solutions
of \eqnref{MoLines} using neural network tangent $\NN{\pb}$ to best fit the underlying space-time discretization \eqnref{ubFE}. The schematic of the \texttt{mcTangent} architecture with sequential data learning for the $k$th data subset and $S = 1$ is illustrated in \cref{fig:NN_architecture}. {We note that, unlike SINDy \cite{brunton2016discovering}, which discovers the dynamic systems from a dictionary of common differential operators, \texttt{mcTangent} aims to approximate high dimensional complex nonlinear tangent slope $\bs{G}$ operator using neural network.}
\begin{remark}
Note that it is not essential that
$\ubar{i}$ must be obtained by the forward Euler scheme \eqnref{ubFE}. In fact, our approach is flexible in the sense that any one-step explicit scheme, denoted as $\mc{F}$, (including explicit Runge-Kutta) is admissible. In such a case, our neural network can be considered as learning the forward Euler approximation of the ground-truth scheme.
\end{remark}
\input{Architecture}
Taking all the batches into account yields the total loss function
\begin{equation}
\eqnlab{loss_sum}
\mc{J} := \frac{1}{\LRp{N_t - S}\LRp{S+1}} \sum_{k=0}^{N_t - S-1} \sum_{i=1}^{S+1} \LRp{ \MSE{\ui{k,i} - \ut{k,i}} + \alpha \MSE{ \ubar{k,i} - \ut{k,i}}}.
\end{equation}
To gain insight into our \texttt{mcTangent} approach, we consider a linear problem in which $\bs{G}\LRp{\pb} = \bs{G} \pb$, and a one-layer linear neural network $\NN{\ui{k,0}} = \bs{W} \ui{k,0} + {\bf b}$. Under a mild condition, our approach should exactly recover the underlying tangent slope, i.e. $\NN{\pb} = \bs{G}\pb$.
Indeed, let $S = 0$ so that the loss function \eqnref{loss_sum} becomes
\begin{equation}
\eqnlab{Linear_1step}
\begin{aligned}
\mc{J} = & \frac{1+\alpha}{N_t} \sum_{k=0}^{N_t - 1} \MSE{\ui{k,1} - \ut{k,1}} = \frac{1+\alpha}{N_t} \nor{{U}^{1} - \Tilde{{U}}^{1}}_{F}^2 \\
= & \frac{\LRp{1+\alpha}\Delta t^2}{N_t} \nor{\bs{G} {U}^{0} - \LRp{\bs{W} {U}^{0} + {\bf b} \mathds{1}^T}}_{F}^2
\end{aligned}
\end{equation}
where ${U}^{t_i}$ and $\Tilde{{U}}^{t_i}$ are matrices with true and predictive solutions as columns, respectively, and $\mathds{1}$ is the unit column vector.
\begin{lemma}
\lemlab{linearOptimality}
The optimal solution $\LRp{\bs{W}^*, {\bf b}^*}$ for the training problem
\[
\min_{\bs{W},{\bf b}}\mc{J}
\]
is given by
\begin{equation}
\begin{aligned}
\eqnlab{optimal_linear_NN}
\bs{W}^* = \bs{G} \overline{\P} \overline{\P}^{\dagger},\quad
{\bf b}^* = \bs{G} \LRp{ \textbf{I} - \overline{\P} \overline{\P}^{\dagger}} \pbbar,
\end{aligned}
\end{equation}
where $\overline{\pb} := \frac{1}{N_t} U^{t_0} \mathds{1}$ is the column-average of matrix ${U}^{t_0}$, $\overline{\P} := U^{t_0} - \overline{\pb}\mathds{1}^T$, and $^\dagger$ denotes the pseudo-inverse. Consequently, the optimal network reads
\[
\NN{\pb} = \bs{G} \overline{\P} \overline{\P}^{\dagger} \pb + \bs{G} \LRp{ \textbf{I} - \overline{\P} \overline{\P}^{\dagger}} \pbbar.
\]
\end{lemma}
\begin{remark}
Lemma \lemref{linearOptimality} tells us that the optimal network exactly recovers the true forward map $\bs{G}$ if $\overline{\P}$ is a full row rank matrix. (In that case, $ \overline{\P} \overline{\P}^{\dagger} = \textbf{I}$.) This holds, for example, when the number of independent data samples is equal to the discretized dimension. We would like to point out that the MC loss term is the same as the ML loss term (up to a constant), and thus does not provide any extra information in this simple case.
When $S > 0$, at the time of writing this paper, we are not able to find a closed form solution as in Lemma \lemref{linearOptimality}. We leave it as future work.
\end{remark}
\begin{remark}
Although we learn the tangent slope using the Forward Euler scheme, it is straightforward to use any explicit scheme, such as Adams–Bashforth and Runge-Kutta methods to accomplish our goal. For example (ignoring extra subscripts for sequential data learning for simplicity), using the two-step Adams-Bashforth scheme, \texttt{mcTangent} solutions read
\[
\ut{i+1} = \ut{i} + \frac{3}{2}\, \Delta t \, \NN{\ut{i}} - \frac{1}{2}\, \Delta t \, \NN{\ut{i-1}},
\]
as oppose to the solutions using the truth tangent slope
\[
\ui{i+1} = \ui{i} + \frac{3}{2}\, \Delta t \, \bs{G}\LRp{\ui{i}} - \frac{1}{2}\, \Delta t \, \bs{G}\LRp{\ui{i-1}}.
\]
Similarly, \texttt{mcTangent} solutions based on the second-order Runge-Kutta scheme reads
\[
\begin{aligned}
\tilde{k}_1 & = \Delta t \, \NN{\ut{i}} \\
\tilde{k}_2 & = \Delta t \, \NN{\ut{i} + k_1} \\
\ut{i+1} & = \ut{i} + \frac{\tilde{k}_1 + \tilde{k}_2}{2},
\end{aligned}
\]
as oppose to the solutions using the truth tangent slope
\[
\begin{aligned}
k_1 & = \Delta t \, \bs{G}\LRp{\ui{i}} \\
k_2 & = \Delta t \, \bs{G}\LRp{\ui{i} + k_1} \\
\ui{i+1} & = \ui{i} + \frac{k_1 + k_2}{2}.
\end{aligned}
\]
Clearly, we have to modify the lost function accordingly, but the idea is the same as forward Euler approach that we have presented above.
\end{remark}
\begin{remark}
Note that we have used forward Euler time discretization for both \cref{eq:ml_pred} and \cref{eq:u_mc_pred} for simplicity, but this is not necessary. We recommend to use time discretizations with the same order of accuracy for both as accuracy gain in incompatible discretizations may not be well paid-off by additional computational demand. For example, if we use low-order accuracy for \cref{eq:ml_pred} but higher-order accuracy for \cref{eq:u_mc_pred}, \texttt{mcTangent} solution could be more accurate with smaller constant in the order of accuracy (still low-order) since it tries to match more accurate solution from \cref{eq:u_mc_pred}. However, the training cost could increase significantly due to several evaluations (and hence differentiations for back-propagation) of the truth tangent slope $
\bs{G}$ in \cref{eq:u_mc_pred}. Clearly high-order accurate approaches could tax the training time significantly.
\end{remark}
\subsection{Model-constrained neural network approach both sequential data and sequential model learnings}
\seclab{sMC}
In \cref{sect:mc}, we present a sequential data learning approach for the proposed model-constrained neural network $\NN{\pb}$ to learn the tangent slope while being constrained to provide the best possible approximate solutions for \eqnref{ubFE} for each time step. In order to improve the long-time predictive capability and accuracy,
this section constructs, in addition to {\em sequential data learning}, a {\em sequential model learning} strategy for training the neural network $\NN{\pb}$ is proposed. Sequential model learning is designed to promote the neural network solutions to respect the underlying discretization scheme for multiple time steps concurrently. In particular, starting from $\ut{k,i}$ we can carry out $R$ steps forward in time using the underlying discretization \eqnref{ubFE} as
\[
\ubar{k,i, r} = \ubar{k,i, r-1} + \Delta t \, \bs{G} \LRp{\ubar{k,i,r-1}}, \quad r = 1, \hdots, R,
\]
and using the neural network approximation \eqnref{ml_pred} as
\[
\ut{k,i, r} = \ut{k,i, r-1} + \Delta t \, \NN{\ut{k,i,r-1}}, \quad r = 1, \hdots, R,
\]
where $\ubar{k,i, 0} = \ut{k,i, 0}= \ut{k,i}$. Here the third superscript $r$ has been introduced to keep track of $R$ sequential forward steps starting from $\ut{k,i}$ for both exact and neural network tangent slopes. In order to ensure that these corresponding $R$ sequential predictions closely match each other, we consider the following loss function
\begin{equation}
\eqnlab{loss_sum_seq_mc}
\mc{J} := \frac{1}{\LRp{N_t - S} \LRp{S+1}} \sum_{k=0}^{N_t - S-1} \sum_{i=1}^{S+1} \LRp{\MSE{\ui{k,i} - \ut{k,i}} + \frac{\alpha}{R} \sum_{r = 1}^{R} \MSE{ \ubar{k,i,r} - \ut{k,i,r}}}.
\end{equation}
\input{Architecture_mc}
The schematic of the \texttt{mcTangent} architecture with both sequential data and model learnings for the $k$th data subset and $S = 2, R = 2$ is depicted in \cref{fig:NN_mc_architecture}. Clearly, when $R = 1$ we recover \eqnref{loss_sum} from \eqnref{loss_sum_seq_mc}. In other words, \eqnref{loss_sum_seq_mc} is a generalization of \eqnref{loss_sum}.
{Intuitively, larger values for $R, S$ increase the predictive capacity of \texttt{mcTangent} solutions, and as an example this will be demonstrated for the Burgers equation in \secref{Burger_eq}.}
However, it is computationally expensive to use large values for both $S$ and $R$. In the numerical results in \cref{sect:numerics}, we study two combinations: $S \geq 1, R = 1$ and $S = 1, R \geq 1$. In order to have a deeper understanding of the role of the loss function \eqnref{loss_sum_seq_mc} in training the neural network tangent and its predictive capability, we shall provide an in-depth heuristic argument in \cref{sect:noise_data_sec} and a rigorous error estimation for \texttt{mcTangent} predictions in \cref{sect:error}.
\subsection{Data randomization}
\seclab{noise_data_sec}
It has been observed \cite{sanchez2020learning} that adding a small amount of noise to training data not only increases the generalization on unseen data but also reduces accumulated errors in predictions. In fact, clean noise-free data does not represent the accumulated error in the predictive state that is fed back to the network to produce subsequent predictions. Moreover, noisy data encourages neural network predictions to be more robust to noise-corrupted inputs and errors.
In order to investigate the significance of different noise additions (adding noise to the training inputs, weights of the neural network, and output labels) on the model generalization, \cite{an1996effects} demonstrates that the reasonable noise level in the outputs does not influence the trained network. Randomizing training data, on the one hand, prevents the neural network from overfitting data, and on the other hand, can make the network insensitive to noise in data in the validation phase.
It is well-known that randomization induces a regularization of the gradient of the loss function with respect to the inputs \cite{reed1992regularization}.
Consequently, the neural network, if a proper noise level is used, is regularized to be a smooth function of the input data. The smoothness reduces the sensitivity to the variation in the input \cite{matsuoka1992noise} and can enhance the stability of long-time predictions \cite{poggio1990networks}.
The work in \cite{bishop1995training} showed that adding noise to data is equivalent to introducing a Tikhonov regularization to the loss function (where the regularization parameter is the noise variance) and thus improving the model generalization. However, the analysis is only valid in the context of infinite training data set, as pointed out in \cite{an1996effects}.
Inspired by the aforementioned work, we randomize the input data for the model-constrained network. We shall show that randomization induces regularizations not only to promote the smoothness of the network but also to enhance the similarity of the derivatives of the network $\NN{\pb}$ and the true tangent slope $\bs{G}\LRp{\pb}$. As shall be seen, the numerical results in \cref{sect:numerics} reveal that
randomization improves significantly the long-term stability and accuracy.
In this paper, we randomize the input $\pb$ of the network as
\begin{equation}
\eqnlab{add_noise}
\vb = \pb + \bs{\epsilon},
\end{equation}
where $\bs{\epsilon}$ is a normal random vector $\bs{\epsilon} \sim \mc{N}\LRp{\bs{0}, \delta^2 \textbf{I}}$. Note that the following heuristic arguments also hold for any random vector with independent components, each of which is a random variable with zero mean and variances $\delta^2$. Let $\mathbb{E}\LRs{\cdot}$ denote the expectation with respect to $\bs{\epsilon}$. Following \cite{an1996effects}, for a generic loss function $\mc{L}\LRp{\pb}$ we have
\begin{equation}
\eqnlab{Expect_form}
\begin{aligned}
\mathbb{E}\LRs{\mc{L}\LRp{\vb}} = & \mc{L}\LRp{\pb} + \mathbb{E}\LRs{\eval{\pp{\mc{L}}{\pb}}_{\pb} \bs{\epsilon}} + \frac{1}{2} \mathbb{E}\LRs{ \bs{\epsilon}^T \eval{\pp{^2 \mc{L}}{\pb^2}}_{\pb} \bs{\epsilon}}
+ o\LRp{\nor{\bs{\epsilon}}^2}
\\
\approx & \mc{L}\LRp{\pb} + \frac{1}{2} \mathbb{E}\LRs{ \bs{\epsilon}^T \eval{\pp{^2 \mc{L}}{\pb^2}}_{\pb} \bs{\epsilon}},
\end{aligned}
\end{equation}
where we have used sufficient small noise $\bs{\epsilon}$ (relatively to $\pb$) so that the high-order term $o\LRp{\nor{\bs{\epsilon}}^2}$, using the standard ``small o" notation, is assumed to be negligible.
We consider $S = 0$ and $R = 1$. (For $S > 0$ and/or $R > 1$ , the sequential inputs to the network contain the error which may not satisfy the condition for \eqnref{Expect_form} to hold.) In this case, the loss function \eqnref{loss_sum} becomes
\begin{equation}
\eqnlab{loss_noise}
\begin{aligned}
\mc{J} = & \frac{1}{N_t} \sum_{k=0}^{N_t - 1} {\underbrace{\MSE{\ui{k,1} - \ut{k,1}}}_{\mc{L}_{\text{ML}} \LRp{\ui{k,0} + \bs{\epsilon}}} + \alpha \underbrace{\MSE{\ubar{k,1} - \ut{k,1}}}_{\mc{L}_{\text{MC}}\LRp{\ui{k,0} + \bs{\epsilon}} } }
\end{aligned}
\end{equation}
We now study the randomized ML loss term $\mc{L}_{\text{ML}} \LRp{\ui{k,0} + \bs{\epsilon}}$
and the randomized MC loss term
$\mc{L}_{\text{MC}}\LRp{\ui{k,0} + \bs{\epsilon}}$
to gain insights into the role of randomization.
The machine learning loss term reads
\[
{\mc{L}_{\text{ML}}\LRp{\ui{k,0} + \bs{\epsilon}}} = \MSE{\ui{k,1} - \LRp{\ui{k,0} + \bs{\epsilon} + \Delta t \NN{\ui{k,0} + \bs{\epsilon}}}}
\]
which is a function of true input $\ui{k,0}$ plus a random noise vector $\bs{\epsilon}$. {\em It is important to note that we do not randomize the true data $\ui{k,1}$ against which we compare the machine prediction $\ut{k,1}$}.
Replacing $\mc{L}$ by $\mc{L}_{\text{ML}}$ in \eqnref{Expect_form} yields
\begin{equation}
\eqnlab{MLlossEx}
\begin{aligned}
\mathbb{E}\LRs{\mc{L}_{\text{ML}}\LRp{\ui{k,0} + \bs{\epsilon}}} \approx & \underbrace{\MSE{\ui{k,1} - \LRp{\ui{k,0} + \Delta t \NN{\ui{k,0}}}}}_{\mc{L}_{\text{ML}}\LRp{\ui{k,0}}} +
\delta^2 \LRs{\mc{P}_1 \LRp{\ui{k,0}} + \mc{P}_2 \LRp{\ui{k,0}}},
\end{aligned}
\end{equation}
where
\begin{equation}
\eqnlab{P1}
\mc{P}_1\LRp{\ui{k,0}} = \mb{Tr}\LRs{ \LRp{ \textbf{I} + \Delta t \eval{ \pp{\Psi}{\pb}}_{\ui{k,0}}}^T \LRp{ \textbf{I} + \Delta t \eval{ \pp{\Psi}{\pb}}_{\ui{k,0}}} },
\end{equation}
with $\mb{Tr}\LRs{\cdot}$ as the trace operator, and
\begin{equation}
\begin{aligned}
\mc{P}_2 \LRp{\ui{k,0}} = & \mb{Tr} \LRs{\Delta t {\eval{\frac{\partial^2 \Psi}{\partial \pb^2}}_{\ui{k,0}}}\odot \LRs{\LRp{\ui{k,0} + \Delta t \NN{\ui{k,0}}} - \ui{k,1}}},
\end{aligned}
\end{equation}
where $\odot$ denotes the dot product of the third order tensor $\Delta t \eval{\frac{\partial^2 \Psi}{\partial \pb^2}}_{\ui{k,0}}$ and the vector $ \LRs{\LRp{\ui{k,0} + \Delta t \NN{\ui{k,0}}} - \ui{k,1}}$.
From \eqnref{MLlossEx}, three observations are in order. First, on average, the randomized ML loss term is approximately the original ML loss term plus two additional terms $\mc{P}_1$ and $\mc{P}_2$ scaled by the variance $\delta^2$ of the noise. Second, the first term $\mc{P}_1$ is positive and thus is a regularization. It enforces the boundedness of the gradient (and hence the smoothness) of the neural network. Third, the second term $\mc{P}_2$ can be either positive or negative. However, when the time step $\Delta t$ is small and/or the ML misfit term $\LRs{\LRp{\ui{k,0} + \Delta t \NN{\ui{k,0}}} - \ui{k,1}}$ is small (e.g. with sufficient training), the contribution of the second term is expected to be dominated by the first and thus is negligible. When neither of these two conditions is satisfied, if the training enforces small ``curvature" of the neural network (i.e. small $\eval{\frac{\partial^2 \Psi}{\partial \pb^2}}_{\ui{k,0}}$) then the second term is also negligible. When this happens, training with randomization provides extra smoothness to the network.
Next, from \eqnref{ml_pred} and \eqnref{u_mc_pred}, { the randomized MC loss term} can be written as
\begin{equation*}
\begin{aligned}
\mc{L}_{\text{MC}}\LRp{\ui{k,0} + \bs{\epsilon}} = \MSE{\ubar{k,1} - \ut{k,1}}
= \Delta t^2 \MSE{ \bs{G}\LRp{\ui{k,0}+ \bs{\epsilon}} - \NN{\ui{k,0}+ \bs{\epsilon}}}.
\end{aligned}
\end{equation*}
Applying \eqnref{Expect_form} with $\mc{L}_{\text{MC}}$ in place of $\mc{L}$ gives
\begin{equation}
\eqnlab{MClossEx}
\begin{aligned}
\mathbb{E}\LRs{\mc{L}_{\text{MC}}\LRp{\ui{k,0} + \bs{\epsilon}}} \approx \underbrace{\Delta t^2 \MSE{ \bs{G}\LRp{\ui{k,0}} - \NN{\ui{k,0}}}}_{\mc{L}_{\text{MC}}\LRp{\ui{k,0}}} + \delta^2 \LRs{\mc{Q}_1\LRp{\ui{k,0}} + \mc{Q}_2\LRp{\ui{k,0}}},
\end{aligned}
\end{equation}
where
\begin{equation}
\mc{Q}_1 \LRp{\ui{k,0}} = \Delta t^2 \mb{Tr}\LRs{\LRp{\eval{\pp{\bs{G}}{\pb}}_{\ui{k,0}} - \eval{\pp{\Psi}{\pb}}_{\ui{k,0}}}^T \LRp{\eval{\pp{\bs{G}}{\pb}}_{\ui{k,0}} - \eval{\pp{\Psi}{\pb}}_{\ui{k,0}}} },
\end{equation}
and
\begin{equation}
\begin{aligned}
\mc{Q}_2\LRp{\ui{k,0}} = & \mb{Tr} \LRs{ \Delta t \LRp{\eval{\frac{\partial^2 \bs{G}}{\partial \pb^2}}_{\ui{k,0}} - \eval{\frac{\partial^2 \Psi}{\partial \pb^2}}_{\ui{k,0}}}\odot \Delta t \LRp{\NN{\ui{k,0}} - \bs{G}\LRp{\ui{k,0}}} }.
\end{aligned}
\end{equation}
As can be seen, the randomized MC loss term is approximately a sum of the original ML loss term and two additional terms. The first term $\mc{Q}_1$ is non-negative and behaves like a regularization to enforce the likeliness of the derivatives with respect to $\pb$ of the neural network tangent $\NN{\pb}$ and the true tangent $\bs{G}\LRp{\pb}$. The second term, though could be either negative or positive, can be negligible with sufficient training so that the MC misfit $\Delta t\LRp{\NN{\ui{k,0}} - \bs{G}\LRp{\ui{k,0}}}$ is relatively small. Another possibility for the insignificance of the second term is when the difference in the ``curvature" of the neural network tangent and the true tangent is sufficiently small. In that case, training with randomization promotes the closeness of not only $\NN{\pb}$ and $\bs{G}\LRp{\pb}$ but their first and second derivatives with respect to $\pb$: {\em confirming the significant advantages obtained from data randomization}.
{Next, combining \eqnref{loss_noise}, \eqnref{MLlossEx}, and \eqnref{MClossEx} yields the following result.
\begin{theorem}
Let the input of the neural network be randomized as in \cref{eq:add_noise}. Then
\begin{align}
&\mathbb{E}\LRs{\mc{J}} = \frac{1}{N_t} \sum_{k=0}^{N_t - 1} \LRp{\mc{L}_{\text{ML}} \LRp{\ui{k,0} } + \alpha \mc{L}_{\text{MC}}\LRp{\ui{k,0}}} \eqnlab{loss_final_noise} \\
& + \frac{\delta^2}{N_t} \sum_{k=0}^{N_t - 1} \LRs{ \mc{P}_1 \LRp{\ui{k,0}} + \mc{P}_2\LRp{\ui{k,0}} + \alpha \LRp{\mc{Q}_1 \LRp{\ui{k,0}} + \mc{Q}_2\LRp{\ui{k,0}}} } + o\LRp{\nor{\bs{\epsilon}}^2}.
\nonumber
\end{align}
\end{theorem}}
The first sum in \cref{eq:loss_final_noise} is the original loss (without randomization) and the second sum consists of additional terms induced by data randomization. These additional terms play a vital role in stimulating the stability and accuracy of the neural network. Indeed, as discussed above, randomizing the machine learning loss term encourages the smoothness of the neural network tangent by penalizing its first and second derivatives implicitly. Note that explicitly penalizing the first derivative of a neural network as in \cite{pan2018long} is possible, but this could be computationally expensive and challenging. Doing so for both the first and second derivatives is not recommended.
The above heuristic analysis of data randomization also reveals the {\em power of the model-constrained term} in training neural network: it promotes the agreement of the neural network tangent and the true tangent up to second order that is otherwise not realizable using the standard data-driven approach with only machine learning loss term.
\subsection{Estimation of prediction errors}
\seclab{error}
In this section, we show how data randomization helps improve the stability and accuracy of long-time predictions. We are interested in predicting solutions of the system \eqnref{MoLines} starting from an initial condition $\ui{0}$ that is not in the training set. To that end, it is natural to compare the \texttt{mcTangent} solutions $\ut{i}$ in \eqnref{ml_pred} with the solutions $\ui{i}$ obtained from the discretized system \eqnref{ubFE}.
Let us define the neural prediction error as
\begin{equation}
\eqnlab{err_v}
\ev{}\LRp{\ut{i}} = \ui{i+1} - \LRs{\ut{i} + \Delta t \NN{\ut{i}}}, \quad \eg{i+1} = \nor{\ev{}\LRp{\ut{i}}}_2.
\end{equation}
From \eqnref{ubFE}, \eqnref{ml_pred}, and \eqnref{err_v} we have
\begin{equation}
\eqnlab{bound_err}
\begin{aligned}
\ev{}\LRp{\ut{i}} = & \LRp{\ui{i} + \Delta t \bs{G} \LRp{\ui{i}}} - \LRp{\ut{i} + \Delta t \NN{\ut{i}}} \\
= & \Delta t \bs{G} \LRp{\ut{i} + \ev{}\LRp{\ut{i-1}}} - \Delta t \NN{\ut{i}} + \ev{}\LRp{\ut{i-1}}
\end{aligned}
\end{equation}
Applying the Taylor expansion for the first term gives
\begin{equation}
\Delta t \bs{G} \LRp{\ut{i} + \ev{}\LRp{\ut{i-1}}} = \Delta t \bs{G} \LRp{\ut{i}} + \Delta t \eval{\pp{\bs{G}}{\pb}}_{\ut{i}} \ev{}\LRp{\ut{i-1}} + o\LRp{\eg{i}}
\end{equation}
Substituting back to \eqnref{bound_err}, we have
\begin{equation}
\eqnlab{bound_err_2}
\begin{aligned}
&\ev{}\LRp{\ut{i}} = \Delta t \LRs{\bs{G} \LRp{\ut{i}} - \NN{\ut{i}}} + \Delta t \eval{\pp{\bs{G}}{\pb}}_{\ut{i}} \ev{}\LRp{\ut{i-1}} + \ev{}\LRp{\ut{i-1}} \\
& + o\LRp{\eg{i}} = \Delta t \LRs{\bs{G} \LRp{\ut{i}} - \NN{\ut{i}}} + \Delta t \LRs{\eval{\pp{\bs{G}}{\pb}}_{\ut{i}} - \eval{\pp{\Psi}{\pb}}_{\ut{i}}} \ev{}\LRp{\ut{i-1}} \\
& + \LRs{ \textbf{I} + \Delta t \eval{\pp{\Psi}{\pb}}_{\ut{i}}} \ev{}\LRp{\ut{i-1}} + o\LRp{\eg{i}}
\end{aligned}
\end{equation}
Applying triangle inequality and Cauchy–Schwarz inequality for \eqnref{bound_err_2} and using \eqnref{err_v} yields
\begin{equation}
\eqnlab{final_err}
\begin{aligned}
\eg{i+1} \leq &
\Delta t\MSEsqrt{\bs{G}\LRp{\ut{i}} - \NN{\ut{i}}}\\
& + \Delta t \MSEsqrt{\eval{\pp{\bs{G}}{\pb}}_{\ut{i}} - \eval{\pp{\Psi}{\pb}}_{\ut{i}}} \eg{i} + \MSEsqrt{1 + \Delta t \eval{\pp{\Psi}{\pb}}_{\ut{i}}} \eg{i} + o\LRp{\eg{i}}, \quad i \ge 0.
\end{aligned}
\end{equation}
We observe in \eqnref{final_err} that the first term on the right-hand side is the model-constrained loss term
being as small as possible at the training data. On the other hand, $\Delta t \MSEsqrt{\eval{\pp{\bs{G}}{\pb}}_{\ut{i}} - \eval{\pp{\Psi}{\pb}}_{\ut{i}}}$ and $\MSEsqrt{1 + \Delta t \eval{\pp{\Psi}{\pb}}_{\ut{i}}}$ are regularized to be bounded and/or small by data randomization (see \cref{sect:noise_data_sec}). A heuristic argument reveals that the prediction error is under control at all times. Indeed, suppose $ \Delta t\MSEsqrt{\bs{G}\LRp{\ut{i}} - \NN{\ut{i}}}$, $\Delta t \MSEsqrt{\eval{\pp{\bs{G}}{\pb}}_{\ut{i}} - \eval{\pp{\Psi}{\pb}}_{\ut{i}}}$, and $\MSEsqrt{1 + \Delta t \eval{\pp{\Psi}{\pb}}_{\ut{i}}}$ are bounded. Since $\eg{0} = 0$, $\eg{1}$ is bounded, and by induction $\eg{i}$ is also bounded for $i \ge 0$. A rigorous version of this argument is given in Theorem \theoref{mainTheo}.
\begin{theorem}
\theolab{mainTheo}
Assume that the second derivative of $\bs{G}\LRp{\pb}$ with respect to $\pb$ is uniformly bounded.
Let
\[
f^{i+1} := \Delta t\MSEsqrt{\bs{G}\LRp{\ut{i}} - \NN{\ut{i}}},
\]
and
\[
g^{i+1} := \Delta t \MSEsqrt{\eval{\pp{\bs{G}}{\pb}}_{\ut{i}} - \eval{\pp{\Psi}{\pb}}_{\ut{i}}} + \MSEsqrt{1 + \Delta t \eval{\pp{\Psi}{\pb}}_{\ut{i}}} + c^i,
\]
where $c^i = \mc{O}\LRp{\eg{i}}$. Then,
the prediction error $\eg{n}$ at time $t_n$ satisfies
\[
\eg{n} \le \sum_{k = 1}^{n}\LRp{\Pi_{i = k+1}^{n}g^i} f^k.
\]
\end{theorem}
\begin{proof}
The proof is a simple application of a discrete Gronwall lemma on \eqnref{final_err}.
\end{proof}
\begin{remark}
Note that the boundedness of the second derivative of $\bs{G}\LRp{\pb}$ with respect to $\pb$ is valid for problem \eqnref{MoLines} with a smooth tangent slope. The boundedness of $f^i$ and $g^i$ is not too restricted if the prediction scenarios are close to the training data. Indeed, as argued in \cref{sect:noise_data_sec}, data randomization enforces the small values for $f^i$ and $g^i$ at the training points. Now, due to the smoothness of $\NN{\pb}$ and $\bs{G}\LRp{\pb}$ and their closeness in both values and derivatives (again by randomization), the continuity guarantees the small values for $f^i$ and $g^i$ during the prediction.
\end{remark}
\begin{remark}
Theorem \theoref{mainTheo} allows us to bound the error between the neural network prediction with the exact solution of the original PDEs \eqnref{eq_base} provided that an error estimation of the solution of the discretized equation \eqnref{ubFE} is given. Indeed, suppose the error in the discretized solution $\pb^n$ and the exact solution $\pb\LRp{t_n}$ at time $t_n$ is bounded by $\mc{O}\LRp{\Delta t + h^p}$, where $h$ is the mesh size and $p$ is the order of accuracy of the underlying spatial discretization. Then by a simple application of triangle inequality we have
\[
\ut{n} - \pb\LRp{t_n} = \mc{O}\LRp{\Delta t + h^p + \sum_{k = 1}^{n}\LRp{\Pi_{i = k+1}^{n}g^i}f^k},
\]
which shows that in order to get the optimal accuracy and computational effort we ideally need to balance not only the temporal and spatial discretization errors but also the error in the neural network. Clearly, balancing the former two is not that difficult from a numerical analysis point of view, but balancing also the network error is challenging as it depends on the actual training process and randomization.
\end{remark}
\section{Numerical results}
\seclab{numerics}
In this section, we present several numerical results using the proposed model-constrained tangent slope neural network (\texttt{mcTangent}) approach for transport equation (\cref{sect:1D_wave}), viscous Burger's equation (\cref{sect:Burger_eq}), and Navier-Stokes equation (\cref{sect:2D_NS}). As shall be shown, \texttt{mcTangent} solutions are\textemdash thanks to the model-constrained term and data randomization\textemdash stable and capable of producing accurate approximations far beyond the training time horizons. In \cref{sect:Comp_cost} we provide detailed information on parameter tuning, randomness setting, and the cost for both training and testing.
Five hyperparameters of interest are the number of training samples, noise level $\delta$, sequential machine learning steps $S$, sequential model-constrained learning steps $R$, and regularization parameter $\alpha$. For convenience, we shall conventionally write them in a group.
For example the $\LRp{d600, 2\%, 1,1,0}$ setting means we consider 600 training data samples, 2\% noise, S = 1, R = 1, and $\alpha = 0$. In order to ensure the fairness between simulations and the comparison among approaches, we use fixed random keys
for training and testing data generation, for adding noise, and for neural network parameter initialization
We implement our approach and perform all computations in \texttt{JAX} \cite{jax2018github}.
We would like to point out that all computations (training, testing, and predicting) are done with single precision accuracy.
\subsection{One-dimensional (1D) wave/transport equation}
\seclab{1D_wave}
The 1D wave equation considered in this section is given by
\[
\pp{u}{t} + c \pp{u}{x} = 0,
\]
with the wave speed $c = 1$, the spatial domain $x \in \LRp{0,1}$, and time horizon $t \in \LRp{0, T}$. The equation is equipped with an initial condition $u(x,0) = u_0(x)$ and periodic boundary condition. We are interested in real-time approximate solutions of the wave equation for any initial condition $u_0(x)$.
{\bf Data generation.}
In this problem, the initial condition samples are drawn from
\[
u_0(x) = \sum_{i=1}^{5} a_i \sin \LRp{2\pi x \, i} + \sum_{i=1}^{5} b_i \cos \LRp{2\pi x \, i},
\]
where $a_i,b_i$ are distributed by the standard normal distribution with zero mean and unit variance, i.e., $a_i,b_i \sim \mc{N}\LRp{0,1}$. We solve the wave equation with the forward Euler scheme for the temporal derivative
and the first order upwind finite difference scheme for spatial derivative. The time horizon is chosen as $T = 5 \times 10^{-2}$. A fine space-time mesh with $n_x = 10000$ points in space and $n_t = 2000$ points in time is deployed to achieve highly accurate solutions. The training data samples are obtained by extracting the high resolution solutions on a coarser uniform space-time mesh with $n_x = 200$ and $n_t = 100$. In this simple problem, we generate a fixed training data set of 200 initial conditions.
Note that we aim to predict long-time solutions, $t \in \LRp{0, 3}$, from the short-time training data in the interval $t \in \LRp{0, 5 \times 10^{-2}}$.
{\bf Neural network architecture.} Because of the linear nature of the problem and the first order upwind finite difference scheme, a linear neural work is sufficient to approximate the resulting tangent slope. The linear neural network is defined as
\[\NN{\ui{i}} = \bs{W} \ui{i} + {\bf b}, \]
where the weights $\bs{W} \in \mc{R}^{n_x \times n_x}$, and the bias ${\bf b} \in \mc{R}^{n_x}$. To train, we use \texttt{ADAM} \cite{kingma2014adam} optimizer with default parameters and the learning rate of $10^{-3}$.
We determine the best combination of weights and biases (and hence the final trained network) as the one that provides the lowest accumulated mean square error for $500$ time steps for the test sample. Specifically, during the training process, at each epoch, we solve for the predictions from the test initial condition with the current-epoch learned network. The accumulated mean-square error between predictive solutions and ground truth solutions is calculated at the $500$th time step to determine the ``optimal" network.
{\bf Long-time predictions.}
Shown in \cref{fig:1D_compare_all} is the mean-square error between true (high resolution) solutions and predicted ones obtained by various neural networks, each of which is trained with both randomized and noise-free training data.
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{figures/1D_wave/1D_results_compares.pdf}
\caption{\textbf{Wave/transport equation}. Comparison between different neural network approaches with/without randomization.}
\figlab{1D_compare_all}
\end{figure}
For pure data-driven machine learning networks ($\alpha = 0$, and thus no model-constrained term), we observe that noise-free data trained networks outperform those trained with noisy data ones.
This is not surprising as for this linear problem, as predicted by Lemma \lemref{linearOptimality}, one can obtain linear networks that accurately learn the tangent slope with sufficiently rich data.
Therefore, the predictions by the learned linear networks are almost the same as the ground truth solutions. On the contrary, training with noisy data causes the neural network to predict solutions with a small amount of error such that it adapts to (possibly overfits) the amount of noise in the ground-truth solutions.
\cref{fig:1D_wave_NN_params} presents the weight matrix and bias vector for two cases $\LRp{d200, 0\%, 10, 1, 0}$ and $\LRp{d200,0\%, 1, 1, 0}$ with noise-free data. It can be seen that both networks are almost identical and both have only an upper diagonal with a large magnitude. We also note that
the bias vector is relatively small
and thus we ignore this bias vector in the subsequent comparisons. We present the test predicted solutions for the setting $\LRp{d200,0\%, 1, 1, 0}$ in \cref{fig:1D_preds_different_periods}. As the network fits the tangent slope for high-resolution data, accurate results are preserved far beyond
the training time horizon, while finite difference results on the same coarse grid show a severe diffusion/dissipation effect. Furthermore, settings with a large number of sequential steps such as $\LRp{d200,0\%,10,1,0}$, $\LRp{d200,1\%,10,1,0}$ and $\LRp{d200,2\%,10,1,0}$ yield more accurate neural networks than their counterparts with $S = 1$. The reason is that long sequential training reduces the prediction error.
For model-constrained neural networks, we use $\alpha = 10^5$ as the regularization parameter
for all cases. We tested with different values for $\alpha$ and almost the same results are obtained for larger values, while smaller values make neural networks perform similarly to the pure data-driven machine learning networks. It can be seen in \cref{fig:1D_compare_all} that training with randomized data returns neural networks, regardless of $S, R$ values, as good as the coarse finite difference approximation with $n_x = 200$. This is expected as we constrain the training with a coarse finite difference model. The trained weight matrices and bias vectors for these neural networks corresponding to three settings $\LRp{d200,1\%, 10,1,10^5}$, $\LRp{d200,1\%, 1,1,10^5}$ and $\LRp{d200,1\%, 1,5,10^5}$ are shown in \cref{fig:1D_wave_FD_NN_params}. Again, the bias vectors do not have a significant role in the predictions. Note that, unlike those from purely data-driven in \cref{fig:1D_wave_NN_params} which have arbitrary structure, the model-constrained weight matrices, after ignoring small elements, have the same structure as the first-order upwind scheme matrix. Among these neural networks, the long sequential model-constrained network with $R = 5$ is closest to the first-order upwind scheme matrix. It is not surprising as the neural network is constrained to satisfy the first-order upwind scheme in multiple time steps. On the other hand, the neural network trained with noise-free data shows instability starting from the $2000$th time step in long-term predictions. This instability is due to the lack of regularizations
as compared to the randomized cases for which regularizations are explicit via the model-constrained term and implicit via randomization (see \cref{sect:noise_data_sec}).
{\tiny
\begin{figure}[htb!]
\centering
\begin{tabular*}{\textwidth}{c c c}
\centering
\rotatebox[origin=c]{90}{\small $S = 10$} &
\raisebox{-0.\height}{
\begin{minipage}{.4\textwidth}
\begin{tikzpicture}
\tikzstyle{every node}=[font=\tiny, node distance=7.5mm]
\node (img) {\includegraphics[width = .95\textwidth]{figures/1D_wave/1D_W__0_0_10_1.pdf}};
\node[below=of img, node distance=0cm, xshift = -3mm, yshift=1cm]
{Column index of $\bs{W}$};
\node[left=of img, node distance=0cm, rotate=90, anchor=center,yshift=-0.7cm]
{Row index of $\bs{W}$};
\end{tikzpicture}
\end{minipage}
} &
\raisebox{-0.\height}{
\begin{minipage}{.35\textwidth}
\begin{tikzpicture}
\tikzstyle{every node}=[font=\tiny, node distance=7.5mm]
\node (img) {\includegraphics[width = 0.95\textwidth]{figures/1D_wave/1D_b__0_0_10_1.pdf}};
\node[below=of img, node distance=0cm, yshift=1cm, xshift = 2mm]
{Index of bias vector ${\bf b}$};
\node[left=of img, node distance=0cm, rotate=90, anchor=center,yshift=-0.7cm]
{Magnitude $b_i$};
\end{tikzpicture}
\end{minipage}
}
\\
\centering
\rotatebox[origin=c]{90}{\small $S = 1$} &
\raisebox{-0.\height}{
\begin{minipage}{.4\textwidth}
\begin{tikzpicture}
\tikzstyle{every node}=[font=\tiny, node distance=7.5mm]
\node (img) {\includegraphics[width = 0.95\textwidth]{figures/1D_wave/1D_W__0_0_1_1.pdf}};
\node[below=of img, node distance=0cm, xshift = -3mm, yshift=1cm]
{Column index of $\bs{W}$};
\node[left=of img, node distance=0cm, rotate=90, anchor=center,yshift=-0.7cm]
{Row index of $\bs{W}$};
\end{tikzpicture}
\end{minipage}
} &
\raisebox{-0.\height}{
\begin{minipage}{.35\textwidth}
\begin{tikzpicture}
\tikzstyle{every node}=[font=\tiny, node distance=7.5mm]
\node (img) {\includegraphics[width = 0.95\textwidth]{figures/1D_wave/1D_b__0_0_1_1.pdf}};
\node[below=of img, node distance=0cm, yshift=1cm, xshift = 2mm]
{Index of bias vector ${\bf b}$};
\node[left=of img, node distance=0cm, rotate=90, anchor=center,yshift=-0.7cm]
{Magnitude $b_i$};
\end{tikzpicture}
\end{minipage}
}
\end{tabular*}
\caption{\textbf{Wave/transport equation}. Pure data-driven trained linear neural network parameters: weight matrix heat maps (\textit{left column}) and bias vector magnitudes (\textit{right column}) with $\alpha = 0, \delta = 0\%$.}
\figlab{1D_wave_NN_params}
\end{figure}
}
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{figures/1D_wave/Low_FD_and_NN_Predictions_based_dt.pdf}
\caption{\textbf{Wave/transport equation}. The predicted solutions at time steps $n_t = 2000, 4000, 6000$ by learned neural network corresponding to ($d200, S = 1, R = 1, \alpha =0, \delta = 0\%$), finite difference solutions on coarse grid with $n_x = 200$, and the high resolution solutions (True).}
\figlab{1D_preds_different_periods}
\end{figure}
\begin{figure}[htb!]
\centering
\begin{tabular*}{\textwidth}{c c c}
\centering
\raisebox{-0.5\height}{\small $S = 10, R = 1$} &
\raisebox{-0.5\height}{\small $S = 1, R = 1$} &
\raisebox{-0.5\height}{\small $S = 1, R = 5$}
\\
\centering
\raisebox{-0.\height}{
\begin{minipage}{.28\textwidth}
\begin{tikzpicture}
\tikzstyle{every node}=[font=\tiny, node distance=7.5mm]
\node (img) {\includegraphics[width = 0.95\textwidth]{figures/1D_wave/1D_W__1e5_1_10_1.pdf}};
\node[below=of img, node distance=0cm, xshift = -2.5mm, yshift=1cm]
{Column index of $\bs{W}$};
\node[left=of img, node distance=0cm, rotate=90, anchor=center,yshift=-0.7cm]
{Row index of $\bs{W}$};
\end{tikzpicture}
\end{minipage}
} &
\raisebox{-0.\height}{
\begin{minipage}{.28\textwidth}
\begin{tikzpicture}
\tikzstyle{every node}=[font=\tiny, node distance=7.5mm]
\node (img) {\includegraphics[width = 0.95\textwidth]{figures/1D_wave/1D_W__1e5_1_1_1.pdf}};
\node[below=of img, node distance=0cm, xshift = -2.5mm, yshift=1cm]
{Column index of $\bs{W}$};
\node[left=of img, node distance=0cm, rotate=90, anchor=center,yshift=-0.7cm]
{Row index of $\bs{W}$};
\end{tikzpicture}
\end{minipage}
} &
\raisebox{-0.\height}{
\begin{minipage}{.28\textwidth}
\begin{tikzpicture}
\tikzstyle{every node}=[font=\tiny, node distance=7.5mm]
\node (img) {\includegraphics[width = 0.95\textwidth]{figures/1D_wave/1D_W__1e5_1_1_5.pdf}};
\node[below=of img, node distance=0cm, xshift = -2.5mm, yshift=1cm]
{Column index of $\bs{W}$};
\node[left=of img, node distance=0cm, rotate=90, anchor=center,yshift=-0.7cm]
{Row index of $\bs{W}$};
\end{tikzpicture}
\end{minipage}
}
\\ ~ \\
\centering
\raisebox{-0.5\height}{
\begin{minipage}{.28\textwidth}
\begin{tikzpicture}
\tikzstyle{every node}=[font=\tiny, node distance=7.5mm]
\node (img) {\includegraphics[width = 0.95\textwidth]{figures/1D_wave/1D_b__1e5_1_10_1.pdf}};
\node[below=of img, node distance=0cm, yshift=1cm, xshift = 2mm]
{Index of bias vector ${\bf b}$};
\node[left=of img, node distance=0cm, rotate=90, anchor=center,yshift=-0.7cm]
{Magnitude $b_i$};
\end{tikzpicture}
\end{minipage}
} &
\raisebox{-0.5\height}{
\begin{minipage}{.28\textwidth}
\begin{tikzpicture}
\tikzstyle{every node}=[font=\tiny, node distance=7.5mm]
\node (img) {\includegraphics[width = 0.95\textwidth]{figures/1D_wave/1D_b__1e5_1_1_1.pdf}};
\node[below=of img, node distance=0cm, yshift=1cm, xshift = 2mm]
{Index of bias vector ${\bf b}$};
\node[left=of img, node distance=0cm, rotate=90, anchor=center,yshift=-0.7cm]
{Magnitude $b_i$};
\end{tikzpicture}
\end{minipage}
} &
\raisebox{-0.5\height}{
\begin{minipage}{.28\textwidth}
\begin{tikzpicture}
\tikzstyle{every node}=[font=\tiny, node distance=7.5mm]
\node (img) {\includegraphics[width = 0.95\textwidth]{figures/1D_wave/1D_b__1e5_1_1_5.pdf}};
\node[below=of img, node distance=0cm, yshift=1cm, xshift = 2mm]
{Index of bias vector ${\bf b}$};
\node[left=of img, node distance=0cm, rotate=90, anchor=center,yshift=-0.7cm]
{Magnitude $b_i$};
\end{tikzpicture}
\end{minipage}
}
\end{tabular*}
\caption{\textbf{Wave/transport equation}. Trained model-constrained linear neural network parameters: weight matrix heat map (\textit{top row}) and bias vector magnitude (\textit{bottom row}) with $\alpha = 1e^5, \delta = 1\%$.}
\figlab{1D_wave_FD_NN_params}
\end{figure}
{\bf Implicit time integration with learned network.}
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{figures/1D_wave/1D_implicit_schemes.pdf}
\caption{\textbf{Wave/transport equation}. Predicted solutions at $t = 0.1$ obtained by forward Euler scheme and backward Euler scheme with the true tangent slope $G$ and learned neural network counterparts with time stepsize $\Delta t' = \frac{50}{3} \Delta t = \frac{25}{3} \times 10^{-3}$.}
\figlab{1D_preds_imp}
\end{figure}
One of the advantages of our proposed tangent slope learning approach is that once trained the learned tangent slope can be used with any time discretization method.
To demonstrate this, we use the learned neural network tangent for the setting $\LRp{d200,0\%,1,1,0}$
with both backward and forward Euler schemes using a time stepsize $\Delta t' = \frac{50}{3} \Delta t$ which is much larger than the training stepsize. It can be seen in \cref{fig:1D_preds_imp} that the forward Euler solutions blow up for both learned and true tangent slopes, which is obvious as the time stepsize is much larger than the stable time stepsize. Both approaches are stable with implicit integration and the results are comparable (though the learned tangent slope was trained with a smaller time step size).
{\bf Direct learning versus \texttt{mcTangent} slope learning.}
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{figures/1D_wave/1D_results_compares_Direct.pdf}
\caption{\textbf{Wave/transport equation}. Comparison between direct neural networks (Direct) and tangent slope neural networks (\texttt{mcTangent}).}
\figlab{1D_compare_Direct}
\end{figure}
\begin{figure}[htb!]
\centering
\begin{tabular*}{\textwidth}{c c c}
\centering
\raisebox{-0.5\height}{\small $S = 10, R = 1, \alpha = 0 $} &
\raisebox{-0.5\height}{\small $S = 1, R = 1, \alpha = 0$} &
\raisebox{-0.5\height}{\small $S = 1, R = 5, \alpha = 1$}
\\
\centering
\raisebox{-0.5\height}{
\begin{minipage}{.28\textwidth}
\begin{tikzpicture}
\tikzstyle{every node}=[font=\tiny, node distance=7.5mm]
\node (img) {\includegraphics[width = 0.95\textwidth]{figures/1D_wave/1D_Direct_W_Direct_0_0_10_1.pdf}};
\node[below=of img, node distance=0cm, xshift = -2.5mm, yshift=1cm]
{Column index of $\bs{W}$};
\node[left=of img, node distance=0cm, rotate=90, anchor=center,yshift=-0.7cm]
{Row index of $\bs{W}$};
\end{tikzpicture}
\end{minipage}
} &
\raisebox{-0.5\height}{
\begin{minipage}{.28\textwidth}
\begin{tikzpicture}
\tikzstyle{every node}=[font=\tiny, node distance=7.5mm]
\node (img) {\includegraphics[width = 0.95\textwidth]{figures/1D_wave/1D_Direct_W_Direct_0_0_1_1.pdf}};
\node[below=of img, node distance=0cm, xshift = -2.5mm, yshift=1cm]
{Column index of $\bs{W}$};
\node[left=of img, node distance=0cm, rotate=90, anchor=center,yshift=-0.7cm]
{Row index of $\bs{W}$};
\end{tikzpicture}
\end{minipage}
} &
\raisebox{-0.5\height}{
\begin{minipage}{.28\textwidth}
\begin{tikzpicture}
\tikzstyle{every node}=[font=\tiny, node distance=7.5mm]
\node (img) {\includegraphics[width = 0.95\textwidth]{figures/1D_wave/1D_Direct_W_Direct_1_1_1_10.pdf}};
\node[below=of img, node distance=0cm, xshift = -2.5mm, yshift=1cm]
{Column index of $\bs{W}$};
\node[left=of img, node distance=0cm, rotate=90, anchor=center,yshift=-0.7cm]
{Row index of $\bs{W}$};
\end{tikzpicture}
\end{minipage}
}
\end{tabular*}
\caption{\textbf{Wave/transport equation}. Heat map of weight matrix of direct linear neural networks.}
\figlab{1D_wave_FD_Direct_NN_params}
\end{figure}
We now compare our tangent slope learning and direct learning. Here, by direct learning we mean learning the map from $\ui{i}$ to $\ui{i+1}$ for two consecutive time steps. Clearly, unlike the former, the latter is tailored, and thus limited, to a particular space-time discretization.
To be fair, we also use the linear network with zero bias for the direct learning approach. \cref{fig:1D_compare_Direct} presents the mean-square error of predictions obtained by direct neural networks and tangent slope networks, both with and without model-constrained terms. As can be seen, both direct and tangent slope neural networks are comparable in terms of accuracy. However, the learned weight matrices of direct neural networks
do not have the pattern of the underlying space-time discretization matrices, and this can be observed from \cref{fig:1D_wave_FD_Direct_NN_params}. That is, while our tangent slope approach preserves the structure of spatial discretization, the direct approach, which seems to be natural, does not.
\subsection{2D Burger's equation}
\seclab{Burger_eq}
We consider the following viscous 2D Burger's equations
\begin{align*}
& \pp{u}{t} + u \pp{u}{x} + v\pp{u}{y} = \nu \LRp{\pp{^2u}{x^2} + \pp{^2u}{y^2} } \\
& \pp{v}{t} + u \pp{v}{x} + v\pp{v}{y} = \nu \LRp{\pp{^2v}{x^2} + \pp{^2v}{y^2} },
\end{align*}
where $x,y \in \LRs{0,1}$ and $ t \in (0,T] $. The boundary condition is periodic and the initial velocity is given by $v(x,y,0) = v_0\LRp{x,y} = 1$ and $u(x,y,0) = u_0\LRp{x,y}$. We take the viscosity coefficient to be $\nu = 10^{-2}$. We aim to predict $x$-velocity $u$ in the time interval $t \in \LRp{0, 1.5}$ given an initial velocity $u_0(x,y)$ at $t = 0$.
{\bf Data generation.}
We draw periodic samples of $\pb$ using the truncated Karhunen-Loève expansion
\[u_0(x,y) = \exp \LRp{\sum_{i=1}^{15} \sqrt{\lambda_i} \, \omega_i(x,y) \, z_i},\]
where $\textbf{z} = \LRc{z_i}_{i=1}^{15} \sim \mathcal{N} \LRp{0, \textbf{I}}$, and $\LRp{\lambda, \omega}$ are eigenpairs obtained by the eigendecomposition of the covariance operator $7^{\frac{3}{2}} \LRp{-\Delta + 49 \textbf{I}}^{-2.5}$, where $\Delta$ is the Laplacian operator, with periodic boundary conditions. Training data corresponding to each initial velocity is generated from a $128 \times 128$ high-resolution spatial mesh
and $1000$ time steps for the time horizon $T = 0.1$
using finite difference method. These high resolution solutions are down-sampled on a coarser mesh of 100 time steps ($\Delta t = 10^{-3}$) and $32 \times 32$ spatial mesh. These down-sampled solutions are treated as true solutions for the training process. Meanwhile, we draw 10 test initial velocity samples independently, and the corresponding test data set of 10 samples is created in the same manner. However, the time horizon $T = 1.5$ for test samples is chosen\textemdash much larger than
the trained time horizon\textemdash with time stepsize $\Delta t = 10^{-3}$. This helps us test the accuracy and stability of neural network solutions beyond the training regime.
{\bf Neural network architecture.}
We use a shallow network of one layer with 5000 neurons for all cases to approximate the tangent slope of Burger's equations.
Note that we have compared the one-layer network with two- and three-layer networks with different numbers of neurons ranging from 100 to 5000. These deeper networks perform poorly with small data sets and are improved with large data sets in which the shallow one has comparable performance. Note that one-layer neural network approximation capabilities are rigorously justified by past universal approximation theories (see, e.g. \cite{Cybenko1989,hornik1989multilayer,Zhou17,johnson2018deep}) and our current work \cite{BuiUniversal21}.
Thus we shall use a one-layer neural network for all numerical results.
In addition, ReLU \cite{nair2010rectified} is used as the activation function.
\texttt{ADAM} optimizer is used with the learning rate of $10^{-4}$ and the training batch size is 40 samples.
For this example, reasonably optimal weights/biases are the ones giving the lowest accumulated mean square error after $1500$ time steps for 10 test data.
We take $\alpha = 10^5$ for the regularization parameter as this gives the best results from our numerical experiments (not shown here).
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{figures/2D_Bur/Burger_compare.pdf}
\caption{\textbf{Burger's equations}. Comparison of mean square error among different neural networks trained with $200$ data with/without noise. Recall that $\alpha = 0$ corresponds to the pure data-driven neural network training without model-constrained terms. Forward solver denotes the numerical solution on $32\times 32$ spatial mesh. }
\figlab{2D_Bur_compare_all}
\end{figure}
{\bf Comparison of different learned neural networks.}
\cref{fig:2D_Bur_compare_all} presents the comparison of mean square error obtained by different learned neural networks with the data set of 200 samples. It can be seen that, in general, the model-constrained neural networks are far better than their pure data-driven counterparts (i.e. with $\alpha = 0$). Additionally, long sequential machine learning trainings with $S = 10$ provide slightly better accuracy than $S = 1$, except for the noisy data with pure data-driven network in which the improvement is significant.
For model-constrained neural networks, long sequential training results with $S = 10$ in two settings $\LRp{d200,0\%, 10, 1, 10^5}$ and $\LRp{d200,2\%, 10, 1, 10^5}$ show an marginal improvement compared to short sequential training with $S = 1$ in two settings $\LRp{d200,0\%, 1, 1, 10^5}$ and $\LRp{d200,2\%, 1, 1, 10^5}$. Therefore, $S = 1$ is sufficient and we use it for the rest of numerical results with model-constrained neural networks. \cref{fig:2D_Bur_compare_all} shows that using $5\%$ noise causes the neural network corresponding to $\LRp{d200,5\%, 1, 1, 10^5}$ to perform poorly, while $1\%$ noise gives almost the same accuracy as $2\%$ noise. It is noticeable that the long sequential model-constrained training with $R = 5$ ($d200,2\%, 1, 5, 10^5$) yields higher accuracy than the others. However, large $R$ is more computationally expensive since many passes through the back-propagation computational graph are needed.
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{figures/2D_Bur/Burger_results_cases.pdf}
\caption{\textbf{Burger's equations}. The mean square error versus the number of time steps for various learned neural networks using 200, 600, and 1000 data samples with $S = 1$. }
\figlab{2D_Bur_d200d600}
\end{figure}
\begin{figure}[htb!]
\centering
\begin{tabular*}{\textwidth}{c c c c c}
\centering
&
\raisebox{-0.5\height}{\small $n_t = 0$} &
\raisebox{-0.5\height}{\small $n_t = 100$} &
\raisebox{-0.5\height}{\small $n_t = 500$} &
\raisebox{-0.5\height}{\small $n_t = 1500$}
\\
\centering
\rotatebox[origin=c]{90}{\small True $\pb$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_FD_127x127_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_FD_127x127_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_FD_127x127_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_FD_127x127_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{\small $\alpha =0,\delta = 0\%$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_data_600_alpha_0_noise_0_S_1_R_1_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_data_600_alpha_0_noise_0_S_1_R_1_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_data_600_alpha_0_noise_0_S_1_R_1_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_data_600_alpha_0_noise_0_S_1_R_1_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{\small $\alpha =0,\delta = 2\%$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_data_600_alpha_0_noise_2_S_1_R_1_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_data_600_alpha_0_noise_2_S_1_R_1_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_data_600_alpha_0_noise_2_S_1_R_1_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_data_600_alpha_0_noise_2_S_1_R_1_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{\small $\alpha = 1e^5,\delta = 0\%$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_data_600_alpha_100000_noise_0_S_1_R_1_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_data_600_alpha_100000_noise_0_S_1_R_1_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_data_600_alpha_100000_noise_0_S_1_R_1_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_data_600_alpha_100000_noise_0_S_1_R_1_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{\small $\alpha =1e^5,\delta = 2\%$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_data_1000_alpha_100000_noise_2_S_1_R_1_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_data_1000_alpha_100000_noise_2_S_1_R_1_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_data_1000_alpha_100000_noise_2_S_1_R_1_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_data_1000_alpha_100000_noise_2_S_1_R_1_step_t_1500.pdf}}
\end{tabular*}
\caption{\textbf{Burger's equations}. Predicted solutions at different time steps ($n_t$) obtained by various learned neural network tangents with $600$ training data samples and $S = 1$ and $\Delta t = 10^{-3}$. {\em Top row}: True high-resolution solution; {\em Second row}: pure data-driven network without data randomization; {\em Third row}: pure data-driven network with noisy data; {\em Fourth row}: model-constrained network without data randomization; {\em Fifth row}: model-constrained network with noisy data.}
\figlab{2D_Bur_samples}
\end{figure}
{\bf Long-time predictions with small and large training data sets.}
As discussed above, since long sequential machine learning training does not provide significant improvement,
we consider $S = 1$ for numerical results using large data sets
in \cref{fig:2D_Bur_d200d600}. As can be seen, compared to $200$ data samples, training with $600$ data samples provides more accurate predictions. Moreover, model-constrained neural networks with randomized data are the most accurate among others (model-constrained with noise-free data and pure machine learning with/without randomized data). We can also observe that using more than $600$ data samples does not provide significant improvements but is more expensive.
Unlike the case with $200$ data samples, long and short sequential model-constrained trainings with $R = 5$ and $R = 1$, respectively, provide similar results for $600$ data samples. This is expected as richer data reduces the significance of the model-constrained term.
As shown in \cref{fig:2D_Bur_samples}, predicted solutions obtained by the model-constrained approach (the fifth row) with data randomization are in good agreement with the ground-truth counterparts. On the contrary, the pure data-driven approach with data randomization (the third row) shows poor long-time predictions. We also observe that
both pure data-driven learning solutions and model-constrained solutions (the second and fourth rows, respectively) without randomization are unstable for long-time predictions. It is not surprising since both do not have sufficient regularizations compared to the randomized cases in which extra regularizations are implicitly performed (see \cref{sect:noise_data_sec}). Moreover, regularizations induced by data randomization shown in \cref{sect:noise_data_sec} stabilize the network predictions and this can be clearly seen by comparing the third and the second rows for the pure data-driven learning approach, and by comparing the fourth and the fifth rows for the model-constrained learning approach.
\cref{fig:2D_Bur_dU_samples} plots the contours of the learned and the true tangent slopes. Clearly, the learned model-constrained tangent slope with data randomization provides the best agreement with the true tangent slope. This is not surprising as both the governing equations (explicit via model-constrained term) and sufficient regularizations (implicit via data randomization) are incorporated.
\begin{figure}[htb!]
\centering
\begin{tabular*}{\textwidth}{c c c c c c}
\centering
&
\raisebox{-0.5\height}{\small True $\bs{G}(\pb)$} &
\raisebox{-0.5\height}{\small $\LRp{0, 0\%}$} &
\raisebox{-0.5\height}{\small $\LRp{0, 2\%}$} &
\raisebox{-0.5\height}{\small $\LRp{10^5, 0\%}$} &
\raisebox{-0.5\height}{\small $\LRp{10^5, 2\%}$}
\\
\centering
\rotatebox[origin=c]{90}{\small $n_t = 100$} &
\raisebox{-0.5\height}{\includegraphics[width=.16 \textwidth]{figures/2D_Bur/Bur_dU_FD_32x32_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.16 \textwidth]{figures/2D_Bur/Bur_dU_data_600_alpha_0_noise_0_S_1_R_1_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.16 \textwidth]{figures/2D_Bur/Bur_dU_data_600_alpha_0_noise_2_S_1_R_1_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.16 \textwidth]{figures/2D_Bur/Bur_dU_data_600_alpha_100000_noise_0_S_1_R_1_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.16 \textwidth]{figures/2D_Bur/Bur_dU_data_1000_alpha_100000_noise_2_S_1_R_1_step_t_100.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{\small $n_t = 500$} &
\raisebox{-0.5\height}{\includegraphics[width=.16 \textwidth]{figures/2D_Bur/Bur_dU_FD_32x32_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.16 \textwidth]{figures/2D_Bur/Bur_dU_data_600_alpha_0_noise_0_S_1_R_1_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.16 \textwidth]{figures/2D_Bur/Bur_dU_data_600_alpha_0_noise_2_S_1_R_1_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.16 \textwidth]{figures/2D_Bur/Bur_dU_data_600_alpha_100000_noise_0_S_1_R_1_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.16 \textwidth]{figures/2D_Bur/Bur_dU_data_1000_alpha_100000_noise_2_S_1_R_1_step_t_500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{\small $n_t = 1500$} &
\raisebox{-0.5\height}{\includegraphics[width=.16 \textwidth]{figures/2D_Bur/Bur_dU_FD_32x32_step_t_1500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.16 \textwidth]{figures/2D_Bur/Bur_dU_data_600_alpha_0_noise_0_S_1_R_1_step_t_1500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.16 \textwidth]{figures/2D_Bur/Bur_dU_data_600_alpha_0_noise_2_S_1_R_1_step_t_1500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.16 \textwidth]{figures/2D_Bur/Bur_dU_data_600_alpha_100000_noise_0_S_1_R_1_step_t_1500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.16 \textwidth]{figures/2D_Bur/Bur_dU_data_1000_alpha_100000_noise_2_S_1_R_1_step_t_1500.pdf}}
\end{tabular*}
\caption{\textbf{Burger's equations}. Contours of True and various learned tangent slopes. Contours are plotted at different time steps $n_t$ for four different combinations of regularization parameter $\alpha$ and noise level $\delta$. For all cases, we use $600$ data samples, $S = 1$, and $\Delta t = 10^{-3}$. \textit{First column}: True tangent slope ; \textit{Second column}: pure data-driven tangent slope without data randomization $\LRp{0, 0\%}$; \textit{Third column}: pure data-driven tangent slope with data randomization $\LRp{0, 2\%}$; \textit{Fourth column}: model-constrained tangent slope without data randomization $\LRp{10^5, 0\%}$; \textit{Fifth column}: model-constrained tangent slope with data randomization $\LRp{10^5, 2\%}$.}
\figlab{2D_Bur_dU_samples}
\end{figure}
{\bf Predictive flexibility in time for \texttt{mcTangent} approach}
As discussed above, one appealing feature of tangent slope learning is that once trained it can be used to solve for approximate solutions with smaller or larger time stepsizes, despite the fact that it is trained based on a particular spatial discretization. On the contrary, direct learning is attached to a space-time discretization. \cref{fig:2D_Bur_smaller_dt_samples} shows the model-constrained tangent slope learning solutions and contours of the corresponding learned tangent slope at various times for the setting $\LRp{d600, 2\%, 1,1, 10^5}$. Here we use half of the training time stepsize $\Delta t'' = \frac{1}{2}\Delta t = 5 \times 10^{-4}$. It can be seen that these predictions are indistinguishable from ones (the fifth row in \cref{fig:2D_Bur_samples} for prediction solutions and the fifth column in \cref{fig:2D_Bur_dU_samples} for predicted tangent slopes) obtained by using the training time stepsize $\Delta t = 10^{-3}$ with the same learned network.
\begin{figure}[htb!]
\centering
\begin{tabular*}{\textwidth}{c c c c c}
\centering
&
\raisebox{-0.5\height}{\small $t = .1$} &
\raisebox{-0.5\height}{\small $t = 0.5$} &
\raisebox{-0.5\height}{\small $t = 0.5005$} &
\raisebox{-0.5\height}{\small $t = 1.5$}
\\
\centering
\rotatebox[origin=c]{90}{\small True $\pb$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_FD_127x127_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_FD_127x127_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_FD_127x127_step_t_500_5.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_FD_127x127_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{\small True $\bs{G}(\pb)$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_dU_FD_32x32_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_dU_FD_32x32_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_dU_FD_32x32_step_t_500_5.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_dU_FD_32x32_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{\small Predicted $\pb$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_smallerdt_data_1000_alpha_100000_noise_2_S_1_R_1000_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_smallerdt_data_1000_alpha_100000_noise_2_S_1_R_1000_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_smallerdt_data_1000_alpha_100000_noise_2_S_1_R_1000_step_t_500.5.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_smallerdt_data_1000_alpha_100000_noise_2_S_1_R_1000_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{\small Predicted $\Psi(\pb)$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_smallerdt_dU_data_1000_alpha_100000_noise_2_S_1_R_1000_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_smallerdt_dU_data_1000_alpha_100000_noise_2_S_1_R_1000_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_smallerdt_dU_data_1000_alpha_100000_noise_2_S_1_R_1000_step_t_500.5.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_smallerdt_dU_data_1000_alpha_100000_noise_2_S_1_R_1000_step_t_1500.pdf}}
\end{tabular*}
\caption{\textbf{Burger's equations}. Predicted solutions and tangent slope using by \texttt{mcTangent} neural networks with $\LRp{d600, 2\%, 1,1, 10^5}$, and time step $\Delta t'' = \frac{1}{2}\Delta t = 5 \times 10^{-4}$. \textit{First row}: True high-resolution solutions; \textit{Second row}: contours of True tangent slope, $\bs{G}(\pb)$; \textit{Third row}: Predicted \texttt{mcTangent} solutions, $\pb$; \textit{Fourth row}: contours of \texttt{mcTangent} tangent slope, $\Psi(\pb)$.}
\figlab{2D_Bur_smaller_dt_samples}
\end{figure}
{\bf Implicit time integration with learned network.}
Another appealing feature of tangent slope learning is that once trained it can be deployed with any time discretization schemes.
We use the learned network from the setting $\LRp{d600, 2\%, 1, 1, 10^5}$ together with the backward Euler method with a larger time stepsize $\Delta t' = 12.5 \Delta t = 1.25 \times 10^{-2}$, where $\Delta t = 10^{-3}$ is the training stepsize. Shown in \cref{fig:2D_Bur_implicit_samples} are predicted solutions at $t=\LRc{0, 0.1, 0.5, 1.5}$ corresponding to $0, 100, 500, 1500$th time steps. We observe that solutions using the forward Euler scheme, regardless of using the true tangent slope or learned one (second and third rows, respectively), are unstable as the time stepsize $\Delta t'$ is too big for stability. On the contrary, using the backward Euler scheme, \texttt{mcTangent} solutions are comparable to the true counterparts (fourth and fifth rows, respectively).
Clearly, due to large time stepsize, both are more diffusive compared to the true solutions with small time stepsize $\Delta t$ in the first row.
\begin{figure}[htb!]
\centering
\begin{tabular*}{\textwidth}{c c c c c}
\centering
&
\raisebox{-0.5\height}{\small $t = 0$} &
\raisebox{-0.5\height}{\small $t = 0.1$} &
\raisebox{-0.5\height}{\small $t = 0.5$} &
\raisebox{-0.5\height}{\small $t = 1.5$}
\\
\centering
\rotatebox[origin=c]{90}{\small True solution $\pb$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_FD_127x127_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_FD_127x127_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_FD_127x127_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_FD_127x127_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{FE - True $\bs{G}$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_Explicit_FD_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_Explicit_FD_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_Explicit_FD_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_Explicit_FD_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{FE - Learned $\Psi$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_Explicit_NN_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_Explicit_NN_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_Explicit_NN_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_Explicit_NN_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{\small BE - True $\bs{G}$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_implicit_FD_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_implicit_FD_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_implicit_FD_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_implicit_FD_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{\small BE - Learned $\Psi$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_implicit_NN_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_implicit_NN_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_implicit_NN_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_implicit_NN_step_t_1500.pdf}}
\end{tabular*}
\caption{\textbf{Burger's equations}. Predicted solutions at different times obtained by Forward Euler (FE) scheme and Backward Euler (BE) scheme using large stepsize $\Delta t' = 12.5 \Delta t = 1.25 \times 10^{-2}$ with the true tangent slope $\bs{G}$ and the learned neural network $\Psi$ for the setting $\LRp{d600, 2\%, 1, 1, 10^5}$.}
\figlab{2D_Bur_implicit_samples}
\end{figure}
{\bf Direct learning versus \texttt{mcTangent} slope learning.}
Recall that by direct learning we mean learning the map from $\ui{i}$ to $\ui{i+1}$ for two consecutive time steps.
We investigate the difficulty and complexity of direct learning.
Specifically, we use a data set with $600$ samples with/without data randomization to learn the neural network with one layer of 5000 neurons that maps velocities from one step to the next. As shown in \cref{fig:2D_Bur_Direct_EF_NNs}, the direct learning approach (with the best combination of hyperparameters) for the setting $\LRp{d600, 2\%, 1,3,2}$ is less accurate for both short-time and long-time predictions compared to the tangent learning counterpart with even a smaller data set of $200$ samples with the setting $\LRp{d200, 2\%, 1,1,10^5}$.
Interestingly, unlike the tangent learning approach, the direct learning approach, both pure data-driven and model-constrained approaches, trained with randomized data is less accurate compared to noise-free data in short sequential training $S = 1$. On the other hand, data randomization does not have visible benefits on long sequential training $S = 10$. Specifically, both $\LRp{d600, 0\%, 10,1,0}$ and $\LRp{d600, 2\%, 10,1,0}$ settings behave similarly.
Also seen in \cref{fig:2D_Bur_Direct_EF_NNs}, among pure data-driven networks ($\alpha = 0$) with direct learning, long sequential machine learning training with $S = 10$ is the most accurate.
Model-constrained network with direct learning for the setting $\LRp{d600, 2\%, 1, 1, 10}$ is much more accurate compared to the pure data-driven network with direct learning for the same setting.
Moreover,
sequential model-constrained networks for $R = 2, 3$ corresponding to two settings $\LRp{d600,2\%, 1, 2, 2}$ and $\LRp{d600,2\%, 1, 3, 2}$ are comparable to much longer sequential machine learning network with $S = 10$ for the setting $\LRp{d600,2\%, 10, 1, 0}$.
In the presented results,
it is important to point out that for direct learning,
care must be taken in choosing a good regularization parameter $\alpha$. For example, $\alpha = 2$ is good for $R = 2, 3$, but $\alpha = 10$ is good for $R = 1$. On the contrary, tangent learning is more robust. In particular, a single $\alpha = 10^5$ works well for all settings.
Solutions predicted by direct and tangent learnings
(both with model-constrained terms)
for $\LRp{d600, 2\%, 1, 3, 2}$ and $\LRp{d200, 2\%, 1, 1, 10^5}$, respectively, are shown in \cref{fig:2D_Bur_Direct_samples}. As can be observed, tangent learning solutions with even smaller data set $\LRp{d200, 2\%, 1, 1, 10^5}$ are much more accurate than the direct learning with $\LRp{d600, 2\%, 1, 3, 2}$. This is due to the fact that direct learning tries to learn a mixed space-time discretization, which is more difficult than learning only the spatial discretization in tangent learning.
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{figures/2D_Bur/Burger_compare_direct.pdf}
\caption{\textbf{Burger's equations}. The mean square error versus time steps obtained by the direct learning approach (Direct) using 600 training samples and the tangent learning approach (\texttt{mcTangent}) using 200 training samples.}
\figlab{2D_Bur_Direct_EF_NNs}
\end{figure}
\begin{figure}[htb!]
\centering
\begin{tabular*}{\textwidth}{c c c c c}
\centering
&
\raisebox{-0.5\height}{\small $t = 0$} &
\raisebox{-0.5\height}{\small $t = 0.1$} &
\raisebox{-0.5\height}{\small $t = 0.5$} &
\raisebox{-0.5\height}{\small $t = 1.5$}
\\
\centering
\rotatebox[origin=c]{90}{\small True $\pb$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_FD_127x127_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_FD_127x127_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_FD_127x127_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_FD_127x127_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{\small \texttt{mcTangent} $\pb$, d200} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_direct_data_200_alpha_100000_noise_2_S_1_R_1_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_direct_data_200_alpha_100000_noise_2_S_1_R_1_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_direct_data_200_alpha_100000_noise_2_S_1_R_1_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_direct_data_200_alpha_100000_noise_2_S_1_R_1_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{Direct $\pb$, d600} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_direct_data_600_alpha_2_noise_2_S_1_R_3_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_direct_data_600_alpha_2_noise_2_S_1_R_3_step_t_100.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_direct_data_600_alpha_2_noise_2_S_1_R_3_step_t_500.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_Bur/Bur_direct_data_600_alpha_2_noise_2_S_1_R_3_step_t_1500.pdf}}
\end{tabular*}
\caption{\textbf{Burger's equations}. Solutions at different times, \textit{First row}: True high-resolution solutions; \textit{Second row}: learned tangent slope neural network solutions with $\LRp{d200, 2\%, 1, 1, 10^5}$; \textit{Third row}: learned direct neural network solutions with $\LRp{d600, 2\%, 1, 3, 2}$.}
\figlab{2D_Bur_Direct_samples}
\end{figure}
\subsection{Navier-Stokes equation}
\seclab{2D_NS}
The vorticity form of the 2D Navier-Stokes equation for viscous and incompressible fluid \cite{FouierOP} can be written as
\begin{equation*}
\begin{aligned}
\partial_t u(x,t) + v(x,t) \cdot \nabla u(x,t) &= \nu \Delta u(x,t) + f(x), & \quad x \in \LRp{0,1}^2, t \in (0, T] \\
\nabla \cdot v(x,t) & = 0, & \quad x \in \LRp{0,1}^2, t \in (0, T] \\
u(x,0) & = u_0(x), & \quad x \in \LRp{0,1}^2
\end{aligned}
\end{equation*}
where $v\LRp{x,t}$ is the velocity field, $u = \nabla \times v$ is the vorticity, $u_0 $ is the initial vorticity, $f(x) = 0.1 \LRp{\sin \LRp{2 \pi \LRp{x_1 + x_2}} + \cos\LRp{2 \pi \LRp{x_1 + x_2}}}$ is the forcing function and $\nu = 10^{-3}$ is the viscosity coefficient. Our goal is to solve for the vorticity $u\LRp{x,y,t}$ given the initial condition $u_0$ at $t = 0$ by a trained tangent network $\Psi$ .
{\bf Data generation.}
Data pair $\LRp{\pb, \bs{y}}$ is generated by a similar procedure outlined for Burger's equation problem in \cref{sect:Burger_eq}.
In particular, we draw samples of $\pb_0$ using the truncated Karhunen-Loève expansion
\[\pb_0 = \sum_{i=1}^{15} \sqrt{\lambda_i} \, {\bf \omega_i}(x) \, z_i,\]
where $z_i \sim \mathcal{N} \LRp{0, 1}, i = 1, \hdots, 15$, and $\LRp{\lambda, {\bf \omega}}$ is eigenpairs obtained by the eigendecomposition of the covariance operator $ 7^{\frac{3}{2}} \LRp{-\Delta + 49 \textbf{I}}^{-2.5}$ with periodic boundary conditions. Next, given initial vorticity $\pb_0$, we solve the Navier-Stokes equation by the stream-function formulation with a pseudospectral method \cite{FouierOP}.
High resolution solutions are obtained on a uniform $128 \times 128$ spatial mesh and uniform 1000 time steps in $\LRp{0, 2}$.
The high-resolution solutions are then down-sampled on a coarser mesh $32 \times 32$ in space and 200 uniform time steps, and they are used as the training data. To verify the accuracy of the learned neural network, we draw 10 test samples independently. It turns out that the Navier-stokes equation is much more challenging than Burger's equation, thus we use 200 time steps for each training data as opposed to 100 for the Burger equation.
Similar to the above, to challenge the learned network we use 1500 time steps for testing, and thus the testing time horizon is far beyond the training time horizon.
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{figures/2D_NS/NS_results_cases.pdf}
\caption{\textbf{Navier-Stokes equation}. The mean-square error versus time steps obtained by the various learned neural network with $S = 1$.}
\figlab{2D_NS_d200d600}
\end{figure}
{\bf Neural network architecture.}
With the same observation for the Burger equation in \cref{sect:Burger_eq}, we use a shallow network of one layer with 5000 neurons using ReLU activation function.
\texttt{ADAM} optimizer with default parameters is used with the learning rate of $2 \times 10^{-4}$, while the training batch size is 2 samples. The chosen ``optimal" network is the one having the lowest accumulated mean square error after $1500$ time steps for 10 testing samples. Following the wave and Burger examples, we pick a relatively large value for the model-constrained regularization parameter $\alpha = 10^5$.
{\bf Long-time predictions.}
\begin{figure}[htb!]
\centering
\begin{tabular*}{\textwidth}{c c c c c}
\centering
&
\raisebox{-0.5\height}{\small $n_t = 0$} &
\raisebox{-0.5\height}{\small $n_t = 200$} &
\raisebox{-0.5\height}{\small $n_t = 1000$} &
\raisebox{-0.5\height}{\small $n_t = 1500$}
\\
\centering
\rotatebox[origin=c]{90}{\small True $\pb$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_FD_127x127_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_FD_127x127_step_t_200.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_FD_127x127_step_t_1000.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_FD_127x127_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{$\alpha =0,\delta = 0\%$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_data_600_alpha_0_noise_0_S_1_R_1_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_data_600_alpha_0_noise_0_S_1_R_1_step_t_200.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_data_600_alpha_0_noise_0_S_1_R_1_step_t_1000.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_data_600_alpha_0_noise_0_S_1_R_1_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{$\alpha =0,\delta = 2\%$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_data_600_alpha_0_noise_2_S_1_R_1_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_data_600_alpha_0_noise_2_S_1_R_1_step_t_200.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_data_600_alpha_0_noise_2_S_1_R_1_step_t_1000.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_data_600_alpha_0_noise_2_S_1_R_1_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{$\alpha =1e^5,\delta = 0\%$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_data_600_alpha_100000_noise_0_S_1_R_1_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_data_600_alpha_100000_noise_0_S_1_R_1_step_t_200.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_data_600_alpha_100000_noise_0_S_1_R_1_step_t_1000.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_data_600_alpha_100000_noise_0_S_1_R_1_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{$\alpha =1e^5,\delta = 2\%$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_data_600_alpha_100000_noise_2_S_1_R_1_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_data_600_alpha_100000_noise_2_S_1_R_1_step_t_200.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_data_600_alpha_100000_noise_2_S_1_R_1_step_t_1000.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_data_600_alpha_100000_noise_2_S_1_R_1_step_t_1500.pdf}}
\end{tabular*}
\caption{\textbf{Navier-Stokes equation}. Predicted solutions at different time steps obtained by various trained networks with 600 data samples. \textit{First row}: ground truth; \textit{Second row}: pure data-driven network solutions with noise-free data; \textit{Third row}: pure data-driven network solutions with randomized data; \textit{Fourth row}: model-constrained network solutions with noise-free data; \textit{Fifth row}: model-constrained network solutions with randomized data.}
\figlab{2D_NS_samples}
\end{figure}
\cref{fig:2D_NS_d200d600} shows the mean-square error of predictions and ground truth solutions as a function of time steps. It can be seen that training with a large data set with $600$ samples provides much more accurate solutions than with small data set with $100$ samples. On the one hand, among learned neural networks trained with $100$ data samples, the model-constrained network with data randomization for $\LRp{d100,2\%, 1,1,10^5}$ setting is far closer to the true solution than the other networks. This implies that the model-constrained approach has a significant contribution to producing accurate predictions in the context of small data.
In the case of richer data set with 600 samples, networks with two settings $\LRp{d600,0\%, 1,1,0}$ and $\LRp{d600,0\%, 1,1,10^5}$ trained with noise-free data show a good performance in the short time predictions, while the long-time predictions deteriorate. Noticeably, between these two networks, the model-constrained one has more accurate predictions starting from the $500$th time step.
In the meantime, with the same data set with 600 samples, pure data-driven neural networks trained with higher noise level data give a higher error, for example, $\LRp{d600,2\%, 1,1,0}$ neural network predictions are less accurate than those obtained from $\LRp{d600,1\%, 1,1,0}$. In contrast, model-constrained network with 2\% noise level $\LRp{d600,2\%, 1,1,10^5}$ is superior to 1\% noise level $\LRp{d600,1\%, 1,1,10^5}$.
Another point is that as we increase the sequential model-constrained value to $R=5$, we obtain good predictions in both short-time and long-time intervals.
Two model-constrained networks with $\LRp{d600,0\%, 1,5,10^5}$ and $\LRp{d600,2\%, 1,5,10^5}$ are comparable to the network with much larger data set $\LRp{d1000,0\%, 1,1,10^5}$ without randomization. \begin{figure}[htb!]
\centering
\begin{tabular*}{\textwidth}{c c c c c}
\centering
&
\raisebox{-0.5\height}{\small $t = 0$} &
\raisebox{-0.5\height}{\small $t = 2$} &
\raisebox{-0.5\height}{\small $t = 10$} &
\raisebox{-0.5\height}{\small $t = 15$}
\\
\centering
\rotatebox[origin=c]{90}{True solutions $\pb$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_Explicit_FD_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_Explicit_FD_step_t_200.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_Explicit_FD_step_t_1000.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_Explicit_FD_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{FE - Learned $\Psi$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_Explicit_NN_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_Explicit_NN_step_t_200.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_Explicit_NN_step_t_1000.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_Explicit_NN_step_t_1500.pdf}}
\\
\centering
\rotatebox[origin=c]{90}{\small BE - Learned $\Psi$} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_implicit_NN_step_t_0.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_implicit_NN_step_t_200.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_implicit_NN_step_t_1000.pdf}} &
\raisebox{-0.5\height}{\includegraphics[width=.20 \textwidth]{figures/2D_NS/NS_implicit_NN_step_t_1500.pdf}}
\end{tabular*}
\caption{\textbf{Navier-Stokes equation}. Comparison of various neural network solutions. \textit{Top row}: True solutions with spectral method (Crank–Nicolson time integration scheme) with the true tangent slope; \textit{Second row}: Forward Euler (FE) scheme (a different scale bar in the third and fourth columns); \textit{Third row}: Backward Euler (BE) scheme with the learned neural network for 20 times larger time stepsize $\Delta t' = 20 \Delta t = 0.2$.}
\figlab{2D_NS_implicit_samples}
\end{figure}
However, the noisy data network $\LRp{d600,2\%, 1,5,10^5}$ outperforms the noise-free one $\LRp{d600,0\%, 1,5,10^5}$ in the long-time predictions.
In summary, model-constrained network with data randomization outperforms all other networks.
Given a test initial vorticity, the plots of predicted solutions obtained by different learned networks are shown in \cref{fig:2D_NS_samples}. As can be seen, the model-constrained network with the setting $\LRp{d600,2\%, 1,1,10^5}$ provides the most accurate solutions as opposed to others trained from the same data set.
{\bf Implicit time integration with learned network.}
We used the learned network for backward Euler scheme with 20 times larger time stepsize, $\Delta t' = 20 \Delta t = 0.2$, compared to training stepsize $\Delta t = 0.01$. As shown in \cref{fig:2D_NS_implicit_samples},
forward Euler scheme with the learned network shows severe instability, while the backward Euler scheme with the learned network solutions is in good agreement with the spectral solution with Crank–Nicolson scheme with a much smaller time stepsize.
\subsection{Information on parameter tuning, randomness, and cost}
\seclab{Comp_cost}
\subsubsection{Parameter tuning}
\seclab{tuning}
The purpose of this section is to determine a good set of hyperparameters including the learning rate, batch size, the number of layers, and the number of neurons on each layer. To that end, we set the random seed to 0 in order to have a fair initialization for all networks. For initialization, weights are drawn randomly from zero-mean Gaussian distribution with a variance of 0.01, while biases are set to zero. For all cases, we take $S = 1, R = 1$, and noise-free data set with $600$ samples. We carry out the tuning process manually for only Burgers and Navier-Stokes examples as the transport example admits an analytical solution. We pick the learning rate in $\LRc{10^{-4},2 \times 10^{-4}, 5 \times 10^{-4}, 10^{-3}}$, batch size in $\LRc{2, 10, 40, 100}$, the number of layers in $\LRc{1,2,3}$, the number of neurons per layers in $\LRc{50, 200, 1000, 5000, 10000}$, and the model-constrained regularization parameter $\alpha$ in $\LRc{10, 10^3, 10^5, 10^7}$. We pick the combination of parameters that provides the best testing accuracy (see also \cref{sect:Burger_eq} and \cref{sect:2D_NS} for the discussion on testing accuracy) in each numerical problem. The chosen parameter set is then used for training different values of $S, R$ and noise level $\delta$.
\subsubsection{Robustness with random initializations and data randomization}
In this section, we study the effect of weights/biases random initialization and data randomization on the performance of the chosen neural network architectures in \cref{sect:tuning}. We provide the study for Burger's equations in \cref{sect:Burger_eq} since the result for the Navier-Stokes equation in \cref{sect:2D_NS} would be similar.
{For random initialization of weights/biases, we initialize the neural network with 32 different random seeds ranging from 0 to 31. For each random seed, we use the same set of hyper-parameters found in \cref{sect:tuning}. As shall be shown, our model-constrained approach is robust in random initialization, that is, all random seeds work equally well.
Thanks to this robustness, we simply initialize weights/biases with random seed 0 and study the effect of 32 different noise random seeds ranging from 0 to 31 for data randomization.}
As an example, we compare the mean and variance of the mean square error between the pure data-driven machine learning case $\LRs{d600, 2\%, 0, 1, 0}$ and the corresponding model-constrained case $\LRs{d600, 2\%, 10^5, 1, 1}$. The mean and variance results in \cref{fig:Burger_randomization} and \cref{fig:Burger_noise} show that \texttt{mcTangent} networks are not only accurate but also more reliable with a smaller variance compared to the pure data-driven counterparts. Consequently\textemdash again thanks to the model-constrained term\textemdash the performance of \texttt{mcTangent} networks are robust to both
weights/biases random initialization and data randomization.
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{figures/Burger_statistic_comparison.pdf}
\caption{The mean and variance of mean square error of predictions for \texttt{mcTangent} and pure data-driven machine learning approaches, obtained by 32 different neural networks corresponding to 32 different weights/biases random initializations.}
\figlab{Burger_randomization}
\end{figure}
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{figures/Burger_noise_comparison.pdf}
\caption{The mean and variance of mean square error of predictions for \texttt{mcTangent} and pure data-driven machine learning approaches, obtained by training a neural network (with weights/biases being initialized with random seed 0) with 32 random realizations of data corresponding to 32 different random seeds.}
\figlab{Burger_noise}
\end{figure}
\subsubsection{Training and testing cost}
\cref{tab:train_cost} presents the training computational cost for Burgers' problem using different values of $S$ and $R$. The \texttt{mcTangent} neural network is learned with 200 training samples with $2\%$ additive noise, $\alpha = 10^5$, learning rate $10^{-4}$, and batch size $40$. It can be seen that in the purely data-driven approach, i.e. $S= 1, R= 0$, the computational time per epoch is small, but the number of required epochs for convergence is larger. It is not surprising that adding larger $R$ and $S$ leads to a significant increase in computational cost per epoch. However, in this problem, since the overall convergence rate (measured in the terms of the number of epochs) is faster, the total amount of time for training model-constrained neural networks are at most three times larger than the pure machine learning method. To be more specific, $S = 1, R = 0$ network requires $11.67$ hours compared to $22.22$ hours for $S = 1, R = 1$ network. On the other hand, $35$ hours and $34$ hours are needed to train the cases $S = 10, R = 1$ and $S = 1, R = 5$, respectively. We note that all model-constrained networks corresponding to $S=1, R = 1$, $S =10, R =1$ and $S = 1, R =5$ provide comparable accuracy levels which are significantly better than that obtained by the pure data-driven machine learning approach corresponding to $S = 1, R = 0$, and this is shown in \cref{fig:2D_Bur_compare_all}.
To verify the computational benefits in the prediction stage, we compare the computational time between the ground truth solution using the truth tangent slope $\bs{G}\LRp{\ut{}}$ and \texttt{mcTangent} tangent slope $\NN{\ut{}}$ in \cref{tab:test_cost}. It can be seen that the \texttt{mcTangent} tangent slope is much faster (more than $10$ times faster) than the truth tangent slope for the 2D Navier-Stokes problem. For the 2D Burgers' problem, the \texttt{mcTangent} tangent slope evaluation is negligibly faster than the truth.
That is, even with small-scale 2D problems with fast finite difference evaluations, the neural network is still faster. It is important to point out that the computational cost for \texttt{mcTangent} neural network remains unchanged, $2\times 10^{-4}$ seconds, for either Burgers or Navier-Stokes equations. We expect the computational gain is much more notable for 3D complex problems where the evaluation of the truth tangent slope is much more demanding. The gain is even more significant for implicit methods as in these cases not only the evaluation of the tangent slope but also the evaluation of its Jacobian is needed. This poses great challenges for traditional numerical methods, but for the \texttt{mcTangent} approach, the evaluations of a feed-forward network and its Jacobian are trivial and fast.
\begin{table}[htb!]
\caption{Training cost for Burgers' equations using different values of $S$ and $R$. \texttt{mcTangent} neural network is learned with 200 samples with $2\%$ additive noise, $\alpha = 10^5$, learning rate $10^{-4}$, and batch size $40$.}
\tablab{train_cost}
\centering
\begin{tabular}{|c|c|c|c|}
\hline
& 1 Epoch (seconds) & Number of Epoch & Training time (hours) \\ \hline
$S=1, R = 0$ & 0.07 & $6.0 \times 10^{5}$ & 11.67 \\ \hline
$S = 1, R = 1$ & 0.20 & $4.0 \times 10^{5}$ & 22.22 \\ \hline
$S = 10, R = 1$ & 0.86 & $1.5 \times 10^{5}$ & 35.83 \\ \hline
$S = 1, R = 5$ & 0.62 & $2.5 \times 10^{5}$ & 34.44 \\ \hline
\end{tabular}
\end{table}
\begin{table}[htb!]
\caption{Computational cost of ground truth tangent slope $\bs{G}(\ut{i})$ (mesh grid: $32 \times 32$), and trained neural network $\Psi\LRp{\ut{i}}$}
\tablab{test_cost}
\centering
\begin{tabular}{|c|c|c|}
\hline
& $G (\ui{})$ (seconds) & $\Psi(\ui{})$ (seconds) \\ \hline
Burgers' equation & $2.1 \times 10^{-4}$ & $2 \times 10^{-4}$ \\ \hline
Navier-Stokes equation & $7.0 \times 10^{-3}$ & $2 \times 10^{-4}$ \\ \hline
\end{tabular}
\end{table}
\section{Conclusions}
\seclab{conclusions}
We have presented a model-constrained tangent slope learning (\texttt{mcTangent}) approach to simulate dynamical systems in real-time. At the heart of \texttt{mcTangent} is a careful craft synergizing several desirable strategies: i) a tangent slope learning to take advantage of the neural network speed and time-accurate nature of the method of lines;
ii) a model-constrained approach to encode the neural network tangent slope with the underlying physics; iii) sequential learning strategies to promote long-time stability and accuracy; and iv) data randomization approach to implicitly regularize the smoothness of the neural network tangent and its likeliness to the truth tangent up second order derivatives in order to further enhance the stability and accuracy of \texttt{mcTangent} solutions. Rigorous results are provided to analyze and justify the proposed approach. Several numerical results for transport equation, viscous Burgers equation, and Navier-Stokes equation are presented to study and demonstrate the capability of the proposed \texttt{mcTangent} learning approach. Further theoretical analysis of \texttt{mcTangent} with both sequential learning strategies is ongoing to provide a deeper understanding of the approach. Strategies to improve the accuracy and to strongly encode the underlying governing equations are also part of future work.
\section*{Disclosure statement}
No potential conflict of interest was reported by the author(s).
\bibliographystyle{plain}
|
\section{Introduction}
\label{sec:intro}
With the advancement of sixth-generation (6G) mobile communication technology, data transmission rate in the conventional communication systems is increasing but approaching the Shannon limit. Meanwhile, the remaining available spectrum resources are becoming increasingly scarce. To solve this dilemma, semantic communication technologies are proposed~\cite{xie2020lite}, which aims to transmit the extracted semantic information relevant to the communications goal. Because the data amount that needs to be transmitted can be reduced significantly while ensuring the effectiveness of communications~\cite{xie2021deep}, semantic communications can be widely used in intelligent wireless networks, to enable smart transportation \cite{lin2020edge}, smart logistic \cite{song2020applications}, smart cities \cite{an2019toward}, smart homes \cite{sezer2015development}, and smart healthcare \cite{thangaraj2016agent}.
Existing semantic communication systems \cite{xie2021deep}, \cite{weng2021semantic} are pretrained with labeled datasets with certain channel models. However, a main drawback is that the accuracy and performance of a pretrained semantic model decrease when the background knowledge or communication environment changes, i.e., mismatch between the knowledge base/channel model used in the training and the actual knowledge base/channel model. To reduce the gap in performance, fine-tuning of the model parameters can be done based on the real channel models \cite{dorner2017deep} and new background knowledge \cite{xie2021deep}. However, edge and Internet of Things (IoT) devices with limited computation power might not have enough resources for fine-tuning. Moreover, the results of fine-tuning depend highly on the amount of labeled data of the new knowledge base. To solve the aforementioned problems, inspired by the model trading framework in collaborative edge learning \cite{lim2020federated}, we can adopt a trading system in which model providers trade the trained model to other devices. Specifically, the semantic model provider has more resources to train quality semantic models with the relevant knowledge base and channel models, and the edge devices can obtain the semantic model (semantic encoder/decoder) from the model providers. Using the semantic model, edge devices can extract semantic information from the collected raw data. This enables semantic information exchange between edge devices. Furthermore, as the semantic information is helpful for the decision making of smart agents~\cite{yun2021attention}, the trading of semantic information should also be studied. Using the semantic models, the edge devices collect and trade the semantic information with interested information buyers. For example, one vehicle can buy semantic information \cite{vancea2018semantic}, \cite{liao2021road} from nearby vehicles/smart sensors about the conditions of the surrounding environment.
To promote the above two types of trade in the semantic communication system, i.e., semantic model trading and semantic information trading, we should design novel and effective incentive mechanisms:
\begin{enumerate}
\item {\it{Semantic model trading:}} To encourage the participation of semantic model providers, incentive mechanisms are designed so that they are rewarded for supplying quality semantic models. In general, edge devices are willing to pay more for semantic models that can achieve better semantic performance. We are the first to propose a deep learning (DL) based auction mechanism to determine an allocation of the semantic model to the edge devices and the price to be paid by the edge devices to the model providers. We show analytically that the DL-based auction attains the properties of truthfulness while maximizing the revenue of the model providers.
\item {\it{Semantic information trading:}} To facilitate the semantic information trading between multiple semantic information buyers and edge devices, e.g., vehicles that are interested in collecting semantic information about the conditions of the surrounding environment \cite{vancea2018semantic}, \cite{liao2021road}, we introduce a double auction mechanism to model the competition between the buyers and edge devices. In the auction, we propose semantic based valuation functions, i.e., the valuation of the information is a function of semantic performance of the edge devices. In particular, the semantic information buyers are willing to pay more for the semantic information with higher accuracy, and hence the edge devices have more incentive to obtain better models from the semantic model trading. Moreover, the proposed double auction mechanism shows the desired properties of individual rationality, incentive compatibility, and budget balance, which are all significant properties to achieve sustainable and rational trading.
\end{enumerate}
While many recent works have focused on improving the performance of semantic communication systems \cite{xie2020lite}, \cite{xie2021deep}, \cite{farsad2018deep}, few works have addressed the designs of incentive mechanisms for semantic communication systems. By achieving the aforementioned two kinds of trade, we propose a novel hierarchical trading system to enhance the economically-sustainable development of semantic communication systems. The main contributions of our paper are:
\begin{itemize}
\item We propose an incentive design framework for the semantic model trading and semantic information trading to support the deployment of semantic communication systems. Our designed mechanisms support the development of semantic communication systems by motivating the participation of model providers to build and share high-quality semantic engines, buyers to acquire relevant and useful semantic information, and semantic information sellers to facilitate other stakeholders in the semantic information exchange.
\item We model the competition in the semantic model trading and semantic information trading with auction mechanisms. Different from conventional auctions, our auction can maximize the revenue of semantic model providers while achieving the properties of individual rationality and incentive compatibility. Simulation results are provided based on a case study on semantic text transmission where we derive the valuation functions based on the sentence similarity score and bilingual evaluation understudy (BLEU) score \cite{papineni2002bleu}.
\item We propose an effective feature reduction method for data transmission under a limited data transfer budget. In contrast to existing works of feature reduction techniques for semantic communication systems \cite{jiang2021deep}, \cite{yang2021semantic}, our method does not increase communication cost and reduce the performance gap between partial feature and full feature.
\end{itemize}
Compared with our previous work \cite{liew2022economics}, the significant extensions in this paper include:
\begin{itemize}
\item In contrast to previous work in which the incentive mechanism is customized for wireless powered devices, we propose a general framework that can be applied to semantic communication systems with different purposes.
\item While the previous work focuses on semantic information transfer, we consider both and joint semantic model trading and semantic information trading in this paper.
\item To model realistic semantic communication systems, multiple semantic information buyers and sellers are considered instead of a single buyer setting in the previous work.
\end{itemize}
Our paper is organised as follows. In Section \ref{sec:relatedwork}, we discuss the related works of semantic communication systems and incentive mechanism design. In Section \ref{sec:sysmodel}, we detail the system model and problem formulation. In Section \ref{sec:casestudy}, we present a case study of semantic model trading and semantic information trading for semantic text transmission. In Section \ref{sec:results}, we present the numerical results, and Section \ref{sec:conclusion} concludes the paper.
\section{Related Work}\label{sec:relatedwork}
\subsection{Vehicular Networks}
With the development of vehicular infrastructure in recent years, vehicles can be seen as important network players with computing, caching and communication capabilities~\cite{li2021adaptive,ye2021joint}. However, as the number of vehicles increases, the vehicular network relies heavily on reliable real-time communication and interactions for complex operations~\cite{wu2020dynamic,nanda2019internet}, such as route planning and collision avoidance. Thus, timely and accurate information updates are vital to the development of the vehicular networks. This implies that the conventional communication paradigm which seeks the lowest latency is no longer a sustainable development direction. To make fast and accurate decisions in vehicular networks, it is important to leverage the semantic meaning of information~\cite{pappas2021goal}. The authors in~\cite{zhu2021video} design a resource allocation algorithm for semantic video transmission in vehicular networks. By using the proposed algorithm \cite{zhu2021video}, the semantic understanding accuracy of the video transmission is optimized by a multi-agent deep Q-network. The simulation results show that the proposed method can achieve as high as 70\% improvement for the density of correctly detected objects, compared with the conventional QoS and QoE based resource allocation methods.
However, it is not realistic to train a usable semantic model for each vehicle, due to the limited computing resources and the dynamic positioning of vehicles~\cite{tayyaba20205g}. Therefore, we will consider a semantic model trading system in this paper. Moreover, considering the importance of semantic information in the vehicular networks, vehicles can then sell the semantic information to potential buyers. The trading of semantic information is gaining attraction especially for the sustainable development of large-scale multi-agent systems.
\subsection{Deep Learning Enabled Semantic Communication Systems}
Conventional communication systems focus on transmitting bits or symbols with minimum error from the transmitter to the receiver, and the performance is evaluated at the bits or symbols level. In contrast to the traditional communication systems, semantic communication system aims to communicate at the semantic level, where performance is evaluated by the recovery of the meanings of the data instead of bits accuracy. Semantic communication systems for text \cite{xie2021deep}, speech signals \cite{weng2021semantic}, and multimodal data \cite{xie2021task} first encode the data by a semantic encoder and send the encoded semantic information to the receivers. The receivers then decode the received signals with semantic decoders to recover the original data. Typically, the semantic encoders and decoders are implemented by end-to-end DL networks and trained with labeled data.
To improve the encoding efficiency, several works focus on reducing the size of the data during transmission. The authors in \cite{jiang2021deep} mask the bits according to the original sentence length to save the transmission resources. For the image classification task, the authors in \cite{yang2021semantic} use the gradient of the neural network to select import features. However, the proposed method requires extra storage cost to store the gradients of weights of the network. Most of the existing data reduction techniques are implemented together with the training process. A drawback is that, after the model is trained and parameters are fixed, further reduction of data size degrades the performance of the networks. To solve this problem, we develop an effective data reduction technique to reduce the performance gap in this paper.
\begin{figure*}[htb]
\centering
\centerline{\includegraphics[width=18cm]{sys_model_journal.eps}}
\caption{The system model which includes semantic model trading and semantic information trading. In the semantic model trading, edge devices trade with semantic model providers to obtain semantic encoders/decoders for semantic communications. In the semantic information trading, edge devices equipped with semantic encoders/decoders trade semantic information with buyers.}
\label{fig:sysmod}
\end{figure*}
\subsection{Incentive Mechanism Design}
In real-world settings, data transmissions are limited by the communication resources such as bandwidth and energy. Incentive mechanisms are designed to encourage certain parties to contribute to a communication network. For example, in a multi-node wireless powered communication network, selfish wireless nodes are not willing to charge other nodes by consuming their resources. To encourage the participation of these nodes, \cite{zheng2021age} proposed incentive schemes to deal with the selfishness of wireless nodes with an Age of Information (AoI) based utility functions. In collaborative edge learning, incentive mechanisms are used to incentivise the data owners to provide the updated model parameters for global model aggregation \cite{lim2021decentralized}.
Given that most of the communication networks are using conventional communication systems, semantic-aware incentive mechanisms are needed to be designed to motivate the participation of all parties in the development of semantic communication systems. We propose auctions as incentive mechanisms in the semantic model trading and semantic information trading, and derive the value of the semantic model and semantic information with semantic-based valuation functions.
\section{System Model and Problem Formulation}
\label{sec:sysmodel}
\subsection{System Model}
We consider a semantic communication network (Fig. \ref{fig:sysmod}) that consists of a set $\mathcal{M} = \{1,2,\ldots,m,\ldots,M\}$ of $M$ edge devices. To perform semantic encoding and decoding, the edge devices have to obtain the trained semantic models from the semantic model providers. Model trading is a common practice in collaborative edge learning, and in particular, federated learning \cite{lim2020federated}, where the model providers (sellers) receive incentives for providing trained models to the participants (buyers). In the case of semantic communications, the models being traded are the semantic encoders and decoders used for semantic information encoding and decoding, respectively. Devices with limited computation and communication resources can obtain high-quality semantic models from model trading. Moreover, it is shown that mismatches in communication channels and background knowledge of the communication environment degrade the performance of a pretrained semantic communication model \cite{xie2021deep}. Therefore, trading with model providers that perform machine learning training based on the relevant background knowledge and communication environment helps to improve the semantic performance of the devices. For example, devices can trade with the model providers that collect training data from the same certain geographical area as the buyers \cite{yang2022semantic}.
To encourage the participation of model providers, incentive mechanisms should be designed to ensure that model providers are appropriately rewarded from the semantic model trading process.
Similar to incentive mechanisms designed for the model trading in FL, the devices have to compete to obtain the semantic models from the semantic model providers. Intuitively, the devices are willing to pay more if the model obtained can achieve high semantic performance.
Besides, there exists a set $\mathcal{N} = \{1,2,\ldots,n,\ldots,N\}$ of $N$ semantic information buyers that are interested to obtain semantic information from the devices. For example, this may be semantic information trading between UAVs in real time \cite{yun2021attention}, and collection of semantic information for image classification tasks for autonomous vehicles \cite{yang2021semantic}. In this case, incentive mechanism design is also needed to facilitate the trading of such semantic information.
In the following, we propose two auction mechanisms for semantic model trading and semantic information trading. In the semantic model trading, we adopt a DL-based auction mechanism to derive the \textit{semantic-aware} valuation of the semantic models. The semantic model trading could be a channel to supply the semantic model for the devices to extract semantic information. Then, the semantic information from devices with higher accuracy is more valuable to semantic information buyers. For semantic information trading, we study the double auction mechanism for information trading between multiple buyers and multiple semantic information sellers. We further investigate how the semantic model obtained from the semantic model trading affects the results in the semantic information trading.
\subsection{Auction for Semantic Model Trading}
\label{smt}
The valuation of the devices for the model provided by the service provider is given by:
\begin{equation}\label{modalval}
v_{m} = A_{p} - A_{m},
\end{equation}
where $A_{p}$ is the accuracy of the model from the model provider, and $A_{m}$ is the accuracy of the current model of device $m$ ($A_{m} = 0$ if the device does not own any model). The accuracy metric can be the text similarity score for semantic text transmission \cite{xie2021deep}, signal-to-distortion ratio (SDR) for semantic speech signal transmission \cite{weng2021semantic}, and answer accuracy in visual question answering (VQA) \cite{xie2021task}. In every round of the single-item auction, the model provider, i.e., the auctioneer collects bids $(\mathfrak{b}_1,\mathfrak{b}_2,\ldots,\mathfrak{b}_M)$ from all smarts devices, i.e., bidders, and then decides the winner, $m^*$, and corresponding payment price, $\theta_{m^*}$. The utility of the device is given by $u_m = v_m - \theta_{m^*}$, if the device is the winner and $u_m = 0$ otherwise.
Traditional single-item auctions such as the first-price auction and Second-Price Auction (SPA) can be used to determine the winner and price. For an auction to be optimal \cite{myerson1981optimal}, it should attain the properties of incentive compatibility and individual rationality. Individual rationality guarantees that the utility of the devices is non-negative by participating in the auction, i.e., $u_m \geq 0$. Incentive compatibility ensures that each device submits bids according to their true valuations, respectively, i.e., $\mathfrak{b}_m = v_m$, regardless of the actions of other devices, and the utility of each device is maximized by submitting the truthful bid. In the first-price auction, the highest bidder wins and pays the exact bid submitted, maximizing the revenue gain of the model provider but does not guarantee incentive compatibility. In SPA, the highest bidder wins but pays the price of the second highest bidder. SPA ensures incentive compatibility but does not maximize the revenue of the model provider.
\begin{algorithm}
\caption{DL-Based Auction (DLA) Algorithm}\label{alg:aucalg}
\label{alg:DLA}
\hspace*{\algorithmicindent} \textbf{Input:} Bids of devices $\boldsymbol{\mathfrak{b}}=(\mathfrak{b}_1,\ldots,\mathfrak{b}_m,\ldots, \mathfrak{b}_M)$\\
\hspace*{\algorithmicindent} \textbf{Output:} Winner and Payment Price
\begin{algorithmic}[1]
\STATE \textbf{Initialization:} $\mathbf{w} = [w_{qs}^m]\in \mathbb{R}_{+}^{M\times QS}, \bm{\beta} = [\beta_{qs}^m] \in \mathbb{R}^{M\times QS} $
\WHILE{Loss function $\hat{R}(\mathbf{w,\bm{\beta}})$ is not minimized}
\STATE Compute transformed bids $\overline{\mathfrak{b}}_m = \Phi_m(\mathfrak{b}_m) = \min_{q\in Q}\max_{s\in S}(w_{qs}^m\mathfrak{b}_m + \beta_{qs}^m)$
\STATE Compute the allocation probabilities $z_m(\overline{\boldsymbol{\mathfrak{b}}}) = softmax(\overline{\mathfrak{b}}_1, \overline{\mathfrak{b}}_2,\ldots,\overline{ \mathfrak{b}}_{M+1};\kappa)$
\STATE Compute the SPA-0 payments $\theta^0_m(\overline{\boldsymbol{\mathfrak{b}}}) = ReLU(\max_{j\neq m} \overline{\mathfrak{b}}_j)$
\STATE Compute the conditional payment $\theta_m = \Phi_m^{-1}(\theta^0_m(\overline{\boldsymbol{\mathfrak{b}}}))$
\STATE Compute the loss $\hat{R}(\mathbf{w} ,\bm{\beta})$
\STATE Update parameters $\mathbf{w}$ and $\bm{\beta}$ using SGD optimizer
\ENDWHILE
\RETURN Winner $m^*$ and payment price $\theta_{m^*}$
\end{algorithmic}
\end{algorithm}
We adopt a DL-based optimal auction mechanism \cite{dutting2019optimal} that can maximize the revenue of the seller while achieving the properties of incentive compatibility and individual rationality. The auctioneer (i.e., the model provider) does not have a priori knowledge about the bidders and optimal decisions in determining the winner. Nevertheless, the model provider can learn from experience and adjust the auction decision using DL-based optimal auction.
The DL-based auction consists of three major functions: monotone increasing function, $\Phi_m$, allocation rule, $z_m$, and conditional payment rule, $\theta_m$.
Firstly, the input bids, $\boldsymbol{\mathfrak{b}}=(\mathfrak{b}_1,\ldots,\mathfrak{b}_m,\ldots, \mathfrak{b}_M)$, are transformed by $Q$ groups of $S$ linear functions, followed by the $\min$ and $\max$ operations, i.e., the transformed bid,
\begin{equation}
\overline{\mathfrak{b}}_m= \Phi_m(\mathfrak{b}_m) = \min_{q\in Q}\max_{s\in S} (w_{qs}^m \mathfrak{b}_m + \beta_{qs}^m),
\end{equation}
where $w_{qs}^m\in \mathbb{R}_+$, $q=1,\dots,Q$, $s=1,\dots,S$ and $\beta_{qs}^m\in\mathbb{R}$, $q=1,\dots,Q$, $s=1,\dots,S$ are the weight and bias to be trained, respectively.
The linear functions are strictly monotonically increasing functions to ensure the properties of incentive compatibility and individual rationality of the auction:
\newtheorem{theorem}{Theorem}
\begin{theorem}
(\cite{dutting2019optimal}) For any set of strictly monotonically increasing function \{$\Phi_1,\dots,\Phi_M$\}, an auction defined by allocation rule $z_m=z_m^0 \circ \Phi_m$ and the payment rule $\theta_m = \Phi_m^{-1} \circ \theta_m^0 \circ \Phi_m$ has the properties of incentive compatibility and individual rationality, where $z^0$ and $\theta^0$ are the allocation and payment rule of a second price auction with zero reserve, respectively, and $\circ$ indicates function composition, i.e., $(f\circ g)(x) = f(g(x))$.
\end{theorem}
To ensure that the auction learnt by the network achieves incentive compatibility and individual rationality, we constrain the allocation and payment rules of the network by following Theorem 1.
After the monotone transformation, the transformed bids are passed to separate networks that approximate the allocation and payment rule.
The allocation rule which follows the second price auction with zero reserve (SPA-0) allocation rule is approximated by a softmax function \cite{bridle1989training} to maximize the allocation probability of the highest bid, i.e.,
\begin{equation}
z_m(\overline{\boldsymbol{\mathfrak{b}}})=\frac{e^{\kappa \overline{\mathfrak{b}}_m}}{\sum^{M+1}_{j=1} e^{\kappa \overline{\mathfrak{b}}_j}},
\end{equation}
where $\overline{\mathbf{\mathfrak{b}}}=(\overline{\mathfrak{b}}_1,\dots,\overline{\mathfrak{b}}_{M+1})$, $\overline{\mathfrak{b}}_{M+1}$ is an additional dummy input, and $\kappa$ determines the quality of the approximation. The higher the value of $\kappa$, the higher the accuracy of approximation but the allocation function is less smooth and harder to optimize.
The SPA-0 payment rule is given by:
\begin{equation}
\theta^0_m(\overline{\boldsymbol{\mathfrak{b}}}) = ReLU(\max_{j\neq m} \overline{\mathfrak{b}}_j),
\end{equation}
where $ReLU(x)=\max(x,0)$ is used to ensure that the payment is non-negative.
To obtain the payment price, the inverse transformation function is applied on the SPA-0 price of the transformed bids, i.e.,
\begin{equation}
\theta_m = \Phi_m^{-1}(\theta^0_m(\overline{\boldsymbol{\mathfrak{b}}})),
\end{equation}
where the inverse transformation function can be expressed by:
\begin{equation}
\Phi_m^{-1}(y)=\max_{q\in Q}\min_{s\in S}(w^m_{qs})^{-1}(y-\beta_{qs}^m).
\end{equation}
To maximize the revenue, the network optimizes a loss function that is the negative value of the seller revenue. The loss function is given by
\begin{equation}
\hat{R}(\mathbf{w} ,\bm{\beta})=-\sum^M_{m=1}z_m(\overline{\boldsymbol{\mathfrak{b}}})\theta_m.
\end{equation}
\subsection{Auction for Semantic Information Trading}
We consider $N$ semantic information buyers and $M$ devices, where the buyers are interested in buying semantic information from the devices. Consider that the devices obtain the semantic models from the semantic model trading, the semantic information buyers are willing to pay more for the semantic information from devices with high accuracy, $A_m$.
We propose a single-round double auction for the one-to-one mapping of the buyers and the sellers. In the double auction, there are
\sloppy
\begin{itemize}
\item A set of semantic information buyers $\mathcal{N} = \{1,\ldots,n,\ldots,N\}$
\item A set of semantic information sellers $\mathcal{M} = \{1,\ldots,m,\ldots,M\}$, the devices that provide semantic information to the buyers
\item A trusted third party, the auctioneer
\end{itemize}
Based on the semantic performance, each buyer has different preferences for the devices. Let $\mathbf{b_n} = (b_n^1,\ldots,b^M_n)$ denote the bid vector of buyer $n$, where $b_n^m$ is the bid of buyer $n$ for device $m$, i.e., the price that buyer $n$ is willing to pay for receiving semantic information from device $m$.
Let $\mathbf{a} = (a_1,\ldots,a_M)$ denote the ask vector of the devices, where $a_m$ is the ask of device $m$, i.e., the price that device $m$ is willing to receive for trading the semantic information. The value of the semantic information from device $m$ to buyer $n$ can be expressed as
\begin{equation}
v_n^m = A_n^m(A_m),
\end{equation}
where $A_n^m$ is the accuracy of the semantic information transmitted by device $m$ to buyer $n$, and $A_m$ (determined by the semantic model trading) is the upper bound of the achievable accuracy of current semantic model.
Let $p_n$ be the price that buyer $n$ pays, the utility of buyer $n$ is given by
\begin{equation}
\label{iotutil}
u_n^b =
\begin{cases}
v_n^m - p_n & \text{if buyer $n$ wins the auction,}\\
0 & \text{otherwise.}
\end{cases}
\end{equation}
Note that to compare the utility of buyer $n$ when it wins different devices, we also use $u_{n,m}^b$ and $u_{n,m'}^b$ to denote the utility of buyer $n$ when it wins to obtain semantic information of device $m$ and $m'$, respectively.
Following \cite{jiao2020toward}, the data collection cost is given by
\begin{equation}
c_m^{d} = d_m \gamma_m,
\label{dcost}
\end{equation}
where $d_m$ and $\gamma_m$ are the data size and unit data cost, respectively.
The computational cost can be formulated as
\begin{equation}
c_m^{cp} = d_m \Gamma_m,
\label{cpcost}
\end{equation}
where $\Gamma$ is the unit computational cost to extract semantic information from the collected data. This cost can be due to the energy consumption \cite{lim2021decentralized} or edge/cloud computation resource rental fee \cite{dinh2020online}. The communication cost for device $m$ to transmit the semantic information is
\begin{equation}
c_m^{cm} = P_m\frac{\mathbb{N}_m}{R}\nu_m,
\label{cmcost}
\end{equation}
where $P_m$ is the communication power, $\mathbb{N}_m$ is the number of bits used to represent the semantic information, $R$ is the transmission rate in bits per second, and $\nu_m$ is the unit energy cost for communication.
The cost of the semantic model is given by
\begin{equation}
c_m^{md} = \frac{\theta_m}{T_m},
\label{mdcost}
\end{equation}
where $\theta_m$ is the price paid for the current semantic model (determined by the model trading auction in Section \ref{smt}) and $T_m$ is the expected number of transmissions with the model.
The total cost for device $m$ to transmit the semantic information is then defined as follows:
\begin{equation}
\begin{aligned}
C_m &= c_m^{d} + c_m^{cp} + c_m^{cm} + c_m^{md} \\
& = d_m \gamma_m + d_m \Gamma_m + P_m\frac{\mathbb{N}_m}{R}\nu_m + \frac{\theta_m}{T_m}.
\label{cost}
\end{aligned}
\end{equation}
Let $y_m$ be the payment to device $m$, the utility of the device $m$ is given by
\begin{equation}
u_m^s =
\begin{cases}
y_m - C_m & \text{if device $m$ wins the auction,}\\
0 & \text{otherwise.}
\end{cases}
\end{equation}
The proposed double auction has two stages, the {\it candidate-determination and pricing} stage, and the {\it candidate-elimination} stage. The algorithms for the two stages are shown in Algorithms \ref{alg:cdp} and \ref{alg:ce} respectively. Note that the DLA refers to the DL-Based Auction in Algorithm \ref{alg:DLA}. In the {\it candidate-determination and pricing} stage, the auctioneer determines the buyer candidates of each device, the prices that the buyer candidates pay, and the payment to be rewarded to the devices.
Let $n$ and $\theta_n$ denote the winning buyer and payment price determined by DLA, respectively. For each device $m$, all bids are sent to DLA to determine the winner and payment price. If the payment price is not lower than the ask $a_m$, i.e., $\theta_n \ge a_m$, then the buyer $n$ is added to the set of buyer candidates $\mathcal{N}_c$ with price $p_n = \theta_n$, and device $m$ is added to the set of seller candidates $\mathcal{M}_c$ with payment $y_m = \theta_n$.
\begin{algorithm}
\caption{Candidate Determination and Pricing}\label{alg:cdp}
\hspace*{\algorithmicindent} \textbf{Input:} $\mathcal{N},\mathcal{M},\mathbf{b},\mathbf{a}$ \\
\hspace*{\algorithmicindent} \textbf{Output:} $\mathcal{N}_c,\mathcal{M}_c,\mathcal{P}_c,\mathcal{Y}_c$, $\hat{\sigma}$
\begin{algorithmic}[1]
\FOR{$ m \in \mathcal{M}$}
\STATE $n, \theta_n= DLA(\{b^m_n,\forall n \in \mathcal{N}\})$
\IF{$\theta_n \geq a_m$}
\STATE $\hat{\sigma}(m) = n$, buyer $n$ is added into $\mathcal{N}_c$, and seller $m$ is added into $\mathcal{M}_c$
\STATE $p_n = y_m = \theta_n$
\STATE $p_n$ and $y_m$ are added into $\mathcal{P}_c$ and $\mathcal{Y}_c$, respectively
\ENDIF
\ENDFOR
\end{algorithmic}
\end{algorithm}
After the first stage, each buyer candidate may win more than one device. In the {\it candidate-elimination} stage, for each buyer candidate, the algorithm selects the best device such that the buyer yields the highest utility in Equation \eqref{iotutil}. If more than one device yields the same highest utility for the buyer, the best device is randomly selected.
In the following, we prove that the double auction mechanism in our model satisfies the properties of individual rationality, incentive compatibility, and budget balanced.
\begin{algorithm}
\caption{Candidate-elimination}\label{alg:ce}
\hspace*{\algorithmicindent} \textbf{Input:} $\mathcal{N}_c, \mathcal{M}_c ,\mathcal{P}_c, \mathcal{Y}_c, \hat{\sigma}, \mathbf{b}$ \\
\hspace*{\algorithmicindent} \textbf{Output:} $\mathcal{N}_w,\mathcal{M}_w,\mathcal{P}_w,\mathcal{Y}_w$
\begin{algorithmic}[1]
\STATE $\mathcal{N}_w \leftarrow \mathcal{N}_c,\mathcal{M}_w \leftarrow \mathcal{M}_c, \mathcal{P}_w \leftarrow \mathcal{P}_c,\mathcal{Y}_w \leftarrow \mathcal{Y}_c, \sigma \leftarrow \hat{\sigma} $
\FOR{any two seller $m, m' \in \mathcal{M}_w, m \ne m'$}
\IF{$\sigma(m) = \sigma(m')$}
\IF{$u_{n,m}^b = u_{n,m'}^b$}
\STATE $m^*\leftarrow$ randomly selected from $\{m,m'\}$
\ELSE
\STATE $m^*\leftarrow$ arg $\min_{k\in\{m,m'\}}u_{n,k}^b$
\ENDIF
\ENDIF
\STATE $\mathcal{M}_w \leftarrow \mathcal{M}_w \setminus \{m^*\}$
\STATE $\mathcal{P}_w \leftarrow \mathcal{P}_w \setminus \{p_{m^*}\}$
\STATE $\mathcal{Y}_w \leftarrow \mathcal{Y}_w \setminus \{y_{m^*}\}$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\newtheorem{lemma}{Lemma}
\begin{theorem}
The proposed double auction mechanism is individually rational. All winning buyers and sellers are rewarded with non-negative utilities i.e. $p_n \le b_n^m$ and $y_m \ge a_m$
\end{theorem}
\begin{proof}
From Algorithm \ref{alg:cdp}, since DLA has the property of individual rationality \cite{dutting2019optimal}, we have $\theta_n \leq b_n^m$. Therefore $p_n \le b_n^m$ and $y_m \ge a_m$, individual rationality is satisfied in the candidate determination and pricing stage. Since Algorithm \ref{alg:ce} does not change the value of $p_n$ and $y_m$, the individual rationality is preserved after the candidate eliminations.
\end{proof}
\begin{theorem}
The proposed double auction mechanism is incentive compatible. All buyers and sellers submit their bids and asks truthfully as they cannot improve their utilities by submitting bids and asks that are different from their true valuations.
\end{theorem}
\begin{proof}
We prove the incentive compatibility by the following lemmas:
\begin{enumerate}
\item The proposed double auction mechansim is truthful for the sellers (as shown in Lemma \ref{lemma:selleric}).
\item The proposed double auction mechanism is truthful for the buyers (as shown in Lemma \ref{lemma:buyeric}).
\end{enumerate}
\end{proof}
\begin{lemma}
The proposed double auction mechanism is truthful for the sellers.
\label{lemma:selleric}
\end{lemma}
\begin{proof} To prove that the proposed double auction mechanism is truthful for the sellers, we discuss the three possible outcomes for the sellers in the following subsets:
\begin{enumerate}
\item Subset $\mathcal{M}_w$, sellers that win the auction,
\item Subset $\mathcal{M}_c \setminus \mathcal{M}_w$, sellers that are selected as candidates but are eliminated during the candidate elimination stage, and
\item Subset $\mathcal{M} \setminus \mathcal{M}_c$, sellers that are not selected as candidates.
\end{enumerate}
In each of the subsets, we discuss the cases where the sellers bid untruthfully. In each case, we show that the sellers cannot achieve higher utilities with the untruthful bids. Note that tilde $\widetilde{\cdot}$ is shown for the notations to indicate the outcomes of the untruthful cases.
\begin{enumerate}[leftmargin=*]
\item For seller $m \in \mathcal{M}_w$:
{\it Case 1.} Seller $m$ does not win the auction with untruthful ask, $\widetilde{u}_m^s = 0 \le u_m^s$.
{\it Case 2.} Seller $m$ wins the auction with untruthful ask. In this case, the payment does not change because the input bids to DLA are not changed, i.e., $\widetilde{u}_m^s = u_m^s$.
\item For seller $m \in \mathcal{M}_c \setminus \mathcal{M}_w$, changing ask does not change the price as discussed in the case of seller $m \in \mathcal{M}_w$. Therefore seller $m \in \mathcal{M}_c \setminus \mathcal{M}_w$ does not win the auction regardless of the value of $a_m$, $\widetilde{u}_m^s = u_m^s = 0$.
\item For seller $m \in \mathcal{M} \setminus \mathcal{M}_c$:
{\it Case 1.} Seller $m$ does not win by asking untruthfully, i.e., $m \notin \widetilde{\mathcal{M}}_w$, therefore the utility remains unchanged, $\widetilde{u}_m^s = u_m^s = 0$.
{\it Case 2.} Seller $m$ wins by asking untruthfully, i.e., $m \in \widetilde{\mathcal{M}}_w$. Let buyer $n$ be the winner of semantic information from $m$ with price $\widetilde{p}_n = \widetilde{y}_m$.
To win the auction, $m$ has to ask lower than the true valuation such that $\widetilde{a}_m < C_m$. As the payment is not affected by $\widetilde{a}_m$, we have $\widetilde{y}_m = y_m$ and since $m$ does not win by asking truthfully, $y_m < C_m$, therefore $m$ suffers negative utility in this case, i.e., $\widetilde{u}_m^s = \widetilde{y}_m - C_m < 0 = u_m^s$.
Therefore we can conclude that the sellers cannot obtain a higher utility by asking untruthfully.
\end{enumerate}
\end{proof}
\begin{lemma}
The proposed double auction mechanism is truthful for the buyers.
\label{lemma:buyeric}
\end{lemma}
\begin{proof}
To prove that the proposed double auction mechanism is truthful for the buyers, we discuss the two possible outcomes for the buyers in the following subsets:
\begin{enumerate}
\item Subset $\mathcal{N}_w$, buyers that win the auction, and
\item Subset $\mathcal{N} \setminus \mathcal{N}_w$, buyers that lose the auction.
\end{enumerate}
In each of the subsets, we discuss the cases where the buyers ask untruthfully. In each case, we show that the buyers cannot achieve higher utilities with the untruthful asks. Note that tilde $\widetilde{\cdot}$ is shown for the notations to indicate the outcomes of the untruthful cases.
\begin{enumerate}[leftmargin=*]
\item For buyer $n \in \mathcal{N}_w$, assuming $n$ wins seller $m$ by bidding truthfully. Let us consider the following cases when buyer $n$ bids untruthfully:
{\it Case 1.} Buyer $n$ loses with untruthful bid, $\widetilde{u}_n^b = 0 \le u_n^b$.
{\it Case 2.} Buyer $n$ wins the same seller $m$ with untruthful bid, given individual rationality property of DLA, we have $\widetilde{u}_n^b \le u_n^b$.
{\it Case 3.} Buyer $n$ wins with a different seller $m'$ with untruthful bid.
Let us consider the following cases when buyer $n$ bids truthfully:
\begin{itemize}[leftmargin=*]
\item Seller $m' \in \mathcal{M}_c$ and $\hat{\sigma}(m')=n$.
Since buyer $n$ wins $m$ in the truthful case, we have $u_{n,m'}^b \le u_{n,m}^b$. Given that DLA has the property of individual rationality, we have $\widetilde{u}_{n,m'}^b \le u_{n,m'}^b$.
Thus we know that $\widetilde{u}_{n,m'}^b \le u_{n,m}^b$.
\item Seller $m' \in \mathcal{M}_c$ and $\hat{\sigma}(m') \ne n$. It means that there is another buyer candidate $n'$ with higher or equal bid for $m'$, i.e., $b_{n'}^{m'} \ge b_n^{m'}$. When buyer $n$ wins $m'$ by bidding untruthfully, since DLA satisfies the individual rationality constraint, we have $\widetilde{u}_{n,m'}^b \le 0$. From Theorem 2, we know that $u_{n,m}^b \ge 0$ (all winning buyers and sellers are rewarded with non-negative utility), thus we have $\widetilde{u}_{n,m'}^b \le u_{n,m}^b$.
\item Seller $m' \notin \mathcal{M}_c$ and buyer $n$ wins $m'$ by bidding untruthfully. Since DLA has the property of individual rationality, we have $\widetilde{u}_{n,m'}^b \le 0$. From Theorem 2, we know that $u_{n,m}^b \ge 0$, thus we have $\widetilde{u}_{n,m'}^b \le u_{n,m}^b$.
\end{itemize}
\item For buyer $n \in \mathcal{N}\setminus\mathcal{N}_w$ with utility $u_n^b = 0$. We consider the following cases when buyer $n$ bids untruthfully.
{\it Case 1.} Buyer $n$ loses with untruthful bid, $\widetilde{u}_n^b = 0 = u_n^b$.
{\it Case 2.} Buyer $n$ wins seller $m$ by bidding untruthfully. Since DLA has the property of individual rationality, we have $\widetilde{u}_n^b \le 0 = u_n^b$.
\end{enumerate}
Therefore we can conclude that the buyers cannot obtain a higher utility by bidding untruthfully.
\end{proof}
\begin{theorem}
The proposed double auction mechanism is budget balanced. The total price paid by the winning buyers is not less than the total payment to the winning sellers, i.e., $\sum_{n \in \mathcal{N}_w} p_n \ge \sum_{m \in \mathcal{M}_w} y_m$.
\end{theorem}
\begin{proof}
According to Algorithms \ref{alg:cdp} and \ref{alg:ce}, the price that winning buyers pay and the payment received by winning sellers are equal for every winning seller-buyer pairs. Thus, we have
\begin{equation}
\sum_{n \in \mathcal{N}_w} p_n - \sum_{m \in \mathcal{M}_w} y_m = 0.
\end{equation}
We can conclude that the double auction mechanism is budget balanced.
\end{proof}
In Algorithm \ref{alg:cdp}, since there are $M$ sellers in set $\mathcal{M}$, the time complexity of the candidate determination and pricing stage is $O(M)$. In Algorithm \ref{alg:ce}, we know that $|\mathcal{M}_c| \le |\mathcal{M}|=M$. In the worst case, the for-loop runs for $\frac{M(M-1)}{2}$ times. Therefore, Algorithm \ref{alg:ce} has the time complexity of $O(\frac{M(M-1)}{2})=O(M^2)$. Overall, the proposed double auction mechanism is a polynomial time algorithm with the time complexity of $O(M^2)$.
\section{Case Study: Semantic Text Transmission}\label{sec:casestudy}
In this section, we apply the proposed auction mechanisms to the semantic text transmission. We derive the valuations of the semantic model trading and semantic information trading for semantic text transmission.
\subsection{Deep Learning Enabled Semantic Communication Systems}
\label{dlsemcom}
We consider the $M$ devices perform text data transmission with DL enabled semantic communication systems, e.g., voice controlled devices (Google Nest Hub \footnote{https://www.cnet.com/home/smart-home/how-to-set-up-your-new-google-nest-hub-or-nest-hub-max/}, Amazon Echo \footnote{https://www.androidauthority.com/amazon-echo-5th-gen-3095027}, and Apple HomePod \footnote{https://www.apple.com/sg/newsroom/2021/10/apple-introduces-homepod-mini-in-new-bold-and-expressive-colors/}). In DL enabled semantic communication system, collected sentences, $\mathbf{S}=[s_1,s_2,\ldots,s_{N_s}]$, are encoded by semantic encoder and channel encoder. The encoded signal can be represented by
\begin{equation}
\mathbf{X} = enc_c(enc_s(\mathbf{S})) \label{enc},
\end{equation}
where $\mathbf{X} \in \mathbb{R}^{N_s \times L \times D}$, $N_s$ is the number of sentences, $L$ is the sentence length, $D$ is the output dimension of channel encoder, $enc_c(\cdot)$ is the channel encoder, and $enc_s(\cdot)$ is the semantic encoder. Note that all inputs are padded to length $L$ before passing to the encoders.
After winner determination of the double auction, winning devices transmit encoded information to the winning buyers. At the buyer, signal received can be expressed as
\begin{equation}
\mathbf{Y} = \mathbf{HX}+\mathbf{A},
\end{equation}
where $\mathbf{H}$ is the channel gain between the transmitter and receiver and $\mathbf{A} \sim \mathcal{N}(0,\sigma^2_n)$ is the additive white Gaussian noise (AWGN). The decoded sentences are given by
\begin{equation}
\widehat{\mathbf{S}} = dec_s(dec_c(\mathbf{Y})),
\end{equation}
where $dec_s(\cdot)$ and $dec_c(\cdot)$ are the semantic decoder and channel decoder of the receiver.
We adopt the network architecture of DeepSC \cite{xie2021deep} where the semantic encoder and decoder are implemented as multiple Transformer \cite{vaswani2017attention} encode and decode layers, and channel encoder as dense layers with different units. Our incentive mechanism can be easily extended to other network architectures by following the same evaluation procedure.
The BLEU score and the sentence similarity are two of the critical performance metrics of the text-based semantic communication system. The BLEU score measures an exact matching of words in the original and recovered sentences without considering their semantic information. In contrast to the BLEU score, the sentence similarity is calculated by the cosine similarity of the extracted semantic features from original and recovered sentences. In our model, a pre-trained Bidirectional Encoder Representations from Transformers (BERT) \cite{devlin-etal-2019-bert} model is used for the semantic features extraction. Let $\mathbf{s}$ and $\widehat{\mathbf{s}}$ denote one sentence from $\mathbf{S}$ and $\widehat{\mathbf{S}}$, respectively. The BLEU score can be expessed as
\begin{equation}
\log \text{BLEU} = \min \left(1-\frac{l_{\widehat{\mathbf{s}}}}{l_{\mathbf{s}}},0\right) + \displaystyle\sum_{i=1}^{I} u_i \log p_i,
\end{equation}
where $l_{\mathbf{s}}$ and $l_{\widehat{\mathbf{s}}}$ are the lengths of the original and recovered sentences respectively, $u_i$ is the weight of $i$-grams, and $p_i$ is the $i$-grams score, which is given by
\begin{equation}
p_i = \frac{\displaystyle\sum_{k=1}^{K_i} \min(C_k (\widehat{\mathbf{s}}), C_k(\mathbf{s}))}{\displaystyle\sum_{k=1}^{K_i} \min(C_k(\widehat{\mathbf{s}}))},
\end{equation}
where $K_i$ is the number of elements in $i$-th grams, and $C_k (\cdot)$ is the frequency count function for the $k$-th element in $i$-th grams.
The sentence similarity is given by
\begin{equation}
similarity(\widehat{\mathbf{s}},\mathbf{s}) = \frac{\mathbf{B}(\mathbf{s}) \cdot \mathbf{B}(\widehat{\mathbf{s}})^T}{\lVert\mathbf{B}(\mathbf{s})\rVert \lVert\mathbf{B}(\widehat{\mathbf{s}})\rVert},
\end{equation}
where $\mathbf{B}(\cdot)$ is a pre-trained BERT model used to measure the sentence similarity.
In general, to obtain a higher BLEU score and similarity score, we need to increase the output dimension $D$ of the encoder \cite{devlin-etal-2019-bert}.
However, increasing $D$ comes at the cost of a larger data size, and the amount of data that devices can send is limited by the communication resources, e.g., energy supply to the devices \cite{liew2022economics}. Specifically, the BLEU score and the similarity score of device $m$ can be expressed as
\begin{equation}
\label{simscore}
s_m = f_{sim}(D) = f_{sim}(\frac{\mathbb{N}_m}{N_s\times L \times b_f}),
\end{equation}
and
\begin{equation}
\label{bleuscore}
BLEU_m = f_{BLEU}(D) = f_{BLEU}(\frac{\mathbb{N}_m}{N_s\times L \times b_f}),
\end{equation}
respectively, where $f_{sim}(\cdot)$ and $f_{BLEU}(\cdot)$ are simple lookup to obtain the scores of the model, $b_f$ is the number of bits used by a unit feature, and $\mathbb{N}_m$ is the total number of bits that the device $m$ can transmit. The values of $f_{sim}(\cdot)$ and $f_{BLEU}(\cdot)$ can be obtained by using different output dimension $D$ to evaluate the similarity score and the BLEU score, respectively. A unit feature is a single entry of $\mathbf{X} \in \mathbb{R}^{N_s \times L \times D}$, and $b_f$ is the number of bits used to represent a float type data. In our model, the data size in Equations \eqref{dcost} and \eqref{cpcost} is given by the number of words collected, i.e., $d=N_s\times L$. The total number of bits affects the communication cost as shown in Equation \eqref{cmcost}.
From Equations \eqref{simscore} and \eqref{bleuscore}, it is clear that the scores are affected by the size of data and model performance. Since each device has a different model performance and data to be sent, the similarity score and the BLEU score are different among the devices.
\subsection{Semantic-Aware Valuation for Auctions}
In the semantic model trading, the devices bid according to the performance of the semantic model (Equation \eqref{modalval}), i.e.,
\begin{equation}
\mathfrak{b}_m = v_m = A_{p} - A_m.
\end{equation}
The accuracy of the model from the model provider can be expressed as follows:
\begin{equation}
A_{p} = \lambda_m s_{p} + \beta_m BLEU_{p},
\end{equation}
where $s_{p}$ and $BLEU_{p}$ are the similarity score and the BLEU score achievable by the model provided, respectively, $\lambda_m$ is the preference for the similarity score by the device $m$, $\beta_m$ is the preference for the BLEU score by the device $m$, and $\lambda_m + \beta_m=1$. If $\beta_m > \lambda_m$, it indicates that the device has more interest in the exact recovery of words whereas $\lambda_m > \beta_m$ indicates higher interest in the matching of the semantic meaning. For example, some medical devices \cite{dhyani2021intelligent} would have higher $\beta_m$ because the exact recovery of medical terms is more important, whereas devices that collect data for text classification \cite{shah2016review} would have higher $\lambda_m$.
The accuracy of the current model of device $m$ is given by:
\begin{equation}
A_{m} = \lambda_m s_{m} + \beta_m BLEU_{m},
\end{equation}
where $s_{m}$ and $BLEU_{m}$ are the similarity score and the BLEU score achievable by the current model.
In the semantic information trading, based on the communication environment and resources, each device can achieve different semantic performance when transmitting information to the buyers. Therefore, based on the semantic performance, each buyer has different preferences for the devices.
The value of the semantic information from device $m$ to buyer $n$ is given by:
\begin{equation}
v_n^m = \lambda_n s_m + \beta_n BLEU_m ,
\end{equation}
where $\lambda_n$ is the preference for the similarity score, and $\beta_n$ is the preference for the BLEU score by the buyer $n$. As the auction is truthful for all buyers and sellers, the buyers and sellers submit bids and asks according to their true valuations, i.e., $b_n^m = v_n^m$ and $a_m = C_m$. Again, the cost of collecting the information by device $m$ can be obtained from Equation \eqref{cost}.
\subsection{Feature Reduction Technique}
\label{featreduce}
Let $\mathbb{N}_m$ denote the number of bits that device $m$ can send to the buyer.
Based on the bit budget $\mathbb{N}_m$, not all features of the encoded information can be sent.
However, the semantic communication model is trained with a fixed number of features with output dimension $D$. A sample of feature representation output by semantic encoder with 16 features is shown in Fig. \ref{fig:textfeat}.
Sentences decoded from partial features have a lower similarity score and BLEU score than that decoded from all features.
Deep neural networks need to fine-tune the model parameters to reduce the gap in performance.
Unfortunately, devices that operate on limited resource might not be able to fine-tune the model in real-time because it is both time and energy consuming.
Therefore an effective feature reduction method is required for these devices to minimize the gap in performance when they have to communicate with a limited bit budget.
\begin{figure*}[htb]
\centering
\centerline{\includegraphics[width=18cm]{text_features.eps}}
\caption{Sample output of semantic encoder with 16 features, input sentence: ``thirdly it criticises the shortcomings but in a positive manner".}
\label{fig:textfeat}
\end{figure*}
\begin{figure*}[htb]
\centering
\centerline{\includegraphics[width=16cm]{controlled_dropout_training_new.eps}}
\caption{Illustration of controlled dropout during the training.}
\label{fig:cdroptrain}
\end{figure*}
\begin{figure}[t]
\begin{minipage}[b]{.48\linewidth}
\centering
\centering
\centerline{\includegraphics[width=3.5cm]{conventional_dropout_new.eps}}
\centerline{(a)}\medskip
\label{fig:convendrop}
\end{minipage}
\hfill
\begin{minipage}[b]{0.48\linewidth}
\centering
\centering
\centerline{\includegraphics[width=3.5cm]{controlled_dropout_new.eps}}
\centerline{(b)}\medskip
\label{fig:rev}
\end{minipage}
\caption{An illustration of (a) conventional dropout and (b) controlled dropout.}
\label{fig:dropout}
\end{figure}
We propose a simple feature reduction method where the performance can be adjusted by a regularization technique \cite{ko2017controlled} during the training of the model. Consider that the model on device is pre-trained with output dimension $D$, under the limited bit budget, the encoded signal, $\mathbf{X} \in \mathbb{R}^{N_s \times L \times D}$ is reduced to $\mathbf{X}' \in \mathbb{R}^{N_s \times L \times D'}$, where $0<D'<D$. At the receiver, the received signal $\mathbf{Y}'\in \mathbb{R}^{N_s \times L \times D'}$ is padded with zeros to become $\mathbf{Y}\in \mathbb{R}^{N_s \times L \times D}$. The proposed data reduction method is illustrated in Fig. \ref{fig:cdroptrain}. To obtain $f_{sim}(\cdot)$ and $f_{BLEU}(\cdot)$, we first train the DeepSC model with the data with dimension $D$ and use the trained model to evaluate the similarity scores for output dimension $d$, $\forall d \in [1,D]$. Then, we can obtain $f_{sim}(\cdot)$ and $f_{BLEU}(\cdot)$ from the evaluation results of test datasets.
To reduce the degradation of performance, we add a controlled dropout \cite{ko2017controlled} layer before the channel decoding layer of the receiver. For example, if index $d_i$ is selected by controlled dropout, all units from (0, 0, $d_i$) to ($N_s-1$, $L-1$, $d_i$) become zeros. The conventional dropout \cite{srivastava2014dropout} technique randomly drops units (Fig. \ref{fig:dropout}(a)) in the training process to solve the overfitting issue of the deep neural networks. In contrast to conventional dropout, controlled dropout drops units intentionally, i.e., dropping a selected index of a dimension, as shown in Fig. \ref{fig:dropout}(b). An illustration of the effect of controlled dropout during training is shown in Fig. \ref{fig:cdroptrain}. In our experiments, we drop units from a certain index of the output dimension. As shown in \cite{ko2017controlled}, we can obtain a better performance than conventional dropout when the index is randomly selected. Following \cite{ko2017controlled}, the index is randomly selected with a dropout rate, $p_{drop}$, $0 < p_{drop} < 1$. Controlled dropout helps the model to generalize to the reduced features during the training.
\section{Numerical Results}\label{sec:results}
In this section, we evaluate the performance of the proposed auction mechanisms and feature reduction method. The values of experiment parameters are presented in Table \ref{tab:param}. The similarity and BLEU scores are sampled according to the simulation settings in \cite{lim2021decentralized}, \cite{luong2018optimal} for the DL-based auction. The dropout rate is set according to \cite{ko2017controlled}. Following \cite{jiao2020toward}, \cite{chen2022performance}, we set the cost-related parameters in the double auction as shown in Table \ref{tab:param}.
\begin{table}
\begin{center}
\caption{Experiment Parameters \cite{lim2021decentralized}, \cite{jiao2020toward}, \cite{ko2017controlled}, \cite{luong2018optimal}, \cite{chen2022performance}}
\begin{tabular}{ |l|l|c|c| }
\hline
Parameters & Values \\
\hline
Similarity score coefficient, $\lambda_n$ & $\sim \mathcal{U}[0,1] $\\
BLEU score coefficient, $\beta_n$ & $1-\lambda_n$ \\
Dropout rate, $p_{drop}$ & 0.1 \\
Reduced output dimension, $D'$ & $\sim \mathcal{U}[1,16] $ \\
Data size, $d_m$ & $\sim \mathcal{U}[10,100]$ \\
Unit data cost, $\gamma_m$ & $0.001$ \\
Unit computational cost, $\Gamma_m$ & $0.001$ \\
Communication power, $P_m$ & $1$ \\
Number of bits transmitted, $\mathbf{N}_m$ & $10000$ \\
Transmission rate, $R$ & $100000$ \\
Unit energy cost, $\nu_m$ & $0.01$ \\
Expected number of transmissions, $T_m$ & $100$ \\
\hline
\end{tabular}
\label{tab:param}
\end{center}
\end{table}
\subsection{Evaluation of DeepSC with Feature Reduction}
We first investigate the improvement of semantic performance under the proposed feature reduction method. With the help of the DeepSC, we set the output dimension of encoder $D$ to 16, and train the model under AWGN channel for 200 epochs. The training and test data is obtained from the proceedings of the European Parliament \cite{koehn2005europarl}. We use $7347$ English sentences in the dataset for our evaluation, and use the rest of the English sentences for training. The performance scores are considered for the evaluation of the proposed double auction mechanism.
As described in Section \ref{featreduce}, we add a controlled dropout layer between the physical layer and receiver. The dropout probability is set as $0.1$ which means $10\%$ of the features are dropped randomly in a controlled setting. We record the similarity score and the BLEU score for the output dimensions from 1 to 16, which is shown in Fig. \ref{fig:odvsscore}. Regardless of the application of controlled dropout, we observe that the performance degrades as the output dimension decreases. The reason is that fewer features are transmitted.
However, when the output dimension changes from $1$ to $15$, the performance of model trained with controlled dropout outperforms constantly the baseline mode. In other words, as the output dimension decreases, the baseline model has a larger performance gap compared to the model with controlled dropout. Specifically, the reduction of the similarity score per output dimension is 0.05 in the baseline model and 0.04 in the proposed model. For the reduction of the BLEU score per output dimension, it is 0.06 for the baseline model and 0.05 for the proposed model. This result shows that the proposed model can maintain a similarity score of 0.80 even after 25\% of feature reduction ($D=12$) whereas the baseline model can only achieve the similarity score of 0.60 with the same output dimension. As shown in Table \ref{tab:sent}, the recovered sentence has higher similarity when the controlled dropout is applied.
However, we notice that the best performance achieved by the baseline model at $D=16$ is slightly higher than that of the proposed model. The BLEU score and the similarity score for our proposed model are 0.89 and 0.91, respectively, but both scores are 0.94 for the baseline model. The reason is that the accuracy is slightly dropped due to the generalization of the feature reduction. Overall, the gap in performance at fewer output dimensions is compensated by the controlled dropout during training.
\begin{figure}[t]
\centering
\begin{tikzpicture}
\begin{axis}[
xlabel={Output Dimension},
xmin=1, xmax=16,
ymin=0, ymax=1.4,
xtick={1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16},
ytick={0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0,1.1,1.2,1.3,1.4},
legend pos=north east,
legend style={fill=white, nodes={scale=0.8, transform shape}},
ymajorgrids=true,
xmajorgrids=true,
grid style=dashed,
legend cell align={left},
every axis plot/.append style={thick},
/pgf/number format/.cd,
1000 sep={},
]
\pgfplotstableread{simscore_snr9.txt}\simscore;
\addplot[
color=red,
]
table
[
x expr=\thisrowno{0},
y expr=\thisrowno{1}
] {\simscore};
\addlegendentry{Similarity Score without Controlled Dropout};
\addplot[
color=red,
dashed,
]
table
[
x expr=\thisrowno{0},
y expr=\thisrowno{2}
] {\simscore};
\addlegendentry{Similarity Score with Controlled Dropout};
\addplot[
color=blue,
mark=*,
]
table
[
x expr=\thisrowno{0},
y expr=\thisrowno{3}
] {\simscore};
\addlegendentry{BLEU Score without Controlled Dropout};
\addplot[
color=blue,
mark=*,
dashed,
]
table
[
x expr=\thisrowno{0},
y expr=\thisrowno{4}
] {\simscore};
\addlegendentry{BLEU Score with Controlled Dropout};
\end{axis}
\end{tikzpicture}
\caption{BLEU and similarity score under different output dimensions.}
\label{fig:odvsscore}
\end{figure}
\begin{table*}
\begin{center}
\caption{Sample Sentences with and without Controlled Dropout}
\begin{tabular}{ |l|l|c|c| }
\hline
\textbf{Original Sentence} & thirdly it criticises the shortcomings but in a positive manner \\
\hline
\textbf{Output Dimension = 15, with controlled dropout} & thirdly it have the shortcomings but in a positive manner \\
\hline
\textbf{Output Dimension = 14, with controlled dropout} & thirdly it have the shortcomings but in a positive manner \\
\hline
\textbf{Output Dimension = 15, without controlled dropout} & thirdly it forward the shortcomings but in a positive manner\\
\hline
\textbf{Output Dimension = 14, without controlled dropout} & thirdly it played the shortcomings but in a off manner \\
\hline
\end{tabular}
\label{tab:sent}
\end{center}
\end{table*}
\subsection{Evaluation of Deep Learning based Auction Mechanism}
Without loss of generality, we consider that the devices do not own any semantic model initially, i.e., $A_{m} = 0$ and $\mathfrak{b}_m = v_m = A_{p}$. To obtain the bid profiles, we consider $A_{p} \sim U[0,0.4]$ and $\sim U[0.5,0.9]$. We collect 1000 training samples with 10 bidders (devices) in each of the samples and perform training for 500 epochs. From Fig. \ref{fig:revmodprovider}, we observe that the DL-based auction can always achieve higher revenue than that of the SPA, regardless of the values of $A_{p}$. The reason is that the DL-based auction mechanism can adapt to different bid profiles by optimizing the parameters in the DL network. Moreover, we observe that, while SPA is incentive compatible, it does not maximize the revenue of the model providers. In contrast to SPA, the DL-based auction maximizes the revenue of model providers while keeping the desired properties of incentive compatibility and individual rationality, which helps to attract more model providers to offer quality semantic encoder/decoder for semantic communications.
\begin{figure}
\centering
\begin{tikzpicture}
\pgfplotsset{
width=7cm, compat=1.3
}
\begin{axis}[
xlabel={Iterations},
ylabel={Revenue of Semantic Model Provider},
xmin=1, xmax=500,
ymin=0.2, ymax=1.4,
xtick={1,100,200,300,400,500},
legend pos=north east,
legend style={fill=none, nodes={scale=0.8, transform shape}},
ymajorgrids=true,
xmajorgrids=true,
grid style=dashed,
legend cell align={left},
every axis plot/.append style={very thick},
/pgf/number format/.cd,
1000 sep={},
]
\pgfplotstableread{auction_res_uni_d50.txt}\resauction;
\addplot[
color=red,
dash pattern=on 6pt off 4pt on 2pt off 4pt,
]
table
[
x expr=\thisrowno{0},
y expr=\thisrowno{1}
] {\resauction};
\addlegendentry{DL-based Auction, $A_{p}\sim U[0,0.4]$};
\addplot[
color=teal,
]
table
[
x expr=\thisrowno{0},
y expr=\thisrowno{2}
] {\resauction};
\addlegendentry{SPA, $A_{p}\sim U[0,0.4]$};
\addplot[
color=blue,
dashed,
dash pattern= on 8pt off 4pt,
]
table
[
x expr=\thisrowno{0},
y expr=\thisrowno{3}
] {\resauction};
\addlegendentry{DL-based Auction, $A_{p}\sim U[0.5,0.9]$};
\addplot[
color=orange,
dashed,
]
table
[
x expr=\thisrowno{0},
y expr=\thisrowno{4}
] {\resauction};
\addlegendentry{SPA, $A_{p}\sim U[0.5,0.9]$};
\end{axis}
\end{tikzpicture}
\caption{Revenue of semantic model providers.}
\label{fig:revmodprovider}
\end{figure}
\subsection{Evaluation of Double Auction Mechanism}
To evaluate the performance of the double auction mechanism, we generate 1000 samples and average the simulation results.
We set the number of sellers to $M=20$ and evaluate the performance under different number of buyers.
Note that in the following discussion, we refer semantic information buyers as buyers and devices as sellers for simplicity.
\begin{figure}
\centering
\begin{tikzpicture}
\pgfplotsset{
width=7cm, compat=1.3
}
\begin{axis} [
ybar=0pt,
bar width=4pt,
xmin=0,xmax=8,
ymin=0,ymax=1.2,
legend style={fill=white, nodes={scale=0.8, transform shape}},
ymajorgrids=true]
\pgfplotstableread{n100_res.txt}\truthful;
\addplot[
fill=blue,
]
table
[
x expr=\thisrowno{0},
y expr=\thisrowno{1}
] {\truthful};
\addlegendentry{Ask};
\addplot[
fill=teal,
]
table
[
x expr=\thisrowno{0},
y expr=\thisrowno{2}
] {\truthful};
\addlegendentry{Bid};
\addplot[
fill=violet,
]
table
[
x expr=\thisrowno{0},
y expr=\thisrowno{3}
] {\truthful};
\addlegendentry{Price};
\end{axis}
\end{tikzpicture}
\caption{Ask, bid, and price of winning seller-buyer pairs.}
\label{bidaskprice}
\end{figure}
\begin{figure}
\centering
\pgfplotsset{scaled y ticks=false}
\begin{tikzpicture}
\pgfplotsset{
xmin=2, xmax=10,
width=6cm, compat=1.3
}
\begin{axis}[
axis y line*=left,
ymin=0, ymax=0.4,
xlabel=Number of Buyers,
ylabel=Average Utility of Winning Sellers,
xtick={2,3,4,5,6,7,8,9,10},
ytick={0,0.1,0.2,0.3,0.4},
ymajorgrids=true,
xmajorgrids=true,
grid style=dashed,
legend cell align={left},
every axis plot/.append style={thick},
/pgf/number format/.cd,
1000 sep={}
]
\addplot[mark=x,red]
coordinates{
(2,0.273609416)
(4,0.274812524)
(6,0.284980794)
(8,0.29956794)
(10,0.315792566)
}; \label{seller}
\end{axis}
\begin{axis}[
axis y line*=right,
yticklabel style={
/pgf/number format/fixed,
/pgf/number format/precision=2
},
ytick={0,0.01,0.02,0.03,0.04},
legend pos=north west,
legend style={fill=white, nodes={scale=0.8, transform shape}},
axis x line=none,
ymin=0, ymax=0.04,
ylabel=Average Utility of Winning Buyers,
]
\addlegendimage{/pgfplots/refstyle=seller}\addlegendentry{Sellers}
\addplot[mark=*,blue]
coordinates{
(2,0.020157724)
(4,0.005530165)
(6,0.004171116)
(8,0.003831985)
(10,0.004775161)
}; \addlegendentry{Buyers}
\end{axis}
\end{tikzpicture}
\caption{Average utility of winning buyers and sellers.}
\label{fig:totalutil}
\end{figure}
\begin{figure}[t]
\centering
\pgfplotsset{scaled y ticks=false}
\begin{tikzpicture}
\pgfplotsset{
xmin=2, xmax=10,
width=7cm, compat=1.3
}
\begin{axis}[
ymin=0, ymax=0.6,
xlabel=Number of Buyers,
ylabel=Average Utility of Winning Sellers,
xtick={2,3,4,5,6,7,8,9,10},
ytick={0,0.1,0.2,0.3,0.4,0.5,0.6},
ymajorgrids=true,
xmajorgrids=true,
grid style=dashed,
legend cell align={left},
legend style={fill=white, nodes={scale=0.8, transform shape}},
every axis plot/.append style={thick},
/pgf/number format/.cd,
1000 sep={}
]
\addplot[mark=x,red]
coordinates{
(2,0.273609416)
(4,0.274812524)
(6,0.284980794)
(8,0.29956794)
(10,0.315792566)
}; \addlegendentry{with Deep Learning}
\addplot[mark=*,blue]
coordinates{
(2,0.244814444)
(4,0.239604142)
(6,0.251322238)
(8,0.21908414)
(10,0.274147555)
}
; \addlegendentry{baseline}
\end{axis}
\end{tikzpicture}
\caption{Average utility of winning sellers with and without deep learning (baseline) in the double auction mechanism.}
\label{utildl}
\end{figure}
\begin{figure}[t]
\centering
\pgfplotsset{scaled y ticks=false}
\begin{tikzpicture}
\pgfplotsset{
xmin=2, xmax=10,
width=7cm, compat=1.3
}
\begin{axis}[
ymin=0, ymax=0.7,
xlabel=Number of Buyers,
ylabel=Average Utility of Winning Sellers,
xtick={2,3,4,5,6,7,8,9,10},
ytick={0,0.1,0.2,0.3,0.4,0.5,0.6,0.7},
ymajorgrids=true,
xmajorgrids=true,
grid style=dashed,
legend cell align={left},
legend style={fill=white, nodes={scale=0.8, transform shape}},
every axis plot/.append style={thick},
/pgf/number format/.cd,
1000 sep={}
]
\addplot[mark=x,red]
coordinates{
(2,0.296050726)
(4,0.256871025)
(6,0.308167955)
(8,0.337936519)
(10,0.364629871)
}; \addlegendentry{with Deep Learning, $\theta_m \geq 0.5$}
\addplot[dashed,red]
coordinates{
(2,0.100113574)
(4,0.217876129)
(6,0.152168025)
(8,0.15119952)
(10,0.154217376)
}
; \addlegendentry{with Deep Learning, $\theta_m < 0.5$}
\addplot[mark=*,blue]
coordinates{
(2,0.252222063)
(4,0.220542363)
(6,0.268727023)
(8,0.248645835)
(10,0.315370008)
}
; \addlegendentry{without Deep Learning, $\theta_m \geq 0.5$}
\addplot[mark=*,blue,dashed]
coordinates{
(2,0.039493662)
(4,0.135004628)
(6,0.156426408)
(8,0.124685486)
(10,0.117285937)
}
; \addlegendentry{without Deep Learning, $\theta_m < 0.5$}
\end{axis}
\end{tikzpicture}
\caption{Average utility of winning sellers with different ranges of $\theta_m$.}
\label{utiltheta}
\end{figure}
\begin{figure}[t]
\subfigure[]{
\centering
\pgfplotsset{scaled y ticks=false}
\begin{tikzpicture}
\pgfplotsset{
xmin=2, xmax=10,
width=4cm, compat=1.17
}
\begin{axis}[
label style={font=\scriptsize},
tick label style={font=\tiny},
yticklabel style={
/pgf/number format/fixed,
/pgf/number format/precision=2},
ymin=0, ymax=0.6,
xlabel=Number of Buyers,
ylabel=Similarity Score,
xtick={2,3,4,5,6,7,8,9,10},
ytick={0,0.1,0.2,0.3,0.4,0.5,0.6},
ymajorgrids=true,
xmajorgrids=true,
grid style=dashed,
legend cell align={left},
legend style={fill=white, nodes={scale=0.5, transform shape}},
every axis plot/.append style={thick},
/pgf/number format/.cd,
1000 sep={}
]
\addplot[mark=x,red]
coordinates{
(2,0.31661977563455185)
(4,0.37063285257855044)
(6,0.326431110125172)
(8,0.3337682992794319)
(10,0.3457011779589282)
}; \addlegendentry{$\theta_m \geq 0.5$}
\addplot[mark=*,blue]
coordinates{
(2,0.1669973423216254)
(4,0.1426884590199269)
(6,0.1688055098946527)
(8,0.1443021703427639)
(10,0.1588024486502578)
}
; \addlegendentry{$\theta_m < 0.5$}
\end{axis}
\end{tikzpicture}
}
\subfigure[]{
\centering
\pgfplotsset{scaled y ticks=false}
\begin{tikzpicture}
\pgfplotsset{
xmin=2, xmax=10,
width=4cm, compat=1.17
}
\begin{axis}[
label style={font=\scriptsize},
tick label style={font=\tiny},
yticklabel style={
/pgf/number format/fixed,
/pgf/number format/precision=2},
ymin=0, ymax=0.6,
xlabel=Number of Buyers,
ylabel=BLEU Score,
xtick={2,3,4,5,6,7,8,9,10},
ytick={0,0.1,0.2,0.3,0.4,0.5,0.6},
ymajorgrids=true,
xmajorgrids=true,
grid style=dashed,
legend cell align={left},
legend style={fill=white, nodes={scale=0.5, transform shape}},
every axis plot/.append style={thick},
/pgf/number format/.cd,
1000 sep={}
]
\addplot[mark=x,red]
coordinates{
(2,0.32052125495177236)
(4,0.3791832170253799)
(6,0.33069722450874767)
(8,0.33690080743964856)
(10,0.3474632178279614)
}; \addlegendentry{$\theta_m \geq 0.5$}
\addplot[mark=*,blue]
coordinates{
(2,0.16595875755783013)
(4,0.14013916301765394)
(6,0.1683687743248769)
(8,0.14187539916126404)
(10,0.15628916855205552)
}
; \addlegendentry{$\theta_m < 0.5$}
\end{axis}
\end{tikzpicture}
}
\caption{(a) Average similarity score of sellers (b) Average BLEU score of sellers with different cost of the semantic model, $\theta_m$.}
\label{fig:bleusim}
\end{figure}
\begin{figure}[t]
\centering
\pgfplotsset{scaled y ticks=false}
\begin{tikzpicture}
\pgfplotsset{
xmin=2, xmax=10,
width=7cm, compat=1.3
}
\begin{axis}[
ymin=0, ymax=9,
xlabel=Number of Buyers,
ylabel=Average Number of Winning Sellers,
xtick={2,3,4,5,6,7,8,9,10},
ytick={0,1,2,3,4,5,6,7,8,9},
ymajorgrids=true,
xmajorgrids=true,
grid style=dashed,
legend cell align={left},
legend style={fill=white, nodes={scale=0.8, transform shape}},
every axis plot/.append style={thick},
/pgf/number format/.cd,
1000 sep={}
]
\addplot[mark=x,red]
coordinates{
(2,1.3)
(4,2.8)
(6,3.3)
(8,5.4)
(10,5.4)
}; \addlegendentry{with Deep Learning, $\theta_m \geq 0.5$}
\addplot[dashed,red]
coordinates{
(2,0.7)
(4,1.2)
(6,2.3)
(8,1.7)
(10,1.9)
}
; \addlegendentry{with Deep Learning, $\theta_m < 0.5$}
\addplot[mark=*,blue]
coordinates{
(2,1.8)
(4,2.9)
(6,3.7)
(8,5.2)
(10,5.5)
}
; \addlegendentry{without Deep Learning, $\theta_m \geq 0.5$}
\addplot[mark=*,blue,dashed]
coordinates{
(2,0.2)
(4,1.1)
(6,1.9)
(8,1.9)
(10,1.8)
}
; \addlegendentry{without Deep Learning, $\theta_m < 0.5$}
\end{axis}
\end{tikzpicture}
\caption{Average number of winning sellers with and without deep learning in the double auction mechanism.}
\label{fig:noseller}
\end{figure}
\begin{figure}[t]
\subfigure[]{
\centering
\pgfplotsset{scaled y ticks=false}
\begin{tikzpicture}
\pgfplotsset{
xmin=0.1, xmax=1,
width=4cm, compat=1.17
}
\begin{axis}[
label style={font=\scriptsize},
tick label style={font=\tiny},
yticklabel style={
/pgf/number format/fixed,
/pgf/number format/precision=2},
ymin=-0.2, ymax=0.2,
xlabel=Ask,
ylabel=Utility,
xtick={0.2,0.4,0.6,0.8,1.0},
ytick={-0.2,-0.1,0,0.1,0.2},
ymajorgrids=true,
xmajorgrids=true,
grid style=dashed,
legend cell align={left},
legend style={fill=white, nodes={scale=0.5, transform shape}},
every axis plot/.append style={thick},
/pgf/number format/.cd,
1000 sep={}
]
\addplot[mark=x,blue,dashed]
coordinates{
(0.1,0.12)
(0.19,0.12)
(0.2,0.12)
(0.3,0)
(0.4,0)
(0.5,0)
(0.6,0)
(0.7,0)
(0.8,0)
(0.9,0)
(1.0,0)
};
\end{axis}
\end{tikzpicture}
}
\subfigure[]{
\centering
\pgfplotsset{scaled y ticks=false}
\begin{tikzpicture}
\pgfplotsset{
xmin=0.1, xmax=1,
width=4cm, compat=1.17
}
\begin{axis}[
label style={font=\scriptsize},
tick label style={font=\tiny},
yticklabel style={
/pgf/number format/fixed,
/pgf/number format/precision=2},
ymin=-0.2, ymax=0.2,
xlabel=Ask,
ylabel=Utility,
xtick={0.2,0.4,0.6,0.8,1.0},
ytick={-0.2,-0.1,0,0.1,0.2},
ymajorgrids=true,
xmajorgrids=true,
grid style=dashed,
legend cell align={left},
legend style={fill=white, nodes={scale=0.8, transform shape}},
every axis plot/.append style={thick},
/pgf/number format/.cd,
1000 sep={}
]
\addplot[mark=x,blue,dashed]
coordinates{
(0.1,0)
(0.18,0)
(0.2,0)
(0.3,0)
(0.4,0)
(0.5,0)
(0.6,0)
(0.7,0)
(0.8,0)
(0.9,0)
(1.0,0)
};
\end{axis}
\end{tikzpicture}
}
\caption{Utility when asking untruthfully by (a) seller $m$ that wins the auction, $m \in \mathcal{M}_w$ (b) seller $m$ that loses the auction, $m \notin \mathcal{M}_w$.}
\label{fig:truthseller}
\end{figure}
To validate that the double auction mechanism is individually rational and budget balanced, we record the values of ask, bid, and price in one of the samples with $M=20$ and $N=10$. The values are shown in Fig. \ref{bidaskprice}. We observe that there are totally 7 winning seller-buyer pairs, and the utilities for all of the winning pairs are positive. This means that the winning sellers are paid higher than their cost, and the winning buyers pay no more than their true valuation for the semantic information. Therefore, both buyers and sellers have incentives to participate in the auction. For the losing sellers and buyers, their utilities are zero. This shows that the property of individual rationality is achieved because all of the buyers and sellers are awarded with a non-negative utility. The price paid by winning sellers is equal to the payment received by the winning buyers. Thus, the budget balanced property is satisfied.
\begin{figure}[t]
\subfigure[]{
\centering
\pgfplotsset{scaled y ticks=false}
\begin{tikzpicture}
\pgfplotsset{
xmin=0.1, xmax=1,
width=4cm, compat=1.17
}
\begin{axis}[
label style={font=\scriptsize},
tick label style={font=\tiny},
yticklabel style={
/pgf/number format/fixed,
/pgf/number format/precision=2},
ymin=-0.2, ymax=0.2,
xlabel=Bid,
ylabel=Utility,
xtick={0.2,0.4,0.6,0.8,1.0},
ytick={-0.2,-0.1,0,0.1,0.2},
ymajorgrids=true,
xmajorgrids=true,
grid style=dashed,
legend cell align={left},
legend style={fill=white, nodes={scale=0.5, transform shape}},
every axis plot/.append style={thick},
/pgf/number format/.cd,
1000 sep={}
]
\addplot[mark=x,blue,dashed]
coordinates{
(0.1,0)
(0.19,0.00177)
(0.2,0)
(0.3,-0.09045307895236071)
(0.4,-0.0775076057582384)
(0.5,-0.09628250261836108)
(0.6,-0.08214270135455187)
(0.7,-0.05417460760169085)
(0.8,-0.10288744231753405)
(0.9,-0.08905493279986437)
(1.0,-0.10054930111460741)
};
\end{axis}
\end{tikzpicture}
}
\subfigure[]{
\centering
\pgfplotsset{scaled y ticks=false}
\begin{tikzpicture}
\pgfplotsset{
xmin=0.1, xmax=1,
width=4cm, compat=1.17
}
\begin{axis}[
label style={font=\scriptsize},
tick label style={font=\tiny},
yticklabel style={
/pgf/number format/fixed,
/pgf/number format/precision=2},
ymin=-0.2, ymax=0.2,
xlabel=Bid,
ylabel=Utility,
xtick={0.2,0.4,0.6,0.8,1.0},
ytick={-0.2,-0.1,0,0.1,0.2},
ymajorgrids=true,
xmajorgrids=true,
grid style=dashed,
legend cell align={left},
legend style={fill=white, nodes={scale=0.8, transform shape}},
every axis plot/.append style={thick},
/pgf/number format/.cd,
1000 sep={}
]
\addplot[mark=x,blue,dashed]
coordinates{
(0.1,0)
(0.2,-0.038536677140890735)
(0.3,-0.12262935378044762)
(0.4,-0.1268948707768313)
(0.5,-0.10954130627602257)
(0.6,-0.11443690397232689)
(0.7,-0.09042409994095482)
(0.8,-0.09685430862396874)
(0.9,-0.10318235494583763)
(1.0,-0.1152497027108065)
};
\end{axis}
\end{tikzpicture}
}
\caption{Utility when bidding untruthfully by (a) buyer $n$ that wins seller $m$ in the auction, $n \in \mathcal{N}_w$ (b) buyer $n$ that loses seller $m$ in the auction, $n \notin \mathcal{N}_w$.}
\label{fig:truthbuyer}
\end{figure}
The average utility of the winning buyers and sellers are presented in Fig. \ref{fig:totalutil}, which is obtained by averaging the values of 1000 samples. Intuitively, as the number of buyers increases, the sellers have more choices to achieve higher utilities. From Fig. \ref{fig:totalutil}, we observe that the auction mechanism helps to increase the average utility of the winning sellers as the number of buyers grows. Thus, our proposed mechanism can attract more sellers to participate in the information exchange with semantic communication systems.
To investigate the impact of DL in the double auction, we compare the average utilities of winning sellers with and without DL mechanism. The results without the DL mechanism (i.e., the baseline) are obtained by using the double auction mechanism proposed in \cite{jin2015auction}. It is shown in Fig. \ref{utildl} that the average utility of the winning sellers is higher when DL mechanism is adopted in the double auction. The reason is that the DL mechanism helps to maximize the revenue of the sellers.
As shown in Fig. \ref{utiltheta}, the average utility of the winning sellers is higher when the sellers set the payment price $\theta_m \geq 0.5$. The reason is that the sellers with higher $\theta_m$ obtain the semantic model which has a higher BLEU score and similarity score. Hence the buyers are willing to pay more to obtain more accurate information. This insight is verified in Fig. \ref{fig:bleusim}, in which we can see that the similarity score and the BLEU score are higher for sellers with $\theta_m \geq 0.5$. The higher similarity and BLEU scores motivate the buyers to submit higher bids to the sellers, which results in higher utilities as shown in Fig. \ref{utiltheta}. Furthermore, it is shown in Fig. \ref{fig:noseller} that there are more sellers with $\theta_m \geq 0.5$ from the winning sellers. In other words, the seller with higher $\theta_m$ has a higher chance to win the auction, regardless of the number of buyers.
To verify the truthfulness of the double auction, the sellers and buyers are randomly chosen to evaluate their utilities when their bid and ask are different from their true valuation. In Fig. \ref{fig:truthseller} (a), seller $m$ wins and gains the utility $u^s_m = 0.12$ when it asks truthfully with $a_m = C_m = 0.19$. It is shown that the utility cannot be improved by other values of ask. From Fig. \ref{fig:truthseller} (b), seller $m$ loses the auction with truthful ask $u^s_m = 0.18$ obtaining zero utility. It is shown that seller $m$ does not obtain a higher utility when asking untruthfully. In Fig. \ref{fig:truthbuyer} (a), buyer $n$ wins seller $m$ when it bids truthfully with $b^m_n = v^m_n = 0.19$ achieving a utility $u^b_{n,m}=0.0018$. There is no other higher utility achieved when it bids untruthfully. Fig. \ref{fig:truthbuyer} (b) shows the scenario when buyer $n$ does not win seller $m$ and achieve a non-positive utility when it bids untruthfully.
From the experiment results, we observe that the sellers that pay higher prices for the semantic models can achieve better similarity and BLEU scores in the double auction. It is shown that the sellers with better performance are more likely to win the auction and obtain higher utilities. Numerical results also show that the proposed double auction is incentive compatible, individually rational, and budget balanced.
\section{Conclusion And Future Directions}\label{sec:conclusion}
In this paper, we have proposed incentive mechanisms for both semantic model trading and semantic information trading. We developed the valuation functions for general semantic communications, and performed a case study of the proposed auctions for semantic text transmission. To improve the system performance, we have proposed an effective feature reduction method to support devices with limited transmission resources. Simulation results show that the proposed method helps to increase significantly the utility of devices in the semantic information trading. Moreover, with the double auction mechanism, we have matched the buyers and devices effectively. It is also shown that the revenue of the semantic model provider can be maximized while keeping the properties of incentive compatibility and individual rationality. For future research directions, we can consider the semantic-aware incentive mechanism design in non-text-based transmission such as wireless images and video transmission, and other semantic-based intelligent tasks.
For future works, considering that the raw data collected from different regions decays over time, we can count the age of information in the value functions of raw data. The difference in the age of information can also be taken into account in the evaluation of transmission accuracy. Moreover, we can consider that the semantic information, which is extracted from different types of raw data, e.g., text, image, and audio, has difference values.
\bibliographystyle{IEEEtran}
|
\section{Discussion}
Our experiments demonstrate LRL via positive FT and BT, high performance relative to an STE, reduced catastrophic forgetting, and fast adaptation to new tasks. Our work clearly advances the domain-specific state-of-the-art in \SC{} by reducing the sample complexity of deep RL to master the minigames.
Compared to STE, our experiments required only 6\% of training samples to achieve 80-90\% of expert performance.
Existing lifelong RL implementations (e.g., \citep{mendez2020lifelong}) have been tested on benchmark problems, but proved to be challenging to scale up to the complex observation and action space of SC-2. Future work can leverage our techniques in the full SC-2 game and reduce the astronomical sample complexity of current deep RL agents in SC-2 (e.g., AlphaStar \citep{vinyals2019grandmaster}).
While the wake-sleep mechanism and model-free GR show promising results, there is room for improvement, and the approach is easily extensible. We highlight several potential extensions. Our current model triggers sleep on a set schedule. Instead, it is possible to self-trigger sleep via recent work in lifelong changepoint detection \citep{faber2021watch}. Currently, we store full observations in the replay buffer, which due to memory constraints, limits the number of samples that can be saved. It would be prudent to intelligently compress experiences in the replay buffer \citep{riemer2019scalable, wang2021acae, caccia2020online,zhang2017deeper,schaul2015prioritized, hayes2019remind}. Past work has looked at improving replay buffers via prioritized replay \citep{schaul2015prioritized} based on detecting dead-end states.
We further hypothesize that increasing the lifetime of an agent by avoiding dead end states is a useful bias. Our methods operate over hand-crafted representations of the observation space; the system could be extended by introducing representation learning from RGB-observations via unsupervised (STAM \citep{smith2019unsupervised,smith2019unsupervised2}) and self-supervised learning \citep{gallardo2021self}. We can further extend our models using mixtures of VAEs (e.g., the Eigentask framework \cite{raghavan2020lifelong}) to dynamically grow and shrink the model, leading towards agents that build implicit models of the tasks they experience. Finally, we used a fixed neural architecture and loss functions throughout the agent's lifetime. GR is a flexible approach that can include progress in other areas of lifelong learning, e.g., dynamically changing the underlying neural architectures \citep{schwarz2018progress}, neuromodulation \citep{brna2019uncertainty}, and regularization \citep{kolouri2019sliced}.
Further, interesting avenues include the application of GR to multi-agent RL \citep{nekoei2021continuous} and federated learning \citep{rostami2017multi}, where communication between agents can enable better coordination between agents and collective acquisition of knowledge. Due to the inherent plastic-yet-stable nature of lifelong learners, our methods can be more robust to data drift, concept drift, or task changes, reducing system downtime and the overhead of retraining, and used to address model obsolescence and reduce the technical debt of machine learning models \citep{sculley2015hidden, alahdab2019empirical, bogner2021characterizing}.
\section{Hidden Replay for Lifelong RL}
\label{sec:model-free-generative-replay}
This section introduces our core contribution of wake-sleep training combined with model-free GR of latent representations. We also explore two other GR architectures along the way.
The fundamental tradeoff in lifelong learning is between \emph{plasticity}, i.e.\!, learning the current task, and \emph{stability}, i.e.\!, remembering past tasks.
We address this tradeoff in lifelong RL by employing an approach that alternates between two distinct phases \citep{raghavan2020lifelong}.
In the wake phase, a \emph{plastic} wake policy $\pi_w$ is optimized for the current task by interaction with the environment and using any off-the-shelf RL method.
Transition tuples are collected during training and stored in a buffer (see Alg. \ref{alg:wakesleep}); each tuple contains $(o, r, a)$, the observation $o$, reward for the previous action $r$, and the policy output $a$ (e.g., the policy logits or one-hot encoded action).
The sleep policy $\pi_s$ provides ``advice'' to explore the current task using the consolidated policy learned from all previous tasks.
Wake phase actions are selected according to a mixture of $\pi_w$ and $\pi_s$, the probability of choosing an action from $\pi_s$ decaying to 0 over time.
We use an off-policy RL algorithm such as VTrace \citep{espeholt2018impala} to accommodate this off-policy action selection in the optimization of $\pi_w$.
In the sleep phase, the \emph{stable} sleep policy $\pi_s$ is optimized with the objective of incorporating new knowledge (the action selection in the wake buffer) while retaining current knowledge.
Similar to generative replay in supervised learning, augmented dataset(s) are created by combining wake transitions with tuples from the generative model $g_s$ and pseudo-labelled by the sleep policy.
This form of replay is model-free because it does not learn the POMDP dynamics and reward and does not generate and replay temporal trajectories.
The sleep policy and the generative model are jointly trained.
The sleep policy is trained to minimize $\mathcal{L}_{xent}$, a standard cross-entropy loss between the outputs (logits) of $\pi_s$, treating the replay augmented dataset as the ground truth.
The generative model is trained using the combined dataset and the standard VAE loss (Eq. \ref{loss:vae}).
The description so far is easily applied to veridical replay, i.e.\!, when generative modelling and replay occurs in the original observation space of the MDP, but this is undesirable as discussed before. The design choice of whether the input or hidden space is replayed distinguishes the model implementations (Alg. \ref{alg:wakesleep} vs Alg. \ref{alg:hidden_replay}, and models in Fig. \ref{fig:all_replay_archs}) from each other.
\begin{minipage}{0.49\textwidth}
\begin{algorithm}[H]
\caption{Wake-Sleep Training in the Sequential and Two-Headed models.}
\label{alg:wakesleep}
\begin{algorithmic}[1]
\STATE {\bfseries Iterates:} Generator $g_s^t$, sleep policy $\pi_{s}^t$, wake policy $\pi_{w}^t$, wake buffer $b^t_w$
\FOR{$t=1,2,\ldots$}
\FOR[Wake Phase]{$K$ times}
\STATE{Sample observation $o$ and reward $r$ from current task}
\STATE Sample $a \sim \pi_s^t(o)$ w.p. $p_{\text{advice}}$, else $a \sim \pi_w^t(o)$
\STATE{Add transition $(o, r, a)$ to FIFO wake buffer $b_w^t$}
\STATE{Update wake policy $\pi^t_w$ on current task reward using $b^t_w$}
\STATE{Decay $p_\text{advice}$}
\ENDFOR\\
\FOR[Sleep Phase]{$N$ iterations}
\STATE{Sample batch $B$ from $b_w^t$ (c.f.\! Line 6).}
\STATE Sample batches $O_s^t \sim g_s^t$
\STATE Pseudo-label $A_s^t = \pi_s^t(O_s^t)$
\STATE (GR Buffer) $D = B \cup (O_s^t, A_s^t)$
\STATE Minimize $\mathcal{L}_{VAE}(g_s^{t+1}) + \alpha\mathcal{L}_{xent}(g_s^{t+1}, \pi_s^{t+1})$ on $D$
\ENDFOR\\
\ENDFOR
\end{algorithmic}
\end{algorithm}
\end{minipage}
\hfill
\begin{minipage}{0.49\textwidth}
\begin{algorithm}[H]
\caption{Wake-Sleep Training in the Hidden Replay model.}
\label{alg:hidden_replay}
\begin{algorithmic}[1]
\STATE {\bfseries Iterates:} Generator $g_s^t$, sleep policy $\pi_{s}^t$, wake policy $\pi_{w}^t$, wake buffer $b^t_w$, random replay buffer $b^t_r$
\FOR{$t=1,2,\ldots$}
\FOR[Wake Phase]{$K$ times}
\STATE{Sample observation $o$ and reward $r$ from current task}
\STATE Sample $a \sim \pi_s^t(o)$ w.p. $p_{\text{advice}}$, else $a \sim \pi_w^t(o)$
\STATE{Add transition $(o, r, a)$ to wake buffer $b^t_w$}
\STATE{Update wake policy $\pi^t_w$ on current task reward using $b^t_w$}
\STATE{Decay $p_\text{advice}$}
\ENDFOR\\
\FOR[Sleep Phase]{$N$ iterations}
\STATE{If ER: Batch $B \sim b^t_w$ (c.f.\! Line 6).}
\STATE{If GR: Feature batches $H^t_s \sim g_s^t$}
\STATE{If GR: Pseudo-label $A^t_s \sim \pi_s^t(H_s^t)$}
\STATE{If RaR: $b^t_r = b^t_r \bigcup b^t_w.sample(K)$}
\STATE{If RaR: $D_r \sim b^t_r$}
\STATE Minimize $\mathcal{L}_{VAE}(g_s^{t+1}) + \alpha\mathcal{L}_{xent}(g_s^{t+1}, \pi_s^{t+1})$ on $(H^t_s, A^t_s)$ and end-to-end minimize (including feature extractor) the same loss on observation-action pairs in $B \bigcup D_r$
\ENDFOR\\
\ENDFOR
\end{algorithmic}
\end{algorithm}
\end{minipage}
\hfill
\begin{figure}[t]
\centering
\includegraphics[width=0.92\textwidth]{graphics/updates_archs_2.png}
\caption{We investigate three replay architectures for model-free generative replay (sleep phase). The red dashed arrows denotes sampling steps starting from the VAE to generate labelled actions. The blue dashed arrows show how the various replay buffers are sampled and fed into the sleep model during training. The black arrows denote the forward pass. \textbf{Top:} Sequential Replay Architecture; \textbf{Middle:} Two-Headed Replay Architecture; \textbf{Bottom:} Hidden Replay Architecture.}
\label{fig:all_replay_archs}
\end{figure}
We compare three replay architectures that we denote Sequential, Two-Headed, and Hidden Replay (shown in Fig. \ref{fig:all_replay_archs}). Each architecture is a specific arrangement of the following three components: a feature extractor, a VAE, and a policy network. The feature extractor maps raw image-like observations to feature vectors using a convolutional neural network (CNN) or a multi-layered perceptron. The policy network is dependent on the action space, e.g., in our experiments we utilize a convolutional architecture to output spatial actions.
In the \hypertarget{text:sequential}{\textbf{Sequential}} architecture (Fig.~\ref{fig:all_replay_archs}, Top), the VAE is upstream of the policy network, and the sleep policy $\pi_s$ operates on the VAE reconstruction rather than the true observation. This architecture is similar to the one proposed in \citet{atkinson2021pseudo}.
During sleep (following the red dashes in Fig.~\ref{fig:all_replay_archs}, Top), sampled latent vectors are decoded to observations that are then labelled by the policy. We hypothesize that the sequential architecture tends to map observations for which the policy chooses similar actions to generic reconstructions. We found that this method produces blurry reconstructions for image-like observations.
The \hypertarget{text:twoheaded}{\textbf{Two-Headed}} architecture (Fig.~\ref{fig:all_replay_archs}, Middle) decouples the decoder/generator from the policy network. Each observation (real or generated) is passed through the feature extractor. The features are then separately passed through the VAE and the policy network. Our experiments showed improved visual reconstructions (see Appendix \ref{sec:reconstructions_app}) due to separation of the influence of the policy on the VAE decoder. However, sampling in this architecture requires two forward passes as shown in the figure (once to decode an observation and once to label it).
Alg. \ref{alg:wakesleep} shows the data flow that is common between these two architectures that perform veridical replay, i.e.\!, the VAE posterior is a density over raw and complex perceptual space. The training is wasteful because the wake policy $\pi_w$ has already learned an effective feature extractor, and latent-to-action partition that works well for the current task (solved in wake phase). The sleep phase in veridical replay does not utilize these learned features and tries to re-learn them, whereas deep RL (as in wake phase) is effective at learning representations to map raw percepts to actions.
\setlength{\columnsep}{5pt}%
\begin{wrapfigure}{R}{0.35\textwidth}
\centering
\includegraphics[width=0.35\textwidth, keepaspectratio]{graphics/hidden_replay_partitioning.png}
\vspace{-0.75cm}
\caption{Hidden replay prevents drift in the features-to-action
mapping from the latent vector space of a deep RL agent (Minigrid tasks). Colors represent actions; $\times$s: drawn from the random replay buffer, $\bullet$: drawn from wake buffer; $\star$: sampled via generative replay.}
\vspace{-0.75cm}
\label{fig:hidden_replay_partitioning}
\end{wrapfigure}
The final architecture we consider is inspired by the success of \hypertarget{text:hidden}{\textbf{Hidden Replay}} in class incremental learning \citep{van2020brain} and its connection to biology. This architecture (Fig.~\ref{fig:all_replay_archs}, Bottom) generates feature vectors directly rather than observations. Sampling features instead of image-like observations benefits from dimensionality reduction. Reconstructing or generating complex image observations is hard and requires increasingly complex generative models \citep{van2017neural,van2016conditional,razavi2019generating}. In RL, especially when working with ``symbolic'' (semantic) (as opposed to ``perceptual'' (RGB)) observations, changing a small detail of the observation can change the optimal action drastically. Observations might be better separated in feature space, and thus, the sleep policy could be less sensitive to an imperfect decoder. In addition, in contrast to the two-headed architecture, the feature extractor does not need to be run when sampling.
For example, when faced with new tasks that require new actions to be taken, hidden replay can learn to add a new action partition to the feature space while rearranging the existing feature space in a manner that preserves partitioning needed to solve previous tasks. Empirically, we observe a clean partition of the feature space according to the optimal action. Even though samples and reconstructions are not perfect, they fall within the same action partition (Fig.~\ref{fig:hidden_replay_partitioning}) as the ground truth. These representations may allow for knowledge transfer between tasks because tasks requiring similar actions can potentially efficiently be mapped into the existing feature space, but validating this requires further studies.
Hidden replay is not without limitations because GR samples are in terms of features without the corresponding observations. This V-shape in the training graph (blue lines in Fig. \ref{fig:all_replay_archs}, Bottom) can cause the feature extractor to drift over time or collapse to a mode (e.g., map all observations to the same feature vector).
As new tasks are encountered, this would cause sampled features
to not align with the features mapped by the policy and VAE components.
To limit feature drift, we store a small (typically $K < 1\%$ of the experience replay buffer) set of raw observations paired with actions in a separate buffer. We denote this RaR (for ``random replay'' buffer). These are accumulated over the lifetime of the agent from all wake replay buffers $b^t_w$ (although a more sophisticated exemplar selection \citep{rebuffi2017icarl,mi2020ader} is also possible). We observed that replaying a few random examples results in less feature drift between task changes (See Appendix ~\ref{sec:justification_of_exemplars} for an ablation). In our experiments, we found that saving 96 random samples per sleep was sufficient to alleviate feature drift in the \SC{} domain and saving 256 random samples per sleep was sufficient for the Minigrid domain. The complete data flow for hidden replay for LRL is shown in Alg. \ref{alg:hidden_replay}, where the design choice of combinations of ER, GR, and RaR buffers are ablated in our experimental ablations.
\textbf{Remark}: When applied to single task RL, our wake-sleep cycle resembles policy iteration interleaved with policy distillation steps, where each iteration $t$ produces an improved $\pi_w^t$ (nudge towards STE) that is then distilled into $\pi_s^t$. Wake experiences are gathered from the most recent sleep policy (``advice''), including from multiple sub-optimal policies similar in spirit to dataset aggregation in DAGGER \citep{ross2011reduction}. Similar to DAGGER, the consolidated policy is executed to sample next trajectories. Our supplementary experiments (Appendices \ref{sec:wake_additional_results} and \ref{sec:ste_additional_results}) show stable convergence of the wake-sleep cycle to optimal policies in single task training (convergence to STE).
\section{Introduction}
Lifelong Reinforcement Learning (LRL) involves training an agent to maximize its cumulative performance on a stream of changing \textit{tasks} over a long lifetime. LRL agents must balance \textit{plasticity vs. stability}: learning the current task while maintaining performance on previous tasks. Deep neural networks are especially prone to instability, often exhibiting \textit{catastrophic forgetting} of previous tasks when trained on multiple tasks that are presented sequentially \citep{kirkpatrick_overcoming_2016}. One approach to meet the challenges of deep LRL is by careful managing the agent's learning experiences, in order to learn (without forgetting) and build internal meta-models (of the tasks, environments, agents, and world). One strategy for managing experiences is to recall data from previous tasks and mix it with data from the current task when training, essentially transforming sequential LRL into batch multi-task reinforcement learning (RL) \citep{brunskill2013sample}. The technique of \textit{experience replay} of past data is common in deep RL \citep[e.g.][]{mnih2013playing} and has been studied in the lifelong learning literature \citep{isele2018selective,rolnick2019experience,hayes2021replay}.
However, storing sufficient examples from all previous tasks may be unreasonable in systems with limited resources or long lifetimes.
Rather than store an ever-growing list of experiences, one can instead employ \textit{generative replay} (GR) \citep{shin2017continual,van_de_ven_generative_2018}, a biologically-inspired replay mechanism that augments learning experiences with self-labelled examples drawn from an internal generative model trained to approximate the distribution of past experiences.
In LRL (see \citet{khetarpal2020towards} for a survey), each task is a Markov decision process (MDP), and the agent must learn a policy for interacting with its environment to maximize its lifetime performance (e.g., average reward over all tasks).
While GR has been shown to be effective in lifelong supervised learning \citep{van_de_ven_generative_2018}, it has been less explored in LRL.
The LRL setting is complicated by the fact that the agent's experiences are observation-action \emph{trajectories} whose distribution depends on the environment dynamics and the policy being executed.
While prior work with world models \citep{ketz2019continual} and pseudo-rehearsal \citep{atkinson2021pseudo} focused on generative replay of (partial) trajectories, this requires modeling the dynamics, which is difficult in complex environments.
In this paper, we present agent architectures for LRL that use GR for memory consolidation and
satisfy two desiderata: (a) Introspective density modelling of the latent representations of policies learned using deep RL, and (b) Model-free end-to-end learning.
The first property avoids the challenges of density modelling of complex high-dimensional perceptual inputs, whereas policy learning using deep RL works well with such perceptual inputs.
The second property avoids the challenges of learning temporal dynamics and reward functions from few learning experiences with sparse rewards.
Our contributions extend powerful \emph{model-free} deep RL and do not learn the dynamics explicitly.
Our first contribution is a model-free approach to generative replay (Section \ref{sec:model-free-generative-replay}) in lifelong RL. We empirically demonstrate that replay of independently sampled observation-action pairs is sufficient to mitigate forgetting. This is noteworthy because observation-action pairs are simpler to generate than trajectories, and because replay of observation-action pairs is directly analogous to replay of input-label pairs in supervised learning, thus unifying generative replay methods for lifelong learning.
Our second contribution is a mechanism for memory consolidation via \emph{wake-sleep} cycles. During \emph{wake phases}, our agent improves a policy for the current task(s) using an off-the-shelf RL algorithm while storing samples of experiences in a buffer. Periodically, the agent enters a \emph{sleep phase}, during which the generative memory is trained to model the agent's new experiences, and the procedural knowledge embodied in the policy learned during the wake phase is consolidated into the agent's ``skill set'' via policy distillation \citep{rusu2015policy} with GR. In the next wake phase, a new exploration policy is seeded with knowledge from previously-learned skills, and the cycle repeats.
We investigate three different replay architectures (Section \ref{sec:model-free-generative-replay}) within the wake-sleep framework, starting from a na\"ive generative and adding ingredients to achieve the aforementioned introspective density modelling of the latent representations and model-free end-to-end learning.
We evaluate our algorithm and architectural contributions through experiments (Section \ref{sec:experiments}) on three different scenarios comprising tasks from the \SC{} (SC-2) and Minigrid domains. We report several key findings showing the impact of the design choices on quantitative metrics that include transfer learning, generalization to unseen tasks, fast adaptation after task change, performance compared to task experts, and catastrophic forgetting.
We observe that our GR prevents drift in the features-to-action mapping of a deep RL agent (that learns the features from observations). We find that \emph{hidden replay} (replay of such features) outperforms other replay mechanisms, which agrees with past results in lifelong supervised learning \citep{van2020brain}. Since deep RL in SC-2 is computationally demanding, we release a dataset of trajectories of trained single-task expert policies for future research \footnote{\url{https://github.com/sri-l2m/l2m\_data}}.
\section{Experiments}
\label{sec:experiments}
\subsection{Evaluation scenarios}
\label{sec:scenarios}
A lifelong learning \textit{scenario} describes the order of task presentation and repetition of tasks. A concrete instantiation of a scenario with tasks (or POMDPs) is called a \textit{syllabus}. Each scenario is used to evaluate the learning agent over all syllabi and average metrics. The three scenarios we consider are depicted in Figure~\ref{fig:scenarios}.
\textbf{Pairwise:} Two tasks are presented once each. This scenario evaluates knowledge transfer between two tasks. \textbf{Alternating:} Two tasks are presented with three repetitions of each pair. This scenario evaluates knowledge retention over longer lifetimes. \textbf{Condensed:} A random permutation of all tasks is presented.
This is a useful setting for understanding how an agent performs across different task orderings.
Given a syllabus, an experiment run alternates between \emph{evaluation blocks} (EBs) and \emph{learning blocks} (LBs) \citep[see][]{new2022lifelong}.
We say that a task is \emph{seen} wrt an EB if it has appeared in any LB preceeding it, otherwise it is \emph{unseen}.
During each EB, the average accumulated reward of the agent is evaluated on all tasks in the syllabus (including unseen tasks); during each LB, the agent learns on a single task for a fixed number of interactions.
The STE for each task serves as a baseline and measures the relative performance of the learner wrt asymptotic optimal.
\subsection{Metrics for Lifelong RL}
We run controlled and quantitative experiments and compare to multiple baselines along three sets of metrics.
First, we compare learning behavior by examining the accumulated rewards captured by periodic evaluations of the agent, i.e., we look at the trend of EB rewards averaged over EBs (and tasks).
Second, we aim to understand how the rewards achieved by an agent compares to
a single task expert (STE).
This forms the basis for our second set of metrics: \emph{relative reward (RR)}.
Note that these metrics also focus on
EBs and do not consider
LBs. Finally, we compare algorithms using the lifelong learning metrics defined by \citep{new2022lifelong}, which take into account both LBs and EBs, and aim to frame metrics in a way that are agnostic to the domain.
We begin by describing the RR metrics.
For each task, we train (offline) the STE
using an off-the-shelf deep RL algorithm.
RR metrics measure how well the agent has learned the seen tasks, how well the agent performs over all tasks in the syllabus, and how well the agent generalizes to unseen tasks.
In general, RR is computed as
\begin{align}
RR(\pi,T,S) &= \frac{1}{|T|}\sum_{t \in T}\frac{1}{|S(t)|}\sum_{s \in S(t)} \frac{r_\pi(s)}{r_{\pi^*_s}(s)}
\label{eq:rr}
\end{align}
where $\pi$ is the policy of the learner, $T$ is a set EBs, $S(t)$ is the set of tasks at EB $t$, $\pi^*_s$ is the STE for task $s$, and $r_{\pi}(s)$ is the accumulated reward (return) achieved on task $s$ from using policy $\pi$. Different metrics are defined by setting $S, T$ to different EBs and tasks.
$\mathbf{RR_\Omega}$: $T$ only includes the final EB, and $S(t)$ is the set of seen tasks in the syllabus. $RR_\Omega$ measures how well the agent has learned the seen tasks in the syllabus ``at the end of the day''. To measure catastrophic forgetting, we denote $\mathbf{RR_\sigma}$ where
$T$ includes all EBs, and $S(t)$ is the set of seen tasks in the syllabus encountered in previous LBs up to and including EB $t$. $\mathbf{RR_\upsilon}$ is complementary in that it measures how well the agent generalizes to unseen tasks, $T$ includes all EBs, and $S(t)$ is the set of unseen tasks not encountered in previous LBs up to and including EB $t$. To measure plasticity, we denote $\mathbf{RR_\alpha}$ where $T$ is the set of all EBs, and $S(t)$ is only the task learned in LB $t$ preceeding EB $t$.
Note that in all cases, higher more-positive values are better for all metrics as all our STEs achieve positive return.
As mentioned, we employ established lifelong learning metrics defined by \citet{new2022lifelong}, which capture catastrophic forgetting, transfer between all pairs of tasks, and speed of learning. In all cases, more-positive values are better for all metrics.
\textbf{Forward Transfer Ratio (FTR)} is the ratio of EB performance on an unseen task. \textbf{Backward Transfer Ratio (BTR)} is the ratio of EB performance on a seen task. A value greater than one indicates positive transfer. \textbf{Relative Performance (RP)} is the ratio of the Area Under the Learning Curve (AUC) between the lifelong learner and STE learning curve. A value greater than one indicates either faster learning by the lifelong learner and/or superior asymptotic performance. \textbf{Performance Maintenance (PM)} is the average difference between task performance in any EB that comes after the LB of a seen task.
A value less than $0$ indicates forgetting. Refer to \citep{new2022lifelong} for more details about the metrics and the precise computation.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{graphics/alternating_curves.png}
\caption{Illustrative examples showing EB rewards of baseline and hidden replay in alternating tasks scenario. Column color denotes which task is trained in LBs between EBs. Left: Example syllabus of two similar tasks; the lifelong learner (LL) performs similarly to the baseline.
Right: Examples of two dissimilar tasks where the baseline exhibits catastrophic forgetting, our hidden replay does not.}
\label{fig:dissimilar_alternating}
\vspace{-0.75cm}
\end{figure}
\subsection{Domains}
Our experiments use two RL domains: \SC{} minigames \citep{vinyals2017starcraft} and Minigrid \citep{gym_minigrid}. We evaluate on existing RL tasks (POMDPs) in these domains.
\textbf{Starcraft 2}:
We selected 3 minigames namely \emph{DefeatRoaches}, \emph{DefeatZerglingsAndBanelings}, and \emph{CollectMineralShards}. We created an additional task in each case as described in Appendix ~\ref{sec:sc2_task_descriptions}. We used PySC2 \citep{vinyals2017starcraft} to interface with SC-2 game.
The agent receives positive rewards for collecting resources and defeating enemies, and negative rewards for losing friendly units.
As observations, we used the unit type, selection status, and unit density two-dimensional inputs.
We follow the action space and convolutional policy network architecture defined in \citet{vinyals2017starcraft}.
Single task experts were trained to convergence ($\sim$ 30M environment steps using the VTrace algorithm \citep{espeholt2018impala}). We consider the pairwise, alternating, and condensed scenarios where each LB consists of 2 million environment steps (about 6\% of the STE samples), and each EB consists of 30 episodes per task. Complete details can be found in the Appendix \ref{sec:imp_sc2}.
\textbf{Minigrid}:
Due to the heavy computational load of experimentation in SC-2, we also show results in the Minigrid domain \citep{gym_minigrid}. We use the TELLA framework \citep{lifelong-learning-systems} to orchestrate the experiments ($\sim$100 repetitions per scenario). Tasks involve an agent that must navigate to a goal while interacting with various entities in the environment. We consider 5 tasks --- SimpleCrossing, DistShift, a custom Fetch, a custom Unlock, and DoorKey --- each with two variants (differing in the size of the grid and/or number of objects) for a total of 10 POMDPs. Descriptions of these tasks can be found in \citet{gym_minigrid}. An observation is a $7\times7$ top-down view of the grid directly in front of the agent, with 3 channels containing information about the object type, object color, and object state in each tile. Rewards are sparse: $-1$ for running into obstacles and lava, $0$ for not reaching the goal, and a reward in $[0, 1]$ for reaching the goal proportional to the number of steps taken. There are only six actions: turn left, turn right, move forward, pick up an object, drop an object, and interact with an object. STEs are trained on each task for 1M environment steps using PPO \citep{schulman2017proximal}. We consider the pairwise and condensed scenarios where the LB lengths vary between 100k and 700k (Appendix~\ref{sec:minigrid_task_lengths}) based on examining the smallest number of steps for the STE to converge to $80\%$ of its peak performance averaged over ten runs. Each EB consists of 100 episodes per task. Further details can be found in Appendix~\ref{sec:imp_minigrid}.
\textbf{Baseline Agents}:
We compare against several non-lifelong baseline agents: 1) Random policy; 2) \emph{Baseline agent} using off-the-shelf deep RL trained on tasks in sequence without any lifelong learning. We used the VTrace algorithm \citep{espeholt2018impala} for SC-2 and PPO \citep{schulman2017proximal} for the Minigrid experiments.
\textbf{Ablations}: In order to compare different replay type, we perform ablations against the combination of buffers used for replay (see Fig. \ref{fig:all_replay_archs}): 1) \emph{Hidden Replay (ER)}, 2) \emph{Hidden Replay (ER+RaR)}, 3) \emph{Hidden Replay (ER+GR)}, 4) \emph{Hidden Replay (ER+RaR+GR)}. In each case, the VAE, sleep policy, and wake policy are trained, but on different augmented datasets as shown in Alg.~\ref{alg:hidden_replay} (samples are drawn from VAE only if GR is enabled).
Similarly, we perform ablations along architectures discussed in Fig.~\ref{fig:all_replay_archs}.
\textbf{Single Task Experiments}:
We conducted experiments that demonstrated that the wake-sleep training phases successfully imitates the STE. These can be found in the Appendix \ref{sec:ste_additional_results} for SC-2 and Minigrid.
\subsection{Evidence of Overcoming Catastrophic Forgetting}
We observed that different syllabi and scenarios can exhibit different levels of catastrophic forgetting due to varying degrees of similarity between the tasks in the syllabus. We first illustrate this issue before showing large scale results averaging over all syllabi.\footnote{We attempt to quantify pairwise similarity between tasks in Appendix \ref{sec:imp_sc2} by evaluating the forward transfer of the STE for each task to all other tasks. Quantifying task similarity remains an open problem.}
In Figure \ref{fig:dissimilar_alternating}, we illustrate examples that
1) the LL agent and non-LL agent perform similarly when syllabi are composed of similar tasks, and both agents exhibit positive FT and BT, and 2) the non-LL agent experiences catastrophic forgetting (zig-zag pattern) of one or both tasks whereas the LL agent shows significantly less forgetting \emph{when transitioning between two dissimilar tasks}.
We note that in the \SC{} domain, evidence of extreme catastrophic forgetting is somewhat rare, a possible weakness in the design of our task set. Nevertheless, LL agents clearly outperform non-LL on almost all the other metrics, signifying the need to look beyond catastrophic forgetting when evaluating LL.
\subsection{Comparison of Different Types of Replay for Sleep}
\begin{figure*}[t]
\centering
\includegraphics[width=.98\textwidth]{graphics/minigrid_replay.png}
\caption{Evaluating different replay types on the hidden replay model for the Minigrid domain. Error bars denote 95\% confidence interval.}
\label{fig:minigrid_replay_ablations_main}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=0.98\textwidth]{graphics/sc2_replay.png}
\caption{Evaluating different replay types on the hidden replay model for the \SC{} domain. Error bars denote 95\% confidence interval.}
\label{fig:sc2_replay_ablations_main}
\end{figure*}
To understand the effect of different replay types on the hidden replay model: experience replay from the wake buffer (ER), generative replay (GR), and random replay (RaR). All approaches still train both the policy network and generator, even if GR is disabled. We summarize the important findings in Fig.~\ref{fig:minigrid_replay_ablations_main} and Fig.~\ref{fig:sc2_replay_ablations_main} \footnote{Further experiments can be found in Appendix \ref{app:replay_ablation}}.
\textbf{Transfer Learning and Generalization to Unseen Tasks}:
Looking at the pairwise scenario in Minigrid and \SC{}, we see that models that use hidden replay with ER+GR, ER+RaR or ER+RaR+GR show much higher FT compared to the baseline. Hidden replay with ER+GR, ER+RaR or ER+RaR+GR show higher $RR_\upsilon$ and the effect is more pronounced for \SC{} condensed scenario, showing generalization to unseen tasks.
\textbf{Mitigating Catastrophic Forgetting}:
Looking at PM across all scenarios in Minigrid and \SC{} and looking at $RR_\Omega$ of the Minigrid condensed scenario, we validate that the hidden replay model requires stabilization with RaR in order to (significantly) reduce catastrophic forgetting, compared to ER+GR alone which is typically worse than ER+RaR and ER+RaR+GR.
\textbf{Adaptation to New Tasks}: Interestingly, we find that $RR_\alpha$ for the condensed scenarios of Minigrid and \SC{}, shows that simply performing distillation during sleep phases produces
a beneficial ``jumpstart'' effect when learning new tasks, compared to the baseline which is also able to learn the new task but requires more learning experiences.
\textbf{Backward Transfer}: The BT metric seems to vary across the board based on the type of replay used. This is in part because when computing the BT metric from task 2 to task 1, it is very sensitive to how well the model learns on task 1. If the model learns task 1 to near optimal performance, it is very difficult to achieve $BT>1$ whereas if task 1 is not well-learned, even a weak model on a similar task 2 can potentially easily achieve $BT>1$. However, if we look at $RR_\sigma$ instead, we see a different story where the models that learn best (ER+RaR and ER+RaR+GR) often exhibit the best performance on seen tasks in opposition to the BT metric.
\textbf{General Observations}:
Adding a small amount of random replay (ER+RaR+GR) fixes many issues with standard generative hidden replay, producing noticeably better results. We find that ER+RaR, even without GR, achieves significant improvement over the baseline agent, and the model further improves when all types of replay (ER+RaR+GR) are combined, albeit by a small amount. This suggests that RaR of a small number of samples is often powerful and
complementary (i.e., often helpful, rarely hurtful) to GR (n.b.\! all methods train the VAE, but only GR samples from it). This leads us to suspect that the replay architecture is more important than the sampling strategy that we investigate next.
\subsection{Comparison of Different Architectures for Sleep}
\begin{figure*}[t]
\centering
\includegraphics[width=0.98\textwidth]{graphics/minigrid_archs.png}
\caption{Evaluating the performance of different architectures for the Minigrid domain. Error bars denote 95\% confidence interval.}
\label{fig:minigrid_ablations_archs_main}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=0.98\textwidth]{graphics/sc2_archs.png}
\caption{Evaluating the performance of different architectures for the \SC{} domain. Error bars denote 95\% confidence interval.}
\label{fig:sc2_ablations_archs_main}
\vspace{-0.5cm}
\end{figure*}
We aim to understand the strengths and weaknesses of the model architectures introduced in this paper (Sequential, Two-Headed, Hidden Replay) and understand how these models improve over simple baselines.
We summarize the important findings in Fig.~\ref{fig:minigrid_ablations_archs_main} and Fig.~\ref{fig:sc2_ablations_archs_main} \footnote{Further experiments can be found in Appendix \ref{app:archs_ablation})}.
\textbf{Transfer Learning and Generalization to Unseen Tasks}: Forward transfer seems to be a strength of the hidden replay model. Looking at FT across all domains and scenarios and looking at the $RR_\upsilon$ of the \SC{} scenarios, we see that the hidden replay model noticeably outperforms the other models.
\textbf{Avoiding Catastrophic Forgetting}:
Looking at PM and $RR_\Omega$ across all scenarios, the hidden replay model often outperforms the other models.
\textbf{Adaptation to New Tasks}: We see that the hidden replay model achieves consistently high (and often the best) RP and $RR_\alpha$ across most scenarios in both domains. This suggests that the hidden replay model is not interfering with new tasks and may be providing a helpful ``jumpstart'' effect.
\textbf{Positive Backward Transfer}: As before, it can be difficult to interpret the BT metric. In general, the BT of the hidden replay model is on par with other models that achieve comparable terminal performance, and the hidden replay model shows strong performance in terms of $RR_\sigma$.
\textbf{Sample Efficiency}
Compared to STE, our experiments using the hidden replay model required only 6\% of training samples to achieve 80-90\% of expert performance in most \SC{} scenarios.
\textbf{General Observations}: In general, across both domains and many scenarios, the hidden replay model seems to outperform the other model architectures tested in most of the metrics of interest. This suggests that the hidden replay model is a potentially powerful model for addressing lifelong RL. We observe (in the case of the Minigrid experiments where all types of replay are used for all architectures), architecture seems to play a more significant role in improving lifelong learning performance than the type of replay.
\subsubsection*{Acknowledgments}
This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under the Lifelong Learning Machines (L2M) program Contract No. HR0011-18-C-0051. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the Defense Advanced Research Projects Agency (DARPA).
Special thanks to Roberto Corizzo, Kamil Faber, Nathalie Japkowicz, Michael Baron, James Smith, Sahana Pramod Joshi, Zsolt Kira, Cameron Ethan Taylor, Mustafa Burak Gurbuz, Constantine Dovrolis, Tyler L. Hayes, Christopher Kanan, and Jhair Gallardo for their useful discussions and feedback related to system design.
\section{Background}
\subsection{Lifelong Reinforcement Learning (LRL) Setup}
A lifelong RL \emph{syllabus} is a sequence of \emph{tasks} that the learning agent experiences one at a time for a fixed number of interactions.
Each task is assumed to be formulated as a Partially Observable Markov Decision Processes (POMDPs) \citep{sutton2018reinforcement}, where
each POMDP is defined by a tuple $(\mathcal{S}, \mathcal{A}, P(s'|s,a), R(s, a), \mathcal{O}, Z(o|s))$ consisting of set of states $\mathcal{S}$, a set of actions $\mathcal{A}$, a transition probability function between states $P(s'|s,a)$, a reward function $R(s, a)$, a set of observations $\mathcal{O}$, and conditional observation probability function $Z(o|s)$.
We assume that all tasks have distinct observation spaces of the same dimensionality, share a common action space (e.g., point-and-click, issue movement or attack commands), but differ in the transition and reward dynamics.
For example, this setup can capture the use case of an embodied agent using the same set of sensors and actuators throughout their lifetime but with different reward functions or objectives.
In single-task RL, the objective is to learn a policy that maps the history of observations to actions, $\pi: \mathcal{O}^h \rightarrow \mathcal{A}$, such that the policy maximizes the expected discounted future reward, given by the value function $V^{\pi}(s) = E_{S_t \sim P, A_t \sim \pi}[\sum_{t = 1}^{\infty} \gamma^{t-1} R(S_t, A_t)]$. We will refer to the task-optimal policy as a \emph{single task expert} (STE).
Lifelong Learning involves training an agent to continually learn from a stream of tasks over long lifetimes.
As the agent encounters new tasks, it must accumulate and leverage knowledge to learn novel tasks faster (positive forward transfer), while maintaining performance on previous tasks (minimizing catastrophic forgetting or negative backwards transfer).
In this paper, we consider the task of lifelong reinforcement learning (LRL), which has been less explored than class-incremental classification \citet{hayes2021replay} in the literature.
The objective in LRL is to learn a policy that maximizes average cumulative reward over all POMDPs in a task set while learning from limited interactions with a sequence of POMDPs drawn from the task set.
This is challenging because the learning problem is non-stationary and the learner needs to accommodate an increasing number of tasks in a model with fixed capacity.
The task changes after an unknown number of interactions with the environment.
The agent has no control over the choice of the next task.
In our setting the agent has no knowledge of the identity of the current task, or the transition points between tasks.
\subsection{Generative Replay (GR)}
Our work is focused on generative replay (GR), which can be broadly described as a form of data augmentation where real-world experiences are used to continually train a generative model over the input space of the problem. Subsequently, data sampled from this generator can be pseudo-labeled or self-labeled to generate paired training examples that may reflect previously trained tasks.
Different types of replay were discussed by \citet{hayes2021replay} along with its role in the brain and in current deep learning implementations. Most of the prior work on generative replay has focused on the problem of class-incremental classification \citet{van_de_ven_three_2019,shin2017continual}, or class-incremental generative modelling \citet{cong2020gan} as a subroutine thereof, but there has been significantly less exploration of GR in LRL.
We use a Variational Auto-Encoder (VAE) \citep{kingma_auto-encoding_2014} as the generative model. The VAE consists of an encoder $q_\phi(z|o)$ mapping observations to a vector-valued latent space $z$, and a decoder $p_\theta(o|z)$ that is trained to reconstruct the original input. We use the Gaussian prior $z \sim N(0, \mathbf{I})$ with 0 mean and unit variance $\mathbf{I}$. To train the VAE, we minimize the standard loss $\mathcal{L}_\text{VAE}$,
\begin{equation}
\label{loss:vae}
\min E_{q_{\phi}(z|o)}[\ln p_{\theta}(o|z)] - D_{KL}[q_{\phi}(z|o)||p(z)]
\end{equation}
World models \citep{ketz2019continual} use model-based generative replay for LRL i.e., the replay generates rollouts or entire trajectories \citep[e.g.][]{ketz_using_2019}. Model-based replay works well in many domains, but it is challenging for domains such as SC-2 due to the complex dynamics and high dimensional observation and action spaces.
The most relevant prior work is that of \citeauthor{atkinson2021pseudo} \citep{atkinson2021pseudo} who explored model-free generative replay for Deep Q-Learning (DQN).
However, \citet{atkinson2021pseudo} uses GANs in the raw perceptual space of the agent (aka veridical replay \citet{hayes2021replay}). Others have explored sophisticated strategies for prioritized experience replay in DQN \citep[e.g.][]{isele2018selective, riemer2019scalable,wang2021acae,caccia2020online,zhang2017deeper,schaul2015prioritized} in LRL, but these are not generative in nature.
We show that na\"ively migrating the generative modelling (similar to \citeauthor{atkinson2021pseudo}) to the feature space of the policy causes instabilities and severe catastrophic forgetting.
To the best of our knowledge, this is the first work on model-free GR for LRL in the feature space of generic deep RL agents.
\section{Appendix}
\subsection{Potential Negative and Positive Societal Impacts}
This approach presents a means of training a multi-agent system to play real-time strategy games using lifelong reinforcement learning. Such a tool could be used by malicious foreign actors to train models for strategic military actions, which could negatively impact the national security of benign nations.
However, this tool can also be used for societal benefits including training intelligent robot swarms to perform useful tasks over long lifetimes (e.g., applications in agricultural and manufacturing robotics). Similarly, it could be used in domains which require incorporation of feedback into the learning (i.e. necessitates RL) and where distribution shift is a significant problem; e.g., for planning treatments in healthcare.
\subsection{Limitations}
The model is not flawless. The following are some limitations we have observed w.r.t. our proposed approach:
\begin{itemize}
\item While catastrophic forgetting is reduced using our proposed models, the agent still fails to perfectly learn and remember tasks, and thus, there is still room for improvement in terms of many of the LL metrics. Specifically, we've observed that the model can struggle to remember over very long periods when trained on long syllabi of tasks.
\item Training the agent is non-trivial, requiring carefully defining reasonable tasks and performing hyperparameter tuning.
\item The sleep model is dependent on having a good quality single-task expert for the wake model, and in some cases, the wake model fails to properly learn a task, disrupting the learning of the sleep model.
\item We have also seen some fragility in the learning of the hidden space for hidden replay and imperfect generated samples in the sequential and two-headed models. We have observed cases where the sampled data drifts enough that it mis-aligns with the true distribution of data and/or the quality of the sampled data degrades significantly. In such cases, the sleep model can become unstable and the proposed replay mechanisms may hurt the model's learning.
\item While we would like to rely mostly on generative replay and experience replay from the last wake cycle for training the model, we've found that the hidden replay model requires a small buffer of random samples across tasks to maintain the hidden space. This introduces sub-optimality in the memory usage of the model. We have also observed that this ``random replay'' has an out-sized effect on improving lifelong learning compared to the generative replay.
\end{itemize}
\subsection{Starcraft 2 Task Variants}
\label{sec:sc2_task_descriptions}
The task variants we consider are as follows:
\begin{itemize}
\item \textbf{Collect Mineral Shards – No Fog of War}: A map with 2 Marines and an endless supply of Mineral Shards. Rewards are earned by moving the Marines to collect the Mineral Shards, with optimal collection requiring both Marine units to be split up and moved independently. Whenever all 20 Mineral Shards have been collected, a new set of 20 Mineral Shards are spawned at random locations (at least 2 units away from all Marines). Fog of war is disabled.
\item \textbf{Collect Mineral Shards – Fog of war:} A map with 2 Marines and an endless supply of Mineral Shards. Rewards are earned by moving the Marines to collect the Mineral Shards, with optimal collection requiring both Marine units to be split up and moved independently. Whenever all 20 Mineral Shards have been collected, a new set of 20 Mineral Shards are spawned at random locations (at least 2 units away from all Marines). Fog of war is enabled, meaning the agent must be able to learn without full knowledge of the current state of the environment.
\item \textbf{DefeatZerglingsAndBanelings – One Group:} A map with 9 Marines on the opposite side from a group of 6 Zerglings and 4 Banelings. Rewards are earned by using the Marines to defeat Zerglings and Banelings. Whenever all Zerglings and Banelings have been defeated, a new group of 6 Zerglings and 4 Banelings is spawned and the player is awarded 4 additional Marines at full health, with all other surviving Marines retaining their existing health (no restore). Whenever new units are spawned, all unit positions are reset to opposite sides of the map
\item \textbf{DefeatZerglingsAndBanelings – Two Groups:} A map with 9 Marines in the center with two groups consisting of 9 Zerglings one one side and 6 Banelings on the other side. Rewards are earned by using the Marines to defeat Zerglings and Banelings. Whenever a group has been defeated, a new group of 9 Zerglings and 6 Banelings is spawned and the player is awarded 6 additional Marines at full health, with all other surviving Marines retaining their existing health (no restore). Whenever new units are spawned, all unit positions are reset to opposite sides of the map.
\item \textbf{DefeatRoaches – One Group:} A map with 9 Marines and a group of 4 Roaches on opposite sides. Rewards are earned by using the Marines to defeat Roaches, with optimal combat strategy requiring the Marines to perform focus fire on the Roaches. Whenever all 4 Roaches have been defeated, a new group of 4 Roaches is spawned and the player is awarded 5 additional Marines at full health, with all other surviving Marines retaining their existing health (no restore). Whenever new units are spawned, all unit positions are reset to opposite sides of the map.
\item \textbf{DefeatRoaches – Two Groups:} A map with 9 Marines in the center and two groups consisting of 6 total Roaches on opposite sides (3 on each side). Rewards are earned by using the Marines to defeat Roaches, with optimal combat strategy requiring the Marines to perform focus fire on the Roaches. Whenever all 6 Roaches have been defeated, a new group of 6 Roaches is spawned and the player is awarded 7 additional Marines at full health, with all other surviving Marines retaining their existing health (no restore). Whenever new units are spawned, all unit positions are reset to starting areas of the map.
\end{itemize}
\subsection{Implementation Details}
\label{exp:archs}
\subsubsection{Starcraft 2}
\label{sec:imp_sc2}
In this section we describe the implementation details for our \SC{} agent and experiments. The wake agent is a VTrace agent \citep{espeholt2018impala}. The feature extractor component of the wake agent consists of an architecture that is a slight modification of the FullyConv model of \citet{vinyals2017starcraft}. We further split this into a feature extractor component and a policy network component to facilitate easy weight copying between the wake and sleep agents. For the hidden replay model, we add a LayerNormalization layer to the output of the feature extractor to constrain the space of the features, which helps to improve feature reconstructions and limit feature drift. The architecture produces features of dimension 8192.
The sleep agent feature extractor and policy network architectures are identical to the wake model. We use a five-layer convolutional neural network architecture with for the encoder and a five-layer deconvolutional architecture for the decoder for the VAEs for the sequential and two-headed models. In both cases, we use ReLU activations between the layers and add batch normalization before the ReLUs. We use two-layered multi-layered perceptrons with ReLU activation functions for the encoder and decoder for the VAE for the hidden replay model. To improve numerical stability, the logvar predicted by the encoder is bounded to take values in $[-5, 5]$ using a scaled-hyperbolic tangent layer.
Three styles of replay are used during sleep for the hidden replay architecture: experience replay on a buffer of 10,000 observation-action pairs from a FIFO queue collected during the wake phase, generative replay on the feature vectors, and random replay. Only experience replay and generative replay is used to train the sleep agent for the sequential and two-headed architectures.
For the sequential architecture: the weight of the imitation loss (cross-entropy on action logits) is 50.0, the weight of the reconstruction loss is 1.0, and the weight of the KL loss of the VAE is 3.0. For the two-headed architecture: the weight of the imitation loss (cross-entropy on action logits) is 30.0, the weight of the reconstruction loss is 1.0, and the weight of the KL loss of the VAE is 2.0. For the hidden replay architecture: the weight of the imitation loss (cross-entropy on action logits) is 50.0, the weight of the reconstruction loss is 200.0, and the weight of the KL loss of the VAE is 5.0. These were determined via hand-tuning.
The agent enters sleep three times per task at even intervals for the sequential and two-headed architectures. The agent enters sleep twice per task at even intervals for the hidden replay architecture. Each training iteration of the wake model is trained on 32 trajectories. Each sleep consists of 4,500 iterations of training using batch sizes of 64 (64 wake samples, 64 generated samples, and 64 random observations from the random replay buffer). Generative replay is not used until the second sleep. 96 observations are selected randomly from the wake buffer every sleep to be added to the random replay buffer for the hidden replay architecture. When the agent wakes up, it resets the PPO model and offers advice with a 80\% probability linearly decaying to 0\% after for one-half the the duration between sleeps. No weight copying is used for the sequential variant, but weight copying (running a mini-evaluation block after sleep and copying the best Eigentask's parameters into the wake agent) and advice are both used for the two-headed and hidden replay variants.
The Adam optimizer is used to train both the wake and sleep agents with a learning rate of 1.0e-3.
\textbf{Evaluation:} We selected 3 minigames involving battles (\emph{DefeatRoaches} and \emph{DefeatZerglingsAndBanelings} tasks) or resource collection (\emph{CollectMineralShards}). We created two variants of each task, differentiated by the starting locations of units, the presence of ``fog-of-war'', and/or the number of friendly and enemy units. Task variants are described in Section~\ref{sec:sc2_task_descriptions}. We used PySC2 \citep{vinyals2017starcraft} to interface with SC-2. We used a subset of the available observation maps namely the unit type, selection status, and unit density two-dimensional observations. The action space is factored into functions and arguments, such as $\text{move}(x, y)$ or $\text{stop}()$. Following \citet{vinyals2017starcraft}, our policy networks output probabilities for the different action factors independently.
The agent receives positive rewards for collecting resources and defeating enemy units, and negative rewards for losing friendly units. Single task experts (STEs) were trained to convergence ($\sim$ 30 million environment steps using the VTrace algorithm \citep{espeholt2018impala}). We consider the pairwise, alternating, and condensed scenarios where each LB consists of 2 million environment steps (about 6\% of the STE samples), and each EB consists of 30 episodes per task.
Each lifetime of a condensed scenario takes approximately 3-4 days running on systems with RTX2080 GPUs.
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/m18_pi_similarity_matrix.png}
\caption{Similarity between tasks measured in terms of forward transfer using single task experts}
\label{fig:similarity_matrix}
\end{figure}
\subsubsection{Minigrid}
\label{sec:imp_minigrid}
In this section, we describe the implementation details for our Minigrid agent and experiments. The wake agent is based on the Stable Baselines 3 PPO agent \citep{schulman2017proximal,stable-baselines3}. There are separate networks for computing the value and policy. Both networks are simple multi-layered perceptrons that first map to a 256-dimensional feature space and then apply a densely connected layer to select one of six actions or estimate the value of a state. ReLUs are used as activation functions between all of the layers. Observations are scaled so the values of every cell are in $[0, 1]$ before being input to the network. For exact parameters for the wake agent, please refer to the appendix.
The sleep agent uses one Eigentask. The feature extractor and expected inputs are the same as those of the wake agent's policy network. A variational autoencoder is used as generative model. The VAE is a four-layer bottleneck-style MLP architecture with a latent dimension size: 128. To improve numerical stability, the logvar predicted by the encoder is bounded to take values in $[-5, 5]$.
Three styles of replay are used during sleep: experience replay on a buffer of 20,000 observation-action pairs from a FIFO queue collected during the wake phase, generative replay on the feature vectors, and random replay (256 random observations added per sleep). The weight of the imitation loss (cross-entropy on discrete actions) is 3.0, the weight of the reconstruction loss is 1.0, and the weight of the KL loss of the VAE is 0.03. The agent enters sleep at the end of each task. Each sleep consists of 20,000 iterations of training using batch sizes of 32 (32 wake samples, 32 generated samples, and 32 random examples drawn from the random replay buffer). Generative replay and random replay are not used until the second sleep. 256 observations are selected randomly from the wake buffer every sleep to be added to the random replay buffer with a maximum buffer size of 4096. When the agent wakes up, it resets the PPO model and offers advice with a 90\% probability linearly decaying to 0\% after 100,000 environment steps for the condensed scenario and 33,333 environment steps for the dispersed scenario. No weight copying is used.
The Adam optimizer is used to train both the wake and sleep agents.
\textbf{Evaluation:} We report results using the Minigrid domain \citep{gym_minigrid}, in which an agent must navigate to a goal while interacting with various entities in the environment. We consider five tasks --- SimpleCrossing, DistShift, a custom Fetch, a custom Unlock, and DoorKey --- each with two variants (differing in the size of the grid and/or number of objects) for a total of ten POMDPs. Descriptions of these tasks can be found in \citet{gym_minigrid}. An observation consists of a $7\times7$ top-down view of the grid directly in front of the agent, with three channels containing information about the object type, object color, and object state in each tile. Rewards are sparse and consist of $-1$ for running into obstacles and lava, $0$ for not reaching the goal, and a positive reward in $[0, 1]$ for reaching the goal that is proportional to the number of steps taken. Unlike SC-2, there are only six actions: turn left, turn right, move forward, pick up an object, drop an object, and interact with an object.
STEs are trained on each task for 1M environment steps using PPO \citep{schulman2017proximal}. We consider the pairwise, alternating, and condensed scenarios where the LB lengths vary between 100k and 700k (Section~\ref{sec:minigrid_task_lengths}) based on examining the smallest number of steps for the STE to converge to $80\%$ of its peak performance averaged over ten runs. Each EB consists of 100 episodes per task.
Each lifetime of a condensed scenario takes approximately 4-6 hours to run on systems with RTX2080 GPUs.
\subsection{Minigrid Task Lengths}
\label{sec:minigrid_task_lengths}
We use the following number of environment steps per task in our minigrid experiments:
\begin{itemize}
\item SimpleCrossingS9N1: 400,000
\item SimpleCrossingS9N2: 500,000
\item DistShiftR2: 200,000
\item DistShiftR3: 200,000
\item CustomFetchS5T1N2: 500,000
\item CustomFetchS8T1N2: 700,000
\item CustomUnlockS5: 200,000
\item CustomUnlockS7: 300,000
\item DoorKeyS5: 200,000
\item DoorKeyS6: 300,000
\end{itemize}
\subsection{Parameters of the PPO agent}
These are the parameters used for the stable baselines 3 PPO agent for the Minigrid tasks:
\begin{itemize}
\item n\_steps: 512
\item batch\_size: 32
\item gae\_lambda: 0.95
\item gamma: 0.99
\item n\_epochs: 10
\item ent\_coef: 5.0e-5
\item learning\_rate: 2.5e-4
\item clip\_range: 0.3
\item vf\_coef: 0.75
\item max\_grad\_norm: 5.0
\item policy\_kwargs: dict(activation\_fn=nn.ReLU, net\_arch=[dict(pi=[256, 256], vf=[256, 256])])
\end{itemize}
\subsection{Evaluation Scenarios Visualization}
\begin{figure}[H]
\centering
\includegraphics[width=0.75\textwidth, keepaspectratio]{graphics/scenarios_2.png}
\caption{Lifelong Learning Scenarios. Grey evaluation blocks (EBs) evaluate the agent on all tasks. Learning blocks (LBs) are color coded by task, and subscripts denote task variants.}
\label{fig:scenarios}
\end{figure}
\subsection{Example Learning Curves for Lifelong Curricula on \SC{}}
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/sc2_learning_curve_cropped.png}
\caption{Learning curve for our lifelong learning agent applied to a six minigame scenario in the \SC{} domain. Performance comparable to the converged single-task expert is \textasciitilde 100.}
\label{fig:sc2_learning_curves}
\end{figure}
\subsection{Comparison of Learning Curves for Lifelong Curricula on Minigrid}
\begin{figure}[H]
\centering
\includegraphics[width=0.45\textwidth]{graphics/minigridcondensed_non-ll.png}
\includegraphics[width=0.45\textwidth]{graphics/minigridcondensed_ll.png}
\caption{Learning curves for a condensed scenario in the Minigrid domain. Left: learning curve of non-LL agent; Right: learning curve of LL agent on the same scenario. Maximum reward is 100. The LL agent is capable of learning more tasks without interference and exhibits jump starts in performance for similar tasks due to the advice mechanism.}
\label{fig:minigrid_learning_curves}
\end{figure}
\subsection{A Sleep Policy Trained with Policy Distillation can Learn to Imitate a Converged Single Task Expert}
\label{sec:ste_additional_results}
\textbf{Single Task Experiments}:
\textbf{SC-2:} We demonstrate that a sleep policy can be trained with policy distillation to mimic the performance of an STE's policy. We consider STEs identical to the wake agents trained for 30M environment steps for each of the six SC-2 variants. We act in accordance with the following procedure: i) Train an STE on each of the six SC-2 POMDPs for 30M steps. ii) Measure the final reward for each STE. iii) Load the trained STE and build a buffer of 10K trajectories. iv) Train the sleep model until convergence on randomly sampled batches of observation-action pairs (batch size: 32). v) Periodically evaluate the sleep model on SC-2 simulations, each consisting of 30 episodes of the task-of-interest.
We show results in Figs.~\ref{fig:sleep_imitating_ste} and \ref{fig:sleep_imitating_ste_app}. The two-headed and hidden replay architectures architectures typically match or exceed the STE performance in terms of rewards. The sequential model learns non-trivial policies, but generally performs worse than the STE. This supports our hypothesis that decoupling the generator from the policy network of the VAE results in better-performing policies.
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/sleep_imitating_ste_combined.png}
\caption{Examining whether sleep policies can be trained to mimic the performance of single task experts for the SC-2 POMDPs with the sequential, two-headed, and hidden replay model without a random replay buffer. The blue curves denote the performance (in terms of rewards) of the sleep policy interpolated along evaluation points. The red line denotes terminal performance of the STE after convergence. Error bars denote standard deviation of rewards.}
\label{fig:sleep_imitating_ste}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/sleep_imitating_ste.png}
\caption{Examining whether the sleep policy can be trained to mimic the performance of single task experts for the SC-2 POMDPs with the hidden replay model without a random replay buffer. The blue curves denote the performance (in terms of rewards) of the sleep policy interpolated along evaluation points. The red line denotes terminal performance of the STE after convergence. Error bars denote standard deviation of rewards.}
\label{fig:sleep_imitating_ste_app}
\end{figure}
\textbf{Minigrid:} We also quantitatively validate that the sleep policy using the hidden replay model is able to distill the an STE's converged policy after 1M environment steps for the Minigrid domain using an experience buffer of the STE's last 20K observation-action pairs (repeated 10 times per task). We find that the average difference in rewards between the STE and sleep policy's performance over all task variants is 0.14 with a standard deviation of 0.11, suggesting that the sleep policy imitates the STE to an acceptable level.
\subsection{A Sleep Policy Trained with Policy Distillation can Learn to Imitate the Wake Policy on a Single Task}
\label{sec:wake_additional_results}
We conduct experiments to show that the converged performance of a wake policy can be achieved by our agent via imitating the sequence of intermediate wake policies over multiple wake-sleep phases. This is an important experiment needed to show stable convergence of the wake-sleep mechanism even when distilling sequences of sub-optimal policies. These experiments involve training an agent on a single task, periodically entering sleep at set intervals. We observed that for most tasks, the curves formed by evaluating the sleep agent and wake agents at periodic evaluation blocks closely track one another, demonstrating stable convergence of the sleep agent's policy.
\textbf{SC-2:} We consider the case where the system performs multiple wake-sleep phases over the course of a single task, requiring the sleep policy to consolidate the increasingly improving policies of the wake agent. Using the SC-2 simulator, each wake policy is trained for 6M environment steps, entering sleep at even intervals ten times over the entire run. Every time the system exits sleep, the wake and sleep policies are evaluated on thirty episodes of the task. We show the wake versus sleep performance (in terms of rewards) of a hidden replay-based sleep policy in Fig.\ \ref{fig:sleep_imitating_wake_app}. The sleep policy's performance typically closely mirrors the wake policy's performance, but there are cases where the agent learns a non-trivial policy, but under-performs the wake policy. While not pictured here, similar behavior was observed for the sequential and two-headed architectures as well.
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/sleep_imitating_wake.png}
\caption{Examining whether the sleep policy can be trained to mimic the performance of wake non-lifelong learner for the SC-2 POMDPs as the non-lifelong learners improve. This experiment examined the performance of the hidden replay model without using observations from the random replay buffer. The blue curves denote the performance (in terms of rewards) of the sleep policy interpolated along evaluation points. The red line denotes performance of the wake policy interpolated along evaluation points. Error bars denote standard deviation of rewards.}
\label{fig:sleep_imitating_wake_app}
\end{figure}
\textbf{Minigrid:} We also conducted a post-hoc analysis of our condensed scenario runs in the Minigrid domain to quantify the sleep policy's capability to imitate the wake policy after learning the first task.
\textit{\textbf{Sequential Model:}} The mean absolute difference between the wake policy's average reward for the last 100 episodes and the sleep policy's average reward for the evaluation block was 0.589 with a standard deviation of 0.320, suggesting the sleep policy when using the sequential architecture struggles to accurately imitate the wake policy in many, but not all runs. In cases where the wake model converges to a policy with average reward $>0.9$ for the last 100 episodes, the mean difference metric decreases to $0.591 \pm 0.384$.
\textit{\textbf{Two-Headed Model:}} The mean absolute difference between the wake policy's average reward for the last 100 episodes and the sleep policy's average reward for the evaluation block was 0.422 with a standard deviation of 0.263, suggesting the sleep policy when using the two-headed model can struggle to accurately imitate the wake policy in some runs. In cases where the wake model converges to a policy with average reward $>0.9$ for the last 100 episodes, the mean difference metric decreases to $0.341 \pm 0.297$.
\textit{\textbf{Hidden Replay Model:}} The mean absolute difference between the wake policy's average reward for the last 100 episodes and the sleep policy's average reward for the evaluation block was 0.209 with a standard deviation of 0.227, suggesting the sleep policy accurately imitates the wake policy in many, but not all runs. In cases where the wake model converges to a policy with average reward $>0.9$ for the last 100 episodes, the mean difference metric decreases to $0.110 \pm 0.145$.
These results suggest that the hidden replay model has an advantage when it comes to imitating a wake model when learning a task from scratch. The sequential model struggles with imitating the wake model when learning a task from scratch; however, our main results in the paper suggest that as the sequential model encounters additional tasks, this issue of successfully imitating the wake policy is somewhat alleviated as the relative reward post learning block metric ultimately is similar to the two-headed model when deployed over an entire syllabus of tasks.
\subsection{Effect of Varying the Number of Wake-Sleep Cycles}
we consider the case where the system encounters multiple tasks in sequence, and we evaluate the effect of the number of wake-sleep phases per task on the lifelong learning metrics. We trained the two-headed replay architecture for 4 lifetimes, each consisting of alternating between an \SC{} combat and collection task with 2M environment steps per task. We compared the effect of one, two, and three sleeps per task and report results in Table \ref{tab:multi-sleep}. Results suggest that multiple sleep phases result in improved or maintained performance in most metrics, but may negatively affect FT.
\begin{table}[H]
\centering
\resizebox{1.0\textwidth}{!}{
\begin{tabular}{|c|c|c|c|c||c|c|c|c|c|}
\hline
Agent & PM & FT & BT & RP & RR\textsubscript{$\Omega$} & RR\textsubscript{$\alpha$} & RR\textsubscript{$\sigma$} & RR\textsubscript{$\upsilon$} \\
\hline
One Sleep Cycle Per Task & $-8.76$ ($ \pm 4.39$) & $1.68$ ($ \pm 2.16$) & $0.76$ ($ \pm 0.06$) & $0.78$ ($ \pm 0.22$) & $0.58$ ($ \pm 0.29$) & $0.7$ ($ \pm 0.22$) & $0.58$ ($ \pm 0.22$) & $0.3$ ($ \pm 0.38$)\\
Two Sleep Cycles Per Task & $-10.47$ ($ \pm 6.41$) & $1.16$ ($ \pm 0.67$) & $0.93$ ($ \pm 0.11$) & $0.87$ ($ \pm 0.11$) & $0.79$ ($ \pm 0.47$) & $0.95$ ($ \pm 0.25$) & $0.81$ ($ \pm 0.31$) & $0.27$ ($ \pm 0.31$)\\
Three Sleep Cycles Per Task & $-8.63$ ($ \pm 7.13$) & $1.06$ ($ \pm 0.86$) & $0.78$ ($ \pm 0.25$) & $0.84$ ($ \pm 0.13$) & $0.67$ ($ \pm 0.19$) & $0.7$ ($ \pm 0.13$) & $0.61$ ($ \pm 0.1$) & $0.19$ ($ \pm 0.19$)\\
\hline
\end{tabular}
}
\caption{Comparing the effect of sleeping multiple times per task on a subset of four alternating scenarios involving transitioning between combat and collection \SC{} tasks. We see in many cases that sleeping multiple times per task can improve metrics.}
\label{tab:multi-sleep}
\end{table}
\subsection{Understanding the Effect of Different Replay Types}
\label{app:replay_ablation}
In Tables \ref{tab:replay_mechanism_ablation_minigrid} and \ref{tab:replay_mechanism_ablation_sc2}, we investigate the effects of turning on and off different replay mechanisms in the hidden replay model. In general, we validate that combining experience replay of the wake buffer, generative replay, and random replay, results in an improved agent; although experience replay in combination with random replay performs surprisingly well.
\begin{table}[H]
\centering
\resizebox{1.0\textwidth}{!}{
\begin{tabular}{|c|c|c|c|c|c||c|c|c|c|}
\hline
Scenario & Agent & PM & FT & BT & RP & RR\textsubscript{$\Omega$} & RR\textsubscript{$\alpha$} & RR\textsubscript{$\sigma$} & RR\textsubscript{$\upsilon$} \\
\hline
& Random Policy & $0.37$ ($ \pm 0.81$) & $1.06$ ($ \pm 0.02$) & $1.13$ ($ \pm 0.18$) & $0.22$ ($ \pm 0.04$) & $0.09$ ($ \pm 0.02$) & $0.09$ ($ \pm 0.02$) & $0.09$ ($ \pm 0.02$) & $0.09$ ($ \pm 0.00$) \\
& Baseline PPO & $-48.2$ ($ \pm 12.04$) & $4.07$ ($ \pm 0.87$) & $0.46$ ($ \pm 0.12$) & $2.17$ ($ \pm 0.61$) & $0.5$ ($ \pm 0.1$) & $0.78$ ($ \pm 0.08$) & $0.64$ ($ \pm 0.08$) & $0.16$ ($ \pm 0.02$)\\
Pairwise & Hidden Replay (ER) & $-45.25$ ($ \pm 13.54$) & $6.34$ ($ \pm 1.4$) & $0.51$ ($ \pm 0.26$) & $2.75$ ($ \pm 0.73$) & $0.47$ ($ \pm 0.06$) & $0.73$ ($ \pm 0.08$) & $0.6$ ($ \pm 0.06$) & $0.11$ ($ \pm 0.02$) \\
& Hidden Replay (ER + RaR) & $-21.11$ ($ \pm 11.06$) & $5.97$ ($ \pm 1.84$) & $0.75$ ($ \pm 0.19$) & $2.22$ ($ \pm 0.91$) & $0.56$ ($ \pm 0.08$) & $0.68$ ($ \pm 0.08$) & $0.62$ ($ \pm 0.08$) & $0.11$ ($ \pm 0.02$)\\
& Hidden Replay (ER + GR) & $-62.32$ ($ \pm 25.2$) & $7.12$ ($ \pm 2.51$) & $0.31$ ($ \pm 0.29$) & $2.51$ ($ \pm 1.61$) & $0.4$ ($ \pm 0.11$) & $0.74$ ($ \pm 0.13$) & $0.57$ ($ \pm 0.11$) & $0.11$ ($ \pm 0.02$)\\
& Hidden Replay (ER + RaR + GR) & $-15.77$ ($ \pm 10.09$) & $5.87$ ($ \pm 1.24$) & $1.21$ ($ \pm 0.55$) & $2.47$ ($ \pm 0.69$) & $0.6$ ($ \pm 0.06$) & $0.69$ ($ \pm 0.08$) & $0.64$ ($ \pm 0.06$) & $0.11$ ($ \pm 0.02$) \\
\hline
& Random Policy & $-0.02$ ($ \pm 0.26$) & $1.08$ ($ \pm 0.02$) & $1.08$ ($ \pm 0.02$) & $0.19$ ($ \pm 0.00$) & $0.09$ ($ \pm 0.00$) & $0.09$ ($ \pm 0.00$) & $0.09$ ($ \pm 0.00$) & $0.09$ ($ \pm 0.00$) \\
& Baseline PPO & $-38.92$ ($ \pm 3.98$) & $4.72$ ($ \pm 0.75$) & $4.06$ ($ \pm 0.73$) & $1.44$ ($ \pm 0.2$) & $0.22$ ($ \pm 0.04$) & $0.64$ ($ \pm 0.04$) & $0.33$ ($ \pm 0.04$) & $0.15$ ($ \pm 0.02$)\\
Condensed & Hidden Replay (ER) & $-57.71$ ($ \pm 3.61$) & $6.45$ ($ \pm 0.81$) & $5.32$ ($ \pm 0.81$) & $2.04$ ($ \pm 0.1$) & $0.17$ ($ \pm 0.02$) & $0.77$ ($ \pm 0.04$) & $0.33$ ($ \pm 0.02$) & $0.12$ ($ \pm 0.02$)\\
& Hidden Replay (ER + RaR) & $-20.96$ ($ \pm 2.81$) & $5.44$ ($ \pm 0.93$) & $1.85$ ($ \pm 0.45$) & $2.11$ ($ \pm 0.1$) & $0.54$ ($ \pm 0.04$) & $0.77$ ($ \pm 0.02$) & $0.61$ ($ \pm 0.02$) & $0.16$ ($ \pm 0.02$)\\
& Hidden Replay (ER + GR) & $-52.91$ ($ \pm 7.26$) & $6.75$ ($ \pm 1.03$) & $4.93$ ($ \pm 1.8$) & $1.97$ ($ \pm 0.24$) & $0.19$ ($ \pm 0.02$) & $0.75$ ($ \pm 0.07$) & $0.33$ ($ \pm 0.02$) & $0.12$ ($ \pm 0.02$)\\
& Hidden Replay (ER + RaR + GR) & $-21.94$ ($ \pm 2.09$) & $5.93$ ($ \pm 0.79$) & $1.32$ ($ \pm 0.2$) & $2.1$ ($ \pm 0.1$) & $0.56$ ($ \pm 0.04$) & $0.8$ ($ \pm 0.02$) & $0.63$ ($ \pm 0.02$) & $0.16$ ($ \pm 0.02$)\\
\hline
\end{tabular}
}
\caption{Effect of different replay mechanisms during sleep on lifelong learning metrics on the Minigrid domain. Results are collected using the hidden replay model on thirty-six syllabi. We report means and 95\% confidence intervals. Hidden replay using only experience replay (ER) and generative replay (GR) is insufficient. Random replay (RaR) is needed to stablize the learning of the feature space, and combining all three types of replay yield the best results.}
\label{tab:replay_mechanism_ablation_minigrid}
\end{table}
\begin{table}[H]
\centering
\resizebox{1.0\textwidth}{!}{
\begin{tabular}{|c|c|c|c|c|c||c|c|c|c|}
\hline
Scenario & Agent & PM & FT & BT & RP & RR\textsubscript{$\Omega$} & RR\textsubscript{$\alpha$} & RR\textsubscript{$\sigma$} & RR\textsubscript{$\upsilon$} \\
\hline
& Random Policy & $0.0$ ($ \pm 0.00$) & $1.0$ ($ \pm 0.00$) & $1.0$ ($ \pm 0.00$) & $0.42$ ($ \pm 0.04$) & $0.32$ ($ \pm 0.09$) & $0.32$ ($ \pm 0.09$) & $0.32$ ($ \pm 0.09$) & $0.32$ ($ \pm 0.22$)\\
& Baseline VTrace & $-8.99$ ($ \pm 3.06$) & $1.11$ ($ \pm 0.23$) & $0.79$ ($ \pm 0.12$) & $0.92$ ($ \pm 0.04$) & $0.9$ ($ \pm 0.14$) & $0.91$ ($ \pm 0.12$) & $0.82$ ($ \pm 0.12$) & $0.39$ ($ \pm 0.17$)\\
& Hidden Replay(ER) & $-7.9$ ($ \pm 5.11$) & $1.14$ ($ \pm 0.4$) & $0.75$ ($ \pm 0.29$) & $0.92$ ($ \pm 0.09$) & $0.87$ ($ \pm 0.11$) & $0.9$ ($ \pm 0.11$) & $0.8$ ($ \pm 0.13$) & $0.37$ ($ \pm 0.2$)\\
Pairwise & Hidden Replay(ER $+$ RaR) & $-3.7$ ($ \pm 4.18$) & $1.75$ ($ \pm 0.66$) & $0.92$ ($ \pm 0.11$) & $1.07$ ($ \pm 0.11$) & $0.78$ ($ \pm 0.18$) & $0.81$ ($ \pm 0.18$) & $0.79$ ($ \pm 0.18$) & $0.48$ ($ \pm 0.26$)\\
& Hidden Replay(ER $+$ GR) & $-9.35$ ($ \pm 7.46$) & $1.39$ ($ \pm 0.64$) & $0.71$ ($ \pm 0.18$) & $1.05$ ($ \pm 0.09$) & $0.71$ ($ \pm 0.15$) & $0.82$ ($ \pm 0.11$) & $0.77$ ($ \pm 0.13$) & $0.44$ ($ \pm 0.31$)\\
& Hidden Replay (ER $+$ RaR $+$ GR) & $-0.57$ ($ \pm 2.93$) & $1.85$ ($ \pm 0.88$) & $0.95$ ($ \pm 0.15$) & $1.04$ ($ \pm 0.11$) & $0.77$ ($ \pm 0.18$) & $0.82$ ($ \pm 0.13$) & $0.8$ ($ \pm 0.15$) & $0.55$ ($ \pm 0.26$)\\
\hline
& Random Policy & $0.0$ ($ \pm 0.00$) & $1.0$ ($ \pm 0.00$) & $1.0$ ($ \pm 0.00$) & $0.42$ ($ \pm 0.04$) & $0.32$ ($ \pm 0.09$) & $0.32$ ($ \pm 0.09$) & $0.32$ ($ \pm 0.09$) & $0.32$ ($ \pm 0.22$)\\
& Baseline VTrace & $-8.99$ ($ \pm 3.06$) & $1.11$ ($ \pm 0.23$) & $0.79$ ($ \pm 0.12$) & $0.92$ ($ \pm 0.04$) & $0.9$ ($ \pm 0.14$) & $0.91$ ($ \pm 0.12$) & $0.82$ ($ \pm 0.12$) & $0.39$ ($ \pm 0.17$)\\
Alternating & Hidden Replay(ER) & $-7.9$ ($ \pm 5.11$) & $1.14$ ($ \pm 0.4$) & $0.75$ ($ \pm 0.29$) & $0.92$ ($ \pm 0.09$) & $0.87$ ($ \pm 0.11$) & $0.9$ ($ \pm 0.11$) & $0.8$ ($ \pm 0.13$) & $0.37$ ($ \pm 0.2$)\\
& Hidden Replay(ER $+$ RaR) & $-4.25$ ($ \pm 2.77$) & $1.75$ ($ \pm 0.66$) & $0.88$ ($ \pm 0.11$) & $0.91$ ($ \pm 0.09$) & $0.81$ ($ \pm 0.15$) & $0.83$ ($ \pm 0.13$) & $0.79$ ($ \pm 0.15$) & $0.48$ ($ \pm 0.26$)\\
& Hidden Replay(ER $+$ GR) & $-10.99$ ($ \pm 4.89$) & $1.39$ ($ \pm 0.64$) & $0.65$ ($ \pm 0.15$) & $0.87$ ($ \pm 0.09$) & $0.83$ ($ \pm 0.18$) & $0.89$ ($ \pm 0.11$) & $0.76$ ($ \pm 0.13$) & $0.44$ ($ \pm 0.31$)\\
& Hidden Replay (ER $+$ RaR $+$ GR) & $-6.13$ ($ \pm 4.64$) & $1.85$ ($ \pm 0.88$) & $0.87$ ($ \pm 0.13$) & $0.91$ ($ \pm 0.09$) & $0.86$ ($ \pm 0.13$) & $0.89$ ($ \pm 0.13$) & $0.82$ ($ \pm 0.15$) & $0.55$ ($ \pm 0.26$)\\
\hline
& Random Policy & $-0.03$ ($ \pm 0.28$) & $1.01$ ($ \pm 0.02$) & $1.01$ ($ \pm 0.02$) & $0.54$ ($ \pm 0.02$) & $0.31$ ($ \pm 0.02$) & $0.3$ ($ \pm 0.02$) & $0.31$ ($ \pm 0.00$) & $0.3$ ($ \pm 0.04$)\\
& Baseline VTrace & $-3.41$ ($ \pm 1.7$) & $1.19$ ($ \pm 0.08$) & $1.17$ ($ \pm 0.29$) & $1.14$ ($ \pm 0.06$) & $0.64$ ($ \pm 0.06$) & $0.74$ ($ \pm 0.06$) & $0.66$ ($ \pm 0.06$) & $0.49$ ($ \pm 0.06$)\\
Condensed & Hidden Replay(ER) & $-8.65$ ($ \pm 3.15$) & $1.42$ ($ \pm 0.15$) & $1.12$ ($ \pm 0.11$) & $1.12$ ($ \pm 0.09$) & $0.73$ ($ \pm 0.09$) & $0.89$ ($ \pm 0.07$) & $0.77$ ($ \pm 0.04$) & $0.55$ ($ \pm 0.07$)\\
& Hidden Replay(ER $+$ RaR) & $-3.53$ ($ \pm 1.67$) & $1.35$ ($ \pm 0.09$) & $1.07$ ($ \pm 0.13$) & $1.19$ ($ \pm 0.07$) & $0.8$ ($ \pm 0.04$) & $0.9$ ($ \pm 0.04$) & $0.81$ ($ \pm 0.04$) & $0.61$ ($ \pm 0.11$)\\
& Hidden Replay(ER $+$ GR) & $-9.17$ ($ \pm 2.8$) & $1.57$ ($ \pm 0.42$) & $1.21$ ($ \pm 0.33$) & $1.15$ ($ \pm 0.07$) & $0.75$ ($ \pm 0.04$) & $0.91$ ($ \pm 0.04$) & $0.79$ ($ \pm 0.02$) & $0.59$ ($ \pm 0.07$)\\
& Hidden Replay (ER $+$ RaR $+$ GR) & $-3.05$ ($ \pm 0.97$) & $1.42$ ($ \pm 0.06$) & $1.0$ ($ \pm 0.02$) & $1.17$ ($ \pm 0.06$) & $0.8$ ($ \pm 0.04$) & $0.9$ ($ \pm 0.04$) & $0.83$ ($ \pm 0.04$) & $0.59$ ($ \pm 0.06$)\\
\hline
\end{tabular}
}
\caption{Effect of different replay mechanisms during sleep on lifelong learning metrics on the \SC{} domain. Results are collected using the hidden replay model on twelve syllabi. We report means and 95\% confidence intervals. Hidden replay using only experience replay (ER) and generative replay (GR) is insufficient. Random replay (RaR) is needed to stablize the learning of the feature space, and combining all three types of replay yield the best results.}
\label{tab:replay_mechanism_ablation_sc2}
\end{table}
\subsection{Understanding the Effect of Different Model Architectures}
\label{app:archs_ablation}
In Tables \ref{tab:algo-minigrid} and \ref{tab:algo-M121518}, we compare the different replay architectures introduced in this paper with several simple benchmark models. We observe that the hidden replay model generally outperforms the sequential and two-headed architectures as well as the simple benchmark models.
\begin{table}[H]
\centering
\resizebox{1.0\textwidth}{!}{
\begin{tabular}{|c|c|c|c|c|c|c||c|c|c|c|}
\hline
Scenario & Agent & PM & FT & BT & RP & RR\textsubscript{$\Omega$} & RR\textsubscript{$\alpha$} & RR\textsubscript{$\sigma$} & RR\textsubscript{$\upsilon$} \\
\hline
& Random Policy & $0.37$ ($ \pm 0.81$) $\bullet$ & $1.06$ ($ \pm 0.02$) $\circ$ & $1.13$ ($ \pm 0.18$) $\bullet$ & $0.22$ ($ \pm 0.04$) $\circ$ & $0.09$ ($ \pm 0.02$) $\circ$ & $0.09$ ($ \pm 0.02$) $\circ$ & $0.09$ ($ \pm 0.02$) $\circ$ & $0.09$ ($ \pm 0.00$) $\circ$\\
& Baseline PPO & $-48.2$ ($ \pm 12.04$) & $4.07$ ($ \pm 0.87$) & $0.46$ ($ \pm 0.12$) & $2.17$ ($ \pm 0.61$) & $0.5$ ($ \pm 0.1$) & $0.78$ ($ \pm 0.08$) & $0.64$ ($ \pm 0.08$) & $0.16$ ($ \pm 0.02$)\\
Pairwise & Hidden Replay (ER) & $-45.25$ ($ \pm 13.54$) & $6.34$ ($ \pm 1.4$) & $0.51$ ($ \pm 0.26$) & $2.75$ ($ \pm 0.73$) & $0.47$ ($ \pm 0.06$) & $0.73$ ($ \pm 0.08$) & $0.6$ ($ \pm 0.06$) & $0.11$ ($ \pm 0.02$) $\circ$\\
& Sequential (ER + RaR + GR) & $13.58$ ($ \pm 9.06$) $\bullet$ & $4.01$ ($ \pm 1.02$) & $9.52$ ($ \pm 5.5$) $\bullet$ & $2.57$ ($ \pm 0.72$) & $0.45$ ($ \pm 0.08$) & $0.37$ ($ \pm 0.08$) $\circ$ & $0.41$ ($ \pm 0.08$) $\circ$ & $0.08$ ($ \pm 0.02$) $\circ$\\
& Two-Headed (ER + RaR + GR) & $6.73$ ($ \pm 6.19$) $\bullet$ & $4.24$ ($ \pm 0.88$) & $3.55$ ($ \pm 4.12$) $\bullet$ & $2.85$ ($ \pm 0.72$) & $0.49$ ($ \pm 0.1$) & $0.46$ ($ \pm 0.08$) $\circ$ & $0.48$ ($ \pm 0.08$) & $0.1$ ($ \pm 0.02$) $\circ$\\
& Hidden Replay (ER + RaR + GR) & $-15.77$ ($ \pm 10.09$) & $5.87$ ($ \pm 1.24$) & $1.21$ ($ \pm 0.55$) & $2.47$ ($ \pm 0.69$) & $0.6$ ($ \pm 0.06$) & $0.69$ ($ \pm 0.08$) & $0.64$ ($ \pm 0.06$) & $0.11$ ($ \pm 0.02$) $\circ$\\
\hline
& Random Policy & $-0.02$ ($ \pm 0.26$) $\bullet$ & $1.08$ ($ \pm 0.02$) $\circ$ & $1.08$ ($ \pm 0.02$) $\circ$ & $0.19$ ($ \pm 0.00$) $\circ$ & $0.09$ ($ \pm 0.00$) $\circ$ & $0.09$ ($ \pm 0.00$) $\circ$ & $0.09$ ($ \pm 0.00$) $\circ$ & $0.09$ ($ \pm 0.00$) $\circ$\\
& Baseline PPO & $-38.92$ ($ \pm 3.98$) & $4.72$ ($ \pm 0.75$) & $4.06$ ($ \pm 0.73$) & $1.44$ ($ \pm 0.2$) & $0.22$ ($ \pm 0.04$) & $0.64$ ($ \pm 0.04$) & $0.33$ ($ \pm 0.04$) & $0.15$ ($ \pm 0.02$)\\
Condensed & Hidden Replay (ER) & $-57.71$ ($ \pm 3.61$) & $6.45$ ($ \pm 0.81$) & $5.32$ ($ \pm 0.81$) & $2.04$ ($ \pm 0.1$) $\bullet$ & $0.17$ ($ \pm 0.02$) & $0.77$ ($ \pm 0.04$) $\bullet$ & $0.33$ ($ \pm 0.02$) & $0.12$ ($ \pm 0.02$)\\
& Sequential (ER + RaR + GR)& $-6.8$ ($ \pm 2.29$) $\bullet$ & $4.49$ ($ \pm 0.57$) & $1.98$ ($ \pm 0.35$) $\circ$ & $2.13$ ($ \pm 0.08$) $\bullet$ & $0.56$ ($ \pm 0.04$) $\bullet$ & $0.62$ ($ \pm 0.02$) & $0.58$ ($ \pm 0.02$) $\bullet$ & $0.12$ ($ \pm 0.02$)\\
& Two-Headed (ER + RaR + GR) & $-9.69$ ($ \pm 2.87$) $\bullet$ & $4.1$ ($ \pm 0.57$) & $1.65$ ($ \pm 0.41$) $\circ$ & $2.15$ ($ \pm 0.1$) $\bullet$ & $0.59$ ($ \pm 0.04$) $\bullet$ & $0.69$ ($ \pm 0.04$) & $0.62$ ($ \pm 0.04$) $\bullet$ & $0.13$ ($ \pm 0.02$)\\
& Hidden Replay (ER + RaR + GR) & $-21.94$ ($ \pm 2.09$) & $5.93$ ($ \pm 0.79$) & $1.32$ ($ \pm 0.2$) $\circ$ & $2.1$ ($ \pm 0.1$) $\bullet$ & $0.56$ ($ \pm 0.04$) $\bullet$ & $0.8$ ($ \pm 0.02$) $\bullet$ & $0.63$ ($ \pm 0.02$) $\bullet$ & $0.16$ ($ \pm 0.02$)\\
\hline
\end{tabular}
}
\caption{Comparison of the different model-free generative replay architectures and the baseline on two different Minigrid scenarios. The metrics are averaged over 36 lifetimes per scenario and the 95\% confidence interval is shown. $\bullet$ denotes statistically significant improvement compared to the baseline with a Kruskal-Wallis test followed by Dunn's test with Bon Ferroni correction w/ a p-value of 0.05. $\circ$ denotes the metric is statistically significantly worse than the baseline. We see that the LL agents significantly outperform the non-LL agent interms of performance maintainence and RR\textsubscript{$\Omega$}, suggesting the LL agents significantly reduce catastrophic forgetting. The hidden replay model achieves significantly higher RP and RR\textsubscript{$\alpha$} suggesting that it enables the model to learn new tasks with minimal interference from learning previous tasks.}
\label{tab:algo-minigrid}
\end{table}
\begin{table}[H]
\centering
\resizebox{1.0\textwidth}{!}{
\begin{tabular}{|c|c|c|c|c|c||c|c|c|c|}
\hline
Scenario & Agent & PM & FT & BT & RP & RR\textsubscript{$\Omega$} & RR\textsubscript{$\alpha$} & RR\textsubscript{$\sigma$} & RR\textsubscript{$\upsilon$} \\
\hline
& Random Policy & $0.0$ ($ \pm 0.00$) & $1.0$ ($ \pm 0.00$) & $1.0$ ($ \pm 0.00$) & $0.56$ ($ \pm 0.04$) $\circ$ & $0.33$ ($ \pm 0.06$) $\circ$ & $0.33$ ($ \pm 0.06$) $\circ$ & $0.33$ ($ \pm 0.06$) $\circ$ & $0.34$ ($ \pm 0.14$)\\
& Baseline VTrace (No Sleep) & $-8.2$ ($ \pm 6.54$) & $1.11$ ($ \pm 0.23$) & $0.85$ ($ \pm 0.21$) & $1.03$ ($ \pm 0.06$) & $0.74$ ($ \pm 0.14$) & $0.82$ ($ \pm 0.12$) & $0.78$ ($ \pm 0.12$) & $0.39$ ($ \pm 0.17$)\\
Pairwise & Sequential (ER $+$ GR)& $-5.51$ ($ \pm 3.41$) & $1.18$ ($ \pm 0.36$) & $0.77$ ($ \pm 0.13$) & $0.89$ ($ \pm 0.06$) & $0.54$ ($ \pm 0.15$) & $0.61$ ($ \pm 0.15$) & $0.57$ ($ \pm 0.15$) & $0.47$ ($ \pm 0.24$)\\
& Two-Headed (ER $+$ GR)& $-2.35$ ($ \pm 7.31$) & $1.46$ ($ \pm 0.48$) & $0.98$ ($ \pm 0.18$) & $0.95$ ($ \pm 0.11$) & $0.76$ ($ \pm 0.18$) & $0.79$ ($ \pm 0.18$) & $0.77$ ($ \pm 0.18$) & $0.41$ ($ \pm 0.2$)\\
& Hidden Replay (ER Only) & $-7.9$ ($ \pm 5.11$) & $1.14$ ($ \pm 0.4$) & $0.75$ ($ \pm 0.29$) & $0.92$ ($ \pm 0.09$) & $0.87$ ($ \pm 0.11$) & $0.9$ ($ \pm 0.11$) & $0.8$ ($ \pm 0.13$) & $0.37$ ($ \pm 0.2$)\\
& Hidden Replay (ER $+$ RaR $+$ GR) & $-0.57$ ($ \pm 2.93$) & $1.85$ ($ \pm 0.88$) & $0.95$ ($ \pm 0.15$) & $1.04$ ($ \pm 0.11$) & $0.77$ ($ \pm 0.18$) & $0.82$ ($ \pm 0.13$) & $0.8$ ($ \pm 0.15$) & $0.55$ ($ \pm 0.26$)\\
\hline
& Random Policy & $0.0$ ($ \pm 0.00$) $\bullet$ & $1.0$ ($ \pm 0.00$) & $1.0$ ($ \pm 0.00$) $\bullet$ & $0.42$ ($ \pm 0.04$) $\circ$ & $0.32$ ($ \pm 0.09$) $\circ$ & $0.32$ ($ \pm 0.09$) $\circ$ & $0.32$ ($ \pm 0.09$) $\circ$ & $0.32$ ($ \pm 0.22$)\\
& Baseline VTrace (No Sleep) & $-8.99$ ($ \pm 3.06$) & $1.11$ ($ \pm 0.23$) & $0.79$ ($ \pm 0.12$) & $0.92$ ($ \pm 0.04$) & $0.9$ ($ \pm 0.14$) & $0.91$ ($ \pm 0.12$) & $0.82$ ($ \pm 0.12$) & $0.39$ ($ \pm 0.17$)\\
Alternating & Sequential (ER $+$ GR)& $-7.44$ ($ \pm 3.43$) & $1.18$ ($ \pm 0.36$) & $0.88$ ($ \pm 0.09$) & $0.8$ ($ \pm 0.09$) & $0.59$ ($ \pm 0.15$) & $0.66$ ($ \pm 0.13$) & $0.6$ ($ \pm 0.13$) & $0.47$ ($ \pm 0.24$)\\
& Two-Headed (ER $+$ GR)& $-7.7$ ($ \pm 4.89$) & $1.46$ ($ \pm 0.48$) & $0.91$ ($ \pm 0.11$) & $0.9$ ($ \pm 0.09$) & $0.88$ ($ \pm 0.22$) & $0.94$ ($ \pm 0.11$) & $0.84$ ($ \pm 0.15$) & $0.41$ ($ \pm 0.2$)\\
& Hidden Replay (ER Only) & $-7.9$ ($ \pm 5.11$) & $1.14$ ($ \pm 0.4$) & $0.75$ ($ \pm 0.29$) & $0.92$ ($ \pm 0.09$) & $0.87$ ($ \pm 0.11$) & $0.9$ ($ \pm 0.11$) & $0.8$ ($ \pm 0.13$) & $0.37$ ($ \pm 0.2$)\\
& Hidden Replay (ER $+$ RaR $+$ GR) & $-6.13$ ($ \pm 4.64$) & $1.85$ ($ \pm 0.88$) & $0.87$ ($ \pm 0.13$) & $0.91$ ($ \pm 0.09$) & $0.86$ ($ \pm 0.13$) & $0.89$ ($ \pm 0.13$) & $0.82$ ($ \pm 0.15$) & $0.55$ ($ \pm 0.26$)\\
\hline
& Random & $-0.03$ ($ \pm 0.28$) $\bullet$ & $1.01$ ($ \pm 0.02$) & $1.01$ ($ \pm 0.02$) & $0.54$ ($ \pm 0.02$) $\circ$ & $0.31$ ($ \pm 0.02$) $\circ$ & $0.3$ ($ \pm 0.02$) $\circ$ & $0.31$ ($ \pm 0.00$) $\circ$ & $0.3$ ($ \pm 0.04$) $\circ$\\
& Baseline VTrace (No Sleep) & $-3.41$ ($ \pm 1.7$) & $1.19$ ($ \pm 0.08$) & $1.17$ ($ \pm 0.29$) & $1.14$ ($ \pm 0.06$) & $0.64$ ($ \pm 0.06$) & $0.74$ ($ \pm 0.06$) & $0.66$ ($ \pm 0.06$) & $0.49$ ($ \pm 0.06$)\\
Condensed & Sequential (ER $+$ GR) & $-3.7$ ($ \pm 1.58$) & $1.15$ ($ \pm 0.04$) & $1.0$ ($ \pm 0.09$) & $0.91$ ($ \pm 0.09$) $\circ$ & $0.53$ ($ \pm 0.13$) & $0.69$ ($ \pm 0.09$) & $0.61$ ($ \pm 0.09$) & $0.49$ ($ \pm 0.11$)\\
& Two-Headed (ER $+$ GR) & $-4.53$ ($ \pm 2.97$) & $1.34$ ($ \pm 0.09$) & $1.15$ ($ \pm 0.11$) & $1.09$ ($ \pm 0.13$) & $0.76$ ($ \pm 0.11$) & $0.87$ ($ \pm 0.09$) & $0.78$ ($ \pm 0.09$) & $0.57$ ($ \pm 0.09$)\\
& Hidden Replay (ER Only) & $-8.65$ ($ \pm 3.15$) $\circ$ & $1.42$ ($ \pm 0.15$) $\bullet$ & $1.12$ ($ \pm 0.11$) & $1.12$ ($ \pm 0.09$) & $0.73$ ($ \pm 0.09$) & $0.89$ ($ \pm 0.07$) & $0.77$ ($ \pm 0.04$) & $0.55$ ($ \pm 0.07$)\\
& Hidden Replay (ER $+$ RaR $+$ GR) & $-3.05$ ($ \pm 0.97$) & $1.42$ ($ \pm 0.06$) $\bullet$ & $1.0$ ($ \pm 0.02$) & $1.17$ ($ \pm 0.06$) & $0.8$ ($ \pm 0.04$) & $0.9$ ($ \pm 0.04$) & $0.83$ ($ \pm 0.04$) $\bullet$ & $0.59$ ($ \pm 0.06$)\\
\hline
\end{tabular}
}
\caption{Comparison of the different model-free generative replay architectures and the baseline on three different \SC{} scenarios. The metrics are averaged over 12 lifetimes per scenario and the 95\% confidence interval is shown. $\bullet$ denotes statistically significant improvement compared to the baseline with a Kruskal-Wallis test followed by Dunn's test with Bon Ferroni correction w/ a p-value of 0.05. $\circ$ denotes the metric is statistically significantly worse than the baseline. In general, the hidden replay model performs very well, showing significant gains in FT. Interestingly, the baseline agent performs well in learning new tasks, but the continual learners exhibit better performance on unseen tasks, indicating that they exhibit better generalization. RR\textsubscript{$\Omega$} also shows the continual learners are able to learn without forgetting.}
\label{tab:algo-M121518}
\end{table}
\subsection{Qualitative Analysis of the Generative Model}
\label{sec:reconstructions_app}
In this section, we qualitatively evaluate the capabilities of each replay architecture with respect to its generative modeling capabilities.
In Fig.~\ref{fig:recon_sequential_collect}, we show reconstructions of the first channel of the ``CollectMineralShards'' SC-2 task using the sequential architecture after 11K sleep batches, imitating a converged wake model. These observations are challenging to reconstruct and generate due to the random distribution of shards, each represented by a few pixels. The sequential architecture produces blurry reconstructions that resemble density maps over the shard locations. In Fig.~\ref{fig:recon_two_headed_collect}, we show reconstructions on the same task for the two-headed architecture. By separating the generative model from the policy network, the model reconstructs the observations with sharper detail; however, the two-headed model still struggles with generating realistic observations, generating observations that degrade to density maps. While the generated observations are sub-optimal in terms of realism, experiments suggest they still provide useful information for training the policy network. We show additional ground-truth, reconstructions, and generated samples for the ``DefeatRoaches'' task using the two-headed architecture in Fig.~\ref{fig:recon_two_headed_defeat}. In this case, there is much more structural regularity in the observations, so the reconstructions and generated samples better resemble ground truth data.
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/m12_reconstructions.png}
\caption{Reconstructions of the first channel of the ``CollectMineralShards'' SC-2 task using the sequential architecture after 11,000 sleep batches, imitating a converged wake model.}
\label{fig:recon_sequential_collect}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/m15_reconstructions.png}
\caption{Reconstructions of the first channel of the ``CollectMineralShards'' SC-2 task using the two-headed architecture after 11,000 sleep batches, imitating a converged wake model. We also show some generated samples and the effect of passing generating via the ``loop-vae'' procedure.}
\label{fig:recon_two_headed_collect}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/m15_reconstructions_defeat.png}
\caption{Reconstructions and generated samples of the first channel of the ``DefeatRoaches'' SC-2 task using the two-headed architecture.}
\label{fig:recon_two_headed_defeat}
\end{figure}
Visualizing the reconstructions and generated samples of the hidden replay model is more challenging. To do so, we use a two-dimensional principle components analysis to compare ``ground-truth'' features (i.e., the output of the feature extractor) with reconstructions and generated features. On the left side of Fig.\ \ref{fig:recon_hidden_replay}, the VAE is able to closely reconstruct the ground-truth features for the ``DefeatRoaches'' task. On the right side, we overlay the feature distribution of ``DefeatZerglingsAndBanelings'' and the generated samples for ``DefeatRoaches'' after learning the task sequence ``DefeatRoaches'' $ \rightarrow$ ``DefeatZerglingsAndBanelings''. Since these tasks are similar, we see an overlap between the generated features of ``DefeatRoaches'' from the previous sleep with the new features of ``DefeatZerglingsAndBanelings'', suggesting the generator is generating meaningful features from the task seen during the previous sleep phase.
\begin{figure}[H]
\centering
\includegraphics[width=0.7\textwidth]{graphics/m18_reconstructions_and_samples.png}
\caption{To evaluate reconstructions and generated samples for the hidden replay architecture, we project data in the feature space via its principle components. Left: Reconstructions of the features for the ``DefeatRoaches'' task. Right: Overlaying the feature distribution of ``DefeatZerglingsAndBanelings'' and the generated samples for ``DefeatRoaches'' after learning the task sequence ``DefeatRoaches'' $ \rightarrow$ ``DefeatZerglingsAndBanelings''.}
\label{fig:recon_hidden_replay}
\end{figure}
\subsection{Additional Results: Understanding the Interplay Between Different Replay Mechanisms}
\label{sec:justification_of_exemplars}
We've observed that generative replay alone is not sufficient for the hidden replay architecture. This is because there is no constraint limiting the drift in the feature space between sleeps, so the feature extractor can completely change between two time steps if the tasks are perceptually dissimilar or dissimilar in policy. To help to overcome this issue, we employ ``random replay'' where we save a small batch of random observation-action pairs in the original observation space after every sleep, and randomly replay them during every sleep. In Fig.\ \ref{fig:exemplar_replay_1}, we show the necessity of combining hidden replay with random replay. On the left, we see that using hidden replay without random replay results in the feature space for task 0 dramatically changing after training on task 1, and PM is low. On the right, we see that the distribution of task 0 before and after training on task 1 is significantly better aligned when hidden replay is used in conjunction with random replay, and PM significantly improves. Additional visualizations for pairs of dissimilar tasks are seen in Fig.~\ref{fig:exemplar_replay_2}. In Table \ref{tab:exemplars_vs_no_exemplars}, we see using random replay generally improves PM and FTR, which agrees with our hypothesis that it helps to maintain the feature space, leading to less forgetting and potentially features that generalize better to other tasks. These results are inline with similar observations made from the main experiments of this paper.
\begin{table}[H]
\centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Scenario & Agent & PM & FT & BT & RP \\
\hline
Pairwise & No Random Replay & $-6.28$ ($ \pm 12.82$) & $1.35$ ($ \pm 0.24$) & $0.80$ ($ \pm 0.37$) & $1.08$ ($ \pm 0.19$) \\
& Random Replay & $-3.03$ ($ \pm 4.92$) & $1.65$ ($ \pm 0.28$) & $0.92$ ($ \pm 0.13$) & $1.04$ ($ \pm 0.17$) \\
\hline
Condensed & No Random Replay & $-9.35$ ($ \pm 3.69$) & $1.37$ ($ \pm 0.15$) & $1.20$ ($ \pm 0.24$) & $1.16$ ($ \pm 0.16$) \\
& Random Replay & $-3.05$ ($ \pm 1.76$) & $1.42$ ($ \pm 0.11$) & $1.00$ ($ \pm 0.03$) & $1.17$ ($ \pm 0.11$) \\
\hline
\end{tabular}
\caption{Comparison of the hidden replay model with and without random replay. Results are averaged over multiple lifetimes, and we also record the standard deviation of the metrics.}
\label{tab:exemplars_vs_no_exemplars}
\end{table}
\begin{figure}[H]
\centering
\includegraphics[width=0.8\textwidth]{graphics/m18_exemplars.png}
\caption{We show the necessity of combining hidden replay with random replay. Left: Using hidden replay without random replay results in the feature space for task 0 dramatically changing after training on task 1, and the feature space for both classes converge to a point. PM is -10.68. Right: The distribution of task 0 before and after training on task 1 are significantly better aligned when hidden replay is used in conjunction with random replay. PM is -1.93.}
\label{fig:exemplar_replay_1}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/m18_exemplars_2.png}
\caption{We show two examples where hidden + random replay maintains the feature space of the previous task after learning a new dissimilar SC-2 task (``Collect'' $\rightarrow$ ``Defeat'' and vice versa).}
\label{fig:exemplar_replay_2}
\end{figure}
\section{Appendix}
\subsection{Potential Negative and Positive Societal Impacts}
This approach presents a means of training a multi-agent system to play real-time strategy games using lifelong reinforcement learning. Such a tool could be used by malicious foreign actors to train models for strategic military actions, which could negatively impact the national security of benign nations.
However, this tool can also be used for societal benefits including training intelligent robot swarms to perform useful tasks over long lifetimes (e.g., applications in agricultural and manufacturing robotics). Similarly, it could be used in domains which require incorporation of feedback into the learning (i.e. necessitates RL) and where distribution shift is a significant problem; e.g., for planning treatments in healthcare.
\subsection{Limitations}
The model is not flawless. The following are some limitations we have observed w.r.t. our proposed approach:
\begin{itemize}
\item While catastrophic forgetting is reduced using our proposed models, the agent still fails to perfectly learn and remember tasks, and thus, there is still room for improvement in terms of many of the LL metrics. Specifically, we've observed that the model can struggle to remember over very long periods when trained on long syllabi of tasks.
\item Training the agent is non-trivial, requiring carefully defining reasonable tasks and performing hyperparameter tuning.
\item The sleep model is dependent on having a good quality single-task expert for the wake model, and in some cases, the wake model fails to properly learn a task, disrupting the learning of the sleep model.
\item We have also seen some fragility in the learning of the hidden space for hidden replay and imperfect generated samples in the sequential and two-headed models. We have observed cases where the sampled data drifts enough that it mis-aligns with the true distribution of data and/or the quality of the sampled data degrades significantly. In such cases, the sleep model can become unstable and the proposed replay mechanisms may hurt the model's learning.
\item While we would like to rely mostly on generative replay and experience replay from the last wake cycle for training the model, we've found that the hidden replay model requires a small buffer of random samples across tasks to maintain the hidden space. This introduces sub-optimality in the memory usage of the model. We have also observed that this ``random replay'' has an out-sized effect on improving lifelong learning compared to the generative replay.
\end{itemize}
\subsection{Starcraft 2 Task Variants}
\label{sec:sc2_task_descriptions}
The task variants we consider are as follows:
\begin{itemize}
\item \textbf{Collect Mineral Shards – No Fog of War}: A map with 2 Marines and an endless supply of Mineral Shards. Rewards are earned by moving the Marines to collect the Mineral Shards, with optimal collection requiring both Marine units to be split up and moved independently. Whenever all 20 Mineral Shards have been collected, a new set of 20 Mineral Shards are spawned at random locations (at least 2 units away from all Marines). Fog of war is disabled.
\item \textbf{Collect Mineral Shards – Fog of war:} A map with 2 Marines and an endless supply of Mineral Shards. Rewards are earned by moving the Marines to collect the Mineral Shards, with optimal collection requiring both Marine units to be split up and moved independently. Whenever all 20 Mineral Shards have been collected, a new set of 20 Mineral Shards are spawned at random locations (at least 2 units away from all Marines). Fog of war is enabled, meaning the agent must be able to learn without full knowledge of the current state of the environment.
\item \textbf{DefeatZerglingsAndBanelings – One Group:} A map with 9 Marines on the opposite side from a group of 6 Zerglings and 4 Banelings. Rewards are earned by using the Marines to defeat Zerglings and Banelings. Whenever all Zerglings and Banelings have been defeated, a new group of 6 Zerglings and 4 Banelings is spawned and the player is awarded 4 additional Marines at full health, with all other surviving Marines retaining their existing health (no restore). Whenever new units are spawned, all unit positions are reset to opposite sides of the map
\item \textbf{DefeatZerglingsAndBanelings – Two Groups:} A map with 9 Marines in the center with two groups consisting of 9 Zerglings one one side and 6 Banelings on the other side. Rewards are earned by using the Marines to defeat Zerglings and Banelings. Whenever a group has been defeated, a new group of 9 Zerglings and 6 Banelings is spawned and the player is awarded 6 additional Marines at full health, with all other surviving Marines retaining their existing health (no restore). Whenever new units are spawned, all unit positions are reset to opposite sides of the map.
\item \textbf{DefeatRoaches – One Group:} A map with 9 Marines and a group of 4 Roaches on opposite sides. Rewards are earned by using the Marines to defeat Roaches, with optimal combat strategy requiring the Marines to perform focus fire on the Roaches. Whenever all 4 Roaches have been defeated, a new group of 4 Roaches is spawned and the player is awarded 5 additional Marines at full health, with all other surviving Marines retaining their existing health (no restore). Whenever new units are spawned, all unit positions are reset to opposite sides of the map.
\item \textbf{DefeatRoaches – Two Groups:} A map with 9 Marines in the center and two groups consisting of 6 total Roaches on opposite sides (3 on each side). Rewards are earned by using the Marines to defeat Roaches, with optimal combat strategy requiring the Marines to perform focus fire on the Roaches. Whenever all 6 Roaches have been defeated, a new group of 6 Roaches is spawned and the player is awarded 7 additional Marines at full health, with all other surviving Marines retaining their existing health (no restore). Whenever new units are spawned, all unit positions are reset to starting areas of the map.
\end{itemize}
\subsection{Implementation Details}
\label{exp:archs}
\subsubsection{Starcraft 2}
\label{sec:imp_sc2}
In this section we describe the implementation details for our \SC{} agent and experiments. The wake agent is a VTrace agent \citep{espeholt2018impala}. The feature extractor component of the wake agent consists of an architecture that is a slight modification of the FullyConv model of \citet{vinyals2017starcraft}. We further split this into a feature extractor component and a policy network component to facilitate easy weight copying between the wake and sleep agents. For the hidden replay model, we add a LayerNormalization layer to the output of the feature extractor to constrain the space of the features, which helps to improve feature reconstructions and limit feature drift. The architecture produces features of dimension 8192.
The sleep agent feature extractor and policy network architectures are identical to the wake model. We use a five-layer convolutional neural network architecture with for the encoder and a five-layer deconvolutional architecture for the decoder for the VAEs for the sequential and two-headed models. In both cases, we use ReLU activations between the layers and add batch normalization before the ReLUs. We use two-layered multi-layered perceptrons with ReLU activation functions for the encoder and decoder for the VAE for the hidden replay model. To improve numerical stability, the logvar predicted by the encoder is bounded to take values in $[-5, 5]$ using a scaled-hyperbolic tangent layer.
Three styles of replay are used during sleep for the hidden replay architecture: experience replay on a buffer of 10,000 observation-action pairs from a FIFO queue collected during the wake phase, generative replay on the feature vectors, and random replay. Only experience replay and generative replay is used to train the sleep agent for the sequential and two-headed architectures.
For the sequential architecture: the weight of the imitation loss (cross-entropy on action logits) is 50.0, the weight of the reconstruction loss is 1.0, and the weight of the KL loss of the VAE is 3.0. For the two-headed architecture: the weight of the imitation loss (cross-entropy on action logits) is 30.0, the weight of the reconstruction loss is 1.0, and the weight of the KL loss of the VAE is 2.0. For the hidden replay architecture: the weight of the imitation loss (cross-entropy on action logits) is 50.0, the weight of the reconstruction loss is 200.0, and the weight of the KL loss of the VAE is 5.0. These were determined via hand-tuning.
The agent enters sleep three times per task at even intervals for the sequential and two-headed architectures. The agent enters sleep twice per task at even intervals for the hidden replay architecture. Each training iteration of the wake model is trained on 32 trajectories. Each sleep consists of 4,500 iterations of training using batch sizes of 64 (64 wake samples, 64 generated samples, and 64 random observations from the random replay buffer). Generative replay is not used until the second sleep. 96 observations are selected randomly from the wake buffer every sleep to be added to the random replay buffer for the hidden replay architecture. When the agent wakes up, it resets the PPO model and offers advice with a 80\% probability linearly decaying to 0\% after for one-half the the duration between sleeps. No weight copying is used for the sequential variant, but weight copying (running a mini-evaluation block after sleep and copying the best Eigentask's parameters into the wake agent) and advice are both used for the two-headed and hidden replay variants.
The Adam optimizer is used to train both the wake and sleep agents with a learning rate of 1.0e-3.
\textbf{Evaluation:} We selected 3 minigames involving battles (\emph{DefeatRoaches} and \emph{DefeatZerglingsAndBanelings} tasks) or resource collection (\emph{CollectMineralShards}). We created two variants of each task, differentiated by the starting locations of units, the presence of ``fog-of-war'', and/or the number of friendly and enemy units. Task variants are described in Section~\ref{sec:sc2_task_descriptions}. We used PySC2 \citep{vinyals2017starcraft} to interface with SC-2. We used a subset of the available observation maps namely the unit type, selection status, and unit density two-dimensional observations. The action space is factored into functions and arguments, such as $\text{move}(x, y)$ or $\text{stop}()$. Following \citet{vinyals2017starcraft}, our policy networks output probabilities for the different action factors independently.
The agent receives positive rewards for collecting resources and defeating enemy units, and negative rewards for losing friendly units. Single task experts (STEs) were trained to convergence ($\sim$ 30 million environment steps using the VTrace algorithm \citep{espeholt2018impala}). We consider the pairwise, alternating, and condensed scenarios where each LB consists of 2 million environment steps (about 6\% of the STE samples), and each EB consists of 30 episodes per task.
Each lifetime of a condensed scenario takes approximately 3-4 days running on systems with RTX2080 GPUs.
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/m18_pi_similarity_matrix.png}
\caption{Similarity between tasks measured in terms of forward transfer using single task experts}
\label{fig:similarity_matrix}
\end{figure}
\subsubsection{Minigrid}
\label{sec:imp_minigrid}
In this section, we describe the implementation details for our Minigrid agent and experiments. The wake agent is based on the Stable Baselines 3 PPO agent \citep{schulman2017proximal,stable-baselines3}. There are separate networks for computing the value and policy. Both networks are simple multi-layered perceptrons that first map to a 256-dimensional feature space and then apply a densely connected layer to select one of six actions or estimate the value of a state. ReLUs are used as activation functions between all of the layers. Observations are scaled so the values of every cell are in $[0, 1]$ before being input to the network. For exact parameters for the wake agent, please refer to the appendix.
The sleep agent uses one Eigentask. The feature extractor and expected inputs are the same as those of the wake agent's policy network. A variational autoencoder is used as generative model. The VAE is a four-layer bottleneck-style MLP architecture with a latent dimension size: 128. To improve numerical stability, the logvar predicted by the encoder is bounded to take values in $[-5, 5]$.
Three styles of replay are used during sleep: experience replay on a buffer of 20,000 observation-action pairs from a FIFO queue collected during the wake phase, generative replay on the feature vectors, and random replay (256 random observations added per sleep). The weight of the imitation loss (cross-entropy on discrete actions) is 3.0, the weight of the reconstruction loss is 1.0, and the weight of the KL loss of the VAE is 0.03. The agent enters sleep at the end of each task. Each sleep consists of 20,000 iterations of training using batch sizes of 32 (32 wake samples, 32 generated samples, and 32 random examples drawn from the random replay buffer). Generative replay and random replay are not used until the second sleep. 256 observations are selected randomly from the wake buffer every sleep to be added to the random replay buffer with a maximum buffer size of 4096. When the agent wakes up, it resets the PPO model and offers advice with a 90\% probability linearly decaying to 0\% after 100,000 environment steps for the condensed scenario and 33,333 environment steps for the dispersed scenario. No weight copying is used.
The Adam optimizer is used to train both the wake and sleep agents.
\textbf{Evaluation:} We report results using the Minigrid domain \citep{gym_minigrid}, in which an agent must navigate to a goal while interacting with various entities in the environment. We consider five tasks --- SimpleCrossing, DistShift, a custom Fetch, a custom Unlock, and DoorKey --- each with two variants (differing in the size of the grid and/or number of objects) for a total of ten POMDPs. Descriptions of these tasks can be found in \citet{gym_minigrid}. An observation consists of a $7\times7$ top-down view of the grid directly in front of the agent, with three channels containing information about the object type, object color, and object state in each tile. Rewards are sparse and consist of $-1$ for running into obstacles and lava, $0$ for not reaching the goal, and a positive reward in $[0, 1]$ for reaching the goal that is proportional to the number of steps taken. Unlike SC-2, there are only six actions: turn left, turn right, move forward, pick up an object, drop an object, and interact with an object.
STEs are trained on each task for 1M environment steps using PPO \citep{schulman2017proximal}. We consider the pairwise, alternating, and condensed scenarios where the LB lengths vary between 100k and 700k (Section~\ref{sec:minigrid_task_lengths}) based on examining the smallest number of steps for the STE to converge to $80\%$ of its peak performance averaged over ten runs. Each EB consists of 100 episodes per task.
Each lifetime of a condensed scenario takes approximately 4-6 hours to run on systems with RTX2080 GPUs.
\subsection{Minigrid Task Lengths}
\label{sec:minigrid_task_lengths}
We use the following number of environment steps per task in our minigrid experiments:
\begin{itemize}
\item SimpleCrossingS9N1: 400,000
\item SimpleCrossingS9N2: 500,000
\item DistShiftR2: 200,000
\item DistShiftR3: 200,000
\item CustomFetchS5T1N2: 500,000
\item CustomFetchS8T1N2: 700,000
\item CustomUnlockS5: 200,000
\item CustomUnlockS7: 300,000
\item DoorKeyS5: 200,000
\item DoorKeyS6: 300,000
\end{itemize}
\subsection{Parameters of the PPO agent}
These are the parameters used for the stable baselines 3 PPO agent for the Minigrid tasks:
\begin{itemize}
\item n\_steps: 512
\item batch\_size: 32
\item gae\_lambda: 0.95
\item gamma: 0.99
\item n\_epochs: 10
\item ent\_coef: 5.0e-5
\item learning\_rate: 2.5e-4
\item clip\_range: 0.3
\item vf\_coef: 0.75
\item max\_grad\_norm: 5.0
\item policy\_kwargs: dict(activation\_fn=nn.ReLU, net\_arch=[dict(pi=[256, 256], vf=[256, 256])])
\end{itemize}
\subsection{Evaluation Scenarios Visualization}
\begin{figure}[H]
\centering
\includegraphics[width=0.75\textwidth, keepaspectratio]{graphics/scenarios_2.png}
\caption{Lifelong Learning Scenarios. Grey evaluation blocks (EBs) evaluate the agent on all tasks. Learning blocks (LBs) are color coded by task, and subscripts denote task variants.}
\label{fig:scenarios}
\end{figure}
\subsection{Example Learning Curves for Lifelong Curricula on \SC{}}
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/sc2_learning_curve_cropped.png}
\caption{Learning curve for our lifelong learning agent applied to a six minigame scenario in the \SC{} domain. Performance comparable to the converged single-task expert is \textasciitilde 100.}
\label{fig:sc2_learning_curves}
\end{figure}
\subsection{Comparison of Learning Curves for Lifelong Curricula on Minigrid}
\begin{figure}[H]
\centering
\includegraphics[width=0.45\textwidth]{graphics/minigridcondensed_non-ll.png}
\includegraphics[width=0.45\textwidth]{graphics/minigridcondensed_ll.png}
\caption{Learning curves for a condensed scenario in the Minigrid domain. Left: learning curve of non-LL agent; Right: learning curve of LL agent on the same scenario. Maximum reward is 100. The LL agent is capable of learning more tasks without interference and exhibits jump starts in performance for similar tasks due to the advice mechanism.}
\label{fig:minigrid_learning_curves}
\end{figure}
\subsection{A Sleep Policy Trained with Policy Distillation can Learn to Imitate a Converged Single Task Expert}
\label{sec:ste_additional_results}
\textbf{Single Task Experiments}:
\textbf{SC-2:} We demonstrate that a sleep policy can be trained with policy distillation to mimic the performance of an STE's policy. We consider STEs identical to the wake agents trained for 30M environment steps for each of the six SC-2 variants. We act in accordance with the following procedure: i) Train an STE on each of the six SC-2 POMDPs for 30M steps. ii) Measure the final reward for each STE. iii) Load the trained STE and build a buffer of 10K trajectories. iv) Train the sleep model until convergence on randomly sampled batches of observation-action pairs (batch size: 32). v) Periodically evaluate the sleep model on SC-2 simulations, each consisting of 30 episodes of the task-of-interest.
We show results in Figs.~\ref{fig:sleep_imitating_ste} and \ref{fig:sleep_imitating_ste_app}. The two-headed and hidden replay architectures architectures typically match or exceed the STE performance in terms of rewards. The sequential model learns non-trivial policies, but generally performs worse than the STE. This supports our hypothesis that decoupling the generator from the policy network of the VAE results in better-performing policies.
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/sleep_imitating_ste_combined.png}
\caption{Examining whether sleep policies can be trained to mimic the performance of single task experts for the SC-2 POMDPs with the sequential, two-headed, and hidden replay model without a random replay buffer. The blue curves denote the performance (in terms of rewards) of the sleep policy interpolated along evaluation points. The red line denotes terminal performance of the STE after convergence. Error bars denote standard deviation of rewards.}
\label{fig:sleep_imitating_ste}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/sleep_imitating_ste.png}
\caption{Examining whether the sleep policy can be trained to mimic the performance of single task experts for the SC-2 POMDPs with the hidden replay model without a random replay buffer. The blue curves denote the performance (in terms of rewards) of the sleep policy interpolated along evaluation points. The red line denotes terminal performance of the STE after convergence. Error bars denote standard deviation of rewards.}
\label{fig:sleep_imitating_ste_app}
\end{figure}
\textbf{Minigrid:} We also quantitatively validate that the sleep policy using the hidden replay model is able to distill the an STE's converged policy after 1M environment steps for the Minigrid domain using an experience buffer of the STE's last 20K observation-action pairs (repeated 10 times per task). We find that the average difference in rewards between the STE and sleep policy's performance over all task variants is 0.14 with a standard deviation of 0.11, suggesting that the sleep policy imitates the STE to an acceptable level.
\subsection{A Sleep Policy Trained with Policy Distillation can Learn to Imitate the Wake Policy on a Single Task}
\label{sec:wake_additional_results}
We conduct experiments to show that the converged performance of a wake policy can be achieved by our agent via imitating the sequence of intermediate wake policies over multiple wake-sleep phases. This is an important experiment needed to show stable convergence of the wake-sleep mechanism even when distilling sequences of sub-optimal policies. These experiments involve training an agent on a single task, periodically entering sleep at set intervals. We observed that for most tasks, the curves formed by evaluating the sleep agent and wake agents at periodic evaluation blocks closely track one another, demonstrating stable convergence of the sleep agent's policy.
\textbf{SC-2:} We consider the case where the system performs multiple wake-sleep phases over the course of a single task, requiring the sleep policy to consolidate the increasingly improving policies of the wake agent. Using the SC-2 simulator, each wake policy is trained for 6M environment steps, entering sleep at even intervals ten times over the entire run. Every time the system exits sleep, the wake and sleep policies are evaluated on thirty episodes of the task. We show the wake versus sleep performance (in terms of rewards) of a hidden replay-based sleep policy in Fig.\ \ref{fig:sleep_imitating_wake_app}. The sleep policy's performance typically closely mirrors the wake policy's performance, but there are cases where the agent learns a non-trivial policy, but under-performs the wake policy. While not pictured here, similar behavior was observed for the sequential and two-headed architectures as well.
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/sleep_imitating_wake.png}
\caption{Examining whether the sleep policy can be trained to mimic the performance of wake non-lifelong learner for the SC-2 POMDPs as the non-lifelong learners improve. This experiment examined the performance of the hidden replay model without using observations from the random replay buffer. The blue curves denote the performance (in terms of rewards) of the sleep policy interpolated along evaluation points. The red line denotes performance of the wake policy interpolated along evaluation points. Error bars denote standard deviation of rewards.}
\label{fig:sleep_imitating_wake_app}
\end{figure}
\textbf{Minigrid:} We also conducted a post-hoc analysis of our condensed scenario runs in the Minigrid domain to quantify the sleep policy's capability to imitate the wake policy after learning the first task.
\textit{\textbf{Sequential Model:}} The mean absolute difference between the wake policy's average reward for the last 100 episodes and the sleep policy's average reward for the evaluation block was 0.589 with a standard deviation of 0.320, suggesting the sleep policy when using the sequential architecture struggles to accurately imitate the wake policy in many, but not all runs. In cases where the wake model converges to a policy with average reward $>0.9$ for the last 100 episodes, the mean difference metric decreases to $0.591 \pm 0.384$.
\textit{\textbf{Two-Headed Model:}} The mean absolute difference between the wake policy's average reward for the last 100 episodes and the sleep policy's average reward for the evaluation block was 0.422 with a standard deviation of 0.263, suggesting the sleep policy when using the two-headed model can struggle to accurately imitate the wake policy in some runs. In cases where the wake model converges to a policy with average reward $>0.9$ for the last 100 episodes, the mean difference metric decreases to $0.341 \pm 0.297$.
\textit{\textbf{Hidden Replay Model:}} The mean absolute difference between the wake policy's average reward for the last 100 episodes and the sleep policy's average reward for the evaluation block was 0.209 with a standard deviation of 0.227, suggesting the sleep policy accurately imitates the wake policy in many, but not all runs. In cases where the wake model converges to a policy with average reward $>0.9$ for the last 100 episodes, the mean difference metric decreases to $0.110 \pm 0.145$.
These results suggest that the hidden replay model has an advantage when it comes to imitating a wake model when learning a task from scratch. The sequential model struggles with imitating the wake model when learning a task from scratch; however, our main results in the paper suggest that as the sequential model encounters additional tasks, this issue of successfully imitating the wake policy is somewhat alleviated as the relative reward post learning block metric ultimately is similar to the two-headed model when deployed over an entire syllabus of tasks.
\subsection{Effect of Varying the Number of Wake-Sleep Cycles}
we consider the case where the system encounters multiple tasks in sequence, and we evaluate the effect of the number of wake-sleep phases per task on the lifelong learning metrics. We trained the two-headed replay architecture for 4 lifetimes, each consisting of alternating between an \SC{} combat and collection task with 2M environment steps per task. We compared the effect of one, two, and three sleeps per task and report results in Table \ref{tab:multi-sleep}. Results suggest that multiple sleep phases result in improved or maintained performance in most metrics, but may negatively affect FT.
\begin{table}[H]
\centering
\resizebox{1.0\textwidth}{!}{
\begin{tabular}{|c|c|c|c|c||c|c|c|c|c|}
\hline
Agent & PM & FT & BT & RP & RR\textsubscript{$\Omega$} & RR\textsubscript{$\alpha$} & RR\textsubscript{$\sigma$} & RR\textsubscript{$\upsilon$} \\
\hline
One Sleep Cycle Per Task & $-8.76$ ($ \pm 4.39$) & $1.68$ ($ \pm 2.16$) & $0.76$ ($ \pm 0.06$) & $0.78$ ($ \pm 0.22$) & $0.58$ ($ \pm 0.29$) & $0.7$ ($ \pm 0.22$) & $0.58$ ($ \pm 0.22$) & $0.3$ ($ \pm 0.38$)\\
Two Sleep Cycles Per Task & $-10.47$ ($ \pm 6.41$) & $1.16$ ($ \pm 0.67$) & $0.93$ ($ \pm 0.11$) & $0.87$ ($ \pm 0.11$) & $0.79$ ($ \pm 0.47$) & $0.95$ ($ \pm 0.25$) & $0.81$ ($ \pm 0.31$) & $0.27$ ($ \pm 0.31$)\\
Three Sleep Cycles Per Task & $-8.63$ ($ \pm 7.13$) & $1.06$ ($ \pm 0.86$) & $0.78$ ($ \pm 0.25$) & $0.84$ ($ \pm 0.13$) & $0.67$ ($ \pm 0.19$) & $0.7$ ($ \pm 0.13$) & $0.61$ ($ \pm 0.1$) & $0.19$ ($ \pm 0.19$)\\
\hline
\end{tabular}
}
\caption{Comparing the effect of sleeping multiple times per task on a subset of four alternating scenarios involving transitioning between combat and collection \SC{} tasks. We see in many cases that sleeping multiple times per task can improve metrics.}
\label{tab:multi-sleep}
\end{table}
\subsection{Understanding the Effect of Different Replay Types}
\label{app:replay_ablation}
In Tables \ref{tab:replay_mechanism_ablation_minigrid} and \ref{tab:replay_mechanism_ablation_sc2}, we investigate the effects of turning on and off different replay mechanisms in the hidden replay model. In general, we validate that combining experience replay of the wake buffer, generative replay, and random replay, results in an improved agent; although experience replay in combination with random replay performs surprisingly well.
\begin{table}[H]
\centering
\resizebox{1.0\textwidth}{!}{
\begin{tabular}{|c|c|c|c|c|c||c|c|c|c|}
\hline
Scenario & Agent & PM & FT & BT & RP & RR\textsubscript{$\Omega$} & RR\textsubscript{$\alpha$} & RR\textsubscript{$\sigma$} & RR\textsubscript{$\upsilon$} \\
\hline
& Random Policy & $0.37$ ($ \pm 0.81$) & $1.06$ ($ \pm 0.02$) & $1.13$ ($ \pm 0.18$) & $0.22$ ($ \pm 0.04$) & $0.09$ ($ \pm 0.02$) & $0.09$ ($ \pm 0.02$) & $0.09$ ($ \pm 0.02$) & $0.09$ ($ \pm 0.00$) \\
& Baseline PPO & $-48.2$ ($ \pm 12.04$) & $4.07$ ($ \pm 0.87$) & $0.46$ ($ \pm 0.12$) & $2.17$ ($ \pm 0.61$) & $0.5$ ($ \pm 0.1$) & $0.78$ ($ \pm 0.08$) & $0.64$ ($ \pm 0.08$) & $0.16$ ($ \pm 0.02$)\\
Pairwise & Hidden Replay (ER) & $-45.25$ ($ \pm 13.54$) & $6.34$ ($ \pm 1.4$) & $0.51$ ($ \pm 0.26$) & $2.75$ ($ \pm 0.73$) & $0.47$ ($ \pm 0.06$) & $0.73$ ($ \pm 0.08$) & $0.6$ ($ \pm 0.06$) & $0.11$ ($ \pm 0.02$) \\
& Hidden Replay (ER + RaR) & $-21.11$ ($ \pm 11.06$) & $5.97$ ($ \pm 1.84$) & $0.75$ ($ \pm 0.19$) & $2.22$ ($ \pm 0.91$) & $0.56$ ($ \pm 0.08$) & $0.68$ ($ \pm 0.08$) & $0.62$ ($ \pm 0.08$) & $0.11$ ($ \pm 0.02$)\\
& Hidden Replay (ER + GR) & $-62.32$ ($ \pm 25.2$) & $7.12$ ($ \pm 2.51$) & $0.31$ ($ \pm 0.29$) & $2.51$ ($ \pm 1.61$) & $0.4$ ($ \pm 0.11$) & $0.74$ ($ \pm 0.13$) & $0.57$ ($ \pm 0.11$) & $0.11$ ($ \pm 0.02$)\\
& Hidden Replay (ER + RaR + GR) & $-15.77$ ($ \pm 10.09$) & $5.87$ ($ \pm 1.24$) & $1.21$ ($ \pm 0.55$) & $2.47$ ($ \pm 0.69$) & $0.6$ ($ \pm 0.06$) & $0.69$ ($ \pm 0.08$) & $0.64$ ($ \pm 0.06$) & $0.11$ ($ \pm 0.02$) \\
\hline
& Random Policy & $-0.02$ ($ \pm 0.26$) & $1.08$ ($ \pm 0.02$) & $1.08$ ($ \pm 0.02$) & $0.19$ ($ \pm 0.00$) & $0.09$ ($ \pm 0.00$) & $0.09$ ($ \pm 0.00$) & $0.09$ ($ \pm 0.00$) & $0.09$ ($ \pm 0.00$) \\
& Baseline PPO & $-38.92$ ($ \pm 3.98$) & $4.72$ ($ \pm 0.75$) & $4.06$ ($ \pm 0.73$) & $1.44$ ($ \pm 0.2$) & $0.22$ ($ \pm 0.04$) & $0.64$ ($ \pm 0.04$) & $0.33$ ($ \pm 0.04$) & $0.15$ ($ \pm 0.02$)\\
Condensed & Hidden Replay (ER) & $-57.71$ ($ \pm 3.61$) & $6.45$ ($ \pm 0.81$) & $5.32$ ($ \pm 0.81$) & $2.04$ ($ \pm 0.1$) & $0.17$ ($ \pm 0.02$) & $0.77$ ($ \pm 0.04$) & $0.33$ ($ \pm 0.02$) & $0.12$ ($ \pm 0.02$)\\
& Hidden Replay (ER + RaR) & $-20.96$ ($ \pm 2.81$) & $5.44$ ($ \pm 0.93$) & $1.85$ ($ \pm 0.45$) & $2.11$ ($ \pm 0.1$) & $0.54$ ($ \pm 0.04$) & $0.77$ ($ \pm 0.02$) & $0.61$ ($ \pm 0.02$) & $0.16$ ($ \pm 0.02$)\\
& Hidden Replay (ER + GR) & $-52.91$ ($ \pm 7.26$) & $6.75$ ($ \pm 1.03$) & $4.93$ ($ \pm 1.8$) & $1.97$ ($ \pm 0.24$) & $0.19$ ($ \pm 0.02$) & $0.75$ ($ \pm 0.07$) & $0.33$ ($ \pm 0.02$) & $0.12$ ($ \pm 0.02$)\\
& Hidden Replay (ER + RaR + GR) & $-21.94$ ($ \pm 2.09$) & $5.93$ ($ \pm 0.79$) & $1.32$ ($ \pm 0.2$) & $2.1$ ($ \pm 0.1$) & $0.56$ ($ \pm 0.04$) & $0.8$ ($ \pm 0.02$) & $0.63$ ($ \pm 0.02$) & $0.16$ ($ \pm 0.02$)\\
\hline
\end{tabular}
}
\caption{Effect of different replay mechanisms during sleep on lifelong learning metrics on the Minigrid domain. Results are collected using the hidden replay model on thirty-six syllabi. We report means and 95\% confidence intervals. Hidden replay using only experience replay (ER) and generative replay (GR) is insufficient. Random replay (RaR) is needed to stablize the learning of the feature space, and combining all three types of replay yield the best results.}
\label{tab:replay_mechanism_ablation_minigrid}
\end{table}
\begin{table}[H]
\centering
\resizebox{1.0\textwidth}{!}{
\begin{tabular}{|c|c|c|c|c|c||c|c|c|c|}
\hline
Scenario & Agent & PM & FT & BT & RP & RR\textsubscript{$\Omega$} & RR\textsubscript{$\alpha$} & RR\textsubscript{$\sigma$} & RR\textsubscript{$\upsilon$} \\
\hline
& Random Policy & $0.0$ ($ \pm 0.00$) & $1.0$ ($ \pm 0.00$) & $1.0$ ($ \pm 0.00$) & $0.42$ ($ \pm 0.04$) & $0.32$ ($ \pm 0.09$) & $0.32$ ($ \pm 0.09$) & $0.32$ ($ \pm 0.09$) & $0.32$ ($ \pm 0.22$)\\
& Baseline VTrace & $-8.99$ ($ \pm 3.06$) & $1.11$ ($ \pm 0.23$) & $0.79$ ($ \pm 0.12$) & $0.92$ ($ \pm 0.04$) & $0.9$ ($ \pm 0.14$) & $0.91$ ($ \pm 0.12$) & $0.82$ ($ \pm 0.12$) & $0.39$ ($ \pm 0.17$)\\
& Hidden Replay(ER) & $-7.9$ ($ \pm 5.11$) & $1.14$ ($ \pm 0.4$) & $0.75$ ($ \pm 0.29$) & $0.92$ ($ \pm 0.09$) & $0.87$ ($ \pm 0.11$) & $0.9$ ($ \pm 0.11$) & $0.8$ ($ \pm 0.13$) & $0.37$ ($ \pm 0.2$)\\
Pairwise & Hidden Replay(ER $+$ RaR) & $-3.7$ ($ \pm 4.18$) & $1.75$ ($ \pm 0.66$) & $0.92$ ($ \pm 0.11$) & $1.07$ ($ \pm 0.11$) & $0.78$ ($ \pm 0.18$) & $0.81$ ($ \pm 0.18$) & $0.79$ ($ \pm 0.18$) & $0.48$ ($ \pm 0.26$)\\
& Hidden Replay(ER $+$ GR) & $-9.35$ ($ \pm 7.46$) & $1.39$ ($ \pm 0.64$) & $0.71$ ($ \pm 0.18$) & $1.05$ ($ \pm 0.09$) & $0.71$ ($ \pm 0.15$) & $0.82$ ($ \pm 0.11$) & $0.77$ ($ \pm 0.13$) & $0.44$ ($ \pm 0.31$)\\
& Hidden Replay (ER $+$ RaR $+$ GR) & $-0.57$ ($ \pm 2.93$) & $1.85$ ($ \pm 0.88$) & $0.95$ ($ \pm 0.15$) & $1.04$ ($ \pm 0.11$) & $0.77$ ($ \pm 0.18$) & $0.82$ ($ \pm 0.13$) & $0.8$ ($ \pm 0.15$) & $0.55$ ($ \pm 0.26$)\\
\hline
& Random Policy & $0.0$ ($ \pm 0.00$) & $1.0$ ($ \pm 0.00$) & $1.0$ ($ \pm 0.00$) & $0.42$ ($ \pm 0.04$) & $0.32$ ($ \pm 0.09$) & $0.32$ ($ \pm 0.09$) & $0.32$ ($ \pm 0.09$) & $0.32$ ($ \pm 0.22$)\\
& Baseline VTrace & $-8.99$ ($ \pm 3.06$) & $1.11$ ($ \pm 0.23$) & $0.79$ ($ \pm 0.12$) & $0.92$ ($ \pm 0.04$) & $0.9$ ($ \pm 0.14$) & $0.91$ ($ \pm 0.12$) & $0.82$ ($ \pm 0.12$) & $0.39$ ($ \pm 0.17$)\\
Alternating & Hidden Replay(ER) & $-7.9$ ($ \pm 5.11$) & $1.14$ ($ \pm 0.4$) & $0.75$ ($ \pm 0.29$) & $0.92$ ($ \pm 0.09$) & $0.87$ ($ \pm 0.11$) & $0.9$ ($ \pm 0.11$) & $0.8$ ($ \pm 0.13$) & $0.37$ ($ \pm 0.2$)\\
& Hidden Replay(ER $+$ RaR) & $-4.25$ ($ \pm 2.77$) & $1.75$ ($ \pm 0.66$) & $0.88$ ($ \pm 0.11$) & $0.91$ ($ \pm 0.09$) & $0.81$ ($ \pm 0.15$) & $0.83$ ($ \pm 0.13$) & $0.79$ ($ \pm 0.15$) & $0.48$ ($ \pm 0.26$)\\
& Hidden Replay(ER $+$ GR) & $-10.99$ ($ \pm 4.89$) & $1.39$ ($ \pm 0.64$) & $0.65$ ($ \pm 0.15$) & $0.87$ ($ \pm 0.09$) & $0.83$ ($ \pm 0.18$) & $0.89$ ($ \pm 0.11$) & $0.76$ ($ \pm 0.13$) & $0.44$ ($ \pm 0.31$)\\
& Hidden Replay (ER $+$ RaR $+$ GR) & $-6.13$ ($ \pm 4.64$) & $1.85$ ($ \pm 0.88$) & $0.87$ ($ \pm 0.13$) & $0.91$ ($ \pm 0.09$) & $0.86$ ($ \pm 0.13$) & $0.89$ ($ \pm 0.13$) & $0.82$ ($ \pm 0.15$) & $0.55$ ($ \pm 0.26$)\\
\hline
& Random Policy & $-0.03$ ($ \pm 0.28$) & $1.01$ ($ \pm 0.02$) & $1.01$ ($ \pm 0.02$) & $0.54$ ($ \pm 0.02$) & $0.31$ ($ \pm 0.02$) & $0.3$ ($ \pm 0.02$) & $0.31$ ($ \pm 0.00$) & $0.3$ ($ \pm 0.04$)\\
& Baseline VTrace & $-3.41$ ($ \pm 1.7$) & $1.19$ ($ \pm 0.08$) & $1.17$ ($ \pm 0.29$) & $1.14$ ($ \pm 0.06$) & $0.64$ ($ \pm 0.06$) & $0.74$ ($ \pm 0.06$) & $0.66$ ($ \pm 0.06$) & $0.49$ ($ \pm 0.06$)\\
Condensed & Hidden Replay(ER) & $-8.65$ ($ \pm 3.15$) & $1.42$ ($ \pm 0.15$) & $1.12$ ($ \pm 0.11$) & $1.12$ ($ \pm 0.09$) & $0.73$ ($ \pm 0.09$) & $0.89$ ($ \pm 0.07$) & $0.77$ ($ \pm 0.04$) & $0.55$ ($ \pm 0.07$)\\
& Hidden Replay(ER $+$ RaR) & $-3.53$ ($ \pm 1.67$) & $1.35$ ($ \pm 0.09$) & $1.07$ ($ \pm 0.13$) & $1.19$ ($ \pm 0.07$) & $0.8$ ($ \pm 0.04$) & $0.9$ ($ \pm 0.04$) & $0.81$ ($ \pm 0.04$) & $0.61$ ($ \pm 0.11$)\\
& Hidden Replay(ER $+$ GR) & $-9.17$ ($ \pm 2.8$) & $1.57$ ($ \pm 0.42$) & $1.21$ ($ \pm 0.33$) & $1.15$ ($ \pm 0.07$) & $0.75$ ($ \pm 0.04$) & $0.91$ ($ \pm 0.04$) & $0.79$ ($ \pm 0.02$) & $0.59$ ($ \pm 0.07$)\\
& Hidden Replay (ER $+$ RaR $+$ GR) & $-3.05$ ($ \pm 0.97$) & $1.42$ ($ \pm 0.06$) & $1.0$ ($ \pm 0.02$) & $1.17$ ($ \pm 0.06$) & $0.8$ ($ \pm 0.04$) & $0.9$ ($ \pm 0.04$) & $0.83$ ($ \pm 0.04$) & $0.59$ ($ \pm 0.06$)\\
\hline
\end{tabular}
}
\caption{Effect of different replay mechanisms during sleep on lifelong learning metrics on the \SC{} domain. Results are collected using the hidden replay model on twelve syllabi. We report means and 95\% confidence intervals. Hidden replay using only experience replay (ER) and generative replay (GR) is insufficient. Random replay (RaR) is needed to stablize the learning of the feature space, and combining all three types of replay yield the best results.}
\label{tab:replay_mechanism_ablation_sc2}
\end{table}
\subsection{Understanding the Effect of Different Model Architectures}
\label{app:archs_ablation}
In Tables \ref{tab:algo-minigrid} and \ref{tab:algo-M121518}, we compare the different replay architectures introduced in this paper with several simple benchmark models. We observe that the hidden replay model generally outperforms the sequential and two-headed architectures as well as the simple benchmark models.
\begin{table}[H]
\centering
\resizebox{1.0\textwidth}{!}{
\begin{tabular}{|c|c|c|c|c|c|c||c|c|c|c|}
\hline
Scenario & Agent & PM & FT & BT & RP & RR\textsubscript{$\Omega$} & RR\textsubscript{$\alpha$} & RR\textsubscript{$\sigma$} & RR\textsubscript{$\upsilon$} \\
\hline
& Random Policy & $0.37$ ($ \pm 0.81$) $\bullet$ & $1.06$ ($ \pm 0.02$) $\circ$ & $1.13$ ($ \pm 0.18$) $\bullet$ & $0.22$ ($ \pm 0.04$) $\circ$ & $0.09$ ($ \pm 0.02$) $\circ$ & $0.09$ ($ \pm 0.02$) $\circ$ & $0.09$ ($ \pm 0.02$) $\circ$ & $0.09$ ($ \pm 0.00$) $\circ$\\
& Baseline PPO & $-48.2$ ($ \pm 12.04$) & $4.07$ ($ \pm 0.87$) & $0.46$ ($ \pm 0.12$) & $2.17$ ($ \pm 0.61$) & $0.5$ ($ \pm 0.1$) & $0.78$ ($ \pm 0.08$) & $0.64$ ($ \pm 0.08$) & $0.16$ ($ \pm 0.02$)\\
Pairwise & Hidden Replay (ER) & $-45.25$ ($ \pm 13.54$) & $6.34$ ($ \pm 1.4$) & $0.51$ ($ \pm 0.26$) & $2.75$ ($ \pm 0.73$) & $0.47$ ($ \pm 0.06$) & $0.73$ ($ \pm 0.08$) & $0.6$ ($ \pm 0.06$) & $0.11$ ($ \pm 0.02$) $\circ$\\
& Sequential (ER + RaR + GR) & $13.58$ ($ \pm 9.06$) $\bullet$ & $4.01$ ($ \pm 1.02$) & $9.52$ ($ \pm 5.5$) $\bullet$ & $2.57$ ($ \pm 0.72$) & $0.45$ ($ \pm 0.08$) & $0.37$ ($ \pm 0.08$) $\circ$ & $0.41$ ($ \pm 0.08$) $\circ$ & $0.08$ ($ \pm 0.02$) $\circ$\\
& Two-Headed (ER + RaR + GR) & $6.73$ ($ \pm 6.19$) $\bullet$ & $4.24$ ($ \pm 0.88$) & $3.55$ ($ \pm 4.12$) $\bullet$ & $2.85$ ($ \pm 0.72$) & $0.49$ ($ \pm 0.1$) & $0.46$ ($ \pm 0.08$) $\circ$ & $0.48$ ($ \pm 0.08$) & $0.1$ ($ \pm 0.02$) $\circ$\\
& Hidden Replay (ER + RaR + GR) & $-15.77$ ($ \pm 10.09$) & $5.87$ ($ \pm 1.24$) & $1.21$ ($ \pm 0.55$) & $2.47$ ($ \pm 0.69$) & $0.6$ ($ \pm 0.06$) & $0.69$ ($ \pm 0.08$) & $0.64$ ($ \pm 0.06$) & $0.11$ ($ \pm 0.02$) $\circ$\\
\hline
& Random Policy & $-0.02$ ($ \pm 0.26$) $\bullet$ & $1.08$ ($ \pm 0.02$) $\circ$ & $1.08$ ($ \pm 0.02$) $\circ$ & $0.19$ ($ \pm 0.00$) $\circ$ & $0.09$ ($ \pm 0.00$) $\circ$ & $0.09$ ($ \pm 0.00$) $\circ$ & $0.09$ ($ \pm 0.00$) $\circ$ & $0.09$ ($ \pm 0.00$) $\circ$\\
& Baseline PPO & $-38.92$ ($ \pm 3.98$) & $4.72$ ($ \pm 0.75$) & $4.06$ ($ \pm 0.73$) & $1.44$ ($ \pm 0.2$) & $0.22$ ($ \pm 0.04$) & $0.64$ ($ \pm 0.04$) & $0.33$ ($ \pm 0.04$) & $0.15$ ($ \pm 0.02$)\\
Condensed & Hidden Replay (ER) & $-57.71$ ($ \pm 3.61$) & $6.45$ ($ \pm 0.81$) & $5.32$ ($ \pm 0.81$) & $2.04$ ($ \pm 0.1$) $\bullet$ & $0.17$ ($ \pm 0.02$) & $0.77$ ($ \pm 0.04$) $\bullet$ & $0.33$ ($ \pm 0.02$) & $0.12$ ($ \pm 0.02$)\\
& Sequential (ER + RaR + GR)& $-6.8$ ($ \pm 2.29$) $\bullet$ & $4.49$ ($ \pm 0.57$) & $1.98$ ($ \pm 0.35$) $\circ$ & $2.13$ ($ \pm 0.08$) $\bullet$ & $0.56$ ($ \pm 0.04$) $\bullet$ & $0.62$ ($ \pm 0.02$) & $0.58$ ($ \pm 0.02$) $\bullet$ & $0.12$ ($ \pm 0.02$)\\
& Two-Headed (ER + RaR + GR) & $-9.69$ ($ \pm 2.87$) $\bullet$ & $4.1$ ($ \pm 0.57$) & $1.65$ ($ \pm 0.41$) $\circ$ & $2.15$ ($ \pm 0.1$) $\bullet$ & $0.59$ ($ \pm 0.04$) $\bullet$ & $0.69$ ($ \pm 0.04$) & $0.62$ ($ \pm 0.04$) $\bullet$ & $0.13$ ($ \pm 0.02$)\\
& Hidden Replay (ER + RaR + GR) & $-21.94$ ($ \pm 2.09$) & $5.93$ ($ \pm 0.79$) & $1.32$ ($ \pm 0.2$) $\circ$ & $2.1$ ($ \pm 0.1$) $\bullet$ & $0.56$ ($ \pm 0.04$) $\bullet$ & $0.8$ ($ \pm 0.02$) $\bullet$ & $0.63$ ($ \pm 0.02$) $\bullet$ & $0.16$ ($ \pm 0.02$)\\
\hline
\end{tabular}
}
\caption{Comparison of the different model-free generative replay architectures and the baseline on two different Minigrid scenarios. The metrics are averaged over 36 lifetimes per scenario and the 95\% confidence interval is shown. $\bullet$ denotes statistically significant improvement compared to the baseline with a Kruskal-Wallis test followed by Dunn's test with Bon Ferroni correction w/ a p-value of 0.05. $\circ$ denotes the metric is statistically significantly worse than the baseline. We see that the LL agents significantly outperform the non-LL agent interms of performance maintainence and RR\textsubscript{$\Omega$}, suggesting the LL agents significantly reduce catastrophic forgetting. The hidden replay model achieves significantly higher RP and RR\textsubscript{$\alpha$} suggesting that it enables the model to learn new tasks with minimal interference from learning previous tasks.}
\label{tab:algo-minigrid}
\end{table}
\begin{table}[H]
\centering
\resizebox{1.0\textwidth}{!}{
\begin{tabular}{|c|c|c|c|c|c||c|c|c|c|}
\hline
Scenario & Agent & PM & FT & BT & RP & RR\textsubscript{$\Omega$} & RR\textsubscript{$\alpha$} & RR\textsubscript{$\sigma$} & RR\textsubscript{$\upsilon$} \\
\hline
& Random Policy & $0.0$ ($ \pm 0.00$) & $1.0$ ($ \pm 0.00$) & $1.0$ ($ \pm 0.00$) & $0.56$ ($ \pm 0.04$) $\circ$ & $0.33$ ($ \pm 0.06$) $\circ$ & $0.33$ ($ \pm 0.06$) $\circ$ & $0.33$ ($ \pm 0.06$) $\circ$ & $0.34$ ($ \pm 0.14$)\\
& Baseline VTrace (No Sleep) & $-8.2$ ($ \pm 6.54$) & $1.11$ ($ \pm 0.23$) & $0.85$ ($ \pm 0.21$) & $1.03$ ($ \pm 0.06$) & $0.74$ ($ \pm 0.14$) & $0.82$ ($ \pm 0.12$) & $0.78$ ($ \pm 0.12$) & $0.39$ ($ \pm 0.17$)\\
Pairwise & Sequential (ER $+$ GR)& $-5.51$ ($ \pm 3.41$) & $1.18$ ($ \pm 0.36$) & $0.77$ ($ \pm 0.13$) & $0.89$ ($ \pm 0.06$) & $0.54$ ($ \pm 0.15$) & $0.61$ ($ \pm 0.15$) & $0.57$ ($ \pm 0.15$) & $0.47$ ($ \pm 0.24$)\\
& Two-Headed (ER $+$ GR)& $-2.35$ ($ \pm 7.31$) & $1.46$ ($ \pm 0.48$) & $0.98$ ($ \pm 0.18$) & $0.95$ ($ \pm 0.11$) & $0.76$ ($ \pm 0.18$) & $0.79$ ($ \pm 0.18$) & $0.77$ ($ \pm 0.18$) & $0.41$ ($ \pm 0.2$)\\
& Hidden Replay (ER Only) & $-7.9$ ($ \pm 5.11$) & $1.14$ ($ \pm 0.4$) & $0.75$ ($ \pm 0.29$) & $0.92$ ($ \pm 0.09$) & $0.87$ ($ \pm 0.11$) & $0.9$ ($ \pm 0.11$) & $0.8$ ($ \pm 0.13$) & $0.37$ ($ \pm 0.2$)\\
& Hidden Replay (ER $+$ RaR $+$ GR) & $-0.57$ ($ \pm 2.93$) & $1.85$ ($ \pm 0.88$) & $0.95$ ($ \pm 0.15$) & $1.04$ ($ \pm 0.11$) & $0.77$ ($ \pm 0.18$) & $0.82$ ($ \pm 0.13$) & $0.8$ ($ \pm 0.15$) & $0.55$ ($ \pm 0.26$)\\
\hline
& Random Policy & $0.0$ ($ \pm 0.00$) $\bullet$ & $1.0$ ($ \pm 0.00$) & $1.0$ ($ \pm 0.00$) $\bullet$ & $0.42$ ($ \pm 0.04$) $\circ$ & $0.32$ ($ \pm 0.09$) $\circ$ & $0.32$ ($ \pm 0.09$) $\circ$ & $0.32$ ($ \pm 0.09$) $\circ$ & $0.32$ ($ \pm 0.22$)\\
& Baseline VTrace (No Sleep) & $-8.99$ ($ \pm 3.06$) & $1.11$ ($ \pm 0.23$) & $0.79$ ($ \pm 0.12$) & $0.92$ ($ \pm 0.04$) & $0.9$ ($ \pm 0.14$) & $0.91$ ($ \pm 0.12$) & $0.82$ ($ \pm 0.12$) & $0.39$ ($ \pm 0.17$)\\
Alternating & Sequential (ER $+$ GR)& $-7.44$ ($ \pm 3.43$) & $1.18$ ($ \pm 0.36$) & $0.88$ ($ \pm 0.09$) & $0.8$ ($ \pm 0.09$) & $0.59$ ($ \pm 0.15$) & $0.66$ ($ \pm 0.13$) & $0.6$ ($ \pm 0.13$) & $0.47$ ($ \pm 0.24$)\\
& Two-Headed (ER $+$ GR)& $-7.7$ ($ \pm 4.89$) & $1.46$ ($ \pm 0.48$) & $0.91$ ($ \pm 0.11$) & $0.9$ ($ \pm 0.09$) & $0.88$ ($ \pm 0.22$) & $0.94$ ($ \pm 0.11$) & $0.84$ ($ \pm 0.15$) & $0.41$ ($ \pm 0.2$)\\
& Hidden Replay (ER Only) & $-7.9$ ($ \pm 5.11$) & $1.14$ ($ \pm 0.4$) & $0.75$ ($ \pm 0.29$) & $0.92$ ($ \pm 0.09$) & $0.87$ ($ \pm 0.11$) & $0.9$ ($ \pm 0.11$) & $0.8$ ($ \pm 0.13$) & $0.37$ ($ \pm 0.2$)\\
& Hidden Replay (ER $+$ RaR $+$ GR) & $-6.13$ ($ \pm 4.64$) & $1.85$ ($ \pm 0.88$) & $0.87$ ($ \pm 0.13$) & $0.91$ ($ \pm 0.09$) & $0.86$ ($ \pm 0.13$) & $0.89$ ($ \pm 0.13$) & $0.82$ ($ \pm 0.15$) & $0.55$ ($ \pm 0.26$)\\
\hline
& Random & $-0.03$ ($ \pm 0.28$) $\bullet$ & $1.01$ ($ \pm 0.02$) & $1.01$ ($ \pm 0.02$) & $0.54$ ($ \pm 0.02$) $\circ$ & $0.31$ ($ \pm 0.02$) $\circ$ & $0.3$ ($ \pm 0.02$) $\circ$ & $0.31$ ($ \pm 0.00$) $\circ$ & $0.3$ ($ \pm 0.04$) $\circ$\\
& Baseline VTrace (No Sleep) & $-3.41$ ($ \pm 1.7$) & $1.19$ ($ \pm 0.08$) & $1.17$ ($ \pm 0.29$) & $1.14$ ($ \pm 0.06$) & $0.64$ ($ \pm 0.06$) & $0.74$ ($ \pm 0.06$) & $0.66$ ($ \pm 0.06$) & $0.49$ ($ \pm 0.06$)\\
Condensed & Sequential (ER $+$ GR) & $-3.7$ ($ \pm 1.58$) & $1.15$ ($ \pm 0.04$) & $1.0$ ($ \pm 0.09$) & $0.91$ ($ \pm 0.09$) $\circ$ & $0.53$ ($ \pm 0.13$) & $0.69$ ($ \pm 0.09$) & $0.61$ ($ \pm 0.09$) & $0.49$ ($ \pm 0.11$)\\
& Two-Headed (ER $+$ GR) & $-4.53$ ($ \pm 2.97$) & $1.34$ ($ \pm 0.09$) & $1.15$ ($ \pm 0.11$) & $1.09$ ($ \pm 0.13$) & $0.76$ ($ \pm 0.11$) & $0.87$ ($ \pm 0.09$) & $0.78$ ($ \pm 0.09$) & $0.57$ ($ \pm 0.09$)\\
& Hidden Replay (ER Only) & $-8.65$ ($ \pm 3.15$) $\circ$ & $1.42$ ($ \pm 0.15$) $\bullet$ & $1.12$ ($ \pm 0.11$) & $1.12$ ($ \pm 0.09$) & $0.73$ ($ \pm 0.09$) & $0.89$ ($ \pm 0.07$) & $0.77$ ($ \pm 0.04$) & $0.55$ ($ \pm 0.07$)\\
& Hidden Replay (ER $+$ RaR $+$ GR) & $-3.05$ ($ \pm 0.97$) & $1.42$ ($ \pm 0.06$) $\bullet$ & $1.0$ ($ \pm 0.02$) & $1.17$ ($ \pm 0.06$) & $0.8$ ($ \pm 0.04$) & $0.9$ ($ \pm 0.04$) & $0.83$ ($ \pm 0.04$) $\bullet$ & $0.59$ ($ \pm 0.06$)\\
\hline
\end{tabular}
}
\caption{Comparison of the different model-free generative replay architectures and the baseline on three different \SC{} scenarios. The metrics are averaged over 12 lifetimes per scenario and the 95\% confidence interval is shown. $\bullet$ denotes statistically significant improvement compared to the baseline with a Kruskal-Wallis test followed by Dunn's test with Bon Ferroni correction w/ a p-value of 0.05. $\circ$ denotes the metric is statistically significantly worse than the baseline. In general, the hidden replay model performs very well, showing significant gains in FT. Interestingly, the baseline agent performs well in learning new tasks, but the continual learners exhibit better performance on unseen tasks, indicating that they exhibit better generalization. RR\textsubscript{$\Omega$} also shows the continual learners are able to learn without forgetting.}
\label{tab:algo-M121518}
\end{table}
\subsection{Qualitative Analysis of the Generative Model}
\label{sec:reconstructions_app}
In this section, we qualitatively evaluate the capabilities of each replay architecture with respect to its generative modeling capabilities.
In Fig.~\ref{fig:recon_sequential_collect}, we show reconstructions of the first channel of the ``CollectMineralShards'' SC-2 task using the sequential architecture after 11K sleep batches, imitating a converged wake model. These observations are challenging to reconstruct and generate due to the random distribution of shards, each represented by a few pixels. The sequential architecture produces blurry reconstructions that resemble density maps over the shard locations. In Fig.~\ref{fig:recon_two_headed_collect}, we show reconstructions on the same task for the two-headed architecture. By separating the generative model from the policy network, the model reconstructs the observations with sharper detail; however, the two-headed model still struggles with generating realistic observations, generating observations that degrade to density maps. While the generated observations are sub-optimal in terms of realism, experiments suggest they still provide useful information for training the policy network. We show additional ground-truth, reconstructions, and generated samples for the ``DefeatRoaches'' task using the two-headed architecture in Fig.~\ref{fig:recon_two_headed_defeat}. In this case, there is much more structural regularity in the observations, so the reconstructions and generated samples better resemble ground truth data.
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/m12_reconstructions.png}
\caption{Reconstructions of the first channel of the ``CollectMineralShards'' SC-2 task using the sequential architecture after 11,000 sleep batches, imitating a converged wake model.}
\label{fig:recon_sequential_collect}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/m15_reconstructions.png}
\caption{Reconstructions of the first channel of the ``CollectMineralShards'' SC-2 task using the two-headed architecture after 11,000 sleep batches, imitating a converged wake model. We also show some generated samples and the effect of passing generating via the ``loop-vae'' procedure.}
\label{fig:recon_two_headed_collect}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/m15_reconstructions_defeat.png}
\caption{Reconstructions and generated samples of the first channel of the ``DefeatRoaches'' SC-2 task using the two-headed architecture.}
\label{fig:recon_two_headed_defeat}
\end{figure}
Visualizing the reconstructions and generated samples of the hidden replay model is more challenging. To do so, we use a two-dimensional principle components analysis to compare ``ground-truth'' features (i.e., the output of the feature extractor) with reconstructions and generated features. On the left side of Fig.\ \ref{fig:recon_hidden_replay}, the VAE is able to closely reconstruct the ground-truth features for the ``DefeatRoaches'' task. On the right side, we overlay the feature distribution of ``DefeatZerglingsAndBanelings'' and the generated samples for ``DefeatRoaches'' after learning the task sequence ``DefeatRoaches'' $ \rightarrow$ ``DefeatZerglingsAndBanelings''. Since these tasks are similar, we see an overlap between the generated features of ``DefeatRoaches'' from the previous sleep with the new features of ``DefeatZerglingsAndBanelings'', suggesting the generator is generating meaningful features from the task seen during the previous sleep phase.
\begin{figure}[H]
\centering
\includegraphics[width=0.7\textwidth]{graphics/m18_reconstructions_and_samples.png}
\caption{To evaluate reconstructions and generated samples for the hidden replay architecture, we project data in the feature space via its principle components. Left: Reconstructions of the features for the ``DefeatRoaches'' task. Right: Overlaying the feature distribution of ``DefeatZerglingsAndBanelings'' and the generated samples for ``DefeatRoaches'' after learning the task sequence ``DefeatRoaches'' $ \rightarrow$ ``DefeatZerglingsAndBanelings''.}
\label{fig:recon_hidden_replay}
\end{figure}
\subsection{Additional Results: Understanding the Interplay Between Different Replay Mechanisms}
\label{sec:justification_of_exemplars}
We've observed that generative replay alone is not sufficient for the hidden replay architecture. This is because there is no constraint limiting the drift in the feature space between sleeps, so the feature extractor can completely change between two time steps if the tasks are perceptually dissimilar or dissimilar in policy. To help to overcome this issue, we employ ``random replay'' where we save a small batch of random observation-action pairs in the original observation space after every sleep, and randomly replay them during every sleep. In Fig.\ \ref{fig:exemplar_replay_1}, we show the necessity of combining hidden replay with random replay. On the left, we see that using hidden replay without random replay results in the feature space for task 0 dramatically changing after training on task 1, and PM is low. On the right, we see that the distribution of task 0 before and after training on task 1 is significantly better aligned when hidden replay is used in conjunction with random replay, and PM significantly improves. Additional visualizations for pairs of dissimilar tasks are seen in Fig.~\ref{fig:exemplar_replay_2}. In Table \ref{tab:exemplars_vs_no_exemplars}, we see using random replay generally improves PM and FTR, which agrees with our hypothesis that it helps to maintain the feature space, leading to less forgetting and potentially features that generalize better to other tasks. These results are inline with similar observations made from the main experiments of this paper.
\begin{table}[H]
\centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Scenario & Agent & PM & FT & BT & RP \\
\hline
Pairwise & No Random Replay & $-6.28$ ($ \pm 12.82$) & $1.35$ ($ \pm 0.24$) & $0.80$ ($ \pm 0.37$) & $1.08$ ($ \pm 0.19$) \\
& Random Replay & $-3.03$ ($ \pm 4.92$) & $1.65$ ($ \pm 0.28$) & $0.92$ ($ \pm 0.13$) & $1.04$ ($ \pm 0.17$) \\
\hline
Condensed & No Random Replay & $-9.35$ ($ \pm 3.69$) & $1.37$ ($ \pm 0.15$) & $1.20$ ($ \pm 0.24$) & $1.16$ ($ \pm 0.16$) \\
& Random Replay & $-3.05$ ($ \pm 1.76$) & $1.42$ ($ \pm 0.11$) & $1.00$ ($ \pm 0.03$) & $1.17$ ($ \pm 0.11$) \\
\hline
\end{tabular}
\caption{Comparison of the hidden replay model with and without random replay. Results are averaged over multiple lifetimes, and we also record the standard deviation of the metrics.}
\label{tab:exemplars_vs_no_exemplars}
\end{table}
\begin{figure}[H]
\centering
\includegraphics[width=0.8\textwidth]{graphics/m18_exemplars.png}
\caption{We show the necessity of combining hidden replay with random replay. Left: Using hidden replay without random replay results in the feature space for task 0 dramatically changing after training on task 1, and the feature space for both classes converge to a point. PM is -10.68. Right: The distribution of task 0 before and after training on task 1 are significantly better aligned when hidden replay is used in conjunction with random replay. PM is -1.93.}
\label{fig:exemplar_replay_1}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{graphics/m18_exemplars_2.png}
\caption{We show two examples where hidden + random replay maintains the feature space of the previous task after learning a new dissimilar SC-2 task (``Collect'' $\rightarrow$ ``Defeat'' and vice versa).}
\label{fig:exemplar_replay_2}
\end{figure}
\subsubsection*{Acknowledgments}
This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under the Lifelong Learning Machines (L2M) program Contract No. HR0011-18-C-0051. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the Defense Advanced Research Projects Agency (DARPA).
Special thanks to Roberto Corizzo, Kamil Faber, Nathalie Japkowicz, Michael Baron, James Smith, Sahana Pramod Joshi, Zsolt Kira, Cameron Ethan Taylor, Mustafa Burak Gurbuz, Constantine Dovrolis, Tyler L. Hayes, Christopher Kanan, and Jhair Gallardo for their useful discussions and feedback related to system design.
\section{Discussion}
Our experiments demonstrate LRL via positive FT and BT, high performance relative to an STE, reduced catastrophic forgetting, and fast adaptation to new tasks. Our work clearly advances the domain-specific state-of-the-art in \SC{} by reducing the sample complexity of deep RL to master the minigames.
Compared to STE, our experiments required only 6\% of training samples to achieve 80-90\% of expert performance.
Existing lifelong RL implementations (e.g., \citep{mendez2020lifelong}) have been tested on benchmark problems, but proved to be challenging to scale up to the complex observation and action space of SC-2. Future work can leverage our techniques in the full SC-2 game and reduce the astronomical sample complexity of current deep RL agents in SC-2 (e.g., AlphaStar \citep{vinyals2019grandmaster}).
While the wake-sleep mechanism and model-free GR show promising results, there is room for improvement, and the approach is easily extensible. We highlight several potential extensions. Our current model triggers sleep on a set schedule. Instead, it is possible to self-trigger sleep via recent work in lifelong changepoint detection \citep{faber2021watch}. Currently, we store full observations in the replay buffer, which due to memory constraints, limits the number of samples that can be saved. It would be prudent to intelligently compress experiences in the replay buffer \citep{riemer2019scalable, wang2021acae, caccia2020online,zhang2017deeper,schaul2015prioritized, hayes2019remind}. Past work has looked at improving replay buffers via prioritized replay \citep{schaul2015prioritized} based on detecting dead-end states.
We further hypothesize that increasing the lifetime of an agent by avoiding dead end states is a useful bias. Our methods operate over hand-crafted representations of the observation space; the system could be extended by introducing representation learning from RGB-observations via unsupervised (STAM \citep{smith2019unsupervised,smith2019unsupervised2}) and self-supervised learning \citep{gallardo2021self}. We can further extend our models using mixtures of VAEs (e.g., the Eigentask framework \cite{raghavan2020lifelong}) to dynamically grow and shrink the model, leading towards agents that build implicit models of the tasks they experience. Finally, we used a fixed neural architecture and loss functions throughout the agent's lifetime. GR is a flexible approach that can include progress in other areas of lifelong learning, e.g., dynamically changing the underlying neural architectures \citep{schwarz2018progress}, neuromodulation \citep{brna2019uncertainty}, and regularization \citep{kolouri2019sliced}.
Further, interesting avenues include the application of GR to multi-agent RL \citep{nekoei2021continuous} and federated learning \citep{rostami2017multi}, where communication between agents can enable better coordination between agents and collective acquisition of knowledge. Due to the inherent plastic-yet-stable nature of lifelong learners, our methods can be more robust to data drift, concept drift, or task changes, reducing system downtime and the overhead of retraining, and used to address model obsolescence and reduce the technical debt of machine learning models \citep{sculley2015hidden, alahdab2019empirical, bogner2021characterizing}.
\section{Experiments}
\label{sec:experiments}
\subsection{Evaluation scenarios}
\label{sec:scenarios}
A lifelong learning \textit{scenario} describes the order of task presentation and repetition of tasks. A concrete instantiation of a scenario with tasks (or POMDPs) is called a \textit{syllabus}. Each scenario is used to evaluate the learning agent over all syllabi and average metrics. The three scenarios we consider are depicted in Figure~\ref{fig:scenarios}.
\textbf{Pairwise:} Two tasks are presented once each. This scenario evaluates knowledge transfer between two tasks. \textbf{Alternating:} Two tasks are presented with three repetitions of each pair. This scenario evaluates knowledge retention over longer lifetimes. \textbf{Condensed:} A random permutation of all tasks is presented.
This is a useful setting for understanding how an agent performs across different task orderings.
Given a syllabus, an experiment run alternates between \emph{evaluation blocks} (EBs) and \emph{learning blocks} (LBs) \citep[see][]{new2022lifelong}.
We say that a task is \emph{seen} wrt an EB if it has appeared in any LB preceeding it, otherwise it is \emph{unseen}.
During each EB, the average accumulated reward of the agent is evaluated on all tasks in the syllabus (including unseen tasks); during each LB, the agent learns on a single task for a fixed number of interactions.
The STE for each task serves as a baseline and measures the relative performance of the learner wrt asymptotic optimal.
\subsection{Metrics for Lifelong RL}
We run controlled and quantitative experiments and compare to multiple baselines along three sets of metrics.
First, we compare learning behavior by examining the accumulated rewards captured by periodic evaluations of the agent, i.e., we look at the trend of EB rewards averaged over EBs (and tasks).
Second, we aim to understand how the rewards achieved by an agent compares to
a single task expert (STE).
This forms the basis for our second set of metrics: \emph{relative reward (RR)}.
Note that these metrics also focus on
EBs and do not consider
LBs. Finally, we compare algorithms using the lifelong learning metrics defined by \citep{new2022lifelong}, which take into account both LBs and EBs, and aim to frame metrics in a way that are agnostic to the domain.
We begin by describing the RR metrics.
For each task, we train (offline) the STE
using an off-the-shelf deep RL algorithm.
RR metrics measure how well the agent has learned the seen tasks, how well the agent performs over all tasks in the syllabus, and how well the agent generalizes to unseen tasks.
In general, RR is computed as
\begin{align}
RR(\pi,T,S) &= \frac{1}{|T|}\sum_{t \in T}\frac{1}{|S(t)|}\sum_{s \in S(t)} \frac{r_\pi(s)}{r_{\pi^*_s}(s)}
\label{eq:rr}
\end{align}
where $\pi$ is the policy of the learner, $T$ is a set EBs, $S(t)$ is the set of tasks at EB $t$, $\pi^*_s$ is the STE for task $s$, and $r_{\pi}(s)$ is the accumulated reward (return) achieved on task $s$ from using policy $\pi$. Different metrics are defined by setting $S, T$ to different EBs and tasks.
$\mathbf{RR_\Omega}$: $T$ only includes the final EB, and $S(t)$ is the set of seen tasks in the syllabus. $RR_\Omega$ measures how well the agent has learned the seen tasks in the syllabus ``at the end of the day''. To measure catastrophic forgetting, we denote $\mathbf{RR_\sigma}$ where
$T$ includes all EBs, and $S(t)$ is the set of seen tasks in the syllabus encountered in previous LBs up to and including EB $t$. $\mathbf{RR_\upsilon}$ is complementary in that it measures how well the agent generalizes to unseen tasks, $T$ includes all EBs, and $S(t)$ is the set of unseen tasks not encountered in previous LBs up to and including EB $t$. To measure plasticity, we denote $\mathbf{RR_\alpha}$ where $T$ is the set of all EBs, and $S(t)$ is only the task learned in LB $t$ preceeding EB $t$.
Note that in all cases, higher more-positive values are better for all metrics as all our STEs achieve positive return.
As mentioned, we employ established lifelong learning metrics defined by \citet{new2022lifelong}, which capture catastrophic forgetting, transfer between all pairs of tasks, and speed of learning. In all cases, more-positive values are better for all metrics.
\textbf{Forward Transfer Ratio (FTR)} is the ratio of EB performance on an unseen task. \textbf{Backward Transfer Ratio (BTR)} is the ratio of EB performance on a seen task. A value greater than one indicates positive transfer. \textbf{Relative Performance (RP)} is the ratio of the Area Under the Learning Curve (AUC) between the lifelong learner and STE learning curve. A value greater than one indicates either faster learning by the lifelong learner and/or superior asymptotic performance. \textbf{Performance Maintenance (PM)} is the average difference between task performance in any EB that comes after the LB of a seen task.
A value less than $0$ indicates forgetting. Refer to \citep{new2022lifelong} for more details about the metrics and the precise computation.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{graphics/alternating_curves.png}
\caption{Illustrative examples showing EB rewards of baseline and hidden replay in alternating tasks scenario. Column color denotes which task is trained in LBs between EBs. Left: Example syllabus of two similar tasks; the lifelong learner (LL) performs similarly to the baseline.
Right: Examples of two dissimilar tasks where the baseline exhibits catastrophic forgetting, our hidden replay does not.}
\label{fig:dissimilar_alternating}
\vspace{-0.75cm}
\end{figure}
\subsection{Domains}
Our experiments use two RL domains: \SC{} minigames \citep{vinyals2017starcraft} and Minigrid \citep{gym_minigrid}. We evaluate on existing RL tasks (POMDPs) in these domains.
\textbf{Starcraft 2}:
We selected 3 minigames namely \emph{DefeatRoaches}, \emph{DefeatZerglingsAndBanelings}, and \emph{CollectMineralShards}. We created an additional task in each case as described in Appendix ~\ref{sec:sc2_task_descriptions}. We used PySC2 \citep{vinyals2017starcraft} to interface with SC-2 game.
The agent receives positive rewards for collecting resources and defeating enemies, and negative rewards for losing friendly units.
As observations, we used the unit type, selection status, and unit density two-dimensional inputs.
We follow the action space and convolutional policy network architecture defined in \citet{vinyals2017starcraft}.
Single task experts were trained to convergence ($\sim$ 30M environment steps using the VTrace algorithm \citep{espeholt2018impala}). We consider the pairwise, alternating, and condensed scenarios where each LB consists of 2 million environment steps (about 6\% of the STE samples), and each EB consists of 30 episodes per task. Complete details can be found in the Appendix \ref{sec:imp_sc2}.
\textbf{Minigrid}:
Due to the heavy computational load of experimentation in SC-2, we also show results in the Minigrid domain \citep{gym_minigrid}. We use the TELLA framework \citep{lifelong-learning-systems} to orchestrate the experiments ($\sim$100 repetitions per scenario). Tasks involve an agent that must navigate to a goal while interacting with various entities in the environment. We consider 5 tasks --- SimpleCrossing, DistShift, a custom Fetch, a custom Unlock, and DoorKey --- each with two variants (differing in the size of the grid and/or number of objects) for a total of 10 POMDPs. Descriptions of these tasks can be found in \citet{gym_minigrid}. An observation is a $7\times7$ top-down view of the grid directly in front of the agent, with 3 channels containing information about the object type, object color, and object state in each tile. Rewards are sparse: $-1$ for running into obstacles and lava, $0$ for not reaching the goal, and a reward in $[0, 1]$ for reaching the goal proportional to the number of steps taken. There are only six actions: turn left, turn right, move forward, pick up an object, drop an object, and interact with an object. STEs are trained on each task for 1M environment steps using PPO \citep{schulman2017proximal}. We consider the pairwise and condensed scenarios where the LB lengths vary between 100k and 700k (Appendix~\ref{sec:minigrid_task_lengths}) based on examining the smallest number of steps for the STE to converge to $80\%$ of its peak performance averaged over ten runs. Each EB consists of 100 episodes per task. Further details can be found in Appendix~\ref{sec:imp_minigrid}.
\textbf{Baseline Agents}:
We compare against several non-lifelong baseline agents: 1) Random policy; 2) \emph{Baseline agent} using off-the-shelf deep RL trained on tasks in sequence without any lifelong learning. We used the VTrace algorithm \citep{espeholt2018impala} for SC-2 and PPO \citep{schulman2017proximal} for the Minigrid experiments.
\textbf{Ablations}: In order to compare different replay type, we perform ablations against the combination of buffers used for replay (see Fig. \ref{fig:all_replay_archs}): 1) \emph{Hidden Replay (ER)}, 2) \emph{Hidden Replay (ER+RaR)}, 3) \emph{Hidden Replay (ER+GR)}, 4) \emph{Hidden Replay (ER+RaR+GR)}. In each case, the VAE, sleep policy, and wake policy are trained, but on different augmented datasets as shown in Alg.~\ref{alg:hidden_replay} (samples are drawn from VAE only if GR is enabled).
Similarly, we perform ablations along architectures discussed in Fig.~\ref{fig:all_replay_archs}.
\textbf{Single Task Experiments}:
We conducted experiments that demonstrated that the wake-sleep training phases successfully imitates the STE. These can be found in the Appendix \ref{sec:ste_additional_results} for SC-2 and Minigrid.
\subsection{Evidence of Overcoming Catastrophic Forgetting}
We observed that different syllabi and scenarios can exhibit different levels of catastrophic forgetting due to varying degrees of similarity between the tasks in the syllabus. We first illustrate this issue before showing large scale results averaging over all syllabi.\footnote{We attempt to quantify pairwise similarity between tasks in Appendix \ref{sec:imp_sc2} by evaluating the forward transfer of the STE for each task to all other tasks. Quantifying task similarity remains an open problem.}
In Figure \ref{fig:dissimilar_alternating}, we illustrate examples that
1) the LL agent and non-LL agent perform similarly when syllabi are composed of similar tasks, and both agents exhibit positive FT and BT, and 2) the non-LL agent experiences catastrophic forgetting (zig-zag pattern) of one or both tasks whereas the LL agent shows significantly less forgetting \emph{when transitioning between two dissimilar tasks}.
We note that in the \SC{} domain, evidence of extreme catastrophic forgetting is somewhat rare, a possible weakness in the design of our task set. Nevertheless, LL agents clearly outperform non-LL on almost all the other metrics, signifying the need to look beyond catastrophic forgetting when evaluating LL.
\subsection{Comparison of Different Types of Replay for Sleep}
\begin{figure*}[t]
\centering
\includegraphics[width=.98\textwidth]{graphics/minigrid_replay.png}
\caption{Evaluating different replay types on the hidden replay model for the Minigrid domain. Error bars denote 95\% confidence interval.}
\label{fig:minigrid_replay_ablations_main}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=0.98\textwidth]{graphics/sc2_replay.png}
\caption{Evaluating different replay types on the hidden replay model for the \SC{} domain. Error bars denote 95\% confidence interval.}
\label{fig:sc2_replay_ablations_main}
\end{figure*}
To understand the effect of different replay types on the hidden replay model: experience replay from the wake buffer (ER), generative replay (GR), and random replay (RaR). All approaches still train both the policy network and generator, even if GR is disabled. We summarize the important findings in Fig.~\ref{fig:minigrid_replay_ablations_main} and Fig.~\ref{fig:sc2_replay_ablations_main} \footnote{Further experiments can be found in Appendix \ref{app:replay_ablation}}.
\textbf{Transfer Learning and Generalization to Unseen Tasks}:
Looking at the pairwise scenario in Minigrid and \SC{}, we see that models that use hidden replay with ER+GR, ER+RaR or ER+RaR+GR show much higher FT compared to the baseline. Hidden replay with ER+GR, ER+RaR or ER+RaR+GR show higher $RR_\upsilon$ and the effect is more pronounced for \SC{} condensed scenario, showing generalization to unseen tasks.
\textbf{Mitigating Catastrophic Forgetting}:
Looking at PM across all scenarios in Minigrid and \SC{} and looking at $RR_\Omega$ of the Minigrid condensed scenario, we validate that the hidden replay model requires stabilization with RaR in order to (significantly) reduce catastrophic forgetting, compared to ER+GR alone which is typically worse than ER+RaR and ER+RaR+GR.
\textbf{Adaptation to New Tasks}: Interestingly, we find that $RR_\alpha$ for the condensed scenarios of Minigrid and \SC{}, shows that simply performing distillation during sleep phases produces
a beneficial ``jumpstart'' effect when learning new tasks, compared to the baseline which is also able to learn the new task but requires more learning experiences.
\textbf{Backward Transfer}: The BT metric seems to vary across the board based on the type of replay used. This is in part because when computing the BT metric from task 2 to task 1, it is very sensitive to how well the model learns on task 1. If the model learns task 1 to near optimal performance, it is very difficult to achieve $BT>1$ whereas if task 1 is not well-learned, even a weak model on a similar task 2 can potentially easily achieve $BT>1$. However, if we look at $RR_\sigma$ instead, we see a different story where the models that learn best (ER+RaR and ER+RaR+GR) often exhibit the best performance on seen tasks in opposition to the BT metric.
\textbf{General Observations}:
Adding a small amount of random replay (ER+RaR+GR) fixes many issues with standard generative hidden replay, producing noticeably better results. We find that ER+RaR, even without GR, achieves significant improvement over the baseline agent, and the model further improves when all types of replay (ER+RaR+GR) are combined, albeit by a small amount. This suggests that RaR of a small number of samples is often powerful and
complementary (i.e., often helpful, rarely hurtful) to GR (n.b.\! all methods train the VAE, but only GR samples from it). This leads us to suspect that the replay architecture is more important than the sampling strategy that we investigate next.
\subsection{Comparison of Different Architectures for Sleep}
\begin{figure*}[t]
\centering
\includegraphics[width=0.98\textwidth]{graphics/minigrid_archs.png}
\caption{Evaluating the performance of different architectures for the Minigrid domain. Error bars denote 95\% confidence interval.}
\label{fig:minigrid_ablations_archs_main}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=0.98\textwidth]{graphics/sc2_archs.png}
\caption{Evaluating the performance of different architectures for the \SC{} domain. Error bars denote 95\% confidence interval.}
\label{fig:sc2_ablations_archs_main}
\vspace{-0.5cm}
\end{figure*}
We aim to understand the strengths and weaknesses of the model architectures introduced in this paper (Sequential, Two-Headed, Hidden Replay) and understand how these models improve over simple baselines.
We summarize the important findings in Fig.~\ref{fig:minigrid_ablations_archs_main} and Fig.~\ref{fig:sc2_ablations_archs_main} \footnote{Further experiments can be found in Appendix \ref{app:archs_ablation})}.
\textbf{Transfer Learning and Generalization to Unseen Tasks}: Forward transfer seems to be a strength of the hidden replay model. Looking at FT across all domains and scenarios and looking at the $RR_\upsilon$ of the \SC{} scenarios, we see that the hidden replay model noticeably outperforms the other models.
\textbf{Avoiding Catastrophic Forgetting}:
Looking at PM and $RR_\Omega$ across all scenarios, the hidden replay model often outperforms the other models.
\textbf{Adaptation to New Tasks}: We see that the hidden replay model achieves consistently high (and often the best) RP and $RR_\alpha$ across most scenarios in both domains. This suggests that the hidden replay model is not interfering with new tasks and may be providing a helpful ``jumpstart'' effect.
\textbf{Positive Backward Transfer}: As before, it can be difficult to interpret the BT metric. In general, the BT of the hidden replay model is on par with other models that achieve comparable terminal performance, and the hidden replay model shows strong performance in terms of $RR_\sigma$.
\textbf{Sample Efficiency}
Compared to STE, our experiments using the hidden replay model required only 6\% of training samples to achieve 80-90\% of expert performance in most \SC{} scenarios.
\textbf{General Observations}: In general, across both domains and many scenarios, the hidden replay model seems to outperform the other model architectures tested in most of the metrics of interest. This suggests that the hidden replay model is a potentially powerful model for addressing lifelong RL. We observe (in the case of the Minigrid experiments where all types of replay are used for all architectures), architecture seems to play a more significant role in improving lifelong learning performance than the type of replay.
\section{Introduction}
Lifelong Reinforcement Learning (LRL) involves training an agent to maximize its cumulative performance on a stream of changing \textit{tasks} over a long lifetime. LRL agents must balance \textit{plasticity vs. stability}: learning the current task while maintaining performance on previous tasks. Deep neural networks are especially prone to instability, often exhibiting \textit{catastrophic forgetting} of previous tasks when trained on multiple tasks that are presented sequentially \citep{kirkpatrick_overcoming_2016}. One approach to meet the challenges of deep LRL is by careful managing the agent's learning experiences, in order to learn (without forgetting) and build internal meta-models (of the tasks, environments, agents, and world). One strategy for managing experiences is to recall data from previous tasks and mix it with data from the current task when training, essentially transforming sequential LRL into batch multi-task reinforcement learning (RL) \citep{brunskill2013sample}. The technique of \textit{experience replay} of past data is common in deep RL \citep[e.g.][]{mnih2013playing} and has been studied in the lifelong learning literature \citep{isele2018selective,rolnick2019experience,hayes2021replay}.
However, storing sufficient examples from all previous tasks may be unreasonable in systems with limited resources or long lifetimes.
Rather than store an ever-growing list of experiences, one can instead employ \textit{generative replay} (GR) \citep{shin2017continual,van_de_ven_generative_2018}, a biologically-inspired replay mechanism that augments learning experiences with self-labelled examples drawn from an internal generative model trained to approximate the distribution of past experiences.
In LRL (see \citet{khetarpal2020towards} for a survey), each task is a Markov decision process (MDP), and the agent must learn a policy for interacting with its environment to maximize its lifetime performance (e.g., average reward over all tasks).
While GR has been shown to be effective in lifelong supervised learning \citep{van_de_ven_generative_2018}, it has been less explored in LRL.
The LRL setting is complicated by the fact that the agent's experiences are observation-action \emph{trajectories} whose distribution depends on the environment dynamics and the policy being executed.
While prior work with world models \citep{ketz2019continual} and pseudo-rehearsal \citep{atkinson2021pseudo} focused on generative replay of (partial) trajectories, this requires modeling the dynamics, which is difficult in complex environments.
In this paper, we present agent architectures for LRL that use GR for memory consolidation and
satisfy two desiderata: (a) Introspective density modelling of the latent representations of policies learned using deep RL, and (b) Model-free end-to-end learning.
The first property avoids the challenges of density modelling of complex high-dimensional perceptual inputs, whereas policy learning using deep RL works well with such perceptual inputs.
The second property avoids the challenges of learning temporal dynamics and reward functions from few learning experiences with sparse rewards.
Our contributions extend powerful \emph{model-free} deep RL and do not learn the dynamics explicitly.
Our first contribution is a model-free approach to generative replay (Section \ref{sec:model-free-generative-replay}) in lifelong RL. We empirically demonstrate that replay of independently sampled observation-action pairs is sufficient to mitigate forgetting. This is noteworthy because observation-action pairs are simpler to generate than trajectories, and because replay of observation-action pairs is directly analogous to replay of input-label pairs in supervised learning, thus unifying generative replay methods for lifelong learning.
Our second contribution is a mechanism for memory consolidation via \emph{wake-sleep} cycles. During \emph{wake phases}, our agent improves a policy for the current task(s) using an off-the-shelf RL algorithm while storing samples of experiences in a buffer. Periodically, the agent enters a \emph{sleep phase}, during which the generative memory is trained to model the agent's new experiences, and the procedural knowledge embodied in the policy learned during the wake phase is consolidated into the agent's ``skill set'' via policy distillation \citep{rusu2015policy} with GR. In the next wake phase, a new exploration policy is seeded with knowledge from previously-learned skills, and the cycle repeats.
We investigate three different replay architectures (Section \ref{sec:model-free-generative-replay}) within the wake-sleep framework, starting from a na\"ive generative and adding ingredients to achieve the aforementioned introspective density modelling of the latent representations and model-free end-to-end learning.
We evaluate our algorithm and architectural contributions through experiments (Section \ref{sec:experiments}) on three different scenarios comprising tasks from the \SC{} (SC-2) and Minigrid domains. We report several key findings showing the impact of the design choices on quantitative metrics that include transfer learning, generalization to unseen tasks, fast adaptation after task change, performance compared to task experts, and catastrophic forgetting.
We observe that our GR prevents drift in the features-to-action mapping of a deep RL agent (that learns the features from observations). We find that \emph{hidden replay} (replay of such features) outperforms other replay mechanisms, which agrees with past results in lifelong supervised learning \citep{van2020brain}. Since deep RL in SC-2 is computationally demanding, we release a dataset of trajectories of trained single-task expert policies for future research \footnote{\url{https://github.com/sri-l2m/l2m\_data}}.
\section{Background}
\subsection{Lifelong Reinforcement Learning (LRL) Setup}
A lifelong RL \emph{syllabus} is a sequence of \emph{tasks} that the learning agent experiences one at a time for a fixed number of interactions.
Each task is assumed to be formulated as a Partially Observable Markov Decision Processes (POMDPs) \citep{sutton2018reinforcement}, where
each POMDP is defined by a tuple $(\mathcal{S}, \mathcal{A}, P(s'|s,a), R(s, a), \mathcal{O}, Z(o|s))$ consisting of set of states $\mathcal{S}$, a set of actions $\mathcal{A}$, a transition probability function between states $P(s'|s,a)$, a reward function $R(s, a)$, a set of observations $\mathcal{O}$, and conditional observation probability function $Z(o|s)$.
We assume that all tasks have distinct observation spaces of the same dimensionality, share a common action space (e.g., point-and-click, issue movement or attack commands), but differ in the transition and reward dynamics.
For example, this setup can capture the use case of an embodied agent using the same set of sensors and actuators throughout their lifetime but with different reward functions or objectives.
In single-task RL, the objective is to learn a policy that maps the history of observations to actions, $\pi: \mathcal{O}^h \rightarrow \mathcal{A}$, such that the policy maximizes the expected discounted future reward, given by the value function $V^{\pi}(s) = E_{S_t \sim P, A_t \sim \pi}[\sum_{t = 1}^{\infty} \gamma^{t-1} R(S_t, A_t)]$. We will refer to the task-optimal policy as a \emph{single task expert} (STE).
Lifelong Learning involves training an agent to continually learn from a stream of tasks over long lifetimes.
As the agent encounters new tasks, it must accumulate and leverage knowledge to learn novel tasks faster (positive forward transfer), while maintaining performance on previous tasks (minimizing catastrophic forgetting or negative backwards transfer).
In this paper, we consider the task of lifelong reinforcement learning (LRL), which has been less explored than class-incremental classification \citet{hayes2021replay} in the literature.
The objective in LRL is to learn a policy that maximizes average cumulative reward over all POMDPs in a task set while learning from limited interactions with a sequence of POMDPs drawn from the task set.
This is challenging because the learning problem is non-stationary and the learner needs to accommodate an increasing number of tasks in a model with fixed capacity.
The task changes after an unknown number of interactions with the environment.
The agent has no control over the choice of the next task.
In our setting the agent has no knowledge of the identity of the current task, or the transition points between tasks.
\subsection{Generative Replay (GR)}
Our work is focused on generative replay (GR), which can be broadly described as a form of data augmentation where real-world experiences are used to continually train a generative model over the input space of the problem. Subsequently, data sampled from this generator can be pseudo-labeled or self-labeled to generate paired training examples that may reflect previously trained tasks.
Different types of replay were discussed by \citet{hayes2021replay} along with its role in the brain and in current deep learning implementations. Most of the prior work on generative replay has focused on the problem of class-incremental classification \citet{van_de_ven_three_2019,shin2017continual}, or class-incremental generative modelling \citet{cong2020gan} as a subroutine thereof, but there has been significantly less exploration of GR in LRL.
We use a Variational Auto-Encoder (VAE) \citep{kingma_auto-encoding_2014} as the generative model. The VAE consists of an encoder $q_\phi(z|o)$ mapping observations to a vector-valued latent space $z$, and a decoder $p_\theta(o|z)$ that is trained to reconstruct the original input. We use the Gaussian prior $z \sim N(0, \mathbf{I})$ with 0 mean and unit variance $\mathbf{I}$. To train the VAE, we minimize the standard loss $\mathcal{L}_\text{VAE}$,
\begin{equation}
\label{loss:vae}
\min E_{q_{\phi}(z|o)}[\ln p_{\theta}(o|z)] - D_{KL}[q_{\phi}(z|o)||p(z)]
\end{equation}
World models \citep{ketz2019continual} use model-based generative replay for LRL i.e., the replay generates rollouts or entire trajectories \citep[e.g.][]{ketz_using_2019}. Model-based replay works well in many domains, but it is challenging for domains such as SC-2 due to the complex dynamics and high dimensional observation and action spaces.
The most relevant prior work is that of \citeauthor{atkinson2021pseudo} \citep{atkinson2021pseudo} who explored model-free generative replay for Deep Q-Learning (DQN).
However, \citet{atkinson2021pseudo} uses GANs in the raw perceptual space of the agent (aka veridical replay \citet{hayes2021replay}). Others have explored sophisticated strategies for prioritized experience replay in DQN \citep[e.g.][]{isele2018selective, riemer2019scalable,wang2021acae,caccia2020online,zhang2017deeper,schaul2015prioritized} in LRL, but these are not generative in nature.
We show that na\"ively migrating the generative modelling (similar to \citeauthor{atkinson2021pseudo}) to the feature space of the policy causes instabilities and severe catastrophic forgetting.
To the best of our knowledge, this is the first work on model-free GR for LRL in the feature space of generic deep RL agents.
\section{Hidden Replay for Lifelong RL}
\label{sec:model-free-generative-replay}
This section introduces our core contribution of wake-sleep training combined with model-free GR of latent representations. We also explore two other GR architectures along the way.
The fundamental tradeoff in lifelong learning is between \emph{plasticity}, i.e.\!, learning the current task, and \emph{stability}, i.e.\!, remembering past tasks.
We address this tradeoff in lifelong RL by employing an approach that alternates between two distinct phases \citep{raghavan2020lifelong}.
In the wake phase, a \emph{plastic} wake policy $\pi_w$ is optimized for the current task by interaction with the environment and using any off-the-shelf RL method.
Transition tuples are collected during training and stored in a buffer (see Alg. \ref{alg:wakesleep}); each tuple contains $(o, r, a)$, the observation $o$, reward for the previous action $r$, and the policy output $a$ (e.g., the policy logits or one-hot encoded action).
The sleep policy $\pi_s$ provides ``advice'' to explore the current task using the consolidated policy learned from all previous tasks.
Wake phase actions are selected according to a mixture of $\pi_w$ and $\pi_s$, the probability of choosing an action from $\pi_s$ decaying to 0 over time.
We use an off-policy RL algorithm such as VTrace \citep{espeholt2018impala} to accommodate this off-policy action selection in the optimization of $\pi_w$.
In the sleep phase, the \emph{stable} sleep policy $\pi_s$ is optimized with the objective of incorporating new knowledge (the action selection in the wake buffer) while retaining current knowledge.
Similar to generative replay in supervised learning, augmented dataset(s) are created by combining wake transitions with tuples from the generative model $g_s$ and pseudo-labelled by the sleep policy.
This form of replay is model-free because it does not learn the POMDP dynamics and reward and does not generate and replay temporal trajectories.
The sleep policy and the generative model are jointly trained.
The sleep policy is trained to minimize $\mathcal{L}_{xent}$, a standard cross-entropy loss between the outputs (logits) of $\pi_s$, treating the replay augmented dataset as the ground truth.
The generative model is trained using the combined dataset and the standard VAE loss (Eq. \ref{loss:vae}).
The description so far is easily applied to veridical replay, i.e.\!, when generative modelling and replay occurs in the original observation space of the MDP, but this is undesirable as discussed before. The design choice of whether the input or hidden space is replayed distinguishes the model implementations (Alg. \ref{alg:wakesleep} vs Alg. \ref{alg:hidden_replay}, and models in Fig. \ref{fig:all_replay_archs}) from each other.
\begin{minipage}{0.49\textwidth}
\begin{algorithm}[H]
\caption{Wake-Sleep Training in the Sequential and Two-Headed models.}
\label{alg:wakesleep}
\begin{algorithmic}[1]
\STATE {\bfseries Iterates:} Generator $g_s^t$, sleep policy $\pi_{s}^t$, wake policy $\pi_{w}^t$, wake buffer $b^t_w$
\FOR{$t=1,2,\ldots$}
\FOR[Wake Phase]{$K$ times}
\STATE{Sample observation $o$ and reward $r$ from current task}
\STATE Sample $a \sim \pi_s^t(o)$ w.p. $p_{\text{advice}}$, else $a \sim \pi_w^t(o)$
\STATE{Add transition $(o, r, a)$ to FIFO wake buffer $b_w^t$}
\STATE{Update wake policy $\pi^t_w$ on current task reward using $b^t_w$}
\STATE{Decay $p_\text{advice}$}
\ENDFOR\\
\FOR[Sleep Phase]{$N$ iterations}
\STATE{Sample batch $B$ from $b_w^t$ (c.f.\! Line 6).}
\STATE Sample batches $O_s^t \sim g_s^t$
\STATE Pseudo-label $A_s^t = \pi_s^t(O_s^t)$
\STATE (GR Buffer) $D = B \cup (O_s^t, A_s^t)$
\STATE Minimize $\mathcal{L}_{VAE}(g_s^{t+1}) + \alpha\mathcal{L}_{xent}(g_s^{t+1}, \pi_s^{t+1})$ on $D$
\ENDFOR\\
\ENDFOR
\end{algorithmic}
\end{algorithm}
\end{minipage}
\hfill
\begin{minipage}{0.49\textwidth}
\begin{algorithm}[H]
\caption{Wake-Sleep Training in the Hidden Replay model.}
\label{alg:hidden_replay}
\begin{algorithmic}[1]
\STATE {\bfseries Iterates:} Generator $g_s^t$, sleep policy $\pi_{s}^t$, wake policy $\pi_{w}^t$, wake buffer $b^t_w$, random replay buffer $b^t_r$
\FOR{$t=1,2,\ldots$}
\FOR[Wake Phase]{$K$ times}
\STATE{Sample observation $o$ and reward $r$ from current task}
\STATE Sample $a \sim \pi_s^t(o)$ w.p. $p_{\text{advice}}$, else $a \sim \pi_w^t(o)$
\STATE{Add transition $(o, r, a)$ to wake buffer $b^t_w$}
\STATE{Update wake policy $\pi^t_w$ on current task reward using $b^t_w$}
\STATE{Decay $p_\text{advice}$}
\ENDFOR\\
\FOR[Sleep Phase]{$N$ iterations}
\STATE{If ER: Batch $B \sim b^t_w$ (c.f.\! Line 6).}
\STATE{If GR: Feature batches $H^t_s \sim g_s^t$}
\STATE{If GR: Pseudo-label $A^t_s \sim \pi_s^t(H_s^t)$}
\STATE{If RaR: $b^t_r = b^t_r \bigcup b^t_w.sample(K)$}
\STATE{If RaR: $D_r \sim b^t_r$}
\STATE Minimize $\mathcal{L}_{VAE}(g_s^{t+1}) + \alpha\mathcal{L}_{xent}(g_s^{t+1}, \pi_s^{t+1})$ on $(H^t_s, A^t_s)$ and end-to-end minimize (including feature extractor) the same loss on observation-action pairs in $B \bigcup D_r$
\ENDFOR\\
\ENDFOR
\end{algorithmic}
\end{algorithm}
\end{minipage}
\hfill
\begin{figure}[t]
\centering
\includegraphics[width=0.92\textwidth]{graphics/updates_archs_2.png}
\caption{We investigate three replay architectures for model-free generative replay (sleep phase). The red dashed arrows denotes sampling steps starting from the VAE to generate labelled actions. The blue dashed arrows show how the various replay buffers are sampled and fed into the sleep model during training. The black arrows denote the forward pass. \textbf{Top:} Sequential Replay Architecture; \textbf{Middle:} Two-Headed Replay Architecture; \textbf{Bottom:} Hidden Replay Architecture.}
\label{fig:all_replay_archs}
\end{figure}
We compare three replay architectures that we denote Sequential, Two-Headed, and Hidden Replay (shown in Fig. \ref{fig:all_replay_archs}). Each architecture is a specific arrangement of the following three components: a feature extractor, a VAE, and a policy network. The feature extractor maps raw image-like observations to feature vectors using a convolutional neural network (CNN) or a multi-layered perceptron. The policy network is dependent on the action space, e.g., in our experiments we utilize a convolutional architecture to output spatial actions.
In the \hypertarget{text:sequential}{\textbf{Sequential}} architecture (Fig.~\ref{fig:all_replay_archs}, Top), the VAE is upstream of the policy network, and the sleep policy $\pi_s$ operates on the VAE reconstruction rather than the true observation. This architecture is similar to the one proposed in \citet{atkinson2021pseudo}.
During sleep (following the red dashes in Fig.~\ref{fig:all_replay_archs}, Top), sampled latent vectors are decoded to observations that are then labelled by the policy. We hypothesize that the sequential architecture tends to map observations for which the policy chooses similar actions to generic reconstructions. We found that this method produces blurry reconstructions for image-like observations.
The \hypertarget{text:twoheaded}{\textbf{Two-Headed}} architecture (Fig.~\ref{fig:all_replay_archs}, Middle) decouples the decoder/generator from the policy network. Each observation (real or generated) is passed through the feature extractor. The features are then separately passed through the VAE and the policy network. Our experiments showed improved visual reconstructions (see Appendix \ref{sec:reconstructions_app}) due to separation of the influence of the policy on the VAE decoder. However, sampling in this architecture requires two forward passes as shown in the figure (once to decode an observation and once to label it).
Alg. \ref{alg:wakesleep} shows the data flow that is common between these two architectures that perform veridical replay, i.e.\!, the VAE posterior is a density over raw and complex perceptual space. The training is wasteful because the wake policy $\pi_w$ has already learned an effective feature extractor, and latent-to-action partition that works well for the current task (solved in wake phase). The sleep phase in veridical replay does not utilize these learned features and tries to re-learn them, whereas deep RL (as in wake phase) is effective at learning representations to map raw percepts to actions.
\setlength{\columnsep}{5pt}%
\begin{wrapfigure}{R}{0.35\textwidth}
\centering
\includegraphics[width=0.35\textwidth, keepaspectratio]{graphics/hidden_replay_partitioning.png}
\vspace{-0.75cm}
\caption{Hidden replay prevents drift in the features-to-action
mapping from the latent vector space of a deep RL agent (Minigrid tasks). Colors represent actions; $\times$s: drawn from the random replay buffer, $\bullet$: drawn from wake buffer; $\star$: sampled via generative replay.}
\vspace{-0.75cm}
\label{fig:hidden_replay_partitioning}
\end{wrapfigure}
The final architecture we consider is inspired by the success of \hypertarget{text:hidden}{\textbf{Hidden Replay}} in class incremental learning \citep{van2020brain} and its connection to biology. This architecture (Fig.~\ref{fig:all_replay_archs}, Bottom) generates feature vectors directly rather than observations. Sampling features instead of image-like observations benefits from dimensionality reduction. Reconstructing or generating complex image observations is hard and requires increasingly complex generative models \citep{van2017neural,van2016conditional,razavi2019generating}. In RL, especially when working with ``symbolic'' (semantic) (as opposed to ``perceptual'' (RGB)) observations, changing a small detail of the observation can change the optimal action drastically. Observations might be better separated in feature space, and thus, the sleep policy could be less sensitive to an imperfect decoder. In addition, in contrast to the two-headed architecture, the feature extractor does not need to be run when sampling.
For example, when faced with new tasks that require new actions to be taken, hidden replay can learn to add a new action partition to the feature space while rearranging the existing feature space in a manner that preserves partitioning needed to solve previous tasks. Empirically, we observe a clean partition of the feature space according to the optimal action. Even though samples and reconstructions are not perfect, they fall within the same action partition (Fig.~\ref{fig:hidden_replay_partitioning}) as the ground truth. These representations may allow for knowledge transfer between tasks because tasks requiring similar actions can potentially efficiently be mapped into the existing feature space, but validating this requires further studies.
Hidden replay is not without limitations because GR samples are in terms of features without the corresponding observations. This V-shape in the training graph (blue lines in Fig. \ref{fig:all_replay_archs}, Bottom) can cause the feature extractor to drift over time or collapse to a mode (e.g., map all observations to the same feature vector).
As new tasks are encountered, this would cause sampled features
to not align with the features mapped by the policy and VAE components.
To limit feature drift, we store a small (typically $K < 1\%$ of the experience replay buffer) set of raw observations paired with actions in a separate buffer. We denote this RaR (for ``random replay'' buffer). These are accumulated over the lifetime of the agent from all wake replay buffers $b^t_w$ (although a more sophisticated exemplar selection \citep{rebuffi2017icarl,mi2020ader} is also possible). We observed that replaying a few random examples results in less feature drift between task changes (See Appendix ~\ref{sec:justification_of_exemplars} for an ablation). In our experiments, we found that saving 96 random samples per sleep was sufficient to alleviate feature drift in the \SC{} domain and saving 256 random samples per sleep was sufficient for the Minigrid domain. The complete data flow for hidden replay for LRL is shown in Alg. \ref{alg:hidden_replay}, where the design choice of combinations of ER, GR, and RaR buffers are ablated in our experimental ablations.
\textbf{Remark}: When applied to single task RL, our wake-sleep cycle resembles policy iteration interleaved with policy distillation steps, where each iteration $t$ produces an improved $\pi_w^t$ (nudge towards STE) that is then distilled into $\pi_s^t$. Wake experiences are gathered from the most recent sleep policy (``advice''), including from multiple sub-optimal policies similar in spirit to dataset aggregation in DAGGER \citep{ross2011reduction}. Similar to DAGGER, the consolidated policy is executed to sample next trajectories. Our supplementary experiments (Appendices \ref{sec:wake_additional_results} and \ref{sec:ste_additional_results}) show stable convergence of the wake-sleep cycle to optimal policies in single task training (convergence to STE). |
\section{Problem Statement}
\section{Background and Related Work}
\begin{comment}
\subsection{Causal Discovery}
Given observational data, Causal Discovery Algorithms (CDAs) find causal relationships among different attributes in the dataset \cite{glymour2019review}.
Causal relations returned by CDAs can be graphically represented via a Directed Acyclic graph (DAG), also known as causal graph, where each node represents a data attribute and each edge represents a causal relation. For example, a directed edge from X to Y signify that X is the cause of Y, i.e., a change in X will cause a change in Y. The set of causal discovery algorithms can be broadly classified into constraint based methods such as PC\cite{Spirtes2000}, Fast Causal Inference (FCI), etc. and score based algorithms such as Greedy Equivalence Search (GES)\cite{chickering2002optimal}, Fast GES \cite{fges}, etc. Among the set of possible causal DAGs, constraint based CDAs rely on clever schedule of conditional independence tests whereas score based CDAs rely on some fitness metric like BIC score to filter to the final causal graph.
In this work, we have used PC algorithm
as it can handle mixed data types and provide asymptotically correct results.
It should be noted that causal discovery algorithms operate under a set of assumptions and have their own limitations. For example, PC algorithm assumes that no confounders (direct common cause of two variables) are missing from the dataset. In a practical setting, such factors can lead to an imperfect causal graph. One way to refine such causal graphs is by injecting domain knowledge. In line with existing work \cite{Wang2016, wang2017visual, hoque2021outcome}, we have employed a visual analytics approach to augment CDAs with human knowledge to get the final causal graph.
\end{comment}
\subsection{Bias Identification}
\label{sec:identification}
The existing literature on bias identification mostly revolves around different fairness metrics. Numerous fairness metrics have been proposed which capture different facets of fairness, such as group fairness, individual fairness, counterfactual fairness, etc.
\cite{dwork2018group,aif360, bechavod2017penalizing, kusner2017counterfactual,arvindTalk}.
Another way to classify fairness metrics can be on the level they operate on. For eg., dataset based metrics are solely computed using the dataset and are independent of any ML model, such as statistical parity difference. On the other hand, classifier based metrics are computed over the predictions of a trained ML model, such as false negative rate difference.
So far, there is not a single best fairness metric. Moreover, some of the fairness metrics can be mutually incompatible, i.e., it is impossible to optimize different metrics simultaneously\cite{kleinberg2016inherent}.
In line with existing visual tools \cite{silva, aif360}, our tool also uses a diverse set of fairness metrics to present a more comprehensive picture.
Many fairness metrics solely focus on the aggregate relationship between the sensitive attribute and the output variable.
This can lead to misleading conclusions as the aggregate trend might disappear or reverse when accounting for other relevant factors. A prime example of this phenomenon, also known as Simpson's paradox \cite{Pearl2009}, is the Berkeley Graduate Admission dataset \cite{bickel1975sex}. There it appeared as if the admission process was biased against women since the overall admit rate for men (44\%) was significantly higher than for women (30\%) \cite{barocas2017fairness}. However, this correlation/association did not account for the fact that women typically applied for more competitive departments than men. After correcting for this factor, it was found that the admission process had a small but statistically significant bias \textit{in favor of} women \cite{bickel1975sex}.
Causal models can be an effective tool for dealing with such a situation as they can decipher the different intermediate factors (indirect effects) along with their respective contributions behind an aggregate trend.
Hence, our tool also employs causal model for bias identification.
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{figure/workflow.png}
\setlength{\belowcaptionskip}{-8pt}
\setlength{\abovecaptionskip}{-4pt}
\caption{The workflow of D-BIAS has 4 parts: (1) The default causal model is inferred using the PC algorithm and SEM. Here, dotted edges represent undirected edges (2) Refine stage - the default causal model is refined by directing undirected edges, reversing edges, etc. (3) Debiasing stage - the causal model is debiased by deleting/weakening biased edges (4) The debiased causal model is used to simulate the debiased dataset.}
\label{fig:workflow}
\end{figure*}
\subsection{Bias Mitigation}
\label{sec:mitigation}
The existing literature on algorithmic bias mitigation can be broadly categorized into the three different stages in which they operate within the ML pipeline, namely pre-processing, in-processing and post-processing. In the pre-processing stage, the dataset is modified such that the bias with respect to an attribute or set of attributes is reduced or eliminated \cite{zemel2013learning, kamiran2009classifying, hajian2013dataModification, nipsIBM, rajabi2021tabfairgan}. This can be achieved by either modifying the output label \cite{kamiran2012, kamiran2009classifying} or by modifying the input attributes \cite{nipsIBM, zemel2013learning}.
In the in-processing stage, the algorithm is designed to take in biased data but still generate unbiased predictions. This can be achieved by tweaking the learning objectives such that accuracy is optimized while discrimination is minimized
\cite{inprocess}. Finally, in the post-processing stage, predictions from ML algorithms are modified to counter bias
\cite{lohia2018bias}.
Our work relates closely with the pre-processing stage where we
make changes to the input attributes and the output label.
There is also a growing set of work at the intersection of bias mitigation and causality \cite{zhang2017causal, wu2018discrimination, chiappa2019path}. The general idea is to generate the causal network, modify it, and then simulate debiased data. These approaches fully rely on automated techniques to yield the true causal network, and assume a priori knowledge about fair/unfair causal relationships.
Our work draws inspiration from this line of work and presents a more general solution where human domain knowledge is leveraged to refine the causal network and generate the debiased dataset.
\subsection{Visual Tools}
Recent years have seen visual tools like \textit{Silva}\cite{silva}, \textit{FairVis}\cite{fairvis}, \textit{FairRankVis}\cite{xie2021fairrankvis}, \textit{DiscriLens}\cite{DiscriLens}, \textit{FairSight}\cite{fairsight}, \textit{WordBias}\cite{ghai2021wordbias}, etc. which are all aimed at tackling algorithmic bias. Although most of these tools are focused on bias identification, a few of them, such as FairSight, also permit debiasing. However, the debiasing strategy used in such tools is fairly basic, like eliminating proxy variable(s). Simple measures like this can lead to high data distortion and can have a high negative impact on data utility. Our work relates more closely with \textit{Silva} which also features a graphical causal model in its interface. Using an empirical study, it showed that users can interpret causal networks and found them helpful in identifying algorithmic bias. However like most other visual tools, Silva is limited to bias identification. Our work advances the state of the art by presenting a tool that supports both bias identification and mitigation, with our debiasing strategy being more nuanced and sensitive toward data distortion.
\begin{comment}
\section{Our Approach}
Most of the existing literature focuses on fully automated approaches to tackle algorithmic bias. Such approaches have raised serious concerns about interpretability, trust, accountability, etc. \textcolor{blue}{Mention fairness can't be automated.}
In this work, we propose \textit{Human in the loop} approach for bias identification and mitigation with the objective of addressing some of these concerns. Here, we have chosen graphical causal model as a medium to present underlying patterns in the data as well as to inject human domain knowledge in the system. \textcolor{blue}{why causality for algorithmic bias?} \textcolor{blue}{generalizability - better than correlation}
In the following, we provide more detailed thoughts on the added fairness, accountability, interpretability, and trust our approach can invoke.
\textbf{Fairness.} In the literature, multiple definitions \cite{arvindTalk} of fairness are proposed which can be mutually incompatible and entail unavoidable tradeoffs \cite{kleinberg2016inherent} (see also Sec. \ref{sec:identification}). There is no firm consensus on a single most appropriate definition of fairness \cite{gajane2017formalizing}. For a given application in a given context, algorithms can't decide a suitable tradeoff between fairness metrics acceptable to all stakeholders. On the other hand, a human trusted by the majority of stakeholders can take an informed decision when presented with the required information. Hence, introducing a human in the loop might improve perceived fairness.
\textbf{Interpretability.} Introducing a human in the loop will only be effective when a person is able to understand the underlying state of the system and provide useful feedback. We define interpretability as the ability to understand how bias flows from sensitive attributes and contaminates other attributes in the dataset. We have used a causal model to capture this flow
(see Fig \ref{fig:teaser}). A study by Borkin et al, \cite{borkin2013makes} showed that network visualizations are quite popular in infographics and are among the most memorable visual artifacts. Thus, causal models, explained by a straightforward network visualization, promise to be intuitive and bring more interpretability. \textcolor{blue}{Also talk about silva}
\textbf{Accountability.} Another major concern with algorithmic bias is accountability. Who should be held accountable if an algorithm does something wrong? Should software developers be held accountable or the people who curated/selected the dataset? Given that software developers themselves aren't fully aware of the behavior of the black box models under different conditions, it would not be entirely justified to blame them. Assuming that training data is the only or at least a major source of bias, having an interactive visual tool that presents the underlying state of the system to a human(s) analyst empowered to guide the debiasing process should bring more accountability.
\textbf{Trust.} People are more likely to trust a system if they can tinker with it. This was shown in \cite{dietvorst2016overcoming}, where it was found that people were more likely to use an algorithm and accept its errors if they were given the opportunity to modify it themselves, even if it meant making it perform imperfectly. Our human in the loop approach allows human(s) to supervise/participate in the entire bias mitigation process and make changes as they feel fit. Hence, our approach should also instill more trust.
We will evaluate our tool on each of these parameters and compare it against a fully automated tool (see Sec. \ref{sec:evaluation}).
\end{comment}
\section{Methodology}
The workflow of our system can be understood from Fig. \ref{fig:workflow}. A detailed discussion for each stage follows next.
\begin{comment}
Given raw tabular data as input, D-BIAS outputs its debiased version. The entire process can be understood in four steps as shown in \autoref{fig:workflow}. In step 1, the raw multivariate data is read in and is used to compute a causal model using automated techniques.
In step 2, the user interacts with the edges of the initial causal model like directing undirected edges to reach to a reliable causal model. In step 3, the user identifies and modifies biased causal relationships. Finally, in step 4, our tool computes a new dataset based on user interactions in the previous step. The user can thereafter click on `Evaluate Metrics' to compute and plot different evaluation metrics.
If the user is satisfied with the evaluation metrics, they can download the debiased dataset and safely use it for any downstream application. Otherwise, the user may return to the causal network and make further changes. This cycle continues until the user is satisfied. A detailed discussion follows next.
\end{comment}
\subsection{Generating the Causal Model}
\label{sec:causal}
A causal model generally consists of two parts: causal graph (skeleton) and statistical model. Given a tabular dataset as input, we first infer an initial causal graph (directed acyclic graph) using a popular causal discovery algorithm, namely the PC algorithm
\cite{Spirtes2000,Colombo2014Order-independentLearning}.
This algorithm relies on conditional independence tests and a set of orientation rules to yield the causal graph (for details, see appendix B). Each node in the causal graph represents a data attribute and each edge represents a causal relation. For example, a directed edge from X to Y signifies that X is the cause of Y, i.e., a change in X will cause a change in Y.
The PC algorithm provides qualitative information about the presence of causal relationship between variables. To quantitatively estimate the strength of each causal relationship,
we use linear structural equation models (SEM) where a node is modeled as a linear function of its parent nodes.
\begin{equation}
y = \sum_{i}^{parents(y)}{\beta_{i} x_i} \;+ \;\varepsilon
\end{equation}
In the above equation, variable y is modeled as a linear combination of its parents $x_i$, their regression coefficients ($\beta_i$) and the intercept term ($\varepsilon$).
If y is a numerical variable, we use linear regression else we use the multinomial logit model to estimate the values of $\beta_i$ and $\varepsilon$. Here, $\beta_i$ represents the strength of causal relation between $x_i$ and y.
We repeat this modeling process for each node with non-zero parent nodes. Such nodes that have at least one edge leading into them are termed as endogenous variables.
Other nodes correspond to exogenous variables or independent variables that have no parent nodes. In \autoref{fig:workflow} (Refined Causal Model), s and a are exogenous variables while b, c and y are endogenous variables. Here, y will be modeled as a function of its parent nodes, i.e., a, b and c. Similarly, b and c will be modeled as function of s.
After this process, we arrive at a causal graph whose different edges are parameterized using SEM. This constitutes a Structural Causal Model (SCM) or simply causal model (see Fig. \ref{fig:workflow} (1)).
This causal model, generated using automated algorithm, might have some undirected/erroneous causal edges due to different factors like sampling bias, missing attributes, etc.
To achieve a reliable causal model, we have taken a similar approach as Wang and Mueller \cite{wang2017visual} where we leverage user knowledge to refine the causal model via operations like adding, deleting, directing, and reversing causal edges (see Fig. \ref{fig:workflow} (2)). For every operation, the system computes a score (Bayesian Information Criterion (BIC)) of how well the modified causal model fits the underlying dataset. Similar to \cite{wang2017visual}, our system assists the user in refining the causal model by providing the difference in BIC score before and after the change. A negative BIC score suggests a better fit.
After achieving a reliable causal model, we enter into the debiasing stage where any changes made to the causal model will reflect on the debiased dataset (see Fig. \ref{fig:workflow} (3)).
\begin{comment}
where subscript $i$ indicates the $i$-th attribute, $\varepsilon$ is the intercept term and $r_{i}$ is the residual term. Residual is the difference between prediction and real value.
The edges for which the algorithm is not able to learn the direction are not used for regression. In causality theory, the regression coefficient $\beta_k$ serves as the measure of the causal influence from $x_k$ to $y$, and is represented as an edge weight in Fig \ref{fig:causal_diagram}.
The magnitude of $\beta_k$ is encoded as line width. The higher the magnitude, the wider the edge's width. The sign of $\beta_k$ is encoded by color. Edges with positive $\beta_k$ are green, otherwise red. The residual term $\varepsilon$ describes the part of $y$ that cannot be explained by the variable considered.
\end{comment}
\subsection{Auditing and Mitigating Social Biases}
From a causal perspective, discrimination can be defined as an unfair causal effect of the sensitive attribute (say, race) on the outcome variable (say, getting a loan)\cite{Pearl2000}. A direct causal path from a sensitive variable to the output variable constitutes disparate treatment while an unfair indirect path via some proxy attribute (say, zipcode) constitutes disparate impact \cite{chiappa2018causal}.
A direct path is certainly unfair but an indirect path may be fair or unfair (as in the case of Berkeley Admission dataset \cite{Pearl2009}). Our system computes all causal paths
and lets the user decide if a causal path is fair or unfair. If a causal path is unfair, the user should identify which constituting causal relationship(s) are unfair and act on them. The user can do this by deleting or weakening such biased causal relationships to reduce/mitigate the impact of the sensitive attribute on the outcome variable. For example, in Fig.\ref{fig:workflow} (Refined Causal Model), s is the sensitive attribute and y is the outcome variable. Here, the user deletes the edge $s \to b$ and weakens the edge $s \to c$ (shown as a thin edge). Once the user has dealt with the biased causal relationships, we achieve what we call the \textit{Debiased Causal Model}.
\begin{algorithm}
\caption{Generate Debiased Dataset}
\label{algo:1}
\begin{algorithmic}[1]
\STATE $ D \leftarrow $ Original Dataset
\STATE $G(V, E) \leftarrow $ refined causal model
\STATE $E_a \leftarrow$ set of edges added during the debiasing stage
\STATE $E_m \leftarrow $ subset of E that were deleted/strengthened/weakened during the debiasing stage
\STATE
\FOR{\textbf{each} e in $E_a$}
\STATE $n \leftarrow $node pointed by head of e
\STATE retrain linear model for n as a function of its parents
\ENDFOR
\STATE
\STATE $V_{sim} \leftarrow \emptyset$ \hfill // Attributes that need to be simulated
\FOR{\textbf{each} e in $(E_m \cup E_a)$}
\STATE $n \leftarrow $node pointed by head of e
\STATE $V_{sim} \leftarrow V_{sim} + n + all\_descendent\_nodes(n)$
\ENDFOR
\STATE $V_{sim} \leftarrow remove\_duplicates(V_{sim})$
\STATE
\STATE $D_{deb} \leftarrow \emptyset $ \hfill // Debiased Dataset
\FOR{\textbf{each} v in topological\_sort(V)}
\IF{v present in $V_{sim}$}
\STATE $D_{deb}[v] \leftarrow $ generate values based on \autoref{eq:gen}
\ELSE
\STATE $D_{deb}[v] \leftarrow D[v]$
\ENDIF
\ENDFOR
\STATE
\STATE // Rescale values for the simulated attributes
\FOR{\textbf{each} v in $V_{sim}$}
\IF{v is a categorical variable}
\STATE $D_{deb}[v] \leftarrow $ rescale values based on Algorithm 2
\ELSE
\STATE // v is a numeric variable
\STATE $\mu, \sigma^2 \leftarrow mean(D[v]), variance(D[v])$
\STATE $\mu_{deb}, \sigma_{deb}^2 \leftarrow mean(D_{deb}[v]), variance(D_{deb}[v])$
\STATE $D_{deb}[v] \leftarrow \mu + (D_{deb}[v] - \mu_{deb})/\sigma_{deb}*\sigma$
\ENDIF
\ENDFOR
\STATE \textbf{Result} $D_{deb}$
\end{algorithmic}
\end{algorithm}
\subsection{Generating the Debiased Dataset}
We simulate the debiased dataset based on the debiased causal model (see Fig. \ref{fig:workflow} (4)).
The idea is that if the user weakens/removes biased edges from the causal network, then the simulated dataset might also contain less biases. The standard way to simulate a dataset based on a causal model is to generate random numbers for the exogenous variables \cite{sofrygin2017simcausal}. Thereafter, each endogenous variable is simulated as a function of its parent nodes (variables) in the causal network.
In this work, we have adapted this procedure to suit our needs, i.e., simulating a fair dataset while having minimum distortion from the original dataset.
Our approach to generate the debiased dataset, as illustrated in Algorithm \ref{algo:1}, can be broken down into 4 steps.
At step 1 (lines 6--9), we focus on the set of edges added during the debiasing stage ($E_{a}$). We retrain regression models corresponding to each of the target nodes of $E_{a}$. This will update the weights (regression coefficients ($\beta_i$)) for all edges leading into any target node of $E_{a}$. At step 2 (lines 11--16), we identify the set of nodes (attributes) that need to be simulated ($V_{sim}$). Unlike the standard procedure, we only simulate selective nodes that are directly/indirectly impacted by the user's interaction to minimize distortion. This set includes the target nodes of all edges that the user has interacted with along with their descendent nodes. For example, in \autoref{fig:workflow}(3), the user deletes the edge $s \to b$ and weakens the edge $s \to c$. So, we will only simulate variables $b$, $c$ and $y$. At step 3 (lines 18--25), we actually simulate all nodes that are a part of $V_{sim}$ using \autoref{eq:gen}. All other nodes are left untouched and their values are simply copied from the original dataset into the debiased dataset.
It should be noted that all parent nodes must be simulated before their child node as the values for a node are computed using their parent nodes. So, we simulate all endogenous variables in a topological order.
For eg., in \autoref{fig:workflow}(4), nodes \textit{b} and \textit{c} will be simulated before node \textit{y}.
\vspace{-7pt}
\begin{equation}
\label{eq:gen}
y = \sum_{i}^{parents(y)}{\alpha_{i}\beta_{i} x_{i}} \;+\;\; \varepsilon \;\;+\; \sum_{i}^{parents(y)}{(1-\alpha_{i})\beta_{i}r_{i}}
\end{equation}
\vspace{-7pt}
In the above equation, node y is simulated as a sum of 3 terms. The first term is the weighted linear combination of parent nodes. Here, $\alpha_i$ is the scaling factor that has a default value of 1 and can range between [0, 2] as determined by user interaction.
Strengthening an edge, sets $\alpha_i>$1; weakening an edge, sets $\alpha_i<$1. For example, if the user weakens the edge between node $x_i$ and y by -35\%, then $\alpha_i$=0.65, strengthening it by +35\% will set $\alpha_i$=1.35, and deleting it will set $\alpha_i$=0. The second term is the intercept that was computed when the regression model for y was last trained. The third term adds randomness in proportion to the degree to which the user has altered an incoming edge. Here, $r_{i}$ is a normal random variable that has a similar distribution as $x_{i}$ ($r_{i} \sim \mathcal{N}({\mu_{x_{i}}, \sigma_{x_{i}}^{2}})$). This term adds fairness as it is random and alleviates distortion for y.
\begin{algorithm}[t]
\caption{Rescale Categorical Variable v}
\begin{algorithmic}[1]
\STATE D[v] $\leftarrow$ categorical variable v in the original dataset
\STATE prob\_mat $\leftarrow$ probability matrix for v computed using Eq. 2
\STATE
\STATE lr $\leftarrow $ 0.1 \hfill // learning rate
\STATE iterations $\leftarrow $ 0
\LOOP
\STATE // DPD: Discrete Probability Distribution
\STATE $dist_{ori} \leftarrow$ DPD( D[v] )
\STATE $dist_{deb} \leftarrow$ DPD( argmax(prob\_mat) )
\STATE diff $\leftarrow \sum{\left\|(dist_{ori} - dist_{deb})/dist_{deb}\right\|}$
\STATE scale\_factor $\leftarrow 1 + lr*\sum{(dist_{ori} - dist_{deb})/dist_{deb})}$
\STATE prob\_mat $\leftarrow$ scale\_factor * prob\_mat
\STATE $dist_{deb} \leftarrow$ DPD( argmax(prob\_mat) )
\STATE new\_diff $\leftarrow \sum{\left\|(dist_{ori} - dist_{deb})/dist_{deb}\right\|}$
\IF{new\_diff $>$ diff or iterations $>$ 50}
\STATE break
\ENDIF
\STATE iterations $\leftarrow$ iterations + 1
\ENDLOOP
\STATE $D_{deb}[v] \leftarrow$ argmax(prob\_mat)
\STATE \textbf{Result} $D_{deb}[v]$
\begin{comment}
def scale_prob_matrix(pmat, col_ori):
learning_rate = 0.1
iterations = 0
while True:
m1 = col_ori.value_counts(normalize=True)
m2 = pmat.idxmax(axis=1).value_counts(normalize=True)
gap = np.sum(np.abs((m1-m2)/m2))
# print(gap)
scale_fac = 1 + learning_rate*((m1-m2)/m2)
pmat = scale_fac * pmat
m2 = pmat.idxmax(axis=1).value_counts(normalize=True)
new_gap = np.sum(np.abs((m1-m2)/m2))
#print(round(new_gap - gap,10))
if new_gap > gap or iterations > 50:
break
gap = new_gap
iterations = iterations + 1
tmp_data.idxmax(axis=1).tolist()
return pmat
\end{comment}
\end{algorithmic}
\label{algo:2}
\end{algorithm}
Fig. \ref{fig:deb_eqn} illustrates the case where the node \textit{Job} has a single parent node (\textit{Gender}). Let's say that the user chooses to delete this edge ($\alpha$=0). Going the conventional route (without the third term), the attribute \textit{Job} will be reduced to a constant value (the intercept ($\varepsilon$)) which is undesirable.
Adding the (third) random term generates the \textit{Job} distribution below the `+' node which is far more balanced (fair) in terms of \textit{Gender} than the `Original' distribution on the top right. However,
the number of people getting the job (or not) is distorted compared the `Original' distribution.
This is corrected in Step 4 (lines 27--37) in Algorithm \ref{algo:1}, where we rescale each simulated variable so that its distribution is close to its original distribution. For numerical variable(s), we simply standardize values to their original mean and standard deviation. For categorical variable(s), the model returns a set of probability scores corresponding to each possible output label for each data point.
We iteratively scale such a probability matrix so that the resulting debiased distribution inches toward the original distribution (see Algorithm \ref{algo:2}).
We continue this process until a fixed number of iterations or when the difference between original and debiased distribution starts increasing. In Fig. \ref{fig:deb_eqn}, we observe that the resulting `Debiased' distribution matches the `Original' distribution in terms of \textit{Job} allocation, while maintaining gender parity.
It should be noted that simulating each attribute adds a corresponding modeling error to the process. This modeling error is typically small but it can potentially overpower the impact of the user's intervention, especially when a user makes a small change, say weakening an edge by 5\%.
In such a case, the results may not be in strict accordance with the user’s expectations.
\subsection{Evaluation Metrics}
\label{sec:eval_metrics}
Once the debiased dataset is generated, it can be evaluated using different metrics that operate at the dataset level and the classifier level. For the second case, the debiased dataset is used to train a ML model chosen by the user, and a set of metrics are computed over the model's predictions. Here, the idea is to evaluate the downstream effects of debiasing.
All the evaluation metrics can be grouped into three broad categories, namely utility, fairness and distortion. It should be noted that there might be a trade-off among the three categories. For eg., reducing bias might cause high data distortion or lower utility. For comparison, we have used a baseline debiasing strategy which just removes the sensitive attribute(s) from the dataset.
\begin{figure}[t]
\centering
\includegraphics[width=0.75\columnwidth]{figure/debiasing.png}
\setlength{\belowcaptionskip}{-4pt}
\setlength{\abovecaptionskip}{-2pt}
\caption{Illustration of the data debiasing process using a toy example where the node Job is caused by a single variable, namely gender. Green color marks the proportion of females who got the job (Y) or not (N).}
\label{fig:deb_eqn}
\end{figure}
\textbf{Fairness.} Our tool presents a diverse set of 5 popular fairness metrics, namely statistical parity difference (Parity diff), individual fairness (Ind. bias), accuracy difference (Accuracy diff), false negative rate difference (FNR diff) and false positive rate difference (FPR diff) \cite{aif360}. Two of these metrics operate at the dataset level (Parity diff, Individual bias) and the rest operate on the classifier’s predictions.
Here, Ind. bias is defined as the mean percentage of a data point's k-nearest neighbors that have a different output label. A lower value for Ind. bias is desirable as it means that similar data points have similar output labels. For the 4 other fairness metrics, we compute some statistic for the two groups say males and females, and then report their absolute difference. This statistic can be ML model dependent, such as accuracy,
or model independent, such as the likelihood of getting a positive output label.
Lower values for such metrics suggests more equality between groups. For computing model based metrics, we omit the sensitive attributes(s) and perform 3-fold cross validation using the user-specified ML model with 50:50 train test split ration, and then report the mean absolute difference between groups across the 3 folds.
\textbf{Utility.} The utility of ML models is typically measured using metrics like accuracy, F1 score, etc.
In our context, we are interested in measuring the utility of a ML model when it is trained using the debiased dataset instead of the original dataset.
To compute the utility for the original dataset, we perform 3-fold cross validation using the user-specified ML model and report the mean accuracy and F1 score. For the debiased dataset, we follow a similar procedure where we train the user-specified ML model using the debiased dataset. However, we use the outcome variable from the original dataset as the ground truth for validation. Sensitive attribute(s) are removed from both datasets before training.
Ideally, we would like the utility metrics for the debiased dataset to be close to the corresponding metrics for the original dataset.
\textbf{Data Distortion.} Data distortion is the magnitude of deviation of the debiased dataset from the original dataset.
Since the dataset can have a mix of continuous and categorical variables, we have used the \textit{Gower distance}\cite{gower} metric.
We compute the distance between corresponding rows of the original and debiased dataset, and then report the mean distance as data distortion. This metric is easy to interpret as it has a fixed lower and upper bound ([0,1]). It is 0 if the debiased dataset is the same as the original while higher values signify higher distortion. Lower values for data distortion are desirable.
\begin{figure}
\centering
\includegraphics[scale=0.25]{figure/adult_income_logs.png}
\caption{Logs view highlighting the changes made to the causal network for the Adult Income dataset. Dotted lines represent deleted edges and nodes in grey represent the impacted nodes.}
\label{fig:logs}
\vspace{-10pt}
\end{figure}
\section{The D-BIAS Tool}
\subsection{Generator Panel}
This is the first component of the tool the user interacts with (see Fig.\ref{fig:teaser} (A)). The user starts off by choosing a dataset from its dropdown menu.
The user then selects the label variable which should be a binary categorical variable as we are considering a classification setting. Next, the user selects all nominal variables which is required for fitting the SEM model. Lastly, the user chooses a p-value and clicks the `Causal Model' button to generate the causal network.
Here, the p-value is used by the PC algorithm to conduct independence tests. We set $p=0.01$ for all our demonstrations. It can be changed to 0.05 or 0.10 for smaller datasets.
The `Debiased Data' button downloads the debiased dataset.
\begin{comment}
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{figure/teaser_pic.png}
\caption{The interface of our D-BIAS visual tool using the Adult Income dataset. (A) The Generator panel: used to create the causal network and download the debiased dataset (B) The Causal Network view: shows the causal relations between the attributes of the data, allows user to inject their prior in the system (C) The Evaluation panel: used to choose the sensitive variable, the ML model and displays different evaluation metrics. }
\label{fig:teaser}
\end{figure*}
\end{comment}
\begin{figure}[t]
\centering
\includegraphics[width=0.75\columnwidth]{figure/subgroup.png}
\caption{The visual interface for selecting subgroups (Group A and B). Each column consists of a list of bar charts/histograms representing all attributes in the dataset. By default, all bars are colored gray. The user can click on multiple bars to select a subgroup. Selected bars are colored blue. Each bar is filled in proportion of their representation in the selected subgroup as a ratio of the entire dataset. In this picture, Group A consists of individuals who went to elite universities and whose age lies between 24-40. It represents 18\% of the dataset.}
\label{fig:custom_group}
\vspace{-10pt}
\end{figure}
\subsection{Causal Network View}
This is the most critical piece of the interface where the user will spend most of the time (see Fig.\ref{fig:teaser}(B)). The center of this view contains the actual causal network which is surrounded by 4 panels on all sides.
\textbf{Causal Network.} All features in the dataset are represented as nodes in the network and each edge represents a causal relation. The width of an edge encodes the magnitude of the corresponding standardized beta coefficient. It signifies the feature importance of the source node in predicting the target node. The color of an edge encodes the sign of the corresponding standardized beta coefficient. Green (red) represents positive (negative) influence of the source node on the target node. If an edge is undirected, it does not have a beta coefficient and is colored orange. Finally, gray color encode edges
that represent relationships which can not be represented by a single beta coefficient. This occurs when the target node is a categorical variable with more than 2 levels.
The causal network supports many interactions to enhance the user's overall experience and productivity. It supports operations like zooming and panning. The user can move nodes around if they are not satisfied with the default layout. On clicking a node, all directly connected edges and nodes are highlighted. Similarly, on clicking an edge, its source and target nodes are highlighted. Moreover, clicking a node or an edge also visualizes their distribution in the \textit{Comparison View} (see \autoref{subsec:eval_panel}).
\begin{figure}[h]
\vspace{-1em}
\centering
\includegraphics[width=0.95\columnwidth]{figure/top_panel.png}
\vspace{-1 em}
\label{fig:top_chart}
\end{figure}
\textbf{Top panel.} The panel right above the causal network (as shown above) allows the selection of an edge by choosing the source and target nodes. Next to the dropdown menus are a series of buttons which allow a user to inject their prior into the system. Going from left to right, they represent operations like adding an edge, deleting an edge, reversing the edge direction
and directing an undirected edge. The toggle at the end represents the current stage (Refine/Debias) and helps transitioning from one to the other.
\begin{figure*}[tb]
\centering
\includegraphics[width=1.85\columnwidth]{figure/adult_income.png}
\setlength{\belowcaptionskip}{-2pt}
\setlength{\abovecaptionskip}{-1pt}
\caption{Case study: Adult Income dataset (a) Causal model generated using automated techniques (b) Refined causal model (c) Clicking the Gender node visualizes its distribution as a bar chart (d) Bivariate distribution between Gender and Marital Status (e) All paths from Gender to Income in the refined causal model (f) Debiased causal model (g) Evaluation metrics to compare our results against the baseline debiasing approach. }
\label{fig:adult}
\end{figure*}
\textbf{Left panel.} The bar to the left of the causal network shows the change in BIC score. This bar is updated each time the user performs operations like directing an undirected edge, adding/deleting an edge, etc. A negative value means that the change made to the causal network is in sync with the underlying dataset; positive values mean the opposite. Negative (positive) values are represented in green (red).
\textbf{Right panel.} The panel to the right of the causal network offers four functionalities. Going from top to bottom, they represent zooming in, zooming out, reset layout and changing weight of an edge.
The slider at the bottom gets activated when the user clicks on an edge during the debiasing stage. It allows the user to weaken/strengthen an edge depending on the selected value between -100\% to 100\%. Moving the slider changes $\alpha_i$ and also impacts the effective beta coefficient for the selected edge ($\alpha_i\beta_i$). This change manifests visually in the form of proportional change in the corresponding edge width. Moving the slider to -100\% will result in deletion of the selected edge.
\begin{figure}[h]
\centering
\includegraphics[scale=0.2]{figure/bottom_panel.png}
\vspace{-2 em}
\label{fig:bottom_chart}
\end{figure}
\textbf{Bottom panel.} As shown below, the bottom panel offers 4 functionalities. The ``Find Paths" button triggers depth first traversal of the causal graph to compute all directed paths between the source and target node as selected in the top panel. This will be especially helpful
when the graph is big and complex. All the computed paths are then displayed below the bottom panel. A user can click on a displayed path to highlight it and see an animated view of the path going from the source to the target node. The ``Logs" button highlights changes made to the causal network during the debiasing stage (see Fig.\ref{fig:logs}). All edges are hidden except for the newly added edges, deleted edges and edges that were weakened/strengthened. Nodes impacted by such operations ($V_{sim}$) are highlighted in grey.
If a user is interested in knowing the exact beta coefficients for all edges, the edge weights toggle will help in doing just that. By default, it is set to `hide' to enhance readability. If turned to `show', the beta coefficients will be displayed on each edge. The filter slider helps user focus on important edges by hiding edges with absolute beta coefficients less than the chosen value.
\subsection{Evaluation Panel}
\label{subsec:eval_panel}
This panel, at the right (see Fig.\ref{fig:teaser}(C)), provides different options and visual plots for comparing and evaluating the changes made to the original dataset. From the top left, users can select the sensitive variable and the ML algorithm from their respective dropdown menus. This selection will be used for computing different fairness and utility metrics.
For the sensitive variable, the dropdown menu consists of all binary categorical variables in the dataset along with a \textit{Custom Group} option. Selecting the \textit{Custom Group} option opens a new window (see Fig. \ref{fig:custom_group}) which allows the user to select groups composed of multiple attributes. This interface facilitates comparison between intersectional groups say black females and white males.
Clicking the ``Evaluate Metrics" button triggers the computation of evaluation metrics that are displayed on the right half of this panel (Performance View). It also visualizes the relationship between the sensitive attribute or selected groups and the outcome variable using a 4-fold display \cite{friendly1994fourfold} on the left half of this panel (Comparison View).
\textbf{Comparison View.} This view comprises two plots aligned vertically where the top plot represents the original dataset and the bottom plot represents the debiased dataset.
This view has two functions. It first aids the user in the initial exploration of different features and relationships. When a node or edge is clicked in the causal network, the summary statistics of the corresponding attributes is visualized (see Fig. \ref{fig:adult}(c) for an example). For binary relationships, we use either a scatter plot, a grouped bar chart or an error bar chart depending on the data type of the attributes. The second function of the Comparison View is to visualize the differences between the original and the debiased dataset. Initially, the original data is the same as the debiased data and so identical plots are displayed. However, when the user injects their prior into the system, the plots for the original and debiased datasets start to differ. We added this view to provide more transparency and interpretability to the debiasing process and also help detect sampling bias in the data.
When the user clicks the ``Evaluate Metrics" button, the Comparison View visualizes the binary relation between the sensitive attribute or selected groups and the outcome (label) variable via the 4-fold plot~\cite{friendly1994fourfold} as shown in Fig.\ref{fig:teaser}(C), left panel).
We chose a 4-fold display over a more standard brightness-coded confusion matrix since the spatial encoding aids in visual quantitative assessments. The left/right half of this display represents two groups based on the chosen sensitive variable (say males and females) or as defined in the Custom Group interface, while the top/bottom half represents different values of the output variable say getting accepted/rejected for a job. Here, symmetry along the y-axis means better group fairness.
\textbf{Performance View.}
This view houses all the evaluation metrics as specified in Sec.\ref{sec:eval_metrics}. It uses a horizontal grouped bar chart to visualize 2 utility and 5 fairness metrics. Lower values for the fairness metrics mean better fairness. Higher values for utility metrics means better utility. Data distortion is visualized using a donut chart. On hovering over any of these charts, a tooltip shows the exact values.
\section{Case Study}
\label{sec:case_study}
We demonstrate the utility of our tool for bias identification and mitigation using the Adult Income dataset. Each data point in the dataset represents a person described by 14 attributes recorded from the US 1994 census. Here, the prediction task to classify if a person's income will be greater or less than \$50k based on attributes like age, sex, education, marital status, etc. We chose this dataset as it is widely used in the algorithmic fairness literature \cite{inprocess,nipsIBM,ghai2022cascaded}. Here, we have chosen a random sample of 3000 points from this dataset for faster computation.
\begin{figure*}
\centering
\includegraphics[width=1.9\columnwidth]{figure/fourfold_stages3.png}
\setlength{\belowcaptionskip}{-8pt}
\setlength{\abovecaptionskip}{-4pt}
\caption{The above picture shows the impact of 3 types of user interaction as captured by the 4-fold display. Due to space constraints, we have only shown a subset of the causal network which connects the Gender node with the Income node. For details, please refer to the description in Section \ref{sec:case_study}.}
\label{fig:adult_fourfold}
\end{figure*}
\textbf{Generating the causal network.}
We start off by selecting the Adult Income dataset from the respective dropdown menu in the Generator panel.
We select \textit{Income} as the label variable and \textit{Work class}, \textit{Marital Status}, \textit{Race}, \textit{Gender}, \textit{Income} as the nominal variables. Next, we click on the \textit{Causal Model} button which generates the default causal model (see Fig. \ref{fig:adult} (a)).
Here, we examine different edges of the causal model and act on them as needed to reach to a reliable causal model. We start with the 7 undirected edges encoded in orange. We direct edges based on our domain knowledge like \textit{Hours per Week} $\to$ \textit{Income},
\textit{Education} $\to$ \textit{Income}, \textit{Education} $\to$ \textit{Hours per week}, etc. After each of these operations, we observe a green bar in the left panel of the Causal Network View. This indicates that the resulting causal model is a better fit over the underlying dataset.
Next, we examine other directed edges. Many of them align with our domain knowledge like \textit{Capital Gain} $\to$ \textit{Income}, \textit{Age} $\to$ \textit{Income}, etc. However, we found a couple of them to be counter-intuitive, namely \textit{Capital Gain} $\to$ \textit{Education} and \textit{Marital Status} $\to$ \textit{Gender}. In a causal relation, cause always precedes effect. Hence, immutable personal characteristics like sex, race, etc., which are assigned at birth, can not be the effect of a later life event like marriage or work class. So, in this case, we chose to reverse both these edges to get to the refined causal model (see Fig. \ref{fig:adult}(b)).
\textbf{Auditing for social biases.}
Once we have reached a reliable causal model, we start auditing for different kinds of biases. We click on different nodes and edges to explore their distributions. For eg., clicking the \textit{Gender} node visualizes its distribution in the \textit{Comparison View}. We observe that females are underrepresented in the dataset (895 females vs 2105 males) (see Fig. \ref{fig:adult} (c)).
Given this representation bias and the fact that gender pay gap is a well-known issue, we decided to investigate further.
We found an indirect path from \textit{Gender} to \textit{Income} via \textit{Hours per week}. This indicates a possible disparity in income based on gender. To probe further, we click the ``Evaluate Metrics" button with \textit{Gender} as the sensitive variable to compute different fairness metrics. We observe significant gender bias as captured by metrics like Accuracy diff (14\%), FPR diff (22\%), FNR diff (17\%), etc.
The 4-fold display for the original data in Fig. \ref{fig:adult_fourfold} (A) reveals that only 12\% of the females earn more than \$50k compared to 31\% for males. Thus there is a significant income disparity based on gender. To have a more comprehensive understanding of the issue, we search for all possible paths by selecting \textit{Gender} and \textit{Income} as the source and target from the Top panel and then clicking the ``Find Paths" button from the bottom panel. This populates 4 different causal paths below the bottom panel (see Fig. \ref{fig:adult} (e)). We will now focus on the different causal relationships in these paths and try to make minimal changes to achieve more fairness.
\textbf{Bias mitigation.} To mitigate gender bias, we first enter the debiasing stage by flipping the Stage toggle from Refine to Debias. From here on, any changes to the causal network will simulate a new debiased dataset.
Among the 4 paths we previously discovered, the top 2 paths have a common causal edge, i.e., \textit{Gender} $\to$ \textit{Marital status}. On clicking this edge, we find that most males in the dataset are married while most females are single (see Fig. \ref{fig:adult}(d)). This pattern indicates sampling bias.
Ideally, we would like no relation between these attributes so we delete this causal edge.
Next, we assess the remaining two causal paths. Based on our domain knowledge, we find the causal edge \textit{Hours per week} $\to$ \textit{Income} to be socially desirable, and the edges \textit{Gender} $\to$ \textit{Work class} and \textit{Gender} $\to$ \textit{Hours per week} to be socially undesirable. We delete the two biased edges to get to the debiased causal model (see Fig. \ref{fig:adult}(f)). To verify all changes made so far, we click on the \textit{Logs} button. As shown in Fig. \ref{fig:logs}, it shows 3 dotted lines for the removed edges and highlight the impacted attributes.
Lastly, we click on the ``Evaluate Metrics" to see the effect of our interventions. The 4-fold display for the debiased data in Fig. \ref{fig:adult_fourfold}(C) shows that the disparity between the two genders has now decreased from 19\% to 6\% (compare Fig. \ref{fig:adult_fourfold}(A)). The percentage of females who make more than \$50k have undergone a massive growth of 75\% (from 12\% to 21\%).
Moreover, as shown in Fig.~\ref{fig:adult}(g), we find that all fairness metrics have vastly improved, with only a slight decrease in the utility metrics and an elevated distortion (12\%).
These results clearly indicate the efficacy of our debiasing approach.
\textbf{Partial debiasing.} Considering the tradeoff between different metrics, one might choose to debias data partially based on their context, i.e., weaken biased edges instead of deleting them or keeping certain unfair causal paths from the sensitive variable to the label variable intact.
For eg., one might choose to delete the edge \textit{Gender} $\to$ \textit{Marital status} and weaken the edges \textit{Gender} $\to$ \textit{Work class} and \textit{Gender} $\to$ \textit{Hours per week} by 25\% and 75\%, respectively. On evaluation, we find this setup to sit somewhere between the original and the fully debiased case (see Fig. \ref{fig:adult_fourfold} (B)). It performs better on fairness than the original dataset (gap: 9\% vs 19\%) but worse than the full debiased version (gap: 9\% vs 6\%). Similarly, it incurs more distortion than the original dataset but less than the full debiased version (11\% vs 12\%).
\begin{table*}
\caption{Evaluation metrics to compare the debiased dataset generated using our tool against the baseline debiasing approach for different datasets.}
\label{tab:table}
\scriptsize%
\centering%
\begin{tabu}{%
r%
*{10}{c}%
*{2}{c}%
}
\toprule
\multirow{2}{*}{Dataset} & Sensitive & \multirow{2}{*}{version} & \multirow{2}{*}{ML model} & \multirow{2}{*}{Accuracy} & \multirow{2}{*}{F1} & Parity & Individual & Accuracy & FNR & FPR & Data \\
& attribute & & & & & difference & Bias & difference & difference & difference & Distortion \\
\midrule \\
\multirow{2}{*}{Synthetic Hiring} & \multirow{2}{*}{Gender} & baseline & \multirow{2}{*}{SVM} & 77\% & 0.59 & 11.12 & 19.09 & 4.14 & 14.26 & 6.82 & 0\% \\
& & debiased & & 77\% & 0.60 & 1.66 & 12.93 & 2.99 & 1.37 & 3.63 & 6\% \\ \\
\multirow{2}{*}{Adult Income} & \multirow{2}{*}{Gender} & baseline & Logistic & 82\% & 0.69 & 19.32 & 17.92 & 14.35 & 17.98 & 22.53 & 0\%\\
& & debiased & Regression & 75\% & 0.63 & 6.24 & 4.8 & 0.88 & 2.33 & 1.9 & 12\% \\ \\
\multirow{2}{*}{COMPAS} & \multirow{2}{*}{Race} & baseline & Random & 67\% & 0.64 & 12.07 & 33.9 & 0.17 & 23.07 & 16.89 & 0\%\\
& & debiased & Forest & 63\% & 0.59 & 11.12 & 2.19 & 0.44 & 0.68 & 1.55 & 13\% \\ \\
\bottomrule
\end{tabu}%
\vspace{-5pt}
\end{table*}
\textbf{Intersectional groups.} D-BIAS facilitates auditing for biases against intersectional groups using the ``Custom Group" option from the sensitive variable dropdown.
Here, we choose \textit{Black Females} and \textit{White Males} as the two groups. At the outset, there is a great disparity between the groups as reflected in the 4-fold display and the fairness metrics (see Fig. \ref{fig:teaser}).
As these subgroups are defined by \textit{Gender} and \textit{Race}, we focus on the unfair causal paths from these nodes to the label variable (\textit{Income}).
For debiasing, we first perform the same operations we did for gender debiasing. Thereafter, we reduce the impact of race by deleting the edges \textit{Race} $\to$ \textit{Work class} and \textit{Race} $\to$ \textit{Marital status} which we deem as socially undesirable. On evaluation (see Fig. \ref{fig:teaser}), we find a significant decrease in bias across all fairness metrics for the debiased dataset compared to the conventional debiasing practice (blue bars) which just trains the ML model with the sensitive attributes (here \textit{Gender} and \textit{Race}) simply removed.
Finally, the two 4-fold displays reveal that the participation of the disadvantaged group more than doubled, while the privileged group experienced only a modest loss.
\textbf{Exacerbating bias.} The flexibility offered by D-BIAS to refine the causal model can be misused to increase bias as well.
Bias can be exacerbated by strengthening/adding biased causal edges and weakening/deleting other relevant causal edges.
For eg.,
one can exacerbate gender bias by strengthening the edges \textit{Gender} $\to$ \textit{Marital status} and \textit{Marital status} $\to$ \textit{Income} by a 100\%. On evaluation, we find that the proportion of females making $>$\$50k has shrunk to just 1\% while the proportion of males has surged to 38\%. In effect, this has broadened the gap between males and females making more than \$50k by about 2x from 19\% to 37\% (see Fig. \ref{fig:adult_fourfold} (D)).
\textbf{Results.} Apart from the Adult Income dataset, we also tested our tool using the synthetic hiring dataset and the COMPAS recidivism dataset (see appendix D for details).
The evaluation metrics for all 3 datasets after full debiasing are reported in Table \ref{tab:table}. As we can observe, our tool is able to reduce bias significantly compared to the baseline approach across the 3 datasets for a small loss in utility and data distortion. \textit{These results validate the potential of HITL approach in mitigating bias}. It is interesting to observe that the F1 score for the synthetic hiring dataset is slightly higher than the baseline approach. However, this is line with the existing literature \cite{ghai2022cascaded} where similar instances have been recorded.
\begin{comment}
\subsection{Expert Interview}
\textcolor{blue}{
We conducted semi-structured interviews with 2 domain experts individually. Expert E1 is a faculty member in the psychology department at a reputed R1 University and expert E2 is a research staff member at a reputed research lab. Each interview lasted anywhere between 45-60 min. In the beginning, each expert was briefed about the overall objective and features of the tool. Thereafter, we demonstrated the exact functioning of the system using a case study. Lastly, we queried them about the design, effectiveness and possible future improvements. Their comments and feedback are summarized as follows.}
Both Experts E1 and E2 felt that the interface was quite \textit{intuitive}. Expert E2 commented ``\textit{It is a much needed tool and is very useful}". Both experts particularly liked the flexibility offered by the tool. If there are multiple paths from the sensitive variable to the label variable, the user can choose to retain/remove some or all the paths. Expert E1 emphasized that ``\textit{trust is critical for sensitive areas like algorithmic bias"}. The visualization and interactivity offered by D-BIAS can potentially lead to more trust.
For future improvement, expert E2 observed that there are lots of possible actions a user can take to include their prior. So, it will be nice to have tooltips for different UI compponents to enhance discoverability for different possible actions. Furthermore, Expert E2 suggested to indicate what value of a given fairness metric can be considered acceptable/non-acceptable. Expert E1 suggested to include better metrics for capturing utility as accuracy can be deceiving when output class labels are unequally distributed.
\end{comment}
\section{User Study}
\label{sec:user_study}
We conducted a user study to evaluate two primary goals: (1) usability of our tool, i.e., if participants can comprehend and interact with our tool effectively to identify and mitigate bias, (2) compare our tool with the state of the art in terms of human-centric metrics like accountability, interpretability, trust, usability, etc.
\textbf{Participants.} We recruited 10 participants aged 24-36; gender: 7 Male and 3 Female; profession: 8 graduate students and 2 software engineers.
The majority of the participants are computer science majors with no background in data visualization or algorithmic fairness. 80\% of the participants trust AI and ML technologies in general.
The participation was voluntary with no compensation.
\textbf{Baseline Tool.} For an even comparison, we looked for existing tools with a visual interface that support bias identification and mitigation. This led us to IBM’s AI Fairness 360 \cite{aif360} toolkit whose visual interface
can be publicly accessed online\footnote{https://aif360.mybluemix.net/data}. However, we didn't go further with this toolkit as the baseline (control group) because it has a significantly different look and feel which is difficult to control for. Instead, we took inspiration from this toolkit and built a baseline visual tool (not to be confused with the baseline debiasing strategy) which mimics its workflow but matches the design of our D-BIAS tool (see Fig.\ref{fig:baseline}).
IBM’s AI Fairness toolkit allows the user to choose from a set of fairness enhancing interventions with varying impact on the evaluation metrics. However, this study is focused on other important aspects such as trust, accountability, etc. So, in order to have a tightly controlled experiment, we imagine a hypothetical automated debiasing algorithm whose performance exactly matches the peak performance of our tool for all evaluation metrics. Here, peak performance refers to the state where all unfair causal edges are deleted.
Using the baseline tool is quite simple (see appendix E). The user first selects the dataset, label variable, etc.
They can then audit for different biases by selecting the sensitive attribute and then clicking on the `Check bias' button. This will compute and present a set of fairness metrics in the same fashion as the D-BIAS tool. Lastly, the user can click the `Debias \& Evaluate' button to debias the dataset and generate its evaluation metrics. A small lag is introduced before displaying evaluation metrics to mimic a real debiasing algorithm.
\textbf{Study design.} We conducted a within subject study where each participant was asked to use the baseline tool and D-BIAS in random order. The study was conducted remotely, i.e., each participant could access and interact with the tools via their own machine. For each tool, a small tutorial was given using the Synthetic Hiring dataset\footnote{We generated a synthetic hiring dataset fraught with gender and racial bias to better evaluate our tool.
For details, please refer to appendix C .} to demonstrate the workflow and features of the tool. Each participant was then given some time to explore and interact with each system. Next, the participants were asked to identify and mitigate bias for the Adult Income dataset.
For the D-BIAS tool, participants were also asked to complete a set of 5 tasks to evaluate usability. Each task was carefully designed to cover our testing goals and had a verifiable correct solution.
Tasks included: generate a causal network, direct undirected edges, identify if bias exists with respect to an attribute, identify proxy variables and finally debias the dataset. After using each tool, the participants were asked to answer a set of survey questions. Lastly, we collected subjective feedback from each participant regarding their overall experience with both the tools.
Throughout the study, participants were in constant touch with the moderator for any assistance.
Participants were encouraged to think aloud during the user study.
\textbf{Survey Design.} Each participant was asked to answer a set of 13 survey questions to quantitatively measure usability, interpretability, workload, accountability and trust. All of these questions can be answered on a 5-point Likert Scale. To capture cognitive workload, we selected two applicable questions from the NASA-LTX task load index \cite{nasa_tlx}, i.e., ``How mentally demanding was the task? and ``How hard did you have to work to accomplish your level of performance?". Participants could choose between 1 = very low to 5 = very high. For capturing usability, we picked 3 questions from the System Usability Scale (SUS) \cite{sus}. For e.g., ``I thought the system was easy to use", ``I think that I would need the support of a technical person to be able to use this system". Participants could choose between 1 = Strongly disagree to 5 = Strongly agree. To capture accountability, we asked two questions based on previous studies \cite{XAL, cai2019effects}. For e.g., ``The credit/blame for mitigating bias effectively is totally due to" (1 = System's capability, 5 = My input to the system). To capture interpretability, we consulted Madsen Gregor scale\cite{madsen2000measuring} and adopted 3 questions for our application. For e.g., ``I am satisfied with the insights and results obtained from the tool?", ``I understand how the data was debiased?" Answers could lie between 1 = Strongly disagree to 5 = Strongly agree. For measuring trust, we referred to McKnight's framework on Trust\cite{trust1, trust2} and other studies \cite{XAL, drozdal2020trust} to come up with 3 questions for our specific case. For e.g., ``I will be able to rely on this system for identifying and debiasing data" (1 = Strongly disagree, 5 = Strongly agree).
\textbf{Results.} Despite not having a background in algorithmic fairness or data visualization, all participants were able to complete all 5 tasks using the D-BIAS tool. This indicates that our tool is easy to use.
The survey data was analyzed to calculate usability, interpretability, workload, accountability and trust ratings for each tool by each participant. The mean ratings along with their standard deviations are plotted in Fig.\ref{fig:user_study}.
Using t-test, we found statistically significant differences for all measures with p$<$0.05. \textit{We found that D-BIAS outperforms the baseline tool in terms of trust, accountability and interpretability.} However, it lags in usability and cognitive workload.
So, if someone is looking for a quick fix or relies on ML algorithms more than humans, automated debiasing is the way to go. Conversely, if trust, accountability or interpretability is important, D-BIAS should be the preferred option. Looking at these results in conjunction with the results reported in Table 1, \textit{we find that our tool enhances fairness while fostering accountability, trust and interpretability.}
\textbf{Subjective feedback.}
After the study, we gathered feedback from each participant about what they liked or disliked about D-BIAS. Most participants liked the overall design, especially the causal network. We got comments like ``\textit{Interface is user friendly}", ``\textit{Causal network gives control and flexibility}", ``\textit{Causal network is very intuitive and easy to understand}", ``\textit{Causal network is a great way to understand relationships between features}". Most participants agreed that after a tutorial session, it should be fairly easy for even non-experts to play with the system. Another important aspect which received a lot of attention was our human-in-the-loop approach. Participants felt that they had a lot more control over the system and that they could change things around. One of the participants commented ``\textit{It feels like I have a say}". Some of the participants said they felt more accountable because the system offered much flexibility and that they had a choice to make.
Many of the participants strongly advocated for D-BIAS over the baseline tool. For e.g., ``\textit{D-BIAS better than automated debiasing any day}", ``\textit{D-BIAS hand's down!}". Few of the participants had a more nuanced view. They were of the opinion that the baseline tool might be the preferred option if someone is looking for a quick fix.
We also received concerns and suggestions for future improvement. Two of the participants raised concern about the tool's possible misuse if the user is biased. Another participant raised concern about scalability for larger datasets. Most participants felt that adding tooltips for different UI components especially the fairness metrics will be a great addon. Two participants wished they could see the exact changes in CSV file in the visual interface itself.
\begin{figure}[t]
\centering
\includegraphics[scale=0.50]{figure/baseline_tool_3.png}
\vspace{-8pt}
\caption{Baseline visual tool used as a benchmark in the user study.}
\label{fig:baseline}
\end{figure}
\section{Discussion, Limitations, and Future work}
\textbf{Efficacy.} The efficacy of our tool depends on how accurately the causal model captures the underlying data generating process and the ensuing refining/debiasing process. As our tool is based on causal discovery algorithms, it inherits all its assumptions such as the causal markov condition and its limitations like sampling biases, etc.\cite{challenges}.
For eg., Caucasians had a higher mean age than African Americans in the COMPAS dataset. So, the PC algorithm falsely detected a causal edge between \textit{Age} and \textit{Race} (see appendix D.2). From our domain knowledge, we can deduce that this error is due to sampling bias.
Ideally, one should use such an insight to gather additional data points for the under sampled group. However, it may not always be possible. In such cases, our tool can be leveraged to remove such patterns in the debiasing stage.
We dealt with a similar case (\textit{Gender} $\to$ \textit{Marital status}) for the Adult Income data (see Sec. \ref{sec:case_study}).
It is also worth noting that our tool is able to reduce disparity between the privileged and the unprivileged group but it may not be able to close the gap entirely. This can be due to missing proxy variables whose influence is unaccounted for or maybe because linear models are too simple to capture the relationship between a node and its parents. Future work might use non-linear SEMs and causal discovery algorithms like Fast Causal Inference (FCI)
that can better deal with missing attributes.
\textbf{Scalability.} As the size of the dataset increases in terms of features and rows, scalability can become an issue.
With dataset size, the time to generate a causal network, debiasing data, finding paths between nodes and computing evaluation metrics will all increase proportionally. Among all these steps, the process to generate the default causal network might be the most computationally expensive. So, future work should employ GPU-based parallel implementation of PC algorithm like cuPC \cite{zarebavani2019cupc} or use inherently faster causal discovery algorithms like F-GES \cite{xie2020visual} to better scale to larger datasets.
On the front end, the causal network will become big and complex as the number of features increase. With limited screen space, the user might find it difficult to comprehend the causal network. To alleviate this issue, we have implemented different visual analytics techniques like zooming, panning, filtering weak edges, finding paths, etc.
Future work might optimize graph layout algorithm and explore other visual analytics techniques like node aggregation to help navigate larger graphs better \cite{xie2020visual}.
\textbf{Applications.} In this paper, we have emphasized how our tool can help identify and remove social biases. However, our approach and tool is not limited to social biases. Our tool can incorporate human feedback to realize policy and institutional goals as well (see appendix D.1). For eg., one might strengthen the edge between the nodes \textit{Education} and \textit{Income} to implant a policy intervention where people with higher education are incentivized.
A ML model trained over the resulting dataset is likely to reflect such policy intervention in its predictions.
Next, we plan to extend our HITL methodology to tackle biases in other domains such as word embeddings.
\begin{figure} [t]
\includegraphics[width=\columnwidth]{figure/user_study_trimmed.png}
\setlength{\belowcaptionskip}{-8pt}
\setlength{\abovecaptionskip}{-4pt}
\vspace{-1em}
\caption{Mean user ratings from the survey data along with their standard deviation for different measures.
}
\label{fig:user_study}
\end{figure}
\textbf{Human factors.} Involving a human in the loop for identifying and debiasing data is a double edged sword. On one hand, it is a key strength of our tool as it provides real world domain knowledge and fosters accountability and trust. On the other hand, it can also be its main weakness if the human operating this tool intentionally/unconsciously injects social biases.
A user can misuse the system in two ways. Firstly, the user can choose to ignore the social biases inherent in the dataset by not acting on the unfair causal edges. Such behaviour renders the system ineffective. Secondly, a biased user can explicitly introduce their own biases in the system by adding/strengthening unfair causal edges.
Since this is a human aided tool, the biases that are inherent to the human user cannot be avoided. Hence, we recommend choosing the user responsibly.
Moreover, we can always check the system logs and hold the person responsible for their action/inaction.
\acknowledgments{
This work was partially funded by NSF grants CNS 1900706, IIS 1527200, IIS 1941613, and NSF SBIR contract 1926949.}
\bibliographystyle{abbrv-doi}
\section{Problem Statement}
\section{Background and Related Work}
\begin{comment}
\subsection{Causal Discovery}
Given observational data, Causal Discovery Algorithms (CDAs) find causal relationships among different attributes in the dataset \cite{glymour2019review}.
Causal relations returned by CDAs can be graphically represented via a Directed Acyclic graph (DAG), also known as causal graph, where each node represents a data attribute and each edge represents a causal relation. For example, a directed edge from X to Y signify that X is the cause of Y, i.e., a change in X will cause a change in Y. The set of causal discovery algorithms can be broadly classified into constraint based methods such as PC\cite{Spirtes2000}, Fast Causal Inference (FCI), etc. and score based algorithms such as Greedy Equivalence Search (GES)\cite{chickering2002optimal}, Fast GES \cite{fges}, etc. Among the set of possible causal DAGs, constraint based CDAs rely on clever schedule of conditional independence tests whereas score based CDAs rely on some fitness metric like BIC score to filter to the final causal graph.
In this work, we have used PC algorithm
as it can handle mixed data types and provide asymptotically correct results.
It should be noted that causal discovery algorithms operate under a set of assumptions and have their own limitations. For example, PC algorithm assumes that no confounders (direct common cause of two variables) are missing from the dataset. In a practical setting, such factors can lead to an imperfect causal graph. One way to refine such causal graphs is by injecting domain knowledge. In line with existing work \cite{Wang2016, wang2017visual, hoque2021outcome}, we have employed a visual analytics approach to augment CDAs with human knowledge to get the final causal graph.
\end{comment}
\subsection{Bias Identification}
\label{sec:identification}
The existing literature on bias identification mostly revolves around different fairness metrics. Numerous fairness metrics have been proposed which capture different facets of fairness, such as group fairness, individual fairness, counterfactual fairness, etc.
\cite{dwork2018group,aif360, bechavod2017penalizing, kusner2017counterfactual,arvindTalk}.
Another way to classify fairness metrics can be on the level they operate on. For eg., dataset based metrics are solely computed using the dataset and are independent of any ML model, such as statistical parity difference. On the other hand, classifier based metrics are computed over the predictions of a trained ML model, such as false negative rate difference.
So far, there is not a single best fairness metric. Moreover, some of the fairness metrics can be mutually incompatible, i.e., it is impossible to optimize different metrics simultaneously\cite{kleinberg2016inherent}.
In line with existing visual tools \cite{silva, aif360}, our tool also uses a diverse set of fairness metrics to present a more comprehensive picture.
Many fairness metrics solely focus on the aggregate relationship between the sensitive attribute and the output variable.
This can lead to misleading conclusions as the aggregate trend might disappear or reverse when accounting for other relevant factors. A prime example of this phenomenon, also known as Simpson's paradox \cite{Pearl2009}, is the Berkeley Graduate Admission dataset \cite{bickel1975sex}. There it appeared as if the admission process was biased against women since the overall admit rate for men (44\%) was significantly higher than for women (30\%) \cite{barocas2017fairness}. However, this correlation/association did not account for the fact that women typically applied for more competitive departments than men. After correcting for this factor, it was found that the admission process had a small but statistically significant bias \textit{in favor of} women \cite{bickel1975sex}.
Causal models can be an effective tool for dealing with such a situation as they can decipher the different intermediate factors (indirect effects) along with their respective contributions behind an aggregate trend.
Hence, our tool also employs causal model for bias identification.
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{figure/workflow.png}
\setlength{\belowcaptionskip}{-8pt}
\setlength{\abovecaptionskip}{-4pt}
\caption{The workflow of D-BIAS has 4 parts: (1) The default causal model is inferred using the PC algorithm and SEM. Here, dotted edges represent undirected edges (2) Refine stage - the default causal model is refined by directing undirected edges, reversing edges, etc. (3) Debiasing stage - the causal model is debiased by deleting/weakening biased edges (4) The debiased causal model is used to simulate the debiased dataset.}
\label{fig:workflow}
\end{figure*}
\subsection{Bias Mitigation}
\label{sec:mitigation}
The existing literature on algorithmic bias mitigation can be broadly categorized into the three different stages in which they operate within the ML pipeline, namely pre-processing, in-processing and post-processing. In the pre-processing stage, the dataset is modified such that the bias with respect to an attribute or set of attributes is reduced or eliminated \cite{zemel2013learning, kamiran2009classifying, hajian2013dataModification, nipsIBM, rajabi2021tabfairgan}. This can be achieved by either modifying the output label \cite{kamiran2012, kamiran2009classifying} or by modifying the input attributes \cite{nipsIBM, zemel2013learning}.
In the in-processing stage, the algorithm is designed to take in biased data but still generate unbiased predictions. This can be achieved by tweaking the learning objectives such that accuracy is optimized while discrimination is minimized
\cite{inprocess}. Finally, in the post-processing stage, predictions from ML algorithms are modified to counter bias
\cite{lohia2018bias}.
Our work relates closely with the pre-processing stage where we
make changes to the input attributes and the output label.
There is also a growing set of work at the intersection of bias mitigation and causality \cite{zhang2017causal, wu2018discrimination, chiappa2019path}. The general idea is to generate the causal network, modify it, and then simulate debiased data. These approaches fully rely on automated techniques to yield the true causal network, and assume a priori knowledge about fair/unfair causal relationships.
Our work draws inspiration from this line of work and presents a more general solution where human domain knowledge is leveraged to refine the causal network and generate the debiased dataset.
\subsection{Visual Tools}
Recent years have seen visual tools like \textit{Silva}\cite{silva}, \textit{FairVis}\cite{fairvis}, \textit{FairRankVis}\cite{xie2021fairrankvis}, \textit{DiscriLens}\cite{DiscriLens}, \textit{FairSight}\cite{fairsight}, \textit{WordBias}\cite{ghai2021wordbias}, etc. which are all aimed at tackling algorithmic bias. Although most of these tools are focused on bias identification, a few of them, such as FairSight, also permit debiasing. However, the debiasing strategy used in such tools is fairly basic, like eliminating proxy variable(s). Simple measures like this can lead to high data distortion and can have a high negative impact on data utility. Our work relates more closely with \textit{Silva} which also features a graphical causal model in its interface. Using an empirical study, it showed that users can interpret causal networks and found them helpful in identifying algorithmic bias. However like most other visual tools, Silva is limited to bias identification. Our work advances the state of the art by presenting a tool that supports both bias identification and mitigation, with our debiasing strategy being more nuanced and sensitive toward data distortion.
\begin{comment}
\section{Our Approach}
Most of the existing literature focuses on fully automated approaches to tackle algorithmic bias. Such approaches have raised serious concerns about interpretability, trust, accountability, etc. \textcolor{blue}{Mention fairness can't be automated.}
In this work, we propose \textit{Human in the loop} approach for bias identification and mitigation with the objective of addressing some of these concerns. Here, we have chosen graphical causal model as a medium to present underlying patterns in the data as well as to inject human domain knowledge in the system. \textcolor{blue}{why causality for algorithmic bias?} \textcolor{blue}{generalizability - better than correlation}
In the following, we provide more detailed thoughts on the added fairness, accountability, interpretability, and trust our approach can invoke.
\textbf{Fairness.} In the literature, multiple definitions \cite{arvindTalk} of fairness are proposed which can be mutually incompatible and entail unavoidable tradeoffs \cite{kleinberg2016inherent} (see also Sec. \ref{sec:identification}). There is no firm consensus on a single most appropriate definition of fairness \cite{gajane2017formalizing}. For a given application in a given context, algorithms can't decide a suitable tradeoff between fairness metrics acceptable to all stakeholders. On the other hand, a human trusted by the majority of stakeholders can take an informed decision when presented with the required information. Hence, introducing a human in the loop might improve perceived fairness.
\textbf{Interpretability.} Introducing a human in the loop will only be effective when a person is able to understand the underlying state of the system and provide useful feedback. We define interpretability as the ability to understand how bias flows from sensitive attributes and contaminates other attributes in the dataset. We have used a causal model to capture this flow
(see Fig \ref{fig:teaser}). A study by Borkin et al, \cite{borkin2013makes} showed that network visualizations are quite popular in infographics and are among the most memorable visual artifacts. Thus, causal models, explained by a straightforward network visualization, promise to be intuitive and bring more interpretability. \textcolor{blue}{Also talk about silva}
\textbf{Accountability.} Another major concern with algorithmic bias is accountability. Who should be held accountable if an algorithm does something wrong? Should software developers be held accountable or the people who curated/selected the dataset? Given that software developers themselves aren't fully aware of the behavior of the black box models under different conditions, it would not be entirely justified to blame them. Assuming that training data is the only or at least a major source of bias, having an interactive visual tool that presents the underlying state of the system to a human(s) analyst empowered to guide the debiasing process should bring more accountability.
\textbf{Trust.} People are more likely to trust a system if they can tinker with it. This was shown in \cite{dietvorst2016overcoming}, where it was found that people were more likely to use an algorithm and accept its errors if they were given the opportunity to modify it themselves, even if it meant making it perform imperfectly. Our human in the loop approach allows human(s) to supervise/participate in the entire bias mitigation process and make changes as they feel fit. Hence, our approach should also instill more trust.
We will evaluate our tool on each of these parameters and compare it against a fully automated tool (see Sec. \ref{sec:evaluation}).
\end{comment}
\section{Methodology}
The workflow of our system can be understood from Fig. \ref{fig:workflow}. A detailed discussion for each stage follows next.
\begin{comment}
Given raw tabular data as input, D-BIAS outputs its debiased version. The entire process can be understood in four steps as shown in \autoref{fig:workflow}. In step 1, the raw multivariate data is read in and is used to compute a causal model using automated techniques.
In step 2, the user interacts with the edges of the initial causal model like directing undirected edges to reach to a reliable causal model. In step 3, the user identifies and modifies biased causal relationships. Finally, in step 4, our tool computes a new dataset based on user interactions in the previous step. The user can thereafter click on `Evaluate Metrics' to compute and plot different evaluation metrics.
If the user is satisfied with the evaluation metrics, they can download the debiased dataset and safely use it for any downstream application. Otherwise, the user may return to the causal network and make further changes. This cycle continues until the user is satisfied. A detailed discussion follows next.
\end{comment}
\subsection{Generating the Causal Model}
\label{sec:causal}
A causal model generally consists of two parts: causal graph (skeleton) and statistical model. Given a tabular dataset as input, we first infer an initial causal graph (directed acyclic graph) using a popular causal discovery algorithm, namely the PC algorithm
\cite{Spirtes2000,Colombo2014Order-independentLearning}.
This algorithm relies on conditional independence tests and a set of orientation rules to yield the causal graph (for details, see appendix B). Each node in the causal graph represents a data attribute and each edge represents a causal relation. For example, a directed edge from X to Y signifies that X is the cause of Y, i.e., a change in X will cause a change in Y.
The PC algorithm provides qualitative information about the presence of causal relationship between variables. To quantitatively estimate the strength of each causal relationship,
we use linear structural equation models (SEM) where a node is modeled as a linear function of its parent nodes.
\begin{equation}
y = \sum_{i}^{parents(y)}{\beta_{i} x_i} \;+ \;\varepsilon
\end{equation}
In the above equation, variable y is modeled as a linear combination of its parents $x_i$, their regression coefficients ($\beta_i$) and the intercept term ($\varepsilon$).
If y is a numerical variable, we use linear regression else we use the multinomial logit model to estimate the values of $\beta_i$ and $\varepsilon$. Here, $\beta_i$ represents the strength of causal relation between $x_i$ and y.
We repeat this modeling process for each node with non-zero parent nodes. Such nodes that have at least one edge leading into them are termed as endogenous variables.
Other nodes correspond to exogenous variables or independent variables that have no parent nodes. In \autoref{fig:workflow} (Refined Causal Model), s and a are exogenous variables while b, c and y are endogenous variables. Here, y will be modeled as a function of its parent nodes, i.e., a, b and c. Similarly, b and c will be modeled as function of s.
After this process, we arrive at a causal graph whose different edges are parameterized using SEM. This constitutes a Structural Causal Model (SCM) or simply causal model (see Fig. \ref{fig:workflow} (1)).
This causal model, generated using automated algorithm, might have some undirected/erroneous causal edges due to different factors like sampling bias, missing attributes, etc.
To achieve a reliable causal model, we have taken a similar approach as Wang and Mueller \cite{wang2017visual} where we leverage user knowledge to refine the causal model via operations like adding, deleting, directing, and reversing causal edges (see Fig. \ref{fig:workflow} (2)). For every operation, the system computes a score (Bayesian Information Criterion (BIC)) of how well the modified causal model fits the underlying dataset. Similar to \cite{wang2017visual}, our system assists the user in refining the causal model by providing the difference in BIC score before and after the change. A negative BIC score suggests a better fit.
After achieving a reliable causal model, we enter into the debiasing stage where any changes made to the causal model will reflect on the debiased dataset (see Fig. \ref{fig:workflow} (3)).
\begin{comment}
where subscript $i$ indicates the $i$-th attribute, $\varepsilon$ is the intercept term and $r_{i}$ is the residual term. Residual is the difference between prediction and real value.
The edges for which the algorithm is not able to learn the direction are not used for regression. In causality theory, the regression coefficient $\beta_k$ serves as the measure of the causal influence from $x_k$ to $y$, and is represented as an edge weight in Fig \ref{fig:causal_diagram}.
The magnitude of $\beta_k$ is encoded as line width. The higher the magnitude, the wider the edge's width. The sign of $\beta_k$ is encoded by color. Edges with positive $\beta_k$ are green, otherwise red. The residual term $\varepsilon$ describes the part of $y$ that cannot be explained by the variable considered.
\end{comment}
\subsection{Auditing and Mitigating Social Biases}
From a causal perspective, discrimination can be defined as an unfair causal effect of the sensitive attribute (say, race) on the outcome variable (say, getting a loan)\cite{Pearl2000}. A direct causal path from a sensitive variable to the output variable constitutes disparate treatment while an unfair indirect path via some proxy attribute (say, zipcode) constitutes disparate impact \cite{chiappa2018causal}.
A direct path is certainly unfair but an indirect path may be fair or unfair (as in the case of Berkeley Admission dataset \cite{Pearl2009}). Our system computes all causal paths
and lets the user decide if a causal path is fair or unfair. If a causal path is unfair, the user should identify which constituting causal relationship(s) are unfair and act on them. The user can do this by deleting or weakening such biased causal relationships to reduce/mitigate the impact of the sensitive attribute on the outcome variable. For example, in Fig.\ref{fig:workflow} (Refined Causal Model), s is the sensitive attribute and y is the outcome variable. Here, the user deletes the edge $s \to b$ and weakens the edge $s \to c$ (shown as a thin edge). Once the user has dealt with the biased causal relationships, we achieve what we call the \textit{Debiased Causal Model}.
\begin{algorithm}
\caption{Generate Debiased Dataset}
\label{algo:1}
\begin{algorithmic}[1]
\STATE $ D \leftarrow $ Original Dataset
\STATE $G(V, E) \leftarrow $ refined causal model
\STATE $E_a \leftarrow$ set of edges added during the debiasing stage
\STATE $E_m \leftarrow $ subset of E that were deleted/strengthened/weakened during the debiasing stage
\STATE
\FOR{\textbf{each} e in $E_a$}
\STATE $n \leftarrow $node pointed by head of e
\STATE retrain linear model for n as a function of its parents
\ENDFOR
\STATE
\STATE $V_{sim} \leftarrow \emptyset$ \hfill // Attributes that need to be simulated
\FOR{\textbf{each} e in $(E_m \cup E_a)$}
\STATE $n \leftarrow $node pointed by head of e
\STATE $V_{sim} \leftarrow V_{sim} + n + all\_descendent\_nodes(n)$
\ENDFOR
\STATE $V_{sim} \leftarrow remove\_duplicates(V_{sim})$
\STATE
\STATE $D_{deb} \leftarrow \emptyset $ \hfill // Debiased Dataset
\FOR{\textbf{each} v in topological\_sort(V)}
\IF{v present in $V_{sim}$}
\STATE $D_{deb}[v] \leftarrow $ generate values based on \autoref{eq:gen}
\ELSE
\STATE $D_{deb}[v] \leftarrow D[v]$
\ENDIF
\ENDFOR
\STATE
\STATE // Rescale values for the simulated attributes
\FOR{\textbf{each} v in $V_{sim}$}
\IF{v is a categorical variable}
\STATE $D_{deb}[v] \leftarrow $ rescale values based on Algorithm 2
\ELSE
\STATE // v is a numeric variable
\STATE $\mu, \sigma^2 \leftarrow mean(D[v]), variance(D[v])$
\STATE $\mu_{deb}, \sigma_{deb}^2 \leftarrow mean(D_{deb}[v]), variance(D_{deb}[v])$
\STATE $D_{deb}[v] \leftarrow \mu + (D_{deb}[v] - \mu_{deb})/\sigma_{deb}*\sigma$
\ENDIF
\ENDFOR
\STATE \textbf{Result} $D_{deb}$
\end{algorithmic}
\end{algorithm}
\subsection{Generating the Debiased Dataset}
We simulate the debiased dataset based on the debiased causal model (see Fig. \ref{fig:workflow} (4)).
The idea is that if the user weakens/removes biased edges from the causal network, then the simulated dataset might also contain less biases. The standard way to simulate a dataset based on a causal model is to generate random numbers for the exogenous variables \cite{sofrygin2017simcausal}. Thereafter, each endogenous variable is simulated as a function of its parent nodes (variables) in the causal network.
In this work, we have adapted this procedure to suit our needs, i.e., simulating a fair dataset while having minimum distortion from the original dataset.
Our approach to generate the debiased dataset, as illustrated in Algorithm \ref{algo:1}, can be broken down into 4 steps.
At step 1 (lines 6--9), we focus on the set of edges added during the debiasing stage ($E_{a}$). We retrain regression models corresponding to each of the target nodes of $E_{a}$. This will update the weights (regression coefficients ($\beta_i$)) for all edges leading into any target node of $E_{a}$. At step 2 (lines 11--16), we identify the set of nodes (attributes) that need to be simulated ($V_{sim}$). Unlike the standard procedure, we only simulate selective nodes that are directly/indirectly impacted by the user's interaction to minimize distortion. This set includes the target nodes of all edges that the user has interacted with along with their descendent nodes. For example, in \autoref{fig:workflow}(3), the user deletes the edge $s \to b$ and weakens the edge $s \to c$. So, we will only simulate variables $b$, $c$ and $y$. At step 3 (lines 18--25), we actually simulate all nodes that are a part of $V_{sim}$ using \autoref{eq:gen}. All other nodes are left untouched and their values are simply copied from the original dataset into the debiased dataset.
It should be noted that all parent nodes must be simulated before their child node as the values for a node are computed using their parent nodes. So, we simulate all endogenous variables in a topological order.
For eg., in \autoref{fig:workflow}(4), nodes \textit{b} and \textit{c} will be simulated before node \textit{y}.
\vspace{-7pt}
\begin{equation}
\label{eq:gen}
y = \sum_{i}^{parents(y)}{\alpha_{i}\beta_{i} x_{i}} \;+\;\; \varepsilon \;\;+\; \sum_{i}^{parents(y)}{(1-\alpha_{i})\beta_{i}r_{i}}
\end{equation}
\vspace{-7pt}
In the above equation, node y is simulated as a sum of 3 terms. The first term is the weighted linear combination of parent nodes. Here, $\alpha_i$ is the scaling factor that has a default value of 1 and can range between [0, 2] as determined by user interaction.
Strengthening an edge, sets $\alpha_i>$1; weakening an edge, sets $\alpha_i<$1. For example, if the user weakens the edge between node $x_i$ and y by -35\%, then $\alpha_i$=0.65, strengthening it by +35\% will set $\alpha_i$=1.35, and deleting it will set $\alpha_i$=0. The second term is the intercept that was computed when the regression model for y was last trained. The third term adds randomness in proportion to the degree to which the user has altered an incoming edge. Here, $r_{i}$ is a normal random variable that has a similar distribution as $x_{i}$ ($r_{i} \sim \mathcal{N}({\mu_{x_{i}}, \sigma_{x_{i}}^{2}})$). This term adds fairness as it is random and alleviates distortion for y.
\begin{algorithm}[t]
\caption{Rescale Categorical Variable v}
\begin{algorithmic}[1]
\STATE D[v] $\leftarrow$ categorical variable v in the original dataset
\STATE prob\_mat $\leftarrow$ probability matrix for v computed using Eq. 2
\STATE
\STATE lr $\leftarrow $ 0.1 \hfill // learning rate
\STATE iterations $\leftarrow $ 0
\LOOP
\STATE // DPD: Discrete Probability Distribution
\STATE $dist_{ori} \leftarrow$ DPD( D[v] )
\STATE $dist_{deb} \leftarrow$ DPD( argmax(prob\_mat) )
\STATE diff $\leftarrow \sum{\left\|(dist_{ori} - dist_{deb})/dist_{deb}\right\|}$
\STATE scale\_factor $\leftarrow 1 + lr*\sum{(dist_{ori} - dist_{deb})/dist_{deb})}$
\STATE prob\_mat $\leftarrow$ scale\_factor * prob\_mat
\STATE $dist_{deb} \leftarrow$ DPD( argmax(prob\_mat) )
\STATE new\_diff $\leftarrow \sum{\left\|(dist_{ori} - dist_{deb})/dist_{deb}\right\|}$
\IF{new\_diff $>$ diff or iterations $>$ 50}
\STATE break
\ENDIF
\STATE iterations $\leftarrow$ iterations + 1
\ENDLOOP
\STATE $D_{deb}[v] \leftarrow$ argmax(prob\_mat)
\STATE \textbf{Result} $D_{deb}[v]$
\begin{comment}
def scale_prob_matrix(pmat, col_ori):
learning_rate = 0.1
iterations = 0
while True:
m1 = col_ori.value_counts(normalize=True)
m2 = pmat.idxmax(axis=1).value_counts(normalize=True)
gap = np.sum(np.abs((m1-m2)/m2))
# print(gap)
scale_fac = 1 + learning_rate*((m1-m2)/m2)
pmat = scale_fac * pmat
m2 = pmat.idxmax(axis=1).value_counts(normalize=True)
new_gap = np.sum(np.abs((m1-m2)/m2))
#print(round(new_gap - gap,10))
if new_gap > gap or iterations > 50:
break
gap = new_gap
iterations = iterations + 1
tmp_data.idxmax(axis=1).tolist()
return pmat
\end{comment}
\end{algorithmic}
\label{algo:2}
\end{algorithm}
Fig. \ref{fig:deb_eqn} illustrates the case where the node \textit{Job} has a single parent node (\textit{Gender}). Let's say that the user chooses to delete this edge ($\alpha$=0). Going the conventional route (without the third term), the attribute \textit{Job} will be reduced to a constant value (the intercept ($\varepsilon$)) which is undesirable.
Adding the (third) random term generates the \textit{Job} distribution below the `+' node which is far more balanced (fair) in terms of \textit{Gender} than the `Original' distribution on the top right. However,
the number of people getting the job (or not) is distorted compared the `Original' distribution.
This is corrected in Step 4 (lines 27--37) in Algorithm \ref{algo:1}, where we rescale each simulated variable so that its distribution is close to its original distribution. For numerical variable(s), we simply standardize values to their original mean and standard deviation. For categorical variable(s), the model returns a set of probability scores corresponding to each possible output label for each data point.
We iteratively scale such a probability matrix so that the resulting debiased distribution inches toward the original distribution (see Algorithm \ref{algo:2}).
We continue this process until a fixed number of iterations or when the difference between original and debiased distribution starts increasing. In Fig. \ref{fig:deb_eqn}, we observe that the resulting `Debiased' distribution matches the `Original' distribution in terms of \textit{Job} allocation, while maintaining gender parity.
It should be noted that simulating each attribute adds a corresponding modeling error to the process. This modeling error is typically small but it can potentially overpower the impact of the user's intervention, especially when a user makes a small change, say weakening an edge by 5\%.
In such a case, the results may not be in strict accordance with the user’s expectations.
\subsection{Evaluation Metrics}
\label{sec:eval_metrics}
Once the debiased dataset is generated, it can be evaluated using different metrics that operate at the dataset level and the classifier level. For the second case, the debiased dataset is used to train a ML model chosen by the user, and a set of metrics are computed over the model's predictions. Here, the idea is to evaluate the downstream effects of debiasing.
All the evaluation metrics can be grouped into three broad categories, namely utility, fairness and distortion. It should be noted that there might be a trade-off among the three categories. For eg., reducing bias might cause high data distortion or lower utility. For comparison, we have used a baseline debiasing strategy which just removes the sensitive attribute(s) from the dataset.
\begin{figure}[t]
\centering
\includegraphics[width=0.75\columnwidth]{figure/debiasing.png}
\setlength{\belowcaptionskip}{-4pt}
\setlength{\abovecaptionskip}{-2pt}
\caption{Illustration of the data debiasing process using a toy example where the node Job is caused by a single variable, namely gender. Green color marks the proportion of females who got the job (Y) or not (N).}
\label{fig:deb_eqn}
\end{figure}
\textbf{Fairness.} Our tool presents a diverse set of 5 popular fairness metrics, namely statistical parity difference (Parity diff), individual fairness (Ind. bias), accuracy difference (Accuracy diff), false negative rate difference (FNR diff) and false positive rate difference (FPR diff) \cite{aif360}. Two of these metrics operate at the dataset level (Parity diff, Individual bias) and the rest operate on the classifier’s predictions.
Here, Ind. bias is defined as the mean percentage of a data point's k-nearest neighbors that have a different output label. A lower value for Ind. bias is desirable as it means that similar data points have similar output labels. For the 4 other fairness metrics, we compute some statistic for the two groups say males and females, and then report their absolute difference. This statistic can be ML model dependent, such as accuracy,
or model independent, such as the likelihood of getting a positive output label.
Lower values for such metrics suggests more equality between groups. For computing model based metrics, we omit the sensitive attributes(s) and perform 3-fold cross validation using the user-specified ML model with 50:50 train test split ration, and then report the mean absolute difference between groups across the 3 folds.
\textbf{Utility.} The utility of ML models is typically measured using metrics like accuracy, F1 score, etc.
In our context, we are interested in measuring the utility of a ML model when it is trained using the debiased dataset instead of the original dataset.
To compute the utility for the original dataset, we perform 3-fold cross validation using the user-specified ML model and report the mean accuracy and F1 score. For the debiased dataset, we follow a similar procedure where we train the user-specified ML model using the debiased dataset. However, we use the outcome variable from the original dataset as the ground truth for validation. Sensitive attribute(s) are removed from both datasets before training.
Ideally, we would like the utility metrics for the debiased dataset to be close to the corresponding metrics for the original dataset.
\textbf{Data Distortion.} Data distortion is the magnitude of deviation of the debiased dataset from the original dataset.
Since the dataset can have a mix of continuous and categorical variables, we have used the \textit{Gower distance}\cite{gower} metric.
We compute the distance between corresponding rows of the original and debiased dataset, and then report the mean distance as data distortion. This metric is easy to interpret as it has a fixed lower and upper bound ([0,1]). It is 0 if the debiased dataset is the same as the original while higher values signify higher distortion. Lower values for data distortion are desirable.
\begin{figure}
\centering
\includegraphics[scale=0.25]{figure/adult_income_logs.png}
\caption{Logs view highlighting the changes made to the causal network for the Adult Income dataset. Dotted lines represent deleted edges and nodes in grey represent the impacted nodes.}
\label{fig:logs}
\vspace{-10pt}
\end{figure}
\section{The D-BIAS Tool}
\subsection{Generator Panel}
This is the first component of the tool the user interacts with (see Fig.\ref{fig:teaser} (A)). The user starts off by choosing a dataset from its dropdown menu.
The user then selects the label variable which should be a binary categorical variable as we are considering a classification setting. Next, the user selects all nominal variables which is required for fitting the SEM model. Lastly, the user chooses a p-value and clicks the `Causal Model' button to generate the causal network.
Here, the p-value is used by the PC algorithm to conduct independence tests. We set $p=0.01$ for all our demonstrations. It can be changed to 0.05 or 0.10 for smaller datasets.
The `Debiased Data' button downloads the debiased dataset.
\begin{comment}
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{figure/teaser_pic.png}
\caption{The interface of our D-BIAS visual tool using the Adult Income dataset. (A) The Generator panel: used to create the causal network and download the debiased dataset (B) The Causal Network view: shows the causal relations between the attributes of the data, allows user to inject their prior in the system (C) The Evaluation panel: used to choose the sensitive variable, the ML model and displays different evaluation metrics. }
\label{fig:teaser}
\end{figure*}
\end{comment}
\begin{figure}[t]
\centering
\includegraphics[width=0.75\columnwidth]{figure/subgroup.png}
\caption{The visual interface for selecting subgroups (Group A and B). Each column consists of a list of bar charts/histograms representing all attributes in the dataset. By default, all bars are colored gray. The user can click on multiple bars to select a subgroup. Selected bars are colored blue. Each bar is filled in proportion of their representation in the selected subgroup as a ratio of the entire dataset. In this picture, Group A consists of individuals who went to elite universities and whose age lies between 24-40. It represents 18\% of the dataset.}
\label{fig:custom_group}
\vspace{-10pt}
\end{figure}
\subsection{Causal Network View}
This is the most critical piece of the interface where the user will spend most of the time (see Fig.\ref{fig:teaser}(B)). The center of this view contains the actual causal network which is surrounded by 4 panels on all sides.
\textbf{Causal Network.} All features in the dataset are represented as nodes in the network and each edge represents a causal relation. The width of an edge encodes the magnitude of the corresponding standardized beta coefficient. It signifies the feature importance of the source node in predicting the target node. The color of an edge encodes the sign of the corresponding standardized beta coefficient. Green (red) represents positive (negative) influence of the source node on the target node. If an edge is undirected, it does not have a beta coefficient and is colored orange. Finally, gray color encode edges
that represent relationships which can not be represented by a single beta coefficient. This occurs when the target node is a categorical variable with more than 2 levels.
The causal network supports many interactions to enhance the user's overall experience and productivity. It supports operations like zooming and panning. The user can move nodes around if they are not satisfied with the default layout. On clicking a node, all directly connected edges and nodes are highlighted. Similarly, on clicking an edge, its source and target nodes are highlighted. Moreover, clicking a node or an edge also visualizes their distribution in the \textit{Comparison View} (see \autoref{subsec:eval_panel}).
\begin{figure}[h]
\vspace{-1em}
\centering
\includegraphics[width=0.95\columnwidth]{figure/top_panel.png}
\vspace{-1 em}
\label{fig:top_chart}
\end{figure}
\textbf{Top panel.} The panel right above the causal network (as shown above) allows the selection of an edge by choosing the source and target nodes. Next to the dropdown menus are a series of buttons which allow a user to inject their prior into the system. Going from left to right, they represent operations like adding an edge, deleting an edge, reversing the edge direction
and directing an undirected edge. The toggle at the end represents the current stage (Refine/Debias) and helps transitioning from one to the other.
\begin{figure*}[tb]
\centering
\includegraphics[width=1.85\columnwidth]{figure/adult_income.png}
\setlength{\belowcaptionskip}{-2pt}
\setlength{\abovecaptionskip}{-1pt}
\caption{Case study: Adult Income dataset (a) Causal model generated using automated techniques (b) Refined causal model (c) Clicking the Gender node visualizes its distribution as a bar chart (d) Bivariate distribution between Gender and Marital Status (e) All paths from Gender to Income in the refined causal model (f) Debiased causal model (g) Evaluation metrics to compare our results against the baseline debiasing approach. }
\label{fig:adult}
\end{figure*}
\textbf{Left panel.} The bar to the left of the causal network shows the change in BIC score. This bar is updated each time the user performs operations like directing an undirected edge, adding/deleting an edge, etc. A negative value means that the change made to the causal network is in sync with the underlying dataset; positive values mean the opposite. Negative (positive) values are represented in green (red).
\textbf{Right panel.} The panel to the right of the causal network offers four functionalities. Going from top to bottom, they represent zooming in, zooming out, reset layout and changing weight of an edge.
The slider at the bottom gets activated when the user clicks on an edge during the debiasing stage. It allows the user to weaken/strengthen an edge depending on the selected value between -100\% to 100\%. Moving the slider changes $\alpha_i$ and also impacts the effective beta coefficient for the selected edge ($\alpha_i\beta_i$). This change manifests visually in the form of proportional change in the corresponding edge width. Moving the slider to -100\% will result in deletion of the selected edge.
\begin{figure}[h]
\centering
\includegraphics[scale=0.2]{figure/bottom_panel.png}
\vspace{-2 em}
\label{fig:bottom_chart}
\end{figure}
\textbf{Bottom panel.} As shown below, the bottom panel offers 4 functionalities. The ``Find Paths" button triggers depth first traversal of the causal graph to compute all directed paths between the source and target node as selected in the top panel. This will be especially helpful
when the graph is big and complex. All the computed paths are then displayed below the bottom panel. A user can click on a displayed path to highlight it and see an animated view of the path going from the source to the target node. The ``Logs" button highlights changes made to the causal network during the debiasing stage (see Fig.\ref{fig:logs}). All edges are hidden except for the newly added edges, deleted edges and edges that were weakened/strengthened. Nodes impacted by such operations ($V_{sim}$) are highlighted in grey.
If a user is interested in knowing the exact beta coefficients for all edges, the edge weights toggle will help in doing just that. By default, it is set to `hide' to enhance readability. If turned to `show', the beta coefficients will be displayed on each edge. The filter slider helps user focus on important edges by hiding edges with absolute beta coefficients less than the chosen value.
\subsection{Evaluation Panel}
\label{subsec:eval_panel}
This panel, at the right (see Fig.\ref{fig:teaser}(C)), provides different options and visual plots for comparing and evaluating the changes made to the original dataset. From the top left, users can select the sensitive variable and the ML algorithm from their respective dropdown menus. This selection will be used for computing different fairness and utility metrics.
For the sensitive variable, the dropdown menu consists of all binary categorical variables in the dataset along with a \textit{Custom Group} option. Selecting the \textit{Custom Group} option opens a new window (see Fig. \ref{fig:custom_group}) which allows the user to select groups composed of multiple attributes. This interface facilitates comparison between intersectional groups say black females and white males.
Clicking the ``Evaluate Metrics" button triggers the computation of evaluation metrics that are displayed on the right half of this panel (Performance View). It also visualizes the relationship between the sensitive attribute or selected groups and the outcome variable using a 4-fold display \cite{friendly1994fourfold} on the left half of this panel (Comparison View).
\textbf{Comparison View.} This view comprises two plots aligned vertically where the top plot represents the original dataset and the bottom plot represents the debiased dataset.
This view has two functions. It first aids the user in the initial exploration of different features and relationships. When a node or edge is clicked in the causal network, the summary statistics of the corresponding attributes is visualized (see Fig. \ref{fig:adult}(c) for an example). For binary relationships, we use either a scatter plot, a grouped bar chart or an error bar chart depending on the data type of the attributes. The second function of the Comparison View is to visualize the differences between the original and the debiased dataset. Initially, the original data is the same as the debiased data and so identical plots are displayed. However, when the user injects their prior into the system, the plots for the original and debiased datasets start to differ. We added this view to provide more transparency and interpretability to the debiasing process and also help detect sampling bias in the data.
When the user clicks the ``Evaluate Metrics" button, the Comparison View visualizes the binary relation between the sensitive attribute or selected groups and the outcome (label) variable via the 4-fold plot~\cite{friendly1994fourfold} as shown in Fig.\ref{fig:teaser}(C), left panel).
We chose a 4-fold display over a more standard brightness-coded confusion matrix since the spatial encoding aids in visual quantitative assessments. The left/right half of this display represents two groups based on the chosen sensitive variable (say males and females) or as defined in the Custom Group interface, while the top/bottom half represents different values of the output variable say getting accepted/rejected for a job. Here, symmetry along the y-axis means better group fairness.
\textbf{Performance View.}
This view houses all the evaluation metrics as specified in Sec.\ref{sec:eval_metrics}. It uses a horizontal grouped bar chart to visualize 2 utility and 5 fairness metrics. Lower values for the fairness metrics mean better fairness. Higher values for utility metrics means better utility. Data distortion is visualized using a donut chart. On hovering over any of these charts, a tooltip shows the exact values.
\section{Case Study}
\label{sec:case_study}
We demonstrate the utility of our tool for bias identification and mitigation using the Adult Income dataset. Each data point in the dataset represents a person described by 14 attributes recorded from the US 1994 census. Here, the prediction task to classify if a person's income will be greater or less than \$50k based on attributes like age, sex, education, marital status, etc. We chose this dataset as it is widely used in the algorithmic fairness literature \cite{inprocess,nipsIBM,ghai2022cascaded}. Here, we have chosen a random sample of 3000 points from this dataset for faster computation.
\begin{figure*}
\centering
\includegraphics[width=1.9\columnwidth]{figure/fourfold_stages3.png}
\setlength{\belowcaptionskip}{-8pt}
\setlength{\abovecaptionskip}{-4pt}
\caption{The above picture shows the impact of 3 types of user interaction as captured by the 4-fold display. Due to space constraints, we have only shown a subset of the causal network which connects the Gender node with the Income node. For details, please refer to the description in Section \ref{sec:case_study}.}
\label{fig:adult_fourfold}
\end{figure*}
\textbf{Generating the causal network.}
We start off by selecting the Adult Income dataset from the respective dropdown menu in the Generator panel.
We select \textit{Income} as the label variable and \textit{Work class}, \textit{Marital Status}, \textit{Race}, \textit{Gender}, \textit{Income} as the nominal variables. Next, we click on the \textit{Causal Model} button which generates the default causal model (see Fig. \ref{fig:adult} (a)).
Here, we examine different edges of the causal model and act on them as needed to reach to a reliable causal model. We start with the 7 undirected edges encoded in orange. We direct edges based on our domain knowledge like \textit{Hours per Week} $\to$ \textit{Income},
\textit{Education} $\to$ \textit{Income}, \textit{Education} $\to$ \textit{Hours per week}, etc. After each of these operations, we observe a green bar in the left panel of the Causal Network View. This indicates that the resulting causal model is a better fit over the underlying dataset.
Next, we examine other directed edges. Many of them align with our domain knowledge like \textit{Capital Gain} $\to$ \textit{Income}, \textit{Age} $\to$ \textit{Income}, etc. However, we found a couple of them to be counter-intuitive, namely \textit{Capital Gain} $\to$ \textit{Education} and \textit{Marital Status} $\to$ \textit{Gender}. In a causal relation, cause always precedes effect. Hence, immutable personal characteristics like sex, race, etc., which are assigned at birth, can not be the effect of a later life event like marriage or work class. So, in this case, we chose to reverse both these edges to get to the refined causal model (see Fig. \ref{fig:adult}(b)).
\textbf{Auditing for social biases.}
Once we have reached a reliable causal model, we start auditing for different kinds of biases. We click on different nodes and edges to explore their distributions. For eg., clicking the \textit{Gender} node visualizes its distribution in the \textit{Comparison View}. We observe that females are underrepresented in the dataset (895 females vs 2105 males) (see Fig. \ref{fig:adult} (c)).
Given this representation bias and the fact that gender pay gap is a well-known issue, we decided to investigate further.
We found an indirect path from \textit{Gender} to \textit{Income} via \textit{Hours per week}. This indicates a possible disparity in income based on gender. To probe further, we click the ``Evaluate Metrics" button with \textit{Gender} as the sensitive variable to compute different fairness metrics. We observe significant gender bias as captured by metrics like Accuracy diff (14\%), FPR diff (22\%), FNR diff (17\%), etc.
The 4-fold display for the original data in Fig. \ref{fig:adult_fourfold} (A) reveals that only 12\% of the females earn more than \$50k compared to 31\% for males. Thus there is a significant income disparity based on gender. To have a more comprehensive understanding of the issue, we search for all possible paths by selecting \textit{Gender} and \textit{Income} as the source and target from the Top panel and then clicking the ``Find Paths" button from the bottom panel. This populates 4 different causal paths below the bottom panel (see Fig. \ref{fig:adult} (e)). We will now focus on the different causal relationships in these paths and try to make minimal changes to achieve more fairness.
\textbf{Bias mitigation.} To mitigate gender bias, we first enter the debiasing stage by flipping the Stage toggle from Refine to Debias. From here on, any changes to the causal network will simulate a new debiased dataset.
Among the 4 paths we previously discovered, the top 2 paths have a common causal edge, i.e., \textit{Gender} $\to$ \textit{Marital status}. On clicking this edge, we find that most males in the dataset are married while most females are single (see Fig. \ref{fig:adult}(d)). This pattern indicates sampling bias.
Ideally, we would like no relation between these attributes so we delete this causal edge.
Next, we assess the remaining two causal paths. Based on our domain knowledge, we find the causal edge \textit{Hours per week} $\to$ \textit{Income} to be socially desirable, and the edges \textit{Gender} $\to$ \textit{Work class} and \textit{Gender} $\to$ \textit{Hours per week} to be socially undesirable. We delete the two biased edges to get to the debiased causal model (see Fig. \ref{fig:adult}(f)). To verify all changes made so far, we click on the \textit{Logs} button. As shown in Fig. \ref{fig:logs}, it shows 3 dotted lines for the removed edges and highlight the impacted attributes.
Lastly, we click on the ``Evaluate Metrics" to see the effect of our interventions. The 4-fold display for the debiased data in Fig. \ref{fig:adult_fourfold}(C) shows that the disparity between the two genders has now decreased from 19\% to 6\% (compare Fig. \ref{fig:adult_fourfold}(A)). The percentage of females who make more than \$50k have undergone a massive growth of 75\% (from 12\% to 21\%).
Moreover, as shown in Fig.~\ref{fig:adult}(g), we find that all fairness metrics have vastly improved, with only a slight decrease in the utility metrics and an elevated distortion (12\%).
These results clearly indicate the efficacy of our debiasing approach.
\textbf{Partial debiasing.} Considering the tradeoff between different metrics, one might choose to debias data partially based on their context, i.e., weaken biased edges instead of deleting them or keeping certain unfair causal paths from the sensitive variable to the label variable intact.
For eg., one might choose to delete the edge \textit{Gender} $\to$ \textit{Marital status} and weaken the edges \textit{Gender} $\to$ \textit{Work class} and \textit{Gender} $\to$ \textit{Hours per week} by 25\% and 75\%, respectively. On evaluation, we find this setup to sit somewhere between the original and the fully debiased case (see Fig. \ref{fig:adult_fourfold} (B)). It performs better on fairness than the original dataset (gap: 9\% vs 19\%) but worse than the full debiased version (gap: 9\% vs 6\%). Similarly, it incurs more distortion than the original dataset but less than the full debiased version (11\% vs 12\%).
\begin{table*}
\caption{Evaluation metrics to compare the debiased dataset generated using our tool against the baseline debiasing approach for different datasets.}
\label{tab:table}
\scriptsize%
\centering%
\begin{tabu}{%
r%
*{10}{c}%
*{2}{c}%
}
\toprule
\multirow{2}{*}{Dataset} & Sensitive & \multirow{2}{*}{version} & \multirow{2}{*}{ML model} & \multirow{2}{*}{Accuracy} & \multirow{2}{*}{F1} & Parity & Individual & Accuracy & FNR & FPR & Data \\
& attribute & & & & & difference & Bias & difference & difference & difference & Distortion \\
\midrule \\
\multirow{2}{*}{Synthetic Hiring} & \multirow{2}{*}{Gender} & baseline & \multirow{2}{*}{SVM} & 77\% & 0.59 & 11.12 & 19.09 & 4.14 & 14.26 & 6.82 & 0\% \\
& & debiased & & 77\% & 0.60 & 1.66 & 12.93 & 2.99 & 1.37 & 3.63 & 6\% \\ \\
\multirow{2}{*}{Adult Income} & \multirow{2}{*}{Gender} & baseline & Logistic & 82\% & 0.69 & 19.32 & 17.92 & 14.35 & 17.98 & 22.53 & 0\%\\
& & debiased & Regression & 75\% & 0.63 & 6.24 & 4.8 & 0.88 & 2.33 & 1.9 & 12\% \\ \\
\multirow{2}{*}{COMPAS} & \multirow{2}{*}{Race} & baseline & Random & 67\% & 0.64 & 12.07 & 33.9 & 0.17 & 23.07 & 16.89 & 0\%\\
& & debiased & Forest & 63\% & 0.59 & 11.12 & 2.19 & 0.44 & 0.68 & 1.55 & 13\% \\ \\
\bottomrule
\end{tabu}%
\vspace{-5pt}
\end{table*}
\textbf{Intersectional groups.} D-BIAS facilitates auditing for biases against intersectional groups using the ``Custom Group" option from the sensitive variable dropdown.
Here, we choose \textit{Black Females} and \textit{White Males} as the two groups. At the outset, there is a great disparity between the groups as reflected in the 4-fold display and the fairness metrics (see Fig. \ref{fig:teaser}).
As these subgroups are defined by \textit{Gender} and \textit{Race}, we focus on the unfair causal paths from these nodes to the label variable (\textit{Income}).
For debiasing, we first perform the same operations we did for gender debiasing. Thereafter, we reduce the impact of race by deleting the edges \textit{Race} $\to$ \textit{Work class} and \textit{Race} $\to$ \textit{Marital status} which we deem as socially undesirable. On evaluation (see Fig. \ref{fig:teaser}), we find a significant decrease in bias across all fairness metrics for the debiased dataset compared to the conventional debiasing practice (blue bars) which just trains the ML model with the sensitive attributes (here \textit{Gender} and \textit{Race}) simply removed.
Finally, the two 4-fold displays reveal that the participation of the disadvantaged group more than doubled, while the privileged group experienced only a modest loss.
\textbf{Exacerbating bias.} The flexibility offered by D-BIAS to refine the causal model can be misused to increase bias as well.
Bias can be exacerbated by strengthening/adding biased causal edges and weakening/deleting other relevant causal edges.
For eg.,
one can exacerbate gender bias by strengthening the edges \textit{Gender} $\to$ \textit{Marital status} and \textit{Marital status} $\to$ \textit{Income} by a 100\%. On evaluation, we find that the proportion of females making $>$\$50k has shrunk to just 1\% while the proportion of males has surged to 38\%. In effect, this has broadened the gap between males and females making more than \$50k by about 2x from 19\% to 37\% (see Fig. \ref{fig:adult_fourfold} (D)).
\textbf{Results.} Apart from the Adult Income dataset, we also tested our tool using the synthetic hiring dataset and the COMPAS recidivism dataset (see appendix D for details).
The evaluation metrics for all 3 datasets after full debiasing are reported in Table \ref{tab:table}. As we can observe, our tool is able to reduce bias significantly compared to the baseline approach across the 3 datasets for a small loss in utility and data distortion. \textit{These results validate the potential of HITL approach in mitigating bias}. It is interesting to observe that the F1 score for the synthetic hiring dataset is slightly higher than the baseline approach. However, this is line with the existing literature \cite{ghai2022cascaded} where similar instances have been recorded.
\begin{comment}
\subsection{Expert Interview}
\textcolor{blue}{
We conducted semi-structured interviews with 2 domain experts individually. Expert E1 is a faculty member in the psychology department at a reputed R1 University and expert E2 is a research staff member at a reputed research lab. Each interview lasted anywhere between 45-60 min. In the beginning, each expert was briefed about the overall objective and features of the tool. Thereafter, we demonstrated the exact functioning of the system using a case study. Lastly, we queried them about the design, effectiveness and possible future improvements. Their comments and feedback are summarized as follows.}
Both Experts E1 and E2 felt that the interface was quite \textit{intuitive}. Expert E2 commented ``\textit{It is a much needed tool and is very useful}". Both experts particularly liked the flexibility offered by the tool. If there are multiple paths from the sensitive variable to the label variable, the user can choose to retain/remove some or all the paths. Expert E1 emphasized that ``\textit{trust is critical for sensitive areas like algorithmic bias"}. The visualization and interactivity offered by D-BIAS can potentially lead to more trust.
For future improvement, expert E2 observed that there are lots of possible actions a user can take to include their prior. So, it will be nice to have tooltips for different UI compponents to enhance discoverability for different possible actions. Furthermore, Expert E2 suggested to indicate what value of a given fairness metric can be considered acceptable/non-acceptable. Expert E1 suggested to include better metrics for capturing utility as accuracy can be deceiving when output class labels are unequally distributed.
\end{comment}
\section{User Study}
\label{sec:user_study}
We conducted a user study to evaluate two primary goals: (1) usability of our tool, i.e., if participants can comprehend and interact with our tool effectively to identify and mitigate bias, (2) compare our tool with the state of the art in terms of human-centric metrics like accountability, interpretability, trust, usability, etc.
\textbf{Participants.} We recruited 10 participants aged 24-36; gender: 7 Male and 3 Female; profession: 8 graduate students and 2 software engineers.
The majority of the participants are computer science majors with no background in data visualization or algorithmic fairness. 80\% of the participants trust AI and ML technologies in general.
The participation was voluntary with no compensation.
\textbf{Baseline Tool.} For an even comparison, we looked for existing tools with a visual interface that support bias identification and mitigation. This led us to IBM’s AI Fairness 360 \cite{aif360} toolkit whose visual interface
can be publicly accessed online\footnote{https://aif360.mybluemix.net/data}. However, we didn't go further with this toolkit as the baseline (control group) because it has a significantly different look and feel which is difficult to control for. Instead, we took inspiration from this toolkit and built a baseline visual tool (not to be confused with the baseline debiasing strategy) which mimics its workflow but matches the design of our D-BIAS tool (see Fig.\ref{fig:baseline}).
IBM’s AI Fairness toolkit allows the user to choose from a set of fairness enhancing interventions with varying impact on the evaluation metrics. However, this study is focused on other important aspects such as trust, accountability, etc. So, in order to have a tightly controlled experiment, we imagine a hypothetical automated debiasing algorithm whose performance exactly matches the peak performance of our tool for all evaluation metrics. Here, peak performance refers to the state where all unfair causal edges are deleted.
Using the baseline tool is quite simple (see appendix E). The user first selects the dataset, label variable, etc.
They can then audit for different biases by selecting the sensitive attribute and then clicking on the `Check bias' button. This will compute and present a set of fairness metrics in the same fashion as the D-BIAS tool. Lastly, the user can click the `Debias \& Evaluate' button to debias the dataset and generate its evaluation metrics. A small lag is introduced before displaying evaluation metrics to mimic a real debiasing algorithm.
\textbf{Study design.} We conducted a within subject study where each participant was asked to use the baseline tool and D-BIAS in random order. The study was conducted remotely, i.e., each participant could access and interact with the tools via their own machine. For each tool, a small tutorial was given using the Synthetic Hiring dataset\footnote{We generated a synthetic hiring dataset fraught with gender and racial bias to better evaluate our tool.
For details, please refer to appendix C .} to demonstrate the workflow and features of the tool. Each participant was then given some time to explore and interact with each system. Next, the participants were asked to identify and mitigate bias for the Adult Income dataset.
For the D-BIAS tool, participants were also asked to complete a set of 5 tasks to evaluate usability. Each task was carefully designed to cover our testing goals and had a verifiable correct solution.
Tasks included: generate a causal network, direct undirected edges, identify if bias exists with respect to an attribute, identify proxy variables and finally debias the dataset. After using each tool, the participants were asked to answer a set of survey questions. Lastly, we collected subjective feedback from each participant regarding their overall experience with both the tools.
Throughout the study, participants were in constant touch with the moderator for any assistance.
Participants were encouraged to think aloud during the user study.
\textbf{Survey Design.} Each participant was asked to answer a set of 13 survey questions to quantitatively measure usability, interpretability, workload, accountability and trust. All of these questions can be answered on a 5-point Likert Scale. To capture cognitive workload, we selected two applicable questions from the NASA-LTX task load index \cite{nasa_tlx}, i.e., ``How mentally demanding was the task? and ``How hard did you have to work to accomplish your level of performance?". Participants could choose between 1 = very low to 5 = very high. For capturing usability, we picked 3 questions from the System Usability Scale (SUS) \cite{sus}. For e.g., ``I thought the system was easy to use", ``I think that I would need the support of a technical person to be able to use this system". Participants could choose between 1 = Strongly disagree to 5 = Strongly agree. To capture accountability, we asked two questions based on previous studies \cite{XAL, cai2019effects}. For e.g., ``The credit/blame for mitigating bias effectively is totally due to" (1 = System's capability, 5 = My input to the system). To capture interpretability, we consulted Madsen Gregor scale\cite{madsen2000measuring} and adopted 3 questions for our application. For e.g., ``I am satisfied with the insights and results obtained from the tool?", ``I understand how the data was debiased?" Answers could lie between 1 = Strongly disagree to 5 = Strongly agree. For measuring trust, we referred to McKnight's framework on Trust\cite{trust1, trust2} and other studies \cite{XAL, drozdal2020trust} to come up with 3 questions for our specific case. For e.g., ``I will be able to rely on this system for identifying and debiasing data" (1 = Strongly disagree, 5 = Strongly agree).
\textbf{Results.} Despite not having a background in algorithmic fairness or data visualization, all participants were able to complete all 5 tasks using the D-BIAS tool. This indicates that our tool is easy to use.
The survey data was analyzed to calculate usability, interpretability, workload, accountability and trust ratings for each tool by each participant. The mean ratings along with their standard deviations are plotted in Fig.\ref{fig:user_study}.
Using t-test, we found statistically significant differences for all measures with p$<$0.05. \textit{We found that D-BIAS outperforms the baseline tool in terms of trust, accountability and interpretability.} However, it lags in usability and cognitive workload.
So, if someone is looking for a quick fix or relies on ML algorithms more than humans, automated debiasing is the way to go. Conversely, if trust, accountability or interpretability is important, D-BIAS should be the preferred option. Looking at these results in conjunction with the results reported in Table 1, \textit{we find that our tool enhances fairness while fostering accountability, trust and interpretability.}
\textbf{Subjective feedback.}
After the study, we gathered feedback from each participant about what they liked or disliked about D-BIAS. Most participants liked the overall design, especially the causal network. We got comments like ``\textit{Interface is user friendly}", ``\textit{Causal network gives control and flexibility}", ``\textit{Causal network is very intuitive and easy to understand}", ``\textit{Causal network is a great way to understand relationships between features}". Most participants agreed that after a tutorial session, it should be fairly easy for even non-experts to play with the system. Another important aspect which received a lot of attention was our human-in-the-loop approach. Participants felt that they had a lot more control over the system and that they could change things around. One of the participants commented ``\textit{It feels like I have a say}". Some of the participants said they felt more accountable because the system offered much flexibility and that they had a choice to make.
Many of the participants strongly advocated for D-BIAS over the baseline tool. For e.g., ``\textit{D-BIAS better than automated debiasing any day}", ``\textit{D-BIAS hand's down!}". Few of the participants had a more nuanced view. They were of the opinion that the baseline tool might be the preferred option if someone is looking for a quick fix.
We also received concerns and suggestions for future improvement. Two of the participants raised concern about the tool's possible misuse if the user is biased. Another participant raised concern about scalability for larger datasets. Most participants felt that adding tooltips for different UI components especially the fairness metrics will be a great addon. Two participants wished they could see the exact changes in CSV file in the visual interface itself.
\begin{figure}[t]
\centering
\includegraphics[scale=0.50]{figure/baseline_tool_3.png}
\vspace{-8pt}
\caption{Baseline visual tool used as a benchmark in the user study.}
\label{fig:baseline}
\end{figure}
\section{Discussion, Limitations, and Future work}
\textbf{Efficacy.} The efficacy of our tool depends on how accurately the causal model captures the underlying data generating process and the ensuing refining/debiasing process. As our tool is based on causal discovery algorithms, it inherits all its assumptions such as the causal markov condition and its limitations like sampling biases, etc.\cite{challenges}.
For eg., Caucasians had a higher mean age than African Americans in the COMPAS dataset. So, the PC algorithm falsely detected a causal edge between \textit{Age} and \textit{Race} (see appendix D.2). From our domain knowledge, we can deduce that this error is due to sampling bias.
Ideally, one should use such an insight to gather additional data points for the under sampled group. However, it may not always be possible. In such cases, our tool can be leveraged to remove such patterns in the debiasing stage.
We dealt with a similar case (\textit{Gender} $\to$ \textit{Marital status}) for the Adult Income data (see Sec. \ref{sec:case_study}).
It is also worth noting that our tool is able to reduce disparity between the privileged and the unprivileged group but it may not be able to close the gap entirely. This can be due to missing proxy variables whose influence is unaccounted for or maybe because linear models are too simple to capture the relationship between a node and its parents. Future work might use non-linear SEMs and causal discovery algorithms like Fast Causal Inference (FCI)
that can better deal with missing attributes.
\textbf{Scalability.} As the size of the dataset increases in terms of features and rows, scalability can become an issue.
With dataset size, the time to generate a causal network, debiasing data, finding paths between nodes and computing evaluation metrics will all increase proportionally. Among all these steps, the process to generate the default causal network might be the most computationally expensive. So, future work should employ GPU-based parallel implementation of PC algorithm like cuPC \cite{zarebavani2019cupc} or use inherently faster causal discovery algorithms like F-GES \cite{xie2020visual} to better scale to larger datasets.
On the front end, the causal network will become big and complex as the number of features increase. With limited screen space, the user might find it difficult to comprehend the causal network. To alleviate this issue, we have implemented different visual analytics techniques like zooming, panning, filtering weak edges, finding paths, etc.
Future work might optimize graph layout algorithm and explore other visual analytics techniques like node aggregation to help navigate larger graphs better \cite{xie2020visual}.
\textbf{Applications.} In this paper, we have emphasized how our tool can help identify and remove social biases. However, our approach and tool is not limited to social biases. Our tool can incorporate human feedback to realize policy and institutional goals as well (see appendix D.1). For eg., one might strengthen the edge between the nodes \textit{Education} and \textit{Income} to implant a policy intervention where people with higher education are incentivized.
A ML model trained over the resulting dataset is likely to reflect such policy intervention in its predictions.
Next, we plan to extend our HITL methodology to tackle biases in other domains such as word embeddings.
\begin{figure} [t]
\includegraphics[width=\columnwidth]{figure/user_study_trimmed.png}
\setlength{\belowcaptionskip}{-8pt}
\setlength{\abovecaptionskip}{-4pt}
\vspace{-1em}
\caption{Mean user ratings from the survey data along with their standard deviation for different measures.
}
\label{fig:user_study}
\end{figure}
\textbf{Human factors.} Involving a human in the loop for identifying and debiasing data is a double edged sword. On one hand, it is a key strength of our tool as it provides real world domain knowledge and fosters accountability and trust. On the other hand, it can also be its main weakness if the human operating this tool intentionally/unconsciously injects social biases.
A user can misuse the system in two ways. Firstly, the user can choose to ignore the social biases inherent in the dataset by not acting on the unfair causal edges. Such behaviour renders the system ineffective. Secondly, a biased user can explicitly introduce their own biases in the system by adding/strengthening unfair causal edges.
Since this is a human aided tool, the biases that are inherent to the human user cannot be avoided. Hence, we recommend choosing the user responsibly.
Moreover, we can always check the system logs and hold the person responsible for their action/inaction.
\acknowledgments{
This work was partially funded by NSF grants CNS 1900706, IIS 1527200, IIS 1941613, and NSF SBIR contract 1926949.}
\bibliographystyle{abbrv-doi}
|
\section{Introduction}
\label{sec:intro}
The relaxation or equilibration of self-gravitating systems is a ubiquitous astrophysical phenomenon that drives the formation and evolution of star-clusters, galaxies and cold dark matter halos. In quasi-equilibrium, the phase-space density of such collisionless systems can be well characterized by a distribution function (DF) which, according to the strong Jeans theorem, is a function of the conserved quantities or actions of the system. When such a system is perturbed out of equilibrium by a time-dependent gravitational perturbation, either external (e.g., encounter with another galaxy) or internal (e.g., bars or spiral arms), the original actions of the stars are modified, and the system has to re-establish a new (quasi-)equilibrium. Since disk galaxies are highly ordered, low-entropy (i.e., cold) systems, they are extremely responsive. Even small gravitational perturbations can induce oscillations in the disk, which manifest as either standing or propagating waves \citep[see][for a detailed review]{Sellwood.13}. Such oscillations consist of an initially coherent response of stars to a gravitational perturbation. This coherent response is called {\it collective} if its self-gravity is included. Over time, though, the coherence {\it dissipates}, which manifests as relaxation or equilibration and drives the system towards a new quasi-equilibrium, free of large scale oscillations. Equilibration in galactic disks is dominated by collisionless effects, including purely kinematic processes like phase-mixing (loss of coherence in the response due to different orbital frequencies of stars), and self-gravitating or collective processes like Landau damping \citep[loss of coherence due to non-dissipative damping of waves by wave-particle interactions,][]{LyndenBell.62} and violent relaxation \citep[loss of coherence due to scrambling of orbital energies in a time-varying potential,][]{LyndenBell.67}. It is noteworthy to point out that without phase-mixing neither Landau damping \citep[][]{Maoz.91} nor violent relaxation \citep[see][]{Sridhar.89} would result in equilibration. A final equilibration mechanism is chaotic mixing, the loss of coherence resulting from the exponential divergence of neighboring stars on chaotic orbits \citep[e.g.,][]{Merritt.Valluri.96,Daniel.Wyse.15,Banik.vdBosch.22}. As long as most of the phase-space is foliated with regular orbits (i.e., the Hamiltonian is near-integrable), chaotic mixing should not make a significant contribution, and phase-mixing may thus be considered the dominant equilibration mechanism.
Disk galaxies typically reveal out-of-equilibrium features due to incomplete equilibration. These may appear in the form of bars and spiral arms, which are large-scale perturbations in the radial and azimuthal directions, responsible for a slow, secular evolution of the disk. In the vertical direction, disks often reveal warps \citep[][]{Binney.92}. In the case of the Milky Way (hereafter MW) disk, which can be studied in much greater detail than any other system, recent data from astrometric and radial-velocity surveys such as SEGUE \citep[][]{Yanny.etal.09}, RAVE \citep[][]{Steinmetz.etal.06}, GALAH \citep[][]{Bland-Hawthorn.etal.19}, LAMOST \citep[][]{Cui.etal.12} and above all Gaia \citep[][]{Gaia_collab.16, Gaia_collab.18a, Gaia_collab.18b} has revealed a variety of additional vertical distortions. At large galacto-centric radii ($>10 \kpc$) this includes, among others, oscillations and corrugations \citep[][]{Xu.etal.15,Schonrich.Dehnen.18}, and streams of stars kicked up from the disk that undergo phase-mixing, sometimes referred to as `feathers' \citep[e.g.,][]{Price-Whelan.etal.15, Thomas.etal.19, Laporte.etal.22}. Similar oscillations and vertical asymmetries have also been reported in the Solar vicinity \citep[e.g.,][]{Widrow.etal.12, Williams.etal.13, Yanny.Gardner.13, Quillen.etal.18, Gaia_collab.18b, Bennett.Bovy.19, Carrillo.etal.19}. One of the most intriguing structures is the phase-space spiral discovered by \cite{Antoja.etal.18}, and studied in more detail in subsequent studies \citep[e.g.,][]{Bland-Hawthorn.etal.19,Li.Widrow.21,Li.21,Gandhi.etal.22}. Using data from Gaia DR2 \citep[][]{Gaia_collab.18a}, \cite{Antoja.etal.18} selected $\sim 900$k stars within a narrow range of galacto-centric radius and azimuthal angle centered around the Sun. When plotting the density of stars in the $(z,v_z)$-plane of vertical position, $z$, and vertical velocity, $v_z$, they noticed a faint, unexpected spiral pattern, which became more enhanced when colour-coding the $(z,v_z)$-`pixels' by the median radial or azimuthal velocities. The one-armed spiral makes 2-3 complete wraps, resembling a snail shell, and is interpreted as a signature of phase-mixing in the vertical direction following a perturbation, which \cite{Antoja.etal.18} estimate to have occurred between 300 and 900 Myr ago. More careful analyses in later studies \citep[e.g.,][etc.]{Bland-Hawthorn.etal.19,Li.21} have nailed down the interaction time to $\sim 500\Myr$ ago.
The discovery of all these oscillations in the MW disk has ushered in a new, emerging field of astrophysics, known as galactoseismology \citep[][]{Widrow.etal.12, Johnston.etal.17}. Similar to how the timbre of musical notes reveals characteristics of the instrument that produced the sound, the `ringing' of a galactic disk can (in principle) reveal its structure (both stellar disk plus dark matter halo). And similar to how the timbre can tell us whether the string of a violin was plucked (pizzicato) or bowed (arco), the ringing of a galactic disk can reveal information about the perturbation that set the disk ringing. Phase spirals are especially promising in this regard: their structure holds information about the gravitational potential in the vertical direction \citep[in particular, the vertical frequency as a function of the vertical action,][]{Antoja.etal.18} and about the type of perturbation that triggered the phase spiral \citep[e.g., bending mode vs. breathing mode, see][and Section~\ref{sec:impulsive_kick} below]{Widrow.etal.14, Darling.Widrow.19a}. In addition, by unwinding the phase spiral one can in principle determine how long ago the vertical oscillations were triggered. By studying phase spirals at multiple locations in the disk, one may even hope to use some form of triangulation to infer the direction or location from which the perturbation emerged (assuming, of course, that the phase spirals at different locations were all triggered by the same perturbation).
However promising galactoseismology may seem, many questions remain: what kind of perturbation can trigger a phase spiral? how long do phase spirals remain detectable, and what equilibration mechanism(s) causes their demise? Can we really constrain the vertical potential of the disk, or does self-gravity of the perturbation make it difficult to achieve? What kind of constraints can we infer regarding the perturber that triggered the phase spiral? Is galactoseismology likely to be confusion limited, i.e., should we expect that each location in the disk experiences oscillations arising from multiple, independent perturbations? If so, how does this impact our ability to extract useful information? Answering these questions necessitates a deep understanding of how the MW disk, and disk galaxies in general, respond to perturbations.
To date, these questions have mainly been addressed using numerical $N$-body simulations or fairly simplified analytical approaches. In particular, numerous studies have investigated how the MW disk responds to interactions with the Sagittarius (Sgr) dwarf galaxy \citep[e.g.,][]{Gomez.etal.13, Donghia.etal.16, Laporte.etal.18, Khanna.etal.19,Hunt.etal.21}. While simulations likes these have demonstrated that the interaction with Sgr can indeed spawn phase spirals in the Solar vicinity \citep[][]{Antoja.etal.18,Binney.Schonrich.18, Darling.Widrow.19b, Laporte.etal.19, Bland-Hawthorn.etal.19, Hunt.etal.21, Bennett.etal.21}, none of them have been able to produce phase spirals that match those observed in the Gaia data. As discussed in detail in \cite{Bennett.etal.21} and \cite{Bennett.Bovy.21}, this seems to suggest that the amplitude and shape of the ``Gaia snail" cannot be produced by Sgr alone. An alternative explanation, explored by \citet{Khoperskov.etal.19}, is that the Gaia snail was created by buckling of the MW's bar. However, this explanation faces its own challenges \citep[see e.g.,][]{Laporte.etal.19, Bennett.Bovy.21}. Triggering the Gaia snail with a spiral arm \citep[][]{Faure.etal.14} is also problematic, in that it requires the spiral arms to have unusually large amplitude \citep[][]{Quillen.etal.18}. Clearly then, despite a large number of studies, pinpointing the origin of the phase spiral in the Solar vicinity still remains an unsolved problem.
Although simulations have the obvious advantage that they can probe the complicated response of a perturbed disk to a realistic perturbation, which often is analytically intractable, especially if the response is large (non-linear), there are also clear disadvantages. Foremost, reaching sufficient resolution to resolve the kind of fine-structure that we can observe with data sets like Gaia requires extremely large simulations with $N > 10^8-10^9$ particles \citep[][]{Weinberg.Katz.07a, Binney.Schonrich.18, Hunt.etal.21}. Although such simulations are no longer beyond our reach \citep[see e.g.,][]{Bedorf.etal.14, Fujii.etal.19, Hunt.etal.21, Peterson.etal.22}, it is clear that using such simulations to explore large areas of parameter space remains a formidable challenge. To overcome this problem, a semi-analytical approach called the {\it backward-integrating restricted N-body method} was developed originally in the context of perturbation by bars \citep[e.g.,][]{Leeuwin.etal.93,Vauterin.Dejonghe.97,Dehnen.00}, and later on used by \cite{Hunt.Bovy.18} and \cite{Hunt.etal.19} to study non-equilibrium features in the MW caused by transient spiral arms. This method is effectively a Lagrangian formalism to solve the collisionless Boltzmann equation (hereafter CBE) by integrating test particles in the perturbed potential in a restricted N-body framework, i.e., without self-consistently developing the potential perturbation from the DF perturbation. Although appropriate for studying the local kinematic distribution of particles, this approach becomes too expensive to study the global equilibration of a system. Hence, it is important to consider alternative analytical methods that can be used to investigate the global response of a disk.
In this vein, this paper presents a rigorous, perturbative, Eulerian formalism to compute the response of a disk to perturbations. In order to gain valuable insight into the physical mechanism of phase-mixing, without resorting to the computational complexity involved in modelling a realistic disk, which we postpone to Paper~II (Banik et al., in preparation), in this first paper in the series we consider perturbations of an infinite slab with a vertical profile, but homogeneous in the lateral directions. Although a poor representation of a realistic galactic disk, this treatment captures most of the essential features of how disks respond to gravitational perturbations. We study the response of the slab to perturbers of various spatial and temporal scales, with a focus on the formation and dissolution of phase spirals resulting from the vertical oscillations and phase-mixing of stars.
This paper is organized as follows. Section~\ref{sec:linear_theory} describes the application of perturbation theory to our infinite, isothermal slab. Section~\ref{sec:impulsive_kick} then uses these results to work out the response to an impulsive, single-mode perturbation, which nicely illustrates how phase spirals originate from vertical oscillations and how they damp out due to lateral mixing. Sections~\ref{sec:localized} and~\ref{sec:non-impulsive} generalize this to responses to localized (wave packet) and non-impulsive perturbations, respectively. In Section~\ref{sec:sat_encounter} we investigate the response to satellite encounters and examine which satellite galaxies in the halo of the MW can trigger bending and/or breathing modes strong enough to trigger phase spirals at the Solar radius (still approximating the MW disk as an infinite, isothermal slab). We summarize our findings in Section~\ref{sec:concl}.
\newpage
\section{Linear perturbation theory for collisionless systems}
\label{sec:linear_theory}
\subsection{Linear perturbative formalism}
Let the unperturbed steady state distribution function (DF) of a collisionless stellar system be given by $f_0$ and the corresponding Hamiltonian be $H_0$. $f_0$ satisfies the CBE,
\begin{align}
[f_0,H_0]=0,
\end{align}
where the square brackets correspond to the Poisson bracket. Now let us introduce a small time-dependent perturbation in the potential, $\Phi_\rmP(t)$, such that the perturbed Hamiltonian becomes
\begin{align}
H=H_0+\Phi_\rmP(t)+\Phi_1(t),
\end{align}
where $\Phi_1$ is the gravitational potential sourced by the response density, $\rho_1 = \int f_1 d^3\bv$, via the Poisson equation,
\begin{align}
\nabla^2\Phi_1=4\pi G\rho_1.
\end{align}
Here $f_1$ is the linear order perturbation in the DF, i.e., the linear response of the system to the perturbation in the potential. The perturbed DF can thus be written as
\begin{align}
f=f_0+f_1.
\end{align}
Assuming that the perturbations are small such that linear perturbation theory holds, the time-evolution of $f_1$ is governed by the following linearized version of the CBE
\begin{align}
\frac{\partial f_1}{\partial t}+[f_1,H_0]+[f_0,\Phi_\rmP]+[f_0,\Phi_1]=0.
\label{CBE_perturb}
\end{align}
In this paper we shall neglect the self-gravity of the disk, i.e., neglect the polarization term, $[f_0,\Phi_1]$, in the lhs of the linearized CBE. We briefly discuss the impact of self-gravity in Section~\ref{sec::caveats}, leaving a more detailed analysis including self-gravity to a forthcoming publication.
\subsection{Hybrid perturbative formalism for an infinite slab}
\label{sec:slab}
We consider the simplified case of perturbations in an infinitely extended slab, uniform in $(x,y)$, but characterized by a vertical density profile $\rho(z)$. Although a rather poor approximation of a realistic galactic disk, this idealized case serves to highlight some of the main characteristics of disk response. We consider perturbations that can be described by a profile in the vertical $z$-direction and by a superposition of plane waves along the $x$-direction, such that $\Phi_\rmP$ and $f_1$ are both independent of $y$. After making a canonical transformation from the phase-space variables $(z,v_z)$ to the corresponding action angle variables $(I_z,w_z)$, Equation~(\ref{CBE_perturb}) becomes
\begin{align}
\frac{\partial f_1}{\partial t}+\frac{\partial H_0}{\partial I_z}\frac{\partial f_1}{\partial w_z}+\frac{\partial H_0}{\partial v_x}\frac{\partial f_1}{\partial x}-\frac{\partial \Phi_\rmP}{\partial w_z}\frac{\partial f_0}{\partial I_z}-\frac{\partial \Phi_\rmP}{\partial x}\frac{\partial f_0}{\partial v_x}=0.
\label{CBE_perturb1}
\end{align}
The unperturbed Hamiltonian $H_0$ can be written as
\begin{align}
H_0 = \frac{v^2_x+v^2_y}{2} + \frac{v^2_z}{2} + \Phi_z(z),
\end{align}
where $v_x$, $v_y$ and $v_z$ are the unperturbed velocities of stars along $x$, $y$ and $z$ respectively, and $\Phi_z(z)$ is the unperturbed potential that dictates the oscillatory vertical motion of the stars.
We expand $\Phi_\rmP$ and $f_1$ as Fourier series that are discrete along $z$ but continuous along $x$:
\begin{align}
\Phi_\rmP(z,x,t)&=\sum_{n=-\infty}^{\infty}\int d k\, \exp{\left[i (n w_z + k x)\right]}\, \Phi_{nk}(I_z,t),\nonumber \\
f_1(z,v_z,x,v_x,v_y,t)&=\sum_{n=-\infty}^{\infty}\int d k\, \exp{\left[i (n w_z + k x)\right]}\, f_{1nk}(I_z,v_x,v_y,t).
\end{align}
Here $z$ can be expressed as the following implicit function of $w_z$ and $I_z$,
\begin{align}
w_z = \Omega_z \int_0^{z} \frac{d z'}{\sqrt{2\left[E_z(I_z)-\Phi_z(z')\right]}}.
\label{z_wz_Iz}
\end{align}
where $\Omega_z=\Omega_z(I_z)$ is the vertical frequency of stars with vertical action $I_z$, given in equation~(\ref{omzvx}) below.
Here and throughout this paper we express any dependence on the continuous wave number $k$ with an index rather than an argument, i.e., $\Phi_{nk}(I_z,t)$ rather than $\Phi_n(k,I_z,t)$. This implies that any function that carries $k$ as an index is in Fourier space.
\begin{figure*}[h!]
\centering
\includegraphics[width=0.95\textwidth]{illustration6.jpg}
\caption{Illustration of the $n=0$, $n=1$ and $n=2$ plane-wave perturbation modes in a laterally uniform and vertically isothermal slab (left-hand panel) and the velocity impulses corresponding to these modes (right-hand panel) in the case of an instantaneous/impulsive perturbation. In the left-hand panel, the rectangular box indicates a random section of the slab, centered on the slab's midplane ($z=0$), while red and blue colors indicate positive and negative $\Phi_\rmP$. For clarity, this color coding is only shown at the extrema (peaks and troughs) of the mode, which has a wave-vector that is pointing in the $x$-direction. The right-hand panel shows an edge-on view of the slab, with arrows indicating the local direction of the velocity impulse caused by the instantaneous perturbation $\Phi_\rmP$, and dots marking locations in the disk where the velocity impulse is zero. Whereas the $n=0$ mode corresponds to a longitudinal perturbation, both $n=1$ and $n=2$ correspond to transverse perturbations; the former is a bending mode, while the latter is a breathing mode (note though that both these modes also cause velocity impulses in the lateral directions). Finally, `A' and `B' mark two specific locations in the slab to which we refer in the text and in Figs.~\ref{fig:impulsive_slab_n1} and~\ref{fig:impulsive_slab_n2}.}
\label{fig:modes}
\end{figure*}
We express the perturber potential and the DF perturbation or response as linear superpositions of Fourier modes. Since we do not take into account the self-gravity of the response itself, i.e., do not self-consistently solve the Poisson equation along with the CBE, these are not dynamical or normal modes of the system. In other words, the oscillation frequencies of the Fourier modes are just the unperturbed frequencies, $\Omega_z$, and do not follow a dispersion relation as in the self-gravitating case. To aid the visualization of the various Fourier modes, Fig.~\ref{fig:modes} illustrates what the $n=0$, $n=1$ and $n=2$ modes for one particular value of the wavenumber $k$ look like. The figure also indicates the direction of the velocity impulses resulting from an instantaneous perturbation of each mode.
Substitution of the above expressions in equation~(\ref{CBE_perturb1}) yields the following evolution equation for $f_{1nk}$
\begin{align}
\frac{\partial f_{1nk}}{\partial t}+i(n\Omega_z+k v_x)f_{1nk}=i\left(n\frac{\partial f_0}{\partial I_z}+k\frac{\partial f_0}{\partial v_x}\right)\Phi_{nk},
\label{f1nk_de}
\end{align}
where we have used that
\begin{align}
\Omega_z = \frac{\partial H_0}{\partial I_z}\,, \;\;\; v_x=\frac{\partial H_0}{\partial v_x}.
\label{omzvx}
\end{align}
The above first order differential equation in time is easily solved using the Green's function technique. With the initial condition, $f_{1nk}(t_\rmi)=0$, we obtain the following integral form for $f_{1nk}$ for a given time dependence of the perturber potential,
\begin{align}
f_{1nk}(I_z,v_x,v_y,t)&=i\left(n\frac{\partial f_0}{\partial I_z}+k\frac{\partial f_0}{\partial v_x}\right)\int_{t_i}^{t}d\tau \exp{\left[-i(n\Omega_z+k v_x)(t-\tau)\right]}\, \Phi_{nk}(I_z,\tau).
\label{f1nk_slabsol}
\end{align}
This solution is analogous to the particular solution for a forced oscillator with natural frequencies, $n\Omega_z$ and $k v_x$, which is being forced by an external perturber potential, $\Phi_{nk}$. The time-dependence of this external perturbation ultimately dictates the temporal evolution of the perturbation in the DF, $f_{1nk}$. A {\it net} response requires gradients in the (unperturbed) DF with respect to the actions and/or velocities. Similar solutions for the response of perturbed, collisionless systems have been derived in a number of previous studies \citep[e.g.,][]{LyndenBell.Kalnajs.72, Tremaine.Weinberg.84, Carlberg.Sellwood.85, Weinberg.89, Weinberg.91, Weinberg.04, Kaur.Sridhar.18, Banik.vdBosch.21a, Kaur.Stone.21, Chiba.Schonrich.21}, often in the context of phenomena like angular momentum transport, radial migration or dynamical friction.
\subsection{Perturbation in an isothermal slab}
\label{sec:iso_slab}
The infinite slab has a non-uniform (uniform) density profile along the vertical (horizontal) direction. Therefore the unperturbed motion of the stars is only vertically bounded by a potential but is unbounded horizontally. This implies that the unperturbed DF, $f_0$, involves a potential $\Phi_z$ only along $z$. For simplicity, we assume it to be isothermal but with different velocity dispersions in the vertical direction, $\sigma_z$, and the in-plane directions, $\sigma_x = \sigma_y \equiv \sigma$, i.e.,
\begin{align}
f_0(v_x,v_y,E_z)= \frac{\rho_c}{{(2\pi)}^{3/2}\sigma_z\,\sigma^2} \, \exp\left[-\frac{E_z}{\sigma^2_z}\right] \, \exp\left[-\frac{v^2_x+v^2_y}{2\sigma^2}\right],
\label{f_iso}
\end{align}
where
\begin{align}
E_z = \frac{1}{2} v^2_z +\Phi_z(z)
\label{E_z}
\end{align}
is the energy involving the $z$-motion. The corresponding density and potential profiles in the vertical direction are given by
\begin{align}
\rho_z(z) = \rho_c \, {\sech}^2(z/h_z),\;\;\;\;\;\;\;\;\;\;\;\Phi_z(z) = 2 \sigma^2_z \,\ln\left[\cosh(z/h_z)\right],
\end{align}
where $h_z$ is the vertical scale height \citep[][]{Spitzer.42,Camm.50}. The vertical action, $I_z$, can be obtained from the unperturbed Hamiltonian, $E_z$, as follows
\begin{align}
I_z = \frac{1}{2\pi} \oint v_z \, d z= \frac{2}{\pi} \int_0^{z_{\rm max}} \sqrt{2[E_z - \Phi_z(z)]} \, d z,
\end{align}
where $\Phi_z(z_{\rm max})=E_z$, i.e., $z_{\rm max} = h_z \cosh^{-1}\left(\exp{\left[E_z/2\sigma^2_z\right]}\right)$. The time period of vertical oscillation is given by
\begin{align}
T_z = \oint \frac{d z}{v_z} = 4\int_0^{z_{\rm max}} \frac{d z} {\sqrt{2\left[E_z-\Phi_z(z)\right]}},
\label{T_z}
\end{align}
and the vertical frequency is $\Omega_z = 2\pi/T_z$. Throughout this paper, to compute the perturbative response of the slab, we shall use typical MW parameter values, i.e., $h_z=0.4$ kpc, $\sigma_z=23$ km/s, and $\sigma=1.5\,\sigma_z=35$ km/s \citep[][]{McMillan.11}.
Substituting the above form for $f_0$ (Equation~[\ref{f_iso}]) in Equation~(\ref{f1nk_slabsol}) and using that $\Omega_z = \Omega_z(I_z) = \partial E_z/\partial I_z$ yields the following closed integral form for $f_{1nk}$:
\begin{align}
f_{1nk}(I_z,v_x,v_y,t) &= -i\left(\frac{n\Omega_z}{\sigma^2_z} + \frac{k v_x}{\sigma^2}\right) \, f_0(v_x,v_y,E_z) \, \int_{t_i}^{t}d \tau \, \exp{\left[-i(n\Omega_z+k v_x)(t-\tau)\right]} \, \Phi_{nk}(I_z,\tau).
\label{f1nk_isosol}
\end{align}
\subsection{Perturber potential}
\label{sec:per_pot}
The slab response depends on the spatio-temporal nature of the perturber. In this paper we consider two different functional forms of the perturber potential described below.
\subsubsection{Separable potential}
\label{sec:sep_per_pot}
In order to capture the essential physics of perturbative collisionless dynamics without much computational complexity, we shall consider the following separable form for the perturber potential:
\begin{align}
\Phi_\rmP(z,x,t) = \Phi_\rmN\, \calZ(z) \calX(x) \calT(t),
\label{Phip_sep}
\end{align}
where $\Phi_\rmN$ has the units of potential, and $\calZ$, $\calX$ and $\calT$ are dimensionless functions of $z$, $x$ and $t$ respectively that specify the spatio-temporal profile of $\Phi_\rmP$. Thus, the Fourier transform of $\Phi_\rmP$ can also be written in the following separable form,
\begin{align}
\Phi_{nk}(I_z,t) = \Phi_\rmN\, \calZ_n(I_z) \calX_k\, \calT(t).
\label{Phip_sep_fourier}
\end{align}
Here $\calZ_n(I_z)$ is the $n^{\rm th}$ Fourier coefficient in the discrete Fourier series expansion of $\calZ(z)$ in the vertical angle, $w_z$, given by
\begin{align}
\calZ_n(I_z) = \frac{1}{2\pi} \int_0^{2\pi} d w_z \, \calZ(z) \, \exp{\left[-in w_z\right]},
\label{Phip_sep_fourier_z}
\end{align}
where we have used the implicit expression for $z$ in terms of $w_z$ and $I_z$ given in equation~(\ref{z_wz_Iz}). $\calX_k$ is the Fourier transform of $\calX(x)$, given by
\begin{align}
\calX_k = \frac{1}{2\pi} \int_{-\infty}^{\infty} d x\, \calX(x) \, \exp{\left[-ikx\right]}.
\label{Phip_sep_fourier_x}
\end{align}
In the following sections, we investigate the slab response to perturbers with various functional forms for $\calX(x)$ and $\calT(t)$, while keeping the form for $\calZ(z)$ arbitrary. We start in Section~\ref{sec:impulsive_kick} with an impulsive ($\calT(t)=\delta(t)$) single-mode ($\calX(x) = \exp[ikx]$) perturbation, followed in Section~\ref{sec:localized} by a perturbation that is temporally impulsive but spatially localized ($\calX(x)=\exp{\left[-x^2/\Delta^2_x\right]}$). In Section~\ref{sec:non-impulsive} we consider the same spatially localized perturbation, but this time temporally extended ($\calT(t)=\exp{\left[-\omega^2_0 t^2\right]}$).
\subsubsection{Satellite galaxy along straight orbit}
\label{sec:sat_per_pot}
As a practical astrophysical application of our perturbative formalism, we also study the response of an isothermal slab to a satellite galaxy or DM subhalo undergoing an impact along a straight orbit with a uniform velocity $\vp$ at an angle $\thetap$ (with respect to the disk normal). We model the impacting satellite as a point perturber, whose potential is given by
\begin{align}
\Phi_\rmP(z,x,t) = -\frac{GM_\rmP}{\sqrt{{\left(z-\vp\cos{\thetap} t\right)}^2+{\left(x-\vp\sin{\thetap} t\right)}^2}}.
\label{Phip_sat}
\end{align}
In this case the spatial and temporal parts are coupled and thus the slab response needs to be evaluated by performing the $\tau$ integral before the $w_z$ and $x$ integrals (to find $\Phi_{nk}$), as shown in Appendix~\ref{App:sat_disk_resp}.
\section{Response to an Impulsive Perturbation}
\label{sec:impulsive_kick}
In order to gain some insight into the perturbative response of the slab, we start by solving equation~(\ref{f1nk_isosol}) for an instantaneous impulse at $t=0$; i.e., $\calT(t) = \delta(t)$. Here the normalization factor $\Phi_\rmN$ has the units of potential times time. With the initial time $t_i<0$, the integral over $\tau$ yields $\exp{\left[-i(n\Omega_z+k v_x)t\right]}$. Further integrating $f_{1nk}$ over $v_x$ and $v_y$ and summing over all $n$ modes, yields the following form for any given $k$ mode of the perturbed DF for a given action $I_z$ and angle $w_z$,
\begin{align}
f_{1k}(I_z,w_z,t)&=\sum_{n=-\infty}^{\infty}\exp{\left[in w_z\right]}\int_{-\infty}^{\infty}d v_y\int_{-\infty}^{\infty}d v_x\, f_{1nk}(I_z,v_x,v_y,t)\nonumber \\
&=A_{\rm norm}\, D_k(t)\, R_k(I_z,w_z,t),
\label{f1_delta}
\end{align}
where
\begin{align}
A_{\rm norm}=\frac{\rho_c}{\sqrt{2\pi}\sigma_z} \exp{\left[-E_z/\sigma^2_z\right]}
\end{align}
is a normalization factor reflecting the vertical structure of the unperturbed disk,
\begin{align}
D_k(t)=\exp{\left[-\frac{k^2\sigma^2t^2}{2}\right]}
\end{align}
is a damping term that describes the temporally Gaussian decay of the response by lateral mixing, and
\begin{align}
R_k(I_z,w_z,t)=-\Phi_\rmN\calX_k\sum_{n=-\infty}^{\infty}\calZ_n(I_z)\left(k^2t+i\frac{n\Omega_z}{\sigma^2_z}\right)\exp{\left[i n \left(w_z - \Omega_z\,t\right)\right]}
\label{Rk_impulse}
\end{align}
is a (linear) response function that includes vertical phase-mixing.
Equation~(\ref{f1_delta}) is the basic `building block' for computing the response of our infinite isothermal slab to a perturbation mode $k$ in the impulsive limit. Using the canonical transformation from $(w_z,I_z)$ to $(z,v_z)$, i.e., using equations~(\ref{z_wz_Iz}) and~(\ref{E_z}), we can transform $f_{1k}(I_z,w_z,t)$ to $f_{1k}(v_z,z,t)$. Upon multiplying this by $\exp{\left[ikx\right]}$ and integrating over $k$, and then integrating further over $v_z$ at a fixed $z$, one obtains the response density as a function of both time and position:
\begin{align}
\rho_1(z,x,t) &=-\frac{\rho_c \Phi_\rmN}{\sqrt{2\pi}\sigma_z} \sum_{n=-\infty}^{\infty} \int_0^{\Tilde{I}_z} d I_z\, \frac{\Omega_z}{\sqrt{2\left[E_z-\Phi_z(z)\right]}} \exp{\left[-E_z/\sigma^2_z\right]} \exp{\left[i n \left(\Tilde{w}_z - \Omega_z\,t\right)\right]} \,\calZ_n(I_z) \nonumber \\
&\times \int d k\,\exp{\left[ikx\right]}\, \exp{\left[-\frac{k^2\sigma^2t^2}{2}\right]}\left(k^2t+i\frac{n\Omega_z}{\sigma^2_z}\right)\calX_k,
\end{align}
where $\Tilde{I}_z$ is the solution of $E_z(I_z)=\Phi_z(z)$, and $\Tilde{w}_z$ is the solution for $w_z(z,I_z)$ from equation~(\ref{z_wz_Iz}).
\begin{figure*}[t!]
\centering
\includegraphics[width=0.95\textwidth]{bending_illustration_n1.jpg}
\caption{The formation of a one-armed phase spiral due to an impulsive $n=1$ bending-mode perturbation. The color-coding in the left-hand panels shows the unperturbed distribution function $f_0(z,v_z)$ (equation~[\ref{f_iso}]) in the isothermal slab at neighboring locations A (top) and B (bottom), separated by a lateral distance of $\pi/k$, with blue (red) indicating a higher (lower) phase-space density. Locations A and B coincide with extrema in the perturbation mode as depicted in Fig.~\ref{fig:modes}. The black and yellow contours indicate the phase-space trajectories for two random values of $E_z$ (or, equivalently, $I_z$). The cyan arrows indicate the velocity impulses resulting from the instantaneous perturbation at different locations in phase-space. Note that, in the case of the $n=1$ mode considered here, at the extrema A and B all velocity impulses $\Delta v_z$ are positive and negative, respectively (cf. Fig~\ref{fig:modes}). The middle panels indicate the response $f_1$ immediately following the instantaneous response (at $t=0$), with blue (red) indicating a positive (negative) response density. Finally, the right-hand panels show the response after some time $t$, computed using equation~(\ref{f1_delta}). Note how the response at A reveals a one-armed phase spiral that is exactly opposite of that at location B, i.e., they exactly cancel each other. Hence, lateral mixing causes damping of the phase spiral amplitude.}
\label{fig:impulsive_slab_n1}
\end{figure*}
In order to gain insight into the slab response for a particular $I_z$ and $w_z$, let us start by analyzing equation~(\ref{f1_delta}) for the $n=0$ mode, an in-plane density wave, for which the perturbation causes an in-plane velocity impulse as depicted in Fig.~\ref{fig:modes}. The response is a standing, longitudinal oscillation in density. The response function for this mode is $R_k(I_z,w_z,t) = \Phi_\rmN \calZ_0(I_z)\calX_k\, k^2 t$, indicating that the amplitude of oscillation initially grows linearly with time. However, this growth is inhibited by the Gaussian damping function $D_k(t) = \exp[-\half k^2 \sigma^2 t^2]$, which describes lateral mixing due to the non-zero velocity dispersion of stars in the $k$-direction. The Gaussian form of this temporal damping term owes its origin to the assumed Gaussian/Maxwellian form of the unperturbed velocity distribution along the plane. Hence, following the perturbation, the $n=0$ mode starts to grow linearly with time, but then rapidly damps away; the response loses its coherence due to mixing in the direction of the wave-vector. In the cold slab limit $(\sigma\to 0)$, without any lateral streaming motion to damp it out, the response will grow linearly in time until it eventually becomes non-linear. This is because in an infinite, laterally homogeneous slab there is no restoring force in the lateral directions, causing the stars to stream uninhibited towards (away from) the minima (maxima) of $\Phi_\rmP$ due to the initial velocity impulse induced. This leads to over- and under-density spikes which cannot be treated using linear theory. Hence, Equation~(\ref{f1_delta}) can only adequately describe the response to an instantaneous $n=0$ mode at early times, or if the damping time $\tau_\rmD \sim (k \sigma)^{-1}$ is shorter than the time-scale of formation of density spikes. The latter is roughly the time needed to cross one quarter of the perturbation's wavelength with the velocity impulse triggered at the zeroes of $\Phi_\rmP$. Therefore, in order for linear theory to be valid, we require that $\sigma > (2/\pi) \, |\Delta v|_{\rm max}$, where $|\Delta v|_{\rm max}=k\, \Phi_\rmN \calZ_0(I_z)\calX_k$. Moreover, upon including self-gravity, it can be found that the $n=0$ mode is linearly stable only for $k>k_J\approx \sqrt{4\pi G \rho_c}/\sigma$ \citep[][]{Binney.Tremaine.08}, or in other words $\lambda<\lambda_J\approx \sigma\sqrt{\pi/G \rho_c}$, where $k_J$ and $\lambda_J=2\pi/k_J$ refer to the Jeans wave-number and Jeans wavelength respectively. In the $\sigma\to 0$ limit, the Jeans wave-length, $\lambda_J\to 0$, and thus the $n=0$ mode becomes globally unstable. Hence, the condition of Jeans stability requires an additional constraint on $\sigma$: $\sigma>\sqrt{4\pi G\rho_c}/k$. The validity of linear perturbation theory thus requires that for each $k$,
\begin{align}
\sigma > \max{\left[\frac{\sqrt{4\pi G \rho_c}}{k},\frac{2k}{\pi}\Phi_\rmN \calZ_0(I_z)\calX_k\right]}.
\end{align}
\begin{figure*}[t!]
\centering
\includegraphics[width=0.95\textwidth]{breathing_illustration_n2.jpg}
\caption{Same as Fig.~\ref{fig:impulsive_slab_n1}, except for a pure $n=2$ breathing mode. Note how in this case the velocity impulses above and below the mid-plane are of opposite sign (cyan arrows in left-hand panels). As a consequence, the response density immediately following the perturbation has a quadrupole signature (middle panels), which ultimately gives rise to two-armed phase spirals (right-hand panels). Note how once again, the phase spirals at A and B are each other's additive inverse.}
\label{fig:impulsive_slab_n2}
\end{figure*}
For $n=1$, the perturbation is a standing, transverse wave on the slab, formally known as the bending wave. The perturbation induces velocity impulses in the direction perpendicular to the slab, as indicated in Fig.~\ref{fig:modes}. At the locations marked A and B, separated by a lateral distance of $\pi/k$, these velocity impulses point in the positive and negative $z$-directions, respectively. The top panels of Fig.~\ref{fig:impulsive_slab_n1} illustrate the impact this has at location A. The left-hand panels indicate the velocity impulses (cyan arrows) in the $(z,v_z)$-plane. Prior to the perturbation, due to the vertical restoring force from the slab, each star executes a periodic oscillation in this plane. The black and yellow contours indicate the corresponding phase-space trajectories for two values of $I_z$, while the heat-map indicates phase-space density (bluer colors indicate higher density). The top-middle panel shows that immediately following the impulse, the phase-space density is boosted (reduced) where $v_z>0$ ($v_z<0$), resulting in a dipole pattern in the phase-space distribution of stars. After the impulse, the stars continue to execute periodic motion in the $(z,v_z)$-plane, but starting from their new position (corresponding to a modified action $I_z$). The angular frequency of this periodic motion is $\Omega_z$, which is a function of the (modified) action, and hence, stars with different actions oscillate in the $(z,v_z)$-plane at different frequencies. As a consequence, the perturbed phase-space density shown in the middle panels is wound-up into a {\it phase spiral} of over- and under-densities as depicted in the right-most panels of Fig.~\ref{fig:impulsive_slab_n1}. The bottom panels of Fig.~\ref{fig:impulsive_slab_n1} show what happens following the impulsive perturbation at location B. Since the velocity impulses are now reversed in direction, the phase spiral that emerges is exactly the opposite of that at location A.
The creation of phase spirals is an outcome of phase-mixing in the $z$-direction and is described by the oscillatory factor, $\exp[i\,n(w_z -\Omega_z t)]$, that is part of the response function $R_k(I_z,w_z,t)$. It consists of two terms: a term that scales as $k^2 t$, which describes the lateral streaming motion of stars due to the non-zero velocity impulses in the lateral directions (see Fig.~\ref{fig:modes}), and a term that scales as $n\Omega_z/\sigma^2_z$ which purely describes the vertical oscillations. As in the case of the $n=0$ mode, the lack of a restoring force in the lateral directions\footnote{If accounting for self-gravity of the response density, there will be non-zero forces in the lateral direction, but these will promote growth rather than act as a restoring force. This ultimately leads to exponential growth (according to linear theory) of unstable modes and Landau damping of stable modes, which occurs exponentially, i.e., more slowly than the Gaussian lateral mixing in the absence of self-gravity.} causes the perturbation to grow linearly with time in the absence of lateral streaming (for a cold disk with $\sigma \approx 0$). Meanwhile, the phase spirals continue to wind-up, which implies that the vertical bending loses its coherence. Over time, phase-mixing in the vertical direction will ensure that the disk regains mirror-symmetry with respect to the midplane, but with a scale-height, $h_z$, that would be a periodic function of $x$, with a wavelength equal to $\pi /k$ (i.e., half the wave-length of the original perturbation).
However, all this ignores lateral mixing due to the unconstrained motion with non-zero velocity dispersion in the $x$ direction. Stars that received an impulse $\Delta v_z>0$ create phase spirals that are exactly the inverse of those created by neighboring stars for which the impulse was negative. Thus lateral mixing between neighboring points on the slab causes a damping of the phase spiral amplitude at any location, a process that is captured by the damping function $D_k(t)$. The lateral mixing timescale is $\tau_\rmD \sim 1/k\sigma$, indicating, as expected, that small scale perturbations (larger $k$) mix faster, and that mixing is more efficient for larger velocity dispersion in the lateral direction. After a few mixing time-scales, the slab will once again be completely homogeneous (laterally), with a scale-height $h_z$ that is independent of location. In addition, the density of stars in the $(z,v_z)$-plane will once again be perfectly symmetric without any trace of a phase spiral. The slab has completely equilibrated, and the only impact that remains of the impulsive perturbation is that the new scale-height is somewhat larger than it was originally, i.e., the impulsive perturbation has injected energy into the disk, which causes it to puff-up in the vertical direction. Hence, the final outcome is as envisioned in the impulsive-heating scenario discussed in the seminal study of \citet{Toth.Ostriker.92}. This persistent effect in the vertical density profile is however only captured in perturbation theory at second order \citep[e.g.,][]{Carlberg.Sellwood.85}; to first order the perturbation simply phase-mixes away in the impulsive limit considered here.
For $n=2$, the perturbation triggers a breathing mode, as depicted in Fig.~\ref{fig:modes}, i.e., at a given location A on the slab, the velocity impulses for this mode are positive (negative) for positive (negative) $z$. As evident from Fig.~\ref{fig:impulsive_slab_n2}, this leads to a quadrupole pattern for the initial perturbed phase-space distribution of stars, which becomes a two-armed phase spiral over time, as opposed to the one-armed phase spiral resulting from the $n=1$ mode. This reveals an important lesson: the structure of a phase spiral depends, among others, on which perturbation mode(s) are triggered. The phase spirals in regions A and B are each other's additive inverse. Hence, once again lateral mixing will cause damping of the phase spiral's amplitude, as described by the damping function $D_k(t)$. \cite{Hunt.etal.21} have shown using N-body simulations that two-armed phase spirals can indeed arise from breathing mode oscillations and that both bending and breathing modes can be excited at different locations on the MW disk by satellite-induced perturbations such as the passage of Sagittarius (see section~\ref{sec::MW_satellites} for detailed discussion).
To summarize, we see that, in case of our infinite slab, equilibration after an impulsive perturbation is driven by a combination of phase-mixing in the vertical direction and free-streaming damping in the horizontal direction. While the former gives rise to phase spirals in the $(\sqrt{I_z} \cos w_z, \sqrt{I_z} \sin w_z)$ or equivalently the $(z,v_z)$ plane, the latter causes them to damp away by lateral mixing. Due to vertical phase-mixing the phase spiral will continue to wrap itself up into a more and more tightly wound pattern, until its structure can no longer be discerned observationally due to finite-$N$ noise and measurement errors in the actions and angles of individual stars (this is an example of coarse-grain mixing). Hence, even without lateral mixing phase spirals are only detectable for a finite duration.
\section{Response to a localized perturbation}
\label{sec:localized}
In the previous section we investigated the slab response to an external disturbance with a single wavenumber $k$. Realistic perturbations are however localized in space and thus consist of many wavenumbers. In this section we shall look into what happens when the slab is hit by an impulsive perturbation that is spatially localized.
For simplicity, we assume that the external perturber behaves as a Gaussian packet with half-width $\Delta_x$ along the $x$ direction, i.e., $\Phi_\rmP$ is given by equation~(\ref{Phip_sep}) with
\begin{align}
\calX(x) = \exp{\left[-x^2/2\Delta_x^2\right]}.
\end{align}
The $\calZ(z)$ term in equation~(\ref{Phip_sep}) denotes the vertical structure of the perturber potential, which is part of what dictates the relative strength of bending and breathing mode oscillations. We shall see in the next section, though, that the relative strength of the modes is mostly dictated by the form of $\calT(t)$. For simplicity, we only consider localization along the $x$ and $z$-directions; along the $y$-direction the perturbation is assumed to extend out to infinity. We emphasize, though, that this assumption does not impact the essential physics of phase-mixing and lateral mixing discussed below.
The Fourier transform of the perturber potential, $\Phi_{nk}$, is given by equation~(\ref{Phip_sep_fourier}), with
\begin{align}
\calX_k = \frac{\Delta_x}{\sqrt{2\pi}}\, \exp{\left[-k^2\Delta_x^2/2\right]}.
\label{Phink_gaussian}
\end{align}
Upon substituting the above expression for $\calX_k$ in equation~(\ref{f1_delta}) we obtain the response for a single $k$ mode, $f_{1k}$. After multiplying this by $\exp{\left[ikx\right]}$, integrating over all $k$ and summing over all $n$ modes, we obtain the following final form for the slab response density in the case of a (laterally) Gaussian perturber,
\begin{align}
f_1(I_z,w_z,x,t) &= \sum_{n=-\infty}^{\infty} \exp{\left[i n w_z\right]} \int_{-\infty}^{\infty} d k\, \exp{\left[ikx\right]}\, f_{1k}(I_z,w_z,t) \nonumber \\
&=A_{\rm norm}\, D(x,t)\, R(I_z,w_z,x,t),
\label{f1_local_impulsive}
\end{align}
where
\begin{align}
A_{\rm norm}=\frac{\rho_c}{\sqrt{2\pi}\sigma_z} \exp{\left[-E_z/\sigma^2_z\right]}
\end{align}
is the same normalization factor as in equation~(\ref{f1_delta}),
\begin{align}
D(x,t)=\frac{\Delta_x}{\sqrt{\Delta_x^2+\sigma^2 t^2}} \exp{\left[-\frac{x^2}{2\left(\Delta_x^2+\sigma^2 t^2\right)}\right]}
\end{align}
is a factor that captures the decay of the response by lateral mixing, and
\begin{align}
R(I_z,w_z,x,t)=-\Phi_\rmN \sum_{n=-\infty}^{\infty}\calZ_n(I_z)\left[\frac{t}{\Delta_x^2+\sigma^2 t^2}\left(1-\frac{x^2}{\Delta_x^2+\sigma^2 t^2}\right)+i\frac{n\Omega_z}{\sigma^2_z}\right]\exp{\left[i n \left(w_z - \Omega_z\,t\right)\right]}
\label{R_local_impulsive}
\end{align}
with $\calZ_n(I_z)$ given by equation~(\ref{Phip_sep_fourier_z}), corresponds to the remaining part of the response that includes vertical phase-mixing.
The above expression (equation~[\ref{f1_local_impulsive}]) for the slab response to a localized disturbance has several important features. Firstly, the profile of the slab response is nearly Gaussian in $x$ since we assumed a Gaussian form (along $x$) for the perturber potential. Secondly, the $D(x,t)$ factor describes the decay of the response amplitude and widening of the response profile due to mixing by lateral streaming. The mixing in this case occurs as a power law in time rather than like a Gaussian as for a single $k$ mode (see equation~[\ref{f1_delta}]), since the power spectrum of the Gaussian perturber is dominated by small $k$ which mix very slowly, at a timescale, $\tau_\rmD\sim 1/k\sigma$. Thirdly, the $R$ factor captures two important effects: (i) a transient response reflecting an initial linear growth due to the perturber-induced velocity impulse, followed by a subsequent decay by lateral mixing, and (ii) vertical oscillations of stars (for $n\neq 0$) at different frequencies resulting in phase-mixing over time and the formation of phase spirals as described in detail in Section~\ref{sec:impulsive_kick}. The $n=0$ modes, i.e., perturbations confined to the slab, damp out faster than the non-zero $n$ modes that manifest the vertical oscillations of stars. Since the perturber was introduced impulsively by means of a Dirac delta function in time, the higher order oscillations are stronger for the same value of $\calZ_n(I_z)$ as the corresponding changes in the vertical actions have larger amplitude. Typically, for $n\geq 2$, $\calZ_n(I_z)$ gets smaller with larger $n$; hence the $n=2$ breathing mode turns out to be the dominant mode of oscillation for impulsive disturbances. The response characteristics however change as we move to non-impulsive or more temporally extended perturbers in the next section.
It takes time for the local response to propagate along the slab by lateral streaming. Initially the perturber's gravity draws in stars towards the center of impact, $x=0$. Thus, immediately following the impulse, the region near the center of impact has a larger concentration of stars, which laterally stream outwards due to non-zero velocity dispersion. This leads to a damping of the response amplitude at small $x$ and growth at large $x$, or equivalently damping and widening of the response profile, which occurs at the rate,
\begin{align}
\calD_x(t) = \frac{d}{d t} \sqrt{\Delta^2_x + \sigma^2 t^2} = \frac{\sigma^2 t}{\sqrt{\Delta^2_x + \sigma^2 t^2}}.
\end{align}
This rate of outward streaming of slab material is initially equal to
\begin{align}
\lim_{t\to 0} \calD_x(t) = \frac{\sigma^2 t}{\Delta_x},
\end{align}
but at later times asymptotes to a constant value,
\begin{align}
\lim_{t\to \infty} \calD_x(t) = \sigma.
\end{align}
To summarize, the response to a spatially localized perturbation can be understood in the context of that to a single mode plane wave perturbation discussed in the previous section, as follows. In both cases, the response involves vertical oscillations that phase-mix away, thus giving rise to phase spirals. However, whereas the plane wave response maintains its sinusoidal profile in the lateral direction with an overall Gaussian decay of the amplitude due to lateral mixing, the response profile in the case of localized perturbation changes its shape and undergoes both decay and widening. This is because in the latter case the response is a linear superposition of responses to many plane wave perturbations with different $k$, each decaying in amplitude over a time-scale, $\tau_\rmD\sim 1/k\sigma$, due to lateral mixing by free-streaming. Since the spatially Gaussian profile considered here has a Gaussian power spectrum and thus more power on large scales (small $k$) that mix more slowly, the combined response from all $k$ modes undergoes much slower lateral mixing (as a power law) than that from a single $k$ mode. The typical timescale of coarse-grained survival (against free-streaming damping) of the phase spiral in this case turns out to be $\sim (f_{\rm max}/f_{\rm res})\,\Delta_x/\sigma$. Here $f_{\rm max}$ is the maximum amplitude of the phase spiral, which is attained at $t=0$, and $f_{\rm res}$ is the resolution limit. The power law nature of free-streaming damping implies that the response to spatially and temporally localized perturbations (e.g., encounters with satellite galaxies) can be sustained in the disk for a long time.
\section{Response to a non-impulsive perturbation}
\label{sec:non-impulsive}
Thus far we have only considered impulsive perturbations of our slab, with $\calT(t)=\delta(t)$. However, a realistic disturbance would not only have a spatial structure, the effects of which we studied in the previous section, but also be extended in time. In this section we investigate the effect of non-impulsive or temporally extended disturbances on the slab oscillations. In particular, we broaden the Dirac delta pulse from the previous section into a Gaussian in time, i.e., $\Phi_\rmP$ is given by equation~(\ref{Phip_sep}) with $\calT(t) = \frac{1}{\sqrt{\pi}}\,\exp{\left[-\omega^2_0 t^2\right]}$, where $\omega_0$ is the pulse frequency. We define the pulse-width or pulse-time as $\tau_\rmP=\sqrt{2}/\omega_0$. We also assume that the pulse is localized and follows a Gaussian profile in $x$ as in the previous section, i.e., $\calX(x)=\exp{\left[-x^2/2\Delta^2_x\right]}$. As before, $\calZ(z)$ in equation~(\ref{Phip_sep}) denotes some generic vertical profile. The (spatial) Fourier transform of this potential, $\Phi_{nk}$, is provided in equation~(\ref{Phip_sep_fourier}) with $\calX_k$ given by equation~(\ref{Phink_gaussian}) and $\calZ_n$ given by equation~(\ref{Phip_sep_fourier_z}). We can substitute this in equation~(\ref{f1nk_slabsol}) and perform the integration over $\tau$ and $v_x$ to obtain the following expression for the response for a single $k$ mode,
\begin{align}
&f_{1k}(I_z,w_z,t)=A_{\rm norm}\, D_k(t)\,R_k(I_z,w_z,t),
\label{f1k_gaussian}
\end{align}
where
\begin{align}
A_{\rm norm}=\frac{\rho_c}{\sqrt{2\pi}\sigma_z} \exp{\left[-E_z/\sigma^2_z\right]}
\end{align}
is the same normalization factor as in equation~(\ref{f1_delta}),
\begin{align}
D_k(t) = \frac{\calQ^3}{2\omega_0}\, \exp{\left[-\calQ^2\frac{k^2\sigma^2t^2}{2}\right]}
\end{align}
is a factor that describes the damping of the response due to lateral mixing, and
\begin{align}
&R_{k}(I_z,w_z,t)=-\Phi_\rmN\calX_k \sum_{n=-\infty}^{\infty} \calZ_n(I_z) \left\{\,
S_{nk} \, \Upsilon_{nk}(t) \, \left(k^2 t + i \frac{n\Omega_z}{\sigma^2_z}\right) \exp{\left[i\, n(w_z-\calQ\,\Omega_z t)\right]} - \calG_{nk}(w_z,t)\right\},
\label{Rk_gaussian}
\end{align}
with $\calZ_n(I_z)$ given by equation~(\ref{Phip_sep_fourier_z}), includes the vertical phase-mixing of the response. Here $\calQ$ is a factor that depends on the pulse frequency, $\omega_0$, and the wavenumber $k$, and is given by
\begin{align}
\calQ=\calQ(\omega_0,k\sigma)=\frac{\omega_0}{\sqrt{\omega^2_0+\frac{k^2\sigma^2}{2}}}.
\end{align}
The mode-strength,
\begin{align}
S_{nk} = \exp{\left[-\frac{1}{\omega^2_0+\frac{k^2\sigma^2}{2}}\frac{n^2\Omega^2_z}{4}\right]}
\label{mode_strength_gaussian}
\end{align}
is a function that indicates the strength of each $n$ mode,
\begin{align}
\Upsilon_{nk}(t) &= 1+\erf\left\{\calQ \left(\omega_0 t-i\frac{n\Omega_z}{2\omega_0}\right)\right\}
\label{growth_gaussian}
\end{align}
describes the temporal build-up of the response and the decay of transient oscillations, and
\begin{align}
\calG_{nk}(w_z,t) = \frac{k^2}{\sqrt{\pi}\,\omega_0\calQ} \exp{\left[-\calQ^2\omega^2_0 t^2\right]} \exp{\left[in w_z\right]}
\label{transient_gaussian}
\end{align}
is another rapidly decaying transient feature. In the $\omega_0 \to \infty$ limit, both $\Upsilon_{n}(t)$ and the mode strength $S_{nk}$ become unity, and $\calG_{nk}(w_z,t) \to 0$, such that we recover the response to impulsive perturbations given in equation~(\ref{f1_delta}) as required.
It is interesting to contrast this response to an extended pulse to that in the impulsive limit. First of all, the damping factor, $D_k(t)$, which still owes its origin to lateral mixing due to non-zero velocity dispersion, now depends not only on $k$ and $\sigma$ but also on the pulse frequency $\omega_0$. The damping time is given by
\begin{align}
\tau_{\rmD} = \frac{1}{k\sigma} \sqrt{1+\frac{k^2\sigma^2}{2\omega^2_0}},
\end{align}
which scales as $\sim 1/k\sigma$ in the impulsive/short pulse ($\omega^2_0 \gg k^2\sigma^2/2$) limit indicating that the response mixes away laterally with small scale perturbations mixing faster. In the adiabatic/long pulse ($\omega^2_0 \ll k^2\sigma^2/2$) limit, though, $\tau_\rmD \to 1/\sqrt{2}\omega_0$, i.e., the damping of the response follows the temporal behaviour of the perturbing pulse itself, independent of $k$.
The mode-strength reveals several important trends: it exponentially damps away with $n^2$, implying that the lower order modes are much stronger for perturbations that are slower \citep[see also][]{Widrow.etal.14} and/or have larger wavelength (smaller $k$). Therefore the $n=1$ bending modes dominate over the $n=2$ breathing modes for a sufficiently slow pulse. Note, though, that if the pulse is too slow ($\omega_0 \to 0$) the mode strength is super-exponentially suppressed, especially at large scales (small $k$), or if the slab has a small lateral velocity dispersion, $\sigma$, compared to that along the vertical direction, $\sigma_z$ (recall that $\Omega_z\sim \sigma_z/h_z$). This is a classic signature of adiabatic shielding of the slab due to the averaging out of the net response to zero by many oscillations of stars within the (very long) perturbation timescale \citep[cf.][]{Weinberg.94a,Weinberg.94b,Gnedin.Ostriker.99}.
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{Gaussian_packet_slab_response_x.jpg}
\caption{Amplitude of the slab response to a Gaussian (in both $x$ and $t$) packet of half-width $\Delta_x=h_z$ as a function of $x$ for different times since the maximum pulse-strength. The two rows indicate two different pulse times, as indicated. We adopt our fiducial MW parameters (see Section~\ref{sec:iso_slab}) and take $I_z=0.5\, h_z\sigma_z$. Solid (dashed) lines show $n=1$ ($n=2$) bending (breathing) modes, while the grey-dotted lines show the perturbing pulse, $\calT(t)\calX(x)$. The response density initially grows and then damps away due to lateral mixing. In the short pulse limit, the response density is Gaussian in $x$, which damps out and widens like a power law in time. The response in the longer pulse behaves like a sinusoid at small $x$ (see Appendix~\ref{App:ad_lim_resp}) and its intensity shows a transient growth followed by exponential damping before it falls off as a power law. The bending (breathing) mode eventually dominates in the slow (fast) pulse limit.}
\label{fig:gaussian_slab_x}
\end{figure*}
Finally, if the perturbation is not impulsive the frequency with which the slab stars oscillate in the vertical direction is modified with respect to their natural frequency according to
\begin{align}
\Omega_z \to \frac{\omega^2_0}{\omega^2_0+\frac{k^2\sigma^2}{2}} \Omega_z,
\end{align}
which goes to $\Omega_z$ in the impulsive limit, as expected. For slower pulses however, the vertical motion of the stars couples to the lateral motion \citep[see also][]{Binney.Schonrich.18}, resulting in a reduced oscillation frequency, especially for smaller wavelengths (larger $k$). In the extremely slow/adiabatic limit, $\Omega_z \to 0$, signalling a lack of vertical phase-mixing. This is easy to understand; a forced oscillator remains in phase with the perturber if the frequency of the latter is much lower than the natural frequency. In fact, in the adiabatic limit, the response only consists of resonant stars, for which $n\Omega_z+k v_x=0$ (see Appendix~\ref{App:ad_lim_resp}), and thus no phase spiral emerges.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{Gaussian_packet_slab_response_t.jpg}
\caption{Amplitude of the slab response to a Gaussian perturbation (in both $x$ and $t$) at two locations in the slab: at the location of impact, $x = 0$, shown in the top panels, and at a distance $x = 10 h_z$ away, shown in the bottom panels. As in Fig.~\ref{fig:gaussian_slab_x}, the spatial Gaussian wave-packet, $\calX(x)$, has a half-width of $\Delta_x = h_z$. Different columns correspond to different values of the Gaussian pulse-widths, $\tau_\rmP$, as indicated. The grey-dotted line in each panel shows the perturbing pulse $\calT(t)$ at $x=0$, while solid and dashed lines show responses for the $n=1$ (bending) and $n=2$ (breathing) modes. The response to shorter pulses shows a transient growth followed by a power law fall-off with time. Response to longer pulses initially grows and then damps away as a Gaussian before finally transitioning to a power law fall-off. For longer pulses, the bending modes dominate in the long run, while for shorter pulses, the breathing modes are stronger.}
\label{fig:gaussian_slab_t}
\end{figure*}
The above response corresponds to a temporally Gaussian pulse for a fixed wavenumber $k$. To get the full response to a localized perturber, we substitute the expression for $\calX_k$ given in equation~(\ref{Phink_gaussian}), in the $k$-response of equation~(\ref{f1k_gaussian}), multiply it by $\exp{\left[ikx\right]}$ and integrate over all $k$. The resultant response is an oscillating function of $w_z$ and has a profile along $x$ which varies with time. For the short pulse/impulsive case, we recover the expression given in equation~(\ref{f1_local_impulsive}). In Fig.~\ref{fig:gaussian_slab_x} we plot the amplitude (relative to the unperturbed DF) of this oscillating response (normalized by the $z$ Fourier component of the perturber potential, $\calZ_n$) as a function of $x$. The columns correspond to four different times since the time of maximum pulse strength, and the rows correspond to two different pulse-times, as indicated. The solid and dashed lines indicate the bending ($n=1$) and breathing ($n=2$) modes, respectively. The short pulse response shown in the upper panels has a Gaussian profile centered on the point of impact at $x=0$ with the initial width very similar to that of the $\Phi_\rmP$ profile (see equations~[\ref{f1_local_impulsive}]-[\ref{R_local_impulsive}]). Over time, this response profile gets weaker and wider like a power law, as the unconstrained lateral motion of the stars causes an outward streaming, and thus decay, of the response. The long pulse response in the lower panels has a different, more extended profile than in the short pulse case; it exhibits some ripples along $x$ besides having an overall smooth behaviour (see Appendix~\ref{App:ad_lim_resp} for the response derived in the adiabatic limit). As time goes on, the response decays away and widens out due to lateral mixing. Unlike the short pulse case, here the response initially decays like $\sim \exp{\left[-\omega^2_0 t^2\right]}$ over a timescale of the pulse-time, $\tau_\rmP=\sqrt{2}/\omega_0$, before attaining a power law decay at large time.
The temporal behaviour of the response becomes even clearer in Fig.~\ref{fig:gaussian_slab_t}, where we plot the amplitude of the response as a function of time at two different positions on the slab (different rows), and for three different pulse-times (different columns). As before, the solid and dashed lines indicate the $n=1$ and $n=2$ modes, respectively. Initially the slab response grows nearly hand in hand with the perturbing pulse. This is captured by the $\Upsilon_{nk}(t)$ term (equation~[\ref{growth_gaussian}]) in the expression for $R_k(I_z,w_z,t)$, which scales as $\exp{\left[-\calQ^2 \omega^2_0 t^2\right]}$ at small $t$, but asymptotes to a constant value of $2$ at late times. As the perturber strength falls off, the response decays as a Gaussian for each $k$, as described by the damping factor, $D_k(t) \propto \exp[-{\calQ}^2k^2\sigma^2 t^2/2]$. The combined response from all $k$ however decays at a different rate. For the shortest pulse, for which the response asymptotes to that given by equation~(\ref{f1_local_impulsive}), the damping factor, $D(x,t) \propto 1/t$ at late times. In the intermediate and long pulse cases, the response initially tends to follow the same $\sim \exp{\left[-\omega^2_0 t^2\right]}$ decay as the perturbing pulse, before finally transitioning to a power law fall-off, which typically occurs as $\sim 1/t$, just as in the short pulse case. Importantly, this transition sets in later for longer lasting pulses, such that the late-time response for slower perturbations is drastically suppressed with respect to faster perturbations. From the bottom panels, it is evident that the region ($x=10 h_z$) farther away from the center of impact responds later, with a time-lag of $\Delta t= 10\, h_z/\sigma$ (timescale of lateral streaming), which is $\sim 115$ Myr for the typical MW parameter values adopted here. The breathing mode is the dominant mode in the short pulse case ($\tau_\rmP = 10 \Myr$) while in both the intermediate ($\tau_\rmP = 50 \Myr$) and long ($\tau_\rmP = 100 \Myr$) pulse scenarios the bending mode eventually dominates. Note, though, that if the pulse becomes too long, the long-term response is adiabatically suppressed. Hence, there is only a narrow window in pulse-widths for which bending modes dominate and cause a detectable response. In the next section we examine whether any of the MW satellites have encounters with the disk over timescales that fall in this regime.
The response formalism for localized, non-impulsive perturbations developed so far can be used to model the response to transient bars and spiral arms. Encounters with such features can cause transient vertical perturbations in the potential over timescales comparable to the vertical oscillation periods of stars, thereby creating phase-spirals. We discuss this in detail in Paper~II for realistic disk galaxies.
\section{Encounters with satellite galaxies}
\label{sec:sat_encounter}
In all cases considered above we have made the simplifying assumption that the perturber potential is separable, i.e., can be written in the form of equation~(\ref{Phip_sep}). However a realistic perturber is seldom of such simple form. For example, the potential due to an impacting satellite galaxy or DM subhalo (approximated as a point perturber) cannot be written in separable form, thereby making the analysis significantly more challenging. In this section, as an astrophysical application of the perturbative formalism developed in this paper, we compute the response of the infinite slab to a satellite encounter. We relegate the far more involved computation of the response of a realistic disk to an impacting satellite to Paper~II.
As shown in Appendix~\ref{App:sat_disk_resp}, the $n\neq 0$ response to a satellite impacting the slab with a uniform velocity $\vp$ along a straight orbit at an angle $\thetap$, at a distance $x$ from the point of impact, can be approximated as
\begin{align}
f_1(I_z,w_z,x,t) = \frac{\rho_c}{\sqrt{2\pi}\sigma_z} \exp{\left[-E_z/\sigma^2_z\right]}\times i\frac{2G\Mp}{\vp} \sum_{n=-\infty}^{\infty} \frac{n\Omega_z}{\sigma^2_z}\, \Psi_n(x,I_z)\, \exp{\left[i\,\frac{n\Omega_z \sin{\thetap}}{\vp}x\right]} \exp{\left[i n\left(w_z-\Omega_z t\right)\right]},
\label{f1_sat}
\end{align}
where
\begin{align}
\Psi_n(x,I_z)&= \frac{1}{2\pi} \int_0^{2\pi} d w_z\, \exp{\left[-i n \left(w_z - \frac{\Omega_z \cos{\thetap} z}{\vp}\right)\right]} K_0\left[\,\left|\frac{n\Omega_z \left(x\cos{\thetap}-z\sin{\thetap}\right)}{\vp}\right|\,\right],
\label{Psi_n}
\end{align}
with $K_0$ the zero-th order modified Bessel function of the second kind. This expression for the response is only valid far away from the point of impact ($x\gtrsim \sigma t$), such that the response can be approximated as a plane wave along $x$, and at late times, after the perturber has moved far enough away from the disk, i.e., for $t \gg (x\sin{\thetap}+z\cos{\thetap})/\vp$).
There are several salient features of this response that deserve special attention. The strength of the response is dictated by the $K_0$ function whose argument depends on $\Omega_z \cos{\thetap}\, x/\vp$ (for small $I_z$), which is basically the ratio of the encounter timescale,
\begin{align}
\tau_{\rm enc} = \frac{x\cos{\thetap}}{\vp},
\label{tau_enc}
\end{align}
and the vertical dynamical time of the stars,
\begin{align}
\tau_z = \frac{1}{\Omega_z} \sim \frac{h_z}{\sigma_z}.
\label{tau_z}
\end{align}
From the asymptotic limits of $K_0$ it follows that the response scales as a power law ($\sim \vp^{-1}$) in the impulsive ($\tau_{\rm enc}\ll \tau_z$) limit and as $\sim\exp{\left[-\left|n\Omega_z\cos{\thetap}\right|x/\vp\right]}$ in the adiabatic ($\tau_{\rm enc}\gg \tau_z$) limit. The response peaks roughly at the maximum of the $K_0$ function, which occurs when the encounter timescale is comparable to the vertical dynamical time of the stars, i.e., when $\tau_{\rm enc}\approx 0.6\,\tau_z$, or in other words the `resonance' condition,
\begin{align}
\frac{x\cos{\thetap}}{\vp} \approx \frac{0.6}{\Omega_z},
\end{align}
is satisfied. For encounters faster than this, the response is suppressed like a power law, while for slower encounters it is exponentially suppressed. The $\vp^{-1}$ scaling of the response in the impulsive limit is a well known feature of impulsive perturbations \citep[e.g.,][]{Spitzer.58, Aguilar.White.85, Weinberg.94a, Weinberg.94b, Gnedin.etal.99, Banik.vdBosch.21b}, and the exponential suppression is a telltale signature of adiabatic shielding\footnote{While the adiabatic response in one degree-of-freedom cases, e.g., the vertical phase spiral in the isothermal slab, is exponentially suppressed, that in multiple degree-of-freedom systems such as inhomogeneous disks is usually not because of resonances \citep[][]{Weinberg.94a,Weinberg.94b}.}, similar to the adiabatic suppression of the mode-strength factor in the response to slow Gaussian pulses discussed in section~\ref{sec:non-impulsive}.
While the response is heavily damped for very slow encounters, something interesting happens in the mildly slow regime, $\tau_{\rm enc}=x\cos{\thetap}/\vp \gtrsim \tau_z$. In this regime, the ratio of the $n=2$ breathing to the $n=1$ bending mode response scales as
\begin{align}
f_{21} \equiv {f_{1,n=2} \over f_{1,n=1}} \sim \sqrt{2}\,\exp{\left[-\frac{\Omega_z\cos{\thetap}\,x}{\vp}\right]}.
\label{f21}
\end{align}
Thus the bending mode response exponentially dominates over that of the breathing mode for slower (smaller $\vp$), more distant (large $x$), and more perpendicular ($\thetap\approx 0$) encounters. The bending mode is also more pronounced for stars with larger $\Omega_z$ or equivalently smaller $I_z$. On the other hand, for encounters with $\tau_{\rm enc}=x\cos{\thetap}/\vp < \tau_z$, the breathing modes dominate.
Finally, the slab response to the impacting satellite, given in equation~(\ref{f1_sat}), consists of oscillating functions of time, lateral distance $x$, and the vertical oscillation amplitude, $\sqrt{2I_z/\nu}$ (see equations~[\ref{Psi_n_app_epi}] and [\ref{Phin_sat_app}]). This implies that the satellite not only induces temporal oscillations, which give rise to phase-mixing and thus phase spirals due to different oscillation frequencies of the stars (see section~\ref{sec:impulsive_kick}), but also spatial corrugations. These vertical and lateral waveforms have wavenumbers given by
\begin{align}
k_z = \frac{n\Omega_z\cos{\thetap}}{\vp},\;\;\;\;\;\;\;{\rm and}\;\;\;\;\;\;\;\;
k_x = \frac{n\Omega_z\sin{\thetap}}{\vp},
\end{align}
respectively. Thus, perpendicular impacts induce only vertical corrugations while planar ones excite waves only laterally. An inclined encounter, on the other hand, spawns corrugations along both directions. Both wavelengths get longer with decreasing mode order, increasing impact velocity, and decreasing vertical frequencies, i.e., increasing actions.
\begin{figure*}
\centering
\begin{subfigure}{0.35\textwidth}
\centering
\includegraphics[width=1\textwidth]{MW_satellites_vp_bp_constraints_before_final.jpeg}
\label{disk_resp_n1_0.5pi}
\end{subfigure}
\begin{subfigure}{0.63\textwidth}
\centering
\includegraphics[width=1\textwidth]{MW_satellites_vp_bp_constraints_final.jpeg}
\label{disk_resp_n1_0.5pi}
\end{subfigure}
\caption{Regions in the space of impact parameter, $\xp\cos{\thetap}$, and velocity, $\vp$, of a satellite galaxy, corresponding to bending (blue) and breathing (red) mode responses in the Solar neighborhood. Response is adiabatically suppressed in the grey region. The circles in the left, middle and right panels indicate the values of $\xp\cos{\thetap}$ and $\vp$ for several MW satellites during their penultimate, last and next disk crossings respectively. The satellites that induce a relative bending mode response, $f_{1,n=1}/f_0\gtrsim 10^{-4}$, for $I_z=h_z\sigma_z$ in the Solar neighborhood, are indicated by red circles, while the others are denoted in grey. All the MW satellites lie outside the breathing region and thus preferentially excite bending modes in the vicinity of the Sun.}
\label{fig:satellite_constraints1}
\end{figure*}
\subsection{Impact of satellite galaxies on the Milky Way disk}
\label{sec::MW_satellites}
The MW halo harbors many satellite galaxies. Some of these are quite massive, with DM halo mass comparable to the disk mass, and either underwent or are about to undergo an encounter with the MW disk within a few hundred Myr from the present day. Hence we expect at least some of them to perturb the disk significantly. Here we use existing data on MW satellites to obtain a rough estimate of the disk response to their encounters with the MW stellar disk.
Our formalism provides physical insight into the trends and scalings of the disk response as a function of impact parameters and velocities of the MW satellites. We emphasize upfront, though, that the precise numerical estimates of the responses are to be taken with a grain of salt. These estimates only serve as a crude, order-of-magnitude attempt to compare the relative disk responses to different satellite galaxies. As discussed in more detail in section~\ref{sec::caveats}, these estimates are subject to a number of oversimplifications and caveats. First of all, the MW disk is modelled as an isothermal slab, and we only consider the {\it direct} impact of the satellites. We ignore indirect effects due to the self-gravity of the response. Our approach also ignores the presence of a dark matter halo, which can impact the disk response in several ways (see section~\ref{sec::caveats}). Because of all these shortcomings, we caution against using the following response estimates for comparison with actual data and/or detailed numerical simulations.
We consider the MW satellites with parallax and proper motion measurements from Gaia DR2 \citep[][]{Gaia_collab_sat.18b} and the corresponding galactocentric coordinates and velocities computed and documented by \cite{Riley.etal.19} \citep[table A.2, see also][]{Li.etal.20} and \cite{Vasiliev.Belokurov.20}. Of these, we only consider the satellites with known dynamical mass estimates \citep[][]{Simon.Geha.07,Bekki.Stanimirovic.09,Lokas.09,Erkal.etal.19}. Adopting the initial conditions for galactocentric positions ($R,z,\phi$) and velocities ($v_R,v_z,v_\phi$) as the median values quoted by \cite{Riley.etal.19} and \cite{Vasiliev.Belokurov.20}, we simulate the orbits of the galaxies in the combined gravitational potential of the MW halo, disk and bulge, which are respectively modelled by a spherical NFW \citep[][]{Navarro.etal.97} profile (virial mass $M_h=9.78\times 10^{11}\Msun$, scale radius $r_h=16$ kpc, and concentration $c=15.3$), a Miyamoto-Nagai \citep[][]{Miyamoto.Nagai.75} profile (mass $M_d=9.5\times 10^{10} \Msun$, scale radius $a=4$ kpc, and scale-height $b=0.3$ kpc), and a spherical \cite{Hernquist.90} profile (mass $M_b=6.5\times 10^9\Msun$ and scale radius $r_b=0.6$ kpc)\footnote{Our MW potential is similar to {\tt GALPY MWPOTENTIAL2014} \citep[][]{Bovy.15} except for the power-law bulge which has been replaced by an equivalent Hernquist bulge.}. The total mass of our fiducial MW model is thus $1.08\times 10^{12}\Msun$. We evolve the positions and velocities of the satellites both forwards and backwards in time from the present day, using a second order leap-frog integrator. For simplicity, we ignore the effect of dynamical friction\footnote{Dynamical friction might play an important role in the orbital evolution of massive satellites like the Large Magellanic Cloud (LMC) and Sgr, pushing their orbital radius farther out in the past.}. From each individual orbit, we note the time, $t_{\rm cross}$, when the satellite crosses the disk (i.e., crosses $z=0$), and record the corresponding distance, $\xp$, from the Sun, which we integrate backwards/forwards in time using a purely circular orbit up to $t_{\rm cross}$. We also record the velocity, $\vp=\sqrt{v^2_R+v^2_z+v^2_\phi}$, and the angle of impact with respect to the disk normal, $\thetap=\cos^{-1}{(v_z/\vp)}$. Finally, we compute the disk response to the satellite encounter using equation~(\ref{f1_sat}). Results are summarized in Table~\ref{tab:MW_sat_resp} and Figs.~\ref{fig:satellite_constraints1} and~\ref{fig:satellite_constraints2}.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{MW_satellites_disk_response_final.jpeg}
\caption{Bending mode strength, $f_{1,n=1}/f_0$ (upper panel), and the corresponding breathing vs bending ratio, $f_{1,n=2}/f_{1,n=1}$ (lower panel), in the Solar neighborhood for the MW satellites, as a function of the disk crossing time, $t_{\rm cross}$, in Gyr, where $t_{\rm cross}=0$ marks today. The previous two and the next impacts are shown. Here we consider $I_z=h_z \sigma_z$, with fiducial MW parameters. In the upper panel, the region with bending mode response, $f_{1,n=1}/f_0<10^{-4}$, has been grey-scaled, indicating that the response from the satellites in this region is far too adiabatic and weak. Note that the response is dominated by that due to Sgr, followed by Hercules, Leo II, Segue 2 and the Large Magellanic Cloud (LMC). Also note that the previous two and next impacts of all the satellites shown here excite bending modes in the Solar neighborhood.}
\label{fig:satellite_constraints2}
\end{figure*}
In Fig.~\ref{fig:satellite_constraints1}, we plot the impact parameter, $\xp\cos{\thetap}$ (with respect to the Sun), as a function of the encounter velocity, $\vp$, of the satellites, for the penultimate (left-hand panel), last (middle panel), and next (right-hand panel) disk crossings. The red (grey) symbols denote the satellites that induce a strong (weak) amplitude of the bending mode response, $f_{1,n=1}/f_0$, for $I_z=h_z\sigma_z = 9.2 \kpc\kms$. As shown in Appendix~\ref{App:detect_crit}, we consider $f_{1,n=1}/f_0=\delta=10^{-4}$ as a rough estimate for the minimum detectable relative response, i.e., the boundary between strong and weak responses to satellite passage. The solid black line indicates the boundary between bending and breathing modes, i.e., where the breathing-to-bending ratio, $f_{21}$ (equation~[\ref{f21}]), is equal to unity. Hence, the blue and red shaded regions indicate where the response is dominated by bending and breathing modes, respectively. The magenta, dashed line roughly denotes the boundary between a strong bending response (blue shaded region) and a response that is adiabatically suppressed (grey shaded region). The latter is defined by the condition $\exp{\left[-\Omega_z \xp \cos{\thetap}/\vp\right]} < \delta = 10^{-4}$.
In Fig.~\ref{fig:satellite_constraints2}, we plot the amplitude of the bending mode response, $f_{1,n=1}/f_0$ (upper panel), and the breathing-to-bending ratio, $f_{21}=f_{1,n=2}/f_{1,n=1}$ (lower panel), in the Solar neighborhood, as a function of the time $t_{\rm cross}$ (in Gyr) when the satellite crosses the plane of the disk, assuming the fiducial MW parameters. Negative and positive $t_{\rm cross}$ correspond to disk crossings in the past and future, respectively, and we once again consider stars with $I_z = h_z \sigma_z = 9.2 \kpc\kms$.
Both Fig.~\ref{fig:satellite_constraints1} and the lower panel of Fig.~\ref{fig:satellite_constraints2} make it clear that {\it all} the disk crossings considered here preferentially excite bending rather than breathing modes in the Solar neighborhood. As shown in Section~\ref{sec:impulsive_kick} these trigger one-armed phase spirals in the Solar neighborhood, in qualitative agreement with the MW snail observed in the Gaia data. However, as is evident from the upper panel of Fig.~\ref{fig:satellite_constraints2}, most satellites only trigger a minuscule response in the disk, with $f_{1,n=1}/f_0 < \delta = 10^{-4}$, either because the satellite has too low mass, or because the encounter, from the perspective of the Sun, is too slow such that the local response is adiabatically suppressed. The strongest response by far is triggered by encounters with Sgr, for which the bending mode response, $f_{1,n=1}/f_0$, is at least $1-2$ orders of magnitude larger than that for any other satellite. Based on our orbit-integration, it had its penultimate disk crossing, which closely coincides with its last pericentric passage, about $900\Myr$ ago, triggering a strong response of $f_{1,n=1}/f_0 \sim 0.04$ in the Solar neighborhood. The last disk crossing, which nearly corresponds to the last apocentric passage, occurred about $300\Myr$ ago, triggering a very weak (adiabatically suppressed) response. Sgr is currently near its pericenter and will undergo the next disk crossing in about $30\Myr$, which we estimate to only trigger a moderately strong response with $f_{1,n=1}/f_0\sim 0.001$. We caution, though, that in addition to the caveats listed above and in Section~\ref{sec::caveats} these estimates ignore dynamical friction and are sensitive to the MW potential and the current phase-space coordinates of the satellites. We have checked that a heavier MW model with a total mass of $1.5\times 10^{12}\Msun$ does not change the relative amplitudes of the satellite responses significantly, but brings most of the disk crossing times closer to the present day since the satellites are more bound in a heavier MW. For example, the previous pericentric and apocentric passages of Sgr occur at $\sim 600$ and $200\Myr$ ago in the heavier case. The only satellite apart from Sgr that triggers a response $f_{1,n=1}/f_0 > \delta = 10^{-4}$ is Hercules, whose disk crossing $\sim 500\Myr$ ago caused a bending-mode response, $f_{1,n=1}/f_0 = 1.2\times 10^{-4}$. Segue 2 induces a response that is marginally below the detection threshold. Disk crossings of LMC and Leo II trigger responses that are comparable in strength to that of Hercules, but the crossing times are too far in the past or future for them to be considered as candidates for triggering the Gaia snail. All in all, it is clear then that Sgr is by far the most likely candidate among the MW satellite galaxies considered here to have triggered the one-armed phase spiral in the Solar neighborhood discovered in Gaia DR2 by \citet{Antoja.etal.18}.
\begin{table*}
\centering
\hspace{-3cm}
\tabcolsep=0.2 cm
\begin{tabular}{c|c|cc|cc|cc}
\hline
MW satellite & Mass & $f_{1,n=1}/f_0$ & $t_{\rm cross}$ & $f_{1,n=1}/f_0$ & $t_{\rm cross}$ & $f_{1,n=1}/f_0$ & $t_{\rm cross}$ \\
name & $(\Msun)$ & & $(\Gyr)$ & & $(\Gyr)$ & & $(\Gyr)$ \\
& & Penultimate & Penultimate & Last & Last & Next & Next \\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) \\
\hline
\highlight{Sagittarius} & $10^9$ & \highlight{$4.3\times 10^{-2}$} & \highlight{$-0.92$} & $1.4\times 10^{-10}$ & $-0.3$ & $8.3\times 10^{-4}$ & $0.03$ \\
Hercules & $7.1\times 10^6$ & -- & $-3.57$ & $1.2\times 10^{-4}$ & $-0.51$ & $6.4\times 10^{-5}$ & $3.16$ \\
Leo II & $8.2\times 10^6$ & -- & $-3.61$ & $3.5\times 10^{-5}$ & $-1.81$ & $9.3\times 10^{-5}$ & $2.34$ \\
Segue 2 & $5.5\times 10^5$ & $5\times 10^{-5}$ & $-0.84$ & $3.4\times 10^{-5}$ & $-0.25$ & $1.8\times 10^{-6}$ & $0.27$ \\
LMC & $1.4\times 10^{11}$ & $1.4\times 10^{-4}$ & $-6.97$ & -- & $-2.37$ & $7.2\times 10^{-5}$ & $0.12$ \\
SMC & $6.5\times 10^9$ & $3.6\times 10^{-8}$ & $-3.22$ & -- & $-1.39$ & $1.2\times 10^{-9}$ & $0.22$ \\
Draco I & $2.2\times 10^7$ & -- & $-2.43$ & $5\times 10^{-7}$ & $-1.23$ & $1\times 10^{-7}$ & $0.24$ \\
Bootes I & $10^7$ & -- & $-1.65$ & $4.1\times 10^{-7}$ & $-0.35$ & -- & $0.87$ \\
Willman I & $4\times 10^5$ & -- & $-0.63$ & $1.4\times 10^{-7}$ & $-0.21$ & $2.5\times 10^{-8}$ & $0.4$ \\
Ursa Minor & $2\times 10^7$ & -- & $-2.26$ & $5.5\times 10^{-8}$ & $-1.16$ & $8.6\times 10^{-9}$ & $0.29$ \\
Ursa Major II & $4.9\times 10^6$ & $4.5\times 10^{-8}$ & $-2$ & $6.2\times 10^{-10}$ & $-0.1$ & -- & $0.9$ \\
Coma Berenices I & $1.2\times 10^6$ & $7\times 10^{-10}$ & $-2.47$ & -- & $-0.25$ & -- & $0.69$ \\
Sculptor & $3.1\times 10^7$ & -- & $-2.7$ & $2\times 10^{-10}$ & $-0.46$ & -- & $1.47$\\
\hline
\end{tabular}
\caption{MW disk response to satellites for stars with $I_z=h_z\sigma_z$ in the Solar neighborhood. Column (1) indicates the name of the MW satellite and Column (2) indicates its dynamical mass estimate from literature \citep[][]{Simon.Geha.07,Bekki.Stanimirovic.09,Lokas.09,Erkal.etal.19,Vasiliev.Belokurov.20}. We assume $10^9\Msun$ for the Sagittarius mass; note that there is a discrepancy between its measured mass of $\sim 4\times 10^8\Msun$ \citep[][]{Vasiliev.Belokurov.20} and the required mass of $10^9-10^{10}\Msun$ for observable phase spiral signatures in N-body simulations \citep[see for example][]{Bennett.etal.21}. Columns (3) and (4) respectively indicate the bending mode response assuming fiducial MW parameters and the crossing time for the penultimate disk-crossing. Columns (5) and (6) show the same for the last disk-crossing, while columns (7) and (8) indicate it for the next one. Only the satellites that trigger a bending mode response, $f_{1,n=1}/f_0\geq 10^{-10}$, in at least one of the three cases are shown. The responses smaller than $10^{-10}$ are considered far too adiabatic and negligible and are marked by dashes. The case most relevant for the Gaia phase spiral is highlighted in red.}
\label{tab:MW_sat_resp}
\end{table*}
We emphasize that the results shown in Figs.~\ref{fig:satellite_constraints1} and~\ref{fig:satellite_constraints2} correspond to stars with a vertical action $I_z=h_z\sigma_z=9.2\, \kpc\kms$. As mentioned above, the strength of the response depends on the ratio of the encounter time scale, $\tau_{\rm enc}$ (equation~[\ref{tau_enc}]) and the vertical oscillation period of stars in the Solar neighborhood, $\tau_z$ (equation~[\ref{tau_z}]). The latter is longer for stars with larger vertical action, and from the perspective of such stars the encounter is more impulsive, resulting in a stronger response. Since the response does not scale linearly with $\tau_{\rm enc}/\tau_z$, the relative response strength of different satellites depends somewhat on the vertical action. We have verified that for $I_z/(h_z\sigma_z) < 3$, which is roughly the range covered by the Gaia phase spiral, the direct response from the encounter with Sgr remains larger than that of any other satellite considered here by at least $1-2$ orders of magnitude. However, for stars with larger actions (larger vertical excursions), the LMC can dominate the response. In particular, for stars with $I_z/(h_z\sigma_z) \gtrsim 6.5$ ($z_{\rm max}\gtrsim 4\, h_z$), which make up the thick disk, the LMC is expected to trigger a stronger response than Sgr during its upcoming disk crossing.
To summarize, our analysis suggests that the MW satellites during their most recent and forthcoming disk crossings preferentially excite bending modes in the Solar neighborhood. This is because satellite encounters are fairly distant from the Sun and thus the encounter time exceeds the vertical oscillation time of the stars. However, as previously discussed in section~\ref{sec:sat_encounter} and as evident from the N-body simulation of MW-Sgr encounter by \cite{Hunt.etal.21} (especially the earlier disk passages of Sgr), a satellite passage can trigger breathing modes closer to the point of impact, where the encounter is more impulsive. Since almost all the MW satellites undergo their disk-crossings at $R\gg 8\kpc$, future observations of the outskirts of the disk might reveal breathing instead of bending mode oscillations if they are excited by any of the satellites considered here.
\subsection{Caveats}
\label{sec::caveats}
The above calculation of the response of the MW disk to perturbations is subject to a number of oversimplifications and caveats discussed below.
The MW disk is modelled as an isothermal slab, which lacks the axisymmetric density profile and velocity structure that characterize a realistic disk. In particular, whereas the lateral motion in our slab is uninhibited, the in-plane motion in a realistic disk consists of an azimuthal rotation combined with a radial epicyclic motion. Among others, this will have important implications for the global disk response and the rate at which phase spirals damp out due to lateral mixing. In Paper~II (Banik et al., in preparation) we apply our perturbative formalism to a realistic self-gravitating disk galaxy with a pseudo-isothermal distribution function \citep[][]{Binney.10}, and consider both external perturbations (encounters with satellites) and internal perturbations (bars and spiral arms).
All responses calculated in this paper only account for the direct response to a perturbing potential. In general, though, the response also has an indirect component that arises from the fact that neighboring regions in the disk interact with each other gravitationally. This self-gravity of the response, which we have ignored, triggers long-lived normal mode oscillations of the slab that are not accounted for in our treatment. Several simulation-based studies have argued that including self-gravity is important for a realistic treatment of phase spirals \citep[e.g.,][]{Darling.Widrow.19a, Bennett.Bovy.21}. Using the Kalnajs matrix method \citep[][]{Kalnajs.77, Binney.Tremaine.08}, we have made some initial attempts to include the self-gravity of the response in our perturbative analysis, along the lines of \citet{Weinberg.91}. Our preliminary analysis shows that the self-gravitating response is a linear superposition of two terms: (i) a continuum of modes given in equation~(\ref{f1nk_slabsol}), dressed by self-gravity, that undergo phase-mixing and give rise to the phase spiral, and (ii) a discrete set of modes called point modes or normal modes \citep[c.f.][]{Mathur.90,Weinberg.91} that follow a dispersion relation. The continuum response can be amplified by self-gravity when the continuum frequencies, $n\Omega_z+k v_x$, are close to the point mode frequencies, $\nu$. Depending on the value of $k$, the normal modes can be either stable or unstable. \cite{Araki.85} find that in an isothermal slab the bending normal mode undergoes fire hose instability below a certain critical wavelength if $\sigma_z/\sigma \lesssim 0.3$ while the breathing normal mode becomes unstable above the Jeans scale. In the regime of stability, the normal modes are undamped oscillation modes in absence of lateral streaming \citep[][]{Mathur.90} but are Landau damped otherwise \citep[][]{Weinberg.91}. For an isothermal slab with typical MW-like parameter values, the point modes are strongly damped since their damping timescale (inverse of the imaginary part of $\nu$) is of order their oscillation period (inverse of the real part of $\nu$), which turns out to be of order the vertical dynamical time, $h_z/\sigma_z$. Moreover, the normal mode oscillations are coherent oscillations of the entire system, independent of the vertical actions of the stars, and are decoupled from the phase spiral in linear theory since the full response is a linear superposition of the two. Based on the above arguments, we conclude that self-gravity has little impact on the evolution of phase spirals in the isothermal slab, at least in the linear regime. We emphasize that \citet{Darling.Widrow.19a}, who found their phase spirals to be significantly affected by the inclusion of self-gravity, assumed a perturber-induced velocity impulse with magnitude comparable to the local velocity dispersion in the Solar neighborhood; hence their results are likely to have been impacted by non-linear effects. Moreover, the self-gravitating response of an inhomogeneous disk embedded in a dark matter halo, as in the simulations of \citet{Darling.Widrow.19a}, can be substantially different from that of the isothermal slab. We intend to include a formal treatment of self-gravity along the lines of \citet{Weinberg.91} in future work.
The disk of our MW is believed to be embedded in an extensive dark matter halo, something we have not taken into account. The presence of such a halo has several effects. First of all, the satellite not only perturbs the disk, but also the halo. In particular, it induces both a local wake and a global modal response\footnote{The torque from the local as well as global halo response is responsible for dynamical friction acting on the satellite.} \citep[e.g.,][]{Weinberg.89, Tamfal.etal.21}. The former typically trails the satellite galaxy, and boosts its effective mass by about a factor of two \cite[][]{Binney.Tremaine.08}, which might boost the (direct) disk response by about the same factor. The global halo response is typically dominated by a strong $l=1$ dipolar mode followed by an $l=2$ quadrupolar mode \citep[][]{Tamfal.etal.21}, which might have a significant impact on the disk. The presence of a halo also modifies the total potential. At large disk radii and vertical heights, the halo dominates the potential and will therefore significantly modify the actions and frequencies of the stars, and consequently the shape of the phase spirals. Finally, since the disk experiences the gravitational force of the halo, a (sufficiently massive) satellite galaxy can excite normal mode oscillations of the disk in the halo \citep[see for example][]{Hunt.etal.21}. We intend to incorporate some of these effects of the MW halo in Paper~II.
\section{Conclusion}
\label{sec:concl}
In this paper we have used linear perturbation theory to compute the response of an infinite, isothermal slab to various kinds of external perturbations with diverse spatio-temporal characteristics. Although a poor description of a realistic disk galaxy, the infinite, isothermal slab model captures the essential physics of perturbative response and collisionless equilibration via phase-mixing in the disk, and thus serves as a simple yet insightful case for investigation.
We use a hybrid (action-angle variables in the vertical direction and position-momentum variables in the lateral direction) linear perturbative formalism to perturb and linearize the collisionless Boltzmann equation and compute the response in the distribution function of the disk to a gravitational perturbation. We have considered external perturbations of increasing complexity, ranging from an instantaneous (laterally) plane-wave perturbation (Section~\ref{sec:impulsive_kick}), an instantaneous localized perturbation, represented as a wave-packet (Section~\ref{sec:localized}), a non-impulsive, temporally extended, localized perturbation (Section~\ref{sec:non-impulsive}), and ultimately an encounter with a satellite galaxy moving along a straight-line orbit (Section~\ref{sec:sat_encounter}). This multi-tiered approach is ideal for developing the necessary insight into the complicated response that is expected from a realistic disk galaxy exposed to a realistic perturbation. We summarize our conclusions below.
\begin{itemize}
\item The two primary Fourier modes of slab oscillation are the $n=1$ bending mode and the $n=2$ breathing mode, which correspond to anti-symmetric and symmetric oscillations about the mid-plane, respectively. For a sufficiently impulsive perturbation, the dominant mode is the breathing mode, which initially causes a quadrupolar distortion in the $(z,v_z)$-phase space, that evolves into a two-armed phase spiral as the stars with different vertical actions oscillate with different vertical frequencies. If the perturbation is temporally more extended (less impulsive), the dominant mode is the bending mode. This causes a dipolar distortion in $(z,v_z)$-phase space that evolves into a one-armed phase spiral \citep[see also][]{Hunt.etal.21, Widrow.etal.14}. Due to vertical phase-mixing, the phase spiral wraps up tighter and tighter until it becomes indistinguishable from an equilibrium distribution in the coarse-grained sense.
\item Besides vertical phase-mixing the survivability of the phase spiral is also dictated by the lateral streaming motion of stars. The initial lateral velocity impulse towards the minima of $\Phi_\rmP$ tends to linearly boost the contrast of the phase spiral. This is however quickly taken over by lateral streaming (with velocity dispersion $\sigma$), which causes mixing between the over- and under-densities, and damps out the phase spiral amplitude. For an impulsive, laterally sinusoidal perturbation, the disk response is also sinusoidal and damps out like a Gaussian (due to the Maxwellian/Gaussian distribution of the unconstrained lateral velocities) over a timescale of $\tau_\rmD \sim 1/k\sigma$, i.e., small scale perturbations damp out faster, as expected.
\item Lateral mixing operates differently for a spatially localized perturbation which can be expressed as a superposition of many plane waves. The response to each of them damps out like a Gaussian (if the perturber is impulsive). Since the power spectrum of a spatially localized perturber with a lateral Gaussian profile is dominated by its largest scales (small $k$) that mix and damp out slower, the net response from all $k$ damps away as $\sim t^{-1}$ (the response profile spreads out as $\sim t$), much slower than the Gaussian damping in case of a sinusoidal perturber.
\item The disk response to a non-impulsive perturbation is substantially different from that to an impulsive one. If the temporal strength of the perturber follows a Gaussian pulse with pulse frequency, $\omega_0$ (e.g., a transient bar or spiral arm), the response grows and decays following the temporal profile of the pulse before eventually attaining a $\sim 1/t$ power law fall-off. The response peaks when the pulse frequency, $\omega_0$, is comparable to the vertical oscillation frequency, $\Omega_z$. The response to more impulsive perturbations ($\omega_0 \gg \Omega_z$) is suppressed as $\sim 1/\omega_0$, whereas much slower ($\omega_0 \ll \Omega_z$) perturbations trigger a super-exponentially ($\sim \exp{\left[-n^2\Omega^2_z/4\omega^2_0\right]}$ at small $k$) suppressed response. In this adiabatic limit, the stars tend to remain in phase with the perturber, oscillating at frequencies much smaller than $\Omega_z$, which inhibits the formation of a phase spiral.
\item The timescale of perturbation dictates the excitability of different modes, with slower (faster) pulses triggering stronger bending (breathing) modes. An encounter with a satellite galaxy that hits the disk with a uniform velocity $\vp$ and an angle $\thetap$ with respect to the normal at a distance $\xp$ away from an observer in the disk, perturbs the potential at an observer's location with a characteristic time scale $\tau_{\rm enc} \sim \xp\cos{\thetap}/\vp$. If $\tau_{\rm enc}$ is long (short) compared to the typical vertical oscillation time, $\tau_z \sim h_z / \sigma_z$, at the observer's location, the dominant perturbation mode experienced is a bending (breathing) mode. Thus, bending modes are preferentially excited not only by low velocity encounters, but also by more distant and more perpendicular ones. Since the velocities of all MW satellites are much larger than $\sigma_z$, the decisive factor for bending {\it vs.} breathing modes is the distance from the point of impact. This is in qualitative agreement with the results from $N$-body simulations of the MW-Sgr encounter performed by \cite{Hunt.etal.21}, which show more pronounced bending (breathing) modes further from (closer to) the location where Sgr impacts the disk. Moreover, for a given encounter, stars with larger actions undergo stronger breathing mode oscillations since they oscillate slower.
\item Besides phase spirals satellite encounters also induce spatial corrugations in the disk response, with vertical and lateral wave-numbers given by $k_z=n\Omega_z\cos{\thetap}/\vp$ and $k_x=n\Omega_z\sin{\thetap}/\vp$, respectively.
\end{itemize}
As an astrophysical application of our formalism, we have investigated the direct response of the MW disk (approximated as an isothermal slab) to several of the satellite galaxies in the halo for which dynamical mass estimates and galactocentric phase-space coordinates from Gaia parallax and proper motion measurements are available. We integrate the orbits of these satellites in the MW potential and note the impact velocity, $\vp$, angle of impact, $\thetap$, with respect to the normal, and the impact distance from the Solar neighborhood, $\xp$, during their penultimate, last and next disk crossings. We use these parameters to compute the direct response to the MW satellites and find that all of them excite bending modes and thus one-armed phase spirals in the Solar neighborhood, similar to that discovered in the Gaia data by \citet{Antoja.etal.18}. In the Solar vicinity, the largest direct response, by far, is due to the encounter with Sgr. The direct responses triggered by other satellites, most notably Hercules and the LMC, are at least $1-2$ orders of magnitude smaller. Hence, we conclude that, if the Gaia phase spiral was triggered by an encounter with a MW satellite, the strongest contender is Sgr. Although Sgr has been considered as the agent responsible for the Gaia phase spiral and other local asymmetries and corrugations, several studies have pointed out that it cannot be the sole cause of all these perturbations \citep[see e.g.,][]{Bennett.etal.21, Bennett.Bovy.21}. Our work argues, though, that the direct response in the Solar neighborhood from the other MW satellites, including the LMC, is not significant enough, at least in the range of actions covered by the Gaia snail. Of course, as discussed in section~\ref{sec::caveats}, the indirect response from the DM halo of the MW might play an important role especially for the more massive satellites such as Sgr and the LMC. Moreover the global response of a realistic disk will be different from that of the isothermal slab model considered here. We investigate the realistic disk response in Paper~II and leave a sophisticated analysis incorporating self-gravity and halo response for future work. It remains to be seen whether a combination of Sgr plus other (internal) perturbations due to for example spiral arms \citep[][]{Faure.etal.14} or the (buckling) bar \citep[e.g.,][]{Khoperskov.etal.19} can explain the fine-structure in the Solar neighborhood, or whether perhaps a solution requires modifying the detailed MW potential. It is imperative, though, to investigate the structure of phase spirals at other locations in the MW disk, in particular whether they are one-armed or two-armed. This would help to constrain both the time-scale and location of the perturbation responsible for the various out-of-equilibrium features uncovered in the disk of our MW.
\section*{Acknowledgments}
The authors are grateful to the anonymous referee for thoughtful comments and to Kathryn Johnston, Jason Hunt, Adrian Price-Whelan, Kaustav Mitra, Elena D'Onghia, Chris Hamilton and Dhruba Dutta-Chowdhury for insightful discussions and valuable suggestions. MW is supported by the National Science Foundation through Grant No. AST-1812689. FvdB is supported by the National Aeronautics and Space Administration through Grant No. 19-ATP19-0059 issued as part of the Astrophysics Theory Program.
|
\section{Introduction}
This article is an extension of the original work of \cite{fisher},
\cite{hotel} and \cite{winter} (the last one correctly discounting the importance of `variance stabilization'), with the aim of explicitly deriving \emph{all} terms
necessary to achieve an $O(n^{-3/2})$ accuracy of the desired approximation. The
resulting formula then becomes substantially more accurate than those found in the
existing literature.
One of the most natural ways to derive the Fisher transformation of the
empirical correlation coefficient $r$ (assuming a sample from a bivariate
Normal distribution) is to start by finding the first few terms of the Edgeworth
expansion \cite{edge} of the sampling distribution of an arbitrary function
of $r$ say $G(r)$. This is done by standardizing $G(r)$ by an $O(n^{-2})$%
-accurate (in terms of its error) expected value $m$ and $O(n^{-3})$%
-accurate variance $V$, and finding its $O(n^{-3/2})$-accurate skewness $%
\Gamma_{3}$ and $O(n^{-2})$-accurate \emph{excess} (subtracting $3$)
kurtosis $\Gamma_{4}$ (where $n$ is the sample size). The approximate
probability density function (PDF) of%
\begin{equation}
Z\coloneqq\frac{G(r)-m}{\sqrt{V}} \label{stand}
\end{equation}
is then given by
\begin{equation}
f_{Z}(z) \coloneqq
\frac{\exp \left( -\dfrac{z^{2}}{2}\right)}{\sqrt{2\pi }}
\times \left(1+\frac{\Gamma_{3}(z^{3}-3z)}{6}+\frac{\Gamma_{4}(z^{4}-6z^{2}+3)}{24}+\frac{\Gamma_{3}^{2}(z^{6}-15z^{4}+45z^{2}-15)}{72}\right) \label{edge}
\end{equation}%
and has an $O(n^{-3/2})$-proportionate error (compared to an $O(n^{-1/2})$
error of the basic Normal approximation).
It is then possible to find $G$ (based on the resulting differential
equation) to make $\Gamma_{3}$ equal to zero (to the $O(n^{-3/2}$) level of
accuracy) for any value of the `true' correlation coefficient $\rho$. This
yields the expected `$\arctanh$' transformation, but also suggests a subtle
correction to it, making the resulting approximation substantially more
accurate.
\section{Key concepts and formulas}
In this section we assume sampling from \emph{any} specific multivariate
distribution; to simplify the notation, our definitions and examples will
use a tri-variate case and call the random variables $X_{1}$, $X_{2}$ and $%
X_{3}$, (generalizing is easy). The corresponding \emph{central} \textsc{%
moments} are then defined by%
\begin{equation}
\mu_{i,j,k} \coloneqq \mathbb{E}\left[ (X_{1}-\mu_{1})^{i}(X_{2}-\mu_{2})^{j}(X_{3}-\mu_{3})^{k}\right] \label{moment}
\end{equation}%
where $i+j+k$ is the moment's \textsc{order}. They can be conveniently
computed based on the moment generating function (MGF) of the corresponding
\emph{centralized} random variables, namely from%
\begin{align}
M(t_{1},t_{2},t_{3})
&\coloneqq \mathbb{E}\left[ \exp (t_{1}(X_{1}-\mu_{1})+t_{2}(X_{2}-\mu_{2})+t_{3}(X_{3}-\mu_{3}))\right]\label{MGF} \\
&\equiv 1+\sum_{i+j+k\geq 2}^{\infty }\frac{\mu_{i,j,k}}{i!\, j!\, k!} t_{1}^{i}t_{2}^{j}t_{3}^{k} \notag
\end{align}
by differentiating it with respect to $t_{1}$, $t_{2}$ and $t_{3}$
correspondingly $i$, $j$ and $k$ times, and then setting each $t_{\ell }$
equal to zero. If an explicit formula for such an MGF cannot be found (the
integration may have no analytic answer), it is sufficient to replace it by
the last line of (\ref{MGF}), with the summation truncated to exclude terms
beyond the fourth order (computationally more feasible).
It is well known (and easy to prove) that, when replacing the individual
random variables by their respective \emph{sample} means, the MGF of the new
set, namely of $\overline{X_{1}}-\mu_{1}$, $\overline{X_{2}}-\mu_{2}$,
$\overline{X_{3}}-\mu_{3}$, is given by
\begin{equation}
M\left( \frac{t_{1}}{n},\frac{t_{2}}{n},\frac{t_{3}}{n}\right)^{n}
\label{MGFM}
\end{equation}%
This implies that finding%
\begin{equation}
\bar{\mu}_{i,j,k}\coloneqq\mathbb{E}\left[ (\overline{X_{1}}-\mu_{1})^{i}(\overline{X_{2}}-\mu_{2})^{j}(\overline{X_{3}}-\mu_{3})^{k}\right]
\label{momM}
\end{equation}%
can be achieved by the same $i$, $j$, $k$-fold differentiation of (\ref{MGFM}%
) and subsequent $t_{\ell }=0$ substitution as before. Thus, for example, we
get%
\begin{equation}
\bar{\mu}_{2,1,1}=\frac{\mu_{2,0,0}\mu_{0,1,1}+2\mu_{1,1,0}\mu_{1,0,1}}{n^{2}}+\frac{\mu_{2,1,1}-\mu_{2,0,0}\mu_{0,1,1}-2\mu_{1,1,0}\mu_{1,0,1}}{n^{3}}
\end{equation}%
etc. Not surprisingly, the complexity of these formulas increases
`exponentially' with the moment's order.
To find an $O(n^{-3/2})$-accurate approximation to the PDF of any function
of sample means, say
$H\left(\overline{X_{1}},\overline{X_{2}},\overline{X_{3}}\right)$,
we need to re-write this function as%
\begin{equation}
H\left(\mu_{1}+\varepsilon (\overline{X_{1}}-\mu_{1}),\mu_{2}+\varepsilon (\overline{X_{2}}-\mu_{2}),\mu_{3}+\varepsilon (\overline{X_{3}}-\mu_{3})\right) \label{eps}
\end{equation}%
and expand it in $\varepsilon $, up to and including $\varepsilon^{3}$
terms (with the understanding that $\varepsilon $ will be set to $1$
eventually); note that the first (constant) term of this expansion is
$H\left( \mu_{1},\mu_{2},\mu_{3}\right) $.
We then compute the expected value of the first \emph{four} powers of the
result, after subtracting $H\left( \mu_{1},\mu_{2},\mu_{3}\right) $ from
it (to simplify the corresponding algebra). This requires further expanding
of these powers in $\varepsilon $, up to and including $\varepsilon^{2}$,
$\varepsilon^{4}$, $\varepsilon^{4}$ and $\varepsilon^{6}$ terms
respectively, before applying (\ref{momM}) to the individual terms.
The resulting four moments of
$H\left( \overline{X_{1}},\overline{X_{2}},\overline{X_{3}}\right) -H\left( \mu_{1},\mu_{2},\mu_{3}\right)$
are then easily converted to the corresponding $m$, $V$, $\Gamma_{3}$ and
$\Gamma_{4}$; these should be further simplified by keeping only the leading
terms of their $\frac{1}{n}$ expansion (with the exception of $V$, where
both $\frac{1}{n}$ and $\frac{1}{n^{2}}$-proportional terms are needed).
This ensures that all terms contributing to the final $O(n^{-3/2})$ accuracy
of the final answer are included, while the rest of them (often incorrect, since
incomplete) have been eliminated.
\section{Examples}
Pearson's correlation coefficient $r$ is defined, using the $\varepsilon$
notation of the previous section, as
\begin{equation}
r
\coloneqq
\frac
{\rho +\varepsilon (\overline{XY}-\rho )-\varepsilon^{2}\overline{X}\cdot \overline{Y}}
{\sqrt{\left( 1+\varepsilon (\overline{X^{2}}-1)-\varepsilon^{2}\overline{X}^{2}\right) \left( 1+\varepsilon (\overline{Y^{2}}-1)-\varepsilon^{2}\overline{Y}^{2}\right) }} \label{r}
\end{equation}%
Note that $r$ is a function of \emph{five} different sample means. To
investigate its sampling distribution, we assume a random independent sample
of size $n$ from a bivariate Normal distribution of $X$ and $Y$, with both
means equal to $0$ and both variances equal to $1$, while the true (or
`population') correlation coefficient is $\rho $ (the $r$ distribution is
the same whatever means and variances we use; we have thus made the simplest
choice).
The MGF of the required five centralized variables, namely $X$, $Y$, $X^{2}-1$,
$Y^{2}-1$ and $XY-\rho$, is the result of the following (rather
routine) double integration
\begin{equation}
\frac{\displaystyle\int\limits_{-\infty }^{\infty }\exp \left( -\dfrac{x^{2}+y^{2}-2\rho xy}{2(1-\rho^{2})}+t_{1}x+t_{2}y+t_{3}(x^{2}-1)+t_{4}(y^{2}-1)+t_{5}(xy-\rho )\right) }{2\pi \sqrt{1-\rho^{2}}}
\end{equation}%
quoted in the Mathematica code of Figure 1 (which calls it $M$). This
expression is then easily converted into the MGF of $\overline{X}$,
$\overline{Y}$, $\overline{X^{2}}-1$, $\overline{Y^{2}}-1$ and
$\overline{XY}-\rho$, and the first four moments of $r-\rho $ (no further transformation
is applied to $r$ in this example) are found; the corresponding $m$, $V$,
$\Gamma_{3}$ and $\Gamma_{4}$, properly truncated in their $\frac{1}{n}$
expansions, then easily follow. The complete program and its output are
displayed in Figure \ref{fig::code}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.8\textwidth]{CODE.jpg}
\caption{Mathematica code}
\label{fig::code}
\end{center}
\end{figure}
The resulting approximation is than given by (\ref{stand}) and (\ref{edge});
the PDF of $Z$ can be easily converted to an approximate PDF of $r$ by
\begin{equation}
f\left( r\right) \simeq \frac{f_{Z}\left( \dfrac{r-m}{\sqrt{V}}\right) }{%
\sqrt{V}}
\end{equation}%
Using $n=35$ and $\rho =-0.85$, we show (in Figure \ref{fig1} - the exact PDF is red,
the approximate one is blue) how the approximation compares to the exact
answer (which, rather atypically for a sample statistic of this complexity,
has an analytic form - see \cite{hotel}). The maximum error of this approximation, realized when computing $\Pr(-0.9685<r<-0.9133)$, is less than $0.8\%$.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.8\textwidth]{fig1.jpg}
\caption{Exact (red) and Edgeworth-series (blue) PDF of $r$}
\label{fig1}
\end{center}
\end{figure}
\subsection{Fisher transformation}
Re-running the same program with the `$G[x\_]\coloneqq x$' line removed yields, for $%
\Gamma_{3}$, the following expression%
\begin{equation}
\frac{3G^{\prime }[\rho ]\left( (1-\rho^{2})G^{\prime \prime }[\rho ]-2\rho
~G^{\prime }[\rho ]\right) }{\sqrt{n}}
\end{equation}%
To make it zero (for any value of $\rho $), $G$ needs to be either a
constant (which is clearly inadmissible) or a solution to $(1-\rho
^{2})G^{\prime \prime }(\rho )-2\rho ~G^{\prime }(\rho )=0$, namely
$G(\rho)=\arctanh(\rho )$, having chosen its simplest form (all other
possibilities would yield the same $Z$).
Running the same program one more time with `$G[x\_]\coloneqq$~ArcTanh$[x]$' then
results in
\begin{align*}
m &= \arctanh(\rho )+\frac{\rho }{2n} \\
V &= \frac{1}{n}+\frac{6-\rho^{2}}{2n^{2}} \\
\Gamma_{3} &= 0\ \ \ \ \ \ \text{(by design)} \\
\Gamma_{4} &= \frac{2}{n}
\end{align*}%
This time%
\begin{equation}
f\left( r\right) \simeq \frac{f_{Z}\left( \dfrac{\arctanh(r)-m}{\sqrt{V}}\right) }{(1-r^{2})\sqrt{V}}
\end{equation}%
which, when using the previous choice of $n$ and $\rho $, yields a PDF
visually indistinguishable from the exact answer; its maximum error (now
involving a large interval of values - thus not the fairest of comparisons)
is $0.36\%$. This accuracy is maintained (actually, rather fortuitously
reduced to $0.05\%$) after dropping the $\Gamma_{4}$ term, whose
contribution is (for this transformation of $r$) practically negligible, due
to its relatively small size.
Note that the `basic' Fisher transformation similarly ignores $\Gamma_{4}$.
but it also drops the $\frac{1}{n}$-proportional correction to $m$ and uses $%
V=\dfrac{1}{n-3}$ in place of our result. This affects, quite adversely, its
accuracy, as seen in Figure \ref{fig::fig2}.%
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.8\textwidth]{fig2}
\caption{Exact (red) and basic-Fisher (blue) PDF of $r$}
\label{fig::fig2}
\end{center}
\end{figure}
The largest error of this approximation is over $3.5\%$ - clearly
inacceptable!
These two examples should suffice to illustrate using the Edgeworth series
in situations going well beyond its original formulation and intended
purpose.
\section{Conclusion}
We have delineated a procedure for constructing an accurate approximation
for a PDF of any function of several sample means, when sampling a specific
univariate or multivariate distribution. It is based on finding, to a
specific accuracy (in terms of their $\frac{1}{n}$ expansions), the
corresponding mean, variance, skewness and excess kurtosis; in the case of a
single-parameter distribution, it is usually possible to find a
transformation of the sample statistics which eliminates skewness, thus
making the approximation both simpler and more accurate. We have tested the
technique against the `classical' example of Fisher transformation
(suggesting a minor modification leading to a significant improvement), but
its main applicability is to situations with no exact solution (constructing
an \emph{approximate} PDF is then the best we can do). We should mention that we
have not attempted to optimize the algorithm (this would require introducing
cumulants), but since our program takes only a few seconds to execute, this
would not appear necessary in most cases of interest.
|
\section{Introduction}
Geolocation and mobility data collected by location-based services (LBS)~\cite{huang2018location}, can reveal human mobility patterns and address various societal research questions~\cite{kolodziej2017local}. For example, Call Data Records (CDR) have been successfully used to provide real-time traffic anomaly and event detection~\cite{toch2019analyzing, wang2020deep}, and a variety of mobility datasets have been used in shaping policies for urban communities~\cite{ferreira_deep_2020} and epidemic management in the public health domain~\cite{oliver2015mobile,oliver2020mobile}. Moreover, users can benefit from personalized recommendations when they are encouraged to share their location data with third parties or other service providers (SPs, i.e., social platforms)~\cite{erdemir_privacy-aware_2020}. Human mobility prediction based on users' traces, a popular and emerging topic, supports a series of important applications ranging from individual-level recommendation systems to large-scale smart transportation~\cite{song2010limits}. For instance, one of the prerequisites for a successful LBS-recommendation system is the ability to predict users' activities or locations ahead of time, tracking their intentions and forecasting where they will go~\cite{gomes2013will}.
While there is no doubt of the usefulness of predictive applications for mobility data, privacy concerns regarding the collection and share of individuals’ mobility traces have prevented the data from being utilized to their full potential~\cite{shokri2011quantifying, beresford2003location, krumm2009survey}. A mobility privacy study conducted by De Montjoye et al.~\cite{de2013unique} illustrates that four spatio-temporal points are enough to identify 95\% of the individuals, which exacerbates the user re-identification risk and could be the origin of many unexpected privacy leakages.
Additionally, with increasingly intelligent devices and sensors being utilized to collect information about users' locations, a malicious third party can derive increasing intimate details about users' lives, from their social life to their preferences.
Hence, a mechanism capable of decreasing the chance of user re-identification against malicious attackers or untrusted SPs can offer enhanced privacy protection in mobility data applications, as human mobility traces are highly unique.
In the past decade, the research community has extensively studied the privacy of geolocated data via various location privacy protection mechanisms (LPPM)~\cite{gedik2005location, gedik2007protecting}. Some traditional privacy-preserving approaches such as k-anonymity and geo-masking have shown to be insufficient to prevent users from being re-identified~\cite{de2013unique,song2010limits,gonzalez2008understanding, malekzadeh2020privacy}. Differential privacy (DP), another popular notion, is shown to be a limited metric for location trace privacy since temporal correlations are not taken into account~\cite{shokri2011quantifying}. ~\cite{erdemir_privacy-aware_2020} also states that DP and k-anonymity are meant to ensure the privacy of a single data point in time. In general, many DP for LBS (DP-L) mechanisms~\cite{andres2013geo, cunningham2021privacy, he2015dpt} attempt to protect the \emph{user's location} instead of protecting the \emph{user's identity}, which is outside the scope of our problem. However, we integrate one exemplary DP-L mechanism~\cite{andres2013geo} with a corresponding analysis of the user re-identification risk to further examine the effectiveness of the DP-L mechanism and achieve a more comprehensive comparison of our proposed mechanism.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{1_introduction/scenario.pdf}
\caption{Privacy protection in users location data collection and sharing.}
\label{fig:scenario}
\end{figure}
\begin{figure*}[t]
\centering
\begin{subfigure}[b]{0.72\textwidth}
\centering
\includegraphics[width=\textwidth]{1_introduction/Proposed.pdf}
\caption{Mo-PAE}
\label{fig:proposed}
\end{subfigure}
\begin{subfigure}[b]{0.21\textwidth}
\centering
\includegraphics[width=\textwidth]{1_introduction/standalone.pdf}
\caption{Optimal-IMs}
\label{fig:baseline}
\end{subfigure}
\caption{(a) Schematic overview of the proposed privacy-preserving adversarial architecture (Mo-PAE), consisting of DRU, MPU, and URU. (b) The baseline LSTM network for optimal classifiers (Optimal-IMs).}
\label{fig:architecture}
\end{figure*}
More recently, some related works have successfully applied machine-learning- or deep-learning-based approaches to explore effective LPPMs. Rao et al. proposed an model based on Generative Adversarial Network (GAN)~\cite{goodfellow2014generative} to generate privacy-preserving synthetic mobility datasets for data sharing and publication~\cite{rao2020lstm}. Feng et al. investigated human mobility data with privacy constraints via federated learning, achieving promising prediction performance while preserving the personal data on the local devices~\cite{feng2020pmf}.
Though these works provide promising architectures to protect location privacy, the mobility data's privacy protection and utility degradation have not been fully investigated, especially in reducing the chance the user re-identification.
Our work extends these machine-learning-based mechanisms and explores the privacy-utility trade-off on mobility data in terms of declining the effectiveness of privacy inference attacks while maintaining its predictability.
Research on human mobility shows that the predictability of users' location trajectories or mobility, and the particular constraints of users' movements, are sufficient to reconstruct and/or identify anonymous or perturbed locations~\cite{shokri2012protecting}.
This specific confrontation makes the trade-off between mobility predictability and users' identity more challenging.
Consider a scenario, shown in Figure~\ref{fig:scenario}, where users share their daily traces to a trusted mobile network operator, which then aggregates these traces in a privacy-preserving approach and sends them to third parties or other SPs with/without users' consent.
These users may want to minimize the risk of being re-identified and trajectory reconstructed by those who will access these released data. However, they would like to keep receiving potential effective services from SPs. Therefore, a privacy-preserving mechanism, which can release required information for the services (i.e., utility) while features or patterns that facilitate fully data reconstruction or user re-identification are obscured (i.e., privacy), is beneficial.
To this end, we propose a {\bf p}rivacy-aware {\bf a}dversarial network to train a feature extractor $Enc_L$ for {\bf mo}bility privacy, as shown in Figure~\ref{fig:architecture}, namely Mo-PAE.
Mo-PAE is based on representation learning and aims to ease data sharing privacy concerns from privacy inference attacks.
Inspired by PAN~\cite{liu2019privacy} (privacy adversarial network), we employ adversarial learning to better balance the potential trade-off between privacy and utility.
In contrast to PAN, which focuses on the privacy of images, our approach is designed for complex time-series data that exhibits spatial-temporal characteristics.
At the core of our architecture lies an auto-encoder (AE) and long short-term memory (LSTM) layers with three branches, corresponding to the three training optimization objectives of the feature extractor $Enc_L$: i) to $maximize$ the loss associated with the reconstructed output by generative learning, ii) to $minimize$ the prediction loss using the learned representation from the $Enc_L$ by discriminative learning, and iii) to $minimize$ the percentage of users who are re-identifiable through their trajectories by discriminative learning.
We explore and quantify the privacy-utility trade-off achieve by Mo-PAE in terms of data reconstruction leakage (i.e., \textit{Privacy I}), user re-identification risk (i.e., \textit{Privacy II}), and mobility predictability (i.e., \textit{Utility}).
The results show that our proposed mechanism achieves a better privacy level with the same utility loss and vice versa.
The contributions of our work are the following:
\begin{itemize}
\item We propose Mo-PAE to learn an effective privacy-preserving feature extraction \textbf{e}ncoder for mobility applications.
\item We provide an extensive analysis of different mobility tasks and quantify the privacy and utility bound of the target mobility dataset, along with a trade-off analysis between these contrasting objectives.
\item We report the analysis of Mo-PAE by a comprehensive evaluation of four real-world representative mobility datasets.
\item We compare our model with, i) a famous DP notion that developed on the idea from Geo-indistinguishability~\cite{andres2013geo} (namely GI-DP);
ii) a state-of-the-art GAN-based mechanism that attempts to generate synthetic privacy-preserving mobility data (namely TrajGAN~\cite{rao2020lstm}); iii) as well as the optimal LSTM-based inference model, and obtaining favorable results.
\end{itemize}
The rest of this paper is structured as follows:
we review the related work in Section \ref{RelW}; the proposed Mo-PAE is described in detail in Section \ref{DoF}; we describe the experimental settings in Section \ref{ExpS}; we demonstrate an evaluation of our mechanism over four mobility datasets with baseline comparisons in Section \ref{FraE}; Section \ref{sec:discussions} reports an in-depth discussion of our setting; finally, we conclude the paper with future work directions in Section \ref{Conl}.
\section{Related work}
\label{RelW}
\subsection{Notions of Location Privacy}
Diverse privacy notions, \emph{direct} or \emph{indirect}, for the LBSs have been proposed and evaluated in the literature. In~\cite{andres2013geo}, various \emph{direct notions} of location privacy and the techniques to achieve them are examined and concluded, including but not limited to expected distance error, $k$-anonymity, differential privacy (DP), and other location-privacy metrics.
First, the expectation of distance error reflects the accuracy when an adversary guesses the user's real location in a location-obfuscation mechanism by using the available side information. In~\cite{shokri2012protecting}, an optimal LPPM strategy and its corresponding optimal inference attack are obtained by formalizing the mutual optimization of user-adversary objectives (location privacy vs correctness of localization).
Second, $k$-anonymity is the most widely used privacy notion for the LBSs~\cite{wang2019achieving}. These systems aim to protect the \emph{user's identity}, requiring that the attacker cannot infer the correct user among a set of $k$ different users.
Third, DP~\cite{dwork2008differential} is an emerging notion initially formulated in the context of statistical databases and aims to protect an individual's data while publishing aggregate information about the dataset. More precisely, a randomization mechanism $M$ gives $\epsilon$-differential privacy for all neighbouring datasets $D$ and $D'$, and the difference between $D$ and $D'$ is within a bound of $e^\epsilon$. One of the popular mechanisms to achieve DP perturb the original query result using random noise that is calibrated with the privacy budget $\epsilon$ and defines a global sensitivity for all neighbouring $D$ and $D'$~\cite{dwork2014algorithmic}. The work in~\cite{errounda2019analysis} reviews research works done in differential privacy targeted toward location data from a data flow perspective, including collection, aggregation, and mining.
~\cite{andres2013geo} proposed a Geo-indistinguishability notion based on differential privacy and a planar Laplace mechanism.
Significantly, different from the systems in $k$-anonymity category aim at protecting \emph{user's identity}, DP mechanisms are interested in protecting the \emph{user's locations}~\cite{andres2013geo, cunningham2021privacy, he2015dpt, gursoy2020utility}.
Apart from three mainstream approaches, the location cloaking mechanism tries to define the uncertainty region and measure the privacy by the size of the cloak and by the coverage of sensitive regions; the inaccuracy of the sensing technology tries to achieve a certain level of privacy by increasing uncertainty; and transformation-based approach tries to make user's location invisible to the service provider.
In the other hand, \emph{indirect} notions of location privacy arise with the emerging machine learning-based mechanism, which assesses the privacy guarantee by measuring the effectiveness of target inference attacks~\cite{chatzikokolakis2015geo, primault2014differentially, gursoy2018utility}.
In general, for any LBS, their main privacy concerns can be concluded in two categories. One is the attack on the \emph{user's identity} which can be re-identified maliciously. For instance, even if the adversary is assumed to be unaware of the user identity of a trace, they can infer \emph{user's identity} or additional sensitive information due to the location information leakage based on publicly accessible background information.
The other attack is the one on \emph{user's location} while the adversary has legible access to \emph{user's identity}. In this manner, \emph{user's locations} are sensitive, which could exert a significant impact on other sensitive personal details, such as religious affiliation, sexual orientation, economic condition, health status, and so on.
In our work, we are interested in protecting \emph{user's identity} as the privacy scope, which is similar to the location privacy notion defined by the $k$-anonymity, and taking the real/distorted \emph{user's location} as input for the personal recommendation model to provide contextual services for their future travels. In general, DP paradigms have the most formal privacy guarantee than the others, however, they are not immune to inference attacks~\cite{clifton2013syntactic, hamm2017minimax}. We will also compare our proposed model with one popular DP paradigm on location privacy to illustrate the ineffectiveness of DP to our research question.
More details on our privacy definitions are in Section~\ref{DoF}.
\subsection{Location Privacy Preserving Mechanisms}
An effective location-privacy preserving mechanism (LPPM) must consider three fundamental elements: i). the privacy requirements of the users (namely \textit{privacy gain}); ii). the adversary's knowledge and capabilities; iii). and maximal tolerated service quality degradation stemming from the obfuscation of true locations (namely, \textit{utility loss})~\cite{shokri2012protecting}.
The literature on location privacy can be roughly classified into three categories: the design of LPPMs~\cite{bindschaedler2016synthesizing}; recovering actual user traces from anonymized or perturbed traces; the formal analysis and the search for an appropriate location privacy metric that allows for the fair comparison between LPPMs~\cite{shokri2012protecting, lippmann2000evaluating}.
Generally, typical LPPMs use some obfuscation methods - like spatial cloaking, cell, merging, location precision reduction, or dummy cells - to manipulate the probability distribution of user's location.
The most popular approach to protect location privacy is to send a space- or time-obfuscated version of the users' actual traces to the trusted or un-trusted third parties~\cite{shokri2012protecting}.
Xiao et al.~\cite{xiao2015protecting} investigate how to obtain location privacy under temporal correlations with an optimal DP-based LPPM.
Another mainstream approach tries to issue the dummy requests from fake locations to the services provider, the location privacy hence protected as these fake locations increase the uncertainty of the adversary about the users' real movements~\cite{shokri2012protecting}.
The other popular alternative utilizes mixed zones or silent periods to hide users' locations, as the adversary cannot link those who enter with those who exit the region when several users traverse the zone simultaneously.
In general, any LPPM would alter the location information, resulting in a severe distortion of data. Therefore, designing an optimized privacy-preserving algorithm with constrained utility degradation according to user privacy requirements is one critical dimension of LPPMs~\cite{bindschaedler2016synthesizing}.
There is no way to optimally address location privacy issues for all types of location-based systems, and the design of a specific LPPM requires carefully considering the application scenario and the realistic privacy requirements of mobile users~\cite{jiang2021location}. Hence, if a user prefers high service quality rather than the concerns of privacy leakage, then a more flexible system could be applied to guarantee the service quality.
Our proposed model, in this way, can perform flexibly in application scenarios when users have different focuses on privacy or service.
\subsection{Privacy Preserving Techniques for Spatial-Temporal Data}
Current privacy-preserving techniques for spatial-temporal data focus on two research streams.
One is the DP approach to grouping and mixing the trajectories from different users so that the identification of individual trajectory data is converted into a k-anonymity problem~\cite{aktay2020google,xiao2015protecting,andres2013geo}.
For example, a recent Privacy-Preserving Trajectory Framework (PPTPF)~\cite{yang_pptpf_2021} applies the k-indistinguishability to anonymize trips for each user by condensing them into $k-1$ trajectories and determining $k-1$ anonymized clusters of trips.
The other stream focuses on synthetic data generation~\cite{rezaei2018protecting,huang_variational_2019,choi2021trajgail, ijcai2018-530}.
Synthetic data generation methods have been extensively studied in recent years as a way of tackling privacy concerns of location-based datasets.
The majority of existing mobility synthesis schemes are mainly categorized into two approaches:
one is a more traditional, simulation-based approach, while the other is a more recent, neural network-based generative modeling approach that utilizes recurrent autoencoders and generative adversarial networks to produce realistic trajectories~\cite{shin2020user}.
Simulation-based approaches generate mobility traces by modeling overall user behavior as a stochastic process, such as a Markov chain model of transition probabilities between locations, and then drawing random walks,
potentially with additional stochastic noise added, as demonstrated in Xiao et al \cite{xiao_loclok_2017}.
These approaches require considerable feature engineering effort and struggle to capture longer-range temporal and spatial dependencies in the data~\cite{luca2021survey} and are thus limited in their ability to preserve
the utility of the original datasets.
In contrast, the generative neural network approach synthesizes user mobility traces by learning via gradient descent back-propagation, then the optimal weights are utilized for decoding a high-dimensional latent vector representation into sequences that closely resemble the original data.
Such traces can maintain important statistical properties of the original data while taking advantage of noise introduced in the reconstruction process, to improve data subject anonymity.
Huang et al \cite{huang_variational_2019} demonstrates the use of a variational autoencoder network to reconstruct trajectory sequences,
while Ouyang et al \cite{ijcai2018-530} utilizes a convolutional GAN, but neither work directly makes a quantitative
assessment of the extent of privacy protection that their algorithms provide \cite{huang_variational_2019, ijcai2018-530}.
The TrajGAN by Rao et al \cite{rao2020lstm} is a state of the art example of the generative trajectory modeling approach, which quantifies its privacy protection by demonstrating a significant decline in the performance of a second user ID classifier model on the synthetic outputs compared to the original input trajectories.
For these reasons, we used TrajGAN as a baseline for comparison.
Our proposed model takes the neural network-based generative modeling approach, but differs from existing methods, where we utilize a combined, multi-task adversarial neural network to simultaneously reconstruct trajectories, predict next locations, and re-identify users, from the same learned latent vector representation.
We seek an optimal trade-off between the three tasks' individual losses by optimizing a sum loss function with per-task weights, improving the controllability of the relative utility and privacy of the outputs.
\section{Design of the Architecture}
\label{DoF}
\subsection{Definition of Important Terms}
\subsubsection{Mobility Trace}
The raw geolocated data or other mobility data commonly contain three elements: user identifiers \textit{u}, timestamps \textit{t}, and location identifiers \textit{l}. Hence, each location record \textit{r} could be denoted as \emph{$r_i$} = [\textit{$u_i$}, \textit{$t_i$}, \textit{$l_i$}], while each location sequence \emph{S} is a set of ordered location records \emph{$S_n$} =~\{\textit{$r_1$, $r_2$, $r_3$, $\cdots$, $r_n$}\}, namely \textit{mobility trace}. In this paper, we assume that different users' mobility traces are collected and aggregated by trusted telecom operators or social platforms and shared with third-party SPs (trusted or untrusted). Therefore, given the past mobility trace \emph{$S_n$} =~\{\textit{$r_1$, $r_2$, $r_3$, $\cdots$, $r_n$}\}, the mobility prediction task is to infer the most likely location \emph{l$_{n+1}$} at the next timestamp \emph{t$_{n+1}$}. The data fed into the proposed architecture are a list of traces with specific sequence length (\emph{SL}), that is \{$S_{sl}^1$, $S_{sl}^2$, $S_{sl}^3$, $\cdots$, $S_{sl}^j$\}. For instance, if the sequence length is 10, that indicates each trace contains 10 history location records \textit{r}, \emph{$S_{10}$} = \{$r_1$, $r_2$, $r_3$, $\cdots$, $r_{10}$\}, and $SL=10$.
\subsubsection{User Re-identification}
The user re-identification risk arises because of the high uniqueness of human traces~\cite{de2013unique} and could be the origin of many unexpected privacy leakages. We assume each trace \emph{S} is originally labeled with a corresponding user identifier \textit{u}, and the user re-identification is to infer the user \textit{u} to whom the target trace \emph{$S_n$} = \{$r_1$, $r_2$, $r_3$, $\cdots$, $r_n$\} belongs. We thereby leverage the user identifiers \textit{u} as the ground-truth values for the user identity classes. This identity information is what we want to protect in the proposed adversarial network; that is, the entire network should convey as little user identifiable information as possible to decrease the user re-identification accuracy. At the same time, the built-in adversary tries to achieve maximum accuracy.
\subsection{Problem Definitions}
Research on human mobility or LBS shows that the predictability of users' traces or mobility, and the particular constraints of users' movements, are sufficient to reconstruct and/or identify anonymous or perturbed locations~\cite{shokri2012protecting}. This confrontation makes the trade-off between keeping mobility predictability and reducing the chance of user re-identification more interesting. For instance, an adversary can re-identify anonymous users' traces given the users' mobility profile~\cite{de2019give}; infer the users' next activities from the frequency of location visits~\cite{gomes2013will}; even obtain the personal home or working address from the trajectories~\cite{primault2018long}. A number of works on location privacy protection try to evaluate various questions about location information leakage~\cite{shokri2012protecting}. In this work, we design a model to protect location privacy regarding users' identity and data integrity while simultaneously minimizing the service quality (i.e., accuracy of next location forecasting) degradation stemming from the obfuscation of true data. Before describing our proposed model in detail, we first give a brief problem definition of the trade-offs of mobility data between utility and privacy.
\textbf{Data Utility:}
Mobility datasets are of great value for understanding human behavior patterns, smart transportation, urban planning, public health issue, pandemic management, etc. Many of these applications rely on the next location forecasting of individuals, which in the broader context, can provide an accurate portrayal of citizens' mobility over time. Mobility prediction not only can be analyzed to understand personalized mobility patterns but can also inform the allocation of public resources and community services. We focus on the capability of \emph{mobility prediction} (\emph{next location forecasting}) in this paper, and leverage the accuracy of the prediction as an important metric for quantifying the data utility. Hence, we define \textit{Utility} as follows:
\textit{Utility} (\textit{U}): the mobility predictability (i.e., forecasting accuracy). A higher forecasting accuracy indicates higher utility. \textit{Utility loss} then represents the accuracy degradation when the proposed privacy protection mechanism is applied.
\textbf{Privacy Protection:}
Mobility datasets shared with trusted or un-trusted third parties/SPs can be utilized to attain personalized preferences. However, these shared traces also contain sensitive information. With increasing intelligent devices and sensors being utilized to collect information about human activities, the traces also increasingly expose intimate details about users' lives, from their social life to their preferences. The capability of user re-identification is important to balance the risks and benefits of mobility data usage, for all data owners, third parties, and researchers. In our proposed architecture, we design two built-in adversaries to infer the ability of generated features to protect users' sensitive information. We then leverage the reduction of data reconstruction risk and user re-identification risk as our privacy metrics to evaluate our proposed architecture. Hence, the definition of \textit{Privacy} is summarized as follows:
\textit{Privacy I} (\textit{PI}): the distance between the reconstructed data $X'$ and the original data $X$, that is, information loss in the reconstruction process. More information loss indicates higher privacy protection.
\textit{Privacy II} (\textit{PII}): the decline of user re-identification accuracy, that is, the user de-identification effectiveness. The higher degradation of user re-identification accuracy indicates higher privacy protection, as fewer users would be re-identified by the adversary during the privacy attack.
Similar to the definition of the \textit{utility loss}, \textit{privacy gain} (in terms of \textit{PI} and \textit{PII}) quantifies the privacy information protected by the designed privacy-preserving mechanism.
\textbf{Privacy vs. Utility Trade-off:}
There is an inherent trade-off between location privacy protection and utility degradation~\cite{jiang2021location}. That is, achieving a better level of privacy protection may require sacrificing the service quality provided by the data. Such trade-off is omnipresent in various privacy protection mechanisms, especially in location obfuscation mechanisms. Higher privacy protection is achieved when the probability of an adversary inferring the true location of the user decrease, however, the result of a query based on the obfuscated location is significantly different from the actual interest of the user. The privacy-utility trade-off, hence, need to be examined and analyzed to guarantee the efficiency of the privacy protection mechanism.
To be specific, a more obfuscated dataset will tend to perform better at preserving privacy, but worse at preserving utility, and vice versa. Hence, monitoring these two performance metrics in tandem allows users to select the optimal privacy-utility trade-off for their use cases, given their hyperparameter selections. Our Mo-PAE model is designed to train a features Encoder $Enc_L(X)$ that could convey more information on the utility but less on privacy and investigate a better trade-off between them. More details will be discussed in the following sub-section.
\subsection{Architecture Overview}
\label{section: module overview}
We discuss the overall design of the Mo-PAE before detailly reviewing the functionality of each unit.
\subsubsection{Overall Design}
\hfill
Our proposed \textbf{p}rivacy-preserving \textbf{a}dversarial feature \textbf{e}ncoder on mobility data, denoted as the \emph{Mo-PAE}, is based on representation learning and adversarial learning and aims to ease data sharing privacy concerns. Figure~\ref{fig:architecture} presents the basic workflow of the proposed Mo-PAE. It composes of three crucial units: data reconstruction risk unit (DRU), mobility prediction unit (MPU), and user re-identification risk unit (URU).
When three units train concurrently, the MPU is regarded as the \textit{utility discriminator} $U_D$, while DRU and MPU act as two built-in adversaries and are regarded as the two \textit{privacy discriminators}, $P_D^1$ and $P_D^2$, respectively. The built-in adversary has been used as an effective adversarial regularization to prevent inference attacks, \textit{e.g.} in the classification setting~\cite{nasr2018machine} or in various GAN models~\cite{mukherjee2021privgan} for privacy-preserving purpose. We design two built-in adversaries to infer the ability of generated features to protect user's sensitive information, and they are trying to maximize the accuracy of privacy inference tasks in training. At the same time, an encoder $Enc_L$ is trained to produce feature representations $f$ from the target mobility data by jointly optimizing these extracted feature weights using the combined losses of the DRU, MPU, and URU simultaneously, during adversarial training. Therefore, in the Mo-PAE, the encoder $Enc_L$, and three discriminators $U_D$, $P_D^1$, $P_D^2$ play a multi-player game to minimaximize the value function $V(Enc_L, U_D, P_D^1, P_D^2)$:
\begin{equation}
\begin{aligned}
& \underset{Enc_L, U_D}{\min} \underset{P_D^1, P_D^2}{\max} V(Enc_L, U_D, P_D^1, P_D^2) = \mathbb{E}_{x\sim X} [log U_D(Enc_L(x))] +\\ & \mathbb{E}_{x\sim X} [log(1- P_D^1(Enc_L(x)))] +
\mathbb{E}_{x\sim X} [log(1- P_D^2(Enc_L(x)))]
\end{aligned}
\label{equ:adv}
\end{equation}
As described in the Eq.\ref{equ:adv}, we design a multi-task adversarial network to learn an LSTM-based encoder $Enc_L(X;\theta)$ with parameter set $\theta \in \Theta$,
which can generate the optimized feature representations $f=Enc_L(X;\theta)$ via lowering the privacy disclosure risk of user identification information and improving the task accuracy (i.e., mobility predictability) concurrently.
Two potential malicious privacy leakages from URU and DRU, are attempted to retrieve sensitive information from the feature representations $f$. As built-in adversaries, they have full access to the feature representations $f$ and the entire encoder network with parameter set $f=Enc_L(X;\theta)$. In this manner, they have the optimal decoder setting.
Hence, the notion of privacy (\emph{privacy gain}), is measured by the decline of the effectiveness of target inference attacks (i.e., user re-identification attack and data reconstruction attack).
\begin{algorithm}[t]
\SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output}
\SetAlgoLined
\Input{Mobility data \textbf{X}, real mobility prediction labels \textbf{Y}, real user identification labels \textbf{Z}, weights: $\lambda_1$, $\lambda_2$, $\lambda_3$}
\Output{Adversarial Encoder $Enc_L(X; \theta_E, \theta_R, \theta_U, \theta_P)$}
Initialize model parameters $\theta_E, \theta_R, \theta_U, \theta_P$;\\
\For{n epochs}{
\For{$k = 1$, $\cdots$$, K_t$}{
1. Sample a mini-batch of mobility trajectories x, prediction labels y, identification labels z\\
2. Update $\theta_E$ with Adam optimizer on mini-batch loss $L_{sum}(\theta_E, \theta_R, \theta_U, \theta_P, \lambda_1, \lambda_2, \lambda_3)$\\
3. Update $\theta_R$ with Adam optimizer on mini-batch loss $L_R(f; \theta_R)_{(x, \hat{x})}$: $\min L_R$\\
4. Update $\theta_U$ with Adam optimizer on mini-batch loss $L_U(f; \theta_U)_{(y, \hat{y})}$: $\min L_U$\\
5. Update $\theta_P$ with Adam optimizer on mini-batch loss $L_P(f; \theta_P)_{(z, \hat{z})}$: $\min L_P$\\
}
Update with the gradient descent on $L_{sum}(\theta_E, \theta_R, \theta_U, \theta_P, \lambda_1, \lambda_2, \lambda_3)$: $\min L_{sum}$
}
\caption{Training of the Mo-PAE (\emph{Model II})}
\label{algo_disjdecomp}
\end{algorithm}
\subsubsection{Details of Mo-PAE}
\hfill
We define the raw mobility data we want to protect as $\mathcal{X}$, trained features as $\mathcal{F}$, and reconstructed data as $\mathcal{X'}$. Given mobility raw data $\mathcal{X}$ for $P_D^1$ (DRU), the ground-truth label $z_i$ for $P_D^2$ (URU), and the ground-truth label $y_i$ for utility $U_D$ (MPU), we train the encoder $Enc_L$ to learn the representation $\mathcal{F}=Enc_L(\mathcal{X}; \theta_E)$. We design a specific loss function, namely \textit{sum loss} $\mathcal{L}_{sum}$, for this optimization process.
Specifically, when reconstructing the data $\mathcal{X'}$, a decoder $Dec_L$ attempts to recreate the data based on the features $\mathcal{F}$, that is $Dec_L(\mathcal{F}; \theta'_D): \mathcal{F}\rightarrow \mathcal{X'}$. This DRU, the first privacy discriminator $P_D^1$, is trained as a built-in adversary and tries to achieve sensitive information as much as possible. Hence, the DRU is primarily trained by minimizing the reconstruction loss $\mathcal{L_R}$:
\begin{equation}
\min \mathcal{L_R} \Rightarrow \mathcal{L_R} = d(\mathcal{X}, \mathcal{X'}) = \underset{\mathcal{F}; \theta'_R}{\argminA} \| Dec_L(\mathcal{F}, \theta'_R) - \mathcal{X}\|^2
\end{equation}
The URU, the second privacy discriminator $P_D^2(\mathcal{F}; \theta')$, is trained to re-identify whom the target trajectory belongs to.
It outputs a probability distribution of predicted user identifiers among Z potential classes.
Then in this privacy discriminator, the user re-identification loss $\mathcal{L_P}$ is primarily trained to minimize, denoted as $\min\mathcal{L_P}$:
\begin{equation}
\min \mathcal{L_P} \Rightarrow \mathcal{L_P} = \underset{\mathcal{F}; \theta'_P}{\argminA} \sum_{i=1}^{\mathbf{Z}} z_i\text{log}(P_D^1(\mathcal{F}; \theta'_P))
\end{equation}
The MPU, the utility discriminator $U_D(\mathcal{F}; \theta')$, is trained to output a probability distribution of the next location of interest, and this distribution has Y potential classes.
Discriminative training of $U_D$ maximizes the prediction accuracy by minimizing the utility loss $\mathcal{L_U}$ concurrently with minimizing the $\mathcal{L}_{sum}$, denoted as $\min\mathcal{L_U}$.
\begin{equation}
\min\mathcal{L_U} \Rightarrow \mathcal{L_U} = \underset{\mathcal{F}; \theta'_U}{\argminA} \sum_{i=1}^{\mathbf{Y}} y_i\text{log}(U_D(\mathcal{F}; \theta'_U))
\end{equation}
The overall training is to achieve a privacy-utility trade-off by adversarial learning on $\mathcal{L_R}$, $\mathcal{L_U}$, and $\mathcal{L_P}$, concurrently. The encoder $Enc_L(\mathcal{X};\theta_E)$ should satisfy high predictability (\textit{min}~$\mathcal{L_U}$) and low user re-identification accuracy (\textit{max}~$\mathcal{L_P}$) of the mobility data when maximizing the reconstruction loss (\textit{max}~$\mathcal{L_R}$) in reverse engineering, where the training objective transformed from Eq.\ref{equ:adv} can be written as:
\begin{equation}
\min \mathcal{L}_{sum} = \underset{\mathcal{L}_{U}}{\text{min}}\ \underset{\mathcal{L}_{R}, \mathcal{L}_{P}}{\text{max}} (\sum_{x=i}^{\mathcal{X}} \left(\mathcal{L}_{U} \left(f_i\right), \mathcal{L}_{P}\left(f_i\right), \mathcal{L}_{R}\left(f_i\right)\right))
\label{equ:sumloss1}
\end{equation}
We use Eq.~\ref{equ:sumloss1} to guide the first version of Mo-PAE, denoted as \emph{Model I}.
In order to fully investigate the range of trade-offs, we leveraged the Lagrange multipliers~\cite{beavis1990optimisation} as hyperparameters to control the privacy-utility tradeoffs in the Mo-PAE, and this weighted-controlled model is denoted as \emph{Model II}. Accordingly, the optimization function of the training objective is:
\begin{equation}
\begin{aligned}
& \min \mathcal{L}_{sum} = \underset{\mathcal{L}_{U}}{\text{min}}\ \underset{\mathcal{L}_{R}, \mathcal{L}_{P}}{\text{max}} (\sum_{x=i}^{\mathcal{X}} \left(\lambda_1\mathcal{L}_{R} \left(f_i\right), \lambda_2\mathcal{L}_{U}\left(f_i\right), \lambda_3\mathcal{L}_{P}\left(f_i\right)\right)) \\
&= \underbrace{- \lambda_1 \left(\max\mathcal{L_R}\left(f_i\right)\right)}_{\text{Privacy\ I}} + \underbrace{\lambda_2 \left(\min\mathcal{L_U}\left(f_i\right)\right)}_{\text{Utility}}
\underbrace{- \lambda_3 \left(\max\mathcal{L_P}\left(f_i\right)\right)}_{\text{Privacy\ II}} \\
&= -\lambda_1 \| Dec_L(\mathcal{F}) - \mathcal{X}\|^2 + \lambda_2 (\sum_{i=1}^{\mathbf{Y}} y_i\text{log}(U_D(\mathcal{F}))) \\
&- \lambda_3 (\sum_{i=1}^{\mathbf{Z}} z_i\text{log}(P_D(\mathcal{F})))
\end{aligned}
\label{equ:sumloss}
\end{equation}
where $y_i$ is the ground-truth label for \textit{Utility}, $z_i$ is the ground-truth value for \textit{Privacy II}; $\lambda_1$, $\lambda_2$ and $\lambda_3$ are non-negative, real-valued weights, as the hyperparameters that control the privacy-utility trade-off in the Mo-PAE.
As shown in the Algorithm~\ref{algo_disjdecomp}, the gradient of the loss (i.e., $\theta_E$, $\theta_R$, $\theta_U$, $\theta_P$) back-propagates through the LSTM network to guide the training of the encoder $Enc_L$. The encoder is updated with the \textit{sum loss} function $\mathcal{L}_{sum}$ until convergence. It is tricky to practically investigate all possible weight combinations, we look for the optimal combinations through training~\cite{malekzadeh2020privacy} with the Eq.\ref{equ:sumloss} by brute-force evaluation. Then we approximate the required data utility reserved and reformulate the optimization problem in Eq.\ref{equ:sumloss} as a maxima privacy optimization problem.
\begin{equation}
\begin{aligned}
& \underset{Enc_L}{\min} \underset{P_D^1, P_D^2}{\max} V_{\lambda \rightarrow U_D}(Enc_L, P_D^1, P_D^2) = \mathbb{E}_{x\sim X} [log(1- P_D^1(Enc_L(x)))] \\ & + \mathbb{E}_{x\sim X} [log(1- P_D^2(Enc_L(x)))]
\end{aligned}
\label{equ:adv2}
\end{equation}
Additionally, another key contribution is the flexibility of the \textit{sum loss} function $\mathcal{L}_{sum}$ which could be regulated to satisfy different requirements on privacy protection level and service quality. That is, different combinations of weights control the relative importance of each unit and guide the overall model to find the maxima or minima given the specific trade-off choices.
\input{3_proposedmodel/table_datasets}
\subsubsection{Composition Units of Architecture}
\hfill
\textit{I. Mobility Prediction Unit (MPU)}:
The MPU unit is composed of three parts, the input part with the multi-modal embedding of trace information, the sequential part with LSTM layers~\cite{hochreiter1997long}, and an output part with the softmax activation function. As per the definition mentioned earlier, the traces in this work are shown as location sequences \emph{S}. First, the location identifiers \textit{l} and timestamps \textit{t} are converted into one-hot vectors. We then employ LSTM layers to model the mobility patterns and sequential transition relations in these mobility traces. As a prominent variant of the recurrent neural network, LSTM networks exhibit brilliant performance in modeling the entire data sequences, especially for learning long-term dependencies via gradient descent~\cite{zhan2019towards}. Following the sequential module, the softmax layer outputs the probability distribution of the prediction results. This probability distribution is converted to the top-n accuracy metrics to illustrate the unit performance.
\textit{II. Data Reconstruction Risk Unit (DRU)}:
The DRU is the encoder $Enc_L$ unit in reverse, also denoted as $Dec_L$, which is regarded as the first \textit{privacy discriminator} $P_D^1$ in the proposed architecture.
It is designed to evaluate the distance $d(\cdot,\cdot)$ (i.e., \textit{Privacy I}) between the reconstructed data $X'$ and the original input data $X$. A malicious party is free to explore any machine learning model and reconstruct the data if they have the shared extracted features $f$. We use a layer-to-layer reverse architecture of our encoder $Enc_L$ to build the \textit{data reconstruction unit} to act as a robust built-in adversary. To compare with baseline models and keep the comparison in a line, we measure the distance $d(\cdot,\cdot)$ between the $X$ and $X'$ by leveraging the \textit{Euclidean} and \textit{Manhattan} distance as our metrics. Both of them are widely used in location privacy literature~\cite{andres2013geo, erdemir_privacy-aware_2020}.
\textit{III. User Re-identification Risk Unit (URU)}:
The URU is regarded as the second \textit{privacy discriminator} $P_D^2$ in the proposed architecture.
The unit is composed of three parts, the input part with the one-hot embedding of user identity, the sequential part with LSTM layers, and an output part with softmax function. First, the user identity list is converted into one-hot vectors. Similar to the MPU, the URU also applies LSTM layers to better extract the spatial and temporal characteristics of the context. A softmax function with a cross-categorical entropy loss function is applied to output a categorical probability distribution of the user re-identification task. We then use the top-N accuracy of this classifier as the metrics of user re-identification privacy risk (i.e., \textit{Privacy II}). The more accurately a classifier can re-identify the user when given a trajectory, the higher the risk of disclosing private data.
Same as $P_D^1$, $P_D^2$ are designed as the built-in adversary to infer the ability of generated features in protecting users' sensitive information.
The overall architecture eventually learns to fool both built-in adversaries, $P_D^1$ and $P_D^2$, while maintaining the mobility predictability. In this manner, both adversaries are assumed to be free to access the exclusive feature representations and the entire encoder network, which allows them to have the optimal decoder setting. We will discuss the effectiveness of two privacy inference attacks in Section~\ref{UPA}.
\section{Experimental Setting}
\label{ExpS}
\subsection{Datasets}
Experiments are conducted on four representative mobility datasets: Mobile Data Challenge Dataset (MDC)~\cite{laurila2012mobile}, Priva'Mov~\cite{mokhtar2017priva}, GeoLife~\cite{zheng2011geolife}, and FourSquare~\cite{dingqi_yang_modeling_2015}. Once imported into our architecture, each dataset was filtered and preprocessed individually to derive their respective train and test sets illustrated in Table~\ref{tab:datasets}. Each bounding box defines the grid size and the grid granularity is 0.01 degrees.
\noindent
\textbf{MDC}: it is recorded from 2009 to 2011 and contains a large amount of continuous mobility data for 184 volunteers with smartphones running a data collection software, in the Lausanne/Geneva area.
Each record of the \textit{gps-wlan} dataset represents a phone call or an observation of a WLAN access point collected during the campaign~\cite{laurila2012mobile}.
\noindent
\textbf{Priva'Mov}: the PRIVA'MOV crowd-sensing campaign took place in the city of Lyon/France from October 2014 to January 2016. Data collection was contributed by roughly 100 participants including university students, staff, and family members. The crowd-sensing application collected GPS, WiFi, GSM, battery, and accelerometer sensor data. For this paper, we only used the GPS traces from the dataset ~\cite{mokhtar2017priva}.
\noindent
\textbf{GeoLife}: it is collected by Microsoft Research Asia from 182 users in the four and a half year period from April 2007 to October 2011 and contains 17,621 trajectories \cite{zheng2011geolife}.
This dataset recorded a broad range of users’ outdoor movements, including life routines like going home and going to work and some entertainment and sports activities, such as shopping, sightseeing, dining, hiking, and cycling. It is widely used in many research fields, such as mobility pattern mining, user activity recognition, location-based social networks, location privacy, and location recommendation.
\noindent
\textbf{FourSquare NYC}: it contains check-ins in NYC and Tokyo collected during the approximately ten months from 12 April 2012 to 16 February 2013, containing 227,428 check-ins from 1,083 subjects in New York City \cite{dingqi_yang_modeling_2015}.
\subsection{Baseline Models}
\label{sec:baseline}
\subsubsection{I. Optimal Inference Models (Optimal-IMs)}
Optimal-IMs comprise three independent inference models: data reconstruction model, mobility prediction model, and user re-identification model. Each model has a similar layer design as the corresponding unit in the Mo-PAE, however, these three models are completely independent and have no effect on each other. Unlike the Mo-PAE, which leverages adversarial learning to finally attain an extracted feature representation $f$ that satisfies the utility requirements and privacy budgets simultaneously, the Optimal-IMs are only trained for optimal inference accuracy at the individual tasks to characterize the original data.
\subsubsection{II. LSTM-TrajGAN (TrajGAN)~\cite{rao2020lstm}} It is an end-to-end deep learning model to generate synthetic data which preserves essential spatial, temporal, and thematic characteristics of the real trajectory data. Compared with other standard geomasking methods, TrajGAN can better prevent users from being re-identified. The TrajGAN work claims to preserve essential spatial and temporal characteristics of the original data, verified through statistical analysis of the generated synthetic data distributions, which aligns with the mobility prediction-based utility metric in our work.
Hence, we train an optimal mobility prediction model for each dataset and evaluate the mobility predictability of synthetic data generated by the TrajGAN. In contrast to the TrajGAN that aims to generate synthetic data, our proposed Mo-PAE is training an encoder $Enc_L$ that forces the extracted representations \textit{f} to convey maximal utility while minimizing private information about user identity via adversarial learning.
\subsubsection{III. GI-DP~\cite{chatzikokolakis2015geo}} The principle of geo-indistinguishability (GI)~\cite{andres2013geo}, is a formal notion of privacy that protects the user's exact location with a level of privacy that depends on radius r, which corresponds to a generalized version of differential privacy (DP). GI-DP is a mechanism for achieving geo-indistinguishability when the user releases his location repeatedly throughout the day. It fulfills desired protection level by perturbing the actual location with random noises and achieving an optimal trade-off between privacy and utility (i.e., service quality). We re-implement the geo-indistinguishability of optimal utility with graph spanner~\cite{chatzikokolakis2015geo}, namely GI-DP in this paper, to attain the released version data that satisfied the DP guarantees. We then train a series of Optimal-IMs to evaluate the effectiveness of target inference attacks on the released version data in a line to compare with our proposed mechanism.
\subsection{Training}
\subsubsection{Training of Mo-PAE}
The main goal of the proposed adversarial network is to learn an efficient feature representation based on the utility and privacy budgets, using all users' mobility histories. In most experiments in this work, the trajectory sequences consist of 10 historical locations with timestamps (i.e., $SL=10$), and the impact of the varying sequence lengths is discussed in Section~\ref{section:impact sl}.
After data pre-processing, 80\% of each user's records are segmented as the training set and the remaining 20\% as the testing set. We utilize the mini-batch learning method with the size of 128 to train the model until the expected convergence. We take a gradient step to optimize the \textit{sum loss} $L_{sum}$ (i.e., Equation \ref{equ:sumloss}) in terms of $L_R$, $L_U$, and $L_P$ concurrently. Meanwhile, the \textit{sum loss} $L_{sum}$ is optimized by using the Adam optimizer. All the experiments are performed with the Tesla V100 GPU; a round of training would take 30 seconds on average, and each experiment trains for 1000 rounds.
\subsubsection{Training of the TrajGAN}
To provide a state-of-the-art machine learning-based model for comparison, we re-implement the TrajGAN model described in \cite{rao2020lstm} using the same hyperparameters, setting latent vector dimension to 100, using 100 LSTM units per layer, a batch size of 256, utilizing the Adam optimizer with learning rate 0.001 and momentum 0.5, and training for 200 epochs (where one epoch is a pass through the entire training set). We train TrajGAN independently on the training split of each benchmark mobility dataset, and then use it to generate synthetic trajectories from the test set. Then we train the proposed Mo-PAE on the same training data and use it to generate a feature extraction from the same test data. Finally, we evaluate the performance of the user re-identification unit and mobility prediction unit on the real and synthetic test sets generated by TrajGAN, and compare the changes in accuracy to assess the relative utility and privacy of the TrajGAN and Mo-PAE.
\subsubsection{Training of the DP-GI}
We re-implement the DP-GI model described in~\cite{chatzikokolakis2015geo} using the default settings. That is, we set $epsilon$ $\epsilon=0.5$, $dilation$ $ \delta=1.1$, the \emph{distance matrix} $d_x$ is defined by Euclidean distance. From ~\cite{chatzikokolakis2015geo}, let $X$ be a set of locations with metric $d_x$, and let $G(X, E)$ be a $\delta-spanner$ of $X$, if a mechanism K for X is $\frac\epsilon{\delta}d_G$-private, then K is $\epsilon d_x$-private. The dilation of G is calculated as:
\begin{equation}
\delta = \underset{x \neq x' \in X}{max} \frac{d_G(x,x')}{d_x(x, x')}
\label{equ:dilation}
\end{equation}
\begin{equation}
d_G(x, x') \geq d_x (x, x') \hspace{0.5cm} \forall x, x' \in X
\label{equ:note}
\end{equation}
We re-implement the GI-DP to attain the released version data that satisfied the DP guarantees. We then train a series of Optimal-IMs to evaluate the effectiveness of target inference attacks on the released version data in a line to compare with our proposed mechanism.
\subsection{Metrics}
We set \textit{Euclidean}~\cite{ball1960short} and \textit{Manhattan} distance~\cite{black1998dictionary} as our evaluation metrics for the DRU to evaluate the quality of the reconstructed data $X'$ generated from extracted features $f$. Both distance metrics are widely used in location privacy literature~\cite{andres2013geo, erdemir_privacy-aware_2020}. For instance, the work introducing Geo-Indistinguishability~\cite{andres2013geo} utilizes a privacy level that depends on the \textit{Euclidean} distance. \textit{Euclidean} distance gives the shortest or minimum distance between two points. In contrast, \textit{Manhattan} distance applies only if the points are arranged in a grid, and both definitions are feasible for the problem we are working on. Note that these two distances have limited capability in showing the quality of the reconstructed data $X'$, however, they intuitively capture the differences between the original data $X$ and the reconstructed data $X'$.
For both MPU and URU, we leverage the top-n accuracy as our evaluation metric. The accuracy of the MPU is one of the most important factors in evaluating the utility of the extracted feature representation \textit{f}, where predictability of the \textit{f} increases as much as it can during the adversarial training. On the other hand, the competing training objective is to decrease the accuracy of the user re-identification unit to enhance the privacy of \textit{f}. The top-n metric computes the number of times the correct label appears among the predicted top $n$ labels. The top-n metric takes n predictions with higher probability into consideration, and it classifies the prediction as correct if one of them is an accurate label. The top-1 to top-5 accuracies are leveraged in our paper to discuss the performance of the proposed model.
\input{4_evaluation/table_compare}
\section{Architecture Evaluation}
\label{FraE}
In this section, we present the comparison results between the proposed Mo-PAE and three baseline models under the same training setting.
\subsection{Performance Comparison}
We first compare our proposed models with the Optimal-IMs, TrajGAN, and DP-GI on four representative mobility datasets, as shown in Table~\ref{tab:compare}. The overall performance is evaluated in terms of the \textit{utility level} provided by the MPU and the \textit{privacy protection} provided by DRU and URU. The \textit{Model I} is our proposed architecture but without applying the Lagrange multipliers(i.e., where each unit is weighted equally). The \textit{Model II} is the one with Lagrange multipliers (i.e., $\lambda_1$, $\lambda_2$, $\lambda_3$ ) and different weights are given to units (i.e., $\lambda_1=0.1,\ \lambda_2=0.8,\ \lambda_3=0.1$ for the results in Table~\ref{tab:compare}). In this table, the sequence length of the input traces is 10, that is $SL=10$. We will discuss why we choose SL=10 and the impact of the SL in Section \ref{sec:discussions}.
As we mention in Section~\ref{sec:baseline}, Optimal-IMs are trained without considering the privacy-utility trade-offs; hence, they can be leveraged to explain the optimal inference accuracy achieved. That is, before any privacy-preserving mechanism applies, the accuracy of the target or private tasks with raw data. For instance, the accuracy of the \emph{Privacy II} (0.9247 (MDC), 0.5643 (Priva'Mov), 0.6572 (GeoLife), 0.8780 (FourSquare)) demonstrates that an adversary can accurately infer user identity from raw data before any privacy protection.
Different from Optimal-IMs, the other models consider the privacy-utility trade-offs, and we measure the privacy protection and data utility by the effectiveness of the inference units. Therefore, the results in Table~\ref{tab:compare} are shown in \emph{utility loss} and \emph{privacy gain}, both of which are in a percentage format (\%), compared with the accuracy of Optimal-IMs. To compare the trade-off between them more intuitively, we list the "Utility-PII trade-off" column, where "trade-offs = Utility (\% for loss) + Privacy II (\% for gain)". The distance indexes (i.e., \emph{"Euc"} and \emph{"Man"}) are leveraged to intuitively represent the difference between the original data $X$ and reconstructed data $X'$, where a larger value indicates numerical differences between $X$ and $X'$. For the distance index, we are interested in the distance between each trace, hence, we consider the quantity of trace for datasets $X_D$ and get these distance indexes by averaging the corresponding record numbers $N_D$, that is (take \emph{"Euc"} for example):
\begin{equation}
Euc (X_D, X_D') = \frac{ \sqrt{\sum_{i=1}^{N_D} (x_i -x_i')^2}}{N_D}, \hspace{0.5cm} N_D = N_D'
\label{eq:euc}
\end{equation}
\begin{figure*}
\centering
\begin{subfigure}[t]{0.8\textwidth}
\centering
\includegraphics[width=\textwidth]{4_evaluation/legend_oneline.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{4_evaluation/mdc.pdf}
\caption{MDC}
\label{fig:mdc}
\end{subfigure}
\begin{subfigure}[t]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{4_evaluation/privamov.pdf}
\caption{Priva\'Mov}
\label{fig:privamov}
\end{subfigure}
\begin{subfigure}[t]{0.23\textwidth}
\centering
\includegraphics[width=\textwidth]{4_evaluation/geolife.pdf}
\caption{Geolife}
\label{fig:geolife}
\end{subfigure}
\begin{subfigure}[t]{0.23\textwidth}
\centering
\includegraphics[width=\textwidth]{4_evaluation/foursquare.pdf}
\caption{Foursquare}
\label{fig:foursquare}
\end{subfigure}
\caption{Pareto Frontier trade-off analysis on four datasets. The hollow squares and diamonds present the results of the proposed models Mo-PAE.
solid points present the results of the TrajGAN.
Blue color means \textit{SL} = 5. Black color means \textit{SL} = 10.}
\label{fig:compare}
\end{figure*}
\input{4_evaluation/table_euc}
Table~\ref{tab:compare} demonstrates that our proposed models, especially \emph{Model II}, outperform the TrajGAN and GI-DP across various datasets. For instance, with the MDC dataset, our \textit{Model II} achieves the best trade-offs when compared with other models, as the utility loss is only 13.43\% but with 65.51\% privacy gain, while 46.32\% utility loss and 20.32\% privacy gain with the TrajGAN, and 97.34\% utility loss and 97.47\% privacy gain with the GI-DP. The extreme performance on the GI-DP illustrates that while the DP paradigm is a robust privacy-preserving technique in protecting \emph{user's location}, it is not appropriate in protecting the \emph{user's identities}.
More intuitively, in the column of "trade-off", \textit{Model II} achieves all the best trade-offs among four datasets (52.08\% (MDC), 24.48\% (Priva'Mov), 34.36\% (GeoLife) and 48.54 (FourSquare)).
\emph{Model I} has worse performance than \emph{Model II}, in general, but is still superior to TrajGAN and DP-GI, where the latter two might even get \emph{negative} trade-offs (i.e., TrajGAN got -26.00\% with MDC and GI-DP got -5.71\% with Priva'Mov). Moreover, for the Priva'Mov dataset, although the utility loss of the TrajGAN is 4.21\% smaller than our \textit{Model II}, both two privacy metrics of the TrajGAN are worse than the \textit{Model II}. Again, our model has better overall trade-offs, as 23.66\% for \textit{Model I} and 24.48\% for \textit{Model II}. The performance on Geolife and FourSquare are similar but inverse, where the utility of our model is better than TrajGAN and with slightly weaker privacy preservation.
In summary, GI-DP always has the highest privacy gain among the four datasets, however, the utility loss is also very high, resulting in inadequate and unexpected privacy-utility trade-offs. This trend also shows that the DP mechanism is not an appropriate metric for the location privacy of \emph{user's identity}, which is also in line with the conclusions from other related work~\cite{shokri2011quantifying, erdemir_privacy-aware_2020}. The comparisons between \textit{Model I} and \textit{Model II} also illustrate the importance of the Lagrange multipliers, which provides flexibility to our proposed architecture, enabling its application in different scenarios and enhancing the privacy-utility trade-offs in this case.
\begin{figure*}[t]
\centering
\includegraphics[width=0.99\textwidth]{4_evaluation/tradeoffs.pdf}
\caption{Impact of Mo-PAE on the user re-identification accuracy (PII) and relative utility loss (\textit{U}) on four datasets.
The orange area represents the utility loss while the light-green area represents privacy gain. The dark-green area represents the trade-offs between utility achievement and privacy budgets. The x-axis shows five different model settings, and the y-axis shows the trade-offs.}
\label{fig:utility-privacy}
\end{figure*}
\subsection{Trade-off Comparison}
In this section, we present the privacy-utility trade-off analysis between the proposed Mo-PAE and TrajGAN in terms of mobility prediction accuracy (i.e., \textit{U}) and user de-identification efficiency (i.e., \textit{PII}). Figure~\ref{fig:compare} presents the trade-off comparisons of the four datasets, where the \textit{hollow squares} and \textit{hollow diamonds} show the trade-offs provided by the proposed Mo-PAE in $SL=5$ and $SL=10$, respectively. The \textit{solid points} present the results of the TrajGAN under the same experimental setting. As can be seen from these results, in all four cases, the synthetic dataset generated by the TrajGAN is not {\em Pareto-optimal}. That is, the proposed Mo-PAE is able to achieve a better privacy level for a dataset with the same utility value. Compared with the TrajGAN, Mo-PAE improves utility and privacy simultaneously on four datasets. Especially for the performance of MDC, the privacy improves 45.21\% more than the TrajGAN, while the utility also increases by 32.89\%. These results illustrate that our proposed model achieves promising performance in training a privacy-sensitive encoder $Enc_L$ for different datasets.
After evaluating the superior performance of our proposed model, we discuss the privacy guarantee that Mo-PAE provided in terms of data reconstruction(\textit{PI}, \textit{"Euc"} in Table~\ref{tab:euc}) and user re-identification(\textit{PII}, \textit{privacy gain} in Figure~\ref{fig:utility-privacy}). As we mentioned in the Section~\ref{RelW}, the privacy guarantee of Mo-PAE differs from that of DP paradigms and is given in the declined effectiveness of inference attacks.
\begin{figure*}[t]
\centering
\begin{subfigure}[b]{0.98\textwidth}
\centering
\includegraphics[width=\textwidth]{4_evaluation/tem_legend.pdf}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{4_evaluation/tem_mdc.pdf}
\caption{MDC}
\label{fig:tem_mdc}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{4_evaluation/tem_pri.pdf}
\caption{Priva\'Mov}
\label{fig:tem_privamov}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{4_evaluation/tem_geo.pdf}
\caption{Geolife}
\label{fig:tem_geolife}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{4_evaluation/tem_fs.pdf}
\caption{Foursquare}
\label{fig:tem_foursquare}
\end{subfigure}
\hfill
\caption{The effect of temporal granularity on the model performance of four mobility datasets.}
\label{fig:temporal}
\end{figure*}
\begin{figure*}
\centering
\hfill
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{4_evaluation/mdc_prediction.pdf}
\caption{$U_D$ Performance on MDC}
\label{fig:context_mdc_prediction}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{4_evaluation/mdc_recognition.pdf}
\caption{$P_D^2$ Performance on MDC}
\label{fig:context_mdc_recognition}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{4_evaluation/privamov_prediction.pdf}
\caption{$U_D$ Performance on Priva'Mov}
\label{fig:context_pri_prediction}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{4_evaluation/privamov_recognition.pdf}
\caption{$P_D^2$ Performance on Priva'Mov}
\label{fig:context_mdc_recognition2}
\end{subfigure}
\hfill
\caption{Mobility prediction accuracy and user re-identification accuracy change with the trace sequence length (SL) in our proposed $U_D$ and $P_D^2$. The color bars indicate the accuracy from top-1 to top-5, the black texts indicate the top-1 accuracy and the purple texts indicate the top-5 accuracy. For instance, the top-1 mobility prediction accuracy on MDC with \textit{SL} = 2 is 0.473, and the top-5 one is 0.802.}
\label{fig:context}
\end{figure*}
\subsection{Privacy Guarantee Analysis: Effectiveness of Privacy Inference attacks}
\label{UPA}
In this section, we discuss the impact of Mo-PAE on the effectiveness of two privacy inference attacks (i.e., \textit{PI} and \textit{PII}), respectively.
\subsubsection{Effectiveness of Data Reconstruction Attacks - PI}
Table~\ref{tab:euc} shows the impact of the proposed mechanisms on the data reconstruction accuracy(\emph{PI}). The \emph{"Euc"} in the table follows the definition in Eq.\ref{eq:euc}. Overall, \emph{Model II} performs better than \emph{Model I} in limiting the accuracy of data reconstruction regardless of the value of weights. Take the result of GeoLife dataset as an example, \emph{Model II-i} achieves bigger distance than \emph{Model I} (i.e., 0.4343 > 0.0057), while it still gets better utility (i.e., -9.94\% > -17.9\%). Nevertheless, both \emph{Model I} and \emph{Model II} have effectively defended the data reconstruction attack (MDC: 0.0697 > 0.0017 > 0.0000; Priva'Mov: 0.0453 > 0.0009 > 0.0003; GeoLife: 0.4343 > 0.0057 > 0.0008; FourSquare: 0.7933 > 0.0069 > 0.0052), while only at the marginal cost of mobility utility (MDC: -12.55\%; Priva'Mov: -2.71\%; GeoLife: -9.94\%; FourSquare: -1.64\%). The data of the Optimal-IMs are in Table \ref{tab:compare}. We list four representative settings here to make a comprehensive comparison of PII and U. From setting i to iv, one can expect more original data features loss to result in a more significant utility loss. This trend is indeed the case with different weights' combinations. However, as the results show, especially for setting i, the privacy of the traces attains decent protection at the marginal cost of mobility utility.
\subsubsection{Effectiveness of User Re-identification Attacks - PII}
Figure~\ref{fig:utility-privacy} presents the impact of the Mo-PAE (\emph{Model II}) on the user re-identification accuracy(\emph{PII}). In this figure, we list five different settings, I to V ($\lambda_1 = 0.1$, over the range of $\lambda_2$ = \{0.5, 0.6, 0.7, 0.8, 0.9\}), respectively.
The \textit{Zero} line (i.e., y = 0\%) in each sub-figure is leveraged to indicate the original accuracy of the raw data (i.e., Optimal-IMs). The \emph{"Privacy Gain Rate"} (blue square line) shows the effectiveness of the Mo-PAE in defending the user re-identification attacks. That is, after applying \emph{Model II}, the decline range of effectiveness of user re-identification attacks. For instance, with the MDC dataset, in setting I, the effectiveness of user re-identification attacks declines as high as 80\%. In the same time, this high privacy protection is at the cost of nearly 55\% of utility (orange triangle line). Things are better in the setting V, where the Mo-PAE can get 60\% privacy protection only at the cost of less than 10\% utility. The x-axis shows five settings of the model, and the y-axis shows the trade-offs (i.e., \textit{trade-offs} = \textit{privacy gain} + \textit{utility loss}). The orange area represents the utility loss while the light-green area represents the privacy gain when compared with Optimal-IMs. The dark-green area represents the trade-offs between utility and privacy budgets.
In summary, these trade-offs are all positive in different model settings on four different datasets.
The performance on the GeoLife data is the best, while less than 20\% utility loss but more than 50\% privacy gains. The performance on MDC and FourSquare also show the promising privacy-utility trade-offs, especially for setting \textit{V} on the FourSquare dataset, both the utility and privacy increase.
The uniqueness of human mobility trajectories is high, and these trajectories are likely to be re-identified even with a few location data points~\cite{de2013unique}. Our results emphasize that the concern of user re-identification risk could be alleviated effectively with our proposed model.
In real applications, the trade-off of Mo-PAE is achieved continuously over time. New trajectories will be encoded with the pre-trained encoder to attain respective feature representation and utilized by SP for following task-oriented scenarios (no need to retrain). The pre-trained encoder and discriminators are assumed to be updated offline within a fixed duration for best performance purposes.
Additionally, while the architecture focuses on specific application scenarios (i.e., mobility prediction), it could generally be applicable to different task-oriented scenarios.
\section{Discussions}
\label{sec:discussions}
In this section, we further discuss the impact of the temporal granularity of traces, the varying sequence length and weights on the composition units on the Mo-PAE performance.
\subsection{Impact of Temporal Granularity}
The timestamp is one of the essential components of the mobility trace, and different choices on the temporal granularity affect the final performance of any dataset. Figure~\ref{fig:temporal} shows the impact of the varying temporal granularity on the proposed architecture. We present the top-1, top-5, and top-10 accuracies for both utility and privacy dimensions. For instance, when temporal granularity is 10-min, it indicates a location record \textit{r} is taken every 10 minutes from the raw data. When using more coarser temporal granularity, the number of points of interest decreases so does the difficulty of mobility prediction. However, the uniqueness of the trajectory decreases due to ignoring many of the unique locations of each user, resulting in lower privacy. To summarize from Figure~\ref{fig:temporal}, the impact of temporal granularity on the Priva'Mov is minimal. In terms of utility (mobility prediction), Priva'Mov is the only dataset for which accuracy decreases with increasing temporal granularity. This subtle decline emphasizes the trajectory features only have a small change when varying granularity, in line with the university students' mobility.
\subsection{Impact of Varying Sequence Lengths}
\label{section:impact sl}
The performance of the utility discriminator $U_D$ (MPU) and the privacy discriminator $P_D^2$ (URU) exert a significant impact on the overall performance of the proposed Mo-PAE. The trace length is the most critical factor affecting these units' performance. We use two representative datasets (i.e., MDC and Priva'Mov) to present the impact of the varying sequence length on both discriminators.
As shown in Figure~\ref{fig:context}, by changing the lengths of trace sequence $SL$ from 1 ($SL=1$) to 50 ($SL=50$), we observe that $SL$ has a significant impact on different tasks' accuracy (i.e., mobility prediction accuracy for $U_D$ and user re-identification accuracy for $P_D^2$) of two different datasets. Overall, the impact in the MDC dataset is much higher than in the Priva'Mov dataset. Comparing the Figure~\ref{fig:context_mdc_prediction} and Figure~\ref{fig:context_pri_prediction}, there is a much sharper increase on the MDC dataset. More specifically, when the sequence length is increased from 2 to 20, the top-1 mobility prediction accuracy on MDC increases from 0.473 to 0.978 (i.e., +50.5\%), while accuracy on Priva'Mov increases from 0.918 to 0.959 (i.e., only +4.1\%). Similarly, more rapid growth appears in the user re-identification accuracy on MDC, which is +68.0\%, while the increase for Priva'Mov is only +30.8\%. We conclude that the mobility predictability and user re-identification accuracy of a dataset might have a special link. The mobility predictability of Priva'Mov is very high, almost higher than 90\%, but the user re-identification accuracy is always lower than 80\%, which also means the uniqueness of trajectories in this dataset is low. This low uniqueness suggests that the users in this dataset might share similar daily routes, which is reasonable, as we know these trajectories are collected from students at the same university. For the MDC dataset, when $SL=10$, the user re-identification accuracy is relatively high, indicating that the locations are more sparse in this dataset. However, the mobility predictability here is also high, which emphasizing that this sparseness does not affect the predictability. These phenomena indicate that the deep training of MPU and URU might share similar extracted features, while our proposed architecture attempts to extract features more suitable for mobility predictability but less suitable for user re-identification.
We note that the varying trace sequence length not only exerts impacts on the model performance but also has a significant influence on the computation time. For instance, the computation time at $SL=50$ costs six times as much as that at $SL=5$. The computation time also varies between datasets. Hence, an appropriate choice of trajectory sequence length can avoid time-consuming computation and achieve expected task inference accuracy. In our work, we place greater focus on the trace sequence lengths ranging from 5 to 10, which exhibits great performance in both the $U_D$ and $P_D$ while also keeping a low computation time.
\begin{figure}
\centering
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{4_evaluation/lamda3_zero_privacy.pdf}
\caption{$\lambda_1 = 1-\lambda_2,~\lambda_3 = 0$}
\label{fig:lambda3_zero_privacy}
\end{subfigure}
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{4_evaluation/lamda1_zero_privacy.pdf}
\caption{$\lambda_3 = 1-\lambda_2,~\lambda_1 = 0$}
\label{fig:lambda1_zero_privacy}
\end{subfigure}
\caption{Varying weights can tune the privacy-utility trade-offs. The primary y-axis (dashed line) represents utility, the secondary y-axis (solid line) represents privacy. The x-axis represents the value of the target $\lambda$. }
\label{fig:lagranian}
\end{figure}
\subsection{Impact of Varying Weights
As we discussed in Section~\ref{section: module overview}, the \textit{sum loss function} $L_{sum}$ of \emph{Model II} is a linear combination of $L_R$, $L_U$, and $L_P$ with different weights (i.e., Lagrange multipliers). We evaluate the influence of different weights' combinations ($\lambda_1$, $\lambda_2$, and $\lambda_3$) on the \emph{Model II}, as the results shown in Figure~\ref{fig:lagranian}.
We compare the overall model performance in $U_D$ and $P_D^1$ by fixing the $\lambda_3 = 0$, and vary the other two multipliers by subjecting to $\lambda_1 = 1- \lambda_2$, as shown in Figure~\ref{fig:lambda3_zero_privacy}. Figure~\ref{fig:lambda1_zero_privacy} illustrates the effect between $U_D$ and $P_D^2$ by setting the $\lambda_1 = 0$. We could observe in both settings that the utility increases with a larger $\lambda_2$, which means when the MPU is given more weight in the Mo-PAE model, it would exert a positive impact on the data utility. We conclude that the privacy-utility trade-offs could be tuned by varying these weights; the results in the Figure~\ref{fig:lagranian} also verify the effectiveness of our adversarial architecture. We note that the balance of three units is far more complicated than the balance of two. From the extensive experiment we conducted, initialing $\lambda_1=0.1$, $\lambda_2=0.6$, $\lambda_3=0.3$ can guide the model achieve the tradeoff most efficiently.
However, as the experiment results show, there is no dataset independent privacy interpretation for $\lambda_1$, $\lambda_2$ and $\lambda_3$, and we leave a more efficient approach using reinforcement learning to initializing these hyperparameters for different datasets in future work.
\section{Conclusion}
\label{Conl}
In this paper, we presented a privacy-preserving architecture Mo-PAE based on adversarial networks. Our model considers three different optimization objectives and searches for the optimum trade-off for utility and privacy of a given dataset. We reported an extensive analysis of our model performances and the impact of its hyperparameters using four real-world mobility datasets. The weights $\lambda_1,\ \lambda_2,$ and $\lambda_3$ bring more flexibility to our framework, enabling it to satisfy different scenarios' requirements according to the relative importance of utility requirements and privacy budgets. We evaluated our framework on four datasets and benchmarked our results against an LSTM-GAN approach and a DP mechanism. The comparisons indicate the superiority of the proposed framework and the efficiency of the proposed privacy-preserving feature extractor $Enc_L$. Expanding this work, we will consider other utility functions for our models, such as community detection based on unsupervised clustering methods or deep embedded clustering methods. In future work, we will leverage automated search techniques, such as deep deterministic policy gradient algorithm and reinforcement learning, for efficiency in searching for the optimal weight combinations.
\section{Introduction}
\input{1_introduction/1_introduction}
\input{2_related/related_new}
\input{3_proposedmodel/3_proposedmodel}
\input{4_evaluation/4_evaluation}
\input{5_conclusion/5_conclusion}
\section*{Acknowledgment}
We wish to acknowledge the constructive suggestions from PETS reviewers. This work was partially supported by EPSRC Open Plus Fellowship EP/W005271/1.
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
\input{intro.tex}
\section{Methods}
\input{methods.tex}
\section{Evaluation}
\input{results.tex}
\bibliographystyle{myrecomb}
\subsection{Background}
CoViT is based on ViT~\cite{dosovitskiy2020image}, a neural network primarily developed for image classification. We develop and apply a preprocessing step that employs MinHash~\cite{broder2000min}, whose role is extracting the informative feature vectors from the genome, which are further fed into a modified ViT for classification. In this section we provide a detailed introduction to the ViT and MinHash.
\subsubsection{Transformers}
\label{transformers}
Transformer is a deep neural network model proposed by Vaswani et al.~\cite{vaswani2017attention}. It became state of the art in many deep learning applications, such as Natural Language Processing. More recently, the ViT which is an image classification deep neural network based on the transformer model, was introduced in ~\cite{dosovitskiy2020image}. ViT architecture is presented in Figure~\ref{fig:TIC}. It comprises an encoder and a multilayer perceptron (MLP) layer with a softmax activation function to perform predictions (in this paper we will refer to this MLP layer as MLP head). An encoder receives a sequence of representations, each of them is of dimensionality $d_{model}$, and outputs a sequence of learnt representations of the same dimensionality. For the sake of clarity we refer to the sequence of input representations as ``feature vectors sequence'' throughout this section. The transformer encoder layer is composed of two main sub-layers, namely the Multi Head Self Attention (MHSA) and the piece-wise MLP. In the following subsections we will present the structure and the purpose of those sub-layers.
\begin{figure*}[!t]
\begin{center}
\includegraphics[scale=0.3]{figs/TransformersInCoViT2.png}
\end{center}
\caption{The Transformer encoder architecture is built as a stack of alternating $L$ multi-head self-attention (MHSA) layers and $L$ MLP blocks. Layer normalization is applied before every block and a residual connection is employed around each of the sub-layers (i.e., MHSA and MLP)}
\label{fig:TIC}
\end{figure*}
\textbf{Multi Head Self Attention:}
This sub-layer, depicted in Figure~\ref{fig:TIC} (right) receives as an input a sequence of feature vectors $f_1, f_2,\dots, f_n$, each of dimensionality $d_{model}$. The sub-layer outputs a sequence of learnt representations $r_1, r_2, \dots, r_n$, a representation per each feature vector. The output learnt representations are of the same dimensionality as the input ones.
The following presents the MHSA data flow. First, each feature vector $f_j$ is passed through $3h$ linear layers $\{q_l, k_l, v_l\}_{l=1}^{h}$, where $h$ denotes the number of heads, and 3 denotes the three types of different representations (i.e., query, key, and value). The purpose of those linear layers is to learn $h$ query representations for $f_j$, $h$ key representations for $f_j$ and $h$ value representations of $f_j$. Each linear layer defines a weight matrix as follows:
$W_l^q \in \mathbb{R}^{d_{model}\times d_{q}}$;\textbf{ } $ W_l^k \in \mathbb{R}^{d_{model}\times d_{k}}$;\textbf{ } $W_l^v \in \mathbb{R}^{d_{model}\times d_{v}}$.
They are used to calculate their respective representations:
$q_l(f_j) = f_j^T\cdot W_l^q$;\textbf{ } $k_l(f_j) = f_j^T\cdot W_l^k$;\textbf{ } $v_l(f_j) = f_j^T\cdot W_l^v$.
The query and key representations are used to calculate the degree of attention that the learnt output representation, $r_j$, will give to other feature vectors. The value representations are used to represent learnt properties of the feature vector that we intend to pass through the network. Formally, this first step of the MHSA layer is defined as follows:
\begin{align*}
& A^l_j = \sum_{i=1}^{n}{similarity(q_l(f_j),k_l(f_i))}\cdot v_l(f_i)\\
& similarity(q_l(f_j),k_l(f_i)) =
softmax_i(\frac{q_l(f_j)^T k_l(f_1)}{\sqrt{d_k}}, \dots, \frac{q_l(f_j)^T k_l(f_1)}{\sqrt{d_k}})
\end{align*}
where $l$ refers to the $l$'th head, and $softmax_i$ is the $i$'th element of the softmax function, defined as
$softmax(x_1, x_2, \dots, x_n)=\frac{1}{\sum_{i=1}^{n}{e^{x_i}}}(e^{x_1}, e^{x_2},\dots, e^{x_n}).
$
Note that for a given head $l$, feature vectors with higher similarity to $f_j$ will have a larger dot product with $q_l(f_j)$ compared to other feature vectors. This drives the network to learn, for each representation, $r_j$, on which feature vectors of the input sequence it should focus its attention to learn the representations of value. To recap, representations $A_j^1, A_j^2 \dots, A_j^h$ depend on feature vectors that receive a high degree of attention.
Next, these representations are concatenated to a form $A_j = A_j^1 \circ A_j^2 \circ \dots \circ A_j^h$.
The last step of the MHSA is applying to $A_j$ a linear layer, to generate the MHSA output representation.
\textbf{Trainable weights and hyper-parameters:}
For each head, $l\in [1, h]$, we define three weight matrices $W_l^q, W_l^k, W_l^v$.
Additional weight matrix $W^o\in \mathbb{R}^{h\cdot d_v\times d_{model}}$ is reserved for the linear layer that processes $A_j$.
The MHSA hyper-parameters include the number of heads $h$, the dimensionality of the query and key representation $d_q=d_k$, and the dimensionality of the value representation $d_v$.
\textbf{Piece-wise Multi Layer Perceptron:}
The piece-wise MLP receives as an input a sequence of representations (feature vectors). For each feature vector it learns an output representation. Unlike the MHSA, in the piece-wise MLP layer each learnt representation depends only on its input feature vector and does not depend on other feature vectors in the sequence. This layer transforms each input feature vector into an intermediate representation of dimensionality $d_{ff}$, and learns the output representation which is of dimensionality $d_{model}$. Formally, to perform this calculation, we define two weight matrices
$
W^{ff}\in \mathbb{R}^{d_{model}\times d_{ff}}$ and \textbf{ } $W^m\in \mathbb{R}^{d_{ff}\times d_{model}}$.
and calculate the representation of $f$ as follows:
\begin{align*}
MLP(f) = (ReLU(f^T\cdot W^{ff}))^T\cdot W^m
\end{align*}
\textbf{Trainable weights and hyper-parameters:}
There are two weight matrices $W^{ff}$ and $W^m$ as defined above. The only hyper-parameter is $d_{ff}$.
\textbf{Layer Normalization:}
Layer normalization~\cite{ba2016layer} is a method used to normalize the activities of the neurons of a layer improving the training speed for neural network models. It directly estimates the normalization statistics from the summed inputs to the neurons within a hidden layer.
\textbf{The Transformer Encoder:}
The Transformer Encoder (TE) comprises a stack of L identical encoder layers, each comprising an MHSA sub-layer followed by a piece-wise MLP sub-layer. Layer normalization is applied before each sub-layer, and a residual connection is employed around each of the two sub-layers.
\textbf{Hyper-parameters of the TE:}
The number of encoder layers - $L$, the dimensionality of the query and key representations - $d_q = q_k$, the dimensionality of the value representation - $d_v$, and the dimensionality of the intermediate representation of the piece-wise MLP - $d_{ff}$.
\subsubsection{The MinHash Scheme}
The min-wise independent permutations (MinHash)~\cite{broder2000min} is a technique for similarity estimation.
MinHash was applied in many tasks in computational biology, including genome assembly~\cite{berlin2015assembling, shafin2020nanopore},
metagenomic gene clustering~\cite{rasheed2013map, muller2017metacache} and genomic distance estimation~\cite{ondov2016mash}. MinHash implements the following sketch function:
Given a set of characters $A$, a compression factor $n$, and a hash function $h$, the elements of the set $A$ are hashed using function $h$ to generate the set $H(A)$. Then the elements of $H(A)$ are sorted, and the smallest $n$ elements are returned as described in Algorithm~\ref{alg:sketch}.
\begin{algorithm}
\caption{The sketch algorithm $sketch(A)$}\label{alg:sketch}
\begin{algorithmic}
\Require set $A = \{a_1,\dots,a_{|A|}\}$, compression parameter $n$ and a hash function $h$
\State $sketch \gets \emptyset$
\State $H(A) \gets (h(a_1), \dots h(a_{|A|}))$
\State Sort H(A) to get $(h(a_{i_1}), h(a_{i_2}), \dots, h(a_{i_{|A|}}))$\\
\Return $(a_{i_1}, a_{i_2}, \dots, a_{i_{n}})$\footnotemark
\end{algorithmic}
\end{algorithm}
\footnotetext{These are the genome representative kmers referred to in Section 2.2.1}
The subset obtained by applying the sketch algorithm provides a good estimate for the Jaccard index~\cite{hancock2004jaccard}, defined as follows: Given two sets $A, B$ the Jaccard index is
$
J(A, B) = \frac{|A\cap B|}{|A\cup B|}
$.
Formally, given two sets, A and B, $\frac{|sketch(A) \cap sketch(B)|}{|sketch(A) \cup sketch(B)|} \approx J(A, B)$.
We use the sketch algorithm to extract feature vectors from a genome, as presented further.
\subsection{CoViT Architecture}
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.48]{figs/archAndFEE.png}
\end{center}
\caption{(a): A general overview of the CoViT Architecture.
(b): Feature Extractor workflow example.
}
\label{fig:covitarch}
\label{fig:fee}
\end{figure}
The top-level architecture of CoViT is shown in Figure~\ref{fig:covitarch}(a). The input for CoViT is an assembled genome in FASTA file format. To convert the genome into a sequence of feature vectors, we build a pipeline of preprocessing steps, comprising the feature extractor and the embedding layer. The feature extractor chooses a set of fragments (i.e., genome subsequences of fixed length), represents each fragment as a two-dimensional matrix and outputs those genome fragment matrices. Each genome fragment matrix is fed into an embedding layer consisting one neuron. This layer converts the genome fragment matrix into a genome fragment vector, designated the feature vector. The feature vectors are fed into the ViT which outputs the most likely placement candidates (by attaching probabilities to each lineage).
The feature extractor together with the embedding layer transform a genome into a sequence of representative feature vectors, which are the numerical representations of genome fragments. We implement the feature extractor using MinHash to preserve similarities in genome. It allows the feature extractor to find features in the query genome that are shared by the known lineages. The embedding layer is used as a dimensionality reduction step. It allows to reduce CoViT latency without affecting its placement accuracy.
\subsubsection{Feature extractor}
The first part of the CoViT pipeline receives the assembled genome and outputs matrices that represent genome fragments. Feature extractor employs MinHash to find similar fragments (i.e., features) in different genomes.
\begin{algorithm}
\caption{the Feature Extractor}\label{alg:featureExtractor}
\begin{algorithmic}
\Require genome $g$, compression parameter $n$, fragment size $f$, k-mer size $k$, hash function $h$
\State $G \gets \{ g_i \equiv g[i: i+k]\} \text{ } s.t.\text{ } i\in [0, n-k+1]$
\State $kmers \gets sketch(G, n, h)$
\State $left \gets floor(\frac{f-k}{2})$
\State $right \gets ceil(\frac{f-k}{2})$
\State $fragments \gets \{g[i-left, i+right] : g_i \in kmers\}$\\
\Return $one-hot-encode(fragments)$
\end{algorithmic}
{\footnotesize One hot encoding of a fragment refers to the operation of transforming each base in the fragment as follows:\newline $A \rightarrow [1, 0, 0, 0]$; \textbf{ }
$C \rightarrow [0, 1, 0, 0]$; \textbf{ }
$G \rightarrow [0, 0, 1, 0]$; \textbf{ }
$T \rightarrow [0, 0, 0, 1]$; \textbf{ }
$N \rightarrow [0, 0, 0, 0]$.\newline For the example please refer to Figure~\ref{fig:ele}(a).}
\end{algorithm}
An example is illustrated in Figure~\ref{fig:fee}(b). Input is a genome sequence $g$ of length $N=19$. We generate all possible kmers $G$ (where $k=4$), and extract ($n=3$) 4-mers from the genome using the MinHash scheme (i.e., $sketch(G,n=3,h)$). The next step is to generate fragments of length $f=8$ (by extending each of three 4-mers by $\frac{f-k}{2}=\frac{8-4}{2}=2$ bases in each direction). The last step of this workflow is to encode each basepair of the fragments using one-hot encoding.
For the genome sequence $g$ of length $N$, the extractor first generates a set $G$ of all possible kmers (genome sub-sequences of length $k$). It then sketches (i.e., applies MinHash \emph{sketch} function to) the set of kmers, $G$, to extract $n$ representative kmers of the genome (Algorithm \ref{alg:sketch}). Those kmers are used as anchors to be expanded to generate fragments. Last part is transforming a fragment into a numerical matrix where the genome basepairs A, G, C and T are encoded using one-hot encoding as described in Algorithm~\ref{alg:featureExtractor}.
\subsubsection{Embedding layer}
The input of the embedding layer is a genome fragment matrix of dimensions $f\times4$ (i.e., $f$ represents the fragment length and 4 is the dimensionality of the basepair one-hot encoding). It outputs a feature vector of size $f$. The embedding layer consists of one neuron which performs a base-wise linear transformation. Specifically, the embedding layer defines 5 learnable parameters, 4 weights $w_A, w_C, w_G, w_T$ and one bias term $b_N$. Given a genome fragment matrix, we transform each base (represented in one-hot encoding) to a numerical token as presented in Figure~\ref{fig:ele}(a), which also contains an example in which we embed a fragment matrix of dimensions $8\times 4$.
\begin{figure}
\begin{center}
\includegraphics[scale=0.4]{figs/histAndELE.png}
\end{center}
\caption{(a): Embedding a fragment matrix of dimensions $8\times4$. (b): The number of sequenced genomes per lineage histogram.
Most of the lineages (more than 800 out of 1,536 available) have less than 64 accessions available. The peak in the last bin accounts for all lineages with more than 960 accessions.
}
\label{fig:ele}
\label{fig:hist}
\end{figure}
\subsection{Setup}
In this section we describe the dataset, evaluation criteria, training process and the machine on which we train the model.
\subsubsection{Dataset}
We use the COVID-19 Data Portal~\cite{covid19data} as a source of sequenced SARS-CoV-2 genomes. There are 4,470,553 assembled genomes classified into 1,536 distinct lineages. The distribution of genomes across the lineages is highly non-uniform. For example, 463,426 samples are categorized as B.1.1.7, while the majority of the lineages has only a small number of samples, as depicted in Figure~\ref{fig:hist}(b). Such non-uniformity makes training the network challenging. To mitigate this, we select at most 1024 samples from each lineage and discard the rest when constructing the dataset. This way we prevent any potential over-fitting towards lineages like B.1.1.7. We also discard all samples that belong to lineages with fewer than 512 accessions, since such a small number of samples limits the model's ability to generalize. These steps form our dataset, which contains 240,734 samples belonging to 270 different lineage classes.
We divide our dataset into three non overlapping subsets: the training set, the validation set, and the test set. The training and validation sets are used to train and optimize the network. The test set is applied to evaluate CoViT's placement accuracy and to compare it with state-of-the-art solutions. Both the validation and test sets consist of 13,500 genomes randomly selected from the dataset. The rest of the samples comprise the training set.
\subsubsection{Classification accuracy and placement rate criteria}
We use the \emph{top-n accuracy} as CoViT placement accuracy criterion. It estimates the likelihood of the \emph{correct} placement (i.e., the lineage to which the query genome belongs to\footnote{Since the COVID-19 Data Portal assigns all sequenced genomes to their respective lineages, the correct results are known.}) appearing among the \emph{n} most probable results (i.e., \emph{n} lineages output by the network at the top of the descending-order probability list). Specifically, we measure the top-1, top-2 and top-5 accuracy of our model as a criteria of its performance. Such criteria make the comparison to the state-of-the-art straightforward.
By \emph{base ambiguity} we mean the ambiguous bases (i.e. R, Y, K, N etc.) in assembled SARS-CoV-2 genomes. Base ambiguity is a result of incomplete coverage. N marks the worst ambiguity because it can mean any base (A, G, C, or T). We further evaluate the accuracy of CoViT placement under different levels of base ambiguity. In our study, we conservatively use only the worst case (i.e., N) ambiguity.
Another criteria we apply in our analysis is the \emph{placement rate}, a fraction of accessions a placement algorithm is able to place (correctly or incorrectly) on the existing phylogenetic tree. The reasons for the placement rate to be lower than 100\% are base ambiguity and sequencing errors.
\subsubsection{Network Training}
We train CoViT using Adam~\cite{kingma2014adam} which is an optimized gradient descent technique, with $\beta_1 = 0.9, \beta_2=0.999$, batch size of 1024, and weight decay of 0.0001. Through various experiments, we found it most beneficial to choose the following hyper-parameters: $encoder repeats=4$, $d_{model}=256$, $d_{v} = 96$, $d_{k}= 96$, $d_{ff}=1536$, $h=18$. We also applied a dropout of rate 0.2 after each sub-layer.
The model was trained to minimize the categorical cross-entropy loss function. After fine-tuning and choosing those hyper-parameters, we train the model until the validation loss saturates.
\subsubsection{Hardware}
\label{HW}
We train CoViT on a desktop computer with Intel i7-9700K CPU with 8 cores, operating at 3.60GHz with 32GB of 2,666 MT/s DDR4 RAM, and NVIDIA's GeForce RTX 2080 Ti GPU with an 11GB / 14Gbps GDDR6 frame buffer, running at 1.545GHz.
\subsection{Results}
\subsubsection{Accuracy}
We measure CoViT performance using top-n accuracy.
We compare CoViT to the state-of-the-art phylogenetic placement tool, UShER~\cite{turakhia2021ultrafast}, which uses a maximum parsimony approach where it searches for a placement that requires the fewest additional mutations. UShER achieves 92.1\% top-1 accuracy over the test set, while CoViT presents a 2.1\% increment over UShER in top-1 accuracy performance, achieving 94.2\%. Moreover, CoViT achieves a top-2 accuracy of 97.9\% and top-5 accuracy of 99.8\% as summarized in Table~\ref{Tab:t1}.
\begin{table}[hbt!]
\caption{Accuracy comparison
\label{Tab:t1}}
\begin{center}
{
\begin{tabular}{||c|c|c|c||}
\hline
Program & top-1 accuracy & top-2 accuracy & top-5 accuracy\\ [0.5ex]
\hline\hline
CoViT & 94.2\% & 97.9\% & 99.8\%\\ [1ex]
\hline
UShER & 92.1\% & - & -\\ [1ex]
\hline
\end{tabular}
}
\end{center}
{\footnotesize
UShER uses another accuracy criterion called \emph{sister node placement}. They report the correct sister node placement in 97.2\% of cases. While this criterion resembles the top-2 accuracy, they are not identical, therefore we do not report it as a top-2 accuracy for UShER.}
\end{table}
\subsubsection{Run time and memory usage}
\label{efficiency}
We measure CoViT run time and memory usage and compare it to SARS-CoV-2 state-of-the-art placement tools. We compare CoViT and UShER, classifying the genomes in the test set, running on the machine specified in~\ref{HW} (see Table~\ref{Tab:t2}). We also present in Table~\ref{Tab:t2} the comparison to IQ-TREE multicore~\cite{nguyen2015iq}, and EPA-ng~\cite{barbera2019epa} running on a server with 160 processors (Intel Xeon CPU E7-8870 v.4, 2.10 gigahertz), each with 20 CPU cores. Those results where obtained while adding just one sequence to the SARS-CoV-2 global phylogeny containing 38,342 leaves~\cite{turakhia2021ultrafast}.
\newcommand{\mc}[2]{\multicolumn{#1}{c}{#2}}
\definecolor{Gray}{gray}{0.85}
\definecolor{LightCyan}{rgb}{0.88,1,1}
\begin{table}[hbt!]
\caption{Run-time and memory requirements comparison}
\label{Tab:t2}
\begin{center}
{
\begin{tabular}{|c|c|c|c|}
\hline
Program & Avg run-time per sample & Peak memory used (GB) & speedup over UShER\\ [0.5ex]
\hline
\hline
\rowcolor{LightCyan}CoViT (GPU) & \textbf{0.055s} & 5.02 & \textbf{3.53}\\
\hline
\rowcolor{LightCyan}CoViT (CPU) & 0.094s & 3.74 & 2.06\\
\hline
\rowcolor{LightCyan}UShER v0.5.6 & 0.195s & \textbf{0.49} & 1\\
\hline
\hline
\rowcolor{yellow}IQ-TREE multicore v2.1.1 & 2791s & 12.85 & $7\cdot10^{-6}$\\
\hline
\rowcolor{yellow}EPA-ng v0.3.8 & 1658s & 791.82 & $6\cdot10^{-5}$\\
\hline
\end{tabular}
}
\end{center}
{\footnotesize The results marked in light cyan are given running of the machine specified in section~\ref{HW}. The results marked in yellow are adopted from~\cite{turakhia2021ultrafast}.}
\end{table}
\subsubsection{Robustness to base ambiguity}
Some low quality SARS-CoV-2 genomes may contain ambiguous bases, which is a result of incompletely covered genome or sequencing errors. It is highly desirable that such genomes are correctly placed even when the level of ambiguity is significant. To test the robustness of CoViT to base ambiguity, we mask random sites (i.e., replace basepairs by N) in each SARS-CoV-2 genome of the test set, with the percentage of ambiguous bases ranging from 0 to 32\%. We then compare CoViT accuracy and placement rate on this \emph{ambiguous} test set (13,500 accessions with randomly masked sites) against the placement rate and accuracy of UShER. Figure~\ref{fig:ambig} presents the results.
At 32\% ambiguity, UShER's placement rate is 0.5\%, which means UShER places only 67 accessions out of 13,500. For the rest of accessions, UShER returns "failed to map" message. For those accessions UShER is able to map, it reaches the top-1 accuracy of 60.5\%.
CoViT maintains a constant 100\% placement rate. At 32\% ambiguity, CoViT obtains the top-1 accuracy of 82.9\%.
CoViT maintains 100\% placement rate up to 50\% ambiguity (not shown in the figure), at which point CoViT's top-1 accuracy stands at 77.3\%. CoViT exhibits such resilience to base ambiguity because neural networks typically maintain high resilience when processing low quality (high error or ambiguity rates) data.
\begin{figure}[hbt!]
\begin{center}
\includegraphics[scale=0.32]{figs/ambig.png}
\end{center}
\caption{Placement rate (left) and accuracy (right) as a function of the genome ambiguity. i.e., the fraction of the ambiguous bases in the input genome.
}
\label{fig:ambig}
\end{figure}
\section{Conclusion}
The massive amount of sequenced SARS-CoV-2 DNA material paves the way for real-time surveillance of the virus evolution and spread. At the same time, it puts a tremendous strain on phylogenetic analysis infrastructures. To prevent losing valuable genomic data and reaching incorrect conclusions, a new computing paradigm for phylogenetic analysis of viral pandemics is highly required.
CoViT is inspired by an enormous success of neural networks in implementing classification tasks.
CoViT extracts sets of features from a SARS-CoV-2 accession genome and uses Vision Transformer, an attention based neural network model, to accurately classify and place it in real-time. CoViT not only outperforms state-of-the-art tools both in accuracy and run-time, by placing a SARS-CoV-2 genome on a phylogenetic tree in 0.055s with 94.2\% accuracy. It also shows a strong resilience to low quality genome data (i.e. high ambiguity of the newly sequenced accessions). CoViT is able to place SARS-CoV-2 genomes with 50\% ambiguity (when N replaces A, G, C, and T in 50\% of the sites in each genome), while maintaining 77.3\% accuracy.
CoViT lifts the computational barrier to pandemic surveillance by improving the placement accuracy. It exhibits highly robust performance under low quality and ambiguous genomic data while greatly decreasing the processing time. Thus it enables real-time evolutionary tracking of COVID-19 and future viral pandemics.
\subsubsection{Data set}
We use the COVID-19 Data Portal~\cite{covid19data} as a source of the sequenced SARS-CoV-2 genomes.
We sort the SARS-CoV-2 lineages based on the number of sequenced genomes (accessions) in each lineage. Neural network training requires a training set of a sufficient size to provide an accurate classification result. However, only a limited number of lineages have sufficient amount of sequenced genomes (accessions), thus limiting the number of lineages CoViT is able to accurately classify a newly sequenced SARS-CoV-2 genome to. The number of lineages with at least 256 accessions reaches only 375 out of 1536 recorded lineages. Most of the lineages has less than 50 accessions available as depicted in Figure~\ref{fig:hist}.
\subsubsection{Classification groups}
We create four groups of lineages based on the number of accessions in each, as follows:
\begin{itemize}
\item CG1: 107 lineages, with at least 2048 sequenced genomes in each. To build the dataset for this group we take from each lineage exactly 2048 samples (i.e., sequenced genomes).
\item CG2: 189 lineages, with at least 1024 sequenced genomes in each. To build the dataset for this group we take from each lineage exactly 1024 samples.
\item CG3: 269 lineages, with at least 512 sequenced genomes in each. To build the dataset for this group we take from each lineage at least 512 samples, and 1024 samples at most.
\item CG4: 375 lineages, with at least 256 sequence genomes in each. To build the dataset for this group we take from each lineage at least 256 samples, and 1024 samples at most.
\end{itemize}
The purpose of this division is to investigate how much the size of the training set affects the accuracy of the placement.
\subsubsection{Classification accuracy criterion}
CoViT classifies a newly sequenced (query) genome by sorting the known lineages in the descending order of their similarity to it. Therefore, one or two lineages of the highest similarity to the query genome are of the most interest.
We use the top-n accuracy as CoViT placement accuracy criterion. It estimates the likelihood of the \emph{correct} result (i.e. the lineage that fits the query genome) appearing among the \emph{n} highest-probability results (i.e. \emph{n} lineages placed by the network at the top of the descending-order similarity list). Specifically, we measure the top-1, top-2 and top-5 accuracy for every classification group.
Since the Covid-19 Data Portal assigns all sequenced genomes to their respective lineages, we know during the network training which results are correct, and therefore are able to tag them.
\subsubsection{Network training}
\label{Network training}
We train CoViT for each classification group, using the following training and validation datasets:
\begin{itemize}
\item 80\% of the sequenced genomes are applied as the training set
\item Remaining 20\% are employed as the validation set
\end{itemize}
We train CoViT on NVIDIA's GeForce RTX 2080 Ti GPU with an 11GB / 14Gbps GDDR6 frame buffer, running at 1.545GHz. CoViT was implemented using the TensorFlow framework in Python. To train the network, we use two main techniques, the Greedy Layer-Wise Pretraining~\cite{bengio2006greedy} and Transfer Learning~\cite{torrey2010transfer}.
\paragraph{Greedy Layer-Wise Pretraining}
Training deep neural networks is challenging due to the vanishing gradient problem (i.e. the weights in layers close to the input layer are not updated in response to errors calculated on the training dataset). The Greedy Layer-Wise Pretraining allows deep neural networks to be successfully trained, achieving state-of-the-art performance. Pretraining involves adding a new hidden layer to a model and refitting, allowing the new model to learn the inputs from the existing hidden layers, while keeping the weights for the existing hidden layers fixed and not trainable. This technique is based on the observation that training a shallow network is easier than training a deep one~\cite{bengio2006greedy}.
\paragraph{Transfer Learning}
A technique that focuses on storing knowledge gained while solving one problem, and applying it to a different but related problem~\cite{west2007spring}. In transfer learning, we first train a base network on a base dataset and task, and then we re-purpose the learned features, or transfer them, to a second target network to be trained on a target dataset and task. This works well if the features are general, meaning suitable to both base and target tasks, rather than being specific to the base task~\cite{yosinski2014transferable}.
\paragraph{Regularization}
To generalize the results on the training set better, we use several regularization techniques in parallel. We employ a dropout after each MSA and MLP sub-layers in the network. Moreover we use L2 regularization~\cite{krogh1991simple} on the trainable weights of the network (not including the biases).
\subsection{Evaluation results}
\subsubsection{Classification accuracy}
We started by building a CoViT model for the CG3 classification group. To develop deeper networks and achieve best results we chose the following hyper-parameters
\begin{center}
\begin{tabular}{||c|c|c|c|c|c|c||}
\hline
$d_{model}$ & $d_{v}$ & $d_{k}$ & $d_{ff}$ & $heads$ & $dropout\_rate$ & $regularization\_coef$ \\ [0.5ex]
\hline\hline
256 & 96 & 96 & 1536 & 18 & 0.1 & 0.0001 \\ [1ex]
\hline
\end{tabular}
\end{center}
Also, we chose the $kmer$ length to be $16$, the fragment length to be $256$ (i.e., equal to the $d_{model}$) and the comperession factor, $n$, is set at $256$. These parameters are applied to all trained models.
We started training a model for the CG3 classification group using Supervised Greedy Layer-Wise Pretraining~\cite{bengio2006greedy}. We started the pretraining process by building a base model consisting of two TE layers. We trained the model for 60 epochs with a batch size of 256 \emph{examples} (i.e., assembled genomes). The average training time per epoch was 8.58[min], achieving the top-1 accuracy of 79.3\% on the validation set. Then we added two new TE layers to the base model, generating a 4-layer TE model. This 4-layer model was trained for 120 epochs, with the batch size of 256 examples. The average training time per epoch was 16.84[min]. We achieved the top-1 accuracy of 92.7\%, top-2 accuracy of 97.2\% and top-5 accuracy of 99.1\% on the validation set. The accuracy as a function of the number of epochs is depicted in Figure~\ref{fig:accs}(c).
To build the CoViT model for the CG1, CG2 and CG4 classification groups we used transfer learning~\cite{torrey2010transfer}. More specifically, we took the 4-layer learnt model for the CG3 as the base for the CG2 and the CG4 models, where the only component to change was the MLP head (presented in section~\ref{transformers}). We then trained each model for a number of epochs specified in Table~\ref{Tab:t1}, with a batch size of 256 examples.
The accuracy as a function of the number of epochs for these remaining classification groups is presented in Figure~\ref{fig:accs}(a), (b) and (d).
Following are the validation set results. The model for CG1 classification group achieves the highest top-1 accuracy of 97.1\%, top-2 accuracy of 99.0\% and top-5 accuracy of 99.8\%. Respectively, the model for CG2 classification group achieves the highest top-1 accuracy of 95.8\%, top-2 accuracy of 98.1\% and top-5 accuracy of 99.8\%.
The model for CG3 classification group achieves the highest top-1 accuracy of 92.7\%, top-2 accuracy of 97.2\% and top-5 accuracy of 99.1\%. The model for CG4 classification group achieves the highest top-1 accuracy of 92.1\%, top-2 accuracy of 96.9\% and top-5 accuracy of 99.1\%.
These accuracy results and additional training information are summarized in Table~\ref{Tab:t1}.
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.43]{figs/Accuracy_fig6.png}
\end{center}
\caption{Accuracy vs. number of epochs for different models and four classification groups: CG1 (a), CG2 (b), CG3 (c) and CG4 (d).
}
\label{fig:accs}
\end{figure}
\begin{table}[ht]
\caption{Accuracy results and training information of the different models on the different classification groups\label{Tab:t1}}
\begin{center}
\begin{tabular}{||c|c|c|c|c|c|c||}
\hline
classification group & transferred from & epochs & avg $\frac{train\_time}{epoch}$ & top-1 acc & top-2 acc & top-5 acc\\ [0.5ex]
\hline\hline
CG1 & CG2 & 15 & 12.89min & 97.1\% & 99.0\% & 99.8\% \\
\hline
CG2 & CG3 & 40 & 13.09min & 95.8\% & 98.1\% & 99.8\% \\
\hline
CG3 & - & 60, 120 & 8.58min, 16.84min & 92.7\% & 97.2\% & 99.1\% \\
\hline
CG4 & CG3 & 155 & 21.52min & 92.1\% & 96.9\% & 99.1\% \\ [1ex]
\hline
\end{tabular}
\end{center}
\end{table}
\subsubsection{Training and inference times}
The following Table~\ref{Tab:t2} summarizes the training and inference times for each classification group. Inference time for the CG1 is 1.45[ms] which is almost three orders of magnitude faster than probably the fastest state of the art placement tool UShER~\cite{turakhia2021ultrafast}.
\begin{table}[ht]
\caption{Training and inference time of the various models.\label{Tab:t2}}
\begin{center}
\begin{tabular}{||c|c|c||}
\hline
classification group & train time per example & inference time per example\\ [0.5ex]
\hline\hline
CG1 & 4.59ms & 1.45ms\\
\hline
CG2 & 4.67ms & 1.47ms\\
\hline
CG3 & 4.70ms & 1.49ms\\
\hline
CG4 & 4.71ms & 1.49ms\\ [1ex]
\hline
\end{tabular}
\end{center}
\end{table}
\subsubsection{Comparison with state-of-the-art}
Several methods have been developed for the purpose of placing genome sequence samples onto an existing reference phylogeny tree~\cite{minh2020iq, barbera2019epa, loytynoja2012accurate, ruan2007treefam}. However, these methods are far too slow to enable real-time genomic placement.
To address the speed of interpretation of viral transmission and evolutionary dynamics, ultrafast bootstrapping~\cite{hoang2018ufboot2, minh2013ultrafast} have been recently developed. While those methods provide significant speedup, they are not applicable to the problem of placing individual samples onto a reference phylogeny.
Below we address several state of the art classification and placement tools, including Kraken~\cite{wood2019improved}, PACIFIC~\cite{acera2021pacific}, UShER~\cite{turakhia2021ultrafast} and PangoLearn~\cite{o2021assignment}.
Kraken is one of the most popular genome classifiers. It uses exact matching of kmers of the genome(s) in question against a known organisms' genomes database. Kraken was not specifically designed to classify SARS-CoV-2 (or another viral pathogen with pandemic potential) lineages and therefore is not sensitive enough to place SARS-CoV-2 variants which differ from one another by only several basepairs.
PACIFIC is arguably the first deep classification neural network based genome classifier. PACIFIC has been shown to classify genomes into a very limited group of classes (5 distinct virus classes and a human genome class).
PangoLearn is an algorithm for lineage assignment which relies on machine learning and uses random forests. The model was trained using SARS-CoV-2 sequences from GISAID~\cite{gisaid}.
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.25]{figs/covitvsusher.png}
\end{center}
\caption{CoViT accuracy of the different models against UShER on a randomly selected set of accessions.
}
\label{fig:cvu}
\end{figure}
We compare the top-1 accuracy and placement speed of four CoViT models (i.e., CG1, CG2, CG3, CG4) against UShER~\cite{turakhia2021ultrafast} by placing 13,500 genome assemblies randomly selected from the Covid-19 Data Portal~\cite{covid19data}.
UShER is one of the latest phylogenetic placement tools. It uses a maximum parsimony approach where it searches the entire reference variant-annotated tree
for a placement that requires the fewest additional mutations.
One reported limitation of UShER is that it only supports mutations caused by replacements (does not support mutations caused by indels).
For models CG1 through CG4, CoViT achieves the average placement accuracy of 96.1\%, 94.2\%, 93.8\% 93.9\% respectively, while UshER's average accuracy is 92.1\% as depicted in Figure~\ref{fig:cvu}. To test the placement speed, we ran both CoViT and UShER on Intel i7-9700K CPU with 8 cores, operating at 3.60GHz. Additionally, CoViT was implemented on NVIDIA’s GeForce RTX
2080 Ti GPU. The placement times for these 13,500 genome assemblies are presented in table~\ref{Tab:t3}. The rest of the table is adopted from \cite{turakhia2021ultrafast}.
\begin{table}[t]
\caption{Placement time of 13,500 SARS-CoV-2 genomes assemblies.
\label{Tab:t3}}
\begin{center}
\begin{tabular}{||c|c|c||}
\hline
program & time & speedup vs UshER\\ [0.5ex]
\hline\hline
CG1 (GPU) & 12.42m & 3.6\\
\hline
CG2 (GPU) & 12.45m & 3.6\\
\hline
CG3 (GPU) & 12.47m & 3.6\\
\hline
CG4 (GPU) & 12.52m & 3.5\\
\hline
CG1 (CPU) & 21.35m & 2.1\\
\hline
CG2 (CPU) & 21.35m & 2.1\\
\hline
CG3 (CPU) & 21.35m & 2.1\\
\hline
CG4 (CPU) & 21.38m & 2.1\\
\hline
UshER & 44.1m & 1\\ [1ex]
\hline
\end{tabular}
\end{center}
\end{table} |
\section{Introduction}
Spin-orbit coupling (SOC) plays an important and, sometimes decisive, role in many condensed matter systems, including two-dimensional (2D) electron and hole gases in semiconductor heterostructures,\cite{zutic:2004,winkler:book} non-centrosymmetric normal metals~\cite{samokhin:2009} and superconductors,\cite{sigrist:1991,mineev_sigrist} bismuth tellurohalides \cite{Bahramy:2011}, a variety of iridates and vanadates, \cite{balents:2014}, surface/edge states of three-dimensional (3D)/2D topological insulators,\cite{hasan:2010,hasan:2011,qi:2011,alicea:2012,moore:2012}
conducting states at oxide interfaces,\cite{cen:2009} 2D transition metal dichalcogenides (TMD),\cite{manzeli:2017,heinz:2018} graphene on TMD substrates,\cite{morpurnat} atomic Bose\cite{lin:2009,lin:2011} and Fermi\cite{wang:2012, cheuk:2012} gases in simulated non-Abelian magnetic fields, etc.
Coupling between electron spins and momenta leads to a number of fascinating consequences, such as the electric-dipole spin resonance (EDSR), \cite{schulte:2005,wilamowski:2007,wilamowski:2008} current-induced spin polarization, \cite{kato:2004,sih:2005,wunderlich:2005} persistent spin helices, \cite{schliemann:2003, bernevig:2006,koralek:2009} quantum spin \cite{bernevig:2006a,bernevig:2006b,koenig:2007} and anomalous Hall effects,\cite{Chang:2013,Checkelsky:2014,Chang:2015} to name just a few. An interesting and still largely open question is the interplay between spin-orbit and electron-electron interactions. Such interplay gives rise to new phases of matter, e.g., topological Mott insulator,\cite{pesin:2010,balents:2013} gyrotropic and multipolar orders in normal metals,\cite{fu:2015} helical Fermi liquid (FL),\cite{lundgren:2015} Gor'kov-Rashba superconductor with mixed singlet-triplet order parameter,\cite{gorkov:2001} topological Kondo insulators, \cite{dzero:2010} etc. It also affects in a non-trivial way many physical phenomena, e.g., optical conductivity,\cite{farid:2006,agarwal:2011} plasmon spectra,\cite{badalyan:2009,raghu:2010,maiti:2014} RKKY interaction,\cite{simon:2007,chesi:2010,badalyan:2010} non-analytic behavior of the spin susceptibility,\cite{zak:2010,zak:2012,miserev:2021} etc., and gives rise to spin-dependent electron-electron interaction.\cite{gindikin:2022}
In this paper, we review recent progress in theoretical understanding and experimental observation of a new type of collective spin modes in 2D FLs with SOC. Such modes are perhaps the most direct manifestation of an interplay between spin-orbit and electron-electron interactions, as their existence hinges on both components being present. Unlike the conventional Silin mode in a partially spin-polarized FL,\cite{silin:1958} these modes exist even in the absence of an external magnetic field; in addition, they modify in a characteristic way the Silin mode if both SOC and magnetic field are present. As long as SOC is weak, the new modes correspond to oscillations of the magnetization which are decoupled from the oscillations of charge. The origin of the new modes can be traced to the effective spin-orbit magnetic field, which depends on the orientation and magnitude of the electron momentum, and also on the position of electron valley in the Brillouin zone (for multi-valley systems, such as graphene with proximity-induced SOC). Some of these modes have already been observed experimentally in Cd$_{1-x}$Mn$_{x}$Te quantum wells (in the presence of the magnetic field)\cite{perez:2007, baboux:2013,baboux:2015,perez:2016,perez:2017,perez:2019} and in the surface state of a three-dimensional (3D) topological insulator (TI) Bi$_2$Se$_3$ (in zero magnetic field);\cite{kung:2017} however, many more predictions are still awaiting their experimental confirmation.
The rest of the paper is organized as follows. In Sec.~\ref{sec:singleparticle}, we introduce three systems considered in the rest of the article: 1) a 2D electron gas (2DEG) with Rashba and/or Dresselhaus SOC, 2) graphene with proximity-induced SOC, and 3) a Dirac helical state on the surface of a 3D TI. In Sec.~\ref{sec:FL}, we describe the single- and two-valley FL theories, which will be applied to study the collective spin modes in 2DEGs and graphene, respectively, given that SOC and/or magnetic field are weak. In Sec.~\ref{sec:noFL}, we explain why a FL theory cannot be applied to the cases of arbitrarily strong SOC and/or magnetic field. Sec.~\ref{sec:Silin} serves as a short reminder of collective modes in a FL without SOC, in general, and of the Silin modes in a partially spin-polarized FL, in particular. In Sec.~\ref{sec:2DEG}, we discuss collective spin modes in a 2DEG. Sec.~\ref{sec:RSOC} describes the FL theory for the case of Rashba SOC. In Sec.~\ref{sec:lattice}, we show that the FL kinetic equation for a 2DEG with Rashba and/or Dresselhaus SOC and in the presence of the magnetic field can be mapped onto an effective tight-binding model for an artificial one-dimensional (1D) lattice, whose sites are labeled by the projections of the angular momentum. Within this mapping, the role of FL interaction is to produce ``defects'', both of the on-site and bond types, and the collective modes arise as bound states due to such defects. In Sec.~\ref{sec:RSOCZ}, we illustrate how this mapping works for the case of a 2DEG with Rashba SOC and in the presence of the magnetic field using the $s$-wave approximation for the Landau function. Sec.~\ref{sec:Dirac} deals with collective spin modes in Dirac systems. In Sec.~\ref{sec:GR} we apply a two-valley version of the FL theory to graphene with proximity-induced SOC. In Sec.~\ref{sec:helical}, we derive the spectrum of inter-band spin excitations in a Dirac surface state within the ladder approximation. In Sec.~\ref{sec:space}, we discuss the spatial dispersion of collective spin modes. Sec.~\ref{sec:damping} is devoted to damping due to both disorder and electron-electron interaction. In Sec.~\ref{sec:exp}, we discuss both the current and future experiment. Sec.~\ref{sec:finiteq} summarizes the results of a series of Raman experiments on Cd$_{1-x}$Mn$_{x}$Te. In Sec.~\ref{sec:Bi2Se3}, we provide a summary of recent Raman spectroscopy of a collective spin mode on the surface of Bi$_2$Se$_3$. Sec.~\ref{sec:ESR_EDSR} contains the theoretical predictions for electron spin resonance (ESR) and EDSR experiments on graphene with proximity-induced SOC, both in zero and strong (compared to SOC) magnetic field. Our conclusions are given in Sec.~\ref{sec:concl}.
\section{Model single-particle Hamiltonians and electron-electron interaction}
\label{sec:models}
\subsection{Model single-particle Hamiltonians for spin-orbit coupling}
\label{sec:singleparticle}
Despite the variety of real-life systems, the effects of SOC on the electron spectrum can be described by just a few low-energy Hamiltonians, constructed by using the symmetry arguments.
In this paper, we will consider three examples of two-dimensional electron systems: a two-dimensional electron gas (2DEG) sandwiched between two dissimilar semiconductors, monolayer graphene with substrate-induced SOC, and the surface state of a three-dimensional (3D) topological insulator (TI).
The effect of SOC on a 2DEG sandwiched between two dissimilar centrosymmetric semiconductors is described by the venerable Rashba Hamiltonian\cite{rashba:1959,bychkov:1984}
\begin{eqnarray}
\hat H_{\text{R}}=\frac{k^2}{2m}+\alpha(\hat{\boldsymbol{\sigma}}\times{\bf k})\cdot \hat {\bf z}=\frac{k^2}{2m}+\alpha\left(k_y\hat\sigma_x-k_x\hat\sigma_y\right),\nonumber\\
\label{HR}
\end{eqnarray}
where $\hat{\boldsymbol{\sigma}}=(\hat\sigma_x,\hat\sigma_y,\hat\sigma_z)$ is a vector of Pauli matrices which describe electron spin, $\hat {\bf z}$ is a unit vector along the normal to the 2DEG plane, and $\alpha$ is a phenomenological parameter (with units of velocity).
The simplest way to arrive at this Hamiltonian is to notice that a combination $(\hat{\boldsymbol{\sigma}}\times{\bf k})\cdot \hat {\bf z}$ is the only scalar which can be formed out of an axial vector ($\hat{\boldsymbol{\sigma}}$) and two polar vectors (${\bf k}$ and $\hat {\bf z}$).
In a bulk non-centrosymmetric semiconductor, e.g., of the A3B5 family (GaAs, CdTe, etc.), symmetry allows for a cubic (Dresselhaus) coupling between momentum and spin.\cite{dresselhaus:1955} A quantum well on the surface of such a semiconductor is described
by a 2D Dresselhaus Hamiltonian with linear coupling between spin and momentum obtained by projecting the bulk Dresselhaus term onto the quantum-well plane.\cite{dyakonov:1986,dyakonov:book} A particular form of the 2D Dresselhaus Hamiltonian depends
on the orientation of the quantum-well plane with respect to crystallographic axes. We will consider the most common case of a quantum well grown along the (001) direction. With the $x$-axis along the (100) direction and $y$-axis along the (010) the direction, the Dresselhaus Hamiltonian reads
\begin{eqnarray}
\hat H_{\text{D}}=\frac{k^2}{2m}+\beta\left(k_x\hat\sigma_x-k_y\hat\sigma_y\right).\label{HD}
\end{eqnarray}
In heterostructures made from non-centrosymmetric semiconductors the Rashba and Dresselhaus types of SOC usually occur simultaneously, and the total spin-orbit part of the Hamiltonian is the sum of the Rashba and Dresselhaus terms. Without loss of generality, we assume that $\alpha,\beta>0$.
Another popular system is graphene adsorbed on a transition-metal-dichalcogenide (TMD) substrate, such as WS$_2$, WSe$_2$ and MoS$_2$.\cite{Avsar:2014,morpurgo_2015,morpurgo_2016,Yang:2016,shi:2017,Dankert:2017,Ghiasi:2017,wakamura:2018,schonenberger:2018,Omar:2018,Benitez:2018,wakamura:2019,Wang:2019,Island:2019}
In this case, the induced SOC is expected to be a mixture of two types:\cite{morpurgo_2015,cummings:2017,garcia:2018} of Rashba SOC
and
valley-Zeeman (VZ) or Ising SOC, which acts as an out-of-plane magnetic field whose direction alternates between the $K$ and $K'$ valleys of graphene. We focus on the case of monolayer graphene with proximity-induced SOC,
described by the following low-energy Hamiltonian:
\begin{eqnarray}
\label{ham0}
\hat{H}_{\text{GR}} &=& v_\mathrm{D}(\tau_z \hat{\nu}_x k_x + \hat{\nu}_y k_y) + \Delta \hat\nu _z
+ \frac{\lambda_{\rm R}}{2} (\tau_z \hat{\sigma}_y \hat{\nu}_x - \hat{\sigma}_x \hat{\nu}_y)\nonumber\\
&& + \frac{\lambda_\mathrm{VZ}}{2} \tau_z \hat{\sigma}_z,
\end{eqnarray}
where $v_\mathrm{D}$ is the Dirac velocity, $\textbf{k}$ is the electron momentum measured either from the $K$ or $K'$ point of the graphene Brillouin zone, $\lambda_\mathrm{R}$ and $\lambda_\mathrm{VZ}$ are the coupling constants of the Rashba and VZ spin-orbit interactions, respectively, $\Delta$ is the gap due to substrate-induced asymmetry between the A and B sites of the honeycomb lattice, $\hat{\nu}_i$ ($i=x,y$) are the Pauli matrices in the sublattice (pseudospin) subspace, and $\tau_z=\pm 1$ labels the $K$ and $K'$ points. (An expression of the type $\hat\sigma_i\hat\nu_j$ is to be understood as a tensor product of two matrices. A single matrix in one of the subspaces implies that it is tensor product with a unity matrix in the other subspace.) In Eq.~\eqref{ham0}, we neglected the intrinsic, Kane-Mele type of SOC,\cite{KM} which is much weaker than the induced ones.\footnote{Theoretical estimates place the Kane-Mele coupling in the range from 1\,$\mu$eV \cite{min:2006,yao:2007} to 25-50\,$\mu$eV,\cite{trickey:2007,konschuh:2010} while
a recent ESR experiment reports the value of $42.2$\, $\mu$eV \cite{sichau:2019}}
Finally, the surface of a 3D TI, e.g., Bi$_2$Se$_3$, harbors a Dirac helical state.
If the hexagonal warping of the energy contours due to underlying crystal lattice can be neglected, this state is described by the Rashba-like Hamiltonian without the parabolic term\cite{fu:2009}
\begin{eqnarray}
\hat H_{\text{TI}}= v_\mathrm{D}(\hat{\boldsymbol{\sigma}}\times{\bf k})\cdot\hat{\bf z}.\label{HTI}
\end{eqnarray}
In all cases presented above, the effect of an in-plane magnetic field of magnitude $B$ and at angle $\gamma$ to the $x$-axis is accounted for by adding the usual Zeeman term
\begin{eqnarray}
\hat H_{\text{Z}}=\frac 12 \Delta_{\text{Z}}\left(\cos\gamma\;\hat\sigma_x+\sin\gamma\;\hat\sigma_y\right) \label{HZ}
\end{eqnarray}
to the corresponding Hamiltonian.
Here, $\Delta_{\text{Z}}=g\mu_BB$ is the Zeeman splitting, $g$ is the effective Land{\'e} factor (assumed to be isotropic) and $\mu_B$ is the Bohr magneton.
The properties of Hamiltonians \eqref{HR}-\eqref{HTI} are well-understood by now and we will not reproduce the known results here. It suffices to say that Rashba and Dresselhaus types of SOC lead to spin textures in the momentum space, which can be interpreted as the effect of an effective magnetic field,
while the VZ type of SOC in \eqref{ham0} acts as an out-of-plane magnetic field whose orientation alternates between the $K$ and $K'$ valleys.
\subsection{Models of electron-electron interaction}
\label{sec:FL}
Since the focus of this article is on the collective modes,
we need to invoke the electron-electron interaction as it is essential to induce collective behaviour. In all cases, we assume that our system is doped, such that the Fermi energy ($E_F$) is significantly larger than the spin-orbit and/or Zeeman splitting of the electron spectrum. An exception is the surface state of a TI, where SOC forms the spectrum rather than modifies the already existing one. With this exception, SOC can be treated as a perturbation imposed on a single-valley (2DEG) or two-valley (graphene) Fermi liquid. Modulo renormalizations by the interaction, the spin-orbit and Zeeman energy scales determine the frequencies of the corresponding collective modes. Therefore, to leading order in these energy scales, one can neglect the effect of SOC and magnetic field on the Landau interaction function. For the single-valley case, the latter is given by the usual form \cite{agd:1963,lifshitz:1980,baym:book}
\begin{eqnarray}
N_F^*\hat f({\bf k},{\bf k}')
=
F^s(\vartheta) +
\hat{\boldsymbol{\sigma}}
\cdot\hat{\boldsymbol{\sigma}}'
F^a(\vartheta),
\label{FSU2}
\end{eqnarray}
where $N_F^*$ is the total (including the spin degeneracy), renormalized density of states at the Fermi energy, $\vartheta$ is the angle between the momenta ${\bf k}$ and ${\bf k}'$ of two quasiparticles on the Fermi surface ($|{\bf k}|=|{\bf k}'|=k_F$), functions $F^s$ and $F^a$ describe the interaction in the symmetric (direct) and asymmetric (exchange) channels, and unprimed/primed $\sigma$-matrices refer to the spin state of the first/second quasiparticle.
The valley degree of freedom in graphene allows for more interaction channels. In addition to direct and exchange interactions between electrons within a single valley, one now also has an exchange interaction between the valleys, and a mixed spin-valley exchange interaction. If doping is low enough to neglect the trigonal warping of the Fermi contours, the Landau function of a two-valley FL can be written as\cite{Raines_FL:2021}
\begin{widetext}
\begin{eqnarray}
\label{FL}
N_F^*\hat f({\bf k},{\bf k}')
&=&
F^s(\vartheta) +
\hat{\boldsymbol{\sigma}}
\cdot\hat{\boldsymbol{\sigma}}'
F^a(\vartheta)
+
\hat{\boldsymbol{\tau}}_{||}
\cdot \hat{\boldsymbol{\tau}}_{||}'
G^{a||}(\vartheta)+\tau_{z}
\tau_{z}'
G^{a\perp}(\vartheta)
+ (\hat{\boldsymbol{\sigma}}
\cdot \hat{\boldsymbol{\sigma}}'
)\left[ \hat{\boldsymbol{\tau}}_{||}
\cdot \hat{\boldsymbol{\tau}}_{||}'
H^{||}(\vartheta)+\hat\tau_{z}
\hat\tau_{z}'
H^{\perp}(\vartheta)\right],
\end{eqnarray}
\end{widetext}
where
where $N_F^*$ is the total (including the spin and valley degeneracies), renormalized density of states, $\hat{\boldsymbol{\tau}}_{||}=\hat x\tau_x+\hat y\tau_y$ and unprimed/primed $\tau$-matrices refer to the valley state of the first/second quasiparticle. Scattering processes in a two-valley FL are depicted diagrammatically in Fig.~\ref{scatt}, where the solid (dashed) lines refer to electrons in the $K$ ($K'$) valley and $\varsigma_{1,2}$ ($\varsigma_{3,4}$) label the spin indices of the initial (final) states. Diagrams $a$ and $b$ describe intra-valley processes, while diagram $c$ describes an inter-valley process with small momentum transfers, in which electrons stay in their respective valleys. Diagram $d$ corresponds to an inter-valley process with large ($\approx |\bf{K}-\bf{K}'|$) momentum transfer, in which electrons are swapped between the valleys. For Coulomb interaction, the corresponding matrix element is smaller than the matrix elements in diagrams $a-c$. If the inter-valley matrix element is neglected, then the rotational symmetry in the valley space is restored,\cite{Raines_FL:2021} i.e., $G^{a||}(\vartheta)=G^{a\perp}(\vartheta)$ and $H^{a||}(\vartheta)=H^{a\perp}(\vartheta)$.
\begin{figure*}
\centering
\includegraphics[scale=0.45]{ScatteringVertices.pdf}
\caption{\label{scatt} Interaction vertices for intra-valley ($a$ and $b$) and inter-valley ($c$ and $d$) scattering processes. The solid (dashed) lines refer to electrons in the $K$ ($K'$) valley. Diagrams $a$ and $b$ also have exchange partners with outgoing states swapped (not shown). Diagram $d$ involves a large momentum transfer $\approx |\bf{K}-\bf{K}'|$. Reproduced with permission from Ref.~\onlinecite{Kumar_TMD:2021}. Copyright 2021 of the American Physical Society.}
\end{figure*}
Collective modes of a FL in the presence of SOC and/or Zeeman magnetic field (i.e., a field that affects only electron spins but not their orbital motion) can be found from the self-consistent FL kinetic equation for the (matrix) occupation number $\hat n({\bf k},{\bf r},t)$
\begin{eqnarray}
\label{kinetic_edsr}
\frac{\partial \hat{n}}{\partial t} + i [\hat{\varepsilon}, \hat{n}]+\frac 12\left\{ \boldsymbol{\nabla}_{\bf k} \hat \varepsilon, \cdot\boldsymbol{\nabla}_{\bf r} \hat n\right\} -\frac 12\left\{\boldsymbol{ \nabla}_{\bf r} \hat \varepsilon, \cdot\boldsymbol{\nabla}_{\bf k} \hat n\right\} = 0,\nonumber\\
\end{eqnarray}
where $[\hat a_1,\hat a_2]=\hat a_1\hat a_2-\hat a_2\hat a_1$, $\left\{\hat{\bf a}_1,\hat{\bf a}_2\right\}=\hat{\bf a}_1\cdot\hat{\bf a}_2+\hat{\bf a}_2\cdot\hat{\bf a}_1$, and $\hat\varepsilon({\bf k},{\bf r},t)$ is the functional of quasiparticle energy. In the most general case, $\hat\varepsilon({\bf k},{\bf r},t)$ can be written as
\begin{eqnarray}
\hat\varepsilon({\bf k},{\bf r},t)=\hat\varepsilon_0({\bf k})+\hat\varepsilon_{\text{SO}}({\bf k})+\hat\varepsilon_{\text{Z}}+\hat\varepsilon_{\text{FL}}({\bf k},{\bf r},t),
\label{KE}
\end{eqnarray}
where $\hat{\varepsilon}_0$ is the equilibrium quasiparticle energy in the absence of SOC and Zeeman field, $\hat\varepsilon_{\text{SO/Z}}$ are the changes in the energy due to SOC/Zeeman field, and
\begin{eqnarray}
\hat\varepsilon_{\text{FL}}({\bf k},{\bf r},t)=
\text{Tr}' \int \frac {d^Dk'}{(2\pi)^D} \hat f({\bf k},{\bf k}') \hat n'({\bf k}',{\bf r},t)
\end{eqnarray}
accounts for the interaction of a given quasiparticle with the rest. The effect of oscillatory magnetic and electric fields, applied in the ESR and EDSR measurements, are accounted for by adding the corresponding terms to the right-hand side (RHS) of Eq.~\eqref{KE}.
A number of comments are in order.
\begin{itemize}
\item[i)]
In the case of graphene with broken A/B symmetry, the kinetic equation \eqref{kinetic_edsr} needs to be modified to include the effect of a non-Abelian Berry curvature, which arises due to the combined effect of broken A/B symmetry and Rashba SOC.\cite{raines:2022} In this case, the term $\boldsymbol{\nabla}_{\bf k}\hat n$ is replaced by the covariant derivative $\mathcal{D} \hat n=\boldsymbol{\nabla}_{\bf k} \hat n-i\left[\mathcal{A},\hat n\right]$, where $\mathcal{A}$ is the non-Abelian Berry connection.\cite{culcer:2005,xiao:2010,bettelheim:2017} This leads to an effective orbital magnetic field which, in its turn, gives rise to an EDSR peak in the Hall conductivity, see Sec.~\ref{sec:GR}.
\item[ii)]Equation \ref{kinetic_edsr} neglects scattering of quasiparticles either by external sources (disorder, phonons) or by other quasiparticles, which lead to damping of collective modes. These effects will be discussed separately in Sec.~\ref{sec:damping}.
\item[iii)] The FL theory cannot describe a collective spin excitation that condenses out of the continuum of spin-flip transitions between the lower and upper cones of the Dirac surface state described by Eq.~\eqref{HTI}. The reason is that the energy of such an excitation is comparable to $2E_F$, while the FL theory can only describe physics at energies much smaller than $E_F$. In this case, one has to use microscopic, rather than phenomenological methods to describe the electron-electron interaction, see Sec.~\ref{sec:Bi2Se3}.
\item[iv)] If the reader is content with our assumption that the effects of SOC and Zeeman field on a FL can be treated perturbatively, they can skip the next section and go directly to Sec.~\ref{sec:Silin}. A more demanding reader is invited to read the next section, which explains why the FL theory cannot deal with strong SOC and/or strong Zeeman field.
\end{itemize}
\subsection{Does the Fermi-liquid theory work for arbitrarily strong spin-orbit coupling and/or magnetic field?}
\label{sec:noFL}
At first glance, the answer to the question in the title of this section is in the affirmative. All one needs to do is to construct a new Landau interaction function, accounting for broken rotational invariance in the spin space and, in the case of SOC, for coupling between momentum and spin. A modification of the Landau function in Eq.~\eqref{FSU2} for the case of a ferromagnetic metal was proposed long time ago in Ref.~\onlinecite{abrikosov:1958}. For the case of Rashba SOC, the modified Landau function was composed in Ref.~\onlinecite{ashrafi:2013}, but it is too long to be displayed here. It suffices to say that it contains eight instead of two components which, in contrast to Eq.~\eqref{FSU2}, depend not only on the angle between ${\bf k}$ and ${\bf k}'$ but also on the magnitudes of these momenta. And this is the first sign of a problem with the FL theory. To make this problem more evident, we consider the case of non-interacting electrons with Rashba spectrum (cf. Fig.~\ref{fig:Rashba_spectrum}a):
\begin{eqnarray}
\varepsilon^{\pm}_k=\frac{k^2}{2m}\pm \alpha k.
\end{eqnarray}
If both Rashba branches of the spectrum are occupied, the Fermi surface consists of two concentric circles with radii $k_{F}^\pm=\mp m\alpha+\sqrt{(m\alpha)^2+2mE_F}$, as shown in Fig.~\ref{fig:Rashba_spectrum}b. Let us calculate the spin susceptibility. In the diagrammatic language, the $ij$ component of the spin susceptibility is given by a polarization bubble with the corresponding Pauli matrices at the vertices. In the absence of SOC, the spin susceptibility comes from the states in the immediate vicinity of the Fermi surface. This is not so in the presence of SOC. The simplest case is when the Zeeman magnetic field is applied along the normal to the 2DEG plane, i.e., $i=j=z$, see Fig.~\ref{fig:Rashba_spectrum}c. In this case the $T=0$ static susceptibility arises solely from transitions between the two Rashba branches:
\begin{eqnarray}
\chi_{zz}&
=&\frac{g^2\mu_B^2}{2}\int_0^\infty \frac{dk k}{2\pi}\frac{n_{F}(\varepsilon^{-}_k)-n_{F}(\varepsilon^+_k)}{\varepsilon^+_k-\varepsilon^-_k}\nonumber\\
&&=\frac{g^2\mu_B^2}{2}\int_{k_{F}^+}^{k_{F}^-} \frac{dk}{4\pi\alpha}=\frac{g^2\mu_B^2}{2}\frac{m}{2\pi},\label{chizz}
\end{eqnarray}
where $n_{F}(\varepsilon)$ is the Fermi function. We see that the integral over $k$ does not come from the vicinity of either Fermi circle, but rather from the entire interval between the Ferm circles of width $\Delta k_F=k_{F}^--k_{F}^+=2m\alpha$.\footnote{The in-plane components, $\chi_{xx}=\chi_{yy}$, consist of two parts: one is the Ferm-surface contribution, as in the absence of SOC, and another one is the contribution from the entire interval $\Delta k_F$.} It is not a problem for a non-interacting case, because the spectrum is known for arbitrary $k$.\footnote{Note that the integral in Eq.~\eqref{chizz} can be also solved by rewriting it as $\chi_{zz}=(1/4\pi\alpha)\int^\infty_0 dk\left[n_{F}(\varepsilon^{+}_k)-n_{F}(\varepsilon^-_k)\right]$ and then integrating by parts. In this way, one obtains the same result as the sum of two contributions from the Fermi circles. This means that spin susceptibility is an anomalous quantity in the field-theoretical sense, i.e., it can be viewed equivalently either as low-energy or high-energy contribution.\cite{anomaly:book} But this equivalence is lost in the presence of interactions.} But it does become a problem for the interacting system. Indeed, the concept of quasiparticles is applicable only to long-lived states in the vicinity of each of the Fermi circle, see Fig.~\ref{fig:Rashba_spectrum}d. But states away from each of the Fermi circles (shown by red shaded regions) are just some complicated many-body states, which the FL theory cannot describe.
Suppose that one ignores this warning and goes ahead with calculating the renormalized $g$-factor, using the Landau function modified by Rashba SOC from Ref.~\onlinecite{ashrafi:2013}. The only modification which matters here is that, because SOC breaks spin-rotational invariance, the $\hat{\boldsymbol{\sigma}}\cdot\hat{\boldsymbol{\sigma}}'F^a$ term in Eq.~\eqref{FSU2} is replaced by a combination $\hat{\boldsymbol{\sigma}}_{||}\cdot \hat{\boldsymbol{\sigma}}'_{||}f^{a||}({\bf k},{\bf k}')+\hat\sigma_{z}\hat\sigma'_{z}f^{a\perp}({\bf k},{\bf k}')$, where $\hat{\boldsymbol{\sigma}}_{||}=\hat\sigma_x\hat{\bf x}+\hat\sigma_y\hat{\bf y}$. With this modification, one follows the same procedure as for a usual FL \cite{agd:1963,lifshitz:1980} and arrives at the integral equation, for e.g., the out-of-plane $g$-factor
\begin{eqnarray}
g_{zz}^*(k)=g-\int^{k^{*-}_{F}}_{k^{*+}_{F}} \frac{dk' k'}{\pi} \frac{f^{a\perp}_0(k,k')}{\varepsilon^{*+}_{k'}-\varepsilon^{*-}_{k'}}g_{zz}^*(k'),\nonumber\\\label{gfactor}
\end{eqnarray}
where $^*$ indicates a renormalized quantity. (Note that the Luttinger theorem guarantees only that the total area of the Fermi surface is not renormalized, i.e., that $k^{+*2}_{F}+k^{-*2}_{F}=k^{+2}_{F}+k^{-2}_{F}$, which does exclude renormalizations of $k_{F}^\pm$ individually.)
A brief inspection of Eq.~\eqref{gfactor} shows that, in general, it does not make sense. Indeed, the integral on its RHS involves energies of quasiparticles at an arbitrary point between two Fermi circles. However, as we explained before, quasiparticle states are not well-defined away from either of the Fermi circles. Therefore, the renormalization of the $g$-factor cannot be calculated within the FL theory. Only if SOC is weak and thus $\Delta k_F\ll k_F$, one can replace $\varepsilon^{*\pm}_k$ by their quasiparticle forms. But in this case the width of the integration interval cancels out with the energy splitting due to SOC, and we are back to the usual result for the renormalized $g$-factor: $g^*=g/(1+F^{a}_0)$. (For weak SOC, the difference between the in-plane and out-of-plane components of the exchange interaction becomes negligible, and we replaced $F^{a,\perp}_0$ by $F^{a}_0$ at the last step.)
A similar conclusion had eventually been reached in regard to a FL in the Zeeman magnetic field: if the Zeeman splitting is comparable to the Fermi energy, the FL theory does not work.\cite{meyerovich:1992a,meyerovich:1994,meyerovich:1994b,mineev:2004} Interestingly, problems with applying the FL theory to the cases of strong spin polarization, ferromagnetism, and SOC had been foreseen by Conway Herring as early as in 1966.\cite{herring:book} The problem discussed above is a generic feature of systems with broken $SU(2)$ symmetry.
What was said above does not imply that an interacting system of electrons with strong SOC belong to the category of non-Fermi liquids. Indeed, momentum-resolved probes, such as angular-resolved photoemission, would find well-defined quasiparticles near each of the Fermi circles. What it means is that the thermodynamic quantities characterizing the spin sector of such a FL cannot be described by a (small) set of Landau parameters, but must involve the information about states away from the Fermi surfaces.
This is not such an unusual situation. For example, renormalization of the effective mass of a Galilean-invariant FL is described by a single Landau parameter: $m^*=m(1+F_1^s)$. However, if the system is not Galilean-invariant (but still isotropic), the last formula changes to $m^*=mQ(1+F_1^s)$, where $Q$ involves integrals of the interaction vertices over the entire momentum space, \cite{baym:book,chubukov:2018} and thus the effective mass in this case cannot be described by any finite number of Landau parameters.
\begin{figure}[htb]
\centering
\includegraphics[width=1\columnwidth]{Rashba_spectrum.pdf}
\caption{a) Spectrum of the Rashba Hamiltonian \eqref{HR}. b) The Fermi surface of the Rashba-split spectrum for non-interacting electrons. c) Diagrammatic representation of the out-of-plane spin susceptibility, $\chi_{zz}$. $\pm$ label the Green's functions of the Rashba subbands. d) Same as in b) for interacting electrons. Red shaded regions denote states away from the Fermi circles which cannot be described by the Fermi-liquid theory.}
\label{fig:Rashba_spectrum}
\end{figure}
\section{Preamble: Silin modes in a partially spin-polarized Fermi liquid}
\label{sec:Silin}
Switching to the subject of collective modes, it is instructive to begin with a short reminder about the collective modes in a single-valley FL without SOC.\cite{silin:1958,lifshitz:1980,baym:book}
A neutral FL, e.g., normal $^3$He, supports long-lived collective excitation in the (charge) density sector: zero-sound waves with acoustic spectrum (line cZS in Fig.~\ref{fig:collmodes_B0}$a$). In a charged FL, the $l=0$ branch of zero-sound waves a replaced by plasmons, which are gapped in 3D (curve $P_{3D}$) and gapless in 2D (curve $P_{2D}$). The spin zero-sound mode does not fare that well. There is no general theorem that such a mode cannot be long-lived. However, as long as the interaction between fermions is repulsive (which corresponds to the attractive exchange interaction) and under reasonable assumptions about the harmonics of the Landau function $F^a(\vartheta)$ in Eq.~\eqref{FSU2}, i.e., $F_0^a<F_1^a<0$, etc., the mode is subsumed by the particle-hole continuum and thus overdamped (line sZS in Fig.~\ref{fig:collmodes_B0}$a$).
The situation changes in the presence of the Zeeman magnetic field, which gaps out the continuum of particle-hole excitations accompanied by spin-flips, see Fig.~\ref{fig:collmodes_B0}$b$. At $q=0$ the gap in the continuum is equal to the renormalized Zeeman energy $\Delta_\mathrm{Z}^*=\Delta_\mathrm{Z}/(1+F_0^a)$, then it decreases as $q$ increases, and vanishes eventually at $q\approx \Delta_\mathrm{Z}/v_F^*$, where $v_F^*$ is the renormalized Fermi velocity. The region below the lower boundary of the continuum is now free of particle-hole excitations and can support long-lived collective modes.
The frequencies of these modes at $q=0$ are found by writing the occupation number as $\hat n({\bf k},t)=n_F+n_F'\left[\Delta_\mathrm{Z}^*/2+{\bf u}({\bf k},t)\cdot\hat{\boldsymbol{\sigma}}\right]$
and the quasiparticle energy as $\hat\varepsilon({\bf k},t)=\varepsilon_0+\Delta_\mathrm{Z}^*\hat {\bf b}\cdot\hat{\boldsymbol{\sigma}}/2+\hat\varepsilon_{\text{FL}}({\bf k},t)$, where $\hat {\bf b}$ is the unit vector in the direction of the magnetic field and $n_F'=\partial n_F(\varepsilon)/\partial\varepsilon$, and using
Eq.~\eqref{kinetic_edsr} without the gradient term.\cite{lifshitz:1980,baym:book} This gives an equation of motion for the vector ${\bf u}$
\begin{eqnarray}
i\partial_t {\bf u}({\bf k}_F,t)=\Delta_\mathrm{Z}^* \hat {\bf b}\times\left[ {\bf u}({\bf k}_F,t)+\int\frac{d\mathcal{O}'}{\mathcal{O}_D} {\bf u}({\bf k}_F',t)F^a(\vartheta)\right],\nonumber\\
\label{silin_eq}\end{eqnarray}
where ${\bf k}_F=k_F {\bf k}/k$ and ${\bf k}_F'=k_F {\bf k}'/k'$, $d\mathcal{O}'$ is the element of the solid angle subtended by the vector ${\bf k}_F'$, and
$ \mathcal{O}_D$ is the full solid angle in $D$ dimensions. Equation \eqref{silin_eq}, which describes precession of the vector ${\bf u}$ around the magnetic field, is an eigenmode equation for the Silin collective modes.\cite{silin:1958} The frequencies of these modes depend on the angular momentum, $\ell$, in 3D or its projection onto the normal to the plane of motion, $m$, in 2D \cite{baym:book}
\begin{eqnarray}
\Omega_n=\Delta_\mathrm{Z}^*(1+F^{a}_n)=\Delta_\mathrm{Z} \frac{1+F^{a}_n}{1+F^{a}_0},\label{silin_freq}
\end{eqnarray}
with $n=\ell$ in 3D and $n=m$ in 2D. Here, $F^{a}_n$ are the harmonics of the Landau function defined by
\begin{eqnarray}
F^a(\vartheta)=\sum_{\ell} (2\ell +1) F_\ell^a \mathcal{P}_\ell(\vartheta)
\end{eqnarray}
and
\begin{eqnarray}
F^a(\vartheta)=\sum_{m} F_m^a e^{im\vartheta}
\end{eqnarray}
in 3D and 2D, respectively, and $\mathcal{P}_\ell(\vartheta)$ are the Legendre polynomials. Although there is, in principle, an infinite number of Silin modes, only the $n=0$ mode couples to an oscillatory magnetic field, applied in an ESR or nuclear magnetic resonance (NMR) measurements. Indeed, the magnetization is expressed solely through the zeroth harmonic of the function ${\bf u}({\bf k}_F,t)$:
$\boldsymbol{\mathcal{S}}=-(1/2)g^*\mu_BN_F^* \int d\mathcal{O} {\bf u}({\bf k}_F,t)/\mathcal{O}_D=(1/2)g^*
\mu_BN_F^*{\bf u}^0(t)$.
As we see from Eq.~\eqref{silin_freq}, the frequency of the $n=0$ mode is not renormalized by the interaction and coincides with the Larmor frequency for free fermions. This is a general property of Hamiltonians with interactions that conserve spin and do not depend on velocities.\cite{ma:1968} Diagrammatically, the results comes about as a cancellation between the self-energy and vertex corrections the spin susceptibility.\cite{ma:1968}
The difference between spin precession of free fermions and FL quasiparticles is illustrated in Fig.~\ref{fig:precession}. Although the spins of free fermions (on the left) and FL quasiparticles (on the right) precess with the same frequencies, the phases of the former are not correlated but the phases of the latter are locked.
The $n=0$ Silin mode disperses down with $q$, at first quadratically, and then grazes the continuum of spin-flip excitations, as shown in Fig.~\ref{fig:collmodes_B0}$b$. The downward sign of the dispersion can be explained by attraction in the exchange channel of the interaction.\cite{perez:2007} Unlike the frequency at $q=0$, which is not renormalized by the interaction, the functional form of the dispersion encapsulates all harmonics of $F^a$. In particular, the quadratic part contains harmonics $F^a_0$ (which can be also extracted from the spin susceptibility, if the effective mass is measured independently from the specific heat) and $F_1^a$ (which cannot). The dispersion of the Silin mode in normal $^3$He was studied extensively by NMR experiments in a spatially varying magnetic field which produced confined waves, see Ref.~\onlinecite{candela:1986} and references therein. In solid-state systems, the Silin modes where measured in the late 60s by ESR on alkali metals. \cite{schultz:1967,platzman:1967} More recently, the dispersion of the Silin mode was measured by high-precision, finite-$q$ Raman spectroscopy in a Cd$_{1-x}$Mn$_{x}$Te quantum wells.\cite{perez:2007, baboux:2013,baboux:2015,perez:2016,perez:2017,perez:2019} In this case, however, the Silin mode bears clear fingerprints of both Rashba and Dresselhaus SOCs, which are discussed in Sec.~\ref{sec:finiteq}. A tantalizing proposal to observe Silin modes in a quantum spin liquid with spinon Fermi surface has recently been put forward in Ref.~\onlinecite{starykh:2020}.
\begin{figure}[H]
\centering
\includegraphics[width=0.8\columnwidth]{Modes_All.pdf}\\
\vspace{0.1in}
\includegraphics[width=0.8\columnwidth]{Modes_Silin.pdf}\\
\vspace{0.1in}
\includegraphics[width=0.8\columnwidth]{Modes_CSW.pdf}
\caption{\label{fig:collmodes_B0} a) Collective modes of an $SU(2)$-invariant Fermi liquid. cZS and sZS denote acoustic zero-sound modes of a neutral Fermi liquid in the charge and spin sectors, respectively. $P_{2D}$ and $P_{3D}$ denote plasmons of a charged Fermi liquid in 2D and 3D, respectively. b) Dispersion of the $n=0$ Silin mode in the presence of the magnetic field. c) Collective modes of a single-valley 2D Fermi liquid with Rashba and/or Dresselhaus spin-orbit coupling. $\Omega_{||,x}$ and $\Omega_{||,y}$ denote the frequencies of two modes with in-plane magnetizations, $\Omega_{\perp}$ is the frequency of the out-of-plane mode.}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=1.0\columnwidth]{precession.pdf}
\caption{Precession of spins in a Fermi gas (left) and Fermi liquid (right). Although the frequencies of the $n=0$ mode are the same in both cases, the phases of precessing spins are uncorrelated in a Fermi gas and locked in a Fermi liquid.}
\label{fig:precession}
\end{figure}
\section{Chiral spin waves in two-dimensional electron gases}
\label{sec:2DEG}
\subsection{Chiral spin waves with Rashba spin-orbit coupling}
\label{sec:RSOC}
We now proceed with discussing the main subject of this paper--collective spin modes in the presence of SOC--starting from the simplest case of a single-valley 2DEG with Rashba SOC, studied first in Ref.~\onlinecite{shekhter:2005}. The single-particle Hamiltonian is given by Eq.~\eqref{HR} and the Landau function is given by Eq.~\eqref{FSU2}. It is convenient to introduce a set of rotated Pauli matrices which depend on the electron momentum\cite{shekhter:2005}
\begin{eqnarray}
\hat\Sigma_{x}({\bf k})&=&-\hat\sigma_z,\;\hat\Sigma_y({\bf k})=\cos\phi_{\bf k}\hat\sigma_x+\sin\phi_{\bf k}\hat\sigma_y,\nonumber\\
\hat\Sigma_z({\bf k})&=&\sin\phi_{\bf k}\hat\sigma_x-\cos\phi_{\bf k}\hat\sigma_y,\label{Sigma}
\end{eqnarray}
where $\phi_{\bf k}$ is the azimuthal angle of ${\bf k}$. When taken at the same momentum, $\hat\Sigma$'s obey the usual algebra $\left[\hat\Sigma_{\alpha}({\bf k}),\hat\Sigma_\beta({\bf k})\right]=i\epsilon^{\alpha\beta\gamma}\hat\Sigma_\gamma({\bf k})$, where $\epsilon^{\alpha\beta\gamma}$ is the Levi-Civita tensor, while $\left[\hat\Sigma_{y}({\bf k}),\hat\Sigma_z({\bf k}')\right]=i\hat\Sigma_x({\bf k})\cos(\phi_{\bf k}-\phi_{{\bf k}'})$. In this basis,
\begin{eqnarray}
\hat\varepsilon_{\text{SO}}=\frac 12 \Delta_\mathrm{R} \hat\Sigma_{z}({\bf k}),
\end{eqnarray}
where $\Delta_\mathrm{R}=2\alpha k_F$. Treating $\hat\varepsilon_{\text{SO}}$ as a static perturbation, we obtain the renormalized value of the Rashba splitting\cite{raikh:1999,shekhter:2005}
\begin{eqnarray}
\Delta_\mathrm{R}^*=\Delta_\mathrm{R}/(1+F^a_1).\label{renorm}
\end{eqnarray}
This energy marks the end point of the continuum of particle-hole excitations which involve spin-flip transitions between the branches of the Rashba spectrum, see Fig.~\ref{fig:collmodes_B0}$c$.
In a non-interacting system, the continuum at $q=0$ occupies a finite region
of width $\Delta\Omega=4m\alpha^2$. However, the FL theory is valid only to first order in $\alpha$, therefore, within this theory the continuum shrinks to a single point at $\Omega=\Delta_\mathrm{R}^*$.
Now we consider a time-dependent perturbation of the occupation number
\begin{eqnarray}
\delta\hat n({\bf k},t)=n_F'\left[\frac 12 \Delta_\mathrm{R}^* \hat\Sigma_{z}({\bf k})+{\bf u}({\bf k},t)\cdot\hat{\boldsymbol{\Sigma}}({\bf k})\right].\nonumber\\
\end{eqnarray}
Substituting this form into Eq.~\eqref{kinetic_edsr} (without the gradient term) and linearizing with respect to ${\bf u}$, we obtain the following equations of motion for angular harmonics of ${\bf u}$:
\begin{subequations}
\begin{eqnarray}
\partial_t u_{x}^m&= &-\Delta_\mathrm{R}^* u_{y}^m \left[1 +\frac{1}{2} \left(F_{m-1}^a + F_{m+1}^a \right) \right],
\label{u1_m}\\
\partial_t u_{y}^m&=& \Delta_\mathrm{R}^*u_{y}^m \left( 1+F_{m}^a \right),
\label{u2_m}\\
\partial_t u_{3}^m&=&0.
\label{u3_m}
\end{eqnarray}
\end{subequations}
The last equation implies that $u_{3}^m=\text{const}$ and can thus be ignored.
The eigenvalues of this system are the frequencies of a new type of collective modes: chiral spin resonances.\cite{shekhter:2005} Their frequencies are
\begin{eqnarray}
\Omega_m=\Delta_\mathrm{R}^*\sqrt{\left[1+\frac 12(F_{m+1}^a + F_{m-1}^a)\right](1+F_m^a)}.\nonumber\\
\label{freqR}
\end{eqnarray}
(Note that $\Omega_{-m}=\Omega_m$ because $F^a_{-m}=F^a_m$.)
As in the case of the Silin mode, there is an infinite number of chiral-spin resonances. Now, however, only two modes with $m=0$ and $m=1$ couple to the external magnetic field. This is because the magnetization
\begin{eqnarray}
\boldsymbol{\mathcal{S}}=-\frac{1}{4} g^*\mu_B N_F^*\text{Tr}\int \frac{d\phi_{\bf k}}{2\pi} \left[{\bf u}({\bf k}_F,t)\cdot\hat{\boldsymbol{\Sigma}}({\bf k}_F)\right]\hat{\boldsymbol{\sigma}}
\end{eqnarray}
contains not only the $0^{\text{th}}$ but also $\pm 1^{\text{st}}$ harmonics of ${\bf u}$ due to the angular dependence of matrices $\hat\Sigma_\alpha({\bf k})$. The $m=0$ mode with frequency
\begin{eqnarray}
\Omega_0\equiv\Omega_\perp=\Delta_\mathrm{R}^*\sqrt{(1+F_0^a)(1+F_1^a)}\label{Operp}
\end{eqnarray}
corresponds to oscillations of the $z$-component of $\boldsymbol{\mathcal{S}}$, while the doubly-degenerate $m=\pm 1$ mode with frequency
\begin{eqnarray}
\Omega_{\pm 1}\equiv\Omega_{||}=\Delta_\mathrm{R}^*\sqrt{\left[1+\frac 12\left(F_0^a+F_2^a\right)\right]\left(1+F_1^a\right)}\label{Opar}
\end{eqnarray}
corresponds to oscillations of the in-plane components of $\boldsymbol{\mathcal{S}}$. Provided that $F_0^a<F_2^a$, we have $\Omega_\perp<\Omega_{||}$, which is the case shown in Fig.~\ref{fig:collmodes_B0}$c$. Since spin is not a conserved quantity anymore, the frequencies of all modes are renormalized away from $\Delta_\mathrm{R}$.
The chiral spin resonances are the $q=0$ end points of the dispersive modes: chiral spin waves (CSW).\cite{ashrafi:2012,maiti:2014} Once the direction of ${\bf q}$ in the plane is chosen, the degeneracy of the two in-plane mode is lifted and there are now two modes with frequencies $\Omega_{||x}({\bf q})$ and $\Omega_{||y}({\bf q})$, polarized along and perpendicular to ${\bf q}$, respectively. The dispersions of these modes are shown schematically in Fig.~\ref{fig:collmodes_B0}$c$. The in-plane modes run into the continuum at some values of $q$ and disappear from the spectrum. The out-of-plane mode behaves similarly to the Silin mode: it disperses down with $q$ and grazes the continuum of particle-hole excitations. We will discuss the spatial dispersion of CSW in more detail in Sec.~\ref{sec:space}.
The equations of motion \eqref{u1_m}-\eqref{u3_m} allow for a simple physical interpretation.\cite{kumar_lattice:2017}
Namely, one can think of vector ${\bf u}^m(t)$ as a classical spin on site $m$ of a linear spin chain, aligned along the $y$-axis as in Fig.~\ref{ESR_tightbinding}. Spins do not interact with each other but are subject to an effective magnetic field due to Rashba SOC, directed along the $z$ axis and of magnitude $\Delta_\mathrm{R}^*/2$. The effective Land{\'e} factor of these spins is anisotropic in the $(x,y)$ plane with components $\gamma^m_x=2+ F_{m+1}^a+F_{m-1}^a$ and $\gamma^m_y=2(1+F_m^a)$.
The spin chain is {\em non-uniform} because $\gamma_1^m$ and $\gamma_2^m$ depend on the lattice site. Both anisotropy and site-dependence of the $g$-factor arise from the FL interaction. Because $u^m_3=\text{const}$,
the spins precess around the Rashba field.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.26]{ESR.pdf}
\caption{\label{ESR_tightbinding} a) An interpretation of collective modes in a Fermi liquid with Rashba spin-orbit coupling in terms of a linear spin chain. b) The spectrum of the system consists of an infinite number of discrete levels, converging towards the continuum at $\Delta_\mathrm{R}^*$. Reprinted with permission from Ref.~\onlinecite{kumar_lattice:2017}. Copyright 2017 by the American Physical Society.}
\end{figure}
The effective-lattice interpretation will be even more instructive in the case of both Rashba SOC and Zeeman field being present, which is discussed in the next section.
\subsection{Effective lattice model\label{sec:lattice}}
Now we add a Zeeman field aligned along the $x$-axis. The total Hamiltonian is the sum of Eqs.~\eqref{HR} and \eqref{HZ}, and the additional term in the quasiparticle energy in Eq.~\eqref{KE} is
\begin{eqnarray}
\hat\varepsilon_{\text{Z}}=\frac{\Delta_\mathrm{Z}^*}{2} \hat\sigma_x=\frac{\Delta_\mathrm{Z}^*}{2}\left[\cos\phi_{\bf k}\hat\Sigma_x({\bf k})+\sin\phi_{\bf k}\hat\Sigma_y({\bf k})\right].\nonumber\\
\label{SZ}
\end{eqnarray}
Correspondingly, the equations of motion for vector ${\bf u}$ become
\begin{widetext}
\begin{subequations}
\begin{eqnarray}
\partial_t u_{x}^m &= &-\Delta_\mathrm{R}^* u_{y}^m \left[ 1 +\frac{1}{2} \left(F_{m-1}^a + F_{m+1}^a \right) \right] - \frac{1}{2i}\Delta_\mathrm{Z}^*\left(u_{y}^{m-1} - u_{y}^{m+1} \right) \left( 1+F_{m}^a \right)
\nonumber\\
&&+ \frac{1}{2i}\Delta_\mathrm{R}^*u_{z}^m \left( F_{m-1}^a - F_{m+1}^a \right) + \frac{1}{2}\Delta_\mathrm{Z}^*\left( u_{z}^{m-1} + u_{z}^{m+1} \right)\left( 1+F_{m}^a \right),
\label{u1_mZ}\\
\partial_t u_{y}^m &=& \Delta_\mathrm{R}^*u_{x}^m \left(1+F_{m}^a \right) + \frac{1}{2i}\Delta_\mathrm{Z}^* \left[ u_{x}^{m-1} \left(1+F_{m-1}^a \right) - u_{x}^{m+1} \left(1+F_{m+1}^a \right) \right],
\label{u2_mZ}\\
\partial_t u_{3}^m &=& -\frac{1}{2} \Delta_\mathrm{Z}^* \left[ u_{x}^{m-1} \left(1+F_{m-1}^a \right) + u_{1}^{m+1} \left(1+F_{m+1}^a \right) \right].
\label{u3_mZ}
\end{eqnarray}
\end{subequations}
\end{widetext}
Note that the angular dependence of the Zeeman energy in the $\Sigma$-basis in Eq.~\eqref{SZ} leads to a non-locality in the $m$-space: the Zeeman terms in the equation of motion shift the angular momentum by $\pm 1$. Now the equations of motion resemble those for the tight-binding (TB) model with three orbitals per site, in which the Rashba and Zeeman terms play the roles of on-site and hopping energies, respectively. To make this analogy more transparent, we eliminate components $u_y^m$ and $u_z^m$ and introduce the ``Bloch wavefunction''
\begin{eqnarray}
\psi_m=i^{-m}u^m_x.\label{psi}
\end{eqnarray}
Assuming the oscillatory time dependence with frequency $\Omega$, the equation for $\psi_m$ is reduced to the TB form:\cite{kumar_lattice:2017}
\begin{widetext}
\begin{eqnarray}
\label{TB}
\Omega^2\psi_m&=& \left[\Delta_\mathrm{R}^{*2} \left( 1+F_m^a \right) \left( 1+\frac{F_{m+1}^a + F_{m-1}^a}{2} \right) + \Delta_\mathrm{Z}^{*2} \left( 1+F_m^a \right)^2 \right] \psi_m\nonumber \\
&&-
\Delta_\mathrm{R}^{*} \Delta_\mathrm{Z}^{*}
\left[
\left( 1 + \frac{F_{m}^a + F_{m+1}^a}{2} \right) \left( 1 + F_{m+1}^a \right)
\psi_{m+1} +
\left( 1 + \frac{F_{m-1}^a + F_m^a}{2} \right) \left(1 + F_{m-1}^a \right)
\psi_{m-1}\right].
\end{eqnarray}
\end{widetext}
In the absence of interaction, Eq.~\eqref{TB} is simplified to
\begin{equation}
\Omega^2\psi_m=\left(\Delta_\mathrm{R}^2+\Delta_\mathrm{Z}^2\right)\psi_m-\Delta_\mathrm{R}\Delta_\mathrm{Z}\left(\psi_{m+1}+\psi_{m-1}\right).\label{v1mB}
\end{equation}
The eigenvalue of this equation
\begin{eqnarray}
\Omega(\phi_{\bf k})=\left[\Delta_\mathrm{R}^2+\Delta_\mathrm{Z}^2-2\Delta_\mathrm{R}\Delta_\mathrm{Z}
\cos\phi_{\bf k}\right]^{1/2}
\label{free}
\end{eqnarray}
is nothing but the difference between the energies of the Rashba subbands at a given direction of ${\bf k}$ (modulo a phase shift inflicted by the transformation \eqref{psi}).
$\Omega(\phi_{\bf k})$ disperses with $\phi_{\bf k}\in(0,2\pi)$, which is a conjugate variable to $m$. Therefore, $\phi_{\bf k}$ plays the role of ``quasimomentum" confined to the first Brillouin zone $(0,2\pi)$.
The minimum and maximum values of $\Omega(\phi_{\bf k})$, $\Omega_{\max}=\Delta_\mathrm{R}+\Delta_\mathrm{Z}$ and $\Omega_{\min}=|\Delta_\mathrm{R}-\Delta_\mathrm{Z}|$,
mark the edges of the ``conduction band'', which is nothing but the particle-hole continuum. At $\Delta_\mathrm{Z}=\Delta_\mathrm{R}$ the gap in the continuum collapses to zero. At this value of the magnetic field, the Fermi contours of the Rashba branches touch at one point and thus a transition between the two branches costs no energy.
We now come back to the interacting version of Eq.~\eqref{TB}, which allows for a simple physical interpretation. Imagine a 1D lattice with sites labeled by index $m=0,\pm 1,\dots$. The Bloch wavefunction $\psi_m$ resides on these sites. The first term on the RHS of Eq.~\eqref{TB} is the energy on site $m$, the last two terms describe hopping between site $m$ and its the nearest neighbors, $m\pm 1$. If we artificially remove all $F_m^a$'s from the equation but keep the renormalized values of $\Delta_\mathrm{R}$ and $\Delta_\mathrm{Z}$, we will obtain the renormalized band (continuum) with boundaries $\Omega^*_{\max}=\Delta_\mathrm{R}^*+\Delta_\mathrm{Z}^*$ and $\Omega^*_{\min}=|\Delta_\mathrm{R}^*-\Delta_\mathrm{Z}^*|$. The role of $F_m^a$'s is to renormalize both on-site energies and hopping matrix elements. Because $F_m^a$ depends on $m$, both the on-site energies and hopping matrix elements vary along the lattice. In other words, $F_m^a$'s introduce both one-site and bond defects.
In principle, $F^a_m$ is non-zero for any $m$. Therefore, each site of our lattice is different from the others, and we can view our lattice as an ordered alloy composed from an infinite variety of chemically distinct atoms. But if we keep only a few first harmonics of $F^a(\vartheta)$ and neglect the rest, the central region of the lattice will contain defects, while the outer regions will be defect-free. Lattice defects produce bound states which split off the band. In this language, therefore, the collective modes of a FL are the bound states of an effective 1D lattice.
Studying a much more transparent problem of bound states, one can understand a more complicated case of a FL with SOC and in the presence of the magnetic field.
The case of Dresselhaus SOC can be analyzed in a similar way. According to Eq.~\eqref{HD}, the corresponding term in the quasiparticle energy reads
\begin{eqnarray}
\hat\varepsilon_{\text{SO}}&=&\frac{\Delta_\mathrm{D}^*}{2}\left(\cos\phi_{\bf k}\hat\sigma_x-\sin\phi_{\bf k}\hat\sigma_y\right)\nonumber\\
&=&\Delta_\mathrm{D}^* \left[\cos 2\phi_{\bf k}\hat\Sigma_y({\bf k})+\sin 2\phi_{\bf k}\hat\Sigma_z({\bf k})\right],\nonumber\\
\end{eqnarray}
where $\Delta_\mathrm{D}^*=\Delta_\mathrm{D}/(1+F_1^a)$.
A double angle in this equation implies that the corresponding terms in the equations of motion shift the angular momentum by $\pm 2$. In the lattice interpretation, this corresponds to hopping between next-to-nearest neighbors.
A complete dictionary of mapping between the FL kinetic equation and TB model is given in Table~\ref{table:I}.
\begin{table*}
\caption{\label{table:I}Mapping of the Fermi-liquid kinetic equation onto an effective 1D tight-binding model.}
\centering
\begin{tabular}{|c|c|}
\hline
{\bf Fermi-liquid kinetic equation} & {\bf 1D tight-binding model} \\
\hline
angular momentum $m$ & lattice site $m$\\
\hline
azimuthal angle of momentum ${\bf k}$ ($\phi_{\bf k}$) & quasimomentum\\
\hline
$m^\text{th}$ harmonic of the spin part of the occupation number $u^m_{x}$ & Bloch wavefunction on site $m$\\
\hline
Rashba spin-orbit coupling & on-site energy\\
\hline
Zeeman splitting & nearest-neighbor hopping\\
\hline
Dresselhaus spin-orbit coupling & next-to-nearest neighbor hopping\\
\hline
harmonics of the Landau function & on-site and bond defects\\
\hline
continuum of spin-flip particle-hole excitations & conduction band\\
\hline
collective modes & bound states\\
\hline
\end{tabular}
\end{table*}
We note that the zero-sound modes of a 2D FL in the absence of SOC and Zeeman field can also be understood as bound states of a 1D lattice. In this case, hopping between sites arises due the gradient term in the kinetic equation, which also shifts the harmonic index by $\pm 1$. For example, the shear zero-sound waves of a 2D FL were analyzed in this way in Ref.~\onlinecite{sodemann:2019}.
\subsection{Effective lattice model for Rashba spin-orbit coupling and Zeeman magnetic field}
\label{sec:RSOCZ}
We now illustrate how the lattice model works for the case when both Rashba SOC and Zeeman magnetic field are present. In addition, we assume that the Landau function is isotropic and thus contains only the zeroth harmonic: $F^a_m=\delta_{m,0}F_0^a$ (the $s$-wave approximation). In a dimensionless form and in the $s$-wave approximation, Eq.~(\ref{TB}) can be written as
\begin{widetext}
\begin{eqnarray}
\omega^2 \psi_m&=&
W\psi_m -J\left(\psi_{m+1}+\psi_{m-1}\right)+U_0\delta_{m,0}
\psi_m
+U_1\left(\delta_{m,1}+\delta_{m,-1}\right)
\psi_m
- \delta J\delta_{m,0}
\left(\psi_1+\psi_{-1}\right)-
\delta J'\left(\delta_{m,1}+\delta_{m,-1}\right)
\psi_0,\nonumber\\
\label{tight binding s-wave}
\end{eqnarray}
\end{widetext}
where
\begin{subequations}
\begin{eqnarray}
\omega&=&\Omega/\Delta_\mathrm{R},\;J=\Delta_\mathrm{Z}^*/\Delta_\mathrm{R},\;W=1+J^2,\label{par1}\\
U_0&=&F_0^a\left[1+J^2\left(2+F_0^a\right)\right],\;U_1= \frac{F_0^a}{2}\label{par2}\\
\delta J&=&J \frac{F^a_0}{2},\;\delta J'=\frac{F^a_0}{2}\left(3+F_0^a\right).\label{par3}
\end{eqnarray}
\end{subequations}
[Note that, according to Eq.~(\ref{renorm}), $\Delta_\mathrm{R}$ is not renormalized in the $s$-wave approximation.] The corresponding lattice is shown pictorially in Fig.~~\ref{ref:chiral}$a$.
\begin{figure}
\centering
\includegraphics[scale=0.35]{s-wave_chiral.pdf}
\caption{\label{ref:chiral} a) Effective lattice for the FL kinetic equation \eqref{tight binding s-wave} with Rashba spin-orbit coupling and in-plane magnetic field, and in the $s$-wave approximation for the Landau function. Filled circles represent sites of ideal lattice with on-site energies $W$, connected by bonds with hopping amplitudes $J$. The $m=0,\pm 1$ sites are occupied by ``impurities'' with on-site energies $U_0$ and $U_{\pm 1}$, respectively, connected by ``defective'' bonds with hopping amplitudes $J\pm \delta J$. b) A simplified version of the effective lattice with on-site disorder only. Reprinted with permission from Ref.~\onlinecite{kumar_lattice:2017}. Copyright 2017 by the American Physical Society.}
\end{figure}
The first two terms on the RHS of Eq.~\eqref{tight binding s-wave} describe an ideal lattice with ``on-site energies'' $W$ and ``hopping amplitudes'' between the nearest neighbors $J$. The rest of the terms represent ``defects''. The third and fourth term correspond to three ``impurities'' on sites $m=0$ and $m=\pm 1$ with on-site energies $U_0$ and $U_1$, respectively. The last two terms describe defective bonds between the $0^{\text{th}}$ and $\pm 1^{\text{st}}$ sites. The amplitudes are equal to $J+\delta J$ for hopping from $0$ to $\pm 1$ and $J+\delta J'$ for hopping in the opposite direction, i.e., the bond defects are {\em chiral}. This means that the effective TB Hamiltonian is non-Hermitian. This does not present any difficulties, however, because the eigenvalues of Eq.~(\ref{tight binding s-wave}) are real.
To solve Eq.~(\ref{tight binding s-wave}), we choose wavefunctions $\psi_0$ and $\psi_{\pm 1}$ as independent variables, and assume that, starting from sites $m=\pm 2$, the wavefunctions of the bound states decreases exponentially with $m$
:\begin{equation}{\psi}_{\pm (|m|+2)} = e^{-(|m|+1)\lambda}
{\psi}_{\pm 1}
\label{Ansatz}
\end{equation}
with $\text{Re}\lambda>0$. This yields a transcendental equation for $\lambda$, whose solutions are presented in Fig.~\ref{fig:R+B_s_wave}. The left panel is for magnetic fields below the gap closing point ($\Delta_\mathrm{Z}^*<\Delta_\mathrm{R}$). There are two collective modes in zero field--these are the same modes as given by Eqs.~\eqref{Operp} and \eqref{Opar}. An in-plane magnetic field lifts the double degeneracy of the $\Omega_{||}$ mode (similar to the case of finite $q$ in Fig.~\ref{fig:collmodes_B0}$c$) and, at finite but small $\Delta_\mathrm{Z}^*$ there are three modes. The in-plane modes run into the continuum at some critical values of $\Delta_\mathrm{Z}^*$, but the out-of-plane mode continues to graze the continuum down the gap-closing point. For $\Delta_\mathrm{Z}^*>\Delta_\mathrm{R}$, the single mode appears again, see Fig.~\ref{fig:R+B_s_wave}, right. Its frequency increases with the magnetic field and, in the limit of $\Delta_\mathrm{Z}^*\gg \Delta_\mathrm{R}$, the mode evolves into the Silin mode with frequency $\Delta_\mathrm{Z}$.
\begin{figure*}[htbp]
\centering
\includegraphics[scale=0.5]{R+B_s_wave.pdf}
\caption{\label{fig:R+B_s_wave} Collective modes of a FL with Rashba SOC and in the presence of the in-plane magnetic field. The Landau function is taken in the $s$-wave approximation: $F^a(\theta)=F_0^a=-0.3$. $\Delta_Z^*$ is the (renormalized) Zeeman energy and $\Delta_\mathrm{R}$ is the Rashba energy splitting. Left: $\Delta_Z^*<\Delta_\mathrm{R}$. Right: $\Delta_Z^*>\Delta_\mathrm{R}$. Inset: Same as in main panel for a wider range of fields. Reprinted with permission from Ref.~\onlinecite{kumar_lattice:2017}. Copyright 2017 by the American Physical Society.}
\end{figure*}
A rather complex spectrum shown in Fig.~\ref{fig:R+B_s_wave} can be understood in terms of simplified versions of the TB model. Namely, the merging of the two in-plane modes with the continuum can be understood qualitatively by ignoring bond defects, i.e., by setting $\delta J=\delta J'=0$. In this case, we have a TB model with identical bonds between all sites and three impurities on sites $m=0,\pm1$, see Fig.~\ref{ref:chiral}$b$.
For an even simpler case of a single impurity in a 1D lattice, it is well-known that there always exists a bound state located either below (for an attractive impurity) or above (for a repulsive impurity) the conduction band. For realistic values $-1<F_0^a<0$, our ``impurities'' are attractive, and thus the bound state is below the band. Given that
a single impurity has at least one bound state, it is natural to expect that a complex of three impurities will have up to three bound states, which is indeed confirmed by an explicit solution of the TB model (see Appendix A.1b in Ref.~\onlinecite{kumar_lattice:2017}). Qualitatively, this can be understood in the continuum limit, in which a three-impurity complex is replaced by a 1D potential well of finite width ($a$) and depth ($U$). Such a well has at least one bound state but may also have two, three, etc. states,
if the product $Ua$ exceeds some critical values.
Our original problem corresponds to a TB model with parameters given by Eqs.~(\ref{par1}-\ref{par3}).
In the limit $\Delta_\mathrm{Z}^*\ll \Delta_\mathrm{R}$ and not too weak interaction $|F_0^a|\sim 1$, the potential energies of the impurity sites are of the order of $1$, which is much larger than the bandwidth $2J=2\Delta_\mathrm{Z}^*/\Delta_\mathrm{R}\ll 1$. Thus we have three strong impurities with the maximum number of bound states, equal to three. As $\Delta_\mathrm{Z}^*$ increases, the bandwidth increases linearly whereas the potential energies increase only as $\Delta_\mathrm{Z}^{*2}$. Therefore, the impurities get relatively weaker (compared to the bandwidth), and we lose first the highest and then next-to-highest bound state. The lowest bound state also disappears but only at the gap-closing point. This seems to contradict the fact that there is at least one bound state in a 1D problem regardless to its parameters. However, this statement is true only for the continuum Schroedinger equation, which does not have the notion of bonds. In our full problem, two bonds are defective and there is an interesting competition between the on-site and bond defects, which does allow the bound state to disappear precisely at $\Delta_\mathrm{Z}^*=\Delta_\mathrm{R}$.\cite{kumar_lattice:2017}
As the field increases beyond the gap-closing point ($\Delta_\mathrm{Z}^*>\Delta_\mathrm{R}$), the $m=0$ impurity becomes stronger while the $m=\pm 1$ impurities remain the same. Therefore, we are back to a single-impurity problem with only one bound state or, equivalently, one collective mode.
From the solution of the kinetic equation one can also deduce the polarization of the collective modes. Figure \ref{fig:polar} shows how the polarization of chiral spin modes changes with the magnetic field.\cite{maiti:2016} In zero field, there are linearly polarized modes. At finite $B<B_c$, where $B_c$ is the field at which the gap in the continuum closes, along the $x$-axis, the longitudinal mode (with the magnetization along the field) remains linearly polarized, while two transverse mode (with the magnetizations perpendicular to the field) are elliptically polarized (as long as they are located outside the continuum). When the single wave emerges on the other side of the gap-closing point ($B>B_c$), it is elliptically polarized, with the magnetization vector precessing around the magnetic field. As the field becomes much larger than $B_c$, the wave transforms gradually into the Silin mode with circular polarization.
\begin{figure}[htbp]
\centering
\includegraphics[width=\columnwidth]{CSW_polarizations.pdf}
\caption{\label{fig:polar} Polarizations of chiral spin modes at different values of the magnetic fields applied along the $x$-axis. $B_c$ is the gap-closing field, at which $\Delta_\mathrm{Z}^*=\Delta_\mathrm{R}^*$.
Reprinted with permission from Ref.~\onlinecite{maiti:2016}. Copyright 2016 by the American Physical Society.}
\end{figure}
\begin{figure*}
\centering
\includegraphics[scale=0.5]
{model_interaction.pdf}
\caption{\label{fig:beyond_swave} Collective modes of a FL with Rashba SOC (left) and in the presence of the in-plane magnetic field (right) for a model form of the Landau function: $F_m^a = F_0^a \exp(-m^2/m_0^2)$ with $F_0^a=-0.3$ and $m_0 = 10$. Reprinted with permission from Ref.~\onlinecite{kumar_lattice:2017}. Copyright 2017 by the American Physical Society.}
\end{figure*}
If the Landau function contains a large or infinite numbers of harmonics, one has to resort to numerical diagonalization of Eq.~(\ref{TB}).
Figure \ref{fig:beyond_swave} shows the result of such diagonalization for a model form of the Landau function, given by
$F_m^a = F_0^a \exp(-m^2/m_0^2)$. In this case, the modes are very densely spaced and form a quasi-continuum.
\section{Collective spin modes in Dirac systems}
\label{sec:Dirac}
\subsection{Graphene with proximity-induced spin-orbit coupling}
\label{sec:GR}
The case of graphene with proximity-induced SOC differs from the one considered in the previous section in two important aspects. First, at the single-particle level, the Hamiltonian \eqref{ham0} contains the valley-Zeeman (VZ) term, in addition to the Rashba term. Second, at the many-body level, the Landau function contains exchange interaction between the valleys.
In this section, we ignore the asymmetry gap as it has no interesting consequences for the spectrum of the collective mode. Furthermore, we will neglect a process in which electrons are swapped between the valleys (see Fig.~\ref{scatt}$d$). Then the Landau function becomes isotropic not only in the spin but also in valley subspaces:
\begin{widetext}
\begin{eqnarray}
\label{FL1}
N_F^*\hat f({\bf k},{\bf k}')
&=&
F^s(\vartheta) +
\hat{\boldsymbol{\sigma}}
\cdot\hat{\boldsymbol{\sigma}}'
F^a(\vartheta)
+
\hat{\boldsymbol{\tau}}
\cdot \hat{\boldsymbol{\tau}}'
G^{a}(\vartheta)+ (\hat{\boldsymbol{\sigma}}
\cdot \hat{\boldsymbol{\sigma}}'
)( \hat{\boldsymbol{\tau}}
\cdot \hat{\boldsymbol{\tau}}')
H(\vartheta).
\end{eqnarray}
\end{widetext}
Since we are interested in collective modes with frequencies $\ll E_F$, the valence band can be projected out by using L{\"o}wdin method,\cite{lowdin:1951} and the effective single-particle Hamiltonian for the conduction band becomes\cite{Kumar_TMD:2021}
\begin{eqnarray}
\label{HR_cc}
\hat H_{c}&=&v_F k + \frac{\lambda_\mathrm{R}}{2} (\hat{\bf k}\times \hat{\boldsymbol{\sigma}})\cdot\hat{\bf z}+\frac{\lambda_\mathrm{VZ}}{2}\tau_z\hat \sigma_z.
\end{eqnarray}
Accordingly, the change in the quasiparticle energy due to SOC is now given by
\footnote{ Note that in $\tau_z$ is just $\pm1$ in Eq. (\ref{HR_cc}, but it is a matrix $\hat\tau_z$ in Eq. (\ref{eso}).}
\begin{eqnarray}
\hat{\varepsilon}_{\text{SO}}({\bf k})&=&\frac{\lambda_\mathrm{R}^*}{2}
(\hat{\bf k}\times \hat{\boldsymbol{\sigma}})\cdot\hat{\bf z}
+ \frac{\lambda_\mathrm{VZ}^*}{2}\hat{\tau}_z \hat{\sigma}_z,\nonumber\\
&=&\frac{\lambda_\mathrm{R}^*}{2}\hat\Sigma_z-\frac{\lambda_\mathrm{VZ}^*}{2}\hat{\tau}_z \hat{\Sigma}_x,\label{eso}
\end{eqnarray}
while the non-equilibrium part of the occupation number can be written as
\begin{eqnarray}
\delta \hat n({\bf k},t)&=&n_F'\left[{\bf u}({\bf k}, t) \cdot\hat{\boldsymbol{\Sigma}}({\bf k}) + {\bf w}({\bf k}, t) \cdot \hat{\boldsymbol{\tau}}\right.\nonumber\\
&&\left. + M_{\alpha\beta}({\bf k}, t) \hat{\tau}_\alpha \hat{\Sigma}_\beta({\bf k})\right],\label{dnsv}
\end{eqnarray}
where the rotated set of Pauli matrices is defined by Eq.~\eqref{Sigma}.Vector ${\bf u}$ and tensor $M_{\alpha\beta}$ describe oscillations of the uniform
\begin{equation}
\boldsymbol{\mathcal{S}}=-
\frac{g^*\mu_B}{8} \int \frac{d^{2}k}{(2\pi)^2} \text{Tr}\left[ \delta \hat n({\bf k},t)\,
\hat{\boldsymbol{\sigma}}\right]
\end{equation}
and valley-staggered magnetization
\begin{eqnarray}
\boldsymbol{\mathcal{M}}
=-
\frac{g^*\mu_B}{8} \int \frac{d^{2}k}{(2\pi)^2} \text{Tr}\left[ \delta \hat n({\bf k},t)\,\hat\tau_z
\hat{\boldsymbol{\sigma}}\right]
,\label{vsmagn}
\end{eqnarray}
respectively.
The vector ${\bf w}$ describes
valley polarization, which is decoupled from
the spin sector and will not be considered below.
Collective modes of our system correspond to coupled oscillations of the uniform and valley-staggered magnetizations. In the absence of VZ such oscillations are decoupled. The spectrum of collective excitations consists of a doublet of in-plane modes and a single out-of plane mode in each of the sectors, to a total of six modes. VZ SOC mixes the $\boldsymbol{\mathcal{S}}$ and $\boldsymbol{\mathcal{M}}$ sectors. Solving the coupled system of equations of motion, one obtains the following expressions for the frequencies of the in-plane modes\cite{Kumar_TMD:2021}
\begin{subequations}
\begin{equation}
\label{R+Z}
\begin{split}
\Omega_{||,\pm}^2 =&
\lambda_\mathrm{R}^{*2}\bigg( \frac{ff_++hh_+}{2} \bigg) + \lambda_\mathrm{VZ}^{*2} \bigg( f_+h_++f_-h_- \bigg)
\pm\Omega_{0}^2,
\end{split}
\end{equation}
where
\begin{widetext}
\begin{equation}
\label{O0}
\begin{split}
\Omega_{0}^2 =& \left[\lambda_\mathrm{R}^{*4}\bigg( \frac{ff_+-hh_+}{2}\bigg)^2
+\lambda_\mathrm{VZ}^{*4} (f_-h_++h_-f_+)^2+\lambda_\mathrm{R}^{*2} \lambda_\mathrm{VZ}^{*2}(ff_-+hh_-)(f_-h_++h_-f_+)\right]^{1/2},
\end{split}
\end{equation}
\end{widetext}
\end{subequations}
\begin{eqnarray}
f&=&1+F^a_1,\;f_+=1+\frac{F^a_0+F^a_2}{2},\; f_-= \frac{F^a_0-F^a_2}{2},\nonumber\\
h&=& 1+H_1,\;h_+=1+\frac{H_0+H_2}{2},\; h_-= \frac{H_0-H_2}{2},\nonumber\\
\label{fh}
\end{eqnarray}
and $\lambda_\mathrm{R}^*=\lambda_\mathrm{R}/(1+F_1^a)$ and $\lambda_\mathrm{VZ}^*=\lambda_\mathrm{VZ}/(1+H_0)$ are the renormalized spin-orbit coupling constants.
The frequencies of the out-of-plane modes are given by
\begin{eqnarray}
\Omega_{\perp,+}^2&=&h\left[(h_++h_-)\lambda_\mathrm{R}^{*2}+f\lambda_\mathrm{VZ}^{*2}\right]\nonumber\\
\Omega_{\perp,-}^2&=&f\left[(f_++f_-)\lambda_\mathrm{R}^{*2}+h\lambda_\mathrm{VZ}^{*2}\right]
\end{eqnarray}
Finite $q$ and/or Zeeman magnetic field lift the degeneracy of the in-plane modes. The spectrum of the collective modes in these cases is a subject of the on-going study.\cite{sam:unpub}
\subsection{Dirac surface state of a 3D topological insulator}
\label{sec:helical}
The Dirac state on the surface of a 3D topological insulator, described by the Hamiltonian \eqref{HTI}, is characterized by locking of the spin and charge degrees of freedom. Indeed, Eq.~\eqref{HTI} implies the charge current and spin densities are related by an operator identity\cite{raghu:2010}
\begin{eqnarray}
\hat{\bf j}=ev_\mathrm{D}\hat{\boldsymbol{\sigma}}\times \hat {\bf z}.\label{js}
\end{eqnarray}
Due to this coupling, the system supports a new kind of collective modes: spin-plasmons.\cite{raghu:2010} Although the dispersion of spin-plasmons at small $q$ is similar to that of usual plasmons in 2DEGs ($\omega\propto \sqrt{q}$), the nature of the two modes is quite different because a spin-plasmon corresponds to coupled oscillations of spin and charge densities. For example, the weight of the spin part at small $q$ is much larger than that of the charge part.
\begin{figure}
\centering
\includegraphics[scale=0.17]{helical.pdf}
\caption{\label{fig:helical} a) Spectrum of the Dirac state on the surface of a 3D topological insulator. b) Spectrum of the collective modes of the surface state. $\omega_p$ denotes the spin-plasmon \cite{raghu:2010}, $\omega_{s,||}$ and $\omega_{s,\perp}$ denote the out- and in-plane collective spin modes, and $\omega_-=2E_F$ is the lower boundary of the continuum of spin-flip particle-hole excitations. c) Ladder diagrams for the spin susceptibility $\chi_{\alpha\beta}$ ($\alpha,\beta=x,y,z$). Panel b) is reprinted with permission from Ref.~\onlinecite{kung:2017}. Copyright 2017 by the American Physical Society.}
\end{figure}
Low-energy excitations of a doped surface state can be described by the effective theory of helical FL, obtained by projecting out the occupied Dirac cone.\cite{lundgren:2015} While the charge sector of such a FL can be described by a few Landau parameters, description of the spin sector runs into difficulties similar to those discussed in Sec.~\ref{sec:noFL}. Namely, the total spin susceptibility contains contribution from high-energy states which cannot be accounted for within the FL theory. Nevertheless, the FS contribution to the spin susceptibility can be expressed through the Landau parameters of a helical FL.\cite{lundgren:2015}
In what follows, we will be interested in a collective mode of the spin-excitonic type observed in Ref.~\onlinecite{kung:2017}; see Sec.~\ref{sec:Bi2Se3} for more details on the experiment. This mode condenses out of the continuum of inter-band particle-hole excitations. In a non-interacting system with Dirac spectrum such a continuum starts at the Pauli threshold $\omega=2E_F$, see Fig.~\ref{fig:helical}$a$, and disperses with $q$ as shown by the shaded region in Fig.~\ref{fig:helical}$b$. The blue shaded region depicts the continuum of gapless intra-band particle-hole excitations. With interactions, one expects to see the spin-plasmon mode (blue curve)
and inter-band collective spin modes (red curves). By the same symmetry arguments as given in Sec.~\ref{sec:RSOC}, there are one out-of-plane mode and two in-plane modes at finite $q$; the latter become degenerate at $q=0$. Inter-band excitations with energies comparable to $E_F$ cannot be described within the FL theory, and one has to treat electron-electron interaction explicitly. The simplest method is the ladder approximation, shown diagrammatically in Fig.~\ref{fig:helical}$c$. For a Hubbard-like interaction and at $q=0$, one obtains the spin susceptibility as \cite{maiti:2014}
\begin{eqnarray}
\label{eq:RPA}
\chi_{\alpha\beta}(\Omega)=
\delta_{\alpha\beta}\frac{g^2\mu_B^2}{2}\frac{\Pi_{\alpha\beta}(\Omega)}
{1+\frac{U}{2}
\Pi_{\alpha\beta}(\Omega)},
\end{eqnarray}
where the components of the polarization bubble $\hat\Pi $ are obtained by analytic
continuation of the corresponding Matsubara expressions
\begin{widetext}
\begin{eqnarray}\label{eq:Pi}
\Pi_{\alpha\beta}(i\Omega_n)=T\sum_{m} \int\frac{d^2k}{(2\pi)^2} {\rm
Tr}\left[\hat\sigma_\alpha\hat G_0({\bf k},i\varepsilon_m+i\Omega_n)
\hat\sigma_\beta \hat G_0({\bf k},i\varepsilon_m)\right].
\end{eqnarray}
\end{widetext}
with $\alpha,\beta=x,y,z$ and
$G_0({\bf k},i\epsilon_m)=
\left(i\epsilon_m-\hat H_{\text{HTI}}\right)^{-1}$
is the Green's function
of the Hamiltonian \eqref{HTI}.
For Dirac spectrum, the momentum integrals for $\Pi_{\alpha\alpha}$ diverge in the ultraviolet and need to be cut off at some momentum $\Lambda$, chosen from the condition that the cubic term in the dispersion due to hexagonal warping\cite{fu:2009} becomes comparable to the linear one. With such a cutoff, we obtain
\begin{widetext}
\begin{eqnarray}
\mathrm{Re}\Pi_{xx}(\Omega)=\mathrm{Re}\Pi_{yy}(\Omega)=\frac 12 \mathrm{Re}\Pi_{zz}(\Omega)=-\frac{1}{4\pi v_\mathrm{D}}\left(\Lambda+\frac{\Omega}{4v_\mathrm{D}}\ln\left\vert\frac{\Omega+2E_F}{\Omega-2E_F}\right\vert\right).
\end{eqnarray}
\end{widetext}
The poles of Eq.~\eqref{eq:RPA} give the frequencies of the collective modes. At weak coupling, the modes are exponentially close to the boundary of the continuum:
\begin{eqnarray}
\Omega_{||}=2E_F-Ve^{-8/u},\;\Omega_\perp=2E_F-Ve^{-4/u},\label{Hubbard}
\end{eqnarray}
where $u=UE_F/2\piv_\mathrm{D}^2$ is the dimensionless coupling constant and $V=4E_F\exp(\Lambda/2k_F)$.
A reader familiar with the semiconductor literature of the 60s may notice that the derivation presented above is
very close to that of ``Mahan excitons'' in degenerate semiconductors.\cite{mahan:1967} Indeed, because of the relation \eqref{js} the in-plane components of the spin susceptibility and conductivity are proportional to each other. \cite{raghu:2010} The same analysis, therefore, would predict that the optical conductivity of a helical state should have a peak at $\Omega=\Omega_{||}$. This is exactly what Mahan obtained by (effectively) resumming the ladder series for the optical conductivity of a degenerate semiconductor; the only difference is that he considered Coulomb rather than Hubbard interaction but this difference is not significant. Indeed, a similar analysis for the Coulomb case shows that the collective modes are exponentially close to the continuum boundary as well, namely\cite{Adamya_thesis}
\begin{eqnarray}
\Omega_{||}&=&2E_F\left\{1-\frac{\alpha}{2}\exp\left[-\frac{\pi}{2\ln \alpha^{-1}}\left(\frac{1}{\alpha}- \ln\frac{\Lambda}{k_F} \right)\right]\right\},\nonumber\\
\Omega_{\perp}&=&2E_F\left\{1-\frac{\alpha}{2}\exp\left[-\frac{\pi}{2\ln \alpha^{-1}}\left(\frac{1}{2\alpha}- \ln\frac{\Lambda}{k_F} \right)\right]\right\},\nonumber\\
\end{eqnarray}
where $\alpha=e^2/v_\mathrm{D}$ is the dimensionless coupling constant of the Coulomb interaction.
Subsequent analysis showed, however, that the ladder approximation is not adequate for this problem:\cite{gavoret:1969,Pimenov:2017} because electrons and holes interact not in vacuum (as it is the case for an undoped semiconductor) but in the presence of the Fermi sea, the excitonic state is overdamped due to Auger-like electron-hole interactions processes, which start at the indirect threshold of $\Omega=E_F$. Moreover, damping due to electron-electron and electron-hole interactions starts already at $\Omega\ll E_F$, and is still operational at $\Omega\sim E_F$.\cite{Sharma:2021,adamya}
Both types of damping give rise to the linewidth proportional to $u^2$ (or $\alpha^2$), whereas the binding energy is exponentially small in $1/u$ (or $1/\alpha)$. Therefore, collective modes (and Mahan excitons) must be totally washed out at weak coupling.
Nevertheless, Eq.~\eqref{eq:RPA} describes very well a spin-resolved collective mode observed by Raman spectroscopy in Bi$_2$Se$_3$, which effectively probes the $zz$-component of the spin susceptibility, see Ref.~\onlinecite{kung:2017} and Sec.~\ref{sec:Bi2Se3}.
The frequency of the observed mode is about $E_F$, which implies that the system is not in the weak-coupling regime. On the other hand,
an order-of magnitude estimate for
%
the linewidth due to all types of electron- and electron-hole interactions is also of order $E_F$ for $\Omega\sim E_F$ and $u\sim 1$ (or $\alpha\sim 1$), so one would expect to see only a very broad peak at best.
Nevertheless, the observed peak is quite sharp: its linewidth is only about $1/20$ of its position. Moreover, the linewidth does not vary with temperature as the latter is raised up to 300 K, which indicates that the broadening is due to impurity scattering. Therefore, damping due to electron- and electron-hole interactions
is significantly weaker than a naive, order-of-magnitude estimate would suggest. We believe that the reason is purely numerical: as analytical and numerical results show, the linewidth due to both processes is only $5\times 10^{-3}E_F$ for a 2D helical state.\cite{Adamya_thesis,adamya} Therefore, the ladder approximation works much better than it might have been expected to.
\section{Spatial dispersion of collective spin modes}
\label{sec:space}
The dispersion of the transverse $n=0$ Silin mode is quadratic in $q$ for $q\ll \Delta_\mathrm{Z}^*/v_F^*$:\cite{silin:1958,ma:1968,mineev:2004, mineev:2005}
\begin{eqnarray}
\Omega(q)=\Delta_\mathrm{Z}+
a_2(\{F^a\})\frac{v_F^{*2}q^2}{\Delta_\mathrm{Z}},\label{qSilin}
\end{eqnarray}
where $a_2(\{F^a\})$ depends on the angular harmonics of the Landau function; explicitly,
$a_2=(1+F_0^a)^2(1+F_1^a)/(F_0^a-F_1^a)$. Note that $a_2<0$ for $-1<F_0^a<F_1^a<0$ and thus dispersion is downward, as shown in Fig.~\ref{fig:collmodes_B0}.
A quadratic scaling of the dispersion $q^2$ follows from the invariance of the spin subspace with respect to rotations about the magnetic field, which requires the dispersion to be isotropic, and from analyticity, which requires that an expansion in $q$ starts from the quadratic term.
If only Rashba SOC is present, the group symmetry of the system is $C_{\infty v}$, i.e., the system is invariant with respect to rotations by an arbitrary angle about the normal to the plane. Therefore, as in the Silin's case, the dispersion is isotropic and, by analyticity, starts with a $q^2$ term. For the lowest three modes,\cite{ashrafi:2012,zhang:2013,maiti:2014,maiti:2017}
\begin{eqnarray}
\Omega_{\alpha}(q)=\Omega_\alpha+
\tilde a^{\alpha}_2(\{F^a\}) \frac{v_F^{*2}q^2}{\Delta_\mathrm{R}},\label{qR}
\end{eqnarray}
where $\alpha\in\{||x,||y,\perp\}$, the $q=0$ frequencies are given by Eqs.~\ref{Operp} and Eqs.~\ref{Opar}, and $\tilde a^{\alpha}_2(\{F^a\})$ some functions of $F_0^a, F_1^a\dots$
If only Dresselhaus SOC is present, the Hamiltonian \eqref{HD} can be transformed back to the Rashba one \eqref{HR} by
by reflecting the spatial coordinates about a mirror plane that contains the (110) axis, upon which $k_x\to k_y$ and $k_y\to k_x$.
Therefore, the spectrum of the collective modes is the same as in Eq.~\eqref{qR} with $\Delta_\mathrm{R}$ being replaced by $\Delta_\mathrm{D}$.
If both Rashba and Dresselhaus SOCs are present, the symmetry is lowered to $D_{2d}$, which only allows rotations by $\pi$ and mirror reflections about the two diagonals. The linear term in the dispersion is absent because by symmetry it can only be $\sqrt{q_x^2+q_y^2}$, which is not allowed by analyticity, while a bilinear term is of the $c_1(q_x^2+q_y^2)+c_2 q_xq_y$ form, where $c_{1,2}$ are constants.
An in-plane magnetic field breaks the rotational symmetry. Now, linear-in-$q$ terms in the dispersion are allowed and, indeed, they have been observed experimentally.\cite{perez:2007,baboux:2013,baboux:2015,perez:2016,perez:2017,perez:2019} The structure of such terms can be deduced just from the symmetries of $C_{\infty v}$ and $D_{2d}$ groups.\cite{baboux:2015,perez:2016,maiti:2017} In both cases, we need to form a scalar out of a polar vector ${\bf q}$ and axial vector ${\bf B}$. In the $C_{\infty v}$ group, this is only possible by forming the Rashba invariant $B_xq_y-B_yq_x=Bq \sin(\phi_{\bf q}-\phi_{{\bf B}})$, were $\phi_{\bf q}$ and $\phi_{{\bf B}}$ are the azimuthal angles of ${\bf q}$ and ${\bf B}$, respectively. Likewise, the only scalar that can be constructed in the $D_{2d}$ group is the Dresselhaus invariant $B_xq_x-B_yq_y=Bq \cos(\phi_{\bf q}+\phi_{{\bf B}})$.
In addition to linear-in-$q$ terms, an in-plane magnetic field also gives rise to the dependence of the mode frequency at $q=0$ on the direction of the magnetic field. However, this effect needs both Rashba and Dresselhaus SOCs to be present. Indeed, since Rashba SOC has continuous rotational symmetry, the direction of the magnetic field is irrelevant in this case, while the case of pure Dresselhaus SOC is reduced back to pure Rashba one. If both types of SOC are present, the $D_{2d}$ symmetry implies that mode frequency depends on the direction of ${\bf B}$ as $\Omega_\alpha(q=0)\propto \sin 2\phi_{{\bf B}}$.\cite{maiti:2017,perez:2017}
To be specific, from this point onwards we will focus on the case when $\Delta_\mathrm{Z}\gg\Delta_\mathrm{R},\Delta_\mathrm{D}\neq 0$, which is relevant for Raman experiments on Cd$_{1-x}$Mn$_{x}$Te quantum wells.\cite{perez:2007,baboux:2013,baboux:2015,perez:2016,perez:2017} In this case, the mode frequency at $q=0$ is given primarily by the (bare) Zeeman energy. The correction due to SOC must be quadratic in both $\Delta_\mathrm{R}$ and $\Delta_\mathrm{D}$ and symmetric on $\Delta_\mathrm{R}\leftrightarrow\Delta_\mathrm{D}$, i.e., to be of the form $\Delta_\mathrm{R}^2+\Delta_\mathrm{D}^2$. The anisotropic term at the $q=0$ exists only of both $\Delta_\mathrm{R}$ and $\Delta_\mathrm{D}$ are non-zero which, in the limit considered, implies that the corresponding term must be proportional to $\Delta_\mathrm{R}\Delta_\mathrm{D}$. The linear-in-$q$ terms due to Rashba and Dresselhaus SOC must be proportional to $\Delta_\mathrm{R}$ and $\Delta_\mathrm{D}$, respectively. Finally, the bilinear-in-$q$ is almost the same as for the Silin mode modulo an anisotropic correction, which arises again only due to the combined effect of Rashba and Dresselhaus SOCs, and thus is also proportional to $\Delta_\mathrm{R}\Delta_\mathrm{D}$. Combining all the arguments given above, we arrive at the following form of the dispersion\cite{perez:2016,maiti:2017,perez:2017}
\begin{eqnarray}
\Omega({\bf q},{\bf B})= \Omega_0({\bf B})+w({\bf q},{\bf B})q+S({\bf q})\frac{v_F^{*2}q^2}{\Delta_\mathrm{Z}},\nonumber\\
\end{eqnarray}
where
\begin{eqnarray}
\Omega_0({\bf B})&=&\Delta_\mathrm{Z}+a_0(\{F^a\})\frac{\Delta_\mathrm{R}^2+\Delta_\mathrm{D}^2}{\Delta_\mathrm{Z}}\nonumber\\
&+& \tilde a_0(\{F^a\}) \frac{\Delta_\mathrm{R}\Delta_\mathrm{D}}{\Delta_\mathrm{Z}}\sin 2\phi_{{\bf B}},\nonumber\\
w({\bf q},{\bf B})&=&v_F a_1(\{F^a\})\left[\frac{\Delta_\mathrm{R}}{\Delta_\mathrm{Z}}\sin(\phi_{\bf q}-\phi_{{\bf B}})\right.\nonumber\\
&&\left.+\frac{\Delta_\mathrm{D}}{\Delta_\mathrm{Z}}\cos(\phi_{\bf q}+\phi_{{\bf B}})\right]\nonumber\\
S({\bf q})&=& a_2(\{F^a\})+\tilde a_2(\{F^a\})\frac{\Delta_\mathrm{R}\Delta_\mathrm{D}}{\Delta_\mathrm{Z}}\sin 2\phi_{\bf q},\label{qq2}
\end{eqnarray}
and the coefficient $a_2(\{F^a\})$ is the same as for a pure Silin mode, Eq.~\eqref{qSilin}. In the equation above we omitted isotropic $q^2$ terms proportional to $\Delta_\mathrm{R}^2$ and $\Delta_\mathrm{D}^2$. The coefficients $a_0$, $\tilde a_0$, and $a_1$ were calculated in Ref.~\onlinecite{maiti:2017} in the $s$-wave approximation for the Landau function, see Eq.~(20) in there.
It was argued in Refs.~\onlinecite{perez:2016,perez:2017} that, to linear order in SOC, the dispersion can be obtained by a canonical transformation of the Hamiltonian, which amounts to replacing ${\bf q}$ in the dispersion of the Silin mode, Eq.~\eqref{qSilin}, by ${\bf q}+{\bf q}_0$, where ${\bf q}_0$ is proportional to $\Delta_\mathrm{R}$ and $\Delta_\mathrm{D}$. This would imply that the coefficients of the $q$ and $q^2$ terms in Eq.~\eqref{qq2} are related as $|a_1|=2|a_2|$. However, a microscopic calculation\cite{maiti:2016} shows that such relation is satisfied only in the weak-coupling limit ($|F_0^a|\ll 1$).
\section{Damping of collective spin modes}
\label{sec:damping}
Spin-orbit coupling is the reason for the collective spin modes, described in this paper, to exist. At the same time, however, it couples spin and momentum and thus enables damping of spin excitations via momentum relaxation due to scattering by non-magnetic degrees of freedom: non-magnetic disorder, phonons, etc.
Non-magnetic disorder in combination with SOC leads to spin relaxation via the Elliott-Yaffet and D'yakonov-Perel' (DP) mechanisms.\cite{dyakonov:book} The former is present in both centro- and non-centrosymmetric systems, while the latter is specific for non-centrosymmetric systems, considered in this article, and in this case the DP mechanism is usually the dominant one. If $\tau$ is a characteristic time of momentum relaxation by disorder (to be defined more precisely later) and $\Delta_{\text{SO}}$ is the energy splitting due to SOC, then the spin dynamics is ballistic with the DP spin relaxation time $\tau_s\sim\tau$ for $\Delta _{\text{SO}}\tau\gg 1$, and diffusive with the DP spin relaxation time $\tau_s\sim 1/\Delta^2_{\text{SO}}\tau$ for $\Delta_{\text{SO}}\tau\ll 1$. On the other hand, the frequency of the collective mode is set by the largest of two energy scales: $\Delta_{\text{SO}}$ and the Zeeman energy $\Delta_\mathrm{Z}$. The spin collective mode can be resolved only when it is underdamped, which corresponds to the region outside the red square in Fig.~\ref{fig:DP}. This requires clean samples and strong SOC/magnetic fields.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{DP1.pdf}
\caption{\label{fig:DP} Damping of collective spin modes by non-magnetic disorder via D'yakonov-Perel' mechanism. $\Delta_{SO}$ and $\Delta$ are the splittings of the energy spectrum due to SOC and magnetic field, respectively, $\tau$ is the transport mean free time due to scattering by disorder. In the region outside the red square the spin collective modes are underdamped. The frequency of the collective mode $(\Omega$) and spin relaxation time ($\tau_s$) are indicated for each region. Within the red square, the modes are overdamped and thus cannot be observed.}
\end{figure}
One concrete example is a single-valley 2D FL with disorder and Rashba SOC. In this case, the width of the chiral spin resonance at $q=0$ is controlled by two relaxation times, renormalized by the FL interaction. For $\Delta_\mathrm{R}^*\tau\gg 1$, the width of the resonance is\cite{shekhter:2005}
\begin{eqnarray}
\Gamma_{\text{i}}=\frac{1}{2x_1\tau_1}+\frac{1}{2x_2(1+x_2/x_0)\tau_2},
\end{eqnarray}
where $x_m=1/(1+F^a_m)$,
\begin{eqnarray}
\frac{1}{\tau_{n}}=\pi N_F^* n_{\text{i}} \int \frac{d\vartheta}{2\pi} (1-\cos n\vartheta)\left\vert u\left(2k_F\sin \frac{\vartheta}{2}\right)\right\vert^2,\nonumber\\
\end{eqnarray}
$n_{\text{i}}$ is the number density of impurities, and $u(Q)$ is the Fourier transform of the single-impurity potential. Note that $\tau_1$ is the transport time that enters the mobility.
Another source of damping is electron-electron interaction. Because the collective modes lie outside the particle-hole continuum, they are not affected by Landau damping, which involves excitation of a single particle-hole pair. However, excitations involving multiple pairs are possible due to the residual interaction between FL quasiparticles, and such excitations give finite width even to modes outside the continuum. For example, plasmons, \cite{dubois:1969,mishchenko:2004,pustilnik:2006,imambekov:2012,principi:2013,Sharma:2021} Silin
mode in a partially spin-polarized FL,\cite{ma:1968,mineev:2004, mineev:2005} and magnons in a ferromagnetic FL\cite{kondratenko:1976,mineev:2004, mineev:2005} are all
damped via this mechanism. Figure \ref{fig:damping} shows diagrams that contribute to damping to lowest order in a dynamically screened Coulomb interaction.\footnote{Although the two last (Aslamazov-Larkin) diagrams appear to be higher order in the Coulomb interaction than the first three, they actually contribute to the same order in the dimensionless coupling constant $e^2/\hbar v_F$.}
Out of the examples listed above, the Silin mode is the closest one to collective spin modes due to SOC. However, there is an important difference.
Namely, conservation of the spin component along the field
ensures that the Silin mode is not damped at
$q\rightarrow0$.\cite{kondratenko:1965} On the contrary, spin is not a conserved quantity in a FL with SOC. Therefore, collective modes in this case are damped even at $q=0$. Evaluation of diagrams in Fig.~\ref{fig:damping} leads to an intuitively clear result: at $T=0$, the width of the resonance is given by\cite{maiti:2015}
\begin{eqnarray}
\Gamma_{\text{ee}}\sim \frac{\Delta_{\text{SO}}^2}{E_F}\lambda^2\ln \lambda^{-1},
\end{eqnarray}
where $\lambda=e^2/v_F$ is the dimensionless coupling constant of the Coulomb interaction. The quadratic dependence on $\Delta_{\text{SO}}$ is an expected scaling of a relaxation rate in a FL.\footnote{A single-particle relaxation rate in a 2D FL has an additional logarithmic factor, but it is canceled out between the diagrams for the spin relaxation rate, as guaranteed by the gauge invariance.}
Modes with finite $q$ are damped by the electron-electron interaction even in the absence of SOC. The same arguments of rotational invariance and analyticity that we used in Sec.~\ref{sec:space} to determine the $q$-dependence of the dispersion, can be applied to the damping rate.
Namely, in the presence of rotational symmetry, the damping rate is proportional to $q^2$.\cite{ma:1968,mineev:2004} If both types of SOC and an in-plane magnetic field are present, the damping rate contains a linear-in-$q$ term, whose structure is the same as of the linear term in the dispersion.\cite{perez:2017,perez:2019}
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{Diagrams_damping.pdf}
\caption{\label{fig:damping} Diagrams contributing to damping of the collective modes by dynamically screened Coulomb interaction (wavy line). The filled circles denote the corresponding vertices. For the case of spin collective modes, the vertices are Pauli matrices.}
\end{figure}
\section{Experiment:\\observations and predictions}
\label{sec:exp}
\subsection{Collective spin waves in $\text{Cd}_{1-x}\text{Mn}_x\text{Te}$ quantum wells}
\label{sec:finiteq}
The dispersion of collective spin waves in Cd$_{1-x}$Mn$_x$Te quantum wells was measured in a series of detailed Raman experiments.\cite{perez:2007,baboux:2013,baboux:2015,perez:2016,perez:2017} This system has both Rashba and Dresselhaus SOCs and thus should have chiral spin waves even in the zero magnetic field. However, the corresponding energy scales are below the resolution of Raman spectroscopy, and one has to apply an in-plane magnetic field to increase the energies of spin-flip excitations. Even a moderate field of 2 T leads to a significant spin polarization due the exchange interaction between magnetic moments of Mn dopants and conduction electron spins, and the effective Zeeman energy is larger than the spin-orbit one. To get some sense of the numbers, for the reported values of $\alpha^*\approx 1.8$\,meV$\cdot$\AA\,and $\beta^*\approx 3.8$\,meV$\cdot$\AA\, (Ref.~\onlinecite{perez:2016}),\footnote{We assigned stars to $\alpha$ and $\beta$ because the experiment measures only the renormalized values of the spin-orbit parameters.} the combined spin-orbit splitting is $\Delta^*_{\text{SO}}=\sqrt{\Delta_\mathrm{R}^{*2}+\Delta_\mathrm{D}^{*2}}\approx 0.1$\,meV at $n=2.7\times 10^{11}\text{cm}^{-2}$, while the measured frequency of the collective mode varies from $\sim 0.4$\, meV at $q=0$ in samples with lower Mn fraction($x=0.013$, Ref.~\onlinecite{perez:2016} ) to $\sim 3$\,mev in samples with higher Mn fraction ($x=0.8$, Ref.~\onlinecite{perez:2007}). Therefore, the effective Zeeman splitting is 4-30 times larger than the spin-orbit one, and the experimental situation corresponds to the right panel of Fig.~\ref{fig:R+B_s_wave} for $\Delta_\mathrm{Z}^*>\Delta_\mathrm{R}^*$ (with $\Delta_\mathrm{R}^*$ replaced by $\Delta^*_{\text{SO}}$), i.e., there is a single collective mode, which evolves into a pure Silin mode in the limit of $\Delta_\mathrm{Z}^*\to\infty$. At lower Mn fractions, however, the effects of SOC on the dispersion and damping of this mode are quite pronounced.
Panels $a$-$d$ in Fig.~\ref{fig:perez_2015} show experimental data for a [001] Cd$_{87}$Mn$_{0.13}$Te quantum well, reproduced from Refs.~\onlinecite{baboux:2015} and Refs.~\cite{perez:2017}. In the experiment, the in-plane magnetic field and vector ${\bf q}$ were kept at $90^{\circ}$ to each other, while the pair of vectors was rotated by angle $\varphi$ measured from the [100] direction, as shown in panel $a$. The Raman signal in panel $b$ exhibits a well-resolved peak which disperses with $q$. Interestingly, the dispersion is not purely quadratic but has a sizable linear term, which is revealed by flipping the direction of ${\bf q}$, as shown in panel $c$. Panels $e$-$g$, reproduced from Ref.~\onlinecite{perez:2017}, show the angular dependence of the mode frequency ($E_0$), spin-wave velocity ($E_1$), and spin-wave stiffness ($E_2$).
Now we compare the experimental results with the theory presented in Sec.~\ref{sec:space}. Under the condition $\phi_{{\bf B}}-\phi_{{\bf q}}=\pm \pi/2$, parameters of the dispersion in Eq.~\eqref{qq2} are reduced to
\begin{eqnarray}
\Omega_0({\bf B})&=&\Delta_\mathrm{Z}+a_0(\{F^a\})\frac{\Delta_\mathrm{R}^2+\Delta_\mathrm{D}^2}{\Delta_\mathrm{Z}}\nonumber\\
&+& \tilde a_0(\{F^a\}) \frac{\Delta_\mathrm{R}\Delta_\mathrm{D}}{\Delta_\mathrm{Z}}\sin 2\phi_{{\bf B}},\nonumber\\
w({\bf B})&=&\pm v_F a_1(\{F^a\})
\frac{\Delta_\mathrm{R}-\Delta_\mathrm{D}\sin 2\phi_{{\bf B}}}{\Delta_\mathrm{Z}}\nonumber\\
S({\bf B})&=& a_2(\{F^a\})-\tilde a_2(\{F^a\})\frac{\Delta_\mathrm{R}\Delta_\mathrm{D}}{\Delta_\mathrm{Z}}\sin 2\phi_{{\bf B}}.\label{qq2b}
\end{eqnarray}
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{perez_2015.pdf}
\includegraphics[scale=0.4]{dispersion_angle_egf1.pdf}
\caption{\label{fig:perez_2015} Experimental results for collective spin waves in $Cd_{1-x}$Mn$_x$Te quantum wells. a) Geometry of the Raman experiment. b) Raman signal. c) $q$-dependences of the spin-wave frequency. d) Angular dependence of the spin-wave velocity [denoted by $w({\bf B})$ in the main text]. Panels $e$-$g$: angular dependence of the mode frequency at $q=0$ [denoted by $\Omega_0({\bf B})$ in the main text], and spin-wave stiffness
[denoted by $S({\bf B})$ in the main text], respectively. Solid and dashed lines represent the numerical and analytical results of Ref.~\onlinecite{perez:2017}. Panels $a$-$d$ and $e$-$f$ are reproduced with permission from Ref.~\onlinecite{baboux:2015} and \onlinecite{perez:2017}, respectively. Copyright 2015, 2017 of the American Physical Society.}
\end{figure}
Figure \ref{fig:Raman} shows the theoretical results for $\Omega_0({\bf B})$ (inset in panel $a$), $w({\bf B})$ (panel $a$), and the $q$- dependence of the dispersion for two opposite orientations of the magnetic field, at $\pi/4$ and $-\pi/4$ (panel $b$). The value of $\Delta_\mathrm{Z}=0.4$\,meV at $B=2$\,T was taken from the experiment, while the Rashba and Dresselhaus coupling constants, and $F_0^a$ were used as fitting parameters. The fitted values of $\alpha$ and $\beta$ (1.9 mev\,\AA\, and 3.8 mev\,\AA) are very close to that reported in Ref.~\onlinecite{perez:2016} ($1.83\pm 0.08$\,meV \AA\, and $3.79\pm 0.11$\,meV \AA, respectively), while the fitted value $F_0^a=-0.41$ is quite close to the Hartree-Fock estimate $F_0^a=-0.3$ for a CdTe quantum well with $n=2.7\times 10^{11}\text{cm}^{-2}$.\cite{maiti:2017} We see that the spin-wave velocity ($w$) is indeed $\pi$-periodic and very close in magnitude the experimental result in panel $f$ of Fig.~\ref{fig:perez_2015}. The $\pi$-modulation of $\Omega_0({\bf B})$, shown in the inset of panel $a$,
is much smaller than that of $w$ because the former effect is second order in SOC.\cite{maiti:2017,perez:2017} This is consistent with panel $e$ of Fig.~\ref{fig:perez_2015}. The linear term in the dispersion is evident from panel $b$ of Fig.~\ref{fig:Raman} and consistent with panel $c$ of Fig.~\ref{fig:perez_2015}. Finally, the experiment also observes small $\pi$-modulation of the spin stiffness (panel $g$), which is consistent with this effect being also second order in SOC.
\begin{figure}
\includegraphics[width=1\columnwidth]{Raman_theory.pdf}
\caption{\label{fig:Raman} Theoretical results for the parameters of the spin-wave dispersion presented in Eq.~(\ref{qq2b}). Reprinted with permission from Ref.~\onlinecite{maiti:2017}. Copyright 2017 by the American Physical Society.}
\end{figure}
\subsection{Collective spin mode on the surface of B$\text{i}_2$S${\text{e}_3}$}
\label{sec:Bi2Se3}
Figure \ref{fig:kung2017} summarizes the results of polarization-resolved Raman spectroscopy of the surface state of a topological insulator Bi$_2$Se$_3$.\cite{kung:2017} To enhance the signal from the surface states, the frequency of incident light was tuned to the transition between two surface Dirac points: near the Fermi energy (SS1) and about $\Delta_S=1.8$\, eV above it (SS2), see panel I in Fig.~\ref{fig:kung2017}. The Raman response shows a well-resolved peak at $\approx 150$\,meV, while the Pauli threshold for this sample is at $2E_F\approx 260$\,meV. The peak is most pronounced for the incident frequency of $1.83$\,eV (panel II$c$), which is the closest to $\Delta_S$. This proves that the signal is indeed due to the surface rather than the bulk states. (Broader peaks at higher frequency, interpreted as the result of excitonic photoluminescence,\cite{Kung:2019} were eliminated by subtracting the hatched areas from the data.) Furthermore, polarization-resolved experiments reveal a magnetic nature of the $150$\,meV excitation. As shown in panel II$d$, the signal is much stronger in the RR channel, when right-polarized photons are scattered into right-polarized ones, i.e, when the angular momenta of the incident and scattered photons differ by $2\hbar$, than in the RL channel, when right-polarized photons are scattered into left-polarized ones, i.e., without a change in the angular momentum. For linearly polarized light, the signal is the strongest in the cross-polarization channel (XY), when the polarization axis is rotated by $\pi/2$. Decomposing the signal into components corresponding to irreducible representations of the $C_{6v}$ group, one finds that the $150$\,meV excitation belongs primarily to the $A_2$ representation, which is the pseudovector representation of $C_{6v}$, see panels III $a$-$c$. Noticeably, the excitation is very robust--it is observed up to 300\,K, see panel III$c$.
Given the findings summarized above, one is prompted to interpret the $150$\,meV excitation as the $q=0$ collective spin mode discussed in Sec.~\ref{sec:helical}, namely, as the out-of-plane mode with frequency $\Omega_\perp$, because in the experiment both the incident and scattered beams were along the normal to the surface. This interpretation is confirmed by the theoretical analysis, which shows that the Raman intensity is proportional to the $zz$-component of the spin susceptibility\cite{kung:2017}
\begin{eqnarray}
{\mathcal R}(\Omega,T)\propto \chi_{zz}(\omega,T)\frac{\gamma_{\text r}}{\left(\Omega_L-\Delta_{S}\right)^2+\gamma_{\text{r}}^2},
\end{eqnarray}
where $\Omega_L$ is the frequency of incident light
and $\gamma_{\text{r}}$ is the linewidth of the resonance.
$\chi_{zz}(\omega,T)$ was calculated within the ladder approximation for a realistic spectrum of the surface state and at finite temperature, and without using the weak-coupling assumption employed in Sec.~\ref{sec:helical}. [For reasons discussed in Sec.~\ref{sec:helical}, the ladder approximation works better than it might have been expected to, thanks to (accidental) numerical smallness of damping due to electron- and electron-hole interactions.] Since the linewidth of the observed peak is independent of temperature, the primary source of damping must be due to disorder. To mimic the effect of damping, the calculated $\chi_{zz}(\omega,T)$ was artificially broadened to produce the observed linewidth of $\approx 8$\,meV. In addition to the linewidth, the coupling constant of a Hubbard-like interaction was treated as a fitting parameter. The theoretical results, presented in panel III$d$, reproduce very well not only the profile of the peak as a function of frequency but also its temperature dependence. In particular, the theory reproduces the pronounced decrease of the resonance frequency with increasing temperature. This happens because the continuum broadens as $T$ increases, which pushes the resonance peak down to lower frequency. The inset of panel III$d$ shows a zoom on the interval between the resonance peak and continuum boundary, which is supposed to be at $2E_F\approx 260$\,meV. We see, however, that the continuum is barely discernible because most of its spectral weight is transferred to the collective mode. The best fit was obtained for the Hubbard coupling $u\approx 0.6$. This is consistent with the
estimate for the screened Coulomb interaction between electrons on the surface of Bi$_2$Se$_3$.
In summary, the results discussed above present strong evidence for a new type of the collective spin mode, arising from the combined effect of SOC and Coulomb interaction. Further experiments measuring the spatial dispersion of this mode would be highly desirable.
\begin{figure*}
\includegraphics[scale=0.28]{Bi2Se3_data.pdf}
\caption{\label{fig:kung2017} I) Bandstructure of Bi$_2$Se$_3$ around the Brillouin zone center. The two groups of Dirac States (SS1 and SS2, respectively) are separated by $\Delta_S\approx 1.8$\,eV. The frequency of incident photons $\Omega_L$ is tuned to a resonance transition between the SS1 and SS2 states.
II) Raman scattering data for different polarization geometries of the incoming and scattered photons, and
different incident photon energies. The collective mode at the Raman shift of $150~$meV is resonantly enhanced by the 1.83 eV photon. III (a)-(c): Raman scattering data decomposed into different channels, corresponding to $E_2$, $A_1$, and $A_2$ irreducible representations of the $C_{6v}$ group, respectively.
III (d): Theoretical results for the Raman response within the ladder approximation. Reprinted with permission from Ref.~\onlinecite{kung:2017}. Copyright 2017 by the American Physical Society.}
\end{figure*}
\subsection{Predictions for future experiments}
\label{sec:ESR_EDSR}
\subsubsection{ESR and EDSR in 2DEGs}
The rate of absorption of electromagnetic wave at normal incidence by a 2DEG with Rashba and/or Dresselhaus SOC can be written as
\begin{eqnarray}
Q=\frac 12\left[\mu_B^2\Omega\chi''_{||}(\Omega)+\sigma'(\Omega)\right]E_{\text{em}}^2,\label{abs}
\end{eqnarray}
where $E_{\text{em}}$ is the amplitude of the electric field of the wave, $\sigma(\Omega)$ is the conductivity of a 2DEG and $\chi_{||}(\Omega)$ is the in-plane component of its spin susceptibility. The first term represents absorption due to ESR, which can be observed even in the absence of SOC. However, its magnitude is proportional to $1/c^{2}$ (in the Gaussian unit system) and thus small. In 2D systems instead of ESR one typically measures electrically-detected spin resonance,
observed as a peak in the longitudinal resistivity under microwave radiation in the regime of integer Hall effect. \cite{stein:1983,graeff:1999,bowers:2000,bowers:2003,sichau:2019}
The second term in Eq.~\eqref{abs}
represents absorption due to EDSR.\cite{rashba:1991,rashbaefros2003,efros2006}
Its origin is an effective magnetic field acting on electron spins due to SOC and with magnitude proportional to $k$. The driving electric field (either from a {\em dc} current or electromagnetic wave) gives rise to a flow of electrons with a non-zero drift velocity, and hence the electron system as a whole experiences an effective magnetic field due to SOC.
This effect gives rise to a range of spectacular phenomena, e.g., a strong enhancement of microwave absorption in a geometry when the driving electric field is in the plane of a 2DEG~\cite{schulte:2005} and a shift of the spin resonance frequency by {\em dc} current.~\cite{wilamowski:2007,wilamowski:2008}
Collective spin modes described in the previous parts of the paper correspond to oscillations of the electron magnetization even in the absence of the external magnetic field. Therefore, they should be detectable both via ESR and EDSR which, in contrast to the conventional setup, should be present even in zero magnetic field. The structure of the signal can be understood qualitatively from Fig.~\ref{fig:R+B_s_wave}, where one just has to replace $\Delta_\mathrm{R}\to\sqrt{\Delta_\mathrm{R}^2+\Delta_\mathrm{D}^2}$ to account for Dresselhaus SOC. In the absence of the magnetic field ($\Delta_\mathrm{Z}^*=0$), the signal consists of two peaks, at frequencies $\Omega_\perp$ and $\Omega_{||}$. At finite field, the peak at $\Omega_{||}$ splits into two. Upon further increase of the field, the $\Omega_{||}$ peaks merge with the continuum and die out, while the $\Omega_\perp$ peak continues to be present all the way till the gap closing point ($\Delta_\mathrm{Z}^*=\sqrt{\Delta_\mathrm{R}^2+\Delta_\mathrm{D}^2}$), and then emerges again at fields above
this point.
To estimate the relative strength of the ESR and EDSR signals, we note that the conductivity in Eq.~\eqref{abs} is the sum of the Drude and spin-orbit parts: $\sigma'(\Omega)=\sigma'_{\text{D}}(\Omega)+\sigma'_{\text{SO}}(\Omega)$.
Since for the Hamiltonians \eqref{HR} and \eqref{HD} the electric current is proportional to magnetization, the spin-orbit part of the conductivity and spin susceptibility are related by
$\sigma'_{\text{SO}}(\Omega)\sim e^2\max\{\alpha^2,\beta^2\} \chi''(\Omega)/\Omega$. Provided that the Drude part at the resonance frequency $\Omega_{\text{r}}$ can be neglected, the ratio of the EDSR to ESR signals can be estimated as\cite{shekhter:2005}
\begin{equation}
\begin{split}
\frac{Q_{\text{EDSR}}}{Q_{\text{ESR}}} &\sim \left(\frac{\max\{\alpha,\beta\}mc}{\Omega_{\text{r}}}\right)^2 \\
&\sim \left(\frac{\lambda_F}{\lambda_C}\right)^2\left\{
\begin{array}{ccc}
1,\;\text{for}\;\Delta_\mathrm{Z}\ll \max\{\Delta_\mathrm{R},\Delta_\mathrm{D}\},\\
\left(\frac{\max\{\Delta_\mathrm{R},\Delta_\mathrm{D}\}}{\Delta_\mathrm{Z}}\right)^2,
\end{array}
\right.
\end{split}
\end{equation}
where $\lambda_F=2\pi/k_F$ is the Fermi wavelength and $\lambda_C=2\pi/mc=2.4\times 10^{-10}$\,cm is the Compton wavelength. For electron number densities in the interval $n=10^{11}-10^{12}$\,cm$^{-2}$, the factor $(\lambda_F/\lambda_C)^2\sim 10^{8}-10^{9}$, and the EDSR signal is stronger than the ESR one by many orders of magnitude, even if SOC is weak.
However, there is a caveat in this estimate, namely, it is valid provided that the Drude part of the conductivity is much smaller then the spin-orbit part, which imposes rather stringent conditions on the strength SOC and sample quality. Near the resonance, the spin-orbit part of the conductivity can be estimated as\cite{shekhter:2005,maiti:2015}
\begin{eqnarray}
\sigma'_{\text{SO}}(\Omega)\sim \frac{e^2}{h} m^*\max\{\alpha^2,\beta^2\}\frac{
\Gamma}{\left(\Omega-\Omega_{\text{r}}\right)^2+
\Gamma^2}.
\end{eqnarray}
Assuming that the linewidth of the resonance is due to D'yakonov-Perel' mechanism in the ballistic regime (cf. Sec.~\ref{sec:damping}), i.e, $\Gamma\sim 1/\tau$,
and that the Drude conductivity is controlled by the same scattering mechanism, i.e., $\sigma'_{\text{D}}(\Omega_{\text{r}})\sim (e^2/h) E_F/\Omega^2_{\text{r}}\tau$, we obtain for the ratio of the two parts of the conductivity right at the resonance:
\begin{eqnarray}
K=\frac{\sigma'_{\text{SO}}(\Omega_{\text{r}})}{\sigma'_{\text{D}}(\Omega_{\text{r}})}\sim
\frac{m^*\max\{\alpha^2,\beta^2\}}{E_F}
\left(\Omega_{\text{r}}\tau\right)^2.\nonumber\\
\label{contrast}
\end{eqnarray}
Therefore, even if the resonance is underdamped, i.e., $\Omega_{\text{r}}\tau\gg 1$, it can be still masked by the Drude part if the first factor on the RHS of Eq.~\eqref{contrast} is sufficiently small.
Estimates\cite{maiti:2015} show that the resonance in zero magnetic field, when $\Omega_{\text{r}}\sim \max\{\Delta_\mathrm{R},\Delta_\mathrm{D}\}$, would be completely masked in a GaAs/GaAlAs heterostructure even with a record-hight mobility of $10^7$\,cm$^2$/V\,s because, due to a relatively weak SOC in this system ($\alpha\sim\beta\sim 1$\,mev\,\AA), $K$ is only $\sim 0.1$. The problem is further exacerbated by the fact that, in the presence of both Rashba and Dresselhaus SOCs, the lower edge of the continuum is located at $\Omega_{-}=|\alpha-\beta|k_F$, which pushes the energies of the collective modes further down.
A better candidate is an InGaAs/InAlAs quantum well, where SOC is much stronger, i.e., $\alpha\sim 100\; \mathrm{meV}\,\mathrm{\AA}$ (Ref.~\onlinecite{nitta:1997}), which helps to compensate for smaller mobilities typical for these structures; the highest mobilities reported for InGaAs/InAlAs samples are in the range $\mu=(2-5)\times 10^5 \;\mathrm{cm}^2$/V\,s (Refs.~\onlinecite{sato:2001} and \onlinecite{yamada:2003}). Also, SOC in these structures is predominantly of the Rashba type,\cite{nitta:1997,luo:1990} which alleviates the problem with a competition between the Rashba and Dresselhaus mechanisms. For a high-mobility InGaAs/InAlAs quantum well, $K\sim 1$ and a detailed calculation confirms that the zero-filed EDSR peak should be visible against the Drude background.\cite{maiti:2015}
\subsubsection{ESR and EDSR in graphene with proximity-induced spin-orbit coupling}
\label{sec:GR_ESR}
\paragraph{Zero magnetic field.}
\begin{figure*}
\centering
\includegraphics[scale=0.5]{ESR_EDSR.pdf}
\caption{\label{Peaks} Theoretical predictions for the zero-field electron spin resonance (ESR) and electric-dipole spin resonance (EDSR) in graphene with proximity-induced spin-orbit coupling (SOC). (a) ESR signal. Vertical axis: the imaginary part of the dynamical spin susceptibility. The frequency on the horizontal axis is
scaled with $\lambda^*_{\text{SOC}}=\sqrt{\lambda_\mathrm{R}^{*2}+\lambda_\mathrm{VZ}^{*2}}$, where $\lambda_\mathrm{R}^*$ and $\lambda_\mathrm{VZ}^*$ are (renormalized) couplings of the Rashba and valley-Zeeman (VZ) types of SOC, respectively. $\Omega_{\pm}$ are the resonance frequencies, given by Eqs.~\eqref{R+Z} and \eqref{O0}. Dashed line: non-interacting system. Red solid line: a two-valley Fermi liquid (FL) with parameters $F^a_0=-0.5500$, $F^a_1=-0.2750$, $F^a_2=-0.1375$, $H_0=-0.5000$, $H_1=-0.2500$, and $H_2=-0.1250$. The ratio $\lambda_\mathrm{VZ}^*/\lambda_\mathrm{R}^*=0.5$. The choice of FL parameters is the same for all panels of the figure.
(b) ESR signal in a FL for several values of $\lambda_\mathrm{VZ}^*/\lambda_\mathrm{R}^*$, as indicated in the legend.
(c) EDSR signal. Vertical axis: the real part of the optical conductivity.
Dashed line: non-interacting system. Solid line: FL.
(d) EDSR signal in a FL for two values of $\lambda_\mathrm{VZ}^*/\lambda_\mathrm{R}^*$, as indicated in the legend. To account for smearing of the resonances by disorder, we added a damping term, $-\delta\hat n({\bf k},t)/\tau_s$, to the right-hand side of the kinetic equation \eqref{kinetic_edsr}. In all panels of Fig.~\ref{Peaks}, $1/\tau_s=0.04\lambda^*_{\text{SOC}}$, where $\lambda^*_{\text{SOC}}=\sqrt{\lambda_\mathrm{R}^{*2}+\lambda_\mathrm{VZ}^{*2}}$. For $\lambda_\mathrm{R}^*=15.0$\,meV and $\lambda_\mathrm{VZ}^*=7.5$\,meV, the spin relaxation time is $1/\tau_s=1$\,ps. Reprinted with permission from Ref.~\onlinecite{Kumar_TMD:2021}. Copyright 2021 by the American Physical Society.}
\end{figure*}
The zero-field ESR and EDSR in graphene with proximity-induced spin-orbit coupling are predicted to have some interesting features.\cite{Kumar_TMD:2021} In the presence of electron-electron interactions and both types of SOC (Rashba and valley-Zeeman), both ESR and EDSR signals consist of two peaks, centered at the frequencies of coupled oscillations of the uniform and valley-staggered magnetizations, see panels $a$ and $c$ in Fig.~\ref{Peaks}. Splitting of the resonance occurs as long as the Landau function in Eq.~\eqref{FL} has more than just the $m=0$ harmonic in the spin-exchange and spin-valley-exchange channels, which is always the case for graphene. Next, even if the (interacting) system has only one type of SOC, there are still two resonance modes, $\Omega_{+}$ and $\Omega_{-}$, but one of them is both ESR- and EDSR-silent, because the spectral weights of the corresponding resonances vanish. This effect is illustrated in panels $b$ and $d$ of Fig.~\ref{Peaks}. Another interesting feature is that the two ESR (EDSR) peaks have different (comparable) magnitudes. Therefore, EDSR is better way to probe the two-peak structure of the resonance.
It is worth pointing out that the relative strengths of the Rashba and VZ components of SOC in graphene on TMD substrates is currently an open issue. While weak antilocalization experiments on monolayer graphene find VZ SOC to be much stronger than the Rashba one,\cite{wakamura:2018,schonenberger:2018,wakamura:2019} the opposite conclusion is reached in, e.g., Refs.~\onlinecite{morpurgo_2015,morpurgo_2016,Yang:2016,Omar:2018}. On the other hand, strong evidence for Rashba SOC being the dominant type in bilayer graphene on WSe$_2$ follows from the dependence of the splitting of the ShdH frequencies on the carrier number density.\cite{morpurgo_2016}
Without getting deeper into this discussion, we note that the ESR and EDSR experiments can be used as an independent test for the dominant type of SOC. Indeed, the coupling between the electric field and electron spins is possible only due to Rashba SOC. Therefore, if the experiment shows no EDSR signal, while the ESR signal contains only a single peak, this would be a clear indication that VZ SOC is the dominant mechanism. On the contrary, if single peaks (at the same frequency) are observed both by EDSR and ESR, this would indicate that Rashba SOC is the dominant mechanism. Finally, if both ESR and EDSR signals are split into two peaks, this would indicate that the Rashba and VZ types of SOC are of comparable strength. A quantitative analysis of the signal shape would allow one not only to obtain the spin-orbit coupling constants, but also to extract up to six FL parameters in the $m=0,1,2$ angular momentum channels, which are hard, if at all possible, to be deduced from other types of measurements.
\paragraph{Strong magnetic field.} The opposite case of a strong (compared to SOC) in-plane magnetic field was analyzed in Ref.~\onlinecite{raines:2022} at $q=0$ and at finite $q$ in Ref.~\onlinecite{ullrich:2021}. If the effect of SOC on the spectrum of collective modes is neglected, the latter consists of the Silin modes, corresponding to oscillations of the uniform magnetization with frequencies as in Eq. \eqref{silin_freq}, and of an additional set of modes, corresponding to oscillations of the valley-staggered magnetization with frequencies
\begin{eqnarray}
\tilde\Omega_m=\frac{1+H_m^\perp}{1+F_0^a}\Delta_\mathrm{Z},
\end{eqnarray}
where $H_m^\perp$ is the $m^{\text{th}}$ harmonic of the function $H^\perp(\vartheta)$ in Eq.~\eqref{FL}.
In the absence of SOC, an external magnetic field couples only to the $m=0$ Silin mode while the electric field does not couple to either of the modes. If both Rashba and valley-Zeeman types of SOC are present and, in addition, the Dirac point is gapped due to the breaking of the A-B symmetry of the honeycomb lattice by the substrate, the external electric field couples to the $m=0$ and $m=2$ Silin modes, and to the $m=0$ and $m=1$ valley-staggered modes. Therefore, the EDSR spectrum consists, in general, of four peaks. In addition, the resonances occurs not only in the longitudinal conductivity, but also in the transverse (Hall) one, although the external magnetic field does not affect the electron orbits. This last effect occurs due to the Berry curvature of the gapped Dirac point, and its mechanism can be understood already for non-interacting electrons as follows.\cite{raines:2022}
Initially, all particle spins are polarized along the external magnetic field, which we take to be along the $\hat{\bf x}$ axis.
Upon application of an external field $\mathbf{E}_{\text{em}}(t)$ the particle spins feel an effective Rashba magnetic field
$\mathbf{B}_{\text{R}} \propto \hat{\bf z} \times \mathbf{j}$, where $ \mathbf{j}$ is the electric current density, and therefore experience a spin torque
$
\mathbf{T} \propto \hat{\bf x} \times \mathbf{B}_{\text{R}} \propto \hat{\bf z} j_x.
$
The $x$ component of the current $j_x$ is composed of
regular and anomalous pieces, shown in the left of Fig.~\ref{fig:diagram},
\begin{equation}
j_x(\omega) = -
\frac{e^2 n E_{\text{em},x}}{i\omega m^*} - e^2N{\mathcal B}_z(k)\label{jx}
E_{\text{em},y},
\end{equation}
where $n$ is the number density, $m^*=k_F/\epsilon'_k\vert_{k=k_F}$ is the effective mass, and $\boldsymbol{\mathcal B}(k)=\mp v_\mathrm{D}^2\Delta/2(v_\mathrm{D}^2k^2+\Delta^2)^{3/2}\hat{\bf z}$ is the Berry curvature of the gapped Dirac points in the $K$ ($K'$) valleys.
The first term in Eq.~\eqref{jx} creates identical torques
in both valleys, while the
second one, being proportional to the Berry curvature, yields valley-staggered torques depicted in the right panel of Fig.~\ref{fig:diagram}.
The component of $\mathbf{E}_{\text{em}}$ along $\mathbf{B}$
causes a valley-uniform torque on the spin, exciting the
Silin mode spin, while
the component of $\mathbf{E}_{\text{em}}$ transverse to $\mathbf{B}$
causes a valley-staggered torque,
and thus excites the valley-staggered spin mode.
Because the charge-to-spin conversion in both cases is proportional to the Rashba coupling, this leads to a term in the conductivity proportional to $\alpha_R^2$.
Furthermore, the Silin mode contributes to $\sigma_{xx}$, while the valley-staggered mode contributes to $\sigma_{xy}$.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{raines.pdf}
\caption{ \emph{Left}: Electric-dipole spin resonance in graphene with proximity-induced spin-orbit coupling in the presence of strong in-plane magnetic field. The regular, $\mathbf{j}_0 \propto \mathbf{E}_{\text{em}}$, and anomalous, $\mathbf{j}_a \propto \boldsymbol{\mathcal{B}}\times \mathbf{E}_{\text{em}}$, currents at the $K$ (orange) and $K'$ (purple) points in the Brillouin zone, induced by
the electric field $\mathbf{E}_{\text{em}}$ of the incident electromagnetic wave.
Here, $\boldsymbol{\mathcal{B}}$ is the valley-staggered Berry curvature.
\emph{Right}: Anomalous fields and torques.
Spins are initially polarized along the static magnetic field $\mathbf{B}$. The anomalous current-induced effective Rashba fields, $\mathbf{B}_{\text{R},a}\propto \lambda_\mathrm{R} \left(\hat{\bf z} \times \mathbf{j}_a\right)$,
produce
valley-specific
torques $\mathbf{T}_a \propto \mathbf{B} \times \mathbf{B}_{\text{R},a}$,
thus exciting the valley-staggered spin mode with an intensity proportional to $|\mathbf{E}_{\text{em}} \times \mathbf{B}|$. Reprinted with permission from Ref.~\onlinecite{raines:2022}. Copyright 2022 by the American Physical Society.
\label{fig:diagram}}
\end{figure}
\section{Conclusions}
\label{sec:concl}
In this review, we summarized recent progress in theoretical understanding and experimental observation of a new type of collective spin modes, arising from an interplay between spin-orbit coupling (SOC) and electron-electron interaction. We focused on three types of real systems: i) a two-dimensional (2D) electron gas (2DEG) with Rashba and/or Dresselhaus SOC, ii) graphene with proximity-induced SOC, and iii) the Dirac state on the surface of a three-dimensional topological insulator. Provided that SOC and/or external magnetic field are weak, i.e., the corresponding energy scales are much smaller than the Fermi energy, collective modes in systems i) and ii) can be described within the single-valley or two-valley versions of the Fermi-liquid (FL) theory, respectively. A transparent physical picture of such collective modes arises due to mapping of a kinetic equation for a 2D FL onto an effective tight-binding model for an artificial one-dimensional lattice, whose sites are labeled by the projections of angular momentum on the normal to the 2DEG plane ($m$). Rashba SOC plays the role of on-site energies, while Zeeman and Dresselhaus terms correspond to hopping between the nearest and next-to-nearest neighbors, respectively, whereas the $m$-dependent components of the Landau interaction function create ``defects'' of both on-site and bond types. Within this mapping, the continuum of particle-hole excitations plays a role of the conduction band, while collective modes are the bound states produced by defects.
We discussed the results of recent Raman experiments on Cd$_{1-x}$Mn$_x$Te quantum wells\cite{perez:2007, baboux:2013,baboux:2015,perez:2016,perez:2017,perez:2019} and the Dirac state on the surface of Bi$_2$Se$_3$,\cite{kung:2017} in which some of the predicted collective modes have been observed, and formulated predictions for future electron spin resonance (ESR) and electric-dipole spin resonance (EDSR) experiments on graphene with proximity-induced SOC.
The new type of collective modes, discussed in this paper, may have potential applications in spintronics, magnonics, optoelectronics, and quantum sensing. Indeed, such modes can be thought of as massive ``particles'', with masses fixed by the FL interaction, moving in a potential profile produced by SOC.\cite{ashrafi:2012} By modulating the strength of SOC along the plane of motion, e.g., by gating, one can confine the modes to waveguides and use them to transmit signals. Despite inherent disorder and other sources of damping, the so far observed modes of this type are quite sharp and robust; for example, the collective mode on the surface of Bi$_2$Se$_3$ is observed up to 300\,K.\cite{kung:2017}
\acknowledgments
It is our great pleasure and honor to dedicate this paper to the 95th birthday of Professor Emmanuel Iosifovich Rashba, whose seminal work on spin-orbit effects in solids laid the foundation of modern spintronics.
One of us (D.L.M.) has had a pleasure of knowing Emmanuel Iosifovich for a long time and have been fortunate to collaborate with him on Ref.~\onlinecite{ashrafi:2013}, which was one of the most gratifying collaborations in his (D.L.M.'s) career. Emmanuel Iosifovich impresses anyone whom he interacts with even briefly by the clarity of his thoughts, encyclopedic knowledge of physics, acute attention to details, and outward friendliness.
We are grateful to our co-authors A. Ashrafi (Magine), G. Blumberg, L. Glazman, A. Goyal, S. Kung, Z. Raines, E. Rashba, and P. Sharma for collaborating with us on the projects discussed in this paper, and to H. Bouchiat, C. Bowers, A. Chubukov, S. Gu{\'e}ron, F. Perez, C. Ullrich,
E. Sherman, O. Starykh,
and T. Wakamura for stimulating discussions. This work was supported by the National Science Foundation under Grant No. NSF DMR-1720816 (D.L.M.), the Department of Energy (DOE) Basic Energy Sciences under Grant No. DE-SC0020353 (A.K.), and the Natural Sciences and Engineering Research Council of Canada (NSERC) Grant No. RGPIN-2019-05486 (S.M.).
\newpage
|
\section{Introduction}\label{sec:intro}
Let $\mathcal{X},\mathcal{Y}$ be real, finite-dimensional Hilbert spaces. In this work, we consider \emph{saddle-point problems} of the form
\begin{equation}\label{eq:saddle}
\min_{x\in\mathcal{X}}\max_{y\in\mathcal{Y}} \Phi(x,y) := \psi(x) + f(x,y) - \zeta(y)
\end{equation}
where
\begin{itemize}
\item $f\colon\mathcal{X}\times\mathcal{Y}\to\mathbb{R}$ is convex-concave and continuously differentiable.\label{ass:f}
\item $\psi\colon\mathcal{X}\to(-\infty,+\infty],\zeta\colon\mathcal{Y}\to(-\infty,+\infty]$ are proper, lower semicontinuous (l.s.c.), and convex.
\end{itemize}
Since many non-smooth optimisation problems can be cast in the form of~\eqref{eq:saddle}, it is a useful and heavily studied tool in itself \cite{adolphs2019local, DURR2013540, akimoto2021saddle, lyashko2011low, nesterov2007dual, hamedani2018optimization}. However, rather than attempt to solve~\eqref{eq:saddle} in its current form, it is far more convenient, even within the afrorementioned references, to follow in the steps of Korpelevi\v{c} \cite{korpelevich1976extragradient} and Popov \cite{popov1980modification}, by casting it as the \emph{Variational Inequality (VI)}:
\begin{equation}\label{eq:VI}
\text{Find $z^*\in\Hilbert$ such that~} \langle F(z^*),z-z^*\rangle + g(z) - g(z^*) \geq 0 \quad \forall z\in\Hilbert,
\end{equation}
where
\begin{equation}\label{eq:saddle to VI}
z=(x,y)\in\Hilbert:=\mathcal{X}\oplus\mathcal{Y},\quad F(z) = (\nabla_x f(x,y), -\nabla_y f(x,y)), \quad g(z) = \psi(x) + \zeta(y),
\end{equation}
and the variable $z^*:=(x^*,y^*)$ shown in~\eqref{eq:VI} characterises the solution $(x^*,y^*)$ to~\eqref{eq:saddle}.
Many methods (see, for instance, \cite{Pang1982IterativeMF, marcotee1995projection, fukishima1992equivalent}) for solving~\eqref{eq:VI} require global Lipschitz continuity of the operator $F$. However, this assumption is often too strong to hold in practice. Even when $F$ is globally Lipschitz continuous, knowledge of its Lipschitz constant is usually required as input to the chosen algorithm and determining this constant is typically more difficult than solving the original problem. Moreover, even if $F$ is globally Lipschitz and its global Lipschitz constant is known, then, as the step-size is inversely related to the Lipschitz constant, a constant step-size rule can be too conservative. This is particularly unnecessary if the generated sequence lies entirely within a region where a \emph{local} Lipschitz constant is small (relative to the size of the global constant).
Therefore, it is beneficial to instead define a step-size sequence which attempts to approximate a local Lipschitz constant with respect to the point iterates. The standard approach then is to generate a step-size sequence via a \emph{backtracking} procedure (see \cite{malitsky2020forward, bellocruz2015variant, censor1998interior, malitsky2018proximal, iusem1997variant, malitsky2020projected} and references therein). While avoiding each of the shortcomings listed above, such methods can become expensive when considering the overall run-time of the algorithm, due to the arbitrarily large number of steps taken during the backtracking procedure within each iteration. An emerging alternative is that of \emph{adaptive step-sizes} \cite{malitsky2020golden, malitsky2020adaptive, alacoglu2020convergence}, which accomplish the same goals as backtracking methods without the need for backtracking, ie, the step-size update is fully explicit. In particular, the \emph{adaptive Golden RAtio ALgorithm (aGRAAL)} \cite{malitsky2020golden} (as stated in Algorithm~\ref{alg:aGRAAL}), named as such because of it's relationship with the \textit{Golden Ratio} $\varphi=\frac{1+\sqrt{5}}{2}$, solves~\eqref{eq:VI}, and is the method we focus on here.
\begin{algorithm}[!htb]
\caption{The Adaptive Golden RAtio ALgorithm (aGRAAL) \cite{malitsky2020golden}.}\label{alg:aGRAAL}
\SetKwInput{Init}{Initialisation}
\KwIn{Initial points $z_0,\overline{z}_0\in\Hilbert$, initial step-size $\lambda_0>0$, $\lambda_{\max}\gg 0$, $\phi\in(1,\varphi]$.}
\Init{Set $z_1=\overline{z}_0$, $\theta_0=1$, $\rho=\frac{1}{\phi}+\frac{1}{\phi^2}$.}
\For{$k=1,2,\dots$}{
Calculate the step-size:
\begin{equation}
\lambda_k = \min\left\{\rho\lambda_{k-1}, \frac{\phi\theta_{k-1}}{4\lambda_{k-1}}\frac{\|z_k - z_{k-1}\|^2}{\|F(z_k) - F(z_{k-1})\|^2}, \lambda_{\max}\right\}.
\end{equation}\\
Compute the next iterates:
\begin{align}\\
\overline{z}_k &= \frac{(\phi - 1)z_k + \overline{z}_{k-1}}{\phi}\\
z_{k+1} &= \prox_{\lambda_k g}(\overline{z}_k-\lambda_k F(z_k))\label{eq:aGRAAL prox}
\end{align}
Update: $\theta_k = \frac{\lambda_k\phi}{\lambda_{k-1}}.$
}
\end{algorithm}
One other way to potentially improve methods is to replace the Euclidean distance in the proximal operator with a non-Euclidean family of distance-like functions called the \emph{Bregman distance}. Such methods, for solving~\eqref{eq:VI}, can be found in existing literature \cite{censor1998interior, nomirovskii2019convergence, vanhieu2022modified, gibali2018new, gibali2020fast, sym12122007, jolaoso2020weak, jolaso2021single, vanhieu2020two}. Interestingly, most of these methods require a Lipschitz assumption but don't require knowledge of the Lipschitz constant. However, these employ a backtracking procedure and/or a non-increasing step-size sequence, whereas the step-size of our new method is fully explicit and is allowed to increase slightly at each iteration.
In this paper, we investigate Bregman modifications to Algorithm~\ref{alg:aGRAAL}. To this end, we begin by proposing the \textit{Bregman-Golden RAtio ALgorithm (B-GRAAL)}, a Bregman version of the fixed step-size \textit{Golden RAtio ALgorithm (GRAAL)}, and prove convergence of our new method in full. We then present an adaptive version of B-GRAAL, or similarly, a Bregman modification to Algorithm~\ref{alg:aGRAAL}, which we refer to as the \textit{Bregman-adaptive Golden RAtio ALgorithm (B-aGRAAL)}. Although we only provide a convergence analysis of B-aGRAAL in a restrictive setting, we observe it to work numerically outside this setting.
One advantage of our new method is the flexibility provided by the Bregman proximal operator. In the context of convex-concave games, for instance, this modified operator arises as the projection onto the probability simplex which has a simple closed form expression with respect to the \emph{Kullback--Leibler (KL) divergence} but not with respect to the standard Euclidean distance. In fact, the Euclidean projection requires an $O(n\log n)$ time algorithm in $n$-dimensions \cite{wang2012projection, chen2011projection, dai2022distributed}, whereas the KL projection only requires $O(n)$ time (see, for instance \cite[Section 5]{BECK2003167} and \cite[Section 4.4]{NIPS2015_f60bb6bb}). Another advantage of these modifications in the constrained optimisation case, is that it is sometimes possible to choose a Bregman distance whose domain is the constraint set so as to make the feasibility of the iterates implicit.
The remainder of this paper is structure as follows. In Section~\ref{sec:prelim}, we collect preliminary results for use in our analysis. In Section~\ref{sec:bgraal}, we present our fixed step-size method and a proof of convergence. In Section~\ref{sec:bagraal}, we present our adaptive method with some partial analysis. Section~\ref{sec:exp} contains experimental results. Firstly, we compare the fixed step-size method with the Euclidean distance and the KL divergence on a matrix game between two players. Secondly, we make the same comparison for the adaptive method on power allocation problem in a Gaussian communication channel. Finally, we apply the adaptive method to an $N$-person Cournot oligopoly model with appropriately chosen Bregman distances over a closed box. We then conclude this paper by presenting some directions for further research.
\section{Preliminaries}\label{sec:prelim}
Throughout this work, $\Hilbert$ denotes a real, finite-dimensional Hilbert space with inner-product $\langle\cdot,\cdot\rangle$ and induced norm $\|\cdot\|$. Given an extended real-valued function $f\colon\Hilbert\to(-\infty,+\infty]$, its \emph{domain} is denoted $\dom f := \{x\in\Hilbert\colon f(x)<+\infty\}$. Its \emph{subdifferential} at $x\in\dom f$ is given by
$$\partial f(x) := \{\nu\in\Hilbert\colon f(x) - f(y) - \langle\nu,x-y\rangle\leq 0\quad\forall y\in\Hilbert\},$$
and defined as $\partial f(x):=\emptyset$ for $x\not\in\dom f$. The \textit{indicator function} of a set $K\subseteq\Hilbert$ is written $\iota_K$ and takes the value $0$ for $x\in K$ and $+\infty$ otherwise.
A proper, l.s.c., convex function $h\colon\Hilbert\to(-\infty,+\infty]$ is called \textit{Legendre}~\cite[Definition~7.1.1]{borwein2010convex} if it is strictly convex on every convex subset of $\dom\partial h:=\{x\in\Hilbert\colon\partial h(x)\neq\emptyset\}$ and differentiable on $\intr\dom h\neq\emptyset$ such that $\|\nabla h(x)\|\to\infty$ whenever $x$ approaches the boundary of $\dom h$. The \textit{convex conjugate} of $h$ written as $h^*\colon\Hilbert\to(-\infty,+\infty]$ is the given by
$$h^*(x^*) := \sup_{x\in\Hilbert}\{\langle x^*,x\rangle - h(x)\}.$$
When $h$ is merely differentiable on $\intr\dom h$, the \emph{Bregman distance} generated by $h$ is the function $D_h\colon \Hilbert\times\intr\dom h\to(-\infty,+\infty]$ given by
$$D_h(x,y) := h(x) - h(y) - \langle\nabla h(y),x-y\rangle.$$
When $h$ is also convex, $D_h$ is non-negative and, when $h$ is $\sigma$-strongly convex, $D_h$ satisfies $$ D_h(x,y)\geq\frac{\sigma}{2}\|x-y\|^2\quad \forall (x,y)\in \Hilbert\times\intr\dom h.$$
We begin by collecting some general properties of the Bregman distance.
\begin{proposition}[Properties of the Bregman distance]\label{prop:prop}
Let $h\colon\Hilbert\to(-\infty,+\infty]$ be Legendre. Then the following assertions hold.
\begin{enumerate}[(a)]
\item\label{item:3} \emph{(three point identity)}
For all $x,y\in\intr\dom h$ and $z\in\dom h$, we have
$$D_h(z,x) - D_h(z,y) - D_h(y,x) = \langle\nabla h(x) - \nabla h(y), y-z\rangle.$$
\item\label{item:dual} For all $x,y\in\intr\dom h$, we have
$$D_h(x,y) = D_{h^*}(\nabla h(y), \nabla h(x))$$
\item\label{item:identity}
For all $x\in\dom h$ and $y,u,v\in\intr\dom h$ such that $\nabla h(y) = \alpha\nabla h(u) + (1-\alpha)\nabla h(v)$ for some $\alpha\in\mathbb{R}$, we have
$$D_h(x,y) = \alpha\Big[D_h(x,u) - D_h(y,u)\Big] + (1-\alpha)\Big[D_h(x,v) - D_h(y,v)\Big].$$
\end{enumerate}
\end{proposition}
\begin{proof}
\ref{item:3}~See, for instance, \cite[Lemma 2.2]{teboulle2018simplified} and the paragraph immediately after.
\ref{item:dual}~See, for instance, \cite[Theorem 3.7(v)]{bauschke1997legendre}.
\ref{item:identity}~By using the definition of $D_h$, together with the assumption $\nabla h(y) = \alpha\nabla h(u) + (1-\alpha)\nabla h(v)$, we obtain
\begin{align*}
D_h(x,y) &= h(x) - h(y) - \langle \nabla h(y),x-y\rangle\\
&= \alpha\Big[h(x) - h(y) - \langle\nabla h(u),x-y\rangle\Big] + (1-\alpha)\Big[h(x) - h(y) - \langle\nabla h(v),x-y\rangle\Big]\\
&= \alpha\Big[h(x) - h(u) - \langle\nabla h(u), x-u\rangle - h(y) + h(u) - \langle\nabla h(u),u-y\rangle\Big]\\
& \qquad + (1-\alpha)\Big[h(x) - h(v) - \langle\nabla h(v), x-v\rangle - h(y) + h(v) - \langle\nabla h(v),v-y\rangle\Big] \\
&= \alpha\Big[D_h(x,u) - D_h(y,u)\Big] + (1-\alpha)\Big[D_h(x,v) - D_h(y,v)\Big].
\end{align*}
This completes the proof.
\end{proof}
\begin{remark}
When $h=\frac{1}{2}\|\cdot\|^2$, Proposition~\ref{prop:prop}\ref{item:identity} recovers the established Euclidean identity
$$\forall x,y\in\Hilbert, \alpha\in\mathbb{R} \qquad \|\alpha x + (1-\alpha)y\|^2 = \alpha\|x\|^2 + (1-\alpha)\|y\|^2 - \alpha(1-\alpha)\|x-y\|^2.$$
\end{remark}
We now turn our attention to operators defined in terms of the Bregman divergence. The \emph{(left) Bregman proximal operator} of a function $f\colon\Hilbert\to(-\infty,+\infty]$ is the (potentially set-valued) operator given by
\begin{equation}\label{eq:breg prox def}
\prox_f^h(y) := \argmin_{x\in\Hilbert}\left\{f(x) + D_h(x,y)\right\}\quad\forall y\in\intr\dom h.
\end{equation}
Since we will only require the left Bregman proximal operator in this work, will omit the qualifier ``left'' from here on-wards. For further details on the analogous ``right Bregman proximal operator'', the reader is referred to \cite{borwein2011characterization}. The \emph{(left) Bregman projection} onto $C$ is the (left) Bregman proximal operator of $\iota_C$, that is,
$$P^h_C(y) := \prox_{\iota_C}^h(y) = \argmin_{x\in C} D_h(x,y)\quad\forall y\in\intr\dom h.$$
Next, we collect properties of the Bregman proximal operator for use in our subsequence algorithm analysis.
\begin{proposition}[Bregman proximal operator]\label{prop:prox}
Let $f\colon\Hilbert\to(-\infty,+\infty]$ be proper, l.s.c, convex and let $h\colon\Hilbert\to(-\infty,+\infty]$ be Legendre such that $\intr\dom h\cap\dom f\neq\emptyset$.
\begin{enumerate}[(a)]
\item\label{item:range} $\range(\prox^h_f)\subseteq\intr\dom h\cap\dom f$.
\item\label{item:sv} $\prox^h_f$ is single-valued on $\dom (\prox^h_f)\subseteq\intr\dom h$. Moreover, if $h+f$ is supercoercive, then $\dom (\prox^h_f)=\intr\dom h$.
\item\label{item:prox thm} Let $y\in\dom(\prox^h_f)$ and $x=\prox_f^h(y)$. Then, for all $u\in\Hilbert$, we have
\begin{equation}\label{eq:prox thm}
f(x) - f(u) \leq \langle\nabla h(y) - \nabla h(x),x-u\rangle.
\end{equation}
\item\label{item:bfne}
Let $y,y'\in\dom(\prox^h_f)$, $x=\prox_f^h(y)$ and $x'=\prox_f^h(y')$. Then
\begin{equation}\label{eq:bfne}
0\leq\langle \nabla h(x) - \nabla h(x^\prime), x-x^\prime\rangle \leq \langle \nabla h(y) - \nabla h(y^\prime), x-x^\prime\rangle.
\end{equation}
\end{enumerate}
\end{proposition}
\begin{proof}
\ref{item:range}~See \cite[Proposition~3.23(v)(b)]{bauschke2003bregman}, noting that the sum rule holds for $f$ and $h$ since $\intr\dom h\cap\dom f\neq\emptyset$ and thus $\dom \partial\left(f+h\right) = \dom\left(\partial f + \nabla h\right) \subseteq \dom f\cap\intr\dom h$.
\ref{item:sv}~The first part follows by combining (a) and \cite[Proposition 3.22 (ii)(d)]{bauschke2003bregman}. For the second part, see \cite[Proposition~3.21(vii)]{bauschke2003bregman}.
\ref{item:prox thm}~The first order optimality condition together with (a) implies $\nabla h(y)-\nabla h (x)\in\partial f(x)$ which establishes the inequality. \ref{item:bfne}~The first inequality in \eqref{eq:bfne} follows from convexity of $h$. To show the second, we apply~\eqref{eq:prox thm} with $u=x^\prime$ to see
$$f(x)-f(x^\prime)\leq\langle\nabla h(y)-\nabla h(x),x-x^\prime\rangle,$$
and similarly,
$$f(x^\prime) - f(x)\leq\langle\nabla h(y^\prime) - \nabla h(x^\prime),x^\prime-x\rangle.$$
Then adding these inequalities gives the desired result.
\end{proof}
\begin{remark}
Parts \ref{item:range}, \ref{item:sv}, \ref{item:bfne} of Proposition~\ref{prop:prox} also apply to the \emph{Bregman resolvent} \cite{borwein2011characterization}, which is defined for a set-valued operator $A\colon\Hilbert\rightrightarrows\Hilbert$ as $R^h_{A} = \left(\nabla h + A\right)^{-1}\circ\nabla h$. To see that the resolvent generalises the proximal operator, we refer to \cite[Proposition 3.22 (ii)(a)]{bauschke2003bregman}.
\end{remark}
\section{The Bregman-Golden Ratio Algorithm}\label{sec:bgraal}
In this section, we consider the \textit{Variational Inequality (VI)} problem:
\begin{equation}\label{eq:VI2}
\text{Find~}z^*\in\Hilbert\text{~such that~}\langle F(z^*), z-z^*\rangle + g(z) - g(z^*) \geq 0 \quad \forall z\in\Hilbert,
\end{equation}
where we assume that
\begin{enumerate}[\textbf{A.\arabic*}]
\item \label{ass:g} $g\colon\Hilbert\to(-\infty,+\infty]$ is proper, l.s.c., convex.
\item\label{ass:h} $h\colon\Hilbert\to(-\infty,+\infty]$ is continuously differentiable, Legendre, and $\sigma$-strongly convex. In addition, we will also require that $D_h(x,x_n)\to 0$ for every sequence $(x_n)\subseteq\intr\dom h$ that converges to some $x\in\dom h$.
\item \label{ass:F} $F\colon\Hilbert\to\Hilbert$ is monotone over $\dom g\cap\intr\dom h\neq\emptyset$.
\item \label{ass:S} $\Omega:=S\cap\dom h \neq \emptyset$ where $S$ denotes the solution set of~\eqref{eq:VI2}.
\end{enumerate}
\begin{remark}
Assumption~\ref{ass:h} is common in the literature concerning Bregman first-order methods~\cite{teboulle2018simplified, Chen1993ConvergenceAO, bauschke2017descent, bauschke2019linear}. In particular, the limit condition holds when $\nabla h$ is continuous and $\dom h$ is open. Indeed, in this case, $\sigma$-strong convexity of $h$ implies that $h^*$ is $\frac{1}{\sigma}$-smooth \cite[Theorem 6]{kakade2009duality}, and so applying Proposition~\ref{prop:prop}\ref{item:dual} gives
$$D_h(x,x_n)=D_{h^*}\left(\nabla h(x_n), \nabla h(x)\right) \leq \frac{1}{2\sigma}\|\nabla h(x_n) - \nabla h(x)\|^2 \to 0.$$
The significance of $\dom h$ being open here is that $h$ is differentiable at $x\in\dom h$, however we observe that the same condition can still hold if $\dom h$ is closed. In particular, it also holds for the KL divergence (see, for instance \cite[Example 2.1]{Chen1993ConvergenceAO}).
\end{remark}
Our proposed algorithm for \eqref{eq:VI2}, the \emph{Bregman-Golden RAtio Algorithm (B-GRAAL)}, is stated in Algorithm~\ref{alg:B-GRAAL}. Once again, $\varphi:=\frac{1+\sqrt{5}}{2}$ denotes the \emph{Golden Ratio}, which satisfies $\varphi^2=\varphi+1$.
\begin{algorithm}[!htb]
\caption{The Bregman-Golden RAtio ALgorithm (B-GRAAL)}\label{alg:B-GRAAL}
\KwIn{Initial points $z_1,\overline{z}_0\in\intr\dom h$ and a step-size $\lambda \in \left(0,\frac{\sigma\varphi}{2L}\right]$.}
\For{$k=1,2,\dots$}{
Compute the next iterates:
\begin{align}
\overline{z}_k &= (\nabla h)^{-1}\left(\frac{(\varphi-1)\nabla h(z_k) + \nabla h(\overline{z}_{k-1})}{\varphi}\right) \label{eq:breg-h}\\
z_{k+1} &=\argmin_{z\in\Hilbert}\left\{\langle F(z_k), z-z_k\rangle + g(z) + \frac{1}{\lambda} D_h(z,\overline{z}_k)\right\}.\label{eq:breg-prox}
\end{align}
}
\end{algorithm}
The following lemma establishes the well-definedness of the sequences generated by the Bregman-GRAAL.
\begin{lemma}\label{lem:contained}
Suppose Assumptions~\ref{ass:g}-\ref{ass:h} hold. Then the sequences $(\overline{z}_k)$ and $(z_k)$ generated by Algorithm~\ref{alg:B-GRAAL} are well-defined. Moreover, $(\overline{z}_k)\subseteq\intr\dom h$ and $(z_k)\subseteq\intr\dom h\cap \dom g$.
\end{lemma}
\begin{proof}
Suppose by way of induction that $z_{k},\overline{z}_{k-1}\in\intr\dom h$ for some $k\geq 1$.
Since the gradient $\nabla h\colon\intr\dom h\to\intr(\dom h^*)$ is a bijection~\cite[Theorem~7.3.7]{borwein2010convex}, it follows that $\nabla h(z_k),\nabla h(\overline{z}_{k-1})\in\intr(\dom h^*)$. As $\intr(\dom h^*)$ is a convex set, \eqref{eq:breg-h} implies that $\nabla h(\overline{z}_k) \in \intr(\dom h^*)$ which establishes that $\overline{z}_k\in\intr\dom h$.
Next, we observe that $z_{k+1} = \prox_{\lambda f}^h(\overline{z}_k)$, where $f(z):=\langle F(z_k), z-z_k\rangle + g(z)$. Note that $\dom f=\dom g$. Since $\lambda f+h$ is $\sigma$-strongly convex, it is supercoercive by \cite[Corollary~11.16]{bauschke2011convex} and so Proposition~\ref{prop:prox}\ref{item:range}-\ref{item:sv} shows that $\prox^h_{\lambda f}$ is single-valued with $\range(\prox^h_{\lambda f})\subseteq\intr\dom h\cap\dom g$ and therefore $z_{k+1}\in\intr\dom h\cap\dom g$.
\end{proof}
\begin{remark}
The Bregman proximal step shown in~\eqref{eq:breg-prox} can be expressed in terms of the Bregman proximal operator: $z_{k+1}=\prox_{\lambda f}^h(\overline{z}_k)$, where $f(z) = \langle F(z_k), z-z_k\rangle + g(z)$. Equivalenty, $z_{k+1} = \prox_{\lambda g}^h(\left(\nabla h\right)^{-1}(\nabla h(\overline{z}_k) - \lambda F(z_k)))$, due to the first-order optimality condition in Proposition~\ref{prop:prox}\ref{item:prox thm}.
\end{remark}
The following lemma is key in our convergence analysis.
\begin{lemma}\label{lem:dec}
Suppose Assumptions~\ref{ass:g}-\ref{ass:S} hold and that $F$ is $L$-Lipschitz continuous on $\dom g\cap\intr\dom h$. Let $z^*\in \Omega$ be arbitrary. Then the sequences $(z_k), (\overline{z}_k)$ generated by Algorithm~\ref{alg:B-GRAAL} satisfy
\begin{multline}\label{eq:energy}
0 \leq (\varphi+1)D_h(z^*, \overline{z}_{k+1}) + \frac{\varphi}{2}D_h(z_{k+1}, z_k) - \varphi D_h(z_{k+1}, \overline{z}_{k+1}) \\
\leq (\varphi+1)D_h(z^*, \overline{z}_k) + \frac{\varphi}{2}D_h(z_k, z_{k-1}) - \varphi D_h(z_k, \overline{z}_k) - \left(1-\frac{1}{\varphi}\right)D_h(z_{k+1}, \overline{z}_k).
\end{multline}
\end{lemma}
\begin{proof}
By first applying Proposition~\ref{prop:prox}\ref{item:prox thm} with $f(z) := \lambda(\langle F(z_k), z-z_k\rangle + g(z))$, $u:=z\in\dom h\cap \dom g$ arbitrary, $x:=z_{k+1}$ and $y:=\overline{z}_k$, followed by the three point idenitity (Proposition~\ref{prop:prop}\ref{item:3}) we obtain
\begin{equation}\label{eq:prox-ineq}
\begin{aligned}
\lambda\left(\langle F(z_k),z_{k+1}-z\rangle + g(z_{k+1}) - g(z)\right) &\leq \langle \nabla h(\overline{z}_k) - \nabla h(z_{k+1}), z_{k+1} - z\rangle\\
&= D_h(z, \overline{z}_k) - D_h(z, z_{k+1}) - D_h(z_{k+1}, \overline{z}_k).
\end{aligned}
\end{equation}
Shifting the index in~\eqref{eq:prox-ineq} (by setting $k\equiv k-1$), setting $z:=z_{k+1}$, and using $\varphi\nabla h(\overline{z}_k) = (\varphi - 1)\nabla h(z_k) + \nabla h(\overline{z}_{k-1})$ followed by the three point identity (Proposition~\ref{prop:prop}\ref{item:3}) gives
\begin{equation}\label{eq:shift prox-ineq}
\begin{aligned}
&\lambda\left(\langle F(z_{k-1}), z_k - z_{k+1}\rangle + g(z_k) - g(z_{k+1})\right) \\
&\quad \leq \langle \nabla h(\overline{z}_{k-1}) - \nabla h(z_k), z_k - z_{k+1}\rangle\\
&\quad= \varphi\langle\nabla h(\overline{z}_k) - \nabla h(z_k), z_k - z_{k+1}\rangle\\
&\quad =\varphi\left[ D_h(z_{k+1}, \overline{z}_k) - D_h(z_{k+1}, z_k) - D_h(z_k, \overline{z}_k)\right].
\end{aligned}
\end{equation}
Let $z^*\in S\cap\dom h$. Setting $z:=z^*$ in~\eqref{eq:prox-ineq}, summing with~\eqref{eq:shift prox-ineq} and rearranging yields
\begin{equation}\label{eq:add prox-ineqs}
\begin{aligned}
\lambda\left(\langle F(z_k), z_k - z^*\rangle + g(z_k) - g(z^*)\right) \leq
& D_h(z^*,\overline{z}_k) - D_h(z^*,z_{k+1}) - D_h(z_{k+1},\overline{z}_k)\\
&+ \varphi\Big[D_h(z_{k+1},\overline{z}_k) - D_h(z_{k+1},z_k) - D_h(z_k,\overline{z}_k)\Big]\\
&+\lambda \langle F(z_k) - F(z_{k-1}), z_k - z_{k+1}\rangle.
\end{aligned}
\end{equation}
We observe that the left-side of~\eqref{eq:add prox-ineqs} is non-negative as a consequence of~\eqref{eq:VI} and~\ref{ass:F}:
\begin{equation}\label{eq:lhs nonneg}
0\leq \langle F(z^*),z_k-z^*\rangle + g(z_k) - g(z^*)\leq \langle F(z_k),z_k-z^*\rangle + g(z_k) - g(z^*).
\end{equation}
To estimate the final term in \eqref{eq:add prox-ineqs}, we use the Cauchy--Schwarz inequality, $L$-Lipschitz continuity of $F$, $\sigma$-strong convexity of $h$ and the inequality $\lambda\leq\frac{\sigma\varphi}{2L}$ to obtain
\begin{equation}\label{eq:lipschitz}
\begin{aligned}
\lambda\langle F(z_k) - F(z_{k-1}), z_k - z_{k+1}\rangle
&\leq \lambda L\|z_k - z_{k-1}\|\|z_k - z_{k+1}\|\\
&\leq \frac{\lambda L}{2}\left(\|z_k - z_{k-1}\|^2+\|z_k - z_{k+1}\|^2\right)\\
&\leq \frac{\varphi}{2}\left(D_h(z_k, z_{k-1}) + D_h(z_{k+1}, z_k)\right).
\end{aligned}
\end{equation}
Combining~\eqref{eq:add prox-ineqs}, \eqref{eq:lhs nonneg} and~\eqref{eq:lipschitz} gives
\begin{multline}\label{eq:combination}
D_h(z^*, z_{k+1}) \leq D_h(z^*, \overline{z}_k) + (\varphi-1)D_h(z_{k+1}, \overline{z}_k) - \varphi D_h(z_k, \overline{z}_k)\\
- \frac{\varphi}{2}D_h(z_{k+1}, z_k) + \frac{\varphi}{2}D_h(z_k,z_{k-1}).
\end{multline}
Now applying Proposition~\ref{prop:prop}\ref{item:identity} with $\nabla h(z_{k+1}) = \frac{\varphi\nabla h(\overline{z}_{k+1}) - \nabla h(\overline{z}_k)}{\varphi-1} = (\varphi+1)\nabla h(\overline{z}_{k+1}) - \varphi\nabla h(\overline{z}_k)$ and rearranging yields
\begin{multline}\label{eq:identity}
(\varphi+1)D_h(z^*, \overline{z}_{k+1}) = D_h(z^*, z_{k+1}) + (\varphi+1)D_h(z_{k+1}, \overline{z}_{k+1}) \\ + \varphi\Big[D_h(z^*, \overline{z}_k) - D_h(z_{k+1}, \overline{z}_k)\Big].
\end{multline}
Combining \eqref{eq:combination} and \eqref{eq:identity}, followed by collecting like-terms, gives
\begin{multline}\label{eq:central}
(\varphi+1)D_h(z^*, \overline{z}_{k+1}) + \frac{\varphi}{2}D_h(z_{k+1}, z_k) - \varphi D_h(z_{k+1}, \overline{z}_{k+1}) \\
\leq (\varphi+1)D_h(z^*, \overline{z}_k) + \frac{\varphi}{2}D_h(z_k, z_{k-1}) - \varphi D_h(z_k, \overline{z}_k) \\
+ D_h(z_{k+1}, \overline{z}_{k+1}) - D_h(z_{k+1}, \overline{z}_k).
\end{multline}
Since $\nabla h(\overline{z}_{k+1})=\frac{\varphi-1}{\varphi}\nabla h(z_{k+1})+\frac{1}{\varphi}\nabla h(\overline{z}_k)$, Proposition~\ref{prop:prop}\ref{item:identity} gives
\begin{multline}\label{eq:distance}
D_h(z_{k+1}, \overline{z}_{k+1}) = \frac{\varphi-1}{\varphi}\big[D_h(z_{k+1}, z_{k+1}) - D_h(\overline{z}_{k+1}, z_{k+1})\big] + \frac{1}{\varphi}\big[D_h(z_{k+1}, \overline{z}_k) - D_h(\overline{z}_{k+1}, \overline{z}_k)\big] \\\leq \frac{1}{\varphi}D_h(z_{k+1}, \overline{z}_k).
\end{multline}
Combining \eqref{eq:central} and \eqref{eq:distance} establishes the second inequality in~\eqref{eq:energy}. To show the first inequality in~\eqref{eq:energy}, we apply the three point identity (Proposition~\ref{prop:prop}\ref{item:3}) to see that
\begin{equation}
\begin{aligned}\label{eq:3pt2}
D_h(z^*,z_{k+1}) +& D_h(z_{k+1},\overline{z}_{k+1}) = D_h(z^*,\overline{z}_{k+1}) + \langle\nabla h(\overline{z}_{k+1}) - \nabla h(z_{k+1}), z^*-z_{k+1}\rangle\\
&= D_h(z^*,\overline{z}_{k+1}) + \frac{1}{\varphi}\langle\nabla h(\overline{z}_k) - \nabla h(z_{k+1}), z^*-z_{k+1}\rangle.\\
&\leq D_h(z^*, \overline{z}_{k+1}) + \frac{\lambda}{\varphi}\left(\langle F(z_k), z^* - z_{k+1}\rangle - g(z_{k+1}) + g(z^*)\right)\\
&= D_h(z^*, \overline{z}_{k+1}) + \frac{\lambda}{\varphi}\langle F(z_k) - F(z_{k+1}), z^* - z_{k+1}\rangle \\ &\qquad +\frac{\lambda}{\varphi}(\langle F(z_{k+1}), z^* - z_{k+1}\rangle - g(z_{k+1}) + g(z^*)) \\
&\leq D_h(z^*, \overline{z}_{k+1}) + \frac{\lambda}{\varphi}\langle F(z_k) - F(z_{k+1}), z^* - z_{k+1}\rangle.
\end{aligned}
\end{equation}
Using $L$-Lipschitz continuity of $F$ and $\sigma$-strong convexity of $h$ gives
\begin{equation}\label{eq:lipschitz2}
\frac{\lambda}{\varphi}\langle F(z_k) - F(z_{k+1}), z^*- z_{k+1}\rangle \leq \frac{1}{2}D_h(z_{k+1}, z_k) + \frac{1}{2}D_h(z^*, z_{k+1}).
\end{equation}
On substituting~\eqref{eq:lipschitz2} back into~\eqref{eq:3pt2} and rearranging, we obtain
\begin{equation}\label{eq:3pt3}
\begin{aligned}
D_h(z_{k+1}, \overline{z}_{k+1})
&\leq D_h(z^*, \overline{z}_{k+1}) + \frac{1}{2}D_h(z_{k+1}, z_k) -\frac{1}{2}D_h(z^*, z_{k+1})\\
&\leq D_h(z^*, \overline{z}_{k+1}) + \frac{1}{2}D_h(z_{k+1}, z_k),
\end{aligned}
\end{equation}
and therefore
\begin{multline*}
0 \leq D_h(z^*, \overline{z}_{k+1}) + \frac{1}{2}D_h(z_{k+1}, z_k) - D_h(z_{k+1}, \overline{z}_{k+1}) \\
\implies 0\leq (\varphi+1)D_h(z^*, \overline{z}_{k+1}) + \frac{\varphi}{2}D_h(z_{k+1}, z_k) - \varphi D_h(z_{k+1}, \overline{z}_{k+1}),
\end{multline*}
which establishes the first inequality of~\eqref{eq:energy} and thus completes the proof.
\end{proof}
The following is our main result regarding convergence of the Bregman GRAAL with fixed step-size.
\begin{theorem}\label{thm:conv}
Suppose Assumptions~\ref{ass:g}-\ref{ass:S} hold and that $F$ is $L$-Lipschitz continuous on $\dom g\cap\intr\dom h$. Then the sequences $(z_k)$ and $(\overline{z}_k)$ generated by Algorithm~\ref{alg:B-GRAAL} converge to a point in $ S\cap\dom h$.
\end{theorem}
\begin{proof}
Let $z^*\in \Omega$ be arbitrary and let $(\eta_k)$ denote the sequence given by
$$\eta_k := (\varphi+1)D_h(z^*, \overline{z}_k) + \frac{\varphi}{2}D_h(z_k, z_{k-1}) - \varphi D_h(z_k, \overline{z}_k)\quad\forall k\in\mathbb{N}.$$
Lemma~\ref{lem:dec} implies that $\lim_{k\to\infty}\eta_k$ exists and $D_h(z_{k+1}, \overline{z}_k) \to 0$ as $k\to\infty$. Referring to~\eqref{eq:distance}, it follows that $D_h(z_{k+1}, \overline{z}_{k+1}) \to 0$. Also, by applying Proposition~\ref{prop:prop}\ref{item:identity} with the identity $\nabla h(z_k) = (\varphi+1)\nabla h(\overline{z}_k) - \varphi \nabla h(\overline{z}_{k-1})$, we obtain
\begin{align*}
D_h(z_{k+1}, z_k)
&= (\varphi+1)\big[D_h(z_{k+1}, \overline{z}_k)
- D_h(z_k,\overline{z}_k)\big]- \varphi\big[D_h(z_{k+1}, \overline{z}_{k-1}) - D_h(z_k, \overline{z}_{k-1})\big] \\
&\leq (\varphi+1)D_h(z_{k+1},\overline{z}_k)+\varphi D_h(z_k,\overline{z}_{k-1}) \to 0.
\end{align*}
Altogether, we have that
$$ \lim_{k\to\infty}\eta_k = (\varphi+1)\lim_{k\to\infty}D_h(z^*, \overline{z}_k),$$
and, in particular, $\lim_{k\to\infty}D_h(z^*, \overline{z}_k)$ exists.
Next, using $\sigma$-strong convexity of $h$, we deduce that $z_{k+1}-\overline{z}_k\to0$, and that $(z_k)$ and $(\overline{z}_k)$ are bounded. Thus, let $\overline{z}\in\Hilbert$ be a cluster point of $(\overline{z}_k)$. Then there exists a subsequence $(\overline{z}_{k_j})$ such that $\overline{z}_{k_j}\to \overline{z}$ and $z_{k_j+1}\to \overline{z}$ as $j\to\infty$.
Now recalling~\eqref{eq:prox-ineq} gives
$$\lambda\left(\langle F(z_{k_j}),z_{{k_j}+1}-z\rangle + g(z_{{k_j}+1}) - g(z)\right) \leq \langle \nabla h(\overline{z}_{k_j}) - \nabla h(z_{{k_j}+1}), z_{{k_j}+1} - z\rangle \quad\forall z\in\Hilbert,$$
and taking the limit-infimum of both sides as $j\to\infty$ shows that $\overline{z}\in\Omega$. Since $z^*\in\Omega$ was chosen in Lemma~\ref{lem:dec} to be arbitrary, we can now set $z^*=\overline{z}$. It then follows that $\lim_{j\to\infty}D_h(z^*,\overline{z}_{k_j}) = 0$, and consequently, $\lim_{j\to\infty}\eta_{k_j} = 0$. Also note that for $n\geq k_j$, we have $\eta_n\leq\eta_{k_j}$ from Lemma \ref{lem:dec}, and therefore
$$(\varphi+1)\lim_{n\to\infty}D_h(z^*, \overline{z}_n) = \lim_{n\to\infty}\eta_n \leq \lim_{j\to\infty}\eta_{k_j} = 0,$$
and therefore $\overline{z}_k\to z^*$ from strong convexity. The fact that $z_k\to z^*$ follows since $z_k-\overline{z}_k\to 0$.
\end{proof}
\begin{remark}
In the special case where $h=\|\cdot\|^2$, Algorithm~\ref{alg:B-GRAAL} recovers the Euclidean GRAAL with fixed step-size from \cite[Section~2]{malitsky2020golden} and the conclusions of Theorem~\ref{thm:conv} recover \cite[Theorem~1]{malitsky2020golden}. Despite this, the proof provided here is new and not the same as the one in \cite[Theorem~1]{malitsky2020golden} even when specialised to the Euclidean case. Indeed, \cite[Theorem~1]{malitsky2020golden} proceeds by establishing the inequality
\begin{equation}\label{eq:mal}
(\varphi+1)\|\overline{z}_{k+1}-z^*\|^2 + \frac{\varphi}{2}\|z_{k+1}-z_k\|^2 \\
\leq (\varphi+1)\|\overline{z}_{k}-z^*\|^2 + \frac{\varphi}{2}\|z_{k}-z_{k-1}\|^2 - \varphi\|z_k-\overline{z}_k\|^2,
\end{equation}
which is different to Lemma~\ref{lem:dec}. Interestingly, \eqref{eq:mal} can be deduced from \eqref{eq:central} in Lemma~\ref{lem:dec} by using the equality
\begin{equation*}
\|z_{k+1}-\overline{z}_{k+1}\|^2 = \frac{1}{\varphi^2}\|z_{k+1} - \overline{z}_k\|^2,
\end{equation*}
which applies in the Euclidean case, in place of \eqref{eq:distance} followed by the identity $\varphi^2=\varphi+1$. Note also that the inequality \eqref{eq:distance} is already weaker than the inequality $\|z_{k+1}-\overline{z}_{k+1}\|^2 \leq \frac{1}{\varphi^2}\|z_{k+1} - \overline{z}_k\|^2$.
\end{remark}
\section{The Adaptive Bregman Golden Ratio Algorithm}\label{sec:bagraal}
In this section, we present an adaptive modification to Algorithm~\ref{alg:B-GRAAL} and analyse its convergence. The method is presented in Algorithm~\ref{alg:B-aGRAAL}. As with the Euclidean aGRAAL, our Bregman adaptive modification has a fully explicit step-size rule. It is presented in Algorithm~\ref{alg:B-aGRAAL}.
\begin{algorithm}[!htb]
\caption{The Bregman adaptive Golden RAtio ALgorithm (B-aGRAAL)}\label{alg:B-aGRAAL}
\SetKwInput{Init}{Initialisation}
\KwIn{Initial points $z_0,\overline{z}_0\in\dom g$, initial step-size $\lambda_0>0$, $\lambda_{\max}\gg 0, \phi\in(1,\varphi]$}
\Init{Set $z_1=\overline{z}_0$, $\theta_0=1$, choose $\rho\in\left[1,\frac{1}{\phi}+\frac{1}{\phi^2}\right]$}
\For{$k=1,2,\dots$}{
Calculate the step-size:
\begin{equation}\label{eq:lambda}
\lambda_k = \min\left\{\rho\lambda_{k-1}, \frac{\sigma\phi\theta_{k-1}}{4\lambda_{k-1}}\frac{\|z_k - z_{k-1}\|^2}{\|F(z_k) - F(z_{k-1})\|^2}, \lambda_{\max}\right\}.
\end{equation}\\
Compute the next iterates:
\begin{align}
\overline{z}_k &= (\nabla h)^{-1}\left( \frac{(\phi - 1)\nabla h(z_k) + \nabla h(\overline{z}_{k-1})}{\phi}\right)\label{eq:bar}\\
z_{k+1} &= \argmin_{z\in\Hilbert}\left\{\langle F(z_k), z-z_k\rangle + g(z) + \frac{1}{\lambda_k} D_h(z,\overline{z}_k)\right\}\label{eq:z}\\
\end{align}\\
Update: $\theta_k = \frac{\lambda_k\phi}{\lambda_{k-1}}.$
}
\end{algorithm}
Observe that the step-size sequence $(\lambda_k)$ in Algorithm~\ref{alg:B-aGRAAL} approximates the inverse of a local Lipschitz constant in the following sense:
\begin{multline}\label{eq:inv loc lip}
\lambda_k \leq \frac{\sigma\phi\theta_{k-1}}{4\lambda_{k-1}}\frac{\|z_k - z_{k-1}\|^2}{\|F(z_k) - F(z_{k-1})\|^2} \leq \frac{\sigma\theta_k\theta_{k-1}}{4\lambda_k}\frac{\|z_k - z_{k-1}\|^2}{\|F(z_k) - F(z_{k-1})\|^2} \\
\implies \lambda_k\|F(z_k) - F(z_{k-1})\| \leq \frac{\sqrt{\sigma\theta_k\theta_{k-1}}}{2}\|z_k - z_{k-1}\|.
\end{multline}
Before giving our main convergence result for Algorithm~\ref{alg:B-aGRAAL}, we require some preparatory lemmas. The first two are concerned with well-definedness of the algorithm and boundedness of the step-size sequence.
\begin{lemma}\label{lem:ad contained}
Suppose Assumptions~\ref{ass:g}-\ref{ass:h} hold. Then the sequences $(\overline{z}_k)$ and $(z_k)$ generated by Algorithm~\ref{alg:B-aGRAAL} are well-defined. Moreover, $(\overline{z}_k)\subseteq\intr\dom h$ and $(z_k)\subseteq\intr\dom h\cap \dom g$.
\end{lemma}
\begin{proof}
Follows by an analogous argument to that of Lemma~\ref{lem:contained} but with $\lambda$ replaced by $\lambda_k$ and $\varphi$ replaced by $\phi$.
\end{proof}
\begin{lemma}\label{lem:bounded}
If $(z_k)$ generated by Algorithm~\ref{alg:B-aGRAAL} is bounded and $F$ is locally Lipschitz, then both $(\lambda_k)$ and $(\theta_k)$ are bounded and separated from $0$. In fact, for any $L>0$ satisfying $\|F(z_k)-F(z_{k-1})\|\leq L\|z_k-z_{k-1}\|$ for all $k\in\mathbb{N}$, we have
$$\lambda_k\geq \frac{\sigma\phi^2}{4L^2\lambda_{\max}}\text{~~and~~} \theta_k\geq\frac{\sigma\phi^3}{4L^2\lambda_{\max}^2}\quad\forall k\in\mathbb{N}. $$
\end{lemma}
\begin{proof}
First we note that $\lambda_k\leq\lambda_{\max}$ by definition, and that $\theta_k\leq\rho\phi\leq 1+\frac{1}{\phi}$.
Since $(z_k)$ is bounded and $F$ is locally Lipschitz continuous, there exists $L>0$ such that $\|F(z_k)-F(z_{k-1})\|\leq L\|z_k-z_{k-1}\|$ for all $k\in\mathbb{N}$. Then an argument analogous to that of~\cite[Lemma 2]{malitsky2020adaptive} shows that $\lambda_k\geq \frac{\sigma\phi^2}{4L^2\lambda_{\max}}$, and consequently, $\theta_k\geq\frac{\sigma\phi^3}{4L^2\lambda_{\max}^2}$.
\end{proof}
Our next result establishes an inequality similar, but not completely analogous, to its fixed step-size counterpart in Lemma~\ref{lem:dec}.
\begin{lemma}\label{lem:adaptive energy}
Suppose Assumptions~\ref{ass:g}-\ref{ass:S} hold. Let $z^*\in \Omega$ be arbitrary. Then the sequences $(z_k)$, $(\overline{z}_k)$ generated by Algorithm~\ref{alg:B-aGRAAL} satisfy.
\begin{multline}\label{eq:ad energy}
\frac{\phi}{\phi-1}D_h(z^*, \overline{z}_{k+1}) + \frac{\theta_k}{2}D_h(z_{k+1}, z_k) - D_h(z_{k+1}, \overline{z}_{k+1}) \\
\leq \frac{\phi}{\phi-1}D_h(z^*, \overline{z}_k) + \frac{\theta_{k-1}}{2}D_h(z_k, z_{k-1})
- \theta_k D_h(z_k, \overline{z}_k) \\+\left(\theta_k-1-\frac{1}{\phi}\right)D_h(z_{k+1}, \overline{z}_k).
\end{multline}
\end{lemma}
\begin{proof}
We proceed in a similar fashion as in the proof to Lemma~\ref{lem:dec}. By first applying Proposition~\ref{prop:prox}\ref{item:prox thm} with $f(z) := \lambda_k(\langle F(z_k), z-z_k\rangle + g(z))$, $u:=z\in\dom h\cap\dom g$ arbitrary, $x:=z_{k+1}$ and $y_0:=\overline{z}_k$, followed by the three-point identity (Proposition~\ref{prop:prop}\ref{item:3}) we obtain
\begin{equation}\label{eq:ad prox-ineq}
\begin{aligned}
\lambda_k\left(\langle F(z_k),z_{k+1}-z\rangle + g(z_{k+1}) - g(z)\right) &\leq \langle \nabla h(\overline{z}_k) - \nabla h(z_{k+1}), z_{k+1} - z\rangle\\
&= D_h(z, \overline{z}_k) - D_h(z, z_{k+1}) - D_h(z_{k+1}, \overline{z}_k).
\end{aligned}
\end{equation}
Shifting the index~\eqref{eq:ad prox-ineq} (by setting $k\equiv k-1$), setting $z:=z_{k+1}$, and using the fact that $\phi\nabla h(\overline{z}_k) = (\phi-1)\nabla h(z_k) + \nabla h(\overline{z}_{k-1})$ followed by the three point identity (Proposition~\ref{prop:prop}\ref{item:3}) gives
\begin{equation}\label{eq:ad shift prox-ineq}
\begin{aligned}
&\lambda_{k-1}\left(\langle F(z_{k-1}), z_k - z_{k+1}\rangle + g(z_k) - g(z_{k+1})\right) \\
&\quad \leq \langle \nabla h(\overline{z}_{k-1}) - \nabla h(z_k), z_k - z_{k+1}\rangle\\
&\quad= \phi\langle\nabla h(\overline{z}_k) - \nabla h(z_k), z_k - z_{k+1}\rangle \\
&\quad = \phi\left[D_h(z_{k+1},\bar{z}_k)-D_h(z_{k+1},z_k)-D_h(z_k,\bar{z}_k)\right].
\end{aligned}
\end{equation}
Now multiplying both sides of~\eqref{eq:ad shift prox-ineq} by $\tfrac{\lambda_k}{\lambda_{k-1}}$ then gives
\begin{multline}\label{eq:ad shift prox-ineq2}
\lambda_k\left(\langle F(z_{k-1}), z_k - z_{k+1}\rangle + g(z_k) - g(z_{k+1})\right) \\ \leq\theta_k\left[D_h(z_{k+1}, \overline{z}_k) - D_h(z_{k+1}, z_k) - D_h(z_k, \overline{z}_k)\right].
\end{multline}
Let $z^*\in S\cap\dom h$. Setting $z=z^*$ in~\eqref{eq:ad prox-ineq}, summing with~\eqref{eq:ad shift prox-ineq} and rearranging yields
\begin{equation}\label{eq:ad add prox-ineqs}
\begin{aligned}
\lambda_k\left(\langle F(z_k), z_k - z^*\rangle + g(z_k) - g(z^*)\right) \leq & D_h(z^*,\overline{z}_k) - D_h(z^*,z_{k+1}) - D_h(z_{k+1},\overline{z}_k)\\
&+\theta_k\Big[D_h(z_{k+1},\overline{z}_k) - D_h(z_{k+1},z_k) - D_h(z_k,\overline{z}_k)\Big]\\
&+\lambda_k\langle F(z_k) - F(z_{k-1}), z_k - z_{k+1}\rangle.
\end{aligned}
\end{equation}
We observe that the left-side of~\eqref{eq:ad add prox-ineqs} is non-negative as a consequence of~\eqref{eq:VI} and~\ref{ass:F}:
\begin{equation}\label{eq:ad lhs nonneg}
0\leq \langle F(z^*),z_k-z^*\rangle + g(z_k) - g(z^*)\leq \langle F(z_k),z_k-z^*\rangle + g(z_k) - g(z^*).
\end{equation}
To estimate the final term of~\eqref{eq:ad add prox-ineqs} we use the Cauchy--Schwarz inequality, the local Lipschitz estimate~\eqref{eq:inv loc lip}, and $\sigma$-strong convexity of $h$ to obtain
\begin{equation}\label{eq:local lipschitz}
\begin{aligned}
\lambda_k\langle F(z_k) - F(z_{k-1}), z_k - z_{k+1}\rangle
&\leq \frac{\sqrt{\sigma\theta_k\theta_{k-1}}}{2}\|z_k - z_{k-1}\|\|z_k - z_{k+1}\|\\
&\leq \frac{\sigma\theta_{k-1}}{4}\|z_k - z_{k-1}\|^2+\frac{\sigma\theta_k}{4}\|z_k - z_{k+1}\|^2\\
&\leq \frac{\theta_{k-1}}{2}D_h(z_k, z_{k-1}) + \frac{\theta_k}{2}D_h(z_{k+1}, z_k).
\end{aligned}
\end{equation}
Combining~\eqref{eq:ad add prox-ineqs}, \eqref{eq:ad lhs nonneg} and~\eqref{eq:local lipschitz} gives
\begin{multline}\label{eq:ad combination}
D_h(z^*, z_{k+1}) \leq D_h(z^*, \overline{z}_k) + (\theta_k-1)D_h(z_{k+1}, \overline{z}_k) - \theta_k D_h(z_k, \overline{z}_k)\\
- \frac{\theta_k}{2}D_h(z_{k+1}, z_k) + \frac{\theta_{k-1}}{2}D_h(z_k,z_{k-1}).
\end{multline}
Now applying Proposition~\ref{prop:prop}\ref{item:identity} with $\nabla h(z_{k+1}) = \frac{\phi\nabla h(\overline{z}_{k+1}) - \nabla h(\overline{z}_k)}{\phi-1}$ and rearranging yields
\begin{multline}\label{eq:ad identity}
\frac{\phi}{\phi-1}D_h(z^*, \overline{z}_{k+1}) = D_h(z^*, z_{k+1}) + \frac{\phi}{\phi-1}D_h(z_{k+1}, \overline{z}_{k+1}) \\ + \frac{1}{\phi-1}\Big[D_h(z^*, \overline{z}_k) - D_h(z_{k+1}, \overline{z}_k)\Big].
\end{multline}
Combining \eqref{eq:ad combination} and \eqref{eq:ad identity}, followed by collecting like-terms and rearranging, gives
\begin{multline}\label{eq:ad central}
\frac{\phi}{\phi-1}D_h(z^*, \overline{z}_{k+1}) + \frac{\theta_k}{2}D_h(z_{k+1}, z_k) - D_h(z_{k+1}, \overline{z}_{k+1}) \\
\leq \frac{\phi}{\phi-1}D_h(z^*, \overline{z}_k) + \frac{\theta_{k-1}}{2}D_h(z_k, z_{k-1})
- \theta_k D_h(z_k, \overline{z}_k) \\+\left(\theta_k - \frac{\phi}{\phi-1}\right)D_h(z_{k+1}, \overline{z}_k) + \frac{1}{\phi-1}D_h(z_{k+1}, \overline{z}_{k+1}).
\end{multline}
Next we apply Proposition~\ref{prop:prop}\ref{item:identity} once again to see that
\begin{multline}\label{eq:ad distance}
D_h(z_{k+1},\overline{z}_{k+1})=\frac{\phi-1}{\phi}\left[D_h(z_{k+1},z_{k+1}) - D_h(\overline{z}_{k+1},z_{k+1})\right] + \frac{1}{\phi}\left[D_h(z_{k+1},\overline{z}_k) - D_h(\overline{z}_{k+1},\overline{z}_k)\right]\\
\leq\frac{1}{\phi}D_h(z_{k+1},\overline{z}_k).
\end{multline}
The final line of~\eqref{eq:ad central} can therefore be estimated as
\begin{equation}\label{eq:last adp}\begin{aligned}
\left(\theta_k - \frac{\phi}{\phi-1}\right)D_h(z_{k+1}, \overline{z}_k) + \frac{1}{\phi-1}D_h(z_{k+1}, \overline{z}_{k+1})
&\leq \left(\theta_k - \frac{\phi}{\phi-1} + \frac{1/\phi}{\phi-1}\right)D_h(z_{k+1}, \overline{z}_k) \\
&= \left(\theta_k -1-\frac{1}{\phi}\right)D_h(z_{k+1}, \overline{z}_k).
\end{aligned}\end{equation}
Substituting \eqref{eq:last adp} into \eqref{eq:ad central} gives \eqref{eq:ad energy}, which completes the proof.
\end{proof}
The following is our main result regarding convergence of the Bregman adaptive GRAAL.
\begin{theorem}\label{th:B-aGRAAL}
Suppose Assumptions~\ref{ass:g}-\ref{ass:S} hold, and $F$ is locally Lipschitz continuous on the bounded set $\dom g\cap\intr\dom h$. Choose $\phi\in\left(1,\varphi\right)$ and $\rho\in\left[1,\frac{1}{\phi}+\frac{1}{\phi^2}\right)$.
Then the sequences $(z_k)$ and $(\overline{z}_k)$ generated by Algorithm~\ref{alg:B-aGRAAL} converge to a point in $\Omega$ whenever $\lambda_{\rm max}>0$ is sufficiently small.
\end{theorem}
\begin{proof}
Since $\dom g\cap\intr\dom h$ is bounded and $F$ is locally Lipschitz, there exists $L>0$ such that $F$ is $L$-Lipschitz continuous on $\dom g\cap\intr\dom h$. Suppose $\lambda_{\rm max}$ is sufficient small so that $\lambda_{\max}\leq\frac{\phi\sqrt{\phi\sigma}}{2L}$ holds. Then, applying Lemma~\ref{lem:bounded} gives $\theta_k\geq\frac{\sigma\phi^3}{4L^2\lambda_{\max}^2}\geq 1$ for all $k\in\mathbb{N}$. Also, since $\theta_k \leq \rho\phi < 1+\frac{1}{\phi}$, there exists an $\epsilon>0$ such that $\theta_k-1-\frac{1}{\phi} \leq -\epsilon$ for all $k\in\mathbb{N}$.
Now let $z^*\in\Omega$ be arbitrary and denote by $(\eta_k)$ the sequence
$$\eta_k := \frac{\phi}{\phi-1}D_h(z^*, \overline{z}_k) +\frac{\theta_{k-1}}{2}D_h(z_k, z_{k-1}) - D_h(z_k, \overline{z}_k).$$
Applying Lemma~\ref{lem:adaptive energy} yields
\begin{equation}\label{eq:new energy}
\eta_{k+1} \leq \eta_k -\epsilon D_h(z_{k+1},\overline{z}_k).
\end{equation}
Next, we should show that $(\eta_k)$ is bounded below. To this end, using the three point identity (Proposition \ref{prop:prop}\ref{item:3}) followed by \eqref{eq:ad prox-ineq} gives
\begin{equation}\label{eq:last term 1}
\begin{aligned}
D_h(z^*,z_{k+1}) &+ D_h(z_{k+1},\overline{z}_{k+1}) = D_h(z^*,\overline{z}_{k+1}) + \langle\nabla h(\overline{z}_{k+1}) - \nabla h(z_{k+1}), z^*-z_{k+1}\rangle\\
&= D_h(z^*,\overline{z}_{k+1}) + \frac{1}{\phi}\langle\nabla h(\overline{z}_k) - \nabla h(z_{k+1}), z^*-z_{k+1}\rangle\\
&\leq D_h(z^*, \overline{z}_{k+1}) + \frac{\lambda_k}{\phi}\left(\langle F(z_k), z^* - z_{k+1}\rangle - g(z_{k+1}) + g(z^*)\right)\\
&= D_h(z^*, \overline{z}_{k+1}) + \frac{\lambda_k}{\phi}\langle F(z_k) - F(z_{k+1}), z^* - z_{k+1}\rangle \\ &\quad +\frac{\lambda_k}{\phi}(\langle F(z_{k+1}), z^* - z_{k+1}\rangle - g(z_{k+1}) + g(z^*)).
\end{aligned}
\end{equation}
Now, as with~\eqref{eq:add prox-ineqs}, the final term in \eqref{eq:last term 1} is non-negative. Since $(\lambda_k)$ and $(\theta_k)$ are bounded and separated from $0$ by Lemma~\ref{lem:bounded}, there exists a constant $M>0$ such that
\begin{equation}\label{eq:bd above}\begin{aligned}
\lambda_k\langle F(z_k) - F(z_{k+1}), z^* - z_{k+1}\rangle
&\leq \frac{\lambda_k}{\lambda_{k+1}}\lambda_{k+1}\|F(z_k) - F(z_{k+1})\|\|z^*-z_{k+1}\| \\
&\leq \frac{\lambda_k}{\lambda_{k+1}}\frac{\sigma\sqrt{\theta_{k+1}\theta_k}}{2}\|z_k-z_{k+1}\|\|z^*-z_{k+1}\|\leq M.
\end{aligned}\end{equation}
Combing \eqref{eq:last term 1} and \eqref{eq:bd above}, noting that $\phi<\frac{\phi}{\phi-1}$, then gives
\begin{equation}\label{eq:bd below}
D_h(z_{k+1},\overline{z}_{k+1}) \leq D_h(z^*, \overline{z}_{k+1}) + \frac{M}{\phi}\leq\phi D_h(z^*,\overline{z}_{k+1}) + M
\implies \eta_{k+1} \geq -M,
\end{equation}
which establishes that $(\eta_k)$ is bounded below.
Next, by telescoping the inequality~\eqref{eq:new energy}, deduce that $(\eta_k)$ is bounded and $D_h(z_{k+1},\overline{z}_k)\to 0$ as $k\to\infty$. Referring to~\eqref{eq:ad distance}, it follows that $D_h(z_{k+1},\overline{z}_{k+1})\to 0$. Also, by applying Proposition~\ref{prop:prop}\ref{item:identity} with the identity $\nabla h(z_k) = \frac{\phi\nabla h(\overline{z}_k) - \nabla h(\overline{z}_{k-1})}{\phi-1}$, we obtain
\begin{align*}
D_h(z_{k+1},z_k) &= \frac{\phi}{\phi-1}\left[D_h(z_{k+1},\overline{z}_k) - D_h(z_k,\overline{z}_k)\right] - \frac{1}{\phi-1}\left[D_h(z_{k+1},\overline{z}_{k-1}) - D_h(z_k,\overline{z}_{k-1})\right]\\
&\leq \frac{\phi}{\phi-1}D_h(z_{k+1},\overline{z}_k) + \frac{1}{\phi-1}D_h(z_k,\overline{z}_{k-1}) \to 0.
\end{align*}
Altogether, noting that $(\theta_k)$ is bounded, we deduce that
$$\lim_{k\to\infty}\eta_k = \frac{\phi}{\phi-1}\lim_{k\to\infty}D_h(z^*,\overline{z}_k),$$
and so, in particular, $\lim_{k\to\infty}D_h(z^*,\overline{z}_k)$ exists.
Next, using $\sigma$-strong convexity of $h$, we deduce that $z_{k+1}-\overline{z}_k\to0$. Let $\overline{z}\in\Hilbert$ be a cluster point of the bounded sequence $(\overline{z}_k)$. Then there exists a subsequence $(\overline{z}_{k_j})$ such that $\overline{z}_{k_j}\to \overline{z}$ and $z_{k_j+1}\to \overline{z}$ as $j\to\infty$.
Now recalling~\eqref{eq:ad prox-ineq} gives
$$\lambda_k\left(\langle F(z_{k_j}),z_{{k_j}+1}-z\rangle + g(z_{{k_j}+1}) - g(z)\right) \leq \langle \nabla h(\overline{z}_{k_j}) - \nabla h(z_{{k_j}+1}), z_{{k_j}+1} - z\rangle \quad\forall z\in\Hilbert,$$
and taking the limit-infimum of both sides as $j\to\infty$ shows that $\overline{z}\in\Omega$. Since $z^*\in\Omega$ was chosen in Lemma~\ref{lem:adaptive energy} to be arbitrary, we can now set $z^*=\overline{z}$. It then follows that $\lim_{j\to\infty}D_h(z^*,\overline{z}_{k_j}) = 0$, and consequently, $\lim_{j\to\infty}\eta_{k_j} = 0$. Also note that for $n\geq k_j$, we have $\eta_n\leq\eta_{k_j}$ from Lemma \ref{lem:dec}, and therefore
$$\frac{\phi}{\phi-1}\lim_{n\to\infty}D_h(z^*, \overline{z}_n) = \lim_{n\to\infty}\eta_n \leq \lim_{j\to\infty}\eta_{k_j} = 0,$$
and $\overline{z}_k\to z^*$ from strong convexity. The fact that $z_k\to z^*$ follows since $z_k-\overline{z}_k\to 0$.
\end{proof}
\begin{remark}
The energy~\eqref{eq:new energy} used in the proof Theorem~\ref{th:B-aGRAAL} is not completely analogous to the one used in the fixed step-size case given in Lemma~\ref{lem:dec}. Notably, the coefficient of the final term of $\eta_k$ in \eqref{eq:new energy} has coefficient $-1$ whereas the corresponding term in Lemma~\ref{lem:dec} has coefficient $-\varphi$. Although it is unlikely to be useful in practice, another interesting feature of the proof it that it requires the maximum step-size to satisfy the upper bound $$\lambda_{\rm max}\leq \frac{\phi\sqrt{\phi\sigma}}{2L} = \sqrt{\frac{\phi}{\sigma}}\frac{\sigma\phi}{2L}. $$
Note that, when $\phi>\sigma$, this upper bound is looser than the upper-bound of $\lambda<\frac{\sigma\phi}{2L}$ required for B-GRAAL (Algorithm~\ref{alg:B-GRAAL}). Thus, it is possible for maximum step-size of B-aGRAAL to be larger than the step-size required for B-GRAAL. This is the case, for instance, if $D_h$ is the KL divergence on the simplex in which case $\sigma=1<\phi$.
Also, the proof of Theorem~\ref{th:B-aGRAAL} required $L$ be a Lipschitz constant for $F$. However, thanks to Lemma~\ref{lem:bounded}, it would have sufficed for $L$ to satisfy the weaker Lipschitz-like inequality $\|F(z_k)-F(z_{k-1})\|\leq L\|z_k-z_{k-1}\|$ for all $k\in\mathbb{N}$. In turn, this would allow for a greater upper bound for $\lambda_{\rm max}$ which is inversely related to $L$.
\end{remark}
\section{Numerical Experiments}\label{sec:exp}
In this section, we present some experimental results for Algorithms~\ref{alg:B-GRAAL} and \ref{alg:B-aGRAAL}, to compare the respective original Euclidean methods against our new methods with respect to various choices of Bregman distances. All experiments are run in Python 3 on a Windows 10 machine with 8GB memory and an Intel(R) Core(TM) i7-10510U CPU @ 1.80GHz processor.
We consider three different problems, all of which can formulated as the variation inequality~\eqref{eq:VI2} for different choices of the function $g$ and the operator $F$. Note that the solutions of the variational inequality~\eqref{eq:VI2} can also be characterised as the monotone inclusion
$$\text{find~}z^*\in\Hilbert\text{~such that~}0\in F(z^*) + \partial g(z^*).$$
Thus, by noting that \eqref{eq:z} implies $0 \in F(z_k) + \partial g(z_{k+1}) + \frac{1}{\lambda_k}\left(\nabla h(z_{k+1}) - \nabla h(\overline{z}_k)\right)$, we monitor the quantity $J_k$ given by
\begin{equation}\label{eq:Jk}
J_k := \frac{1}{\lambda_k}\left(\nabla h(\overline{z}_k) - \nabla h(z_{k+1})\right) + F(z_{k+1}) - F(z_k) \in F(z_{k+1}) + \partial g(z_{k+1})
\end{equation}
as a natural residual for Algorithm~\ref{alg:B-aGRAAL}. The analogous expression for Algorithm~\ref{alg:B-GRAAL} is given by replacing $\lambda_k$ in \eqref{eq:Jk} with $\lambda$.
In all experiments, we run each algorithm on for the same (fixed) number of iterations on $10$ random instances of the chosen problem. The figures in each section show the decrease in residual over time, with each faint line representing one instance and the bold line showing the mean behaviour. We use the parameters $\phi=1.5,\lambda_{\max}=10^6$ throughout, and initial step-size and iterates as described in each section.
\subsection{Matrix Games}\label{subsec:matrix}
To test the fixed step-size B-GRAAL (Algorithm~\ref{alg:B-GRAAL}), we first consider the following matrix game between two players
\begin{equation}\label{eq:matrix}
\min_{x\in\Delta^n}\max_{y\in\Delta^n}\langle Mx,y\rangle,
\end{equation}
where $\Delta^n := \{x\in\mathbb{R}^n_+\colon \sum_{i=1}^n x_i = 1\}$ denotes the unit simplex and $M\in\mathbb{R}^{n\times n}$ a given matrix. This problem is of the form specified in \eqref{eq:saddle} and so can be formulated as the variational inequality~\eqref{eq:VI}.
In particular, we consider the specific problem in the form of \eqref{eq:matrix} of placing a server on a network $G = (V,E)$ with $n$ vertices in a way that minimises its response time. In this problem, a request will originate at some vertex $v_j\in V$, which is not known ahead of time, and the objective is to place the server at a vertex $v_i\in V$ such the response time, as measured by the graphical distance $d(v_i,v_j)$, is minimised. We consider the case where the request location $v_j\in V$ is a decision made by an adversary. The decision variable $x\in\Delta^n$ (resp.\ $y\in\Delta^n$) models mixed strategies for placement of the server (resp.\ request origin). In other words, $x_t$ (resp.\ $y_t$) is the probability of the server (resp.\ request origin) being located at the node $v_t$ for $t=1,\dots,n$. The matrix $M$ is taken to be the distance matrix of the graph $G$, that is $M_{ij} = d(v_i,v_j)$ for all vertices $v_i,v_j\in V$. In this way, the objective function in~\eqref{eq:matrix} measures the expected response time, which we would like to minimise while our adversary seeks to maximise it.
We compare Algorithm~\ref{alg:B-GRAAL} with the squared norm $h(z) = \frac{1}{2}\|z\|^2$, which generates the squared Euclidean distance $D_h(u,v) = \frac{1}{2}\|u-v\|^2$, and the negative entropy $h(z) = \sum_{i=1}^{2n} z_i\log z_i$, which generates the KL divergence $D_h(u,v) = \sum_{i=1}^{2n}u_i\log\frac{u_i}{v_i} + v_i - u_i$, where $z=(x,y)\in\mathbb{R}^{2n}$ is the concatenation of the two vectors $x$ and $y$. Both of these choices for $h$ are $1$-strongly convex on $\Delta^n$. A potential advantage of the the KL projection onto the simplex is that it has the simple closed-formed expression $x\mapsto\frac{x}{\|x\|_1}$. whereas the Euclidean projection has no closed form and takes $O(n\log n)$ time (see, for instance \cite{wang2012projection, chen2011projection, dai2022distributed}). We run two experiments with $n=500$ and $n=1000$, and the results are shown respectively in Figures~\ref{fig:matrix500} and~\ref{fig:matrix1000}. Initial points are chosen as $x_0=y_0=\left(\frac{1}{n},\dots,\frac{1}{n}\right)\in\Delta^n$, and $z_0=(x_0,y_0)$, then $\overline{z}_0$ as a random perturbation of $z_0$. The Lipschitz constant of $F$ is computed as $L=\|M\|_2$, and the algorithm step-size is taken as $\lambda=\frac{\varphi}{2L}$. Under these conditions, convergence to a solution is guranteeed by Theorem~\ref{thm:conv}.
The residual, given by $\min_{i\leq k}\|J_i\|^2$, is shown in Figures~\ref{sfig:500} and \ref{sfig:1000}. The time per iteration is also shown in Figures~\ref{sfig:500time} and \ref{sfig:1000time}. Despite the KL projection being faster than the Euclidean projection, overall, the Euclidean method performed better in this instance.
We now move onto experiments for the adaptive Algorithm~\ref{alg:B-aGRAAL}.
\begin{figure*}[t!]
\centering
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{500error.png}
\caption{Residual over iterations}
\label{sfig:500}
\end{subfigure}%
~
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{500time.png}
\caption{Run time per iteration}
\label{sfig:500time}
\end{subfigure}
\caption{Matrix game results for $n=500$}
\label{fig:matrix500}
\end{figure*}
\begin{figure*}[t!]
\centering
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{1000error.png}
\caption{Residual over iterations}
\label{sfig:1000}
\end{subfigure}%
~
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{1000time.png}
\caption{Run time per iteration}
\label{sfig:1000time}
\end{subfigure}
\caption{Matrix game results for $n=1000$}
\label{fig:matrix1000}
\end{figure*}
\subsection{Gaussian Communication}
We now turn our attention to maximising the information capacity of a noisy \emph{Gaussian communication channel}~\cite[Chapter 9]{thomas2006elements}. In this problem, the goal is to allocate a total power of $P$ across $m$ channels, represented by $p\in\mathbb{R}^m_+$, to maximise the total information capacity of the channels in the presence of allocated noise, represented by $n\in\mathbb{R}^m_+$. The information capacity of the $i$th channel, denoted $C_i(p_i,n_i)$, is the function of power $p_i$ and noise level $n_i$ and is given by
$$C_i(p_i,n_i) = \log\left(1+\frac{\beta_i p_i}{\mu_i+n_i}\right),$$
where $\mu_i>0$ and $\beta_i>0$ are given constants.
Assuming a total power level of $P$ and a total noise level of $N$, optimising for the worst-case scenario by treating the noise allocation as an adversary gives the convex-concave game
$$\max_{p\in\Delta^m_P}\min_{n\in\Delta^m_N} \sum_{i=1}^m C_i(p_i,n_i),$$
where $\Delta_T^n := \{x\in\mathbb{R}^n_+\colon \sum_{i=1}^n x_i = T\}$ is a \textit{scaled} simplex. This problem is also of the form specified in \eqref{eq:saddle} and so can also be formulated as the variational inequality~\eqref{eq:VI}.
The Lipschitz constant of the operator $F$ for this problem is not straightforward to compute, so we apply the adaptive algorithms. Similarly to Section~\ref{subsec:matrix}, we compare the Euclidean and KL versions of Algorithm~\ref{alg:B-aGRAAL}. Since $x\mapsto x\log x$ is $\frac{1}{M}$-strongly convex for $0<x\leq M$, the strong convexity constant of the negative entropy over $\Delta^m_P\times\Delta^m_N$ is $\min\left\{\frac{1}{P},\frac{1}{N}\right\}$. In our experiments, we set $(P,N)=(500,50)$ and generate $\beta\in(0,P]^m$ and $\mu\in(1,N+1]^m$ uniformly. The initial points are chosen as $p_0=\left(\frac{P}{m},\dots,\frac{P}{m}\right)\in\Delta^m_P,n_0=\left(\frac{N}{m},\dots,\frac{N}{M}\right)\in\Delta^m_N$, and $z_0=(p_0,n_0)$, and the initial step-size is taken as $\lambda_0=\frac{\|z_0-\overline{z}_0\|^2}{\|F(z_0)-F(\overline{z}_0)\|^2}$ where $\overline{z}_0$ is again a small random perturbation of $z_0$. It is also worth noting that for the KL version, we had to multiply $\lambda_0$ by a small constant ($10^{-2}$) to avoid numerical instability issues.
\begin{figure*}[t!]
\centering
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{100error.png}
\caption{Residual over iterations}
\label{sfig:100}
\end{subfigure}%
~
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{100times.png}
\caption{Run time per iteration}
\label{sfig:100time}
\end{subfigure}
\caption{Gaussian communication channel results for $m=100$}
\label{fig:gauss100}
\end{figure*}
\begin{figure*}[t!]
\centering
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{200error.png}
\caption{Residual over iterations}
\label{sfig:200}
\end{subfigure}%
~
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{200times.png}
\caption{Run time per iteration}
\label{sfig:200time}
\end{subfigure}
\caption{Gaussian communication channel results for $m=200$}
\label{fig:gauss200}
\end{figure*}
We run two experiments, for $m=100$ and $m=200$, and plot the results in Figures~\ref{fig:gauss100} and~\ref{fig:gauss200} respectively. The KL method is slower than the Euclidean method in terms of the number of iterations and time. However, unlike in the previous section, both methods reach a similar final accuracy.
\subsection{Cournot Completion}
Our final example is a standard $N$-player Cournot oligopoly model \cite[Example 2.1]{bravo2018bandit}. This is a system in which $N$ independent firms supply the market with a quantity of some common good or service. More formally, each firm seeks to maximise their utility subject to their capacity, that is,
\begin{equation}\label{eq:utility}
\displaystyle{\max_{0\leq x_i\leq C_i}} u_i(x_i,x_{-i}) = x_i P(x_T) - c_i x_i,
\end{equation}
where
\begin{itemize}
\item $x_i\geq 0$ is the quantity of the good supplied by the $i$th firm, $i=1,\dots,N$.
\item $x_{-i} = (x_1,\dots,x_{i-1},x_{i+1},\dots,x_N)$ is the quantity of the good supplied by all other firms.
\item $x_T := \sum_{i=1}^N x_i$ is the total amount supplied.
\item $C_i>0$ is the production capacity of the $i$th firm.
\item $c_i>0$ is the production cost of the $i$th firm.
\item $P\colon\mathbb{R}_+\to\mathbb{R}$ is the inverse demand curve.
\end{itemize}
In this section, we consider solutions to this problem in the sense of \emph{Nash equilibria}, which is equivalent to the variational inequality~\eqref{eq:VI2} with
$$F = \left(-\frac{\partial u_1}{\partial x_1}, \dots, -\frac{\partial u_N}{\partial x_N}\right),\quad K = \left[0,C_1\right]\times\dots\times\left[0,C_N\right],\quad g = \iota_K.$$
By choosing a function $h$ such that $\dom h=K$, it is possible to implicitly enforce the capacity constraint in this problem and avoid performing projections. Two examples, over a single closed interval $[\alpha,\beta]$, present themselves which satisfy our assumptions:
$$ h_1(x) = (x-\alpha)\log(x-\alpha) + (\beta-x)\log(\beta-x), \quad
h_2(x) = -\sqrt{(x-\alpha)(\beta-x)}.$$
When $\alpha=0$ and $\beta=1$, $h_1$ is called the \emph{Fermi-Dirac Entropy}, and similarly when $\alpha=-1$ and $\beta=1$, $D_{h_2}$ is called the \emph{Hellinger distance} (\cite[Example 2.2]{bauschke2019linear}, \cite[Example 1]{bauschke2017descent}, \cite[Example 2.1]{teboulle2018simplified}). Then we sum over the independent functions of one variable, with appropriately set intervals $\alpha=0,\beta=C_i$, to create the Bregman distance over the closed box $K$. We also compute the strong convexity constants by minimising $\nabla^2 h_1$ over $(\alpha,\beta)$, which gives $\sigma = \frac{4}{\beta-\alpha}$, and similarly for $h_2$, $\sigma=\frac{2}{\beta-\alpha}$.
In our experiments, $P$ is taken to be a linear inverse demand curve given by $P(x) = a-bx$ for $a,b>0$. All parameters are generated by a log-normal distribution, except the cost vector $c\in\mathbb{R}^N$, which is generated uniformly in $\left[\frac{C_1}{100},\frac{C_1}{5}\right]\times\dots\times\left[\frac{C_N}{100},\frac{C_N}{5}\right]$. We run two experiments, with $N=2000$ and $N=5000$, the results of which are shown in Figures~\ref{fig:nash2000} and~\ref{fig:nash5000}. The initial points are chosen as $z_0 = \frac{C}{2}$, and $\lambda_0=1$.
\begin{figure*}[t!]
\centering
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{2000error.png}
\caption{Residual over iterations}
\end{subfigure}%
~
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{2000times.png}
\caption{Time per iteration}\label{sfig:2000time}
\end{subfigure}
\caption{Results for $N=2000$}
\label{fig:nash2000}
\end{figure*}
\begin{figure*}[t!]
\centering
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{5000error.png}
\caption{Residual over iterations}
\end{subfigure}%
~
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{5000times.png}
\caption{Time per iteration}\label{sfig:5000time}
\end{subfigure}
\caption{Results for $N=5000$}
\label{fig:nash5000}
\end{figure*}
We observe here that the final accuracy depends heavily on the choice of Bregman distance. In both figures, we observe that the Hellinger method is not making any progress after approximately $200$ iterations, while the Euclidean method acheives a modest final accuracy. Meanwhile the Bit method is very fast and accurate, converging to a near $0$ tolerance in roughly $300$ iterations in all instances. Finally, we note that all methods are roughly equal in terms of time per iteration. This is to be expected, since the Euclidean method requires a projection which evaluates $\min\{\max\{x_i,0\}, C_i\}$ for each component, whereas the Bregman methods require evaluating $\nabla h,(\nabla h)^{-1}$ over each component --- all of which are taking $O(N)$ time.
\section{Conclusion}\label{sec:conc}
In this paper, we extended the adaptive method aGRAAL (Algorithm~\ref{alg:aGRAAL}) to the Bregman distance setting. We proposed two such extensions: The first, Algorithm \ref{alg:B-GRAAL}, generalises the fixed step-size GRAAL and converges under the same assumptions for a strongly-convex Bregman function $h\colon\Hilbert\to\mathbb{R}$. The second, Algorithm \ref{alg:B-aGRAAL}, generalises Algorithm \ref{alg:aGRAAL}, and converges in a more restrictive setting. We first examined the performance of Algorithm~\ref{alg:B-GRAAL}, and found that the KL version is less favourable than the Euclidean version, despite the reduced time per iteration. We then tested Algorithm~\ref{alg:B-aGRAAL} on a convex-concave game for Gaussian communication channels, where our new method performed worse with respect to the KL divergence when compared to the Euclidean method, although the run-time per iteration was again significantly shorter as was expected. Finally, we examined a Cournot completion model, where one of Bregman based methods reached a much higher accuracy very quickly.
We conclude by outlining directions for further research:
\begin{itemize}
\item It would be interesting to know whether Algorithm~\ref{alg:B-aGRAAL} can be shown to converge in a more general setting than what we have shown, and if so, under what circumstances. The difficulties in our analysis arose from two issues: first, the estimate derived in~\eqref{eq:ad distance} for the Bregman case is weaker than the Euclidean equality $\|z_{k+1}-\overline{z}_{k+1}\|^2 = \frac{1}{\phi^2}\|z_{k+1}-\overline{z}_k\|^2$ used in~\cite{malitsky2020golden}; and second, the inability to bound $\theta_k$ below without additional assumptions (as the bound in Lemma~\ref{lem:bounded} can be arbitrarily small in general).
\item Throughout this paper, $h$ was assumed strongly convex, but whether or not this assumption can be relaxed is unclear. One potential replacement for strong convexity is considered in \cite{bauschkedykstra, laude2020bregman} where $h$ is twice differentiable such that the Hessian matrix $\nabla^2 h(z)$ is positive definite for all $z\in\intr\dom h$. Within the scope of twice differentiable functions, such a condition lies between strict and strong convexity, the main consequence being that $\nabla h$ is locally Lipschitz and $h$ is locally strongly convex. Indeed, for $\sigma$-strongly convex $h$ with $\eta$-Lipschitz gradient, one can derive the estimates
$$\frac{\sigma}{2}\|z-z'\|^2 \leq D_h(z,z') \leq \frac{\eta}{2}\|z-z'\|^2\quad\forall z,z'.$$
If such inequalities hold on a local scale, then it would remain to be seen whether these coefficients can be estimated and used in the same way that $\lambda_k$ approximates an inverse of the local Lipschitz constant of $F$.
\item In the context of the convex composite optimisation problems of the form $\min_{x\in \Hilbert}f(x) + g(x)$ where $g$ is nonsmooth and $f$ is smooth, the Bregman proximal gradient algorithm \cite{teboulle2018simplified, bauschke2017descent} is known to converge in a more general setting than Lipschitz continuity. Specifically, $L$-Lipschitz continuity of $\nabla f$ can be relaxed to convexity of the function $Lh-f$, which indeed holds if $\nabla f$ is Lipschitz and $h$ is strongly-convex. It would be interesting to see if a similar relaxation of Lipschitz continuous can be used in the context of the algorithms discussed here.
\end{itemize}
\paragraph{Acknowledgements.}
DJU and MKT are supported in part by Australian Research Council grant DE200100063.
|
\section*{Acknowledgments}\end{small}{\noindent #1}\vspace{5pt}}
\newcommand{\datastatement}[1]{\begin{small}\section*{Data Availability Statement}\end{small}{\noindent #1}\vspace{5pt}}
\defcitealias{GrudicHopkins2018}{G18}
\newcommand{{Alfv\'en}}{{Alfv\'en}}
\title[Rapid BH Growth in Dense Clouds]{Hyper-Eddington Black Hole Growth in Star-Forming Molecular Clouds and Galactic Nuclei: Can It Happen?}
\author[Y.\ Shi et al.]{
\vspace{0.2cm}
\parbox[t]{\textwidth}{
Yanlong Shi,$^{1}$\thanks{E-mail: <EMAIL>}
Kyle Kremer,$^{1,2}$
Michael Y. Grudi\'c,$^{2}$
Hannalore J. Gerling-Dunsmore,$^{3,4}$
Philip F. Hopkins$^{1}$
} \vspace*{4pt} \\
$^{1}$TAPIR, Mailcode 350-17, California Institute of Technology, Pasadena, CA 91125, USA\\
$^{2}$The Observatories of the Carnegie Institution for Science, Pasadena, CA 91101, USA\\
$^{3}$ Department of Astrophysical and Planetary Sciences, University of Colorado, 391 UCB, Boulder, CO 80309-0391, USA\\
$^{4}$JILA, University of Colorado and National Institute of Standards and Technology, 440 UCB, Boulder, CO 80309-0440, USA
}
\date{}
\begin{document}
\maketitle
\begin{abstract}
Formation of supermassive black holes (BHs) remains a theoretical challenge. In many models, especially beginning from stellar relic ``seeds,'' this requires sustained super-Eddington accretion. While studies have shown BHs can violate the Eddington limit on accretion disk scales given sufficient ``fueling'' from larger scales, what remains unclear is whether or not BHs can actually capture sufficient gas from their surrounding ISM. We explore this in a suite of multi-physics high-resolution simulations of BH growth in magnetized, star-forming dense gas complexes including dynamical stellar feedback from radiation, stellar mass-loss, and supernovae, exploring populations of seeds with masses $\sim 1-10^{4}\,M_{\odot}$. In this initial study, we neglect feedback from the BHs: so this sets a strong upper limit to the accretion rates seeds can sustain. We show that stellar feedback plays a key role. Complexes with gravitational pressure/surface density below $\sim 10^{3}\,M_{\odot}\,{\rm pc^{-2}}$ are disrupted with low star formation efficiencies so provide poor environments for BH growth. But in denser cloud complexes, early stellar feedback does not rapidly destroy the clouds but does generate strong shocks and dense clumps, allowing $\sim 1\%$ of randomly-initialized seeds to encounter a dense clump with low relative velocity and produce runaway, hyper-Eddington accretion (growing by orders of magnitude). Remarkably, mass growth under these conditions is almost independent of initial BH mass, allowing rapid IMBH formation even for stellar-mass seeds. This defines a necessary (but perhaps not sufficient) set of criteria for runaway BH growth: we provide analytic estimates for the probability of runaway growth under different ISM conditions.
\end{abstract}
\vspace{-0.5cm}
\begin{keywords}
black hole physics -- accretion, accretion discs -- quasars: supermassive black holes -- galaxies: star formation
\end{keywords}
\vspace{-0.3cm}
\section{Introduction}
\label{sec:intro}
Observations have demonstrated the existence of supermassive black holes (BHs) with masses $M_{\rm bh} \sim 10^9 M_\odot$ in quasars at very high redshift ($z \gtrsim 7$) when the Universe was less than a billion years old \cite[e.g.,][]{FanNarayanan2001,WangYang2021}, which implies that these BHs must accrete rapidly from their ``seeds'' \citep{InayoshiVisbal2020}. The physical origin of these seeds remains deeply uncertain, but popular models including direct collapse of super-massive stars with masses $\sim 10^{4}-10^{6}\,M_{\odot}$ \cite[e.g.,][]{BegelmanVolonteri2006,ReganVisbal2017,CorbettMoranGrudic2018,ChonOmukai2020}, runaway mergers in globular clusters \cite[e.g.,][]{PortegiesZwartBaumgardt2004,2020ApJ...891...94B,2020MNRAS.493.2352A,KremerSpera2020,RizzutoNaab2021,ShiGrudic2021,FragioneKocsis2021}, remnants from Population III stars \cite[e.g.,][]{MadauRees2001,RyuTanaka2016}, and relics of ``standard'' stellar evolution (e.g.\ Population II) stars generally produce seeds with masses $\ll 10^{4}\,M_{\odot}$. Given that the $e$-folding time of a BH growing at the Eddington limit\footnote{Throughout, we will follow standard convention and define the Eddington {\em luminosity} as the usual $L_{\rm Edd} = 3.2\times10^{4}\,L_{\odot}\,(M_{\rm bh}/M_{\odot})$, and the ``Eddington mass-accretion rate'' as the accretion rate which would produce $L_{\rm Edd}$ given a canonical reference radiative efficiency $\epsilon_{r} = 0.1$ ($L = \epsilon_{r}\,\dot{M}\,c^{2}$), so $\dot{M}_{\rm Edd} \approx M_{\rm bh}/(45\,{\rm Myr})$.} with a canonical radiative efficiency of $\sim 0.1$ is $\sim 50\,$Myr, almost all of these models require a sustained period of super or hyper-Eddington accretion in the early Universe to be viable \cite[e.g.,][]{PezzulliValiante2016}. This is especially important at masses $\ll 10^{5}\,M_{\odot}$, as various studies have shown that once larger ``super-massive'' mass scales are reached, the gravity of the BH can capture gas from larger radii and lead to runaway growth \citep{LiHernquist2007,DiMatteoColberg2008,2012MNRAS.424.1461L,2013ApJ...771..116J,WeinbergerSpringel2018,2018MNRAS.478.5063H,ZhuLi2020,2021ApJ...917...53A}. But unless one invokes exotic formation mechanisms, a sustained rapid accretion phase is necessary to grow BHs from the stellar ($\sim 10-100\,M_{\odot}$) to super-massive ($\gg 10^{4}\,M_{\odot}$) mass scale \citep{2012MNRAS.424.1461L,2016MNRAS.457.3356V}
There is a well-established and rapidly-growing body of work demonstrating that compact objects can, in fact, exceed the naive ``Eddington accretion rate'' $\dot{M}_{\rm Edd}$ by large factors (up to $\gtrsim 1000$) on scales of the accretion disk itself (recently, see e.g.\ theoretical arguments by \citealt{2016MNRAS.459.3738I,2019ApJ...880...67J,2020ApJ...905...92P,2021PASJ...73..450K,2022PASJ..tmp...17B}, empirical arguments in \citealt{2021A&A...645A..78B,2022MNRAS.509.3599T}, or for reviews, \citealt{PezzulliValiante2016,2019ffbh.book..195M,2019ConPh..60..111S,2019BAAS...51c.352B} and references therein). But these studies generally assume a constant hyper-Eddington ($\sim 10^{3}\,\dot{M}_{\rm Edd}$) influx of gas from larger scales onto the accretion disk as their ``outer boundary condition.'' What remains deeply unclear is whether or not a seed BH -- especially at stellar mass scales -- could actually capture gas from the interstellar medium at a sufficient rate to sustain this accretion, and for long enough that the total mass supplied would be able to grow the BH by many $e$-foldings. There has been some theoretical work on the topic, but it has generally either considered idealized models where the gas around the seed sits in a common potential well and accretes instead of being multi-phase and turbulent, rapidly forming stars \citep[see e.g.][]{2020MNRAS.497..302T,2022arXiv220106584P}, or considered only galactic ($\gg$\,pc) scales \citep[e.g.][]{2022arXiv220108766M} where especially with BHs already $\gg 10^{4}\,M_{\odot}$, sustaining super-Eddington inflow to a nuclear region at least appears viable \citep{2018MNRAS.478.5063H,ReganDownes2019,ZhuLi2020,2021ApJ...917...53A}.
The problem is that in the realistic ISM, order-of-magnitude estimates such as those in \citet{2013ApJ...771..116J} suggest that the rate of gravitational capture of gas from the surrounding ISM -- the Bondi-Hoyle rate \citep{HoyleLyttleton1939,Bondi1952} -- should be extremely small unless the seed is already super-massive. Consider the standard expression
\begin{align}
\dot{M}_{\rm Bondi} \approx \frac{4\pi\,G^2\,M_{\rm bh}^2\rho}{(c_{\rm s}^2+\delta V^2)^{3/2}}.
\label{equ:bondi-hoyle-rate}
\end{align}
where $\rho$, $c_{\rm s}$, and $\delta V$ are the density, sound speed, and gas-BH relative velocity. In the diffuse/warm ISM, this gives $\dot{M}_{\rm Bondi}/\dot{M}_{\rm Edd} \sim 10^{-6}\,(M_{\rm bh}/10\,M_{\odot})\,(n/{\rm cm^{-3}})$ -- vastly sub-Eddington. In dense ($n \gtrsim 100\,{\rm cm^{-3}}$) cold molecular gas (sound speed $\sim 0.1\,{\rm km\,s^{-1}}$), $\dot{M}_{\rm Bondi}$ would be much larger {\em if the gas were laminar and the BH stationary} -- this is akin to the idealized non-turbulent models above. The problem is that realistic cold molecular gas in the ISM is clumpy and dynamical and turbulent, with star formation and stellar feedback generating large random motions -- i.e.\ large $\delta V$ \citep{larson:gmc.scalings,goodman:1998.lws.dependence.on.tracers,evans:1999.sf.gmc.review,stanimirovic:1999.smc.hi.pwrspectrum,elmegreen:2004.obs.ism.turb.review}. As we show below, assuming relative velocities are of order typical gravitational/virial velocities in the cloud then gives $\dot{M}_{\rm Bondi}/\dot{M}_{\rm Edd} \sim 10^{-4}\,(\langle n_{\rm cl} \rangle / 100\,{\rm cm^{-3}})^{1/2}\,(M_{\rm bh}/10\,M_{\odot})\,(10^{6}\,M_{\odot}/M_{\rm cl})$ -- once again, vastly sub-Eddington. Previous analytic and simulation models of this ``turbulent Bondi-Hoyle problem'' in idealized driven turbulence have argued that vorticity and turbulent magnetic fields will suppress the {\em average} accretion rates even relative to this (pessimistic) result \citep{KrumholzMcKee2006,BurleighMcKee2017}.
However, it is also clear from many studies of star formation that turbulence in dense gas also promotes the existence of extremely dense shocks and clumps in the gas \citep[see e.g.][]{klessen:2000.pdf.supersonic.turb,elmegreen:sf.review,vazquez-semadeni:2003.turb.reg.sfr,MacLowKlessen2004,federrath:2008.density.pdf.vs.forcingtype,goodman:2009.dendrogram.sims,federrath:2010.obs.vs.sim.turb.compare,hopkins:2012.intermittent.turb.density.pdfs,squire.hopkins:turb.density.pdf}, which can have low internal velocity dispersions and play a crucial role in turbulent fragmentation and star formation \citep{mckee:2007.sf.theory.review,hennebelle:2008.imf.presschechter,hopkins:excursion.ism,hopkins:excursion.imf,hopkins:excursion.clustering,hopkins:excursion.imf.variation,hopkins:frag.theory,guszejnov:gmc.to.protostar.semi.analytic,murray:2017.turb.collapse}. So it is possible that a more realistic model might allow for hyper-Eddington accretion in rare (but not impossible) cases in these environments. In this study, we therefore extend the series of simulations of dense, star forming environments used previously to study star and star cluster formation in \cite{GrudicHopkins2018,guszejnov:2018.isothermal.nocutoff,GrudicGuszejnov2018,grudic:max.surface.density,grudic:sfe.gmcs.vs.obs,grudic:2019.imf.sampling.fx.on.gmc.destruction,ShiGrudic2021}, to explore BH seed growth in dynamic, star-forming environments akin to dense giant molecular clouds (GMCs) and galactic nuclei.
In this first study, we neglect feedback from the accreting BHs themselves. This is obviously a major simplification, especially for BHs accreting above the Eddington limit -- however, the form and strength of feedback from BHs in this regime remains highly uncertain (see references above), and we wish to identify whether or not sustaining hyper-Eddington accretion is even remotely possible on these scales. Clearly, accretion {\em without} BH feedback represents a relatively strong upper limit to the maximum possible BH seed growth. We can then use the conditions identified here as necessary for such accretion to run simulations including BH feedback, with various parameterizations.
In \S~\ref{sec:simulations}, we describe our simulation methods. Then in \S~\ref{sec:results} we present results, including BH mass evolution in different clouds and its dependence on different initial conditions (ICs). In \S~\ref{sec:discussions}, we analyze the effects of different physics and simulation ICs, give simple analytic formulae for the conditions required for runaway accretion, and discuss some major caveats of our work in \S~\ref{sec:caveats}. Finally, we conclude in \S~\ref{sec:conclusions}.
\begin{footnotesize}
\ctable[caption={{\normalsize Initial conditions (ICs) of our ``fiducial'' reference simulations. Here we show three groups of simulations with low, medium, and high initial mean surface density ($\Bar{\Sigma}_0$). In each group, the clouds have radii ($R_{\rm cl}$) of 5, 50, 500 pc.. Subsequent columns give the approximate initial total cloud mass ($M_{\rm cl}$), initial free-fall time ($t_{\rm ff,\,c}$), gas cell mass/resolution ($m_{\rm gas}$), Plummer-equivalent force softening for star particles ($\epsilon_{\rm soft}^{\rm star}$), and for BHs ($\epsilon_{\rm soft}^{\rm bh}$), and additional notes.}}\label{tab:clouds},center,star,
]{llllllll}{
}{
\toprule
$\Bar{\Sigma}_0$ [$M_{\odot}$/pc$^2$] & $R_{\rm cl}$ [pc] & $M_{\rm cl}$ [$M_{\odot}$] & $t_{\rm ff,\,c}$ [Myr] & $m_{\rm gas}$ [$M_{\odot}$] & $\epsilon_{\rm soft}^{\rm star}$ [pc] & $\epsilon_{\rm soft}^{\rm bh}$ [pc] & Notes \\
\midrule
130 & 5 & $10^{4}$ & 2 & 0.005 & 0.04 & 0.04 & \\
130 & 5 & $10^{4}$ & 2 & 0.04 & 0.09 & 0.09 & No-feedback (low-resolution) variant\\
130 & 50 & $10^{6}$ & 6 & 0.5 & 0.21 & 0.21 & \\
130 & 500 & $10^{8}$ & 20 & 50 & 0.96 & 0.31 & \\
\midrule
1300 & 5 & $10^{5}$ & 0.6 & 0.4 & 0.19 & 0.19 & \\
1300 & 50 & $10^{7}$ & 2 & 40 & 0.89 & 0.31 & \\
1300 & 500 & $10^{9}$ & 6 & 500 & 2.06 & 0.31 & \\
\midrule
13000 & 5 & $10^{6}$ & 0.2 & 0.5 & 0.21 & 0.21 & \\
13000 & 5 & $10^{6}$ & 0.2 & 4 & 0.41 & 0.31 & Varied metallicity test series\\
13000 & 50 & $10^{8}$ & 0.6 & 6 & 0.48 & 0.31 & Highest resolution; $M_{\rm bh}\in(10, 100)\,M_\odot$\\
13000 & 50 & $10^{8}$ & 0.6 & 50 & 0.96 & 0.31 & \\
13000 & 50 & $10^{8}$ & 0.6 & 400 & 1.91 & 0.31 & Varied BH seed number test series\\
13000 & 500 & $10^{10}$ & 2 & 40000 & 8.89 & 0.31 & \\
\bottomrule
}
\end{footnotesize}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{figures/visualization-11}
\vspace{-20 pt}
\caption{An example visualization of the simulations at different scales. The simulation starts with a massive dense gas complex of $M_{\rm cl}=10^8~M_{\odot}$, $R_{\rm cl}=50~{\rm pc}$, and $256^3$ resolution, and we choose the snapshot where the seeds have the highest accretion rate. Here we show the distribution of the gas surface density as the black-red-orange color map. As stars, we show the BH seeds that undergo most significant growth. \textbf{Left}: Visualization of the whole complex; we see the BHs that accrete most are located at the dense region near the center of the system. \textbf{Middle}: 10 times zoom-in centered on the BH that shows the most significant mass growth at this time. \textbf{Right}: Further zoom-in of the region near the BH and the velocity field of gas (vectors, length proportional to magnitude of $|\vec{v}_{\rm bh}-\vec{v}_{\rm gas}|$) in the vicinity of the BH. Here the circle denotes the sink radius. We see a dense clump intersecting the sink radius, with a the relative velocity that is quite small near the BH.}
\label{fig:visualization}
\end{figure*}
\section{Simulations}
\label{sec:simulations}
Our simulation numerical methods are identical to those described and tested fully in \citet{GrudicHopkins2018,GrudicGuszejnov2018,HopkinsWetzel2018,GrudicGuszejnov2021,GrudicKruijssen2021}, modulo the addition of BH seeds described below, so we briefly summarize here. We use the code \textsc{GIZMO}\footnote{A public version of {\small GIZMO} is available at \href{http://www.tapir.caltech.edu/~phopkins/Site/GIZMO.html}{\url{http://www.tapir.caltech.edu/~phopkins/Site/GIZMO.html}}} \cite[]{Hopkins2015} in Meshless Finite Mass (MFM) mode, with magnetohydrynamics (MHD) solved as in \citet{hopkins:mhd.gizmo,hopkins:cg.mhd.gizmo}, self-gravity with adaptive Lagrangian force-softening, radiative cooling from $1-10^{10}$\,K, including molecular, metal-line, fine-structure, photo-electric, ionization and other processes as well as star formation in dense, locally-self-gravitating gas \citep{hopkins:virial.sf,GrudicHopkins2018}, and stellar feedback following the FIRE-2 implementation of the Feedback In Realistic Environments (FIRE\footnote{\href{http://fire.northwestern.edu}{\url{http://fire.northwestern.edu}}}) physics \citep{HopkinsWetzel2018,hopkins:fire3.methods}. In these models ``star particles'' each repreent IMF-averaged ensembles of stars (rather than resolving individual stars and proto-stars as in \citealt{grudic:starforge.methods,guszejnov:2020.starforge.jets}), which evolve along standard stellar evolution models to return mass, metals, momentum, and energy to the ISM in the form of supernovae and O/B and AGB winds \citep{hopkins:sne.methods} as well as acting on the gas via radiative heating, photo-ionization, and radiation pressure \citep{hopkins:radiation.methods}. Simulations with these methods have been previously used to study many properties of GMCs, galactic nuclei, and star clusters, including their observed star formation efficiencies, cluster dynamics and mass profiles, young massive cluster internal structure, globular cluster demographics, and gas emission properties \citep[see references above and e.g.][]{GrudicHopkins2018,GrudicGuszejnov2018,GrudicGuszejnov2021,GrudicKruijssen2021,FukushimaYajima2021}.
We extend these simulations by adding a population of ``seed'' BHs (sink particles) to the ICs, which can accrete gas from the surrounding medium, but otherwise feel only gravitational dynamics (we do not model BH feedback or BH-BH mergers).
\subsection{Black Hole Accretion}
\label{sec:simulations:bh-accretion}
Our BH seeds/sink particle prescription is a simplified version of that numerically presented in \citet{grudic:starforge.methods}. Gas is accreted onto a sink if it meets three criteria:
\begin{enumerate}
\item It is within the sink radius $r_{\rm sink}$ of the BH: $r=|\vec{r}_{\rm gas}-\vec{r}_{\rm bh}|< r_{\rm sink}$.
\item It is bound to the BH, including kinetic, thermal, and magnetic support: $u_{\rm thermal} + (1/2)\,v_{\rm A}^{2} + (1/2)\,\delta V^{2} < G\,M_{\rm sink}/r$, where $u_{\rm thermal}$ is the specific thermal energy, $v_{\rm A}$ the {Alfv\'en}\ speed, and $\delta V^{2} \equiv |\vec{v}_{\rm gas} - \vec{v}_{\rm bh}|^{2}$.
\item Its angular momentum is insufficient to support a circular orbit with radius larger than $r_{\rm sink}$ \cite[]{BateBonnell1995}, i.e.\ $j_{\rm gas} <\sqrt{G\,M_{\rm sink}\,r_{\rm sink}}$ where $j_{\rm gas}$ is the specific angular momentum of the gas cell (evaluated at its center-of-mass location).
\end{enumerate}
If a gas cell somehow meets all these criteria with two BHs simultaneously, it will accrete onto whichever is closer.
We must choose $r_{\rm sink}$ in each simulation. This is usually set to something like the simulation resolution (typical inter-cell separation $\delta r$), and would ideally resolve the Bondi radius, $R_{\rm Bondi} \sim G\,M_{\rm bh}/(c_{\rm s}^{2} + \delta V^{2})$, i.e.\ $r_{\rm sink} \sim R_{\rm Bondi} \gtrsim \delta r$. But in our Lagrangian, dynamical simulations (1) the spatial resolution is not fixed, but scales as $\delta r \sim (\rho/m_{\rm gas})^{1/3}$, and (2) the Bondi radius fluctuates dramatically (as we will show), and varies between seeds. In the ``worst case'' scenario, assume accretion is coming from the low-density diffuse intra-cloud medium (density $\rho \sim \langle \rho \rangle \sim 3\,M_{\rm cl}/4\pi\,R_{\rm cl}^{3}$) with virial or free-fall level relative velocities $\delta V \sim v_{\rm cl} \sim (G\,M_{\rm cl}/R_{\rm cl})^{1/2} \gg c_{s}$. This would give $R_{\rm Bondi} \sim (M_{\rm bh}/M_{\rm cl})\,R_{\rm cl}$, so resolving the Bondi radius ($\delta r \lesssim R_{\rm Bondi}$ in the same diffuse mean-density gas) would require a prohibitive number of cells $N_{\rm cells} \sim (R_{\rm cl}/\delta r)^{3} \gtrsim (M_{\rm cl}/M_{\rm bh})^{3}$. However, as we noted above and will show more rigorously below, the accretion rates from such diffuse gas are orders-of-magnitude below Eddington, and (even if well-resolved) would contribute essentially nothing to the total BH accretion in our simulations. Therefore consider instead the ``best-case'' scenario for accretion: since the turbulence in the molecular clouds has rms Mach numbers $\mathcal{M}_{\rm cl} \sim v_{\rm cl}/c_{s} \sim 10-100$, radiative shocks can produce regions with very high densities $\rho \sim \langle \rho \rangle\,\mathcal{M}_{\rm cl}^{2}$, and low relative velocities $\delta V \lesssim c_{s}$ \citep{Vazquez-Semadeni1994,PadoanNordlund1997,MacLowKlessen2004}. Under these conditions, the Bondi radii will be well-resolved ($\delta r \lesssim R_{\rm Bondi}$) so long as $N \gtrsim \mathcal{M}^{-8}\,(M_{\rm cl}/M_{\rm bh})^{3}$ -- a huge relief ($\propto \mathcal{M}^{8}$) in resolution requirements (which would be easily satisfied by every simulation in this paper). As we will show, regions akin to this idealized example dominate the actual accretion in the simulations.
In practice, we choose a sink radius by estimating a ``characteristic'' Bondi radius $b_{\rm c}$ by assuming $M_{\rm bh,\,c}=100\,M_\odot$, and considering two limits: $\delta V \lesssim c_{\rm s}$ (assuming a mean temperature of $100\,$K, typical in our simulations) so $b_1 = G\,M_{\rm bh,\,c}/c_{\rm s}^{-2}$, and $\delta V \sim v_{\rm cl} \gg c_{\rm s}$ so $b_2 \approx (M_{\rm bh}/M_{\rm cl}) R_{\rm cl}$, and then take $b_{\rm c}=\min (b_1, b_2)$. We have verified in post-processing that in all cases which produce ``interesting'' runaway BH growth, the Bondi radii {\em during the phase where the BH actually accretes rapidly} is at least marginally resolved, as expected from the argument above.
\subsection{Initial Conditions}
\label{sec:simulations:ics-setups}
We sample spherical, turbulent, and non-rotating molecular clouds or cloud complexes with different initial mean surface density ($\bar{\Sigma}_0 \equiv M_{\rm cl}/\pi\,R_{\rm cl}^{2} \approx 100, 10^3, 10^4\,M_\odot/{\rm pc}^2$) and initial radius ($R_{\rm cl}=5, 50, 500\,{\rm pc}$) following the setup and results of \citetalias{GrudicHopkins2018}, where each group with the same surface density was shown to have similar star formation efficiency. Note that these parameters are motivated by massive, dense star-forming cloud and ``clump'' complexes seen in high-redshift galaxies and starburst galaxy nuclei, with only the smaller and lowest-$\bar{\Sigma}_{0}$ clouds analogous to massive GMCs in the Milky Way. Each initial cloud is uniformly magnetized, we also set $E_{\rm turb}/|E_{\rm grav}|=1$ and $E_{\rm mag}/E_{\rm grav}=0.1$, where $E_{\rm turb}$, $E_{\rm mag}$, and $E_{\rm grav}$ are the turbulence (kinetic) energy, magnetic field density, and gravitational binding energy respectively. The clouds serve as the mass reservoirs for BH accretion.
We then insert an ensemble of BH seeds into the IC. Typically, for every seed, the mass ranges within $1~M_\odot \le M_{\rm bh} \le 10^4~M_{\odot}$ and are uniformly distributed in $\log M_{\rm bh}$. The initial position of seeds are sampled randomly but statistically uniformly within the cloud. The initial velocity is sampled such that in each dimension it is uniformly distributed in $[-V_{\rm circ},V_{\rm circ}]$ while the total magnitude is suppressed below $V_{\rm circ}$ to ensure the seeds are bound to the cloud, where $V_{\rm circ}^2 = GM_{\rm cl}(<r)/r$ is the local circular velocity at radius $r$ (assuming uniform mass distribution). We resample seeds which would be within a small distance to the cloud ``edge'' with an outward radial velocity, since these would trivially escape without interesting dynamics.
Rather than simulating only a few BH seeds in one cloud, we include a large number of seeds in every IC so that we can sample many different seed masses and positions and kinematics. However, to avoid significant interactions among the BHs and heavy computational costs, the number of BH seed is controlled to be either below 10000, or the number such that the total BH mass does not exceed $5\%$ of the cloud initial mass $M_{\rm cl}$. For low mass clouds, we decrease the lower and upper bounds of BH seed mass sampling to ensure a sufficient number of BH seeds, which also helps ensure the Bondi radii are resolved (e.g., for $M_{\rm cl}= 10^{4}\,M_{\odot}$, $1\,M_\odot \le M_{\rm bh} \le 100 \,M_\odot$; for $M_{\rm cl}= 10^{5}\,M_{\odot}$, $10\,M_\odot \le M_{\rm bh} \le 10^3 \,M_\odot$, for $M_{\rm cl} \gtrsim 10^{6}\,M_{\odot}$, $10^{2}\,M_\odot \le M_{\rm bh} \le 10^{4} \,M_\odot$).
We use adaptive force softening to avoid divergences in our gravity evaluation or extremely small time steps. For the newly formed stars, which have the same mass as gas particles, the minimum softening length is $r_{\rm soft}^{\rm star} \sim ( m_{\rm gas}/\rho_{\rm sf})^{1/3}$, where $m_{\rm gas}$ is the mass resolution of the cloud and $\rho_{\rm sf}$ is the numerical minimum density for star formation ($1000\,{\rm cm}^{-3}$ in the simulation). For BHs, the softening radius is set as $r_{\rm soft}^{\rm bh} = \min (r_{\rm soft}^{\rm star}, b_{\rm c})$, where $b_{\rm c}$ is the characteristic Bondi-Hoyle accretion radius (introduced in \S~\ref{sec:simulations:bh-accretion}). In the simulation $r_{\rm sink}=r_{\rm soft}^{\rm bh}$, so the setup ensures the code resolves the Bondi-Holye accretion radius and the BHs interact reasonably with star particles.
For reference, we show the initial-conditions in Table~\ref{tab:clouds}. The clouds are divided into three groups with different initial mean surface density $\Bar{\Sigma}_0 = M_{\rm cl}/(\pi R_{\rm cl}^2)$, though within each group the clouds have the same set of initial radii. The fiducial resolution (number of initial gas cells) of our simulations is $128^3$, while a few low ($64^3$) and high ($256^3$) resolution runs of a subset of clouds are also included for comparison. For each fiducial simulation, the termination time scale is $2\, t_{\rm ff}$, where $t_{\rm ff}=\pi\sqrt{R_{\rm cl}^3/(8G M_{\rm cl})}$ is the initial free-fall time scale of the cloud, while for low-resolution ones the termination time is $5\,t_{\rm ff}$. Finally, BH mergers are disabled since the event rate is not significant.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{figures/mass-growth.pdf}
\vspace{-10 pt}
\caption{Mass growth of the total mass of stars as well as (up to) five BHs that show the most significant mass growth, in the ``default'' simulations from Table~\ref{tab:clouds}, as a function of time (in units of the initial homogeneous-cloud free-fall time $t_{\rm ff}$). Here solid lines representing the mass growth of individual BHs, dotted lines representing the total mass growth of all stars summed. We show properties of complexes with different radii/masses ($r_{\rm cl}$): 5\,pc (blue), 50\,pc (orange), 500\,pc (green); and different initial mean surface density ($\Bar{\Sigma}_0$): 130\,$M_{\odot}/{\rm pc}^2$ (left), 1300\,$M_{\odot}/{\rm pc}^2$ (middle), 13000\,$M_{\odot}/{\rm pc}^2$ (right). We also show the highest-resolution run with $M_{\rm cl}=10^8\,M_\odot$ and $R_{\rm cl}=50\,{\rm pc}$ (red lines in the right panel).
The low-density complexes feature almost no BH growth. Higher-density systems generally feature a small number of seeds which ``run away'' and can grow by orders of magnitude.
}
\label{fig:mass-growth}
\end{figure*}
\section{Results}
\label{sec:results}
In this section we show the major results of the simulations. As a first impression, we present the morphology of one example GMC in Fig.~\ref{fig:visualization}, which has $M_{\rm cl}=10^8\,N_\odot$, $R_{\rm cl}=50\,{\rm pc}$, and resolution of $256^3$. After $0.55\,t_{\rm ff}$ of evolution, the GMC has become quite turbulent. We also show the 5 BHs that show the most significant mass growth during the period, which are generally located near the center of the GMC. For the BH that grows most rapidly during the period (the orange star in Fig.~\ref{fig:visualization}), we show the zoomed-in distribution of gas\footnote{For illustration purposes, the color is scaled nonlinearly with the density field so as to better illustrate its morphology.} and its velocity field in the middle and right-hand panels. Near the BH's sink radius (0.313 pc), there is a dense gas clump which has very low velocity compared to the gas at the edge of the view ($\sim 50\,{\rm km/s}$); this is rapidly accreted in the time between the snapshot shown and the next simulation snapshot.
This essentially fits our expectations from Bondi-Hoyle theory, applied {\rm locally} at scales of order the Bondi radius: high gas density and low relative velocity between the BH and nearby gas create the ideal conditions for growth.
\subsection{Seed growth in different clouds}
\label{sec:growth.vs.cloud}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/bh-mass-cdf.pdf}
\vspace{-20 pt}
\caption{Distribution of fractional mass growth in different ICs: we plot the cumulative fraction of the BH seeds which accreted some amount of mass. We denote initial surface density with color, and initial cloud radius/mass with linestyle (as labeled). Most BHs do not accrete significantly (see lower-end cutoffs). But the most high-density, massive clouds generally feature a very small number of seeds which runaway to enormous masses. }
\label{fig:cdf}
\end{figure}
As described in the previous section, in each cloud we sampled a large number of BH seeds to study their mass growth. In Fig.~\ref{fig:mass-growth} we present the mass evolution of (up to) 5 BHs in each simulation that show the most significant mass growth. As we show below, these are {\em not} necessarily the most massive seeds in the ICs.
For clouds with low initial surface density ($\Bar{\Sigma}_0=127\,M_{\odot}/{\rm pc^2}$) the mass growth is modest: essentially no BHs grow by more than a factor $\sim 2-3$, and in general even the most-rapidly growing only increase in mass by tens of percent. At the larger surface densities we sample, the mass of the most-rapidly-growing BHs typically increases by at least two orders of magnitude.
Ignoring the low surface density complexes, if we consider clouds with fixed $\Bar{\Sigma}_{0}$ but different sizes $R_{\rm cl}$ (or equivalently, masses $M_{\rm cl} \equiv \pi\,\Bar{\Sigma}_{0}\,R_{\rm cl}^{2}$), we see that the final masses of the single most-rapidly-growing BHs increase with the total cloud mass, reaching as high as $\sim 3-10\%$ of the total cloud mass. Interestingly, for the lower-mass complexes, we often see several BHs exhibiting similar growth, while for the most massive complexes ($R_{\rm cl} = 500\,{\rm pc}$), one BH runs away early and then proceeds to accrete a huge amount of gas, ``stealing'' the gas supply from other seeds in the cloud.
From the same plot, we also see that the BHs typically grow their mass quickly in a short range of time ($\Delta t \lesssim t_{\rm ff}$) starting at some time near $t \sim t_{\rm ff}$. However, for clouds with higher surface density, we see the time range becomes slightly longer; the BHs in those clouds also start to grow somewhat earlier. Moreover, as we will show below in more detail in some illustrative examples, BH growth always {\em follows} the formation of a significant mass of stars. All these features inspire us to study the effect of star formation and stellar feedback in different clouds, which is discussed in \S~\ref{sec:discussions:feeback}.
As a different way to study the probability of mass growth, we show the cumulative distribution of the final-initial mass difference for all the BHs in Fig.~\ref{fig:cdf}. For most of the BHs there is no large significant mass growth except for a small fraction ($\lesssim 10\%$) of them, which we will discuss in more detail below.
\subsection{Dependence on ICs}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/initial-mass-dm.pdf}
\vspace{-20 pt}
\caption{Dependence of BH growth on initial seed BH mass.
\emph{Top}: We aggregate all seeds in the three fiducial runs ($R_{\rm cl}=5,\,50,\,500\,$pc) for our highest-density clouds, and plot both the total number of BHs in different $\sim 0.1$\,dex-wide bins of initial BH mass ($\sim 1000$ per bin, summing across the simulations), and the number in each bin which capture/accrete a mass $\Delta M_{\rm bh}$ in excess of some fixed amount. The probability of runaway accretion depends only very weakly on initial seed mass, over this range.
\emph{Bottom}: Same as above, but for the highest resolution cloud ($M_{\rm cl}=10^{8}\,M_\odot$, $R_{\rm cl}=50\,{\rm pc}$) which has lighter seeds (10--100 $M_\odot$). Again, there is shallow dependence on the initial seed mass.
}
\label{fig:initial-mass}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=.8\linewidth]{figures/initial-position-velocity.pdf}
\vspace{-10 pt}
\caption{Initial value of the BH seed velocity $v_{\rm ini} = |\vec{v}_{\rm bh}|$ (relative to the initial complex center-of-mass velocity, in units of the characteristic gravitational velocity of the complex $v_{\rm cl} = \sqrt{G\,M_{\rm cl}/R_{\rm cl}}$) and initial position $r_{\rm ini} = |\vec{r}_{\rm bh}|$ (relative to the initial complex center-of-mass, in units of its radius $R_{\rm cl}$), for all seeds in clouds of a given initial surface density (aggregating the three different-mass/size runs). Seeds with mass accretion are color-coded by the amount of mass growth (final-to-initial mass difference), while the rest are not (as grey dots). As long as the seeds do not begin too close to the ``edge'' of the cloud or with too large an initial velocity (in which case they tend to ``drift away'' rather than accrete), there is no strong preference for e.g.\ seeds to be at the cloud center for runaway accretion to occur. }
\label{fig:position-velocity}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/metallicity-dependence.pdf}
\vspace{-20 pt}
\caption{Metallicity effects on accretion. We show the cumulative accreted mass distribution for all BHs as Fig.~\ref{fig:cdf}, but for a set of $M_{\rm cl}=10^6\,M_\odot$ and $R_{\rm cl}=5\,{\rm pc}$ systems simulated at low resolution systematically varying the initial metallicity. There is no strong systematic metallicity dependence.}
\label{fig:metallicity}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/bh-number-dependence.pdf}
\vspace{-20 pt}
\caption{Number of BHs which exceed some final-to-initial mass ratio $r$ (as a proxy for ``runaway'' growth), in simulations of an otherwise identical complex (with properties labeled) where we vary the initial seed number systematically. We require $\sim 50-100$ seeds before there is reliably at least one seed that undergoes runaway growth. Increasing the seed number further, the fraction of seeds which experience at least {\em some} significant growth is $\sim 1\%$, but we see that typically no more than one to a few BHs run away most dramatically (regardless of the seed number provided it is sufficient). }
\label{fig:bh-number-dependence}
\end{figure}
In Fig.~\ref{fig:initial-mass} we show the dependence of mass growth ($\Delta M_{\rm bh}$) on the initial mass of the BH seeds. As we showed above, most seeds did not grow significantly. But more strikingly -- and perhaps surprisingly, given the strong dependence of the Bondi-Hoyle rate on $M_{\rm bh}$ -- we see that there is almost no correlation between the initial seed mass and BH mass growth. The particular simulation here considers seeds from $10^{2}-10^{4}\,M_{\odot}$, but we find the same (in the extended tests described below) for initial seed masses down to $\sim 10\,M_{\odot}$.
In Fig.~\ref{fig:position-velocity}, we present the initial velocity magnitude (relative to the cloud center-of-mass), initial position, and mass growth for all BHs in the simulation. As we can see, there is no strong dependence on either the initial position or velocity magnitude, {\em provided} the BH is (a) reasonably bound to the cloud (initial velocity not larger than $\sim v_{\rm cl} \sim (G\,M_{\rm cl}/R_{\rm cl})^{1/2}$), and (b) the BH does not begin too close to the edge of the cloud with a velocity directed away from the (irregular) centers of collapse (in which case the BH tends to drift away from the dense regions, rather than interact with them).
Another factor that could change the result is the initial metallicity $Z$, which self-consistently alters the cooling physics, stellar evolution tracks, and radiative feedback (opacities) in the simulations. We test this simulating GMCs with $M_{\rm cl}=10^6\,M_\odot$ and $R_{\rm cl}=5\,{\rm pc}$ (from the high surface density group) with varying initial $Z$ in Fig.~\ref{fig:metallicity}. By comparing the distribution functions of BH final-initial mass difference, we see that all those clouds produce statistically similar results for runaway BH accretion, independent of $Z$. We note that there are caveats regarding uncertainties in stellar evolution and treatment of molecular chemistry at extremely low metallicities in these models -- these are reviewed in detail in \citet{GrudicHopkins2018} -- but for all but truly metal-free clouds (where Pop-III evolution may be quite different) we regard this as robust. We also note that \citet{CorbettMoranGrudic2018} showed that the fragmentation and turbulent clumping in even metal-free clouds under high-density conditions like those of interest here are quite similar, independent of different molecular chemistry networks used for the thermochemistry.
We also change the number of BH seeds in the ICs ($N_{\rm bh,tot}$) and check the number of seeds that undergo significant mass growth, in Fig.~\ref{fig:bh-number-dependence}. Here we use different criteria to denote ``significant'': the final-initial mass ratio of the BH is above a constant $r$ and $r=100, 500, 2500$. If we simulate an initial number of seeds $N_{\rm bh,\,ini} \lesssim 64$, it becomes unlikely to see even a single seed undergo runaway growth, while for $N_{\rm bh,\,ini} \gtrsim 100$, we are essentially guaranteed that at least one seed will experience runaway growth. We find the same applying a more limited version of this test to other clouds. Thus there appears to be a threshold $\sim 1\%$ probability for a randomly-drawn seed to undergo runaway growth. However, if we increase the number of seeds further, the absolute number of BHs undergoing runaway accretion clearly saturates at a finite value, of $\sim$\,a few to ten with factor $10-100$ growth and $\sim 1$ with extreme runaway growth. Thus a given cloud can only support at most a few runaway BHs.
\subsection{Effects of other numerical parameters}
\subsubsection{Mass \&\ Force Resolution}
Resolution could influence our results both directly and indirectly. Ideally we would wish to ensure $m_{\rm gas} \ll M_{\rm bh}$, and that the Bondi radii are resolved (see \S~\ref{sec:simulations:bh-accretion}), but there are of course physics we cannot hope to resolve in our larger cloud complexes, such as the formation of individual stars (e.g.\ predicting the IMF itself). Nonetheless, we have tested our results for several clouds at different resolution levels: $64^3$, $128^3$, and $256^{3}$ (See Appendix~\ref{app:resolution-convergence}). For most clouds (especially those with high initial mean surface density), we see no significant different in our statistical/qualitative conclusions across these resolution levels. Similarly, we have re-run two of our clouds (one low and one high density) with factor $\sim 3$ different force-softening values for the collisionless (star and BH) particles, and see no significant effect. Thus our conclusions appear robust, but we caution again that qualitatively new physical processes would need to be simulated if we went to higher resolution (which are represented here by our sub-grid models for e.g.\ IMF-averaged stellar evolution).
\subsubsection{BH Sink Radii}
As noted above, in the simulation the sink radius for BH accretion is set as the smaller value of the Bondi-Hoyle accretion radius with either $\delta V \sim c_{\rm s}$ or $\delta V \sim v_{\rm cl}$. Analysis of our simulation shows that runaway accretion almost always occurs in regions where $\delta V \sim c_{\rm s} \lesssim 0.1\,v_{\rm cl}$, with enhanced densities $\gtrsim 100\,(3\,M_{\rm cl}/4\pi\,R_{\rm cl}^{3})$, where (as noted above) the Bondi radii are orders-of-magnitude larger than one would calculate for the diffuse GMC gas with low relative velocities. As a result the simulations relatively easily resolve the Bondi radius where accretion is relevant. Nonetheless we have re-run a small subset varying $r_{\rm sink}$ by more than an order of magnitude in either direction. If we make $r_{\rm sink}$ much too small -- significantly smaller than the rms inter-cell separation (spatial resolution in the gas) at the highest cloud densities ($\gtrsim 100\,(3\,M_{\rm cl}/4\pi\,R_{\rm cl}^{3})$), then we artificially see suppressed accretion (simply because we require the separation between BH and gas cell center-of-mass be $<r_{\rm sink}$ for capture). If we make the sink radius more than an order of magnitude larger than our default value, we see spurious accretion, usually in the very early timesteps of the simulations (before turbulence is fully-developed), where diffuse gas with low relative velocities is rapidly accreted. But for more reasonable variations in $r_{\rm sink}$, even spanning an order of magnitude, our results are robust. And as we show below, the accretion corresponds fairly well with analytic post-processing expectations, further lending confidence to this estimate.
\subsubsection{Initial BH Velocities}
We have considered some limited experiments where we add a systematic ``boost'' or arbitrarily increase the initial velocities of the BH seeds in the initial conditions. As expected, if the seeds are moving well in excess of the escape velocity relative to the cloud, they escape rapidly and never capture a large amount of gas. So the rms velocities of ``interesting'' BH seeds can only exceed $v_{\rm cl}$ by a modest (at most order-one) factor. On the other hand, reducing the BH seed velocities to zero has very little effect (other than introducing some spurious transient accretion in the first few timesteps when there is no relative gas-BH motion), because they quickly acquire random velocities of order the gravitational velocity $v_{\rm cl}$ from the fragmenting cloud potential.
\section{Discussion}
\label{sec:discussions}
\subsection{Effects stellar feedback \&\ global cloud properties}
\label{sec:discussions:feeback}
Intuitively, stellar feedback can alter BH accretion in two ways. i) Feedback expels gas, which makes it harder for BHs to capture that gas. ii) Feedback can make the cloud more turbulent and create more dense regions. As an example, we include low-resolution simulations with and without feedback physics for the same ICs in Fig.~\ref{fig:feedback_effects}. As we see, for this low-surface density cloud, feedback effectively blows gas away after two free-fall time scales. As a result, feedback suppresses both BH accretion and star formation -- BH growth in particular is suppressed by more than an order of magnitude, the difference between there being a few versus essentially no ``runaway'' BHs.
Star formation and stellar feedback in GMCs have been well studied in previous simulations that are related to this work (e.g.\ \citealt{GrudicHopkins2018}), as well as similar studies with different numerical methods which have reached very similar conclusions for star formation (e.g.\ \citealt{2019MNRAS.487..364L}). One important conclusion of these studies is that the integrated star formation efficiency, and effects of feedback, depend strongly on $\bar{\Sigma}_{0}$. Briefly: a young stellar population of mass $M_{\ast}$ produces a net feedback momentum flux (from the combination of radiation and stellar winds) $\dot{P} \sim \langle \dot{p}/m_{\ast} \rangle\,M_{\rm cl,\,\ast} \sim 10^{-7}\,{\rm cm\,s^{-2}}\,M_{\rm cl,\,\ast}$, while the characteristic gravitational force of the cloud on its gas is $F_{\rm grav} \sim G\,M_{\rm cl,\,tot}\,M_{\rm cl,\,gas} / R_{\rm cl}^{2} \sim G\,M_{\rm cl,\,gas}\,\bar{\Sigma}_{0}$. So the gas reservoir of a cloud is rapidly unbound and ejected when its stellar mass exceeds $M_{\rm cl,\,ast}/M_{\rm cl,\,gas} \gtrsim G\,\bar{\Sigma}_{0} / \langle \dot{p}/m_{\ast} \rangle \sim \bar{\Sigma}_{0} / (1000\,M_{\odot}\,{\rm pc^{-2}})$. So for our low-$\bar{\Sigma}_{0}$ clouds, almost all of the gas is rapidly un-bound after just a small fraction of the GMC forms into stars, preventing it from being accreted by BHs.
We can see this reflected in Fig.~\ref{fig:velocity-different-clouds}, which shows the gas rms bulk velocity $\langle |\vec{v}_{\rm gas}|^{2} \rangle^{1/2}$, gas sound speed $\langle c_{\rm s} \rangle$, and BH rms velocity $\langle |\vec{v}_{\rm bh}|^{2}|\rangle^{1/2}$ as a function of time in different ICs, in units of the characteristic cloud gravitational velocity $v_{\rm cl} \sim (G\,M_{\rm cl}/R_{\rm cl})^{1/2}$. Not surprisingly, the rms velocity of BHs remains of order the gravitational velocities. The gas bulk velocities are dominated by gravity at first so remain of order $v_{\rm cl}$, but when feedback begins to disrupt the cloud they increase in magnitude by a factor $\sim 10$. This effect depends primarily on $\bar{\Sigma}_{0}$, as expected from the argument above.
Similarly, the sound speed of the clouds initially drops extremely quickly owing to cooling until it reaches molecular temperatures (we arbitrarily start from higher temperature, but this has no effect on our results), with $c_{\rm s} \ll v_{\rm cl}$, but then rises once feedback disrupts the cloud owing to a combination of (1) photo-ionization, (2) shocks from stellar winds bubbles, and (3) lower gas densities increasing the cooling time. Since e.g.\ the characteristic photo-ionized $c_{\rm s} \sim 10\,{\rm km\,s^{-1}}$ is roughly constant, the importance of this effect depends primarily on $v_{\rm cl}$, which ranges from $\sim 3\,{\rm km\,s^{-1}}$ in our lowest-mass, lowest-density simulation, to $\sim 300\,{\rm km\,s^{-1}}$ in our highest-mass, highest-density simulation.
In our low-density, low-mass clouds, we see disruption occurs very early (less than $\sim 2\,t_{\rm ff}$), with the gas bulk velocities and sound speeds reaching $\gg v_{\rm cl}$. This makes gravitational capture of gas by seeds nearly impossible. For the intermediate-density clouds, we see the disruption is significantly delayed, and the magnitude of the post-disruption velocities is reduced (with $c_{\rm s} \lesssim v_{\rm cl}$ even during disruption). For the highest-density clouds, there is no real disruption but just dispersal of some residual gas after star formation completes.
We have also considered the impact of stellar feedback on the volume and mass fraction in dense clumps (Fig.~\ref{fig:density-different-clouds}). Specifically, we calculated the volume and mass (in units of the initial cloud volume and total mass) of regions/clumps within the cloud that satisfy $\rho > 100\,\langle \rho \rangle_{0}$ (where $\langle \rho \rangle_{0} \equiv 3\,M_{\rm cl}/(4\pi\,R_{\rm cl}^{3})$ is the initial mean cloud density). The volume/mass of dense clumps increases in all cases rapidly at early times as the cloud collapses, but in the low-density clouds it is rapidly truncated by feedback. In contrast, we see in the higher-density clouds a sustained ``lifetime'' of dense gas: this is driven by shocks and turbulence from feedback from the stars that have formed, but have insufficient power to completely disrupt the cloud. In the highest-density case we even see dense clumps re-emerge several times after $t\gtrsim 3\,t_{\rm ff}$ due to large-scale stellar feedback events -- these correspond to large wind/HII region shells colliding to form dense regions (see e.g.\ \citealt{MaGrudic2020} for more detailed discussion).
Another obvious requirement for runaway BH growth to ``interesting'' IMBH or even SMBH masses is that the total cloud mass is sufficiently large, such that the mass of dense ``clumps'' accreted is interesting. As we show below, the characteristic gas clump masses at high densities which meet the conditions for runaway BH growth are typically $\sim 1\%$ of the cloud mass, neatly explaining the maximum final BH masses seen in \S~\ref{sec:growth.vs.cloud}. This requires a total cloud mass $\gtrsim 10^{5}-10^{6}\,M_{\odot}$ for growth to true IMBH (let alone SMBH) scales. Interestingly, since $v_{\rm cl} \sim 15\,{\rm km\,s^{-1}}\,(M_{\rm cl}/10^{6}\,M_{\odot})^{1/4}\,(\bar{\Sigma}_{0}/10^{3}\,{\rm M_{\odot}\,pc^{-2}})^{1/4}$, this plus the surface density condition above simultaneously ensure that complexes are not over-heated or disrupted by photo-ionized gas.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{figures/feedback_effects.png}
\vspace{-10 pt}
\caption{Effects of stellar feedback on gas morphology and seed BH mass growth. We show two low-resolution runs of a low-density, low-mass cloud with $M_{\rm cl}=10^4~M_\odot$, $R_{\rm cl}=5~{\rm pc}$ cloud, including (middle) and excluding (right) stellar feedback effects, as Fig.~\ref{fig:visualization}. Gas is more concentrated near the center for the no-feedback run while most gas is expelled by stellar winds in the feedback run. We show the cumulative BH and stellar mass growth for these runs as Fig.~\ref{fig:mass-growth} (left): BH growth and star formation are both suppressed at the order-of-magnitude level in these lower-density clouds.}
\label{fig:feedback_effects}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{figures/velocity-gas-bh-cs.pdf}
\vspace{-10 pt}
\caption{Behavior of the rms gas bulk velocity dispersion ($\equiv | \vec{v}_{\rm gas} - \langle \vec{v}_{\rm gas} \rangle| $), sound speeds ($c_{\rm s}$), and BH seed particle velocity dispersions relative to the complex center-of-mass ($\equiv | \vec{v}_{\rm bh} - \langle \vec{v}_{\rm gas} |$), all in units of the characteristic gravitational/circular velocity of the complex ($v_{\rm circ} = v_{\rm cl} = \sqrt{G\,M_{\rm cl}/R_{\rm cl}}$), as a function of time (in units of the initial complex free-fall time ($t_{\rm ff}$). We compare the fiducial runs with different surface density (left-to-right) and radii (top-to-bottom). For each we show the median-absolute-value (line) and $14-86\%$ inclusion interval (shaded). Low-density, low-mass clouds are rapidly disrupted by stellar feedback producing large gas bulk velocities and high sound speeds. This is suppressed in high-density, high-mass systems, enabling continued BH accretion.}
\label{fig:velocity-different-clouds}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/density-evolution.pdf}
\vspace{-20 pt}
\caption{Evolution of the abundance of dense clumps in the simulations. We compare complexes with different initial mean surface density $\Bar{\Sigma}_0$ (different colors; only $R_{\rm cl}=50\,{\rm pc}$ shown for simplicity). For each, we measure the total gas mass (in units of the initial mass $M_{\rm cl}$) with a local gas density exceeding $\rho > 100\,\langle \rho\rangle_{0}$ (where $\langle \rho \rangle_{0} \equiv 3\,M_{\rm cl}/4\pi\,R_{\rm cl}^{3}$ is the initial mean density). In all cases this rises rapidly as gravitational collapse and turbulence develop. In low-density systems, it almost immediately drops to zero after stellar feedback unbinds the gas. In higher-density systems, the presence of dense gas is sustained and even rejuvenated later in the system evolution owing to the presence of strong shocks.}
\label{fig:density-different-clouds}
\end{figure}
\subsection{How Does Runaway Growth Occur?}
\label{sec:runaway}
We now consider the local conditions for runaway growth. In a small ``patch'' of cloud on scales $\sim r_{\rm sink}$ (small compared to the cloud but large compared to the BH accretion disk), it is not unreasonable to approximate the rate of gravitational capture of gas by a sink via the Bondi formula (Eq.~\eqref{equ:bondi-hoyle-rate}), given the local value of $\rho$, $\delta V$, and $c_{\rm s}$.
Recall, from \S~\ref{sec:intro} and \S~\ref{sec:simulations:bh-accretion}, that if we consider the typical or diffuse/volume-filling conditions within the cloud, i.e.\ $\rho \sim \langle \rho\rangle_{0} \sim 3\,M_{\rm cl}/4\pi\,R_{\rm cl}^{3}$ and $\delta V \sim v_{\rm cl} \gg c_{\rm s}$, we would obtain
\begin{align}
\langle \dot{M}_{\rm bh} \rangle_{\rm diffuse}& \sim \frac{G^2 M_{\rm bh}^2 \rho}{\delta V^3} \sim \frac{G^{1/2}\,M_{\rm bh}^{2}}{M_{\rm cl}^{1/2}\,R_{\rm cl}^{3/2}} \sim \left(\frac{\langle n_{\rm cl} \rangle}{{\rm cm^{-3}}}\right)\,\left( \frac{M_{\rm bh}}{M_{\rm cl}} \right)\,\dot{M}_{\rm Edd}
\end{align}
where $\langle n_{\rm cl}\rangle = \langle \rho \rangle_{0}/m_{p}$. If we further assume that the timescale for accretion $\Delta t$ is of order the cloud lifetime, $\sim t_{\rm eff} \sim \sqrt{R_{\rm cl}^3/GM_{\rm cl}}$, then the total mass accreted $\sim \dot{M}_{\rm bh}\,\Delta t$ would be
\begin{align}
\Delta M_{\rm bh} \sim \langle \dot{M}_{\rm bh} \rangle_{\rm diffuse}\,t_{\rm ff} \sim \frac{M_{\rm bh}}{M_{\rm cl}}\,M_{\rm bh}
\end{align}
In other words, unless the ``seed'' is already a large fraction of the entire GMC complex mass (i.e.\ is not really a seed in any sense), then the diffuse accretion will be highly sub-Eddington and the BH will grow only by a tiny fractional amount. This immediately explains why {\em most} of the seeds we simulate indeed grow negligibly.
However, in a highly turbulent cloud we argued above that two effects that may boost the mass growth: i) the dense clumps appear with $\rho \gg \langle \rho\rangle_{0}$, and ii) the turbulence velocity contributes to the relative velocity $\delta \vec{V}$ so locally, low $\delta V$ is possible.
In Fig.~\ref{fig:bondi-hoyle-check} we follow one particular but representative example of a sink which undergoes runaway growth, considering how the relevant factors in the local Bondi rate evolve in the immediate vicinity of the sink. The thermal sound speed is negligible at basically all times in the cold molecular phases compared to $\delta V$, as expected. Runaway growth therefore occurs when the BH happens to encounter a region which simultaneously features a strong local density enhancement, $\rho \sim 10^{3}-10^{4}\,\langle \rho\rangle_{0}$, and low relative velocity $\delta V \lesssim 0.1\,v_{\rm cl}$, below the escape velocity of gas from the sink radius (so it is indeed captured). This boosts the local Bondi accretion rate by a factor of $\sim 10^{7}$, compared to our estimate above for the ``diffuse'' cloud medium. Visual examination shows this resembles Fig.~\ref{fig:visualization} -- the BH happens (essentially by chance) to be moving with a very low relative velocity to a dense clump created nearby by intersecting shock fronts (with Mach $\sim 30-100$ shocks, i.e.\ $v_{\rm shock} \sim 10\,{\rm km\,s^{-1}}$, producing the large density enhancement), and begins accreting it extremely rapidly. Since the Bondi rate scales as $\propto M_{\rm sink}^{2}$, this runs away and most of the clump mass ($\sim 10^{5}\,M_{\odot}$) is rapidly accreted and the clump is tidally disrupted and then captured before it fragments internally to form stars.
Examination show this pattern is typical of the seeds which experience runaway accretion in the simulations.
Analytically, therefore, let us assume that during evolution, a BH encounters a dense clump with local density $\rho_{\rm c}$, clump radius $r_{\rm c}$, mass $\delta M_{\rm c}$, at relative velocity $\delta V_{\rm c}$ (and define $C^{2} = \delta V_{\rm c}^{2} + c_{\rm s}^{2}$, where we can generally assume $C \sim \delta V_{\rm c} \gtrsim c_{\rm s}$, even in regions where $\delta V_{\rm c}$ is relatively low), and accretes in Bondi-Hoyle-like fashion. Fig.~\ref{fig:analytic-scaling} summarizes the resulting accretion for various assumptions. Integrating the Bondi accretion rate for some time $\Delta t$ (assuming the background is constant), we have
\begin{align}
\label{equ:runaway-accretion} \frac{1}{M_{\rm bh,0}} - \frac{1}{M_{\rm bh,final}} = \frac{1}{M_{\rm bh,0}} - \frac{1}{M_{\rm bh,0} + \Delta M_{\rm bh}} & \sim \frac{4\pi\,G^2 \rho_{\rm c}}{C^{3}} \Delta t
\end{align}
where $M_{\rm bh,\,0}$ is the ``initial'' BH mass. This diverges (formally $\Delta M_{\rm bh} \rightarrow \infty$), so in practice the entire clump will be accreted ($\Delta M_{\rm BH} \rightarrow M_{\rm c}$), in a finite time $\Delta t \rightarrow t_{\rm acc} \sim C^{3}/(4\pi\,G^{2}\,\rho_{\rm c}\,M_{\rm bh,0})$. In practice, the time $\Delta t$ will be limited by the shortest of either the dense clump lifetime (usually not shorter than its freefall time $t_{\rm ff,\,c} \sim 1/\sqrt{G\,\rho_{\rm c}}$), the timescale for the clump to fragment and form stars (also no shorter than $t_{\rm ff,\,c}$), or the crossing time $t_{\rm cross} \sim r_{\rm c} / \delta V_{\rm c}$ for the mutual interaction. A simple calculation shows that the ratio $t_{\rm cross}/t_{\rm ff,\,c} \sim (\delta M_{\rm c}/M_{\rm cl})^{1/3}\,(\rho_{\rm c}/\langle \rho\rangle_{0})^{1/6}\,(v_{\rm cl}/\delta V_{\rm c})$. Inserting numbers or considering Fig.~\ref{fig:analytic-scaling} shows that for the conditions of greatest interest for rapid accretion ($\delta V_{\rm c} \ll v_{\rm cl}$, $\rho_{\rm c} \gg \langle \rho\rangle_{0}$, and clump masses $\delta M_{\rm c}$ not incredibly small compared to the mass of the cloud so large BH growth is possible), we usually expect $t_{\rm cross} \gtrsim t_{\rm ff,\,c}$. So considering a ``worst-case'' scenario, then, accretion can run away to accrete the entire clump when $t_{\rm acc} \lesssim t_{\rm ff,\,c}$, which in turn requires:
\begin{align}
\label{eqn:critical.condition} \frac{\delta V_{\rm c}}{v_{\rm cl}} \lesssim 0.1\,\left( \frac{M_{\rm bh,\,0}}{10^{-5}\,M_{\rm cl}} \right)^{1/3}\,\left( \frac{\rho_{\rm c}}{100\,\langle \rho \rangle_{0}} \right)^{1/6}
\end{align}
This corresponds reasonably well to the conditions where, in the simulations, we indeed see runaway growth -- regions with enhanced $\rho_{\rm c}$, and (crucially here), low local $\delta V_{\rm c}$. This also naturally explains why we see only a very weak dependence on initial BH mass -- provided this condition is met (which does not depend strongly on $M_{\rm bh,\,0}$), then the ``growth limiter'' is not the BH mass or Bondi rate (which depends strongly on $M_{\rm BH}$), but the mass of the clump $M_{\rm c}$ (which is, of course, entirely independent of the mass of the BH). Moreover, accretion events can occur sequentially, so once a BH ``jumps'' in mass by accreting a clump, its ``effective'' mass will be larger making it easier to accrete subsequent clumps (in an extreme form of competitive accretion).
Still, if the BHs were truly extremely small (e.g.\ $\ll 10\,M_{\odot}$), or the clouds extremely massive, then the probability of such an event would become small rapidly -- this may explain, in part, why for the most massive complexes we see fewer BHs grow (but those that do grow, grow to even larger masses).
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/bondi-hoyle-check.pdf}
\vspace{-20 pt}
\caption{A representative case study of the environment around one BH seed which undergoes runaway growth (complex properties labeled).
{\em Top:} BH/sink particle mass versus time (in units of the initial complex free-fall time $t_{\rm ff}$).
{\em Middle:} Median sound speed ($c_{s}$; dotted), and $5-90\%$ range of relative gas bulk velocities ($|\vec{v}_{\rm gas}-\vec{v}_{\rm bh}|$; shaded) of all gas cells which fall within the sink radius $r_{\rm sink}$, and the escape velocity from the sink ($\sim \sqrt{G\,M_{\rm sink}/r_{\rm sink}}$). Velocities are in units of the characteristic complex gravitational velocity $v_{\rm cl}=\sqrt{G\,M_{\rm cl}/R_{\rm cl}}$.
{\em Bottom:} Mean density of gas within the sink radius, in units of the initial complex mean density $\langle \rho \rangle_{0} \equiv 3\,M_{\rm cl}/4\pi\,R_{\rm cl}^{3}$.
The runaway growth occurs when the BH intercepts an overdense clump $\rho \gtrsim 100\,\langle \rho\rangle_{0}$, with a low relative velocity $\delta V \lesssim 0.1\,v_{\rm cl}$ so that it is gravitationally captured. Thermal pressure support/sound speed is relatively unimportant, since this is occurring in cold molecular gas.}
\label{fig:bondi-hoyle-check}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/analytic-scaling.pdf}
\vspace{-20 pt}
\caption{Illustrative analytic predictions for BH accretion, based on a simple model of Bondi-Hoyle accretion in encounters with sub-clumps that have local density $\rho_{\rm c}$, mass $\delta M_{\rm c}$, and relative BH-clump velocity $\delta V_{\rm c}$. Here we assume the initial seed mass $M_{\rm bh} = 10^{-4}\,M_{\rm cl}$.
{\em Top:} Assuming a fixed $\delta V_{\rm c}$ (in units of $v_{\rm cl}$), we plot behavior as a function of $\delta M_{\rm c}$ and $\rho_{\rm c}$.
{\em Bottom:} Behavior as a function of $\delta M_{\rm c}$ and $\delta V_{\rm c}$, at fixed $\rho_{c}$.
The blue shaded region denotes where the internal free-fall time of the clump ($t_{\rm ff,\,clump}$) is shorter than the clump-BH crossing timescale ($t_{\rm crossing}$), so will be the rate-limiting timescale for accretion.
Red shaded range shows where the total mass accreted ($\Delta M_{\rm bh}$) over the shorter of $\Delta t = {\rm min}(t_{\rm ff,\,clump},\ t_{\rm crossing})$ would be less than the clump mass ($\delta M_{\rm c}$), so the accretion does not fully ``run away.''
Green lines show contours where $\log_{10}(\Delta M_{\rm bh} / M_{\rm bh,\,0})$ is constant and equal to the value shown.
The region where $\Delta M_{\rm bh}/M_{\rm bh}$ contour is horizontal denotes ``runaway,'' defined by where the BH will accrete the entire clump mass.}
\label{fig:analytic-scaling}
\end{figure}
Finally, in Appendix~\ref{app:bh_accretion_rate_estimation}, we use this to make an order-of-magnitude estimate of the probability of a seed encountering a ``patch'' (i.e.\ clump) of gas meeting the criteria above. Assuming e.g.\ uncorrelated Gaussian velocity fields and lognormal density fields, we estimate that the probability of seeds encountering dense clumps is not low, but the probability of such an encounter also having low relative velocity meeting the condition above is, giving a net probability in the range $\sim 0.001-0.01$. This is remarkably similar (given the simplicity of these assumptions) to our estimate of $\sim 0.01$ from the simulations where we varied the number of seeds systematically, as discussed above.
\section{Caveats}
\label{sec:caveats}
\subsection{Feedback from Accreting Black Holes}
The most important caveat of this study is that we did not include any ``sub-grid'' model for BH accretion or feedback in the simulations. So ``BH accretion rate'' here should really be understood to be ``rate of gravitational capture of gas by the BH-disk system'' (akin to ``Bondi-Hoyle-like mass inflow rate'') and ``BH mass'' or ``sink mass'' represents a sum of the actual BH mass and its bound/captured material (whether that material has actually formed an accretion disk is another question itself).
This is not, of course, because we expect feedback to be unimportant for the BHs which rapidly capture gas: indeed, models of super-Eddington accretion disks (models whose ``outer boundary condition'' is something like the sink radii or ``inner boundary condition'' of our simulations) predict both strong radiative (luminosities near or somewhat above the Eddington luminosity) and kinetic (broad-angle MHD outflows from the disk) feedback (see references in \S~\ref{sec:intro}). While it is conceivable that under sufficiently-dense conditions, the surrounding material could continue to accrete (see e.g.\ \citealt{QuataertGruzinov2000,TakeoInayoshi2018,ReganDownes2019}), this could also completely shut down BH growth and even star formation in the surrounding cloud \citep{SchawinskiKhochfar2006}.
However, crucial details of such accretion and feedback processes remain deeply uncertain. This includes (i) the rate at which material can go from being ``gravitationally captured'' to actually accreted onto the BH (which determines the luminosity and other feedback); (ii) whether star formation and/or fragmentation occurs in the captured disk material if too much mass is captured; and (iii) for a given accretion rate, the radiated spectrum and energy, and the energy and momentum and mass and opening angle of accretion-disk winds. Our intention here is therefore to first identify a set of {\em necessary}, but perhaps not sufficient, pre-conditions for runaway hyper-Eddington seed growth in on ISM scales. Clearly, if a BH cannot sustain super-Eddington gravitational capture rates of sufficient total mass in the first place, then it is unlikely that adding feedback would do anything other than further decrease the (already minimal) accretion. This allows us to already rule out large segments of parameter space as viable locations for hyper-Eddington accretion (e.g.\ Milky Way-like low-density or low-mass clouds, systems with insufficient statistical sampling of ``seeds,'' highly-unbound seeds). In future work (in preparation), we will use this study as the basis for a follow-up set of simulations which do include BH feedback, systematically surveying parameterized models for the major uncertainties described above, but using the results of this study to specifically begin from conditions where we know that, {\em absent} BH feedback, rapid accretion is possible.
\subsection{Other Caveats}
There are also of course other caveats in our simulations themselves. While we survey a factor of $\sim 100$ in mass resolution and see robust results, we are certainly not able to ``converge'' in a strict sense, especially given some ISM micro-physics where the relevant dynamics occur on sub-au scales. We cannot resolve or predict the IMF or stellar evolution tracks of individual stars, let alone their late-stage evolution and potential collapse into BH relics. This is especially unfortunate as one might imagine one source of ``seed'' BHs bound to the cloud would be extremely massive stars that form in that cloud with very short lifetimes that might implode or collapse directly to massive BH remnants, rather than exploding as SNe. A new generation of simulations like STARFORGE might be able to address some of these, but the resolution required has thus far limited their explicit simulations to precisely the low-density, low-mass clouds of least interest here \citep{GrudicGuszejnov2021,GrudicKruijssen2021}.
It is also possible that physics we neglect plays an important role. For example, on galactic scales, cosmic rays can influence the ISM significantly, although many have argued that because of their diffusive nature (smooth CR density gradients), they play little dynamical role (except perhaps via ionization) in the dense ISM clouds of interest here \citep{farber:decoupled.crs.in.neutral.gas,hopkins:cr.mhd.fire2,hopkins:cr.multibin.mw.comparison,bustard:2020.crs.multiphase.ism.accel.confinement}.
More realistic initial conditions and boundary conditions for clouds (embedded in a full ISM, for example) could also be important \citep{lane:2022.turbcloud}. This is perhaps especially relevant for our most massive complexes. When we simulate regions with $R_{\rm cl} \sim 500\,$pc and $\bar{\Sigma}_{0} \sim 10^{4}\,{\rm M_{\odot}\, pc^{-2}}$ -- i.e.\ gas masses as large as $\sim 10^{10}\,M_{\odot}$, these are not really ``clouds'' as we think of GMCs in the Milky Way. Rather, these values are motivated by typical sizes and densities observed in systems like starburst and/or ultra-luminous infrared galaxy nuclei \citep[see e.g.][]{kennicutt98,gao:2004.hcn.compilation,narayanan:2008.sfr.densegas.corr,bigiel:2008.mol.kennicutt.on.sub.kpc.scales}, and seen in the common massive clump-complexes or nuclei of high-redshift massive galaxies \citep{tacconi:high.molecular.gf.highz,krumholz:2010.highz.clump.survival,narayanan:2011.z2.kslaw,orr:ks.law}. But under these conditions, there is usually also a large pre-existing stellar population and dark matter halo, defining the potential of the nuclear gas -- properly simulating these regimes would really require full galaxy-formation simulations. It is likely that this added potential would make the starburst even less able to disrupt, leaving behind a dense nuclear bulge \citep[e.g.][]{sanders88:quasars,tacconi:ulirgs.sb.profiles,rj:profiles,hopkins:cusps.mergers,hopkins:cusps.fp,hopkins:cusps.evol,hopkins:cusps.ell,hopkins:cusp.slopes}.
\section{Conclusions}
\label{sec:conclusions}
We have simulated populations of dynamic, accreting BH seeds with masses $\sim 10^{1}-10^{4}\,M_{\odot}$ in massive cloud complexes (meant to resemble the most massive GMCs, high-redshift and starburst galaxy nuclei), with self-gravity, realistic cooling, detailed models for star formation and stellar feedback in the form of radiation, outflows, and supernovae, but neglecting the feedback from the BHs themselves. Our goal is to identify whether, under any conditions, such seeds can capture gas from the dense, cold ISM at the rates required to power hyper-Eddington accretion, and whether this can be sustained for long enough periods that it is conceivable such BHs could grow to IMBH or even SMBH mass scales. This forms a necessary, but not sufficient, condition for hyper-Eddington growth models for the origin of IMBHs and SMBHs.
Based on our analysis above, we can draw the following conclusions (again, absent BH feedback):
\begin{enumerate}
\item Sustained hyper-Eddington gravitational capture from the ISM can occur, under specific conditions (detailed below). This occurs when BH seeds coincidentally find themselves in regions with locally enhanced densities (local densities well in excess of $\sim 100$ times the complex-mean), with (by chance) very low relative velocities (less than $\sim 10\%$ of the characteristic gravitational velocity of the complex). The dense clump is then captured extremely quickly (on less than its internal dynamical time), which can set of a ``runaway'' of competitive accretion by which the seed grows even more massive (reaching up to $\sim 1\%$ of the complex gas mass).
\item Provided the right conditions are met, this process is only very weakly dependent on the initial seed mass, even for stellar-mass seeds in the $\sim 10-100\,M_{\odot}$ range. Thus, the ``seed'' does not need to already be an IMBH.
\item Much like with star formation, stellar feedback plays a dual role. Stellar feedback overall suppresses star formation and unbinds gas, suppressing BH growth (especially in lower-density clouds). But in higher-density, more-massive complexes, feedback produces regions like colliding shocks/shells which promote exactly the conditions needed for runaway BH growth.
\end{enumerate}
For this runaway accretion to occur, we show that there are several necessary ``global'' criteria the molecular complex must meet, including:
\begin{enumerate}
\item The complex must have a high surface density/gravitational pressure, $\bar{\Sigma}_{0} \gtrsim 1000\,{\rm M_{\odot}\,pc^{-2}}$. Otherwise, stellar feedback disrupts the medium too efficiently, both reducing the time available for accretion but also unbinding dense gas instead of allowing it to remain trapped and thus potentially creating situations with low relative velocities.
\item The complex must also be sufficiently high-mass, $M_{\rm cl} \gtrsim 10^{6}\,M_{\odot}$. This is to ensure both that there is sufficient total mass supply that if hyper-Eddington accretion occurs, the final mass is ``interesting'' (reaching IMBH, let alone SMBH mass scales), but also required, along with the $\bar{\Sigma}_{0}$ criterion, to ensure that the escape velocity of the cloud will be large enough that ionizing radiation does not rapidly unbind material or disrupt the complex and prevent accretion.
\item The BH seeds must be ``trapped'' by the complex, with systematic relative velocities not significantly larger than the characteristic gravitational velocity of the cloud. This means, for example, that a BH moving isotropically in the background galaxy bulge, intersecting a cloud, would be unable to accrete, while BHs with small relative velocities to the cloud are viable.
\item We require at least $\sim 100$ seeds, in complexes meeting all the criteria above, to have an order-unity probability of one showing sustained hyper-Eddington accretion. Thus even when all the criteria are met, the conditions are ``rare'' on a per-seed basis. Once the number of seeds is sufficiently large, the finite number of locations where runaway can occur, plus the competitive accretion dynamics noted above, mean that the number which actually do experience runaway growth saturates at one to a few.
\end{enumerate}
In future work, we will use this preliminary study to inform a more focused study which does include BH feedback, systematically exploring the uncertainties in BH feedback models but focused on cloud conditions where -- at least in the absence of said feedback -- we find runaway growth is possible.
Given that this is really a theoretical ``proof of concept'' and we do not yet include these crucial physics (which we expect may change the key conclusions), we hesitate to make specific observational predictions. Nonetheless, even if BH feedback did nothing to further suppress runaway BH growth, there are some important conclusions we can draw regarding observations of both active (star-forming) clouds and ``relic'' star clusters.
\begin{enumerate}
\item Runaway accretion would not occur in Milky Way/Local Group GMCs or cloud complexes: the necessary conditions much more closely resemble starburst galaxy nuclei and the most massive dense star-forming clumps observed in high-redshift massive galaxies.
\item As a result, the ``relic'' star clusters from regions which could produce runaway accretion will not be typical star clusters or globular clusters. They are much more akin to nuclear star clusters (at the low-mass end) and dense galactic bulges (at the high-mass end). Even if the high-redshift clumps are off-center, these complexes would quickly spiral to the center of the galaxies to form proto-bulges \citep{noguchi:1999.clumpy.disk.bulge.formation,dekel:2009.clumpy.disk.evolution.toymodel}, which is important for SMBH seed formation mechanisms as it is almost impossible for seeds of the masses we predict here to ``sink'' to galaxy centers via dynamical friction in a Hubble time at high redshift, if they are not ``carried'' by more massive star cluster complexes \citep{ma:2021.seed.sink.inefficient.fire}.
\item Regardless of which clusters could have hosted this runaway process, we again find the probability is low on a ``per seed'' basis. Therefore, whether we expect an IMBH/SMBH ``relic'' in the descendants depends fundamentally on the population of seeds and their dynamics. While we find stellar-mass seeds are viable, it is not obvious if these could come from the stars forming in the cloud itself (e.g. from the relics of the stars formed during the process). Most stellar-mass seeds form relatively late after star formation ($\gtrsim 30\,$Myr), in explosions (which could disrupt the cloud), and have large natal kicks (excessive relative velocity). It is possible, if kicks were somehow avoided, that the most massive stars which reach the end of the main sequence more rapidly (at $\sim 3\,$Myr) and collapse directly to BHs could be viable seeds, but then these are much more rare. Alternatively, the seeds could come from the ``pre-existing'' background of stars, as especially in e.g.\ galactic nuclei or $\sim$\,kpc-scale clump complexes in massive galaxies we expect a very large population of background stellar-mass BHs. The key then is their kinematics (i.e.\ whether a sufficient number can be ``captured'' to locally interact as we model).
\item Almost by definition, the required conditions make it very difficult to observe this process ``in action.'' Complexes which meet the criteria above are, by definition, Compton-thick (and since the accretion occurs in over-dense sub-regions, these are even more heavily obscured). Moreover, if the maximum luminosity of accreting BHs (even if they are undergoing hyper-Eddington accretion) is not much larger than the traditional Eddington limit (as most models predict; see \S~\ref{sec:intro}), then the bolometric and even X-ray luminosities of the clouds/complexes will be dominated by the stars (not the runaway accreting BHs), unless the BH accretes an enormous fraction ($\sim 10\%$) of the entire cloud mass.
\item Even if such enormous accretion were to occur (or if the luminosity could exceed Eddington), by the time the BH luminosity could ``outshine'' even a fraction of the stellar luminosity of the complex, its luminosity would be so large that it would not be a ULX-type source. Rather (especially, again, recalling that the complexes of interest are generally in or around distant galactic nuclei), it would much more closely resemble an off-center, obscured AGN (or a dual AGN, if the galaxy already has an accreting SMBH). Large populations of such AGN sources are, of course, well-known, and there are much more mundane ways to produce them (via galaxy mergers or irregular kinematics), but it is perhaps conceivable that a small fraction of them could be systems like what we simulate here.
\end{enumerate}
\acknowledgments{We thank Xiangcheng Ma and Linhao Ma for useful discussions and revisions of this draft. Support for the authors was provided by NSF Research Grants 1911233, 20009234, 2108318, NSF CAREER grant 1455342, NASA grants 80NSSC18K0562, HST-AR-15800. Numerical calculations were run on the Caltech compute cluster ``Wheeler,'' allocations AST21010 and AST20016 supported by the NSF and TACC, and NASA HEC SMD-16-7592.}
\datastatement{The data supporting the plots within this article are available on reasonable request to the corresponding author. A public version of the GIZMO code is available at \href{http://www.tapir.caltech.edu/~phopkins/Site/GIZMO.html}{\url{http://www.tapir.caltech.edu/~phopkins/Site/GIZMO.html}}.}
|
\section{Introduction}\label{sec:intro}
This paper is in a sense a continuation of our previous work \cite{KT2018, KT2020-ASPM, KT2020-Tsukuba}.
There we studied a certain type of zeta functions and their values at positive as well as negative integer arguments,
the prototype being the so-called Arakawa-Kaneko zeta function investigated in \cite{AK1999}, whose values at positive integers
can be written in terms of multiple zeta values and at negative integers in terms of poly-Bernoulli numbers.
Through our study in this context in the case of level 4, we are naturally
led to the investigation of a class of multiple $L$-values of level (or conductor) 4, which in our notation is given by
\[ L_\sh(k_1,\ldots,k_r;\chi_4,\ldots,\chi_4)=\sum_{1\leq m_1<\cdots<m_r \atop m_j \equiv j \bmod 2} \frac{(-1)^{(m_r-r)/2}}{m_1^{k_1}\cdots m_r^{k_r}}. \]
We discovered that these $L$-values with special indices satisfy a notable linear relation with combinatorial numbers (`Entringer
numbers') as coefficients. Also we found a (conjectural) connection of double $L$-values to modular forms of level four, a relation
of previously studied level 2 $L$-values (`multiple $T$-values') to our level 4 values,
and a formula for the generating function of `height one' $L$-values in terms of the Appell hypergeometric
function $F_1$, being in contrast to Gauss's ${}_2F_1$ in the previous cases.
In the next section, we present these findings on the level 4 multiple $L$-values in detail. Various iterated integral expressions are our basic tools.
In Section~\ref{sec:zeta-polyEuler}, we introduce and study the motivating Arakawa-Kaneko type zeta function of level 4
and the corresponding multiple poly-Euler numbers (different from those introduced and studied in Sasaki \cite{Sasaki2012}
and Ohno-Sasaki \cite{OS2012,OS2013,OS2017}). In the final Section~\ref{sec:zetafunct} we connect the one variable multiple
${\widetilde{T}}$-function to the zeta function introduced in \S\ref{sec:zeta-polyEuler} and obtain several family of relations among multiple
${\widetilde{T}}$-values. The formulas and the techniques used in these two sections are more or less parallel to our previous studies.
\section{Some results and conjectures on multiple $L$-values of level four}\label{sec:ttilde}
\subsection{Definition}\label{subsec:def}
In \cite{AK2004}, Arakawa and the first named author defined two types of multiple $L$-values as follows. Let $f_j\,:\,\mathbb{Z} \to \mathbb{C}$ $(j=1,\ldots,r)$ be some periodic functions
(in the sequel we only consider the Dirichlet character of conductor 4 and its square).
For integers $k_1,\ldots,k_r\in \mathbb{Z}_{\geq 1}$, define
\begin{align}
& L_\sh(k_1,\ldots,k_r;f_1,\ldots,f_r)=\sum_{1\leq m_1<\cdots<m_r} \frac{f_1(m_1)f_2(m_2-m_1)\cdots f_r(m_r-m_{r-1})}{m_1^{k_1}m_2^{k_2}\cdots m_r^{k_r}}, \label{MLV-sh}\\
& L_\ast(k_1,\ldots,k_r;f_1,\ldots,f_r)=\sum_{1\leq m_1<\cdots<m_r} \frac{f_1(m_1)f_2(m_2)\cdots f_r(m_r)}{m_1^{k_1}m_2^{k_2}\cdots m_r^{k_r}}. \label{MLV-star}
\end{align}
If $k_r\geq 2$, these series are absolutely convergent, and if $k_r=1$, these are interpreted as
\begin{align*}
& L_\sh(k_1,\ldots,k_r;f_1,\ldots,f_r)=\lim_{M\to \infty}\sum_{1\leq m_1<\cdots<m_r<M} \frac{f_1(m_1)f_2(m_2-m_1)\cdots f_r(m_r-m_{r-1})}{m_1^{k_1}m_2^{k_1}\cdots m_r^{k_r}}, \\
& L_\ast(k_1,\ldots,k_r;f_1,\ldots,f_r)=\lim_{M\to \infty}\sum_{1\leq m_1<\cdots<m_r<M} \frac{f_1(m_1)f_2(m_2)\cdots f_r(m_r)}{m_1^{k_1}m_2^{k_1}\cdots m_r^{k_r}},
\end{align*}
when the limits exist. As the notation suggests, these values satisfy (under a suitable condition, see~\cite{AK2004}) shuffle and stuffle (or harmonic) product rules respectively.
Let $\chi_4$ be the (unique) primitive Dirichlet character of conductor $4$, and consider the case of $f_j=\chi_4$ for all $j$.
Then it is readily seen by definition that
\begin{equation}
L_\sh(k_1,\ldots,k_r;\chi_4,\ldots,\chi_4)=\sum_{1\leq m_1<\cdots<m_r \atop m_j \equiv j \bmod 2} \frac{(-1)^{(m_r-r)/2}}{m_1^{k_1}\cdots m_r^{k_r}},\label{MLV-chi-4}
\end{equation}
which is convergent even when $k_r=1$.
To ease notation, and introducing the factor $2^r$ for later convenience, we define the `multiple ${\widetilde{T}}$-values' for any tuple of
positive integers $(k_1,\ldots,k_r)$
\begin{equation}\label{ttilde-def} {\widetilde{T}}(k_1,\ldots,k_r):=2^r L_\sh(k_1,\ldots,k_r;\chi_4,\ldots,\chi_4)
=2^r\sum_{1\leq m_1<\cdots<m_r \atop m_j \equiv j \bmod 2} \frac{(-1)^{(m_r-r)/2}}{m_1^{k_1}\cdots m_r^{k_r}}. \end{equation}
This is in contrast to our previously studied object
\begin{equation}\label{t-def} T(k_1,\ldots,k_r)=2^rL_\sh(k_1,\ldots,k_r;\chi_4^2,\ldots,\chi_4^2)
=2^r\sum_{1\leq m_1<\cdots<m_r \atop m_j \equiv j \bmod 2} \frac{1}{m_1^{k_1}\cdots m_r^{k_r}},\end{equation}
which we called `multiple $T$-values.' For this, the last entry $k_r$ should be larger than 1. More generally, we introduced in~\cite{KT2020-ASPM, KT2020-Tsukuba}
a level 2 analogue of the multiple polylogarithms for $k_1,\ldots,k_r\in \mathbb{Z}_{\ge1}$:
\begin{equation}
A(k_1,\ldots,k_r;z)=2^r\sum_{0<m_1<\cdots <m_r\atop m_i\equiv i\bmod 2}
\frac{z^{m_r}}{m_1^{k_1}\cdots m_{r}^{k_{r}}}\quad (|z|<1). \label{Def-A-polylog}
\end{equation}
In particular,
\begin{equation}\label{A1exp}
A(1;z)=2\sum_{n=0}^\infty \frac{z^{2n+1}}{2n+1}=2\tanh^{-1}(z)=\log\left(\frac{1+z}{1-z}\right)=\int_0^z
\frac{2dt}{1-t^2}.
\end{equation}
As is easily seen (parallel to the case of usual multiple polylogarithms), the function $A(k_1,\ldots,k_r;z)$
satisfies the derivative formula (see \cite[Lemma 5.1]{KT2020-ASPM})
\begin{align}\label{deriv-a-multi}
\frac{d}{dz}A(k_1,\ldots,k_r;z)
& =
\begin{cases}
\displaystyle{\frac{1}{z}}\,A(k_1,\ldots,k_{r-1},k_r-1;z) & (k_r\geq 2),\\
\displaystyle{\frac{2}{1-z^2}}\,A(k_1,\ldots,k_{r-1};z) & (k_r=1),
\end{cases}
\end{align}
and hence $A(k_1,\ldots,k_r;z)$ is realized as an iterated integral starting with \[ A(1;z)=\int_0^z \frac{2dt}{1-t^2}\] as follows.
We introduce a compact notation of writing an iterated integral.
Let $a$ and $b$ be points in the closed unit disk. For differential 1-forms
$ \Omega_1, \Omega_2,\ldots,\Omega_k$, we understand by the expression
\[ \int_a^b \Omega_1\circ \Omega_2\circ \cdots\cdots \circ\Omega_k\]
the iterated integral
\[ \int_a^b\left(\int_a^{t_k}\cdots\left(\int_a^{t_3}\left(\int_a^{t_2}\Omega_1\right)\Omega_2\right)\cdots\cdots\Omega_k\right),\]
where $\Omega_i=\Omega_i(t_{i})dt_i$.
Here and in the following, every path of integration is considered inside the unit disk. Under this notation, from \eqref{A1exp}
and \eqref{deriv-a-multi}, we immediately have
\begin{equation*}
A(k_1,\ldots,k_r;z)= \int_0^z \Omega_1\circ \Omega_2\circ \cdots\cdots \circ\Omega_k,
\end{equation*}
where $\Omega_j=2dt_j/(1-t_j^2)$ if $j\in\{1,k_1+1,\ldots,k_1+\cdots+k_{r-1}+1\}$ and $\Omega_j=dt_j/t_j$
otherwise. Note that the total number $k$ of differential forms is the {\it weight} $k_1+\cdots+k_r$ of the index
$(k_1,\ldots,k_r)$, and the number of $2dt/(1-t^2)$ is the {\it depth} $r$, the total number of components of the index.
Since
\begin{equation}
T(k_1,\ldots,k_r)=A(k_1,\ldots,k_r;1)\ \ (k_r\geq 2),\label{A-T-rel}
\end{equation}
we have an integral expression of multiple $T$-values (see also \cite[Theorem~2.1]{KT2020-Tsukuba})
\begin{align}
&T(k_1,k_{2},\ldots,k_{r}) \nonumber \\
&=\int_0^1\frac{2dt}{1-t^2}\circ\underbrace{\frac{dt}{t}\circ\cdots\circ\frac{dt}{t}}_{k_1-1}\circ
\frac{2dt}{1-t^2}\circ\underbrace{\frac{dt}{t}\circ\cdots\circ\frac{dt}{t}}_{k_2-1}\circ\cdots\cdots\circ
\frac{2dt}{1-t^2}\circ\underbrace{\frac{dt}{t}\circ
\cdots\circ\frac{dt}{t}}_{k_r-1} \nonumber \\
&=\mathop{\int\cdots\int}\limits_{0<t_1<\cdots <t_k<1}\frac{2dt_1}{1-t_1^2}\underbrace{\frac{dt_2}{t_2}\cdots\frac{dt}{t}}_{k_1-1}
\frac{2dt}{1-t^2}\underbrace{\frac{dt}{t}\cdots\frac{dt}{t}}_{k_2-1}\cdots\cdots\frac{2dt}{1-t^2}\underbrace{\frac{dt}{t}
\cdots\frac{dt_{k}}{t_{k}}}_{k_r-1}.\label{T-integ-exp}
\end{align}
For simplicity, we have suppressed the subscripts of variables of most of the differential forms.
In the same vein, the multiple ${\widetilde{T}}$-values can also be given as integrals. This fact is fundamental to almost all of our proofs of Theorems.
First, as in the case of usual multiple zeta values and $T$-values, by expanding $1/(1+t^2)$ into geometric series and
integrating term by term, we obtain:
\begin{prop}
For $k_1,\ldots,k_r\in \mathbb{Z}_{\geq 1}$, we have
\begin{align}
&{\widetilde{T}}(k_1,k_{2},\ldots,k_{r}) \nonumber \\
&=\int_0^1\frac{2dt}{1+t^2}\circ\underbrace{\frac{dt}{t}\circ\cdots\circ\frac{dt}{t}}_{k_1-1}\circ
\frac{2dt}{1+t^2}\circ\underbrace{\frac{dt}{t}\circ\cdots\circ\frac{dt}{t}}_{k_2-1}\circ\cdots\cdots\circ
\frac{2dt}{1+t^2}\circ\underbrace{\frac{dt}{t}\circ
\cdots\circ\frac{dt}{t}}_{k_r-1} \nonumber \\
&=\mathop{\int\cdots\int}\limits_{0<t_1<\cdots <t_k<1}\frac{2dt_1}{1+t_1^2}\underbrace{\frac{dt_2}{t_2}\cdots\frac{dt}{t}}_{k_1-1}
\frac{2dt}{1+t^2}\underbrace{\frac{dt}{t}\cdots\frac{dt}{t}}_{k_2-1}\cdots\cdots\frac{2dt}{1+t^2}\underbrace{\frac{dt}{t}
\cdots\frac{dt_{k}}{t_{k}}}_{k_r-1}.\label{integ-exp}
\end{align}
\end{prop}
A typical consequence, which we use later several times, is the following identity.
\begin{cor} For any $n\ge1$, we have
\begin{equation}\label{11111}
\frac{{\widetilde{T}}(1)^n}{n!}={\widetilde{T}}(\underbrace{1,\ldots,1}_n).
\end{equation}
\end{cor}
If we make a change of variables $t\to (1-u)/(1+u)$ appeared in \cite{KT2020-Tsukuba} for $T$-values,
we obtain from \eqref{integ-exp} another integral expression when $k_r>1$ (in the case of `admissible' index). For the definition of the `dual' of an index,
see for instance \cite[\S3.1]{KT2020-Tsukuba}.
\begin{cor}
Suppose $k_r>1$ and let $(l_1,\ldots,l_s)$ be the dual index of $(k_1,k_{2},\ldots,k_{r})$. Then we have
\begin{align}
&{\widetilde{T}}(k_1,k_{2},\ldots,k_{r}) \nonumber\\
&=\int_0^1\frac{2du}{1-u^2}\circ\underbrace{\frac{2du}{1+u^2}\circ\cdots\circ\frac{2du}{1+u^2}}_{l_1-1}\circ
\frac{2du}{1-u^2}\circ\underbrace{\frac{2du}{1+u^2}\circ\cdots\circ\frac{2du}{1+u^2}}_{l_2-1}\circ\cdots \nonumber\\
& \qquad\qquad\qquad\cdots\circ
\frac{2du}{1-u^2}\circ\underbrace{\frac{2du}{1+u^2}\circ
\cdots\circ\frac{2du}{1+u^2}}_{l_s-1} \nonumber \\
&=\mathop{\int\cdots\int}\limits_{0<u_1<\cdots <u_k<1}\frac{2du_1}{1-u_1^2}\underbrace{\frac{2du_2}{1+u_2^2}\cdots\frac{2du}{1+u^2}}_{l_1-1}
\frac{2du}{1-u^2}\underbrace{\frac{2du}{1+u^2}\cdots\frac{2du}{1+u^2}}_{l_2-1}\cdots \label{integ-exp2} \\
&\qquad\qquad\qquad\cdots\frac{2du}{1-u^2}\underbrace{\frac{2du}{1+u^2}\cdots\frac{2du_k}{1+u_k^2}}_{l_s-1}. \nonumber
\end{align}
\end{cor}
Furthermore, by the series expressions~\eqref{ttilde-def} and~\eqref{Def-A-polylog}, we see that
\begin{equation}\label{TtbyA}
{\widetilde{T}}(k_1,\ldots,k_r)=i^{-r}A(k_1,\ldots,k_r;i)\quad (i=\sqrt{-1}),
\end{equation}
and hence we have yet another iterated integral representation of ${\widetilde{T}}(k_1,\ldots,k_r)$:
\begin{align}\label{integ-exp3}
&{\widetilde{T}}(k_1,\ldots,k_r) \nonumber\\
&=i^{-r} \int_0^i\frac{2dt}{1-t^2}\circ\underbrace{\frac{dt}{t}\circ\cdots\circ\frac{dt}{t}}_{k_1-1}\circ
\frac{2dt}{1-t^2}\circ\underbrace{\frac{dt}{t}\circ\cdots\circ\frac{dt}{t}}_{k_2-1}\circ\cdots\cdots\circ
\frac{2dt}{1-t^2}\circ\underbrace{\frac{dt}{t}\circ
\cdots\circ\frac{dt}{t}}_{k_r-1}.
\end{align}
\subsection{The space of multiple ${\widetilde{T}}$-values}\label{subsec:space}
As usual, let us consider the $\mathbb{Q}$-vector space
\[ \widetilde{\mathcal{T}}=\sum_{k=0}^\infty\,\widetilde{\mathcal{T}}_k\] spanned by all multiple ${\widetilde{T}}$-values, where
\[ \widetilde{\mathcal{T}}_0=\mathbb{Q},\quad
\widetilde{\mathcal{T}}_{k}=\sum_{1\leq r \leq k \atop {{k_1,\ldots,k_{r}\geq 1 \atop k_1+\cdots+k_r=k}}}
\mathbb{Q}\cdot {\widetilde{T}}(k_1,\ldots,k_r)\quad (k\geq 1).
\]
\begin{prop} The space $\widetilde{\mathcal{T}}$ is a $\mathbb{Q}$-algebra under the usual multiplication of real numbers.
\end{prop}
\begin{proof}
This is a standard consequence of the integral expression~\eqref{integ-exp}, the product being
described by the {\it shuffle product} rule.
\end{proof}
\begin{example}
By~\eqref{integ-exp}, the shuffle product of ${\widetilde{T}}$-values takes exactly the same form
as in the case of multiple zeta (and $T$-) values. In our case, the ${\widetilde{T}}$-values for non-admissible indices
also converge and we do not need any regularization procedure. For instance, we have
\begin{align*}
{\widetilde{T}}(1){\widetilde{T}}(2)&=2{\widetilde{T}}(1,2)+{\widetilde{T}}(2,1),\\
{\widetilde{T}}(2)^2&=4{\widetilde{T}}(1,3)+2{\widetilde{T}}(2,2).
\end{align*}
\end{example}
The first natural question would be the dimension $d_k$
over $\mathbb{Q}$ of each subspace $\widetilde{\mathcal{T}}_k$ of weight $k$ elements.
We have conducted numerical experiments with Pari-GP,
and obtained the following conjectural table.
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|} \hline
$k$ & 0&1&2& 3& 4& 5& 6& 7& 8& 9& 10& 11\\
\hline
$d_k$ & 1& 1& 2& 3& 6& 8& 16& 22& 44& 59& 118& 162\\
\hline
$d_{k,ev}$ & 1& 0& 1& 2& 3& 4& 8& 12& 22& 30& 59& 84\\
\hline
$d_{k,od}$ & 0& 1& 1& 1& 3& 4& 8& 10& 22& 29& 59& 78\\
\hline
\end{tabular}
\end{center}
\vspace{7pt}
It seems that the space of multiple ${\widetilde{T}}$-values of {\it even} depth and that of {\it odd} depth
are disjoint. In the table above we also display the conjectural dimensions $d_{k,ev}$ and $d_{k,od}$ of the spaces of even and odd
depth multiple ${\widetilde{T}}$-values of weight $k$ respectively. As far as the table made by numerical experiments up to weight $11$ goes,
the predicted relation $d_k=d_{k,ev}+d_{k,od}$ holds.
We can also read off the relation $d_k=2d_{k-1}$ and $d_{k,ev}=d_{k,od}$ if $k>1$ is even in this range,
but we are not aware of any reason to believe that this holds true in general.
Through the numerical experiments, we found it very likely that the multiple $T$-values lie in the space $\widetilde{\mathcal{T}}$
of multiple ${\widetilde{T}}$-values and moreover in the subspace spanned by {\it even} depth multiple ${\widetilde{T}}$-values.
This prediction, first announced as a conjecture in a conference, was confirmed soon after (independently) by
R.~Umezawa and M.~Hirose. They pointed out that, essentially, this follows from the path-composition formula for the iterated integrals.
We give here seemingly the most direct proof using that formula.
\begin{theorem}[Umezawa~\cite{U}]\label{Teven}
Any multiple $T$-value can be written as a linear combination of multiple ${\widetilde{T}}$-values
of even depth.
\end{theorem}
\begin{proof}
We use the formula \eqref{T-integ-exp}:
\begin{equation*}
T(k_1,\ldots,k_r)= \int_0^1 \Omega_1\circ \Omega_2\circ \cdots\cdots \circ\Omega_k,
\end{equation*}
where $\Omega_j=2dt_j/(1-t_j^2)$ if $j\in\{1,k_1+1,\ldots,k_1+\cdots+k_{r-1}+1\}$ and $\Omega_j=dt_j/t_j$
otherwise. Now, by decomposing the path of integration into two paths, first from 0 to $i=\sqrt{-1}$ and from $i$ to 1,
and using the path composition formula for iterated integrals (see for instance \cite[Prop.~1.5.1]{Chen}), we may write this as
\[ T(k_1,\ldots,k_r)=\sum_{j=0}^k \left(\int_0^i \Omega_1\circ \cdots \circ\Omega_j\right)\left(\int_i^1 \Omega_{j+1}\circ
\cdots\circ\Omega_k\right). \]
Suppose that $\int_0^i \Omega_1\circ \cdots \circ\Omega_j$ has depth $l$, i.e., there are $l$ $2dt/(1-t^2)$'s
among $ \Omega_1, \ldots, \Omega_j$. Then by \eqref{integ-exp3}, $\int_0^i \Omega_1\circ \cdots \circ\Omega_j$
is equal to $i^l$ times a multiple ${\widetilde{T}}$-value of depth $l$.
On the other hand, by changing the variables $t\to (-iu+1)/(u-i)$, we see that the integral
$\int_i^1 \Omega_{j+1}\circ\cdots\circ\Omega_k$ is equal to $i^{-(k-j-r+l)}$ times a real iterated integral
$\int_0^1 \Omega_{j+1}'\circ\cdots\circ\Omega_k'$, where $\Omega_h'=2du/(1-u^2)$ if
$\Omega_h=2dt/(1-t^2)$ and $\Omega_h'=2du/(1+u^2)$ if $\Omega_h=dt/t$.
Since the number of $2du/(1-u^2)$ (depth) among $\Omega_h'$ is $r-l$ and the total number (weight) is $k-j$,
by \eqref{integ-exp2}, we conclude that the integral $\int_i^1 \Omega_{j+1}\circ\cdots\circ\Omega_k$ is equal to $i^{-(k-j-r+l)}$ times
a multiple ${\widetilde{T}}$-value of weight $k-j$ and depth $k-j-r+l$. Therefore, we conclude that
the product $\left(\int_0^i \Omega_1\circ \cdots \circ\Omega_j\right)\left(\int_i^1 \Omega_{j+1}\circ
\cdots\circ\Omega_k\right)$ is $i^{-k+j+r}$ times a product of multiple ${\widetilde{T}}$-values of
depths $l$ and $k-j-r+l$, which is by the shuffle product $i^{-k+j+r}$ times a sum of multiple ${\widetilde{T}}$-values of
depth $k-j-r+2l$. Since $T(k_1,\ldots,k_r)$ is real and the power $-k+j+r$ of $i$ and the depth $k-j-r+2l$
have the same parity, we conclude that the $T$-value $T(k_1,\ldots,k_r)$ is a sum of ${\widetilde{T}}$-values of
even depth.
\end{proof}
\begin{example} Consider the `height one' multiple $T$-values, i.e., $T$-values of the form $T(\underbrace{1,\ldots,1}_{r-1},k+1)$ with $r,k\ge1$.
Starting with the iterated integral expression \eqref{T-integ-exp} and proceeding as in the above proof, we have
\begin{align*}
&T(\underbrace{1,\ldots,1}_{r-1},k+1)
=\int_0^1\underbrace{\frac{2dt}{1-t^2}\circ\cdots\circ\frac{2dt}{1-t^2}}_{r}\circ
\underbrace{\frac{dt}{t}\circ\cdots\circ\frac{dt}{t}}_{k}\\
&=\left(\int_0^i\underbrace{\frac{2dt}{1-t^2}\circ\cdots\circ\frac{2dt}{1-t^2}}_{r}\right)
\left(\int_i^1 \underbrace{\frac{dt}{t}\circ\cdots\circ\frac{dt}{t}}_{k}\right)\\
&+\sum_{j=1}^r \left(\int_0^i\underbrace{\frac{2dt}{1-t^2}\circ\cdots\circ\frac{2dt}{1-t^2}}_{r-j}\right)
\left(\int_i^1\underbrace{\frac{2dt}{1-t^2}\circ\cdots\circ\frac{2dt}{1-t^2}}_{j}\circ
\underbrace{\frac{dt}{t}\circ\cdots\circ\frac{dt}{t}}_{k}\right)\\
&+\sum_{j=1}^k \left(\int_0^i\underbrace{\frac{2dt}{1-t^2}\circ\cdots\circ\frac{2dt}{1-t^2}}_{r}\circ
\underbrace{\frac{dt}{t}\circ\cdots\circ\frac{dt}{t}}_{j}\right)
\left(\int_i^1 \underbrace{\frac{dt}{t}\circ\cdots\circ\frac{dt}{t}}_{k-j}\right).
\end{align*}
Now, by~\eqref{integ-exp3}, we have
\begin{align*}
\int_0^i\underbrace{\frac{2dt}{1-t^2}\circ\cdots\circ\frac{2dt}{1-t^2}}_{r-j}
&=i^{r-j} {\widetilde{T}}(\underbrace{1,\ldots,1}_{r-j}),\\
\int_0^i\underbrace{\frac{2dt}{1-t^2}\circ\cdots\circ\frac{2dt}{1-t^2}}_{r}\circ
\underbrace{\frac{dt}{t}\circ\cdots\circ\frac{dt}{t}}_{j}
&=i^r {\widetilde{T}}(\underbrace{1,\ldots,1}_{r-1},j+1),
\end{align*}
and by the change of variable $t\to (-iu+1)/(u-i)$
\begin{align*}
\int_i^1\underbrace{\frac{2dt}{1-t^2}\circ\cdots\circ\frac{2dt}{1-t^2}}_{j}\circ
\underbrace{\frac{dt}{t}\circ\cdots\circ\frac{dt}{t}}_{k}
&=i^{-k}\int_0^1\underbrace{\frac{2du}{1-u^2}\circ\cdots\circ\frac{2du}{1-u^2}}_{j}\circ
\underbrace{\frac{2du}{1+u^2}\circ\cdots\circ\frac{2du}{1+u^2}}_{k}\\
&=i^{-k} {\widetilde{T}}(\underbrace{1,\ldots,1}_{k-1},j+1),\\
\int_i^1\underbrace{\frac{dt}{t}\circ\cdots\circ\frac{dt}{t}}_{k-j}
&=i^{-k+j} \int_0^1\underbrace{\frac{2du}{1+u^2}\circ\cdots\circ\frac{2du}{1+u^2}}_{k-j}\\
&=i^{-k+j} {\widetilde{T}}(\underbrace{1,\ldots,1}_{k-j}).
\end{align*}
Putting all these together and using the shuffle product formula
\[ {\widetilde{T}}(\underbrace{1,\ldots,1}_{r})
{\widetilde{T}}(\underbrace{1,\ldots,1}_{k})=\binom{r+k}{r}{\widetilde{T}}(\underbrace{1,\ldots,1}_{r+k}), \] we have
\begin{align}
&T(\underbrace{1,\ldots,1}_{r-1},k+1) = i^{r-k} \binom{r+k}{r}{\widetilde{T}}(\underbrace{1,\ldots,1}_{r+k}) \nonumber \\
&\qquad+ i^{r-k} \sum_{j=1}^r i^{-j} {\widetilde{T}}(\underbrace{1,\ldots,1}_{r-j}) {\widetilde{T}}(\underbrace{1,\ldots,1}_{k-1},j+1)
+ i^{r-k} \sum_{j=1}^k i^{j} {\widetilde{T}}(\underbrace{1,\ldots,1}_{k-j}) {\widetilde{T}}(\underbrace{1,\ldots,1}_{r-1},j+1).\label{TbyTtilde}
\end{align}
Taking the real and imaginary parts of this equation, we have the following set of relations.
When $r+k$ is even,
\begin{align}
&(-1)^{(k-r)/2}\,T(\underbrace{1,\ldots,1}_{r-1},k+1) =\binom{r+k}{r}{\widetilde{T}}(\underbrace{1,\ldots,1}_{r+k}) \nonumber\\
&\quad + \sum_{j=1\atop j:even}^r (-1)^{j/2}\, {\widetilde{T}}(\underbrace{1,\ldots,1}_{r-j}) {\widetilde{T}}(\underbrace{1,\ldots,1}_{k-1},j+1)
+ \sum_{j=1\atop j:even}^k (-1)^{j/2}\, {\widetilde{T}}(\underbrace{1,\ldots,1}_{k-j}) {\widetilde{T}}(\underbrace{1,\ldots,1}_{r-1},j+1),\\
&\qquad\sum_{j=1\atop j:odd}^r (-1)^{(j+1)/2} \,{\widetilde{T}}(\underbrace{1,\ldots,1}_{r-j}) {\widetilde{T}}(\underbrace{1,\ldots,1}_{k-1},j+1)\nonumber\\
&\qquad\qquad\qquad + \sum_{j=1\atop j:odd}^k (-1)^{(j-1)/2}\, {\widetilde{T}}(\underbrace{1,\ldots,1}_{k-j}) {\widetilde{T}}(\underbrace{1,\ldots,1}_{r-1},j+1)=0,
\end{align}
and when $r+k$ is odd,
\begin{align}
(-1)^{(k-r-1)/2}\,T(\underbrace{1,\ldots,1}_{r-1},k+1) &= \sum_{j=1\atop j:odd}^r (-1)^{(j+1)/2}\, {\widetilde{T}}(\underbrace{1,\ldots,1}_{r-j}) {\widetilde{T}}(\underbrace{1,\ldots,1}_{k-1},j+1) \nonumber \\
&+ \sum_{j=1\atop j:odd}^k (-1)^{(j-1)/2}\, {\widetilde{T}}(\underbrace{1,\ldots,1}_{k-j}) {\widetilde{T}}(\underbrace{1,\ldots,1}_{r-1},j+1),
\end{align}
\begin{align}
\binom{r+k}{r}{\widetilde{T}}(\underbrace{1,\ldots,1}_{r+k})&+\sum_{j=1\atop j:even}^r (-1)^{j/2}\, {\widetilde{T}}(\underbrace{1,\ldots,1}_{r-j}) {\widetilde{T}}(\underbrace{1,\ldots,1}_{k-1},j+1)\nonumber \\
&\qquad + \sum_{j=1\atop j:even}^k (-1)^{j/2}\, {\widetilde{T}}(\underbrace{1,\ldots,1}_{k-j}) {\widetilde{T}}(\underbrace{1,\ldots,1}_{r-1},j+1)=0.
\end{align}
For small values of $r$ and $k$ with $r+k\le 4$, we obtain from these, together with the shuffle product, the followings:
\begin{align*}
T(2)&=2{\widetilde{T}}(1,1),\\
T(3)&=T(1,2)={\widetilde{T}}(1,2)+{\widetilde{T}}(2,1),\\
{\widetilde{T}}(3)&=3{\widetilde{T}}(1,1,1),\\
T(4)&=T(1,1,2)=2{\widetilde{T}}(1,3)+{\widetilde{T}}(2,2)+{\widetilde{T}}(3,1)-4{\widetilde{T}}(1,1,1,1),\\
{\widetilde{T}}(4)&=2{\widetilde{T}}(1,1,2)+2{\widetilde{T}}(1,2,1)+{\widetilde{T}}(2,1,1),\\
T(1,3)&=-2{\widetilde{T}}(1,3)+6{\widetilde{T}}(1,1,1,1).
\end{align*}
The particular case of $r=1$ (multiplied by $i^k$) of \eqref{TbyTtilde} will be used later in the proof of Theorem~\ref{Ent-rel}:
\begin{align}\label{key}
i^kT(k+1)&=i(k+1){\widetilde{T}}(\underbrace{1,\ldots,1}_{k+1})+{\widetilde{T}}(\underbrace{1,\ldots,1}_{k-1},2)
+\sum_{j=1}^k i^{j+1} {\widetilde{T}}(\underbrace{1,\ldots,1}_{k-j}) {\widetilde{T}}(j+1).
\end{align}
\end{example}
\subsection{Height one multiple ${\widetilde{T}}$-values}\label{subsec:ht1}
As another application of the iterated integral expression \eqref{integ-exp},
we can compute the generating function of `height one' multiple ${\widetilde{T}}$-values.
Analogous results are known in the case of multiple zeta values as well as multiple $T$-values (\cite{Ao, Dr, KT2020-Tsukuba}):
\[ 1-\sum_{m,n=1}^\infty \zeta(\underbrace{1,\ldots,1}_{n-1},m+1)X^m Y^n
=\frac{\Gamma(1-X)\Gamma(1-Y)}{\Gamma(1-X-Y)}={}_2F_1(X,Y;1;1)^{-1}\]
and
\[ 1-\!\!\sum_{m,n=1}^\infty T({1,\ldots,1}_{n-1},m+1)X^m Y^n=
\frac{2\,\Gamma(1-X)\Gamma(1-Y)}{\Gamma(1-X-Y)}{}_2F_1(1-X,1-Y;1-X-Y;-1), \]
where ${}_2F_1(a,b;c;z)$ is the Gauss hypergeometric function.
In contrast to these cases, now the Appell hypergeometric function $F_1$ emerges.
\begin{theorem}\label{ht1gen}
We have
\[ \sum_{m, n\ge1}{\widetilde{T}}(\underbrace{1,\cdots,1}_{n-1},m)X^{m-1}Y^{n-1}
=\frac2{1-X}F_1(1-X;1-iY,1+iY;2-X;i,-i), \]
where $F_1$ stands for the Appell hypergeometric function
\[ F_{1}(a;b_{1},b_{2};c;x,y)=\sum _{m,n=0}^{\infty }{\frac {(a)_{m+n}(b_{1})_{m}(b_{2})_{n}}{(c)_{m+n}\,m!\,n!}}\,x^{m}y^{n}. \]
\end{theorem}
\begin{proof}
From the integral expression~\eqref{integ-exp}, we have
\begin{align*}
{\widetilde{T}}(\underbrace{1,\ldots,1}_{n-1},m)&=\underset{0<t_1<\cdots<t_n<u_1<\cdots<u_{m-1} <1}{\int\cdots\int}
\frac{2dt_1}{1+t_1^2}\cdots\frac{2dt_n}{1+t_n^2}\frac{du_1}{u_1}\cdots\frac{du_{m-1}}{u_{m-1}} \\
&=\int_0^1\left(\frac1{(n-1)!}\left(\int_0^{t_n}\frac{2dt}{1+t^2}\right)^{n-1}\cdot\frac1{(m-1)!}\left(\int_{t_n}^1\frac{du}{u}\right)^{m-1}\right)
\frac{2dt_n}{1+t_n^2}.
\end{align*}
Here, by using
\[ \int_0^{t_n}\frac{2dt}{1+t^2}=\frac1{i}\log\left(\frac{1+it_n}{1-it_n}\right)\quad(\text{principal value, }0\le t_n\le1), \]
we obtain
\[ {\widetilde{T}}(\underbrace{1,\ldots,1}_{n-1},m)=\frac{i^{1-n}}{(n-1)!(m-1)!}\int_0^1\left(\log\left(\frac{1+it}{1-it}\right)\right)^{n-1}
\left(\log\frac1t\right)^{m-1}\frac{2dt}{1+t^2} \]
and hence
\begin{align*}
&\sum_{m,n\ge1}{\widetilde{T}}(\underbrace{1,\ldots,1}_{n-1},m)X^{m-1}Y^{n-1}\\
&=\int_0^1\left(\sum_{n\ge1}\left(\log\left(\frac{1+it}{1-it}\right)\right)^{n-1}\frac1{(n-1)!}\left(\frac{Y}{i}\right)^{n-1}\right)
\left(\sum_{m\ge1}\left(\log\frac1t\right)^{m-1}\frac{X^{m-1}}{(m-1)!}\right)\frac{2dt}{1+t^2}\\
&=\int_0^1\left(\frac{1+it}{1-it}\right)^\frac{Y}{i}t^{-X}\frac{2dt}{1+t^2}\\
&=2\int_0^1t^{-X}(1-it)^{iY-1}(1+it)^{-iY-1}dt.
\end{align*}
Recall the integral expression of the Appell hypergeometric series $F_1$ (see {\it e.g.} \cite[Chap.14 (p. 300)]{WW}):
\begin{align*}
& F_1(a;b_1,b_2;c;x,y)
=\sum_{m,n=0}^\infty \frac{(a)_{m+n}(b_1)_m(b_2)_n}{(c)_{m+n}\,m!\,n!} x^my^n\quad(|x|<1,\,|y|<1)\\
&=\frac{\Gamma(c)}{\Gamma(a)\Gamma(c-a)}
\int_0^1t^{a-1}(1-t)^{c-a-1}(1-xt)^{-b_1}(1-yt)^{-b_2}dt\quad(\Re(a)>0,\,\Re(c-a)>0).
\end{align*}
We put
\[ a=1-X, c=a+1=2-X, b_1=1-iY, b_2=1+iY, x=i,y=-i \]
and obtain
\begin{align*}
&\int_0^1t^{-X}(1-it)^{iY-1}(1+it)^{-iY-1}dt\\
&=\frac{\Gamma(1-X)\Gamma(1)}{\Gamma(2-X)}
F_1(1-X;1-iY,1+iY;2-X;i,-i) \\
&=\frac1{1-X}F_1(1-X;1-iY,1+iY;2-X;i,-i).
\end{align*}
From this the result follows.
\end{proof}
\subsection{A relation of multiple ${\widetilde{T}}$-values involving Entringer numbers}\label{subsec:Ent}
To state the next theorem, we review the `Entringer number', which counts the number of `down-up'
permutations in the symmetric group $S_{n+1}$ starting with $j+1$.
Alternatively, the Entringer numbers $\{{\mathbb{E}}(n,j)\mid n,j\in \mathbb{Z}_{\geq 0},\ 0\leq j\leq n\}$ are
defined inductively by
\begin{align*}
& {\mathbb{E}}(0,0)=1,\quad {\mathbb{E}}(n,0)=0\ \ (n>0),\\
& {\mathbb{E}}(n,j)={\mathbb{E}}(n,j-1)+{\mathbb{E}}(n-1,n-j)\ \ (n\geq 1,\ 1\leq j\leq n).
\end{align*}
Note in particular that ${\mathbb{E}}(n,j)\in \mathbb{Z}_{\geq 0}$. For more details, see Entringer \cite{Entringer1666} and also Stanley \cite{Stanley2010}.
\renewcommand{\arraystretch}{1.1}
\begin{table}[h]
\begin{center}
\begin{tabular} {|c|c|c|c|c|c|c|c|c|c|c|} \hline
\backslashbox{$\ \ n$}{$j\ \ $} & $ 0 $ & $ 1 $ & $ 2 $ & $ 3 $ & $ 4 $ & $ 5 $ & $ 6 $ & $ 7 $
\\ \hline
$0$ & $1$ & & & & & & &
\\ \hline
$1$ & $0$ & $1$ & & & & & &
\\ \hline
$2$ & $0$ & $1$ & $1$ & & & & &
\\ \hline
$3$ & $0$ & $1$ & $2$ & $2$ & & & &
\\ \hline
$4$ & $0$ & $2$ & $4$ & $5$ & $5$ & & &
\\ \hline
$5$ & $0$ & $5$ & $10$ & $14$ & $16$ & $16$ & &
\\ \hline
$6$ & $0$ & $16$ & $32$ & $46$ & $56$ & $61$ & $61$ &
\\ \hline
$7$ & $0$ & $61$ & $122$ & $178$ & $224$ & $256$ & $272$ & $272$
\\ \hline
\end{tabular}
\end{center}
\label{tab:Bnk}
\caption{${\mathbb{E}}(n,j) \: (0\le n,\,j \leq 7) $}
\end{table}
Incidentally, let ${\mathbb{E}}_0=1$ and define
\begin{equation}
{\mathbb{E}}_n=\sum_{j=0}^{n-1}{\mathbb{E}}(n-1,j)\ (={\mathbb{E}}(n,n))\quad (n\geq 1). \label{Euler-2}
\end{equation}
Then ${\mathbb{E}}_n$ is sometimes called the `Euler number' (see \cite{Stanley2010})
and is equal to the total number of `down-up' (or `up-down') permutations in $S_n$.
A generating function of ${\mathbb{E}}_n$ is
\begin{equation}
\sec x+\tan x=\sum_{n=0}^\infty {\mathbb{E}}_n\frac{x^n}{n!} \label{gene-Euler-2}
\end{equation}
(see \cite[Theorem 1.1]{Stanley2010}). Hence $(-1)^n{\mathbb{E}}_{2n}=E_{2n}$ is the
usual Euler number defined by
\begin{equation}\label{euler}
\sum_{n=0}^\infty E_n \frac{x^n}{n!} = \frac1{\cosh x}
\end{equation}
(see N\"orlund \cite[Chap.\,2]{N1924}, note that $\cosh x$ is an even function) and the odd-indexed ${\mathbb{E}}_{2n+1}$ coincides with the `tangent number' (see \cite{N1924}).
\begin{table}[h]
\begin{center}
\begin{tabular} {|c|c|c|c|c|c|c|c|c|c|c|c|} \hline
$\ \ n$ & $ 0 $ & $ 1 $ & $ 2 $ & $ 3 $ & $ 4 $ & $ 5 $ & $ 6 $ & $ 7 $ & $ 8 $ & $ 9 $ & $10$
\\ \hline
$\ \ {\mathbb{E}}_n$ & $1$ & $1$ & $1$ & $2$ & $5$ & $16$ & $61$ & $272$ & $1385$ & $7936$ & $50521$
\\ \hline
\end{tabular}
\end{center}
\label{tab:En-9}
\caption{${\mathbb{E}}_n \: (0\le n \leq 10) $}
\end{table}
Using ${\mathbb{E}}_n$, the well-known formulas for Riemann zeta values as well as Dirichlet $L$-values of conductor 4
(sometimes referred to as the Dirichlet beta values) can be re-written as formulas for our $T$- and ${\widetilde{T}}$-values
in a uniform manner as
\begin{equation}
\begin{rcases}
T(n+1) & (n:\,\text{\rm odd}) \\
{\widetilde{T}}(n+1) & (n:\,\text{\rm even})
\end{rcases} =\frac{{\mathbb{E}}_{n}}{n!}\left(\frac{\pi}{2}\right)^{n+1}
\label{Euler-Ttilde}
\end{equation}
for any $n\in \mathbb{Z}_{\geq 0}$
(see Comtet \cite{Comtet1974}).
Coming back to the Entringer number, we have the following curious and beautiful relations. Note that
the values on the right are not appearing in \eqref{Euler-Ttilde}, {\it i.e.} the `difficult' values presumably
not rational multiples of powers of $\pi$.
\begin{theorem}\label{Ent-rel}\ \ For any $n\in \mathbb{Z}_{\geq 1}$, we have
\begin{align}
& \sum_{j=1}^{n}{\mathbb{E}}(n,j){\widetilde{T}}(\underbrace{1,\ldots,1,\overset{\tiny j}{\check{2}},1,\ldots,1}_{n}) =
\begin{cases}
\widetilde{T}(n+1) & (n:\,\text{\rm odd}),\\
T(n+1) & (n:\,\text{\rm even}).
\end{cases}
\label{Entringer-Ttilde}
\end{align}
\end{theorem}
\begin{example}
Examples in low weights are
\begin{align*}
T(3)&={\widetilde{T}}(2,1)+{\widetilde{T}}(1,2),\\
{\widetilde{T}}(4)&={\widetilde{T}}(2,1,1)+2{\widetilde{T}}(1,2,1)+2{\widetilde{T}}(1,1,2),\\
T(5)&=2{\widetilde{T}}(2,1,1,1)+4{\widetilde{T}}(1,2,1,1)+5{\widetilde{T}}(1,1,2,1)+5{\widetilde{T}}(1,1,1,2),\\
{\widetilde{T}}(6)&=5{\widetilde{T}}(2,1,1,1,1)+10{\widetilde{T}}(1,2,1,1,1)+14{\widetilde{T}}(1,1,2,1,1)\\
& \quad +16{\widetilde{T}}(1,1,1,2,1)+16{\widetilde{T}}(1,1,1,1,2),\\
T(7)&=16{\widetilde{T}}(2,1,1,1,1,1)+32{\widetilde{T}}(1,2,1,1,1,1)+46{\widetilde{T}}(1,1,2,1,1,1)\\
& \quad +56{\widetilde{T}}(1,1,1,2,1,1)+61{\widetilde{T}}(1,1,1,1,2,1)+61{\widetilde{T}}(1,1,1,1,1,2).
\end{align*}
\end{example}
\begin{proof}
We give a proof by induction on $n\geq 1$. The case $n=1$ becomes the trivial identity ${\widetilde{T}}(2)={\widetilde{T}}(2)$.
Let $k\geq 2$ and assume that the assertions for $n\leq k-1$ hold and consider the case $n=k$.
From the identity~\eqref{key}, by taking out the term with $j=k$ on the right-hand side and noting $(k+1){\widetilde{T}}(\underbrace{1,\ldots,1}_{k+1})
={\widetilde{T}}(1){\widetilde{T}}(\underbrace{1,\ldots,1}_{k})$ corresponds to the term $j=0$, we have
\begin{align*}
& i^{k} T(k+1) -i^{k+1} {\widetilde{T}}(k+1) =\sum_{j=0}^{k-1}i^{j+1}{\widetilde{T}}(j+1){\widetilde{T}}(\underbrace{1,\ldots,1}_{k-j})+{\widetilde{T}}(\underbrace{1,\ldots,1}_{k-1},2).
\end{align*}
We use the real part of this identity:
\begin{align}
& {\rm Re}\left(i^{k} T(k+1) -i^{k+1} {\widetilde{T}}(k+1)\right) =\sum_{j=0 \atop \textrm{$j$\,:\,odd}}^{k-1}(-1)^{(j+1)/2}\,{\widetilde{T}}(j+1){\widetilde{T}}(\underbrace{1,\ldots,1}_{k-j})+{\widetilde{T}}(\underbrace{1,\ldots,1}_{k-1},2). \label{eq-5-1-1}
\end{align}
By the induction hypothesis, we can substitute
$$\widetilde{T}(j+1)=\sum_{h=1}^{j}{\mathbb{E}}(j,h){\widetilde{T}}(\underbrace{1,\ldots,1,\overset{\tiny h}{\check{2}},1,\ldots,1}_{j})\quad (\textrm{$j$\,:\,odd},\ 1\leq j\leq k-1)$$
into \eqref{eq-5-1-1} and obtain
\begin{align*}
& {\rm Re}\left(i^{k} T(k+1) -i^{k+1} {\widetilde{T}}(k+1)\right) \notag \\
& \ \ =\sum_{j=1 \atop \textrm{$j$\,:\,odd}}^{k-1}(-1)^{(j+1)/2}\sum_{h=1}^{j}{\mathbb{E}}(j,h){\widetilde{T}}(\underbrace{1,\ldots,1,\overset{\tiny h}{\check{2}},1,\ldots,1}_{j})
{\widetilde{T}}(\underbrace{1,\ldots,1}_{k-j})+{\widetilde{T}}(\underbrace{1,\ldots,1}_{k-1},2).
\end{align*}
Since multiple ${\widetilde{T}}$-values satisfy the same shuffle product formulas as the usual multiple zeta values, we have the following relation:
\begin{align*}
& {\widetilde{T}}(\underbrace{1,\ldots,1,\overset{\tiny h}{\check{2}},1,\ldots,1}_{m}){\widetilde{T}}(\underbrace{1,\ldots,1}_{n}) =\sum_{l=h}^{n+h}\binom{l}{h}\binom{m+n-l}{m-h}{\widetilde{T}}(\underbrace{1,\ldots,1,\overset{\tiny l}{\check{2}},1,\ldots,1}_{m+n}).
\end{align*}
Therefore we obtain
\begin{align*}
& {\rm Re}\left(i^{k} T(k+1) -i^{k+1} {\widetilde{T}}(k+1)\right)-{\widetilde{T}}(\underbrace{1,\ldots,1}_{k-1},2) \notag \\
& \ \ =\sum_{j=1 \atop \textrm{$j$\,:\,odd}}^{k-1}(-1)^{(j+1)/2}\sum_{h=1}^{j}{\mathbb{E}}(j,h)\sum_{l=h}^{k-j+h}\binom{l}{h}\binom{k-l}{j-h}{\widetilde{T}}(\underbrace{1,\ldots,1,\overset{\tiny l}{\check{2}},1,\ldots,1}_{k})\\
& \ \ =\sum_{l=1}^{k}\left(\sum_{j=1 \atop \textrm{$j$\,:\,odd}}^{k-1}(-1)^{(j+1)/2}\sum_{h=1}^{j}{\mathbb{E}}(j,h)\binom{l}{h}\binom{k-l}{j-h}\right){\widetilde{T}}(\underbrace{1,\ldots,1,\overset{\tiny l}{\check{2}},1,\ldots,1}_{k})\\
& \ \ =\sum_{l=1}^{k}\sum_{h=1}^{k-1}\sum_{j=h \atop \textrm{$j$\,:\,odd}}^{k-1}(-1)^{(j+1)/2}\binom{l}{h}\binom{k-l}{j-h}{\mathbb{E}}(j,h){\widetilde{T}}(\underbrace{1,\ldots,1,\overset{\tiny l}{\check{2}},1,\ldots,1}_{k})\\
& \ \ =\sum_{l=1}^{k}\sum_{h=1}^{l}\sum_{j=h \atop \textrm{$j$\,:\,odd}}^{k}(-1)^{(j+1)/2}\binom{l}{h}\binom{k-l}{j-h}{\mathbb{E}}(j,h){\widetilde{T}}(\underbrace{1,\ldots,1,\overset{\tiny l}{\check{2}},1,\ldots,1}_{k})\\
& \qquad -\delta_{k,\textrm{odd}}\sum_{l=1}^{k}(-1)^{(k+1)/2}{\mathbb{E}}(k,l){\widetilde{T}}(\underbrace{1,\ldots,1,\overset{\tiny l}{\check{2}},1,\ldots,1}_{k}),
\end{align*}
where $\delta_{k,\textrm{odd}}=1\ ({\rm resp.}\ 0)$ if $k$ is odd (resp. even).
By replacing $j$ with $h+j$, the last expression becomes
\begin{align*}
& \sum_{l=1}^{k}\sum_{h=1}^{l}\sum_{j=0 \atop \textrm{$h+j$\,:\,odd}}^{k-l}(-1)^{(h+j+1)/2}\binom{l}{h}\binom{k-l}{j}{\mathbb{E}}(h+j,h){\widetilde{T}}(\underbrace{1,\ldots,1,\overset{\tiny l}{\check{2}},1,\ldots,1}_{k})\\
& \qquad -\delta_{k,\textrm{odd}}\sum_{l=1}^{k}(-1)^{(k+1)/2}{\mathbb{E}}(k,l){\widetilde{T}}(\underbrace{1,\ldots,1,\overset{\tiny l}{\check{2}},1,\ldots,1}_{k}).
\end{align*}
Here, we need
\begin{lemma}\label{Lem-5-2}\ \ For $k,l\in \mathbb{Z}_{\geq 0}$,
\begin{align}
& \sum_{h=0}^{l}\sum_{j=0}^{k-l}\,i^{h+j}\binom{l}{h}\binom{k-l}{j}{\mathbb{E}}(h+j,h) =i^{k-1}{\mathbb{E}}(k,l)+\delta_{k,l}\,i+\delta_{l,0}, \label{eq-Lem-5-2}
\end{align}
where $\delta_{k,l}$ is the Kronecker delta.
\end{lemma}
Multiplying both sides of \eqref{eq-Lem-5-2} by $i$ and taking the real part, we have
$$\sum_{h=1}^{l}\sum_{j=0 \atop \textrm{$h+j$\,:\,odd}}^{k-l}(-1)^{(h+j+1)/2}\binom{l}{h}\binom{k-l}{j}{\mathbb{E}}(h+j,h) =\delta_{k,\textrm{even}}(-1)^{k/2}{\mathbb{E}}(k,l)-\delta_{k,l},$$
where $\delta_{k,\textrm{even}}=1\ ({\rm resp.}\ 0)$ if $k$ is even (resp. odd).
Therefore we obtain
\begin{align*}
& {\rm Re}\left(i^{k} T(k+1) -i^{k+1} {\widetilde{T}}(k+1)\right) \notag \\
& \ \ =\sum_{l=1}^{k}\left(\delta_{k,\textrm{even}}(-1)^{k/2}{\mathbb{E}}(k,l)-\delta_{k,l}\right){\widetilde{T}}(\underbrace{1,\ldots,1,\overset{\tiny l}{\check{2}},1,\ldots,1}_{k}) +{\widetilde{T}}(\underbrace{1,\ldots,1}_{k-1},2)\\
& \qquad -\delta_{k,\textrm{odd}}\sum_{l=1}^{k}(-1)^{(k+1)/2}{\mathbb{E}}(k,l){\widetilde{T}}(\underbrace{1,\ldots,1,\overset{\tiny l}{\check{2}},1,\ldots,1}_{k})\\
& \ \ =\delta_{k,\textrm{even}}(-1)^{k/2}\sum_{l=1}^{k}{\mathbb{E}}(k,l){\widetilde{T}}(\underbrace{1,\ldots,1,\overset{\tiny l}{\check{2}},1,\ldots,1}_{k})\\
& \qquad -\delta_{k,\textrm{odd}}(-1)^{(k+1)/2}\sum_{l=1}^{k}{\mathbb{E}}(k,l){\widetilde{T}}(\underbrace{1,\ldots,1,\overset{\tiny l}{\check{2}},1,\ldots,1}_{k}).
\end{align*}
This gives the desired identities when $n=k$ for both even and odd $k$.
\end{proof}
\begin{proof}[Proof of Lemma \ref{Lem-5-2}]
We use the following generating function of the Entringer numbers (see \cite[(2.2)]{Stanley2010}):
\begin{align}
&\sum_{k=0}^\infty\sum_{l=0}^{k}\,{\mathbb{E}}(k,l)\frac{x^{k-l}}{(k-l)!}\frac{y^l}{l!}=\frac{\cos x+\sin y}{\cos (x+y)}. \label{generating E-kl}
\end{align}
Let $I_1(x,y)$ be the generating function of the left-hand side of \eqref{eq-Lem-5-2}:
\begin{align*}
I_1(x,y) & =\sum_{k=0}^\infty \sum_{l=0}^{k} \sum_{h=0}^l \sum_{j=0}^{k-l} i^{h+j}\binom{l}{h}\binom{k-l}{j}{\mathbb{E}}(h+j,h)\frac{x^{k-l}}{(k-l)!}\frac{y^l}{l!}.
\end{align*}
By \eqref{generating E-kl}, we have
\begin{align*}
I_1(x,y) & =\sum_{h=0}^\infty \sum_{j=0}^\infty i^{h+j} {\mathbb{E}}(h+j,h) \sum_{l=0}^\infty\binom{l}{h} \frac{y^l}{l!}\sum_{k=l}^{\infty}\binom{k-l}{j}\frac{x^{k-l}}{(k-l)!}\\
& =\sum_{h=0}^\infty \sum_{j=0}^\infty i^{h+j} {\mathbb{E}}(h+j,h) \sum_{l=0}^\infty\binom{l}{h} \frac{y^l}{l!}\sum_{k=0}^{\infty}\binom{k}{j}\frac{x^{k}}{k!}\\
& =\sum_{h=0}^\infty \sum_{j=0}^\infty i^{h+j} {\mathbb{E}}(h+j,h) \frac{x^j}{j!}\frac{y^{h}}{h!}e^{x+y}\\
& =\sum_{m=0}^\infty \sum_{h=0}^{m} i^{m} {\mathbb{E}}(m,h) \frac{x^{m-h}}{(m-h)!}\frac{y^{h}}{h!}e^{x+y}\\
& =\frac{\cos(ix)+\sin(iy)}{\cos(i(x+y))}e^{x+y}\\
& =\frac{e^x+e^{-x}+ie^y-ie^{-y}}{1+e^{-2x-2y}}.
\end{align*}
On the other hand, the generating function of the right-hand side of \eqref{eq-Lem-5-2} is
\begin{align*}
I_2(x,y) & =\sum_{k=0}^\infty \sum_{l=0}^{k} \left(i^{k-1}{\mathbb{E}}(k,l)+\delta_{k,l}i+\delta_{l,0}\right)\frac{x^{k-l}}{(k-l)!}\frac{y^l}{l!}\\
& =-i\sum_{k=0}^\infty \sum_{l=0}^{k} {\mathbb{E}}(k,l)\frac{(ix)^{k-l}}{(k-l)!}\frac{(iy)^l}{l!}+ie^y+e^x\\
& =-i\left(\frac{e^x+e^{-x}+ie^{y}-ie^{-y}}{e^{x+y}+e^{-x-y}}\right)+ie^y+e^x\\
& =\frac{e^x+e^{-x}+ie^y-ie^{-y}}{1+e^{-2x-2y}}.
\end{align*}
Hence $I_1(x,y)=I_2(x,y)$, which gives the proof of \eqref{eq-Lem-5-2}, and now the proof of Theorem~\ref{Ent-rel} is complete.
\end{proof}
\subsection{Relation to modular forms}\label{subsec:modular}
In this subsection, we present our experimental discovery on a possible connection between double ${\widetilde{T}}$-
(as well as $T$-) values and modular forms
of level 4 (and 2), or precisely speaking, connection to some `period polynomials' associated to those modular forms.
This is certainly an analogous phenomenon to our previously studied relations between double zeta values and modular forms on the full modular group (\cite{GKZ}).
We however could not give a proof and leave it to the interested readers.
For $N=2$ and $4$, even integer $k\ge4$, and $1\le j\le (k-2)/2$, define the polynomial $\widetilde S_{N,k,j}(X)$ with rational coefficients by
\begin{align*}
\widetilde S_{N,k,j}(X)&=\frac{N^{k-2j-1}}{k-2j} X^{k-2} B_{k-2j}^0\left(\frac1{NX}\right)-\frac1{2j}B_{2j}^0(X)\\
&\qquad \quad -\frac{kB_{2j}B_{k-2j}}{2j(k-2j)B_k}\left(\frac{1-2^{-2j}}{1-2^{-k}}\frac{X^{k-2}}{N}-\frac{1-2^{-k+2j}}{1-2^{-k}}\frac1{N^{2j}}\right),
\end{align*}
where $B_{n}$ is the Bernoulli number and $B_{n}^0(X)$ is the usual Bernoulli polynomial with the term $nB_1X^{n-1}$ removed:
\[ B_{n}^0(X) =\sum_{0\le j\le n \atop \text{$j$:even}}\binom{n}{j} B_j X^{n-j}. \]
We further define $P_{N,k,j}(X)$ and $P_{N,k,j}^{(\pm)}(X)$ by using $\widetilde S_{N,k,j}(X)$ as
\[ P_{N,k,j}(X)
=(-2X+2)^{k-2}\widetilde S_{N,k,j}\left(\frac{X+1}{-2X+2}\right) \]
and
\[ P_{N,k,j}^{(\pm)}(X)=\frac12\left(P_{N,k,j}(X)\pm P_{N,k,j}(-X)\right). \]
We can now state our conjecture.
\begin{conj}\label{conj-modular}
1) For $N=2$ or $4$, even integers $k\ge4$, and integers $j$ with $1\le j\le (k-2)/2$, write the polynomial $P_{N,k,j}^{(+)}(X+1)$ as
\[P_{N,k,j}^{(+)}(X+1) = \sum_{i=0}^{k-2} a_i \binom{k-2}{i} X^i. \]
(Each coefficient $a_i$ depends on $N$, $k$, and $j$.)
Then we have the following relation among the double ${\widetilde{T}}$-values:
\[ \sum_{i=0}^{k-2} a_i\, {\widetilde{T}}(i+1,k-i-1)=0. \]
The $\mathbb{Q}$-vector space $V_{4,k}$ spanned by $ P_{4,k,j}^{(+)}(X)\ (1\le j \le (k-2)/2)$ is of dimension $\left[(k-2)/4\right]$, which we conjecture
to be equal to the number of independent relations among double ${\widetilde{T}}$-values of weight $k$. The polynomials $P_{2,k,j}^{(+)}(X)$ are contained in
$V_{4,k}$, and span the subspace of dimension $\left[(k-2)/6\right]$.
2) For $N=2$ or $4$, even integers $k\ge4$, and integers $j$ with $1\le j\le (k-2)/2$, write the polynomial $P_{N,k,j}^{(-)}(X+1)$ as
\[P_{N,k,j}^{(-)}(X+1) = \sum_{i=0}^{k-3} b_i \binom{k-2}{i} X^i. \]
(Here too we suppress the dependence on $N$ etc. in the notation of $b_i$. Note that the degree of the odd polynomial $P_{N,k,j}^{(-)}(X)$ is at most $k-3$.)
Then we have the following relation among the double $T$-values:
\[ \sum_{i=0}^{k-3} b_i\, T(i+1,k-i-1)=0. \]
In this case, the $\mathbb{Q}$-vector space $W_k$ spanned by $P_{4,k,j}^{(-)}(X)\ (1\le j \le (k-2)/2)$ is the same as that spanned by $P_{2,k,j}^{(-)}(X)\ (1\le j \le (k-2)/2)$,
and the conjectural dimension of $W_k$ is $\left[k/4\right]-1$. The conjectural number of independent relations among double $T$-values is $k/2-2$.
\end{conj}
\begin{remark}
i) The polynomial $\widetilde S_{N,k,j}(X)$ is the period polynomial $r^{+}(R_{\Gamma_0(N),k-2,2j-1})(X)$ in the work of Fukuhara and Yang \cite{FY}.
A period polynomial is a polynomial associated to a cusp form whose coefficients are `periods' of the given cusp form.
They computed period polynomials explicitly for some specific cusp forms $R_{\Gamma_0(N),k-2,2j-1}$ on the congruence subgroup $\Gamma_0(N)$, and exhibited several
properties of those. For more details, see their paper \cite{FY} and the references therein. For basics of modular forms, see e.g. Miyake~\cite{Miyake}.
ii) The number $\left[(k-2)/4\right]$ appeared in 1) of the above conjecture is equal to the difference $\dim S_k(\Gamma_0(4))-\dim S_k(\Gamma_0(2))$ of dimensions of the spaces of
cusp forms of weight $k$ on the congruence subgroups $\Gamma_0(4)$ and $\Gamma_0(2)$.
And the difference $\left[(k-2)/4\right]-\left[(k-2)/6\right]$ is equal to the dimension of the space of new forms of weight $k$ on $\Gamma_0(4)$. Also,
the number $\left[k/4\right]-1$ in 2) is equal to $\dim S_k(\Gamma_0(2))$, whereas $k/2-2=\dim S_k(\Gamma_0(4))$.
We could not find how to produce the remaining $k/2-2-(\left[k/4\right]-1)=\left[(k-2)/4\right]$ relations of double $T$-values
via a similar procedure.
iii) By Fukuhara and Yang~\cite[Cor.~1.9]{FY2} (resp. \cite[Cor.~1.5]{FY}), the polynomials $\widetilde S_{4,k,j}(X)$ (resp. $\widetilde S_{2,k,j}(X)$)
span the $k/2-2=\dim S_k(\Gamma_0(4))$ (resp. $\left[k/4\right]-1=\dim S_k(\Gamma_0(2))$) dimensional space.
\end{remark}
\begin{example}
i) For $N=4$, $k=6$, and $j=1$, we have
\[ \widetilde S_ {4,6,1}(X)=-\frac12 X^4+\frac12 X^2-\frac1{32},\quad P_{4,6,1}^{(+)}(X)=X^4-10X^2+1\]
and
\begin{align*} P_{4,6,1}^{(+)}(X+1) &= - 8 - 16X - 4X^2+ 4X^3 +X^4\\
& =-8\binom{4}{0}-4\binom{4}{1}X -\frac{2}{3} \binom{4}{2} X^2+1\cdot\binom{4}{3}X^3+1\cdot \binom{4}{4}X^4\\
&=-\frac13\left(24\binom{4}{0}+12\binom{4}{1}X +2\binom{4}{2} X^2-3\binom{4}{3}X^3-3 \binom{4}{4}X^4\right).
\end{align*}
Accordingly, we numerically (to high precision) have
\[ 24{\widetilde{T}}(1,5)+12{\widetilde{T}}(2,4)+2{\widetilde{T}}(3,3)-3{\widetilde{T}}(4,2)-3{\widetilde{T}}(5,1)=0.\]
ii) For $N=2$, $k=8$, and $j=2$, we have
\begin{align*} \widetilde S_ {2,8,2}(X)&=-\frac1{17} X^6+\frac14X^4-\frac18 X^2+\frac1{136},\\
P_{2,8,2}^{(+)}(X)&=-\frac{2}{17}(5X^6-61X^4-61X^2+5)
\end{align*}
and
\begin{align*} P_{2,8,2}^{(+)}(X+1) &=\frac2{17}(112+ 336X+ 352X^2+144X^3- 14X^4 - 30X^5-5 X^6 )\\
&=\frac2{17}\left(112\binom{6}{0}+56\binom{6}{1}X+\frac{352}{15}\binom{6}{2}X^2+\frac{36}{5}\binom{6}{3}X^3\right.\\
&\qquad\qquad \left. -\frac{14}{15}\binom{6}{4}X^4-5\binom{6}{5}X^5-5\binom{6}{6}X^6\right).
\end{align*}
We compute
\[ 112{\widetilde{T}}(1,7)+56{\widetilde{T}}(2,6)+\frac{352}{15}{\widetilde{T}}(3,5)+\frac{36}{5}{\widetilde{T}}(4,4)-\frac{14}{15}{\widetilde{T}}(5,3)-5{\widetilde{T}}(6,2)-5{\widetilde{T}}(7,1)=0 \]
to very high precision.
iii) For $N=4$, $k=8$, and $j=1$, we have
\[\widetilde S_ {4,8,1}(X)=\frac{208}{51} X^6-\frac{16}3 X^4 +\frac76 X^2-\frac{19}{408},\quad
P_{4,8,1}^{(-)}(X)= -\frac{640}{17}\left(X^5 -8 X^3 + X\right) \]
and
\begin{align*} &P_{4,8,1}^{(-)}(X+1) = \frac{640}{17}\left(6+18X+14X^2-2X^3-5X^4-X^5\right) \\
&=\frac{64}{51}\left(180\binom{6}{0}+90\binom{6}{1}X +28\binom{6}{2} X^2-3\binom{6}{3}X^3-10 \binom{6}{4}X^4-5\binom{6}{5} X^5\right).
\end{align*}
The corresponding conjectural relation is
\[ 180T(1,7)+90 T(2,6)+28 T(3,5)-3 T(4,4)-10 T(5,3)-5 T(6,2)=0.\]
\end{example}
At the end of this section, we mention a connection to our previously obtained `weighted sum formula'.
If we start with the polynomials $X^{k-2}$ or $1$ instead of $\widetilde S_{N,k,j}(X)$, we obtain the weighted sum formulas
for double ${\widetilde{T}}$- and $T$- values. More precisely, for even $k\ge 4$, take the even and odd parts of the polynomial
\[ (-2X+2)^{k-2}\left(\frac{X+1}{-2X+2}\right)^{k-2} = (X+1)^{k-2},\]
namely $\left((X+1)^{k-2}\pm (-X+1)^{k-2}\right)/2$ and make the shift $X\to X+1$ to obtain
\[ \frac12 \left((X+2)^{k-2}\pm (-X)^{k-2}\right) =\frac12 \left( \sum_{j=0}^{k-2} 2^{k-j}\binom{k-2}{j} X^j\pm X^{k-2}\right) .\]
Then, this `corresponds' to the weighted sum formula
\[ \sum_{j=0}^{k-2} 2^{k-j-2}{\widetilde{T}}(j+1,k-1-j)+{\widetilde{T}}(k-1,1)=(k-1)T(k) \]
proved in \cite[Prop.~4.2]{AK2004} in the `$+$' case, and
\[ \sum_{j=0}^{k-3} 2^{k-j-2}T(j+1,k-1-j)=(k-1)T(k) \]
proved in \cite[Th.~3.2]{KT2020-Tsukuba} in the `$-$' case. If we start with $1$ instead, the resulting polynomial is essentially the same.
\section{Certain zeta functions and poly-Bernoulli and Euler numbers}\label{sec:zeta-polyEuler}
In our previous work \cite{AK1999, KT2020-ASPM}, we studied zeta functions
\begin{align}
\xi(k_1,\ldots,k_r;s)&=\frac{1}{\Gamma(s)}\int_0^\infty {t^{s-1}}\frac{{\rm Li} (k_1,\ldots,k_r;1-e^{-t})}{e^t-1}\,dt \label{Def-xi} \\
\intertext{and}
\psi(k_1,\ldots,k_r;s)&=\frac{1}{\Gamma(s)}\int_0^\infty {t^{s-1}}\frac{A(k_1,\ldots,k_r;\tanh(t/2))}{\sinh t}\,dt,
\label{Def-psi}
\end{align}
both converge in ${\rm Re}(s)>0$ and are analytically continued to entire functions.
Here,
\[ {\rm Li}(k_1,\ldots,k_r;z)=\sum_{0< m_1<\cdots<m_r} \frac{z^{m_r}}{m_1^{k_1}\cdots m_r^{k_r}}\qquad (k_1,\ldots,k_r\in \mathbb{Z};\ |z|<1) \]
is the multiple polylogarithm and $A(k_1,\ldots,k_r;z)$ is its `level 2' analogue already appeared in \eqref{Def-A-polylog}.
We now introduce a level 4 analogue of these zeta functions by using the level 4 variant ${\mathscr{A}}(k_1,\ldots,k_r;z)$ of multiple
polylogarithm, which is defined by the iterated integral as follows. The idea is just to replace the starting point $0$
of the iterated integral of $A(k_1,\ldots,k_r;z)$ with $i=\sqrt{-1}$.
\begin{definition} \label{Def-2-7}
For $k_1,\ldots,k_r\in \mathbb{Z}_{\geq 1}$, define
\begin{align}\label{Def-aa-multi}
{\mathscr{A}}(k_1,\ldots,k_r;z)
& =
\begin{cases}
\displaystyle{\int_{i}^{z}\frac{1}{u}\,{{\mathscr{A}}(k_1,\ldots,k_{r-1},k_r-1;u)}\,du} & (k_r\geq 2),\\
\displaystyle{\int_{i}^{z}\frac{2}{1-u^2}\,{\mathscr{A}}(k_1,\ldots,k_{r-1};u)\,du} & (k_r=1),
\end{cases}
\end{align}
with ${\mathscr{A}}(\emptyset;u)=1$. In particular,
\begin{equation}\label{aa1} {\mathscr{A}}(1;z)=\int_{i}^{z}\frac{2}{1-u^2}=A(1;z)-A(1;i)=2\tanh^{-1}(z)-\frac{\pi i}{2}. \end{equation}
\end{definition}
Note that if $k_r\ge2$, then we may consider the value of ${\mathscr{A}}(k_1,\ldots,k_r;z)$ at $z=1$, and in fact this was already
appeared in the proof of Theorem~\ref{Teven}. We state the formula again as a proposition.
\begin{prop} For $k_1,\ldots,k_r\in\mathbb{Z}_{\ge1}$ with $k_r\ge2$, let $(l_1,\ldots,l_s)$ be the dual index of
$(k_1,\ldots,k_r)$. Then we have
\begin{equation}\label{aaat1}
{\mathscr{A}}(k_1,\ldots,k_r;1)=i^{r-k}{\widetilde{T}}(l_1,\ldots,l_s).
\end{equation}
\end{prop}
\begin{proof} As already done in the proof of Theorem~\ref{Teven}, this can be shown by the change of variables $t\to (-iu+1)/(u-i)$
in the iterated integral
\begin{align*}
&{\mathscr{A}}(k_1,\ldots,k_r;1)\\
&=\int_i^1\frac{2dt}{1-t^2}\circ\underbrace{\frac{dt}{t}\circ\cdots\circ\frac{dt}{t}}_{k_1-1}\circ
\frac{2dt}{1-t^2}\circ\underbrace{\frac{dt}{t}\circ\cdots\circ\frac{dt}{t}}_{k_2-1}\circ\cdots\cdots\circ
\frac{2dt}{1-t^2}\circ\underbrace{\frac{dt}{t}\circ
\cdots\circ\cdots\circ\frac{dt}{t}}_{k_r-1}
\end{align*}
and by the formula \eqref{integ-exp2}.
\end{proof}
We also record here a formula of $A(\underbrace{1,\ldots,1}_{r-1},k+1;z)$ expressed in terms of ${\widetilde{T}}$-values, ${\mathscr{A}}(\underbrace{1,\ldots,1}_{j-1},k+1;z)\ (1\le j\le r)$, and $\log z$.
Specialization $z=1$ gives \eqref{TbyTtilde}.
\begin{prop} For $r,k\ge1$, we have
\begin{align*}
&A(\underbrace{1,\ldots,1}_{r-1},k+1;z)\\
&=\sum_{j=1}^r i^{r-j}{\widetilde{T}}(\underbrace{1,\ldots,1}_{r-j}){\mathscr{A}}(\underbrace{1,\ldots,1}_{j-1},k+1;z)
+\sum_{j=0}^k i^r {\widetilde{T}}(\underbrace{1,\ldots,1}_{r-1},j+1)\frac{(\log z-\log i)^{k-j}}{(k-j)!}.
\end{align*}
\end{prop}
\begin{proof} This can be shown in the same manner using the path composition formula
as in the calculation of \eqref{TbyTtilde}. We omit the detail here.
\end{proof}
Now we define the zeta function associated to our ${\mathscr{A}}$.
\begin{definition}\label{Def-3-8}\
For $k_1,\ldots,k_r\in \mathbb{Z}_{\geq 1}$, set
\begin{equation}
\lambda(k_1,\ldots,k_r;s)=\frac{1}{\Gamma(s)}\int_0^\infty {t^{s-1}}\frac{{\mathscr{A}} (k_1,\ldots,k_r;\tanh(t/2+\pi i/4))}{\cosh t}\,dt\quad ({\rm Re}(s)>0). \label{Def-lambda}
\end{equation}
\end{definition}
Before discussing the convergence of the integral in ${\rm Re}(s)>0$, let us first explain our motivation of introducing these functions,
in other words, explain how we chose the integrants of these. For this, we recall `poly-Bernoulli numbers'.
Poly-Bernoulli numbers, having two versions $B_n^{(k)}$ and $C_n^{(k)}$, were defined by
the first named author in \cite{Kaneko1997} and
in Arakawa-Kaneko \cite{AK1999} by using generating series. For an integer $k\in \mathbb{Z}$,
the sequences $\{B_n^{(k)}\}$ and $\{C_n^{(k)}\}$ of rational numbers are given by
\begin{align*}
&\frac{{\rm Li}_{k}(1-e^{-t})}{1-e^{-t}}=\sum_{n=0}^\infty B_n^{(k)}\frac{t^n}{n!},\qquad \frac{{\rm Li}_{k}(1-e^{-t})}{e^t-1}=\sum_{n=0}^\infty C_n^{(k)}\frac{t^n}{n!},
\end{align*}
where ${\rm Li}_{k}(z)={\rm Li}(k;z)$ (in the notation above) is the classical polylogarithm function (or rational function when $k\le0$) given by
\begin{equation}
{\rm Li}_{k}(z)=\sum_{m=1}^\infty \frac{z^m}{m^k}\quad (|z|<1). \label{Def-polylog}
\end{equation}
Since ${\rm Li}_1(z)=-\log(1-z)$, we see that
$B_n^{(1)}$ and $C_n^{(1)}$
are usual Bernoulli numbers, where the only difference being $B_n^{(1)}=1/2$ and $C_n^{(1)}=-1/2$.
A multiple version $C_n^{(k_1,\ldots,k_r)}$ of $C_n^{(k)}$ is the multi-poly-Bernoulli numbers defined in Imatomi-Kaneko-Takeda \cite{IKT2014} by the generating series
\begin{equation}\label{mpBern}
\frac{{\rm Li}(k_1,\ldots,k_r;1-e^{-t})}{e^t-1}=\sum_{n=0}^\infty C_n^{(k_1,\ldots,k_r)}\frac{t^n}{n!}.
\end{equation}
The function $\xi(k_1,\ldots,k_r;s)$ was introduced as the one interpolating $\{C_n^{(k_1,\ldots,k_r)}\}$ at negative integer arguments:
$$\xi(k_1,\ldots,k_r;-n)=(-1)^n C_n^{(k_1,\ldots,k_r)}\quad (n\in \mathbb{Z}_{\geq 0}).$$
Note that the left-hand side of the definition \eqref{mpBern} is exactly the same as the function appearing in the integral of the definition
\eqref{Def-xi} of $\xi(k_1,\ldots,k_r;s)$, and we observe that this function can be realized as
\[ \frac{{\rm Li}(k_1,\ldots,k_r;1-e^{-t})}{e^t-1}=\frac{d}{dt}{\rm Li}(k_1,\ldots,k_{r-1},k_r+1;1-e^{-t}) \]
and moreover that the function $1-e^{-t}$ is the inverse of ${\rm Li}(1;t)=-\log(1-t)$:
\[ {\rm Li}(1;1-e^{-t})=t. \]
Likewise, we see that the function appearing in \eqref{Def-psi} is
\[ \frac{A(k_1,\ldots,k_r;\tanh(t/2))}{\sinh t}=\frac{d}{dt}A(k_1,\ldots,k_{r-1},k_r+1;\tanh(t/2)), \]
where $\tanh(t/2)$ is the inverse of $A(1;t)$ (see \eqref{A1exp}).
Our newly introduced function ${\mathscr{A}}(k_1,\ldots,k_r; z)$ satisfies, by definition~\eqref{Def-aa-multi}, the same derivative formula as~\eqref{deriv-a-multi}
and therefore, if we follow the same line, it would be natural to consider the function
\[ \frac{d}{dt}{\mathscr{A}}(k_1,\ldots,k_{r-1},k_r+1;h(t)), \]
where $h(t)$ is the inverse of ${\mathscr{A}}(1;t)$.
\begin{lemma}\label{defofh} The function
\begin{equation}
h(x)=\tanh (x)+\frac{i}{\cosh (x)}, \label{Def-tau}
\end{equation}
which is also equal to
\begin{equation}
\tanh\left(\frac{x}{2}+\frac{\pi i}{4}\right), \label{tau-exp1}
\end{equation}
is the inverse of ${\mathscr{A}}(1;x)$. We have
\begin{equation*}
\frac{d}{dx}h(x)=-\frac{i}{\cosh x}h(x), \label{tau-deri}
\end{equation*}
and thus for $k_1,\ldots,k_r\in \mathbb{Z}_{\geq 1}$,
\begin{equation}
\frac{d}{dx}{\mathscr{A}} (k_1,\ldots,k_r;h(x))=
\begin{cases}
-i\,\dfrac{{\mathscr{A}} (k_1,\ldots,k_r-1;h(x))}{\cosh x} & (k_r\geq 2)\\
-i{\mathscr{A}}(k_1,\ldots,k_{r-1};h(x)) & (k_r=1).
\end{cases}
\label{deriv-aa}
\end{equation}
\end{lemma}
\begin{proof}
We know from \eqref{aa1} that
\[ {\mathscr{A}}(1;x)=2\tanh^{-1}(x)-\frac{\pi i}{2}, \]
and so $\tanh\left(x/2+\pi i/4\right)$ is the inverse of ${\mathscr{A}}(1;x)$:
\[ {\mathscr{A}}(1;\tanh\left(\frac{x}{2}+\frac{\pi i}{4}\right))=2\tanh^{-1}(\tanh\left(\frac{x}{2}+\frac{\pi i}{4}\right))-\frac{\pi i}{2}=x.\]
Using the duplication formulas, we compute
\begin{align*}
\tanh (x)+\frac{i}{\cosh (x)}&=\frac{2\sinh (x/2) \,\cosh (x/2) +i(\cosh^2 (x/2)-\sinh^2 (x/2))}{\cosh^2(x/2) +\sinh^2 (x/2)}\\
& =\frac{i(\cosh (x/2)-i\sinh (x/2))^2}{(\cosh (x/2) +i\sinh (x/2))(\cosh (x/2) -i\sinh (x/2))}\\
& =\frac{(1+i)e^{x/2} -(1-i)e^{-{x/2}}}{(1+i)e^{x/2} +(1-i)e^{-{x/2}}}=\frac{e^{{x/2}+\pi i/4} -e^{-{x/2}-\pi i/4}}{e^{{x/2}+\pi i/4} +e^{-{x/2}-\pi i/4}}\\
&=\tanh\left(\frac{x}{2}+\frac{\pi i}{4}\right).
\end{align*}
The derivative formula follows from the definition \eqref{Def-aa-multi}.
\end{proof}
It is amusing to note that this
$h(x)$ is essentially equal to the generating function~\eqref{gene-Euler-2} of the Euler numbers $\{{\mathbb{E}}_n\}$ appeared in \S\ref{subsec:Ent}:
$$\frac{1}{i}h(ix)=\sec x+ \tan x.$$
By using \eqref{deriv-aa}, we may obtain the following estimate and the convergence of the integral \eqref{Def-lambda} in ${\rm Re}(s)>0$ follows.
\begin{lemma}\label{Lem-3-5}
For $k_1,\ldots,k_r\in \mathbb{Z}_{\geq 1}$,
\begin{equation}
{\mathscr{A}} (k_1,\ldots,k_r;h(x))=O\left(x^{k_1+\cdots+k_r}\right)\quad (x\to \infty). \label{Order-aa}
\end{equation}
\end{lemma}
\begin{proof}
We proceed by double induction on $r$ and $k_r$ .
When $r=1$ and $k_1=1$, the left-hand side is equal to $x$ and the assertion is obvious.
For $k_1\geq 2$, by \eqref{deriv-aa} and $\cosh x \to \infty$ $(x\to \infty)$, there exists $C>0$ such that
$$|{\mathscr{A}} (k_1;h(x))|\leq \int_0^{x}\frac{1}{|\cosh u|}|{\mathscr{A}} (k_1-1;h(u))|\,du\leq C\int_0^{x}u^{k_1-1}\,du=\frac{C}{k_1}x^{k_1}$$
for sufficiently large $x>0$. Thus by induction we obtain the assertion when $r=1$.
Consider the case $r\geq 2$. When $k_r=1$, then we immediately obtain the assertion from \eqref{deriv-aa} and the
case $r-1$ because
\[ {\mathscr{A}} (k_1,\ldots,k_r;h(x))=-i \int_0^x {\mathscr{A}} (k_1,\ldots,k_{r-1};h(u))\,du. \]
Note that $h(0)=i$ and ${\mathscr{A}} (k_1,\ldots,k_r;h(0))=0$.
If $k_r\geq 2$, then, again by \eqref{deriv-aa} we may argue similarly as in the case of $r=1$ and we obtain the assertion
by induction.
\end{proof}
\begin{example} By the standard integral representation of $L(s,\chi_4)$,
\[ L(s,\chi_4)=\frac1{2\Gamma(s)}\int_0^\infty \frac{t^{s-1}}{\cosh t}\, dt\quad ({\rm Re}(s)>0), \]
we have
$$\lambda(1;s)=\frac1{\Gamma(s)}\int_0^\infty \frac{t^{s}}{\cosh t}\, dt=2sL(s+1,\chi_4).$$
\end{example}
As in our previous cases (of `level 1' \cite{IKT2014} and `level 2' \cite{KPT}), we may define multi-poly-Euler numbers as follows
and connect them to values of the function $\lambda$ at negative integer arguments.
\begin{definition}\label{Def-3-1}
For $k_1,\ldots,k_r\in \mathbb{Z}_{\geq 1}$, define multi-poly-Euler numbers $\{{\mathcal{E}}_n^{(k_1,\ldots,k_r)}\}$ by
\begin{equation}
\frac{{\mathscr{A}} (k_1,\ldots,k_r;\tanh\left(t/2+\pi i/4\right))}{\cosh t}=\sum_{n=0}^\infty {\mathcal{E}}_n^{(k_1,\ldots,k_r)}\,\frac{t^n}{n!}. \label{Def-Poly-Euler}
\end{equation}
\end{definition}
\begin{remark}\label{Rem-Poly-Euler}
The case of $r=1$ and $k_1=1$ becomes
\begin{equation*}
\frac{t}{\cosh t}=\sum_{n=0}^\infty {\mathcal{E}}_n^{(1)}\,\frac{t^n}{n!},
\end{equation*}
namely ${\mathcal{E}}_{m+1}^{(1)}=(m+1)E_m$ $(m\in \mathbb{Z}_{\geq 0})$, $E_m$ being Euler numbers defined in \eqref{euler}.
Also it should be noted that
$$\frac{t}{\cosh t}=-2\sum_{a=1}^{4}\frac{\chi_4(a)te^{at}}{e^{4t}-1}=-2\sum_{n=0}^\infty B_{n,\chi_4}\frac{t^n}{n!},$$
where $B_{n,\chi_4}$ is the generalized Bernoulli number associated to the character $\chi_4$ of conductor 4.
Hence ${\mathcal{E}}^{(k)}_{n}$ is also regarded as the generalized poly-Bernoulli number of conductor $4$.
\end{remark}
\begin{remark}
Sasaki \cite{Sasaki2012}, about a decade ago, defined the poly-Euler numbers by using a different generating function as
\begin{equation}
\frac{{\rm Li}_{k}(1-e^{-4t})}{4t\cosh t}=\sum_{n=0}^\infty E_n^{(k)}\frac{t^n}{n!}, \label{Def-poly-Euler}
\end{equation}
and studied their properties together with Ohno in Ohno-Sasaki \cite{OS2012,OS2013,OS2017}.
When $k=1$, this is
$$\frac{1}{\cosh t}=\sum_{n=0}^\infty E_n^{(1)}\frac{t^n}{n!}$$
and $E_n^{(1)}$ coincides with the classical Euler number $E_n$.
Also Hamahata \cite{Hama2014} defined the poly-Euler polynomials along the same line.
In \cite{OS2017}, Ohno and Sasaki further defined the zeta function of Arakawa-Kaneko type by
$$L_k(s)=\frac{1}{\Gamma(s)}\int_0^\infty t^{s-1}\frac{{\rm Li}_k(1-e^{-4t})}{8\cosh t}\,dt\quad ({\rm Re}(s)>0,\ k\in \mathbb{Z}_{\geq 1})$$
which satisfies
$$L_k(-n)=\frac{(-1)^n n E_{n-1}^{(k)}}{2}\quad (n\in \mathbb{Z}_{\geq 1}).$$
However, as far as the authors realize, it is unclear whether $L_k(s)$ connects poly-Euler numbers and certain multiple series as $\xi$ and $\psi$ functions did.
Poly-Bernoulli numbers $\{B_n^{(-k)}\}$ and $\{C_n^{(-k)}\}$ satisfy a kind of duality relations (\cite{Kaneko1997, Kaneko2010}), while those for poly-Euler numbers $\{E_n^{(-k)}\}$ (defined by \eqref{Def-poly-Euler}) are unknown.
From the viewpoint of the current paper, we shall define poly-Euler numbers with non-positive indices $\{ {\mathcal{E}}_n^{(-k)}\}$ $(k\geq 0)$ and a related zeta function,
and study their properties in our forthcoming paper \cite{KKT2022} with Komori.
\end{remark}
\begin{remark}\label{Poly-Euler-parity}
From \eqref{deriv-aa} and
$${\mathscr{A}} (k_1,\ldots,k_{r};h(0))={\mathscr{A}} (k_1,\ldots,k_{r};i)=0,$$
we can show by induction that the identity
\begin{equation}
{\mathscr{A}} (k_1,\ldots,k_r;h(-x))=(-1)^{k_1+\cdots+k_r}{\mathscr{A}} (k_1,\ldots,k_r;h(x)) \label{Parity-aa-2}
\end{equation}
holds (take the derivatives of both sides). Hence, if $n$ and $k_1+\cdots+k_r$ are of different parity then
${\mathcal{E}}_n^{(k_1,\ldots,k_r)}=0$.
\end{remark}
\begin{example}\label{Exam-3-3}\
Consider the case of $r=1$ and $k_1=2$.
As in the proof of \cite[Theorem 3.1]{Kaneko1997}, using \eqref{tau-deri}, we have
\begin{align*}
\frac{{\mathscr{A}}(2;h(t))}{\cosh t}&=-\frac{i}{\cosh t}\int_0^{t}\frac{{\mathscr{A}}(1;h(v))}{\cosh v}\,dv\\
& =-i\sum_{m=0}^{\infty}{\mathcal{E}}_m^{(1)} \frac{t^{m-1}}{m!}\sum_{n=0}^\infty {\mathcal{E}}_n^{(1)}\frac{t^{n+1}}{(n+1)!}\\
& =-i\sum_{N=0}^\infty \sum_{j=0}^{N}\binom{N}{j}{\mathcal{E}}_{N-j}^{(1)}\frac{{\mathcal{E}}_{j}^{(1)}}{j+1}\,\frac{t^N}{N!}.
\end{align*}
Hence
$${\mathcal{E}}_N^{(2)}=-i\sum_{j=0}^{N}\binom{N}{j}\frac{{\mathcal{E}}_{N-j}^{(1)}{\mathcal{E}}_{j}^{(1)}}{j+1}\quad (N\in \mathbb{Z}_{\geq 0}).$$
Since ${\mathcal{E}}_0^{(1)}=0$ and ${\mathcal{E}}_{N}^{(1)}=NE_{N-1}$ $(N\geq 1)$, we have ${\mathcal{E}}_{2m}^{(1)}=0$ $(m\in \mathbb{Z}_{\geq 0})$ and
$${\mathcal{E}}_{1}^{(1)}=1,\ {\mathcal{E}}_{3}^{(1)}=-3,\ {\mathcal{E}}_{5}^{(1)}=25,\ {\mathcal{E}}_{7}^{(1)}=-427,\ \ldots$$
Therefore we have ${\mathcal{E}}_{2m+1}^{(2)}=0$ $(m\in \mathbb{Z}_{\geq 0})$ and
$${\mathcal{E}}_0^{(2)}=0,\ \ {\mathcal{E}}_2^{(2)}=-i,\ \ {\mathcal{E}}_4^{(2)}=9i,\ \ {\mathcal{E}}_6^{(2)}=-145i,\ \ldots$$
\end{example}
Using the well-known method of contour integration (see, for example, Washington \cite[Theorem 4.2]{Wash1997}), and noting Lemma \ref{Lem-3-5}, we can easily establish the following.
\begin{prop}\label{Prop-3-9}
For $k_1,\ldots,k_r\in \mathbb{Z}_{\geq 1}$, $\lambda(k_1,\ldots,k_r;s)$ can be analytically continued to the whole complex plane and satisfies
$$\lambda(k_1,\ldots,k_r;-n)=(-1)^n {\mathcal{E}}_n^{(k_1,\ldots,k_r)}\quad (n\in \mathbb{Z}_{\geq 0}).$$
\end{prop}
\section{Relations between the $\lambda$-function and the one variable multiple ${\widetilde{T}}$-function}\label{sec:zetafunct}
As in \cite{AK1999}, we consider the one variable function
${\widetilde{T}}(k_1,\ldots,k_{r-1},s)$ defined by replacing $k_r$ with a variable $s$ in \eqref{ttilde-def}:
\begin{equation*}\label{ttilde-funct} {\widetilde{T}}(k_1,\ldots,k_{r-1},s):=
2^r\sum_{1\leq m_1<\cdots<m_r \atop m_j \equiv j \bmod 2} \frac{(-1)^{(m_r-r)/2}}{m_1^{k_1}\cdots m_{r-1}^{k_{r-1}}m_r^{s}}. \end{equation*}
Then, based on the following integral expression, we can obtain in an almost similar manner as in the previous cases
exactly the same relations among
this function, $\lambda$-function, and the ${\widetilde{T}}$-values. Since the proofs are similar, we only give brief outlines.
\begin{lemma}[cf. \cite{AK1999}\,Theorem 3;\ \cite{KT2020-ASPM}\,Lemma 5.4] \label{Lem-4-1} For $l_1,\ldots,l_{r-1}\in \mathbb{Z}_{\geq 1}$ and ${\rm Re}(s)>1$,
\begin{align}
{\widetilde{T}}(l_1,\ldots,l_{r-1},s) &= \frac{1}{\Gamma(l_1)\cdots\Gamma(l_{r-1})\Gamma(s)}\int_{0}^\infty \cdots \int_{0}^\infty x_1^{l_1-1}\cdots x_{r-1}^{l_{r-1}-1}x_r^{s-1}\notag\\
& \qquad \times \prod_{j=1}^{r}\frac{1}{\cosh(x_j+\cdots+x_r)}dx_1\cdots dx_r. \label{TT-integral}
\end{align}
\end{lemma}
\begin{proof}
Exactly the same method as in the proof of \cite[Theorem 3]{AK1999} using
\begin{align*}
& \frac{1}{\cosh(x_j+\cdots+x_r)}=\frac{2e^{-x_j-\cdots-x_r}}{1+e^{-2(x_j+\cdots+x_r)}} =2\sum_{m_j=0}^\infty (-1)^{m_j}e^{-(2m_j+1)(x_j+\cdots+x_r)}
\end{align*}
works, and the iterated integral can be computed to obtain the assertion.
\end{proof}
\begin{theorem}[cf. \cite{AK1999}\,Theorem 8;\ \cite{KT2020-ASPM}\,Theorem 5.3] \label{Th-4-2} For $r, k\in \mathbb{Z}_{\geq 1}$,
\begin{align}
& \lambda(\underbrace{1,\ldots,1}_{r-1},k;s) \notag \\
&=(-1)^{k-1}i^{1-k}\sum_{a_1,\ldots,a_k\geq 0 \atop a_1+\cdots+a_k=r}
\binom{s+a_k-1}{a_k}\cdot {\widetilde{T}}(a_1+1,\ldots,a_{k-1}+1,a_k+s)\notag\\
&\quad +i^{1-k}\sum_{j=0}^{k-2}(-1)^{j}\,{\widetilde{T}}(\underbrace{1,\ldots,1}_{k-2-j},r+1)\cdot
{\widetilde{T}}(\underbrace{1,\ldots,1}_{j},s).\label{eq-Th-4-2}
\end{align}
\end{theorem}
\begin{proof}
The method of the proof is similar to that of \cite[Theorem 8]{AK1999} and \cite[Theorem 5.7]{KT2020-ASPM} (see also \cite[Theorem 7]{Sasaki2012}).
Given $r,k\ge1$, introduce the following integral
\begin{align*}
J_\nu^{(r,k)}(s)& =\frac{1}{\Gamma(s)}\! \int_{0}^\infty \!\!\!\cdots\! \int_{0}^\infty
\frac{{\mathscr{A}}(\overbrace{1,\ldots,1}^{r-1},\nu;h(x_\nu+\cdots+x_k))}{\prod_{l=\nu}^{k}
\cosh(x_l+\cdots+x_k)} x_k^{s-1}\,dx_\nu\cdots dx_k\quad (1\leq \nu \leq k).
\end{align*}
We compute $J_1^{(r,k)}(s)$ in two different ways. First, since
\[ {\mathscr{A}}(\underbrace{1,\ldots,1}_{r};h(x_1+\cdots+x_k))=\frac{{\mathscr{A}}(1;h(x_1+\cdots+x_k))^r}{r!}=
\frac{(x_1+\cdots+x_k)^r}{r!} \]
by the shuffle product and Lemma~\ref{defofh}, we have
\begin{align*}
J_1^{(r,k)}(s)& =\frac{1}{\Gamma(s)\, r!}\int_{0}^\infty \cdots \int_{0}^\infty
\frac{(x_1+\cdots+x_k)^rx_k^{s-1}}{\prod_{l=1}^{k}\cosh(x_l+\cdots+x_k)}\,dx_1\cdots dx_k\\
& =\frac{1}{\Gamma(s)}\!\!\sum_{a_1+\cdots+a_k=r}
\frac{1}{a_1!\cdots a_k!}\int_{0}^\infty\!\!\! \cdots\! \int_{0}^\infty x_1^{a_1}\cdots x_{k-1}^{a_{k-1}}x_k^{s+a_k-1}\\
& \qquad \times \frac1{\prod_{l=1}^{k}\cosh(x_l+\cdots+x_k)}\,dx_1\cdots dx_k\\
&=\sum_{a_1+\cdots+a_k=r}\frac{\Gamma(s+a_k)}{\Gamma(s)a_k!}
\times \frac1{\Gamma(a_1+1)\cdots\Gamma(a_{k-1}+1)\Gamma(s+a_k)}\\
&\qquad\times \int_{0}^\infty\!\!\! \cdots\! \int_{0}^\infty
\frac{x_1^{a_1}\cdots x_{k-1}^{a_{k-1}}x_k^{s+a_k-1}}
{\prod_{l=1}^{k}\cosh(x_l+\cdots+x_k)}\,dx_1\cdots dx_k.
\end{align*}
Using Lemma~\ref{Lem-4-1} for the last integral, we obtain
\begin{align}
\label{I1-1} J_1^{(r,k)}(s) &= \sum_{a_1+\cdots+a_k=r}\binom{s+a_k-1}{a_k} \cdot {\widetilde{T}}(a_1+1,\ldots,a_{k-1}+1,a_k+s).
\end{align}
Secondly, by Lemma \ref{defofh},
we compute
\begin{align*}
J_\nu^{(r,k)}(s)
&=\frac{2^r}{\Gamma(s)} \int_{0}^\infty\!\!\! \cdots\! \int_{0}^\infty
\left[ i{\mathscr{A}}(\underbrace{1,\ldots,1}_{r-1},\nu+1;h(x_\nu+\cdots +x_k))
\right]_{x_\nu=0}^\infty\\
&\qquad \times \frac1{\prod_{l=\nu+1}^k \cosh(x_l+\cdots+x_k)}\, x_k^{s-1}\,dx_{\nu+1}\cdots dx_k\\
&=i{\mathscr{A}}(\underbrace{1,\ldots,1}_{r-1},\nu+1;1)\cdot
{\widetilde{T}}(\underbrace{1,\ldots,1}_{k-\nu-1},s) -i\,J_{\nu+1}^{(r,k)}\\
&=i{\mathscr{A}}(\underbrace{1,\ldots,1}_{r-1},\nu+1;1)\cdot
{\widetilde{T}}(\underbrace{1,\ldots,1}_{k-\nu-1},s)\\
&\ \ -i^2{\mathscr{A}}(\underbrace{1,\ldots,1}_{r-1},\nu+2;1)\cdot
{\widetilde{T}}(\underbrace{1,\ldots,1}_{k-\nu-2},s) +i^2\,J_{\nu+2}^{(r,k)}.
\end{align*}
Therefore, repeating this operation, we obtain
\begin{align}
J_1^{(r,k)}(s)&=\sum_{\nu=1}^{k-1}(-1)^{\nu-1}i^\nu {\mathscr{A}}(\underbrace{1,\ldots,1}_{r-1},\nu+1;1)\cdot {\widetilde{T}}(\underbrace{1,\ldots,1}_{k-\nu-1},s)+(-1)^{k-1}i^{k-1}J_{k}^{(r,k)}(s) \notag \\
&=\sum_{j=0}^{k-2}(-1)^{k-j}i^{k-j-1}{\mathscr{A}}(\underbrace{1,\ldots,1}_{r-1},k-j;1){\widetilde{T}}(\underbrace{1,\ldots,1}_{j},s)\notag \\
&\qquad +(-1)^{k-1}i^{k-1}\lambda(\underbrace{1,\ldots,1}_{r-1},k;s),\label{I1-2}
\end{align}
by setting $j=k-\nu-1$ and noting
\[ J_k^{(r,k)}(s)=\lambda(\underbrace{1,\ldots,1}_{r-1},k;s). \]
Comparing \eqref{I1-1} and \eqref{I1-2}, we obtain the assertion.
\end{proof}
\begin{theorem}[cf. \cite{AK1999}\ Theorem 9\ (i);\ \cite{KT2020-ASPM}\ Theorem 5.5]\label{T-5-1}\ For $r,k\in \mathbb{Z}_{\geq 1}$ and $m\in \mathbb{Z}_{\geq 0}$,
\begin{align}
&\lambda(\underbrace{1,\ldots,1}_{r-1},k;m+1)\notag\\
&=i^{1-k}\sum_{a_1,\ldots,a_k\geq 0 \atop a_1+\cdots+a_k=m}\binom{a_k+r}{r}
\cdot {\widetilde{T}}(a_1+1,\ldots,a_{k-1}+1,a_k+r+1). \label{ee-5-4}
\end{align}
\end{theorem}
\begin{proof}
By Lemma \ref{defofh}, we have
\begin{align*}
& \lambda(\underbrace{1,\ldots,1}_{r-1},k;m+1)\\
&=\frac{1}{m!i}\int_0^\infty \frac{t_k^{m}}{\cosh t_k} \int_0^{t_k}
\frac{{\mathscr{A}}(\overbrace{1,\ldots,1}^{r-1},k-1;h(t_{k-1})}{\cosh t_{k-1}}
\, dt_{k-1}dt_k\\
&=\frac{1}{m!i^2}\int_0^\infty \frac{t_k^{m}}{\cosh t_k} \int_0^{t_k}
\frac{1}{\cosh t_{k-1}}\int_0^{t_{k-1}}\\
&\qquad\qquad\frac{{\mathscr{A}}(\overbrace{1,\ldots,1}^{r-1},k-2;h(t_{k-2}/2))}{\cosh t_{k-2}}
\, dt_{k-2}dt_{k-1}dt_k\\
&=\cdots\\
&=\frac{1}{m!i^{k-1}}\int_0^\infty \int_0^{t_k}\cdots \int_0^{t_2}
\frac{t_k^m\, {\mathscr{A}}(\overbrace{1,\ldots,1}^{r};h(t_1))}
{\cosh(t_k)\cdots\cosh(t_1)}\, dt_1\cdots dt_k\\
&=\frac{i^{1-k}}{m!r!}\int_0^\infty \int_0^{t_k}\cdots \int_0^{t_2}
\frac{t_k^m\,t_1^r}
{\cosh(t_k)\cdots\cosh(t_1)}\, dt_1\cdots dt_k.
\end{align*}
By the change of variables
\[t_1=x_k, t_2=x_{k-1}+x_k,\ldots, t_k=x_1+\cdots+x_k,\]
and using Lemma \ref{Lem-4-1},
we obtain
\begin{align*}
&\lambda(\underbrace{1,\ldots,1}_{r-1},k;m+1)\\
&=\frac{i^{1-k}}{m!r!}\int_0^\infty\cdots \int_0^\infty
\frac{(x_1+\cdots+x_k)^m\,x_k^r}
{\prod_{l=1}^k\cosh(x_l+\cdots+x_k)}\, dx_1\cdots dx_k\\
&=i^{1-k}\sum_{a_1+\cdots+a_k=m}\binom{a_k+r}{r}
\cdot {\widetilde{T}}(a_1+1,\ldots,a_{k-1}+1,a_k+r+1).
\end{align*}
\end{proof}
Setting $s=m+1$ in Theorem~\ref{Th-4-2} and comparing with Theorem~\ref{T-5-1},
we obtain the following\footnote{Ce Xu has kindly
informed us of their recent paper \cite{XZ} with Jianqiang Zhao, where they study more general
level 4 multiple $L$-values and obtain a similar theorem (\cite[Theorem~4.4]{XZ}) in their
setting, of which the following is a special case.} which corresponds to \cite[Corollary 11]{AK1999}.
\begin{theorem}[cf. \cite{AK1999}\ Corollary 11;\ \cite{KT2020-ASPM}\ Theorem 5.7] \label{Th-4-7}\ For $m,r\geq 1$ and $k\ge2$,
\begin{align}
&\sum_{a_1,\ldots,a_k\geq 0 \atop a_1+\cdots+a_k=m}\binom{a_k+r}{r}\cdot {\widetilde{T}}(a_1+1,\ldots,a_{k-1}+1,a_k+r+1) \notag\\
& \ +(-1)^k \sum_{a_1,\ldots,a_k\geq 0 \atop a_1+\cdots+a_k=r}\binom{a_k+m}{m}\cdot {\widetilde{T}}(a_1+1,\ldots,a_{k-1}+1,a_k+m+1)\notag\\
& =\sum_{j=0}^{k-2}(-1)^j {\widetilde{T}}(\underbrace{1,\ldots,1}_{k-j-2},r+1)\cdot {\widetilde{T}}(\underbrace{1,\ldots,1}_{j},m+1). \label{eq-Th-4-7}
\end{align}
\end{theorem}
{\bf Acknowledgements.}\
{The authors are very grateful to Minoru Hirose, who provided a Pari-GP program to numerically compute ${\widetilde{T}}$-values efficiently.
They also express their gratitude to him and Ryota Umezawa for pointing out the proof of Theorem~\ref{Teven}.
This work was supported by JSPS KAKENHI Grant Numbers JP16H06336, JP21H04430 (Kaneko), and JP21K03168 (Tsumura).}
|
\section{Introduction}
The main aim of quantum metrology is to understand how utilizing quantum effects can enhance the estimation precision of a parameter beyond the classical limits \cite{intro_met1,intro_met2,intro_met3,intro_met4}. The latter is given by the shot-noise limit (SNL), which restricts the estimation precision to scale as $\Delta \theta \sim (N\nu)^{-\frac{1}{2}}$, where $\theta$ is a parameter one wishes to measure precisely, $N$ is the number of particles being measured, and $\nu$ is the number of measurement repetitions. Making use of entangled states, one can enhance the precision by a factor of $k^{-\frac{1}{2}}$, where $k$ is the number of entangled particles \cite{useful_entanglement} and in the limit of a maximally entangled many-body state, i.e.~when $k=N$, one can achieve an estimation precision which scales as $N^{-1} \nu^{-\frac{1}{2}}$ for a decoherence-free system. This is the so-called \textit{Heisenberg scaling}, which provides an improvement of $\sim N^{-\frac{1}{2}}$ over the SNL \cite{H_limit}. Metrological enhancement over the SNL has been demonstrated by utilizing entangled states such as squeezed states \cite{squeezing1,squeezing2}, \textit{NOON} states \cite{NOON_1}, and others \cite{H_limit,GHZ,GHZ_met}, and has been implemented using a variety of setups such as squeezed states interferometers \cite{squeezed_interfero, squeezed_interfero2}, cavity QED~\cite{gietka2019,coherenceqedmet_2019,atomscavitymet_2020}, ion traps~\cite{Dalvitions_2006,metrologyiontrap_2018,amrscienceiontrap_2021}, and distributed quantum networks~ \cite{distributed_qmet1,distributed_qmet2,distributed_qmet3,zhou2022}. In distributed quantum networks, a single-mode photonic input state is fed to a global beam splitter network, which consists of $M$ modes. The output state is then correlated between all of the modes before the parameter is imprinted and then measured at each mode. In this protocol, the parameter estimation has been shown to scale as $M^{-1}$, which corresponds to Heisenberg-like scaling for the system. This demonstrates that aside from the scaling with the number of particles, one can also utilize scaling with the number of modes to further enhance the parameter estimation.
In standard setups for distributed quantum networks for photons, multi-mode correlations can be readily created through the use of beam splitters. However, in recent years precise control of atomic systems has become possible as well through the advent of atom cooling, trapping, and engineering techniques \cite{laser_cool_trap,evap_cool,laser_cool_trap_2}. Despite this, atomic systems have received relatively little attention in the context of distributed quantum networks~\cite{kasevich_2022} and in this work we want to bridge that gap and explore the metrological applications of multi-mode cold atomic systems by investigating how one can enhance the estimation precision by manipulating the controllable parameters of atoms trapped in an optical lattice.
\section{Heisenberg limit in multi-mode, multi-particle atomic systems}
The ultimate bound to the estimation precision for a parameter $\theta$ over all possible measurements is given by the quantum Cram\'er-Rao bound, $\Delta\theta \ge \Delta\theta_{\text{QCR}} = 1/\sqrt{\nu F(\theta)}$, where $F(\theta)$ is the quantum Fisher information \cite{FI,QCRB1}. Given a Hamiltonian $H_\theta$ that depends on the parameter $\theta$ and evolves in time according to $U_\theta = \exp{(-iH_\theta t)}$, one can determine the quantum Fisher information by introducing a local generator $\hat{h}_\theta = i(\partial_\theta U_\theta)U_\theta^\dagger$, which characterizes the sensitivity of a state $\rho_\theta$ to an infinitesimal change in $\theta$, from $\rho_\theta \rightarrow \rho_{\theta + d\theta}$. For a system with an initial state that is a pure state $|\psi\rangle$, the quantum Fisher information can then be written as $F(\theta) = 4\langle\psi|\Delta^2 \hat{h}_\theta |\psi\rangle$ \cite{FI, FI2} and it is maximized when the initial state is given by an optimal state of the form $|\psi_{\text{opt}}\rangle = \frac{1}{\sqrt{2}}(|h_{\text{max}}\rangle + |h_{\text{min}}\rangle)$. In this case one gets
\begin{equation}
F_{\text{max}}(\theta) = (h_{\text{max}} - h_{\text{min}})^2,
\label{eq:FI_max}
\end{equation}
\noindent
where $h_{\text{max}}$ and $h_{\text{min}}$ are the maximal and minimal eigenvalues of $\hat{h}_\theta$ associated with states $|h_{\mathrm{max}}\rangle$ and $|h_\mathrm{min}\rangle$, respectively~\cite{intro_met4}. In the case where the Hamiltonian is composed of non-commuting terms, a compact expression for the local generator may not be available. An alternative expression to the local generator can be written in terms of the eigenvalues $E_k$ and eigenvectors $|\phi_k \rangle$ of the Hamiltonian \cite{Q_met_general_param} given as
\begin{align}
\hat{h}_\theta &= \hat{h}_\theta^{(L)} + \hat{h}_\theta^{(O)}, \label{eq:generator}\\
\hat{h}_\theta^{(L)} &= t\sum_{k=1}^{n_s}\frac{\partial E_k}{\partial \theta}|\phi_k \rangle \langle \phi_k| \label{eq:generator_L}, \\
\hat{h}_\theta^{(O)} &= 2\sum_{l \neq k}e^{-\frac{itE_{kl}}{2}}\sin\left(\frac{tE_{kl}}{2}\right)\langle \phi_l|\partial_\theta\phi_k \rangle|\phi_k \rangle \langle \phi_l|, \label{eq:generator_O}
\end{align}
where $E_{kl} = E_{k} - E_l$ and $n_s$ is the total number of states, while $\hat{h}_\theta^{(L)}$ and $\hat{h}_\theta^{(O)}$ are the linear and oscillating parts of the local generator, respectively. Note that we use the convention $\hbar = 1$ throughout the manuscript.
To describe a multi-particle, multi-mode atomic system let us first consider a general Hamiltonian of the form
\begin{equation}
H = \gamma \sum_m m \hat{a}^\dagger_m \hat{a}_m \label{eq:gen_Ham},
\end{equation}
where $\gamma$ is the parameter we wish to measure precisely, $\hat{a}^\dagger_m$ and $\hat{a}_m$ are creation and annihilation operators, and the mode-label $m$ runs from $1$ to $M$. The quantum Fisher information for a system described by Hamiltonian~\eqref{eq:gen_Ham} is maximized by a state
\begin{equation}
|\psi_{\text{opt}}\rangle = \frac{1}{\sqrt{2}}(|N0...0\rangle + |0...0N\rangle),
\label{eq:optimum_state}
\end{equation}
which is a superposition of all the atoms occupying the first and the last site (for the sake of brevity we will call this a generalized $NOON$ state), and can be calculated to be
\begin{equation}
F_{\text{max}}(\gamma) = T^2(N(M-1))^2 = F_{\text{HL}} \label{eq:HS},
\end{equation}
where $T$ is the length of the time interval during which the information about the unknown parameter was being imprinted. This will be our definition of the Heisenberg limit throughout this work. In particular, we will consider a 1D lattice system in a uniform linear potential, where $M$ is the total number of lattice sites, such that the enhancement proportional to $(M-1)$ is reminiscent to the scaling obtained from distributed photonic networks where $M$ refers to the number of modes. However, in these photonic systems, the maximum quantum Fisher information only scales with $\sim \bar{n}M^2$, where $\bar{n}\equiv N/M$ is the average photon number per mode \cite{distributed_qmet1}. Thus, once the Fisher information is expressed in terms of total amount of photons, the quadratic dependence on the number of modes disappears.
In the following we propose a possible realization of such a system using cold atoms and explore different strategies that can take advantage of the presence of multiple modes.
\begin{figure}[tb]
\centering
\includegraphics[width=\linewidth]{img/schematic_new1.pdf}
\caption{(top) Schematic of the TBH model. (bottom) Schematic of the driven TBH model in the effective time-independent picture as given in Eq.~\eqref{eq:DBH}. All parameters in Eq.~\eqref{eq:DBH}
are scaled with respect to the tunneling coefficient, $J$, and are fixed for all simulations (except for $U$) with values: $J = 1$, $\gamma=33J$, $V_0 = 30.4J$, $\theta = \pi$, $\phi_{m+1} = \phi_m-\pi$, $\phi_0 = -\pi/2$.
}
\label{fig:FINOON}
\end{figure}
\section{Tilted Bose-Hubbard Model}
A Hamiltonian with a term given in Eq.~\eqref{eq:gen_Ham} can be realised in a one-dimensional lattice system that is exposed to a uniform linear potential. Such a system is also known as a tilted Bose-Hubbard (TBH) model and can be described by
\begin{equation}
H_{\text{TBH}} = -J\sum_{\langle i,j \rangle} \hat{a}^\dagger_{i} \hat{a}_j + \gamma \sum_j j \hat{n}_j + \frac{U}{2}\sum_j \hat{n}_j(\hat{n}_j - 1)
\label{eq:TBH},
\end{equation}
where $J$ is the tunneling coefficient, $\langle\bullet\rangle$ denotes nearest neighbor sites, $U$ quantifies the on-site interaction strength between the particles, $\gamma$ is the strength of the tilt, and $\hat{n}_m = \hat{a}^\dagger_m \hat{a}_m$ is the number operator. A schematic of the system is shown in Fig.~\ref{fig:FINOON}(top) and we assume that we only have a finite number of sites, $j=1\dots M$. Although the TBH Hamiltonian contains additional terms when compared to Eq.~\eqref{eq:gen_Ham}, it is known that terms which are not dependent on the unknown parameter cannot alter the maximum attainable precision \cite{disturbance}. Thus, the Bose-Hubbard system in Eq.~\eqref{eq:TBH} will have a maximum quantum Fisher information that is still bounded by Eq.~\eqref{eq:HS}.
The Hamiltonian in Eq.~\eqref{eq:TBH} can be realized using ultracold atoms trapped in a 1D optical lattice under the influence of a linear potential of strength $\gamma$, which, for example, can be gravitational or magnetic in nature \cite{H_H,harper_hofstadter2,gravimeter}. Therefore, a precise measurement of the parameter $\gamma$ would correspond to making a precise measurement of the field. The interaction strength $U$ can be tuned by employing Feshbach resonances \cite{Feshbach_M,Feshbach_O}, and is designed to be small relative to $\gamma$ such that excitations to higher bands are suppressed.
Assuming one can prepare the generalized $NOON$ state in Eq.~\eqref{eq:optimum_state} at $t=0$, the interaction can then be set to $U=0$, and the lattice depth to a value such that $J \ll \gamma $, which freezes the spatial dynamics. The remaining dynamics is therefore purely in the phase difference between the two states in the superposition and is of the form $|\psi_{\text{opt}}(t)\rangle \sim \frac{1}{\sqrt{2}}(|N0...0\rangle + e^{-it\gamma N(M-1)}|0...0N\rangle)$ which is a state that yields $F_{\text{HL}}$ at all times. In this case all the information about the unknown parameter is stored in the relative phase between the two components of the wavefunction. It could be retrieved, for example, by performing an interference measurement, i.e., transferring the information stored in the relative phase to the occupation of each sites, which might be experimentally challenging. Furthermore, \textit{NOON} states are known to be very fragile against losses and thus difficult to prepare. Although several approaches have been suggested in the literature, especially for two-site systems, they usually suffer from having low fidelity as the number of particles is increased \cite{Weiss:09,Fogarty:13,NOON_state_repulse,Naldesi:22}.
We therefore explore in the following the prospects for metrology with a TBH model using a more realistic initial state.
\begin{figure}[tb]
\centering
\includegraphics[width=\linewidth]{img/fig_2.pdf}
\put(-230,125){\textcolor{black}{(a)}}
\put(-100,125){\textcolor{black}{(b)}}
\caption{(a) Time dependence of the scaled $F^{(M)}$ for $M=\{2,3,4\}$ corresponding to solid lines $\{$blue, red, orange$\}$, respectively (the unscaled $F^{(M)}$ is shown in the inset). (b) Growth of $F^{(M)}_{\text{max}}$ scaled by $F^{(M=2)}_{\text{max}}$ with $M$. The inset shows the linear scaling of $\tau$ with respect to $M$ (see text for details).}
\label{fig:FIFock}
\end{figure}
\section{Fisher information with an initial Fock state}
Let us start by considering an initial state where the particles are all placed in the lowest energy site of the lattice
\begin{equation}
|\psi_{\text{Fock}}\rangle = |N0...0\rangle.
\end{equation}
\noindent
We first investigate the case where interactions are switched off and show later that when $U>0$, additional improvement to the quantum Fisher information can be observed due to the correlations introduced by the interactions.
Given that $\gamma \gg J$, the spatial dynamics is frozen as before and since $4\langle\psi_{\text{Fock}}|\Delta^2 \hat{h}_\theta |\psi_{\text{Fock}}\rangle=0$, the Fisher information will be fixed to this value unless one restores the spatial dynamics. To introduce atomic dynamics to the system that depends on $\gamma$, we therefore consider a periodic drive with frequency $\omega = \gamma$, which involves the knowledge about the unknown parameter.
Such an approach would require, for example, an adaptive protocol where the knowledge about $\omega$ would be updated with every round of the protocol~\cite{wiseman_milburn_2009,2011Adaptivemetrology,2017Adaptivemetrologymarkovian,PRXQuantum.2.040301}. The driven Hamiltonian can then be written as
\begin{equation}
H_{\text{DBH}} = H_{\text{TBH}} + V_0\sum_m \hat{n}_m \sin{\left(\omega t + \phi_m + \frac{\theta}{2}\right) \label{eq:DBH}},
\end{equation}
where $V_0$ is the driving amplitude, $\phi$ is a site-dependent phase, and $\theta$ is a constant phase. A schematic of the driven system in the effective time-independent picture is shown in Fig.~\ref{fig:FINOON}(bottom). These kinds of driving terms can be experimentally realized by an off-resonant laser-assisted tunneling scheme~\cite{harper_hofstadter1, harper_hofstadter2}. In order to suppress decoherence due to particle loss into the higher bands, we focus on the limit $\gamma \gg U$~\cite{Floquet_general,FT_bukov}
and compute the resulting quantum Fisher information \cite{FI_eq2} of the system in Eq.~\eqref{eq:DBH} using
\begin{equation}
F(\gamma) = 4\left(\langle\partial_\gamma\psi_{\gamma}|\partial_\gamma\psi_{\gamma}\rangle - \left| \langle\psi_{\gamma}|\partial_\gamma\psi_{\gamma}\rangle \right|^2 \right).
\label{eq:FI_time}
\end{equation}
\noindent
Fig.~\ref{fig:FIFock}(a) shows the evolution of the quantum Fisher information $F^{(M)}$ using the initial state $|\psi_{\text{Fock}}\rangle$ in the non-interacting driven TBH model, where $M$ denotes the total number of lattice sites. Here, $F^{(M)}$ is scaled by $T^2$ and the number of particles is set to $N=1$ as it only has a linear contribution to the quantum Fisher information in the case of $U=0$, that is, $F^{(M,N)} = NF^{(M)}$. The inset of Fig.~\ref{fig:FIFock}(a) shows the oscillating behavior of the unscaled $F^{(M)}$ with $T$, and thus in the vicinity of the first peak, the $F^{(M)}$ starts to scale poorly compared to $T^2$.
This means that we only need to consider the dynamics up until the first peak of $F^{(M)}/T^2$ and we denote this peak as $F^{(M)}_{\text{max}} = \text{max}(F^{(M)}/T^2)$. Defining the time at which $F^{(M)}_{\text{max}} $ is attained as $\tau$, one can see in the inset of Fig.~\ref{fig:FIFock}(b) that it has a linear dependence on $M$, since more modes increases the time after which the state is transferred to the other end of the lattice. The quantum Fisher information enhancement relative to a two-level system, $F^{(M)}_{\text{max}}/F^{(M=2)}_{\text{max}}$, is shown in Fig.~\ref{fig:FIFock}(b) and clearly shows a quadratic dependence on $M$ for larger $M$. This illustrates that one can indeed make use of a larger system size in order to increase the quantum Fisher information even without introducing non-classical correlations.
\begin{figure*}[tb]
\centering
\includegraphics[width=\textwidth]{img/fock_plots.pdf}
\put(-510,285){\textcolor{black}{(a)}}
\put(-387,285){\textcolor{black}{(b)}}
\put(-257,285){\textcolor{black}{(c)}}
\put(-129,285){\textcolor{black}{(d)}}
\put(-510,135){\textcolor{black}{(e)}}
\put(-387,135){\textcolor{black}{(f)}}
\put(-257,135){\textcolor{black}{(g)}}
\put(-129,135){\textcolor{black}{(h)}}
\caption{(a) Density plot of the scaled $F^{(M,N)}$ as a function of $T$ and $U$ for $N=M=3$. (b) Growth of $F^{(M,N)}_{\text{max}}$ with $M$ for $N=\{3,4,5,6,7\}$. The legend applies for plots (b)-(d). (c) $F^{(M,N)}_{\text{max}}$ scaled by $F^{(M=2,N)}_{\text{max}}$ and (d) scaled by the Heisenberg limit. Shown in (c) and (d) are all the data points that we were able to numerically obtain. (e) and (f) show the variance of $\hat{h}_\gamma^{(L)}$ and $\hat{h}_\gamma^{(O)}$, respectively. In (g) the correlation function, $\mathcal{G}^{(N)}$, is calculated for varying $U$ and $T$ and (h) shows the variance of $\hat{h}_\gamma^{(O)}$ overlaid by the correlation function. (See text for details).}
\label{fig:FIFock2}
\end{figure*}
\section{Effect of interactions}
Next, we will consider an interacting system in which the additional non-classical correlations can be created that should improve the quantum Fisher information. The initial state is the same as in the previous section and we imprint the information about the unknown parameter and create the correlations at the same time~\cite{Hayes_2018,2021understandingCQM}. A representative surface plot of the evolution of $F^{(M,N)}/T^2$ with varying interaction strength $U$ for $N=M=3$ is shown in Fig.~\ref{fig:FIFock2}(a). Similar to the non-interacting case, we also observe a sinusoidal-like evolution of $F^{(M,N)}$ therefore we limit our interest up to the time $\tau$ when the first peak of $F^{(M,N)}/T^2$ appears. What is notable, however, is that a certain value of the interaction strength exists, which we will call $\bar{U}$, where the increase in $F^{(M,N)}$ over the non-interacting case is maximal. For the parameters used in Fig.~\ref{fig:FIFock2}(a) this corresponds to $\bar{U}\approx1.92J$ and in general we observe an increase of $F^{(M,N)}_{\text{max}}$ as $N$ and $M$ increase (see Fig.~\ref{fig:FIFock2}(b)). When compared to a two-level system $F^{(M=2,N)}_{\text{max}}$, we still observe an enhancement as $M$ increases, however, increasing the number of particles may not always yield a larger enhancement as shown in Fig.~\ref{fig:FIFock2}(c). Finally, when compared to the Heisenberg limit, the two-level system yields a stronger enhancement over any $M>2$ (for small $M$) and we again observe that the increase in $N$ does not always provide a larger enhancement as can be seen in Fig.~\ref{fig:FIFock2}(d). While it might look like that for increasing number of modes $F^{(M,N)}_{\text{max}}/F_{HL}\rightarrow \sim 0.2$, our limited computational resources currently do not allow us to explore this
To better understand the behavior of the quantum Fisher information $F^{(M,N)}$, we take the high-frequency approximation \cite{FT_dalibard_goldman,FT_dalibard_goldman_reso,Floquet_general,FT_bukov} of the driven, tilted Bose-Hubbard model up to first-order in $1/\omega$, which leads to an effective, time-independent description of the system given as,
\begin{align}
H_{\text{eff}} =& -J_{F}\sum_{j} (\hat{a}^\dagger_{j+1} \hat{a}_j e^{-i\phi_j} + h.c.) + \frac{U}{2}\sum_j \hat{n}_j(\hat{n}_j - 1)\nonumber \\
& + K\left(\frac{1}{\omega}\right)\sum_j \left(\hat{n}_{j+1} - \hat{n}_j \right) + \mathcal{O}\left( \frac{1}{\omega^2} \right)
\label{eq:H_eff}
\end{align}
\noindent
where $J_F=J\mathcal{J}_1(2V/\omega)$ is the renormalized tunneling coefficient and $\mathcal{J}_1(x)$ is a Bessel function of the first kind. Since the effective Hamiltonian in Eq.~\eqref{eq:H_eff} is time-independent, the Fisher information can now be calculated from the local generator $\hat{h}_\gamma$ using Eq.~\eqref{eq:generator} and we look at the contributions from the linear part $\hat{h}_\gamma^{(L)}$ and the oscillating part $\hat{h}_\gamma^{(O)}$ separately. The respective variances are plotted in Figs.~\ref{fig:FIFock2}(e) and (f) and one can see that $\langle(\Delta\hat{h}_\gamma^{(L)})^2\rangle$ dominates over $\langle(\Delta\hat{h}_\gamma^{(O)})^2\rangle$ at long times due to its quadratic dependence on time. The sinusoidal-like behavior that is seen in the inset of Fig.~\ref{fig:FIFock}(a) originates from $\langle(\Delta\hat{h}_\gamma^{(O)})^2\rangle$ and $\tau$ sets the timescale for the appearance of the first maximum of this term. The dependence of these generators to the eigenstates of $H_{\text{eff}}$ is discussed in the Appendix.
Finally, to illustrate that this enhancement of the quantum Fisher information is related to an increase in non-classical correlations, we make use of the $N$-th order correlation function given as \cite{NOON_state_correlator}
\begin{equation}
\mathcal{G}^{(N)} = \left| \frac{1}{C}\langle \psi(t)| \left(\hat{a}^\dagger_M \hat{a}_1 \right)^N | \psi(t) \rangle \right|,
\label{eq:correltion_function}
\end{equation}
\noindent
where $C = N!/2$ is a normalization constant. $\mathcal{G}^{(N)}$ quantifies the $N$-particle correlation between the two outermost sites, $j=1,M$. One can show that for a generalized $NOON$ state, $|\psi_{\text{opt}}(t)\rangle$, the correlation function is $\mathcal{G}^{(N)} = 1$ at all times. On the other hand, for an initial Fock state, $|\psi_{\text{Fock}}\rangle$, this is not the case. A surface plot of $\mathcal{G}^{(N)}$ as a function of $(T,U)$ is shown in Fig~\ref{fig:FIFock2}(g) for $|\psi_{\text{Fock}}\rangle$ and $M=N=3$. One can see that $\mathcal{G}^{(N)}$ has the same qualitative features as $\langle(\Delta\hat{h}_\gamma^{(O)})^2\rangle$ and by superimposing $\langle(\Delta\hat{h}_\gamma^{(O)})^2\rangle$ on $\mathcal{G}^{(N)}$ (see Fig.~\ref{fig:FIFock2}(h)) one can see that the maxima of $\langle(\Delta\hat{h}_\gamma^{(O)})^2\rangle$ lies in the region of large $\mathcal{G}^{(N)}$. This comparison however does not always hold true as can be seen for small $U$, when the correlator $\mathcal{G}^{(N)}$ suggests that there can be large correlation, but the quantum Fisher information is relatively small. However, we have confirmed for all combinations of $N=\{3,4,5\}$ and $M=\{3,4\}$ the regions of large $\langle(\Delta\hat{h}_\gamma^{(O)})^2\rangle$ also correspond to large correlations (not shown). As one goes to larger $N$ for a fixed $M$, the maximum correlation no longer approaches unity which could explain why the scaled $F^{(M,N)}_{\text{max}}$ in Fig~\ref{fig:FIFock2}(c)-(d) does not always increase even if $N$ is increased for a fixed $M$.
\section{Conclusions}
In this work, we have investigated the use of a multi-mode atomic system in the context of distributed quantum sensing. We have shown that the driven, tilted Bose-Hubbard model can make use of the additional degree of freedom of the number of lattice sites or number of modes, $M$, in order to increase the quantum Fisher information of the system with respect to some unknown parameter. A generalized $NOON$ state maximizes the quantum Fisher information at all times when the spatial dynamics is frozen, however, extracting the information from such a state might require a complicated measurement procedure. An initial Fock state cannot saturate the Heisenberg limit but can still benefit from the quadratic scaling in $M$. In this case, the occupation of each site can be used as an optimal estimator. This quadratic scaling was made possible through the introduction of the periodic drive which translates the information about the parameter into the tunneling dynamics of the particles. On the other hand, for a distributed photonic network its quadratic scaling in $M$ becomes linear once the quantum Fisher information is expressed in terms of the total number of particles. Additionally, by introducing interactions to the system we have shown that parameter imprinting and creation of correlations can be achieved simultaneously in contrast to the distributed photonic networks where the creation of the correlated state is performed before the parameter imprinting.
We emphasize that the enhancement with respect to the number of modes is not an unexpected result. Suppose we have a harmonic oscillator where the energy spacing is $\omega_0$, and suppose further that there is a non-linear process that can drive the system from the ground state to the $M$-th level with energy $\omega_M = M\omega_0$, then using the error propagation formula we get that the uncertainty in $\omega_0$ is given by $\Delta\omega_0 = \Delta\omega_M / M$. This is the same enhancement we observe for the tilted, Bose-Hubbard model using the generalized $NOON$ state. The only difference here is that no non-linear process is used to couple neighboring modes and thus demonstrates the advantage of utilizing a ladder-like system with multiple modes in metrology. It is clear that the existence of the tilt (or any dispersion which is of the form $~m^\alpha$, where $\alpha \ge 1$ and $m$ is the mode index) is what allows the enhancement with the number of modes, therefore, other ladder-like systems can also exploit this enhancement. An example of such a system is the periodically forced Bose-Hubbard model \cite{periodically_forced} given as, $H_{\text{PF}} = H_0 + \Gamma\cos{(\omega t) \sum_j j\hat{n}_j} $ where the difference between this and the tilted Bose-Hubbard model is that the tilt is no longer linear but instead it is periodic in time. What is interesting here is that if we precisely measure $\omega$ instead of $\Gamma$, we find that the Heisenberg limit is $(\Gamma NM)^2T^4$, where we still have the enhancement in $N$ and $M$ but now the scaling with time is super-quadratic and can be achieved only if one incorporates an additional optimal control Hamiltonian \cite{FI_eq2}. These ladder-like systems provide an avenue to study quantum metrological setups which have a potential to further increase precision measurements.
\section{Acknowledgements}
The authors would like to acknowledge Lewis Ruks for fruitful discussions.
This work was supported by the Okinawa Institute of Science and Technology Graduate University (OIST). The authors are also grateful for the the Scientific Computing and Data Analysis (SCDA) section of the Research Support Division at OIST.
\noindent
|
\section{Introduction}
We seek a theory of quantum gravity that can describe our own universe. One possible approach to this challenge is to analyze the well-known theories~\cite{Maldacena:1997re} that instead describe asymptotically anti-de Sitter space (AdS). In particular, one would like to identify elements of the AdS/CFT correspondence that have broader validity. The entanglement wedge prescription is a crucial part of the AdS/CFT dictionary: it identifies the spacetime region dual to a given CFT subregion. We will propose a more general prescription that associates an entanglement wedge to arbitrary regions in arbitrary spacetimes.
\subsection{Entanglement Wedge of a Boundary Region}
The Ryu-Takayanagi (RT) proposal~\cite{Ryu:2006bv} was first introduced in the context of the AdS/CFT correspondence, as a method for computing the von Neumann entropy of a stationary CFT state reduced to any spatial region $A$ of the boundary. The proposal was refined to allow for time-dependent states by Hubeny, Rangamani, and Takayanagi~\cite{Hubeny:2007xt}. Quantum corrections were first included by Faulkner, Lewkowycz, and Maldacena~\cite{Faulkner:2013ana}. Engelhardt and Wall~\cite{Engelhardt:2014gca} introduced the present, most powerful formulation of RT in terms of the Quantum Extremal Surface QES$_{\rm min}(A)$ anchored on $A$ with smallest generalized entropy among all such surfaces. This prescription can be applied in the semiclassical regime to any order in $G\hbar$.
The {\em entanglement wedge}, $\mathrm{EW}(A)$, is defined as the homology wedge, \emph{i.e.}, the causal development of any spatial region bounded by $\mbox{QES}_{\rm min}(A)$ and $A$. There is strong evidence that in a large class of states,\footnote{See~\cite{Hayden:2018khn,Akers:2020pmf} for subtleties and appropriate refinements of the entanglement wedge. Our proposal admits analogous refinements, which may be important for its generalization to time-dependent settings; see Sec.~\ref{sec-covariantdiscussion}.} $\mathrm{EW}(A)$ is precisely the bulk dual to the boundary region $A$~\cite{Wall:2012uf,Jafferis:2015del}.\footnote{See \cite{Bousso:2012sj,Hubeny:2012wa,Czech:2012bh} for early work on this problem.} That is, any local bulk operator in $\mathrm{EW}(A)$ can be implemented, or ``reconstructed,'' by a CFT operator in the algebra associated to the region $A$; and no bulk operator outside $\mathrm{EW}(A)$ can be so implemented. This result is known as entanglement wedge reconstruction.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=.4]{Island1.pdf}
\end{center}
\caption{Entanglement island. Hawking radiation emitted from a black hole (grey circle) is collected in the reservoir $a$ (red) located near the boundary region $A$. Before the Page time, the (ordinary) entanglement wedge $EW(A)$ consists only of the blue region adjacent to $A$. After the Page time, $EW(A)$ also contains an island $b$ in the black hole interior.}
\label{fig-island}
\end{figure}
Its definition does not require the entanglement wedge to be connected. Any component of $\mathrm{EW}(A)$ that is not connected to $A$ is called an {\em entanglement island}~\cite{Penington:2019npb,Almheiri:2019psf}. An island arises naturally in the course of black hole evaporation, if the radiation is stored in a distant, localized reservoir $a$~\cite{Bousso:2019ykv}. Consider a small boundary region $A$ whose entanglement wedge contains $a$, as shown in Fig.~\ref{fig-island}. After the Page time, \emph{i.e.}, when the entropy of the radiation exceeds the Bekenstein-Hawking entropy of the remaining black hole, $\mathrm{EW}(A)$ will also include an island $b$ just inside of the black hole horizon. The inclusion of $b$ in $\mathrm{EW}(A)$ decreases the generalized entropy since the radiation entropy in $a$ is purified by the Hawking partners in $b$, and the price paid is only the Bekenstein-Hawking entropy of the island boundary (roughly that of the black hole).
\subsection{Going Beyond AdS}
\label{sec-beyond}
With certain assumptions, the entanglement wedge prescription can be derived from the gravitational path integral~\cite{Lewkowycz:2013nqa}. This implies that the entanglement wedge prescription can be used to compute the entropy not only of subregions of the conformal boundary of AdS, but of more general systems that are coupled to a gravitating spacetime, regardless of whether an exact description is known.\footnote{The earliest example of this phenomenon was the observation that, if a nongravitating external system $\mathsf{R}$ purifies matter near the QES that bounds $\mathrm{EW}(A)$, then access to $\mathsf{R}$ can increase the size of the reconstructible bulk region \cite{Hayden:2018khn, Akers:2019wxj}. In other words, $\mathrm{EW}(A\cup \mathsf{R})$ can differ from $\mathrm{EW}(A)$.}
This is a crucial point: the entanglement wedge prescription is not confined to the context of AdS/CFT. Like the Generalized Second Law~\cite{Bekenstein:1972tm,Wall:2011hj} and the Covariant Entropy Bound~\cite{Bousso:1999xy,Bousso:2015mna}, the entanglement wedge prescription is applicable in general spacetimes. It reveals aspects of the (usually unknown) full quantum gravity theory, extracted from the gravitational path integral.\footnote{When a full theory \emph{is} available, it can be used to verify the predictions of the entanglement wedge prescription. For example, by the AdS/CFT duality, the Page curve for an evaporating AdS black hole follows independently from the scrambling properties and the unitarity of the CFT.}
The full generality of the entanglement wedge prescription --- its applicability beyond the AdS/CFT correspondence --- was long obscured by the lack of nontrivial bound\-ary-anchored extremal surfaces in non-AdS spacetimes or auxiliary systems. Entanglement islands, however, are detached from the conformal boundary by definition. Hence, islands furnish nontrivial examples of entanglement wedges of systems that have nothing to do with the conformal boundary of AdS.
\begin{figure}[t]
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width = 0.97\linewidth]{Island2}
\end{subfigure}
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width = 0.95\linewidth]{AsymptoticEW}
\end{subfigure}
\caption{Examples that motivate and constrain our definition of a \emph{generalized entanglement wedge}, a map $E: a\to E(a)$, that takes bulk regions as its input. \emph{Left:} Hawking radiation arrives on a portion $A$ of future null infinity. After the Page time, the (original) entanglement wedge of $A$ includes an island: $\mathrm{EW}(A)=A\cup b$. When the radiation is still in the distant spacetime region $a$ (green shaded), it should already possess a generalized entanglement wedge, such that $E(a)\supset a\cup b$. \emph{Right:} the (original) entanglement wedge $\mathrm{EW}(A)$ of an AdS boundary region $A$ can be regarded as the bulk algebra of operators encoded in the CFT algebra of operators in $A$. The latter is generated by local operators in $A$, which are dual to quasi-local bulk operators in the near-boundary region $a$. Hence $a$ should possess a generalized entanglement wedge such that $E(a)=\mathrm{EW}(A)$.}
\label{fig-motivate}
\end{figure}
For example, if the Hawking radiation is transferred to an external system $\mathsf{R}$, then after the Page time, $\mathsf{R}$ will possess a nontrivial entanglement wedge: $\mathrm{EW}(\mathsf{R}) = \mathsf{R} \cup b$. (Islands were first discovered in this setting.) This should also be the case for Hawking radiation far from the black hole in asymptotically flat spacetimes (see Fig.~\ref{fig-motivate}, left panel). In both cases, the presence of an entanglement island $b$ is crucial to deriving the Page curve and thus the unitarity of the Hawking process. And in neither case is the entanglement wedge computed for a portion of the conformal boundary of AdS.
What should be regarded as the proper input for computing an entanglement wedge: a portion of the conformal boundary? Nongravitating systems outside the spacetime? In this paper, we will propose that the answer is neither: rather, the entanglement wedge prescription should be formulated as a map that takes any gravitating spacetime region as input, and outputs another (equal or larger) bulk spacetime region. The entanglement wedges of nongravitating systems and of boundary regions should be interpreted as limits of this more general prescription that arise if the input region is weakly gravitating or approaches the asymptotic boundary.
Because the real world has gravity, such a proposal might be testable at least in principle, unlike a boundary-to-bulk map. Importantly, it should extend the notion of an entanglement wedge to arbitrary spacetimes, including cosmology.
For static spacetimes, we propose that the generalized entanglement wedge $E(a)$ is simply the smallest generalized entropy region such that $a \subset E(a)$.\footnote{Throughout, the symbols $\subset$ and $\supset$ are understood to allow set equality.} We motivate this prescription using tensor network toy models of quantum gravity and show that it satisfies a number of desirable properties that suggest it correctly quantifies the range of holographic encoding.
\begin{figure}[t]
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width = 0.6\linewidth]{staticEW}
\end{subfigure}
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width = 0.75\linewidth]{tensorEW}
\end{subfigure}
\caption{\textit{Left:} in this example, the edge of $a$ is a nonconvex surface on a static slice of Minkowski space, and the edge of $E(a)$ is its convex hull. \textit{Right:} a tensor network analogue of the region $a$. For random tensor networks, the map from the boundary of $E(a)$ (blue) to the boundary of $a$ (green) will be an approximate isometry.}
\label{fig:introfig}
\end{figure}
For general regions in time-dependent spacetimes, the correct definition of the generalized entanglement wedge seems less clear. We discuss one possible definition in some detail and show that it satisfies some but not all of the properties satisfied by $E$.
\subsection{Outline}
In Sec.~\ref{sec-motivation} we motivate our prescription. We argue in Sec.~\ref{sec-why} that it must be possible to assign a larger entanglement wedge at least to some gravitating regions. Universality then suggests that all gravitating regions have an entanglement wedge. In Sec.~\ref{sec-tensor} we analyze the example of tensor network toy models and argue that a sensible entanglement wedge may be assigned to bulk regions.
In Sec.~\ref{sec-prescription}, we define the entanglement wedge $E(a)$ of a bulk region $a$. We focus on the ``static'' case, when $a$ and $E(a)$ are part of time-reversal invariant initial data. That is, we propose a generalization of the Ryu-Takayanagi prescription~\cite{Ryu:2006bv} including quantum-corrections~\cite{Faulkner:2013ana}. We establish some definitions and notation in Sec.~\ref{sec-preliminaries}. In Sec.~\ref{sec-gew} we define $E(a)$ as the spatial region with smallest generalized entropy, among all regions that contain $a$.
In Sec.~\ref{sec-properties} we show that $E$ obeys nontrivial properties consistent with its interpretation as an entanglement wedge: no-cloning, nesting, and strong subadditivity.
In Sec.~\ref{sec-covariant}, we discuss possible generalizations of our prescription to the time-dependent case, analogous to the Hubeny-Rangamani-Takayanagi prescription~\cite{Hubeny:2007xt} and its quantum extension by Engelhardt and Wall~\cite{Engelhardt:2014gca}. After introducing relevant concepts in Sec.~\ref{sec-covdefs}, we consider $E_n(a)$, the wedge with the smallest generalized entropy among all quantum-normal wedges that contain $a$ and share the conformal boundary (if any) of $a$. We show that $E_n$ reduces to the entanglement wedge prescription for boundary regions and static regions. Moreover, $E_n$ satisfies an appropriate no-cloning theorem. However, $E_n$ fails to satisfy strong subadditivity and nesting. We briefly discuss some alternative proposals.
\subsection{Discussion}
\label{sec-discussion}
By analogy with the interpretation of the usual entanglement wedge, our results suggest that information in $E(a)$ can be manipulated or summoned by a bulk observer in $a$. This is a striking implication.
We expect that such operations would not admit an interpretation in terms of a continuous, classical spacetime. Indeed, even simple operations in AdS, such as the instantaneous, local addition of a particle deep in the bulk, would violate the Bianchi identity and hence cannot be represented as a continuous geometry without introducing time-folds. Yet, this operation can be implemented instantaneously on a global boundary slice $\sigma$. Continuity of the boundary manifold across $\sigma$ then suggests that the bulk exists as a semiclassical spacetime in the past and future of $\sigma$, but not spacelike to $\sigma$. We expect that the reconstruction, from a bulk region $a$, of operators in $E(a)\cap a'$ involves comparably violent breakdowns of the spacetime geometry, at least outside of $a$.
We stress that the full physical significance of our proposal is not yet clear. In particular, we are not proposing a specific reconstruction map for operators in $E(a)\cap a'$, or any other generalization of the known entries of the holographic dictionary beyond the entanglement wedge prescription. But given its nontrivial properties, we expect $E$ to play a role in formulating quantum gravity as a holographic duality beyond AdS/CFT, for arbitrary spacetimes.
\subsection{Relation to other work}
\label{sec-otherwork}
In Ref.~\cite{Dong:2020uxp}, a definition was given for entanglement islands of \emph{low-energy} bulk fields in a weakly-gravitating region. The fields have an explicit momentum cut-off, and higher energy operators, such as those that would create black holes, are explicitly excluded. Hence, when applied to an asymptotic bulk region, the prescription~\cite{Dong:2020uxp} would not give the entanglement wedge of the corresponding boundary region. Thus, while we believe that Ref.~\cite{Dong:2020uxp} correctly treats the problem formulated there, its construction cannot be the right answer to the question we pose.
In Ref.~\cite{Grado-White:2020wlb}, a ``restricted maximin prescription'' was proposed for the entanglement wedge of a boundary region in a \emph{cut-off} spacetime. Again, this prescription is intended to answer a different question to the one we are interested in here. As we describe in Sec.~\ref{sec-covariantdiscussion}, a closely related prescription --- with the edge of $a$ playing the role of the cut-off surface --- yields a time-dependent generalization of our proposal that obeys strong subadditivity, but which violates the nesting property required for the entanglement wedges of bulk regions.
Finally, the present work does not support claims that there is no Page curve for an evaporating black hole \cite{Raju:2021lwh}. The Page curve is expected to describe the entropy of low-energy Hawking radiation modes, not all the information in principle accessible far from the black hole. For example, let $a$ be the exterior of a large sphere in asymptotically flat space. Then the generalized entanglement wedge we define, $E(a)$, can include the entire interior enclosed by $a$. But this does not mean that the information in $E(a)\backslash a$ can be accessed via low-energy operators in $a$.
\section{Motivation}
\label{sec-motivation}
\subsection{Why a Bulk Region Should Have an Entanglement Wedge}
\label{sec-why}
We begin by arguing that at least {\em some} gravitating bulk regions constitute legitimate input to the entanglement wedge prescription. We will discuss the two examples shown in Fig.~\ref{fig-motivate}.
First, consider the Hawking radiation emitted by a black hole in asymptotically flat space, after the Page time. The radiation must contain the same information, whether it has been extracted into an auxiliary nongravitating system, or arrived at a portion $A$ of future null infinity, or is still traversing a distant, weakly gravitating region $a$. Operationally, the presence of an island $b$ in the former cases indicates unitarity: sufficiently careful experiments would show that the ultimate state of the radiation is pure. But if such experiments could only succeed in the complete absence of gravity, they would fail in the real world, so the question of unitarity would be operationally meaningless.
These considerations make it implausible that the radiation has an entanglement island only if its self-gravity is completely turned off, but not if its self-gravity is arbitrarily small. It follows that the weakly gravitating region $a$ must be assigned an entanglement wedge that includes the island $b$:
\begin{equation}
E(\mbox{distant Hawking radiation}~a) \supset a \cup b~.
\end{equation}
Let us pause for a moment to examine what is happening here. Careful measurements on one part of the bulk spacetime, the Hawking radiation, are allowing us to extract information that resides (from the semiclassical viewpoint) in a distant spacelike-separated region: the black hole interior. This is a striking example of a fundamental nonlocality that appears to be present in quantum gravity; sufficiently complex operators do not have to respect the semiclassical structure of spacetime. Importantly this fundamental nonlocality is not solely a feature of black hole physics; instead it is an essential aspect of holography. This brings us to our second example (Fig.~\ref{fig-motivate}, right panel).
In the AdS/CFT correspondence, local CFT operators in a conformal boundary subregion $A$ are dual to (quasi-)local bulk operators near $A$~\cite{Banks:1998dd,Hamilton:2006az}. The notion of ``near'' can be made precise by defining a bulk region $a$ as the union of the entanglement wedges of tiny boundary regions containing slightly smeared local boundary operators.\footnote{High-dimensional local operators in the CFT will not correspond to low-energy fields in the bulk effective field theory. Instead, they will generically create large black holes. However these black holes will still be localized near the asymptotic boundary. See, \emph{e.g.}, the discussion in \cite{Harlow:2018tng}.} But local CFT operators generate the entire algebra of the CFT. It follows that the algebra generated by {\em bulk operators} in the near-boundary bulk region $a$ encodes the \emph{entire} entanglement wedge of $A$.\footnote{In 2+1 boundary dimensions and higher, Wilson loops play an important role; these are nonlocal operators that (in non-Abelian gauge theories) cannot be written as a product of gauge-invariant local operators. However, any Wilson loop will be contained in a thin boundary strip with arbitrarily shallow entanglement wedge. Therefore it should also be accessible as a bulk operator in the corresponding near-boundary region, and should still be contained in the associated bulk algebra.} Information apparently located deep in the bulk must be secretly, and highly nonlocally, encoded in quasilocal bulk operators near the asymptotic boundary.\footnote{In fact, this encoding is exponentially simpler than the encoding of the black hole interior in Hawking radiation \cite{Harlow:2013tf, Brown:2019rox}.}
Thus, the traditional entanglement wedge prescription for boundary subregions, $\mathrm{EW}(A)$, can be reinterpreted as a statement about the algebra generated by quasi-local bulk operators in the asymptotic bulk region $a$, without any reference to a CFT dual. It follows that an asymptotic bulk region $a$ in asymptotically AdS spacetime must possess an entanglement wedge --- indeed, the same entanglement wedge that would have been assigned to its conformal boundary $A$:
\begin{equation} \label{eq:asymptoticspecialcase}
E(\mbox{asymptotic AdS bulk region}~a) = \mathrm{EW}(A)~.
\end{equation}
If two such apparently dissimilar bulk regions --- Hawking radiation and asymptotic regions in AdS --- encode a larger entanglement wedge, then Occam's razor suggests that
\begin{itemize}
\item \emph{any} bulk region should have an associated entanglement wedge; and
\item \emph{only} bulk regions possess an entanglement wedge. (Boundary regions and auxiliary systems can be viewed as idealized limits of bulk regions that are near asymptotic infinity or very weakly gravitating respectively.)
\end{itemize}
\subsection{Motivation From Tensor Networks }
\label{sec-tensor}
\begin{figure}[t]
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width = 0.75\linewidth]{tensor1}
\end{subfigure}
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width = 0.75\linewidth]{tensor2}
\end{subfigure}
\caption{\textit{Left:} Two tensors $T_1$ and $T_2$ in a tensor network. The in-plane legs (horizontal) connect different tensors together, while the bulk legs (vertical) input matter degrees of freedom. \textit{Right:} By bending the in-plane leg connecting $T_1$ and $T_2$ around, we can reinterpret it as two maximally entangled bulk legs, one on each tensor.}
\label{fig-tensors}
\end{figure}
Additional motivation for such a proposal comes from tensor networks, a useful toy model of static states in quantum gravity~\cite{Swingle:2009bg, Pastawski:2015qua, Hayden:2016cfa, Bao:2018pvs}. In these models, each tensor represents a patch of the gravitational spacetime, with the logarithm of the dimension of an edge capturing the area (in Planck units) of the surface connecting two neighbouring patches. Bulk degrees of freedom within the patch are described by additional ``out-of-plane'' legs that feed into the network.
Importantly, the distinction between out-of-plane and in-plane tensor network legs is not precise. Mathematically, highly entangled Rindler modes on either side of a cut play an identical role in the tensor network as in-plane legs representing area, as shown in Fig.~\ref{fig-tensors}. The same effect is seen in gravity where the Rindler modes and geometrical area are equivalent under renormalization group flow. As we increase the UV cut-off on the bulk quantum field theory, the amount of Rindler entanglement across a surface increases; this increase is cancelled in a computation of generalized entropy by the decrease in $A/4G$ created by the renormalization of Newton's constant.
If we extrapolate the UV cut-off all the way to the Planck scale, it is natural to conjecture that the area term, or analogously the in-plane tensor network legs, go away entirely, with the entanglement coming entirely from out-of-plane bulk legs. In such a limit, the tensor network consists purely of a set of projection operators that map the bulk state to the boundary state.
Why does this suggest that a region $a$ encodes the larger region $E(a)$, as suggested in Fig.~\ref{fig:introfig}? If in-plane legs are really bulk Planck scale bulk legs, then an operator $O$ can be reconstructed by a (potentially very high energy) operator within the region $a$ if the action of $O$ is the same as an operator $\tilde O$ that acts on the out-of-plane legs within $a$, along with the in-plane legs at the boundary of $a$. After all, both sets of degrees of freedom should be in principle accessible from $a$, even if the Planckian degrees of freedom
may not be described by an effective field theory.
In a tensor network, $E(a)$ corresponds to the smallest-entropy cut through the network such that $a$ is contained entirely within the cut (see Fig. \ref{fig:introfig}, right panel). Let us assume for simplicity that the dominant contribution to the entropy of any given cut comes from the maximally entangled in-plane legs. (In gravity, this corresponds to taking the semiclassical limit where the area term gives the dominant contribution to generalized entropy.) Now consider the map $V$ induced by the tensor network from the boundary of $E(a)$, together with bulk legs in $E(a) \backslash a$, to the boundary of $a$. By definition, any intermediate cut through this subnetwork describing $V$ has larger area (i.e. much larger bond dimension) than the input to $V$. For networks with sufficiently generic tensors, this ensures that the map $V$ is an (approximate) isometry,\footnote{Concretely, this will be the case with very high probability if the tensors are drawn from a Haar random ensemble \cite{Hayden:2016cfa}.} and hence that the operator $\tilde O = V^\dagger O V$, which acts on the in-plane legs at the boundary of $a$, successfully reconstructs any operator $O$ acting in $E(a) \backslash a$.
\section{Generalized Entanglement Wedges in Time-reflection\\ Symmetric Slices}
\label{sec-prescription}
\subsection{Preliminaries}
\label{sec-preliminaries}
\begin{defn}\label{def:trs}
A spacetime $M$ is \emph{time-reflection symmetric} if there exists a $\mathbb{Z}_2$ action on $M$ that exchanges past and future timelike directions and preserves all points in a Cauchy surface $\Sigma$.
\end{defn}
\begin{conv}
In this section, we will only be concerned with the time-reflection symmetric Cauchy surface $\Sigma$, regarded as a manifold with Riemannian metric $h$. We assume that $\Sigma$ is inextendible.
\end{conv}
\begin{defn}
We use $\partial s$ to denote the boundary of a set $s\subset \Sigma$ in the induced topology of $\Sigma$. Also $\cl s\equiv s \cup \partial s$ and $\setint s = s \cap \cl s^C$, where the superscript $C$ denotes the set complement in $\Sigma$. \end{defn}
\begin{defn}
A \emph{wedge} $a$ is any open subset of $\Sigma$ that is the interior of its closure: $a=\setint \cl a$. (The name is slightly more natural in the more general time-dependent context; it is adopted for compatibility with the existing phrase ``entanglement wedge.'')
\end{defn}
\begin{defn}
The intersection of two wedges is a wedge; but the union, complement, and relative complement need not be. Given two wedges $a$ and $b$, the \emph{wedge union}, \emph{wedge complement}, and \emph{wedge relative complement} are wedges defined by
\begin{align}
a\Cup b & \equiv \setint\cl(a\cup b)~,\\
a' & \equiv \setint a^C~, \\
a\setminus b & \equiv a \cap b'~.
\end{align}
\end{defn}
\begin{defn}\label{def:area}
The \emph{area} of a wedge $a$ is the geometric area of $\partial a$ and is denoted $\A(a)$.
\end{defn}
\begin{defn}\label{def:sgen}
The {\em generalized entropy} of a wedge $a$ is
\begin{equation}\label{sgendef}
S_{\rm gen}(a)\equiv \frac{\A(\partial a)}{4G\hbar} + S(a) + \ldots
\end{equation}
Here
\begin{equation}
S(a) = - \tr \rho_a \log \rho_a
\end{equation}
is the von Neumann entropy of
\begin{equation}
\rho_a = \tr_{a'} \rho~,
\end{equation}
the density operator of the quantum fields restricted to $a$.
\end{defn}
\begin{rem}
The area in Eq.~\eqref{sgendef} can be thought of as a counterterm that cancels the leading divergence in the von Neumann entropy $S$. Additional counterterms for subleading divergences have been omitted. See the appendix of Ref.~\cite{Bousso:2015mna} for details.
\end{rem}
\begin{defn}
Let $\tilde \Sigma\equiv (M,\tilde h)$ denote the \emph{conformal completion} of $\Sigma$, obtained by conformally mapping $\Sigma$ to a subset of a compact set (for example, a sphere) and taking the closure of the image. $\tilde \Sigma$ may be a manifold-with-boundary~\cite{Wald:1984rg}. The boundary of the image, $\partial \tilde \Sigma$, is called the \emph{conformal boundary} of $\Sigma$. We will not distinguish notationally between a set $a\subset \Sigma$ and its image in $\tilde\Sigma$. However, such a distinction is crucial for boundaries of sets.
\end{defn}
\begin{defn}\label{def:staticconformaledge}
The boundary of a wedge $a\subset \tilde \Sigma$ will be denoted $\delta a$. As shown in Fig. \ref{fig-conf-bdy}, we define the {\em conformal boundary} of $a$ as the set
\begin{equation}
\tilde\partial a\equiv \delta a\cap \partial\tilde \Sigma~;
\end{equation}
thus,
\begin{equation}
\delta a = \partial a\sqcup \tilde\partial a~.
\end{equation}
\end{defn}
\begin{figure}[t]
\centering
\includegraphics[width = 0.5\linewidth]{ConfBdy}
\caption{The boundary $\delta a \in \tilde\Sigma$ of a wedge $a$ in the conformal completion $\tilde\Sigma$ decomposes into the original boundary $\partial a \in \Sigma$ and a conformal boundary $\tilde \partial a \in \partial \tilde\Sigma$.}
\label{fig-conf-bdy}
\end{figure}
\subsection{Definition and Basic Properties}
\label{sec-gew}
\begin{defn}\label{ewstatic}
Given a wedge $a\subset\Sigma$, we define its {\em generalized entanglement wedge}, $E(a)$, as the wedge that satisfies
\begin{equation}
a\subset E(a)\subset \Sigma~~~\mbox{and}~~~\tilde\partial a = \tilde\partial E(a)
\end{equation}
and which has the smallest generalized entropy among all such wedges. We assume without proof that $E(a)$ exists, and for convenience we shall assume that it is unique.
\end{defn}
\begin{lem}[Monotonicity]\label{monolem}
Let $a$ and $b$ be wedges with the same conformal boundary, $\tilde\partial a=\tilde\partial b$. Then
\begin{equation}
a\subset b\implies \S[E(a)]\leq \S[E(b)]~.
\end{equation}
That is, the generalized entropy of the entanglement wedge increases monotonically with the input wedge under inclusion, if the conformal boundary is held fixed.
\end{lem}
\begin{proof}
$E(b)\supset b\supset a \implies \S[E(b)]\geq \S[E(a)]$ by Def.~\ref{ewstatic}.
\end{proof}
\begin{defn}[Entanglement wedge of a boundary region] \label{defn:bdyews}
Let $B \subset \partial \tilde\Sigma$ be a subregion of the conformal boundary $\partial \tilde\Sigma$. The (ordinary) \emph{entanglement wedge} $\mathrm{EW}(B)\subset \Sigma$~\cite{Ryu:2006bv,Faulkner:2013ana} is the wedge with conformal boundary $B$ and smallest generalized entropy among all such sets.
\end{defn}
\begin{lem}[$\mathrm{EW}$ as a special case of $E$] \label{lem:genew=ews}
If the wedge $a$ lies in the (ordinary) entanglement wedge of its conformal boundary, $a \subset \mathrm{EW}(\tilde\partial a)$, then its generalized entanglement wedge is
\begin{equation}
E(a) = \mathrm{EW}(\tilde\partial a).
\end{equation}
\end{lem}
\begin{proof}
The result follows immediately from Definitions~\ref{ewstatic} and \ref{defn:bdyews}.
\end{proof}
\begin{rem} [Asymptotic bulk regions]
Let $a_n$ be an infinite sequence of wedges with $a_{n+1} \subset a_n$, $\tilde\partial a_{n+1} = \tilde \partial a_n$ and $\cap_n a_n = \varnothing$. Then $E(a_n) = \mathrm{EW}(\tilde \partial a_n)$ for all sufficiently large $n$.
\end{rem}
\subsection{No-Cloning, Strong Subadditivity, and Nesting}
\label{sec-properties}
\begin{figure}[t]
\begin{center}
\includegraphics[width = 0.6\linewidth]{staticNoCloning.pdf}
\end{center}
\caption{A configuration excluded by Theorem \ref{staticnocloning} (No-cloning): if $a \cap E(b)=\varnothing$ and $b\cap E(a)=\varnothing$ then $E(a) \cap E(b)=\varnothing$, because otherwise removing the overlap from each entanglement wedge would fail to increase their combined generalized entropy.}
\label{fig:nocloning}
\end{figure}
\begin{thm}[No cloning]\label{staticnocloning}
Let $a,b$ be wedges that satisfy
\begin{equation}\label{staticexclusions}
a \cap E(b)=\varnothing~, ~~ b\cap E(a)=\varnothing~.
\end{equation}
Then
\begin{equation}\label{staticnocloneq}
E(a)\cap E(b)=\varnothing~.
\end{equation}
\end{thm}
\begin{proof}
As shown in Fig. \ref{fig:nocloning},
\begin{align}
\A[E(a)\setminus E(b)] + \A[E(b)\setminus E(a)] \leq \A[E(a)] + \A[E(b)]~.
\end{align}
(This need not be an equality; for example, there may be a disconnected shared boundary in $E(a)\cap E(b)$.) Strong subadditivity implies the same inequality for the von Neumann entropies; hence
\begin{align}
\S[E(a)\setminus E(b)] + \S[E(b)\setminus E(a)] \leq \S[E(a)] + \S[E(b)]~.
\end{align}
That is, the generalized entropy of $E(a)$ or $E(b)$ decreases when $E(a)\cap E(b)$ is removed. By the assumption of the Lemma, $a \subset E(a) \setminus E(b)$ and $b \subset E(b) \setminus E(a)$, so this contradicts Def.~\ref{ewstatic} unless $E(a)\cap E(b)=\varnothing$.
\end{proof}
\begin{figure}[t]
\begin{center}
\includegraphics[width = 0.6\linewidth]{staticSSA.pdf}
\end{center}
\caption{An illustration of the setup from Theorem \ref{thm:ssastatic} (Strong subadditivity): The combined $\S$ of $E(a \Cup b)$ and $E(b \Cup c)$ is at least big as the combined $\S$ of their intersection and union. In turn, these upper bound the combined $\S$ of $E(b)$ and $E(a \Cup b \Cup c)$ respectively.}
\label{fig:SSA}
\end{figure}
\begin{thm}[Strong subadditivity]\label{thm:ssastatic}
Let $a$, $b$ and $c$ be mutually disjoint open subsets of $\Sigma$. Then
\begin{equation}\label{ssastatic}
\S[E(a \Cup b)] + \S[E(b \Cup c)] \geq \S[E(b)] + \S[E(a \Cup b \Cup c)]~.
\end{equation}
\end{thm}
\begin{proof}
The setup is illustrated in Fig. \ref{fig:SSA}. Rearranging components of surfaces, we have
\begin{align}
\A[E(a \Cup b)] & + \A[E(b \Cup c)] \geq \nonumber \\
& \A[E(a \Cup b)\cap E(b \Cup c)] + \A[E(a \Cup b)\Cup E(b \Cup c)]~.
\end{align}
(This need not be an equality since the wedge union can erase boundary portions.) Strong subadditivity implies the same inequality for the von Neumann entropies, and hence for the generalized entropy:
\begin{equation}
\S[E(a \Cup b)] + \S[E(b \Cup c)]\geq
\S[E(a \Cup b)\cap E(b \Cup c)] + \S[E(a \Cup b)\Cup E(b \Cup c)]
\end{equation}
The first set on the right contains $b$ and the second contains $a\Cup b\Cup c$, so Def.~\ref{ewstatic} implies Eq.~\eqref{ssastatic}.
\end{proof}
\begin{thm}[Nesting]
\begin{equation}\label{thm:ewnstatic}
a\subset b \implies E(a)\subset E(b)~.
\end{equation}
\end{thm}
\begin{figure}[t]
\begin{center}
\includegraphics[width = 0.4\linewidth]{staticNesting}
\end{center}
\caption{A hypothetical counterexample to Theorem \ref{thm:ewnstatic} (Nesting): if $a \subset b$ but $E(a) \not\subset E(b)$, we can replace the two entanglement wedges by $E(a) \cap E(b)$ and $E(a) \Cup E(b)$ respectively, decreasing their combined area.}
\label{fig:nesting}
\end{figure}
\begin{proof}
The setup is illustrated in Fig. \ref{fig:nesting}. Since $a\subset b$, we have $a\subset E(a)\cap E(b)$, so by Def.~\ref{ewstatic}
\begin{align}\label{ewnstaticproof1}
\S[E(a)] & \leq \S[E(a)\cap E(b)]~,\\
\S[E(b)] & \leq \S[E(a)\Cup E(b)]~.
\end{align}
But
\begin{equation}
\A[E(b)] + \A[E(a)] \geq \A[E(a)\cap E(b)] + \A[E(a)\Cup E(b)]~.
\end{equation}
(This need not be an inequality since the wedge union can erase boundary portions.) Strong subadditivity implies the same inequality for the von Neumann entropies, and hence for the generalized entropy:
\begin{equation}
\S[E(b)] + \S[E(a)] \geq \S[E(a)\cap E(b)] + \S[E(a)\Cup E(b)]~.
\end{equation}
To avoid a contradiction, all three of the above inequalities must be saturated. The assumed uniqueness of $E(a)$ implies $E(a)\cap E(b)=E(a)$, and hence $E(a)\subset E(b)$.
\end{proof}
\section{Time-dependence}
\label{sec-covariant}
In this Section we discuss possible generalizations of our proposal to the case where $a$ does not lie on a time-reflection symmetric Cauchy surface.
We will sketch approaches, obtain some partial results, and outline specific challenges. We focus primarily on one possible generalization -- the smallest-generalized-entropy quantum normal wedge $E_n(a)$ -- which turns out to obey a no-cloning theorem but not strong subadditivity or nesting. We briefly comment on other possible generalizations that preserve strong subadditivity and nesting respectively, but do not find any single definition that preserves all three.
\subsection{Preliminaries}
\label{sec-covdefs}
We begin by introducing natural generalizations of the key objects introduced in Sec.~\ref{sec-preliminaries}. Let $M$ be a globally hyperbolic Lorentzian spacetime with metric $g$. The chronological and causal future and past, $I^\pm$ and $J^\pm$; and the future and past domains of dependence and Cauchy horizons, $D^\pm$ and $H^\pm$; are defined as in Wald~\cite{Wald:1984rg}. Given $s\subset M$, we use $\partial s$ to denote the boundary of $s$ in $M$, and $s'$ to denote the interior of the set of points that are spacelike related to all points in $s$, that is, points outside the causal future and past of $s$.
\begin{defn}\label{def:covwedge}
A {\em wedge} is a set $a\subset M$ that satisfies $a=a''$ (see Fig. \ref{fig-wedges}, left).
\end{defn}
\begin{figure}[t]
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width = 0.8\linewidth]{wedge}
\end{subfigure}
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width = 0.8\linewidth]{IntersectingWedges}
\end{subfigure}
\caption{\emph{Left:} A spherical wedge $a$, its complement wedge $a'$, and their shared edge $\eth a$ in Minkowski space. The wedge $a$ is diamond-shaped in this spacetime diagram; a Cauchy slice of $a$ is shown in dark green. This wedge is ``normal,'' since outgoing orthogonal lightrays expand. \emph{Right:} The intersection of two wedges is itself a wedge, with an edge that decomposes as $\eth(a \cap b) = \cl\left[(\eth a\cap b) \sqcup (H^+(a)\cap H^-(b))\sqcup \set{a\leftrightarrow b}\right]$.}
\label{fig-wedges}
\end{figure}
\begin{rem}\label{wilem}
It can be shown that the intersection of two wedges $a,b$ is a wedge (see Fig. \ref{fig-wedges}, right); and the \emph{complement wedge} $a'$ is a wedge:
\begin{equation}
(a\cap b)'' = a\cap b~;~~a'''=a'~.
\end{equation}
\end{rem}
\begin{defn}
Given two wedges $a$ and $b$, we define the {\em relative complement of $b$ in $a$} as $a\cap b'$, and the {\em wedge union} as $a\Cup b\equiv (a'\cap b')'$ (see Fig. \ref{fig:union}); these objects are wedges by the above remark.
\end{defn}
\begin{rem
The wedge union satisfies $a\Cup b \supset a\cup b$. It is minimal in the sense that any wedge that contains $a\cup b$ must contain $a\Cup b$.
\end{rem}
\begin{defn
The {\em edge} $\eth a$ of a wedge $a$ is
defined by $\eth a\equiv \partial a \cap \partial a'$. Note that $a$ is fully characterized by specifying $\eth a$ and one spatial side of $\eth a$.
\end{defn}
\begin{defn}
The \emph{area} and \emph{generalized entropy} of a wedge $a$ are defined as in the static case, Defs.~\ref{def:area} and \ref{def:sgen}, with $\partial a$ replaced by $\eth a$. Note that neither depend on a choice of Cauchy slice of $a$.
\end{defn}
\begin{defn}\label{def:conformaledge}
Given a wedge $a$, we distinguish between its edge $\eth a$ in $M$ and its edge $\delta\hspace*{-0.2em}\bar{}\hspace*{0.2em} a$ as a subset of the conformal completion $\tilde M$. The latter can contain an additional piece, the {\em conformal edge}
\begin{equation}
\tilde\eth a\equiv \delta\hspace*{-0.2em}\bar{}\hspace*{0.2em} a\cap \partial\tilde M~.
\end{equation}
\end{defn}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.8\linewidth]{WedgeUnion}
\end{center}
\caption{The wedge union (thick purple diamond) of two wedges (blue diamonds) is defined as the wedge complement of the intersection of their wedge complements. Two examples are shown.}
\label{fig:union}
\end{figure}
\begin{defn}
Given a wedge $a$ and a point $p\in \eth a$, the past (future) {\em quantum expansion}, $\Theta^-(a,p)$ ($\Theta^+(a,p)$), is the shape derivative of the generalized entropy under outward deformations of $a$ along the past (future) null vector field orthogonal to $\eth a$ at $p$. A precise definition can be given in terms of a functional derivative:
\begin{equation}
\Theta^\pm(a,p)\equiv 4G\hbar \, \frac{\delta S_{\rm gen}[a(X^\pm(p))]}{\delta X^\pm(p)}~.
\end{equation}
Here $X^\pm$ are null coordinates orthogonal to $\eth a$, and $a(X^\pm(p))$ are wedges obtained by deforming $\eth a$ along them. See Ref.~\cite{Bousso:2015mna} for further details.
\end{defn}
\begin{rem}
By Eq.~\eqref{sgendef}, the quantum expansion can be decomposed into a classical and a quantum piece:
\begin{equation} \label{eq:qexpansion}
\Theta^\pm(a,p) = \theta(p)+4G\hbar\, \frac{\delta S[a(X^\pm(p))]}{\delta X^\pm(p)}~.
\end{equation}
The first term is the classical expansion~\cite{Wald:1984rg}, which depends only on the shape of $\eth a$ near $p$. The second term in Eq.~\eqref{eq:qexpansion} is nonlocal.
\end{rem}
\begin{lem}\label{subsetexplem}
Let $a\subset b$ be wedges whose edges coincide in an open neighborhood $O$. At any point $p\in \eth a \cap O$,
\begin{equation
\Theta^\pm(a,p)\geq \Theta^\pm(b,p)~.
\end{equation}
\end{lem}
\begin{proof}
Since $\eth a$ and $\eth b$ coincide near $p$, their shape, and hence their classical expansion $\theta$, will agree at $p$.
The Lemma then follows from strong subadditivity of the von Neumann entropy, applied to an infinitesimal deformation of $a$, and of $b$, at $p$; see Fig. 3a in Ref.~\cite{Bousso:2015mna}.
\end{proof}
\begin{defn}
The wedge $a$ is called {\em quantum-normal} at $p\in \eth a$ if $\Theta^+(a,p)\geq 0$ and $\Theta^-(a,p)\geq 0$. Other combinations of signs correspond to \emph{quantum-antinormal} ($\leq,\leq$), \emph{quantum-trapped} ($\leq,\geq$), \emph{quantum-antitrapped} ($\geq,\leq$) and \emph{quantum extremal} ($=,=$). Marginal cases arise if one expansion vanishes at $p$. In relations that hold for all $p\in \eth a$, we drop the argument $p$.
\end{defn}
\begin{conj}[Quantum Focussing Conjecture] \label{conj-qfc}
At all orders in $G\hbar$~\cite{Bousso:2015mna},
\begin{equation}\label{qfc}
\frac{\delta \Theta(a,p)}{\delta X^\pm(\bar{p})}\leq 0~.
\end{equation}
\end{conj}
\begin{rem}
In the ``off-diagonal'' case, $p\neq \bar{p}$, the Quantum Focussing Conjecture (QFC) follows from Lemma~\ref{subsetexplem}. A quantum field theory limit of the QFC, the Quantum Null Energy Condition~\cite{Bousso:2015mna}, was proven in Refs.~\cite{Bousso:2015wca, Balakrishnan:2017bjg}; see also Refs.~\cite{Koeller:2015qmn, Wall:2017blw, Ceyhan:2018zfg, Balakrishnan:2019gxl}. The general diagonal case, $p=\bar{p}$, remains a conjecture.
\end{rem}
\begin{lem}\label{qnormalwi}
The intersection of two quantum-normal wedges is quantum-normal.
\end{lem}
\begin{proof}
Let $a$ and $b$ be quantum-normal. We use the decomposition of $\eth(a \cap b)$ described in Fig. \ref{fig-wedges}. For $p \in \eth a\cap b$ (or $p \in a\cap \eth b$), the condition $\Theta^\pm(a \cap b, p) \geq 0$ follows from the normalcy of $a$ (or $b$) at $p$ and Lemma~\ref{subsetexplem}. For $p\in H^\pm (a)\cap H^\mp(b)$, let $a_p \subset a$ and $b_p \subset b$ be deformations of $a$ and $b$ along $H^\pm (a)$ and $H^\mp (b)$ respectively such that $\eth a_p, \eth b_p$ coincide with $H^\mp (a)\cap H^\pm(b)$ in a neighbourhood of $p$. Since $\Theta^\pm(a) \geq 0$ and $\Theta^\mp(b) \geq 0$ in a neighbourhood of the points lightlike separated from $p$, Conjecture~\ref{conj-qfc} implies that $\Theta^\pm(a_p,p)\geq 0$ and $\Theta^\mp(b_p,p) \geq 0$. Hence $\Theta^\pm(a \cap b, p) \geq 0$ by Lemma~\ref{subsetexplem}.
\end{proof}
\subsection{Smallest-Generalized-Entropy Normal Wedge}
\label{sec-ssnw}
\begin{defn}\label{qew}
Given a wedge $a$, let $E_n(a)$ denote the quantum-normal wedge that contains $a$, has the same conformal boundary as $a$, and has the smallest generalized entropy among all such wedges. As in the static case (Def.~\ref{ewstatic}), we assume without proof that this wedge exists, and we assume for convenience that it is unique.
\end{defn}
\begin{lem}[Monotonicity]
Let $a$ and $b$ be wedges with the same conformal boundary, $\tilde\eth a=\tilde\eth b$. Then
\begin{equation}
a\subset b\implies \S[E_n(a)]\leq \S[E_n(b)]~.
\end{equation}
\end{lem}
\begin{proof}
The proof is analogous to the static case, Lemma \ref{monolem}.
\end{proof}
\begin{lem}[Quantum Expansion] \label{qeqe}
By definition, $E_n$ is quantum normal. Marginal cases arise at points $p\in\eth[E_n(a)]$ that do not lie on $\eth a$:
\begin{itemize}
\item $E_n(a)$ is marginally quantum-antitrapped, $\Theta^-[E_n(a),p]=0$, at $p\in H^+(a')$.
\item $E_n(a)$ is marginally quantum-trapped, $\Theta^+[E_n(a),p]=0$, at points $p\in H^-(a')$.
\item $E_n(a)$ is quantum extremal, $\Theta^+[E_n(a),p]=\Theta^-[E_n(a);p]=0$, at points $p\in a'$.
\end{itemize}
\end{lem}
\begin{proof}
By quantum normalcy of $E(a)$, $\S$ decreases when $\eth E(a)$ is deformed towards $\eth a$ at $p$. Lemma \ref{subsetexplem} ensures quantum normalcy of $\eth E(a)$ away from $p$ under this inward deformation. This conflicts with the $\S$-minimization requirement in Def.~\ref{qew}, unless the quantum expansion at $p$ in the other null direction immediately becomes negative under the deformation.
\end{proof}
\begin{defn}[Entanglement wedge of a boundary region] \label{defn:bdyqew}
Let $B \subset \partial \tilde{M}$ be a subregion of a Cauchy surface of the conformal boundary $\partial \tilde{M}$. The \emph{entanglement wedge of} of $B$, $\mathrm{EW}(B)\subset M$, is the smallest-generalized-entropy quantum-extremal wedge with conformal edge $B$~\cite{Engelhardt:2014gca}.
\end{defn}
\begin{lem}[Reduction of $E_n$ to $\mathrm{EW}$]\label{lem:qgenew=ew}
Let $a$ be contained in the ordinary quantum entanglement wedge of its conformal edge $\tilde\eth a$, \emph{i.e.}, $a \subset \mathrm{EW}(\tilde\eth a)$. Then
\begin{align}
E_n(a) = \mathrm{EW}(\tilde\eth a).
\end{align}
\end{lem}
\begin{proof}
By the arguments in the proof of Lemma \ref{qeqe}, the smallest-$\S$ quantum-normal wedge with conformal edge $\tilde\eth a$ is necessarily extremal, and hence is given by $\mathrm{EW}(\tilde\eth a)$. By assumption $a \subset \mathrm{EW}(\tilde\eth a)$, so this continues to be true when we restrict to normal wedges containing $a$.
\end{proof}
\begin{lem}[Reduction of $E_n$ to $E$] \label{lem:entoe}
Let $\Sigma_0$ be a time-reflection symmetric Cauchy surface, and let $a$ be a wedge with $\eth a\subset \Sigma_0$. Then $E_n(a)$ reduces to the static entanglement wedge $E(a)$:\footnote{In Sec.~\ref{sec-prescription}, $\partial$ denotes the boundary in the topology of $\Sigma_0$, and we revert to this usage in this lemma and its proof. Elsewhere in the present section, $\partial$ denotes the boundary in $M$.}
\begin{equation}
\eth E_n(a)= \partial E(a)~.
\end{equation}
\end{lem}
\begin{proof}
Let $T$ be the time-reflection operator.
The set $E(a) \cap T[E(a)]$ is a wedge by Remark~\ref{wilem}, normal by Lemma \ref{qnormalwi}, and it contains $a$. It is also time-reflection symmetric, so its edge lies on $\Sigma_0$. This conflicts with Conj.~\ref{conj-qfc} unless $E(a)=T[E(a)]$. Hence $\eth E_n(a) \subset \Sigma_0$. Moreover, $\partial E(a\cap \Sigma_0)$ must be pointwise quantum-normal or antinormal by time-reflection symmetry. If it were anti-normal at some point $p$, we could decrease $\S$ by deforming $\partial E(a\cap\Sigma_0)$ outward on $\Sigma_0$ at $p$, contradicting its definition.
\end{proof}
\begin{thm}[No Cloning for $E_n$
Let $a,b$ be wedges that satisfy
\begin{equation}\label{qexclusions}
a\subset E_n(b)'~, ~~ b\subset E_n(a)'~.
\end{equation}
Then
\begin{equation}\label{qnocloneq}
E_n(a)\subset E_n(b)'~,
\end{equation}
or equivalently, $E_n(b)\subset E_n(a)'$.
\end{thm}
\begin{proof}
For arbitrary wedges $c$ and $d$, let
\begin{equation}\label{wp}
c_{/d}\equiv (c\cap d)\Cup(c\cap d')~.
\end{equation}
Lemma~\ref{qeqe} and Conj.~\ref{conj-qfc} can be shown to imply
\begin{equation}
\S[E_n(a)'_{/E_n(b)}]\leq \S[E_n(a)]~,\,\,\,\mathrm{and}\,\,\,\S[E_n(b)'_{/E_n(a)}]\leq \S[E_n(b)]~.
\end{equation}
We add these inequalities and use strong subadditivity of the von Neumann entropy to obtain
\begin{equation}\label{qseaebp}
\S[E_n(a)\cap E_n(b)'] + \S[E_n(a)'\cap E_n(b)] \leq \S[E_n(a)]+\S[E_n(b)]~.
\end{equation}
We will now show that $E_n(a)\cap E_n(b)'$ is quantum-normal. Let us begin by substituting $a\to E_n(a)$ and $b\to E_n(b)'$ in the proof of Lemma~\ref{qnormalwi}. The Lemma assumed that $a$ and $b$ are normal, whereas now $E_n(b)'$ is quantum-antinormal. However, our assumption, Eq.~\eqref{qexclusions}, implies that $\Theta^+[E_n(b)',p]=0$ or $\Theta^-[E_n(b)',p]=0$ at any points $p$ where these expansions actually appears in the proof. Indeed, since $E_n(a)$ is spacelike to $b$, $\Theta^\pm[E_n(b)',p]=0$ for $p\in E_n(a)\cap \eth E_n(b)$ by Lemma~\ref{qeqe}. If $p$ is the endpoint of a geodesic generator of $H^-[E_n(b)']$ that intersects $H^+[E_n(a)]$, then $\Theta^-[E_n(b)',p]=0$ by Lemma~\ref{qeqe}, or else there would be a causal curve from $\eth E_n(a)$ to $b$, in conflict with Eq.~\eqref{qexclusions}.
By the same reasoning $E_n(a)'\cap E_n(b)$ is also quantum-normal. Moreover, $a\subset E_n(a)\cap E_n(b)'$ and $b\subset E_n(a)'\cap E_n(b)$ by Eq.~\eqref{qexclusions}. By Def.~\ref{qew}, $E_n(a)$ and $E_n(b)$ are the unique smallest-$\S$ quantum-normal wedges containing, respectively, $a$ and $b$. Eq.~\eqref{qseaebp} thus implies $E_n(a)\cap E_n(b)'=E_n(a)$, which is equivalent to Eq.~\eqref{qnocloneq}.
\end{proof}
\subsection{Discussion}
\label{sec-covariantdiscussion}
$E_n$ does not satisfy subadditivity or nesting. This suggests that $E_n$ is not quite the correct generalization of $E$. A counterexample to both properties is shown in Fig.~\ref{ShellCE}. A purely classical counterexample to strong subadditivity is furnished by taking $b$ to be a rectangle of length and width $\Delta y\gg \tau \gg \Delta x\gg \ell_P$ at $t=0$ in two-dimensional Minkowski space, and $a$ and $c$ to be identical rectangles obtained by moving $b$ by $1+2\Delta x$ in the $\pm x$ direction and by $\tau$ in the $t$ direction.
\begin{figure}[t]
\begin{center}
\includegraphics[width = 0.6\linewidth]{Anomalies}
\end{center}
\caption{Collapse of a null shell purified by a distant reference system $r$. The wedge $a$ is marginally quantum trapped, so $a=E_n(a)$. One can arrange that $\S[E_n(a)]+\S[E_n (r)]\gg \S[E_n(a\cup r)]$. If $r$ is the Python's Lunch~\cite{Brown:2019rox} (dashed red) of an asymptotic region $b$, one can arrange that $E_n(a\cup b)\not\supset E_n(b)$. Hence $E_n$ satisfies neither subadditivity nor nesting.}
\label{ShellCE}
\end{figure}
Ideally, we would like to find a prescription $e(a)$ that satisfies all of the following properties:
\begin{enumerate}
\item $e$ reduces to the time-reflection symmetric prescription given in Sec.~\ref{sec-prescription}: $e(a) = E(a)$.
\item $e$ reduces to the QES prescription for the entanglement wedge of boundary regions: $a\subset EW(\delta\hspace*{-0.2em}\bar{}\hspace*{0.2em} a)\implies e(a) = EW(\delta\hspace*{-0.2em}\bar{}\hspace*{0.2em} a)$.
\item $e$ is well-defined.
\item $e$ satisfies no-cloning, strong subadditivity, and nesting, in the sense shown for $E$ in Sec.~\ref{sec-prescription}.
\end{enumerate}
We have seen that $E_n$ satisfies the first three requirements as well as no-cloning. We are aware of alternate candidate prescriptions satisfying other subsets of these requirements, but of no \emph{single} prescription that satisfies all. For example, one can ensure nesting in addition to the requirements satisfied by $E_n$, by taking the wedge union $\Cup_{b \subset a} E_n(b)$ over all subwedges $b \subset a$. However, this definition is somewhat artificial, and it still does not obey strong subadditivity.
Ref.~\cite{Grado-White:2020wlb} proposed a restricted maximin prescription for the entanglement wedge of a cut-off boundary region. This prescription can be converted into a proposal for the generalized entanglement wedge of a bulk region, which (like $E_n$) satisfies the first three requirements above. In addition, it obeys strong subadditivity; however, it violates nesting.\footnote{We also expect no-cloning to hold but we have not verified this.} Explicitly, let $\eth a$ be the cut-off surface denoted $\gamma$ in Ref.~\cite{Grado-White:2020wlb}, and let $a$ be its exterior, \emph{i.e.}, the portion discarded by the cut-off.
Choosing $A=\gamma$ in their notation, the prescription of Ref.~\cite{Grado-White:2020wlb} defines a wedge $\mathcal{E}(a) \subset a'$ bounded by $\eth a$ and its restricted maximin surface. The wedge union $e(a) = a \,\Cup \,\mathcal{E}(a)$ is then a candidate for the generalized entanglement wedge. The proof of strong subadditivity from Ref.~\cite{Grado-White:2020wlb} directly implies strong subadditivity for the generalized entanglement wedges, so long as $e(a \Cup b) \cap c = e(b \Cup c) \cap a = \varnothing$. (This ensures that the maximin surfaces lie in $a'\cap b' \cap c'$, the spacetime region that survives all three cutoffs. From an information-theoretic perspective, it is roughly analogous to the conditions in Theorem \ref{staticnocloning} which ensure that $a$, $b$ and $c$ are genuinely independent degrees of freedom.)
However, the statement of nesting proved in Ref.~\cite{Grado-White:2020wlb}, while appropriate in that context, is very different from the general statement $e(a) \subset e(b)$ that must hold for $e$ to be an entanglement wedge of a bulk region. In fact, it is easy to see that the latter statement fails, since $e(a) = a$ for any region $a$ with edge $\eth a$ piecewise lightlike.
However, it may not be appropriate to seek a single definition of the entanglement wedge of bulk regions. In recent work, it was shown that the QES prescription for the entanglement wedge $EW$ of boundary regions must be refined, by replacing the von Neumann entropy with certain one-shot entropies~\cite{Akers:2020pmf}. For static spacetimes, instead of a single entanglement wedge, the refined prescription defines \emph{two} wedges, called the max-EW and the min-EW. They describe the bulk regions that are respectively fully encoded in, and partially influenced by, the boundary region. This refinement has been generalized to the full time-dependent setting~\cite{Akers:2022toappear}. For a boundary input region, the difference between its max- and min-EW is always a quantum effect. However, if one tries to generalize this prescription further to allow for bulk input regions, it appears that in time-dependent settings, the max- and min-EW may differ even at the classical level. We are hopeful that this may resolve the difficulties that we encountered in this section when trying to define a single generalized entanglement wedge with all our desired properties. We leave the details of any such prescription to future work.
\subsection*{Acknowledgements}
We thank Chris Akers, Ben Freivogel, Adam Levine, and Juan Maldacena for valuable discussions. This work was supported in part by the Berkeley Center for Theoretical Physics; by the Department of Energy, Office of Science, Office of High Energy Physics under QuantISED Award DE-SC0019380 and under contract DE-AC02-05CH11231. RB was supported by the National Science Foundation under Award Number 2112880. GP was supported by the Simons Foundation through the ``It from Qubit'' program; by AFOSR award FA9550-22-1-0098; and by an IBM Einstein Fellowship at the Institute for Advanced Study.
\bibliographystyle{JHEP}
|
\section{Details for Simulation and Additional Numerical Results}
\section*{Appendix}
\section{Proof of the theoretical results}\label{sec:app:proof}
\subsection{Proof of Theorem \ref{thm:1}}
\begin{proof}
Since $\mathbf{e}_j ^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}$ is the $j$th component of the solution of $\mathcal{E}(\boldsymbol{\beta}, \widehat \boldsymbol{\alpha}_{\beta_j}, \widehat \boldsymbol{\gamma}_{\beta_j})$, by Lemma \ref{lemma:1} and Lemma \ref{lem:a2} (see Appendix \ref{sec:proof:lem:a1}), we have $\widehat \boldsymbol{\beta}$ is a consistent estimator of $\boldsymbol{\beta}_0$ and
$$\| \widehat\boldsymbol{\beta}-\boldsymbol{\beta}_0 \|_2 = O_p(\sqrt{s \log p/n}) = o_p(n^{-1/4}) \,.$$
since $s = o_p(\sqrt{n}/\log np)$.
We now show the asymptotic normality of $\sqrt{n} (\mathbf{e}_j ^{\scriptscriptstyle \sf T} \widehat{\boldsymbol{\beta}}- \mathbf{e}_j ^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0)$. Let $\bar {h}_j(\mathbf{x}) = \exp(\mathbf{x} ^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{\beta_j})$, $\bar {r}_j(\mathbf{x}) = g(\mathbf{x} ^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{\beta_j})$, $\widehat{h}_j(\mathbf{x}) = \exp(\mathbf{x} ^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\alpha}_{\beta_j})$, and $\widehat{r}_{j}(\mathbf{x}) = g(\mathbf{x} ^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\gamma}_{\beta_j})$. Noting that $\widehat \boldsymbol{\beta}$ is consistent for $\boldsymbol{\beta}_0$, by Theorem 5.21 of \cite{van2000asymptotic}, we have
\begin{equation*}
\begin{aligned}
& \sqrt{n} (\mathbf{e}_j^{\scriptscriptstyle \sf T} \widehat{\boldsymbol{\beta}}- \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0) \\
= &\sqrt{n} \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbf{A}_i \bar{h}_{j}(\mathbf{X}_i) \{ Y_i - \bar{r}_{j}(\mathbf{X}_i)\} + \sqrt{n} \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbf{A}_i \{ \bar{r}_{j}(\mathbf{X}_i) - g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0) \} \\
& + \sqrt{n} \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{e}_j^{\scriptscriptstyle \sf T} (\boldsymbol{\Sigma}_{\breve{\boldsymbol{\beta}}}^{-1}- \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1}) \mathbf{A}_i \bar{h}_{j}(\mathbf{X}_i) \{ Y_i - \bar{r}_{j}(\mathbf{X}_i)\} + \sqrt{n} \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{e}_j^{\scriptscriptstyle \sf T} (\boldsymbol{\Sigma}_{\breve{\boldsymbol{\beta}}}^{-1}- \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1}) \mathbf{A}_i \{ \bar{r}_{j}(\mathbf{X}_i) - g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0) \} \\
& - \sqrt{n} \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbf{A}_i \bar{h}_{j}(\mathbf{X}_i) \{ \widehat{r}_{j}(\mathbf{X}_i) - \bar{r}_{j}(\mathbf{X}_i)\} + \sqrt{n} \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbf{A}_i \{ \widehat{r}_{j}(\mathbf{X}_i) - \bar{r}_{j}(\mathbf{X}_i) \} \\
& - \sqrt{n} \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{e}_j^{\scriptscriptstyle \sf T} (\boldsymbol{\Sigma}_{\breve{\boldsymbol{\beta}}}^{-1}- \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1}) \mathbf{A}_i \bar{h}_{j}(\mathbf{X}_i) \{ \widehat{r}_{j}(\mathbf{X}_i) - \bar{r}_{j}(\mathbf{X}_i)\} + \sqrt{n} \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{e}_j^{\scriptscriptstyle \sf T} (\boldsymbol{\Sigma}_{\breve{\boldsymbol{\beta}}}^{-1}- \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1}) \mathbf{A}_i \{ \widehat{r}_{j}(\mathbf{X}_i) - \bar{r}_{j}(\mathbf{X}_i) \} \\
& + \sqrt{n} \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbf{A}_i \{ \widehat{h}_{j}(\mathbf{X}_i) - \bar{h}_{j}(\mathbf{X}_i) \} \{ Y_i - \bar{r}_{j}(\mathbf{X}_i)\}\\
& - \sqrt{n} \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbf{A}_i \{ \widehat{h}_{j}(\mathbf{X}_i) - \bar{h}_{j}(\mathbf{X}_i) \} \{ \widehat{r}_{j}(\mathbf{X}_i) - \bar{r}_{j}(\mathbf{X}_i)\} \\
& + \sqrt{n} \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{e}_j^{\scriptscriptstyle \sf T} (\boldsymbol{\Sigma}_{\breve{\boldsymbol{\beta}}}^{-1}-\boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1}) \mathbf{A}_i \{ \widehat{h}_{j}(\mathbf{X}_i) - \bar{h}_{j}(\mathbf{X}_i) \} \{ Y_i - \widehat{r}_{j}(\mathbf{X}_i)\} + o_p(1)\\
:= & \mathbf{V}_{ 0 j } + \Delta_{21} + \Delta_{22} + \Delta_{23} + \Delta_{24} + \Delta_{25} + \Delta_{26} + o_p(1) \,,
\label{eq: expan normal beta}
\end{aligned}
\end{equation*}
where $\breve\boldsymbol{\beta}$ lies between $\widehat\boldsymbol{\beta}$ and $\boldsymbol{\beta}_0$. Again using Assumptions \ref{asu:1}, \ref{asu:model} and that $\rho = O(1)$,
$$\mathbf{V}_{0 j} = \sqrt{n} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathcal{E}(\boldsymbol{\beta}_0, \bar \boldsymbol{\alpha}_{\beta_j}, \bar \boldsymbol{\gamma}_{\beta_j}) = \frac{1}{\sqrt{n}}\sum_{i=1}^n \mathbf{e}_j^{\scriptscriptstyle \sf T} \mathcal{W}_{{\scriptscriptstyle \Ssc},\beta_j}(\mathbf{D}_i) + \frac{\rho}{\sqrt{N}}\sum_{i=n+1}^{n+N} \mathbf{e}_j^{\scriptscriptstyle \sf T} \mathcal{W}_{{\scriptscriptstyle \Tsc},\beta_j}(\mathbf{D}_i)$$
converges to $N(0,\sigma_j^2)$ in distribution by central limit theory where
\begin{align*}
\mathcal{W}_{{\scriptscriptstyle \Ssc},\beta_j}(\mathbf{D}_i)&= \boldsymbol{\Sigma}_{\bar{\boldsymbol{\beta}}}^{-1} \mathbf{A}_i \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\alpha}_{\beta_j}) \{ Y_i - g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{\beta_j})\} \text{ and }
\mathcal{W}_{{\scriptscriptstyle \Tsc},\beta_j}(\mathbf{D}_i) = \boldsymbol{\Sigma}_{\bar{\boldsymbol{\beta}}}^{-1} \mathbf{A}_i \{g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{\beta_j}) - g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0) \} \,,
\end{align*}
and $\sigma_j^2 = {\rm Var}\big(\mathbf{e}_j^{\scriptscriptstyle \sf T} \mathcal{W}_{{\scriptscriptstyle \Ssc},\beta_j}(\mathbf{D}_i)\big) + \rho^2 {\rm Var}\big(\mathbf{e}_j^{\scriptscriptstyle \sf T} \mathcal{W}_{{\scriptscriptstyle \Tsc},\beta_j}(\mathbf{D}_i)\big)$.
Then we consider the remaining terms separately. First, we have
\begin{align*}
\|\boldsymbol{\Sigma}_{\breve{\boldsymbol{\beta}}}^{-1}- \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1}\| & = \|\boldsymbol{\Sigma}_{\breve{\boldsymbol{\beta}}}^{-1}\| \|\boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1}\| \|\boldsymbol{\Sigma}_{\boldsymbol{\beta}_0} -\boldsymbol{\Sigma}_{\breve{\boldsymbol{\beta}}} \| = O_p( \|\boldsymbol{\beta}_0 - \widehat\boldsymbol{\beta}\|_2 )
= O_p(\sqrt{s \log p/n}) \,,
\end{align*}
where the second equality comes from the boundness of $\|\boldsymbol{\Sigma}_{\breve{\boldsymbol{\beta}}}^{-1}\| \|\boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1}\|$.
Then we have
\begin{align*}
|\Delta_{21}| & = \sqrt{n} |\mathbf{e}_j^{\scriptscriptstyle \sf T} (\boldsymbol{\Sigma}_{\breve{\boldsymbol{\beta}}}^{-1}- \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1}) \mathcal{E}(\boldsymbol{\beta}_0, \bar \boldsymbol{\alpha}_{\beta_j}, \bar \boldsymbol{\gamma}_{\beta_j}) | \leq \sqrt{n} \|\mathbf{e}_j\|_2 \|\boldsymbol{\Sigma}_{\breve{\boldsymbol{\beta}}}^{-1}- \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1}\| \| \mathcal{E}(\boldsymbol{\beta}_0, \bar \boldsymbol{\alpha}_{\beta_j}, \bar \boldsymbol{\gamma}_{\beta_j})\|_2 \\
& =\sqrt{n} O_p(\sqrt{s \log p/n}) O_p (1/\sqrt{n}) = o_p(1).
\end{align*}
By the moment conditions and the concentration inequality, we have
$$\| \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\bar \boldsymbol{\beta}}^{-1} \mathbf{A}_i \{ Y_i - \bar{r}_{j}(\mathbf{X}_i)\} \bar{h}_{j}(\mathbf{X}_i) \mathbf{X}_i \|_{\infty} \leq \phi\lambda_{\beta_j,\alpha} = O_p ( \sqrt{\log p / n} )$$
and
$$ \| \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbf{A}_i \bar{h}_{j}(\mathbf{X}_i) \dot g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{\beta_j})\mathbf{X}_i^{\scriptscriptstyle \sf T} - \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1}\mathbf{A}_i \dot g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{\beta_j})\mathbf{X}_i \|_{\infty} \leq \phi\lambda_{\beta_j,\gamma} = O_p ( \sqrt{\log p / n}).$$
Then
\begin{align*}
|\Delta_{22}| & \leq C \sqrt{n} |\widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbf{A}_i \bar{h}_{j}(\mathbf{X}_i) \dot g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{\beta_j})\mathbf{X}_i^{\scriptscriptstyle \sf T} ( \widehat \boldsymbol{\gamma}_{\beta_j} - \bar{\boldsymbol{\gamma}}_{\beta_j} ) - \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbf{A}_i \dot g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{\beta_j})\mathbf{X}_i^{\scriptscriptstyle \sf T} ( \widehat \boldsymbol{\gamma}_{\beta_j} - \bar{\boldsymbol{\gamma}}_{\beta_j})| \\
& \leq C \sqrt{n} \|\widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbf{A}_i \bar{h}_{j}(\mathbf{X}_i) \dot g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{\beta_j})\mathbf{X}_i - \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbf{A}_i \dot g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{\beta_j})\mathbf{X}_i \|_{\infty} \| \widehat \boldsymbol{\gamma}_{\beta_j} - \bar{\boldsymbol{\gamma}}_{\beta_j} \|_1 \\
& = \sqrt{n} O_p(\sqrt{\log p/n}) o_p \big( (\log~p)^{-1/2} \big) = o_p(1); \\
|\Delta_{23}| & = \sqrt{n} \Big| \mathbf{e}_j^{\scriptscriptstyle \sf T} (\boldsymbol{\Sigma}_{\breve{\boldsymbol{\beta}}}^{-1}- \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1}) \big\{ \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{A}_i \bar{h}_{j}(\mathbf{X}_i) \{ \widehat{r}_j(\mathbf{X}_i) - \bar{r}_j(\mathbf{X}_i)\} - \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{A}_i \{ \widehat{r}_j(\mathbf{X}_i) - \bar{r}_j(\mathbf{X}_i)\} \big\} \Big| \\
& \leq \sqrt{n} \|\mathbf{e}_j\|_2 \|\boldsymbol{\Sigma}_{\breve{\boldsymbol{\beta}}}^{-1}- \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1}\| \big\| \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{A}_i \bar{h}_j(\mathbf{X}_i) \{ \widehat{r}_j(\mathbf{X}_i) - \bar{r}_j(\mathbf{X}_i)\} - \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{A}_i \{ \widehat{r}_j(\mathbf{X}_i) - \bar{r}_j(\mathbf{X}_i)\} \big\|_2 \\
& \leq C \sqrt{n} \|\mathbf{e}_j\|_2 \|\boldsymbol{\Sigma}_{\breve{\boldsymbol{\beta}}}^{-1}- \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1}\| \Big( \big\| \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{A}_i \bar{h}_j(\mathbf{X}_i) \mathbf{X}_i^{\scriptscriptstyle \sf T} ( \widehat \boldsymbol{\gamma}_{\beta_j} - \bar \boldsymbol{\gamma}_{\beta_j})\big\|_2 + \big\| \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{A}_i \mathbf{X}_i^{\scriptscriptstyle \sf T} ( \widehat \boldsymbol{\gamma}_{\beta_j} - \bar \boldsymbol{\gamma}_{\beta_j} ) \big\|_2 \Big) \\
& \leq C' \sqrt{n} \|\boldsymbol{\Sigma}_{\breve{\boldsymbol{\beta}}}^{-1}- \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1}\| \Big( \big[\widehat\Psc_{\scriptscriptstyle \Ssc}\{\mathbf{X}_i^{\scriptscriptstyle \sf T}(\widehat \boldsymbol{\gamma}_{\beta_j} - \bar\boldsymbol{\gamma}_{\beta_j})\}^2\big]^{1/2} + \big[\widehat\Psc_{\scriptscriptstyle \Tsc}\{\mathbf{X}_i^{\scriptscriptstyle \sf T}(\widehat \boldsymbol{\gamma}_{\beta_j} - \bar\boldsymbol{\gamma}_{\beta_j})\}^2\big]^{1/2} \Big) \\
& = \sqrt{n} O_p(\sqrt{s \log p/n}) \Big( \big[\widehat\Psc_{\scriptscriptstyle \Ssc} (\widehat \boldsymbol{\gamma}_{\beta_j} - \bar\boldsymbol{\gamma}_{\beta_j}) \mathbf{X}_i \mathbf{X}_i^{\scriptscriptstyle \sf T}(\widehat \boldsymbol{\gamma}_{\beta_j} - \bar\boldsymbol{\gamma}_{\beta_j})\big]^{1/2} + \big[\widehat\Psc_{\scriptscriptstyle \Tsc} (\widehat \boldsymbol{\gamma}_{\beta_j} - \bar\boldsymbol{\gamma}_{\beta_j}) \mathbf{X}_i \mathbf{X}_i^{\scriptscriptstyle \sf T}(\widehat \boldsymbol{\gamma}_{\beta_j} - \bar\boldsymbol{\gamma}_{\beta_j}) \big]^{1/2} \Big) \\
& = O_p(\sqrt{s \log p}) \|\widehat \boldsymbol{\gamma}_{\beta_j} - \bar\boldsymbol{\gamma}_{\beta_j}\|_2 ( \| \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{X}_i \mathbf{X}_i \|_{\rm op}^{1/2} + \| \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{X}_i \mathbf{X}_i \|_{\rm op}^{1/2}) \\
& = O_p(\sqrt{s \log p}) O_p( \sqrt{s \log p/n} )
= o_p(1); \\
|\Delta_{24}| & \leq C \sqrt{n} \big|\widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbf{A}_i \{ Y_i - \bar{r}_{j}(\mathbf{X}_i)\} \bar{h}_{j}(\mathbf{X}_i) \mathbf{X}_i^{\scriptscriptstyle \sf T} (\widehat \boldsymbol{\alpha}_{\beta_j} - \bar\boldsymbol{\alpha}_{\beta_j}) \big| \\
& \leq C \sqrt{n} \big \|\widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbf{A}_i \{ Y_i - \bar{r}_{j}(\mathbf{X}_i)\} \bar{h}_{j}(\mathbf{X}_i) \mathbf{X}_i \big\|_{\infty} \|\widehat \boldsymbol{\alpha}_{\beta_j} - \bar\boldsymbol{\alpha}_{\beta_j}\|_1 \\
& = \sqrt{n} O_p(\sqrt{\log p/n}) o_p \big( (\log~p)^{-1/2} \big) = o_p(1); \\
|\Delta_{25}|& = \sqrt{n} \big| \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbf{A}_i \{ \widehat{h}_{j}(\mathbf{X}_i) - \bar{h}_{j}(\mathbf{X}_i) \} \{ \widehat{r}_{j}(\mathbf{X}_i) - \bar{r}_{j}(\mathbf{X}_i)\} \big| \\
& \leq \sqrt{n} \| \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1}\|_2 \|\Delta_a\|_2 = \sqrt{n} o_p(n^{-1/2}) = o_p(1); \\
|\Delta_{26}| & \leq C \sqrt{n} \big \| \boldsymbol{\Sigma}_{\breve{\boldsymbol{\beta}}}^{-1}-\boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \big \| \big[\widehat\Psc_{\scriptscriptstyle \Ssc}\{\mathbf{X}_i^{\scriptscriptstyle \sf T}(\widehat \boldsymbol{\alpha}_{\beta_j} - \bar\boldsymbol{\alpha}_{\beta_j})\}^2\big]^{1/2} \\
& = \sqrt{n} O_p(\sqrt{s \log p/n}) O_p(\sqrt{s \log p/n}) = o_p(1).
\end{align*}
As a result,
$\sqrt{n} (\mathbf{e}_j^{\scriptscriptstyle \sf T} \widehat{\boldsymbol{\beta}}- \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0) = \mathbf{V}_{0j} + o_p(1)$ weakly converges to $N(0,\sigma_j^2)$.
\end{proof}
\subsection{Proof of Theorem \ref{thm:2}}
\label{proof thm2}
\begin{proof}
Before showing the consistency of $\widehat \ROC$ and $\widehat \AUC$, we show the consistency of $\widehat \TPR(c)$ and $\widehat \mathbf{FPR}(c)$ first. Recall that
$$
\widehat \TPR(c) = \widehat \TPR(c; \widehat \boldsymbol{\beta}, \widehat\boldsymbol{\alpha}_{{\sf roc}, t(c)}, \widehat\boldsymbol{\gamma}_{{\sf roc}, t(c)}, \widehat\boldsymbol{\alpha}_{{\sf roc}, -\infty}, \widehat\boldsymbol{\gamma}_{, {\sf roc},-\infty}) = \frac{\widehat \TP(c; \widehat \boldsymbol{\beta}, \widehat\boldsymbol{\alpha}_{{\sf roc}, t(c)}, \widehat\boldsymbol{\gamma}_{{\sf roc}, t(c)})}{\widehat \TP(-\infty; \widehat \boldsymbol{\beta}, \widehat\boldsymbol{\alpha}_{{\sf roc}, -\infty}, \widehat\boldsymbol{\gamma}_{{\sf roc}, -\infty})} \,,
$$
where
\begin{align*}
\widehat \TP(c; \widehat \boldsymbol{\beta}, \widehat\boldsymbol{\alpha}_{{\sf roc}, t(c)}, \widehat\boldsymbol{\gamma}_{{\sf roc}, t(c)}) = & \widehat\Psc_{\scriptscriptstyle \Ssc}\mathbb{I}( \mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq c) \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\alpha}_{{\sf roc}, t(c)}) \{Y_i-g(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\gamma}_{{\sf roc}, t(c)})\} \\
& + \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) g(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\gamma}_{{\sf roc}, t(c)}) \,.
\end{align*}
For notional simplicity, we let
$\widehat \boldsymbol{\theta}_{{\sf roc}, t(c)} = (\widehat\boldsymbol{\alpha}_{{\sf roc}, t(c)} ^{\scriptscriptstyle \sf T}, \widehat\boldsymbol{\gamma}_{{\sf roc}, t(c)}^{\scriptscriptstyle \sf T})^{\scriptscriptstyle \sf T}$ and $\bar \boldsymbol{\theta}_{{\sf roc},t(c)} = (\bar \boldsymbol{\alpha}_{{\sf roc}, t(c)} ^{\scriptscriptstyle \sf T}, \bar \boldsymbol{\gamma}_{{\sf roc}, t(c)}^{\scriptscriptstyle \sf T})^{\scriptscriptstyle \sf T}$. Recall the definitions of $\TP$ and $\TPR$ are
\[
\TP(c,\boldsymbol{\beta}_0) = \mathbb{P}_{\scriptscriptstyle \Tsc}(Y=1,\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0 \geq c) \text{ and } \TPR(c,\boldsymbol{\beta}_0) = \frac{\TP(c,\boldsymbol{\beta}_0)}{\TP(-\infty,\boldsymbol{\beta}_0)} = \mathbb{P}_{\scriptscriptstyle \Tsc}(\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0 \geq c \mid Y=1)\,,
\]
where we include $\boldsymbol{\beta}_0$ as the parameter of $\TP(\cdot,\cdot)$ to reflect the dependence of $\TP$ with $\boldsymbol{\beta}_0$. Then we expand $\widehat \TP(c; \widehat \boldsymbol{\beta}, \widehat \boldsymbol{\theta}_{{\sf roc}, t(c)})$ as $\widehat \TP(c; \widehat \boldsymbol{\beta}, \widehat \boldsymbol{\theta}_{{\sf roc}, t(c)}) =$
\begin{equation*}
\begin{aligned}
& \widehat\Psc_{\scriptscriptstyle \Ssc}\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},c}) \{Y_i-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},c})\} + \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},c}) \\
& + \widehat\Psc_{\scriptscriptstyle \Ssc}\{ \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) \} \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \{Y_i-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \\
& + \widehat\Psc_{\scriptscriptstyle \Tsc} \{ \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) \} g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)}) \\
& - \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq t(c) ) \exp(X_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \{g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}) -g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \\
& + \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq t(c)) \{ g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}) - g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \\
& + \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq t(c)) \{ \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\alpha}_{{\sf roc},t(c)}) -\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \} \{Y_i-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \\
& - \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) \{ \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\alpha}_{{\sf roc},t(c)}) -\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \} \{g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}) - g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \\
& - \widehat\Psc_{\scriptscriptstyle \Ssc}\{ \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) \} \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}\} \{g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}) - g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \\
& + \widehat\Psc_{\scriptscriptstyle \Ssc} \{ \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) \} \{ \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\alpha}_{{\sf roc},t(c)}) -\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \} \{Y_i-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \\
& + \widehat\Psc_{\scriptscriptstyle \Tsc} \{ \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) \} \{ g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}) - g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \\
& - \widehat\Psc_{\scriptscriptstyle \Ssc} \{\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq t(c)) \} \exp(X_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \{g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}) -g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \\
& + \widehat\Psc_{\scriptscriptstyle \Tsc} \{\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq t(c)) \} \{ g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}) - g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \\
& + \widehat\Psc_{\scriptscriptstyle \Ssc} \{\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq t(c)) \} \{ \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\alpha}_{{\sf roc},t(c)}) -\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \} \{Y_i-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \\
& + {\widehat \TP (c, \boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},t(c)}) - \widehat \TP (c, \boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c})}\\
:= & \widehat \TP (c, \boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c}) + U_{31} + U_{32} + \Delta_{31} + \Delta_{32} + \Delta_{33} + \Delta_{34} + \Delta_{35} + \Delta_{36} + \Delta_{37} + V_{31} + V_{32} + V_{33} + W.
\end{aligned}
\end{equation*}
Under Assumption \ref{asu:model}, we have $$\mathbb{E} \widehat \TP (c, \boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c}) = \TP(c,\boldsymbol{\beta}_0) = \mathbb{P}_{\scriptscriptstyle \Tsc}(Y=1,\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0 \geq c)\,.$$
From standard empirical process theory \citep{pollard1990empirical}, one may show that
$$\sup_{\mathbf{X},\boldsymbol{\beta}} |\widehat \TP (c, \boldsymbol{\beta}, \bar \boldsymbol{\theta}_{{\sf roc},c}) - \TP (c, \boldsymbol{\beta})| = O_p(N^{-1/2})$$
and $\sqrt{N}\{ \widehat \TP (c, \boldsymbol{\beta}, \bar \boldsymbol{\theta}_{{\sf roc}, c}) - \TP (c, \boldsymbol{\beta}) \}$ converges weakly to a zero-mean Gaussian process in $(c, \boldsymbol{\beta})$. In addition, to establish the asymptotic distribution of $U_{31}$ and $U_{32}$, we verify that there exists $\delta >0$, such that the classes of
functions indexed by $\boldsymbol{\beta}$ and $c$
$$\mathcal{B}_1 = \left \{ \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \{Y_i-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\}: \|\boldsymbol{\beta} - \boldsymbol{\beta}_0 \|_2 \leq \delta \right\}$$
$${\rm and} \; \mathcal{B}_2 = \left \{ \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)}) : \|\boldsymbol{\beta} - \boldsymbol{\beta}_0 \|_2 \leq \delta \right\}$$
are Donsker classes. We first note that $\left \{ \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) : \|\boldsymbol{\beta} - \boldsymbol{\beta}_0 \|_2 \leq \delta \right\}$ is a Vapnik-Chervonenkis class [\cite{van2000asymptotic}, Page 275]. The functions $\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \{Y_i-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\}$ and $ g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})$ are
continuously difffferentiable and uniformly bounded by constant by Assumption~\ref{asu:1}. It follows that $\mathcal{B}_1$ and $\mathcal{B}_2$ are Donsker classes. Define
$$
\begin{aligned}
& J_1(c,\boldsymbol{\beta}) = \mathbb{E}_{\scriptscriptstyle \Ssc}[\mathbb{I}(\mathbf{A} ^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) \exp(\mathbf{X} ^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \{Y-g(\mathbf{X}^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\}] \\
& \widehat J_1(c,\boldsymbol{\beta}) = \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \{Y_i-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \\
& J_2(c,\boldsymbol{\beta}) = \mathbb{E}_{\scriptscriptstyle \Tsc}[\mathbb{I}(\mathbf{A} ^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) g(\mathbf{X}^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})] \; {\rm and} \; \widehat J_2(c,\boldsymbol{\beta}) = \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)}).
\end{aligned}
$$
Then by Theorem 19.5 of \cite{van2000asymptotic},
$$\sqrt{n} \left \{ \widehat J_1(c,\boldsymbol{\beta}) - J_1(c,\boldsymbol{\beta}) \right\} \; {\rm and} \; \sqrt{N} \left \{ \widehat J_2(c,\boldsymbol{\beta}) - J_2(c,\boldsymbol{\beta}) \right\}$$
converge weakly to a mean zero Gaussian process index by $\boldsymbol{\beta}$. Thus, $\sqrt{n} \left \{ \widehat J_1(c,\boldsymbol{\beta}) - J_1(c,\boldsymbol{\beta}) \right\}$ and $\sqrt{N} \left \{ \widehat J_2(c,\boldsymbol{\beta}) - J_2(c,\boldsymbol{\beta}) \right\}$ are stochastically equicontinuous at $\boldsymbol{\beta}$. In addition, note that $J_1(c,\boldsymbol{\beta})$ and $J_2(c,\boldsymbol{\beta})$ are continuously differentiable at $\boldsymbol{\beta}$. It then follows that
\begin{align*}
& \sqrt{n} \big \{ \widehat J_1(c,\widehat \boldsymbol{\beta}) - \widehat J_1(c,\boldsymbol{\beta}_0) \big\} \\
& = \sqrt{n} \big \{ \widehat J_1(c,\widehat \boldsymbol{\beta}) - J_1(c,\widehat \boldsymbol{\beta}) \big\} - \sqrt{n} \big \{ \widehat J_1(c,\boldsymbol{\beta}_0) - J_1(c,\boldsymbol{\beta}_0) \big\} + \sqrt{n} \big \{ J_1(c,\widehat \boldsymbol{\beta}) - J_1(c,\boldsymbol{\beta}_0) \big\} \\
& = \sqrt{n} \big \{ \widehat J_1(c,\boldsymbol{\beta}_0) - J_1(c,\boldsymbol{\beta}_0) \big\} - \sqrt{n} \big \{ \widehat J_1(c,\boldsymbol{\beta}_0) - J_1(c,\boldsymbol{\beta}_0) \big\} + \sqrt{n} \big \{ J_1(c,\widehat \boldsymbol{\beta}) - J_1(c,\boldsymbol{\beta}_0) \big\} + o_p(1) \\
& = \sqrt{n} \dot J_1(c,\boldsymbol{\beta}_0)^{\scriptscriptstyle \sf T} (\widehat \boldsymbol{\beta} - \boldsymbol{\beta}_0) + o_p(1),
\end{align*}
and
$$
\sqrt{N} \big \{ \widehat J_2(c,\widehat \boldsymbol{\beta}) - \widehat J_2(c,\boldsymbol{\beta}_0) \big\}
= \sqrt{N} \dot J_2(c,\boldsymbol{\beta}_0)^{\scriptscriptstyle \sf T} (\widehat \boldsymbol{\beta} - \boldsymbol{\beta}_0) + o_p(1)
$$
converge in distribution to Gaussian distribution with mean zero where $\dot J_i(c,\cdot)$ is the derivative of $ J_i(c,\cdot)$ with respect to $\boldsymbol{\beta}$.
Since $n\asymp N$, we have
$\sqrt{n} U_{31} = \sqrt{n} \big \{ \widehat J_1(c,\widehat \boldsymbol{\beta}) - \widehat J_1(c,\boldsymbol{\beta}_0) \big\}$ and $\sqrt{n} U_{32} = \sqrt{n} \big \{ \widehat J_2(c,\widehat \boldsymbol{\beta}) - \widehat J_2(c,\boldsymbol{\beta}_0) \big\}$ converge to mean zero Gaussian distribution.
We then control the remaining terms $\Delta_{3i},i=1,\cdots,6$ and $V_{3i},i=1,2,3$, separately.
Using the first order Taylor's expansion,
for $j=1, \dots, m$,
we have
\begin{equation*}
\begin{aligned}
|\Delta_{31} + \Delta_{32}| & \leq C \big\| \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) \exp(X_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \dot{g}(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)}) \mathbf{X}_i^{\scriptscriptstyle \sf T} - \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) \dot{g}(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)}) \mathbf{X}_i^{\scriptscriptstyle \sf T} \big \|_{\infty} \\
&\times \| \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}- \bar \boldsymbol{\gamma}_{{\sf roc},t(c)}\|_1 \\
& \leq C \mathcal{M}_{{\sf roc},\alpha}(\bar\boldsymbol{\alpha}_{{\sf roc},t(c)}; t(c)) \| \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}- \bar \boldsymbol{\gamma}_{{\sf roc},t(c)}\|_1 \\
& = O_p( \lambda_{{\sf roc},\alpha} \| \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}- \bar \boldsymbol{\gamma}_{{\sf roc},t(c)}\|_1 ) \\
& = O_p\left[n_j/n \sqrt{ \log~np /n_j} \{ s \sqrt{\log p/n} + R_{{\sf roc}, r}(\log ~np/n_j)^{1/2- r/2}\} \right]= o_p (n^{-1/2}),
\end{aligned}
\end{equation*}
where the last inequality comes from the moment condition and the last equality comes from
Assumption \ref{asu:5}, which implies $R_{{\sf roc}, r} = o\{ n^{1/2 - r/2} (\log~np )^{r/2-1} \}$ and
$s = o \{n^{1/2}(\log~np)^{-1/2}(\log~p)^{-1/2}\}$.
Similarly, for $j=1, \dots, m$, we have
\begin{equation*}
\begin{aligned}
& |\Delta_{33}| \leq C \big | \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq t(c)) \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \{Y_i-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \mathbf{X}_i^{\scriptscriptstyle \sf T} (\widehat \boldsymbol{\alpha}_{{\sf roc},t(c)} - \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \big| \\
& \leq C \big \| \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq t(c)) \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \{Y_i-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \mathbf{X}_i \big \|_{\infty} \| \widehat \boldsymbol{\alpha}_{{\sf roc},t(c)} - \bar \boldsymbol{\alpha}_{{\sf roc},t(c)} \|_1 \\
& \leq C \mathcal{M}_{{\sf roc},\gamma}(\bar\boldsymbol{\gamma}_{{\sf roc},t(c)}; t(c)) \| \widehat \boldsymbol{\alpha}_{{\sf roc},t(c)} - \bar \boldsymbol{\alpha}_{{\sf roc},t(c)} \|_1\\
& = O_p( \lambda_{{\sf roc},\gamma} \| \widehat \boldsymbol{\alpha}_{{\sf roc},t(c)} - \bar \boldsymbol{\alpha}_{{\sf roc},t(c)} \|_1) = o_p (n^{-1/2}).
\end{aligned}
\end{equation*}
Besides, for $j=1, \dots, m$,
\begin{equation*}
\begin{aligned}
& |\Delta_{34}| \leq C \big | \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq t(c)) \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \mathbf{X}_i^{\scriptscriptstyle \sf T}( \widehat \boldsymbol{\alpha}_{{\sf roc},t(c)} - \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \dot{g}(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)}) \mathbf{X}_i^{\scriptscriptstyle \sf T}( \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)} - \bar \boldsymbol{\gamma}_{{\sf roc},t(c)} ) \big| \\
& \leq C \big \| \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq t(c)) \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \dot{g}(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)}) \mathbf{X}_i \mathbf{X}_i^{\scriptscriptstyle \sf T} \big \|
\| \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)} - \bar \boldsymbol{\gamma}_{{\sf roc},t(c)} \|_2
\| \widehat \boldsymbol{\alpha}_{{\sf roc},t(c)} - \bar \boldsymbol{\alpha}_{{\sf roc},t(c)} \|_2 \\
& \leq C' \| \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)} - \bar \boldsymbol{\gamma}_{{\sf roc},t(c)} \|_2
\| \widehat \boldsymbol{\alpha}_{{\sf roc},t(c)} - \bar \boldsymbol{\alpha}_{{\sf roc},t(c)} \|_2 \\
& = O_p \left[\sqrt{ s \log p/n \left\{ s \log p/n + R_{{\sf roc}, r}(\log ~np/n_j )^{1- r/2}\right\}} \right]
= o_p (n^{-1/2}) \,,
\end{aligned}
\end{equation*}
where the third inequality comes from the boundness of
{ eigenvalues of $\mathbb{E}_{\scriptscriptstyle \Ssc}[ \mathbf{X}_i \mathbf{X}_i^{\scriptscriptstyle \sf T}]$ given in Assumption \ref{asu:1}} and the last equality comes from
Assumption \ref{asu:5} which implies that
$s = o\big( n^{1/2} (\log~p)^{-1} \big)$ and
$s R_{{\sf roc}, r} = o\big( (n_{\min }/\log ~np )^{1- r/2} /\log p \big)$.
Before we proceed, let us figure out the asymptotic property of the function $\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c)$. By Theorem \ref{thm:1}, we know that $\|\widehat \boldsymbol{\beta} - \boldsymbol{\beta}_0\|_2 = O_p(1/\sqrt{n})$. By the boundness of $\mathbf{A}$ from Assumption~\ref{asu:1}, we have $|\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} - \mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0| = O_p(1/\sqrt{n})$ and there exists some $0<\delta=O_p(1/\sqrt{n})$ such that
$$
\begin{aligned}
|\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c)| & \leq \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c \geq \mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0) + {\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\leq c \leq \mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0)} \\
& \leq 2 \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 - \delta \leq c \leq \mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 + \delta).
\end{aligned}
$$
We then have
$$\mathbb{E} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 - \delta \leq c \leq \mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 + \delta) = \mathbb{P}(c - \delta \leq\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \leq c+ \delta) = O_p(\delta) = O_p(1/\sqrt{n}) \,,$$
where the second inequality comes from Assumption \ref{asu:1} that probability density of $\mathbf{A}_i^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0$ is continuously differentiable. Then by Central Limit Theorem we have
$$\widehat\Psc_{\scriptscriptstyle \Ssc}| \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) | = O_p(1/\sqrt{n}) \; {\rm and} \; \widehat\Psc_{\scriptscriptstyle \Tsc}| \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c)| = O_p(1/\sqrt{n}).$$
We then control $\Delta_{3i},i=5,6,7$ as follows.
\begin{align*}
|\Delta_{35}| & = \big | \widehat\Psc_{\scriptscriptstyle \Tsc} \{ \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) \} \{ g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}) - g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \big | \\
& \leq \widehat\Psc_{\scriptscriptstyle \Tsc} | \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) | \max_{i=1,\ldots, n} | g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}) - g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})| \\
& \leq C \widehat\Psc_{\scriptscriptstyle \Tsc} | \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) | \max_{i=1,\ldots, n} \|\mathbf{X}_i\|_{\infty} \| \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}- \boldsymbol{\gamma}_{{\sf roc},t(c)}\|_1 \\
& = O_p ( n^{-1/2} \| \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}- \boldsymbol{\gamma}_{{\sf roc},t(c)}\|_1 ) \\
& = O_p\big( n^{-1/2} ( s \sqrt{\log ~p/n} + R_{{\sf roc}, r}(\log ~np/n_j )^{1/2- r/2}) \big) = o_p(n^{-1/2}),
\end{align*}
where the last equality comes from
Assumption \ref{asu:5}, which implies that
$s = o\left\{ (n /\log~p)^{1/2} \right\}$ and
$R_{{\sf roc}, r} =o\left\{ (n_{\min}/\log ~np )^{1/2- r/2} \right\}$.
Similarly, we have
\begin{align*}
|\Delta_{36}| & = \big | \widehat\Psc_{\scriptscriptstyle \Ssc} \{ \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) \} \{ \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\alpha}_{{\sf roc},t(c)}) -\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \} \{Y_i-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \big | \\
& \leq C \widehat\Psc_{\scriptscriptstyle \Ssc} \big | \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) \big | \max_{i=n+1,\ldots, N} \|\mathbf{X}_i\|_{\infty} \| \widehat \boldsymbol{\alpha}_{{\sf roc},t(c)} - \bar \boldsymbol{\alpha}_{{\sf roc},t(c)} \|_1 \\
& = O_p ( n^{-1/2} \| \widehat \boldsymbol{\alpha}_{{\sf roc},t(c)} - \bar \boldsymbol{\alpha}_{{\sf roc},t(c)} \|_1 ) = o_p(n^{-1/2})\,; \\
|\Delta_{37}| & = \big | \widehat\Psc_{\scriptscriptstyle \Ssc}\{ \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) \} \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}\} \{g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}) - g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \big| \\
& \leq C \widehat\Psc_{\scriptscriptstyle \Ssc} \big | \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) \big | \max_{i=n+1,\ldots, N} \|\mathbf{X}_i\|_{\infty} \| \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)} - \bar \boldsymbol{\gamma}_{{\sf roc},t(c)} \|_1 \\
& = O_p ( n^{-1/2} \| \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}- \boldsymbol{\gamma}_{{\sf roc},t(c)}\|_1 ) = o_p(n^{-1/2}).
\end{align*}
Finally, we need to control $V_{3i},i=1,2,3$. We observe that
$$|\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq t(c))| = \mathbb{I}(\max\{c,t(c)\} > \mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq \min\{c,t(c)\} )$$
and
\begin{align*}
\mathbb{E}|\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq t(c))| & = \mathbb{P}(\max\{c,t(c)\} > \mathbf{A}^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq \min\{c,t(c)\} )\\
& = O_p( \max\{n_{\min}, (n-n_{\min})/(m-1) \}/n ) = O_p(n_{\min}/n )
\end{align*}
by the fact that $m = \lceil n/n_{\min} \rceil$ and
Assumption \ref{asu:1} that probability density of $\mathbf{A} ^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0$ is continuously differentiable.
Then by Central Limit Theorem we have
$$\widehat\Psc_{\scriptscriptstyle \Ssc} |\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq t(c))| = O_p(n_{\min}/n )\,.$$
Hence,
\begin{equation*}
\begin{aligned}
|V_{31}| & = \big|\widehat\Psc_{\scriptscriptstyle \Ssc} \{\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq t(c)) \} \exp(X_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},t(c)}) \{g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\gamma}_{{\sf roc},t(c)}) -g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},t(c)})\} \big| \\
& \leq C \widehat\Psc_{\scriptscriptstyle \Ssc} |\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq t(c))| \|\mathbf{X}_i\|_{\infty} \|\widehat \boldsymbol{\gamma}_{{\sf roc},t(c)} -\bar \boldsymbol{\gamma}_{{\sf roc},t(c)}\|_1 \\
&\leq C \widehat\Psc_{\scriptscriptstyle \Ssc} |\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq t(c))| \max_{i=1,\cdots,n}\|\mathbf{X}_i\|_{\infty} \|\widehat \boldsymbol{\gamma}_{{\sf roc},t(c)} - \bar \boldsymbol{\gamma}_{{\sf roc},t(c)}\|_1 \\
& = O_p( n_{\min} \|\widehat \boldsymbol{\gamma}_{{\sf roc},t(c)} - \bar \boldsymbol{\gamma}_{{\sf roc},t(c)}\|_1 /n) \\
& = O_p\big( n_{\min} ( s \sqrt{\log ~p/n} + R_{{\sf roc}, r}(\log ~np/n_j )^{1/2- r/2})/n \big) = o_p(n^{-1/2})
\end{aligned}
\end{equation*}
by Assumption \ref{asu:5}, which implies
$s = o\left\{n/(n_{\min} \sqrt{\log p}) \right\}$ and
$ R_{{\sf roc}, r} = o \left\{ n^{1/2} n_{\min}^{-1/2-r/2} (\log ~np)^{-1/2+ r/2} \right\}$.
Similarly, we have $V_{32} = o_p(n^{-1/2})$ and $V_{33} = o_p(n^{-1/2})$.
{
We then utilize the doubly robust property to analyze the term $W$.
When the imputation model is correctly specified, $\bar\boldsymbol{\gamma}_{{\sf roc}, t(c)} = \bar\boldsymbol{\gamma}_{{\sf roc}, c}$, $\forall c$,
\begin{align*}
W &= \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0\geq c)\{Y_i - g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{{\sf roc}, c})\} \{\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc}, t(c)}) - \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\alpha}_{{\sf roc}, c}) \},
\end{align*}
and $\mathbb{E}(W) = 0$.
Further, we have
\begin{align*}
{\rm Var}(W) = \frac{1}{n} (\bar\boldsymbol{\alpha}_{{\sf roc}, t(c)} - \bar\boldsymbol{\alpha}_{{\sf roc}, c} )^{\scriptscriptstyle \sf T} \mathbb{E} _{\scriptscriptstyle \Ssc}\left[
\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0\geq c)\{Y_i - g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{{\sf roc}, c})\}^2 \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\alpha}_{{\sf roc}, c}^\prime)^2 \mathbf{X}_i\mathbf{X}_i^{\scriptscriptstyle \sf T}
\right] (\bar\boldsymbol{\alpha}_{{\sf roc}, t(c)} - \bar\boldsymbol{\alpha}_{{\sf roc}, c}),
\end{align*}
where $\bar\boldsymbol{\alpha}_{{\sf roc}, c}^\prime$ is on the line connecting $\bar\boldsymbol{\alpha}_{{\sf roc}, t(c)}$ and $ \bar\boldsymbol{\alpha}_{{\sf roc}, c}$.
By Assumption \ref{asu:3}, we have $\|\bar\boldsymbol{\alpha}_{{\sf roc}, c}^\prime\|_1$ is bounded, and by Assumption \ref{asu:1}, we have $\lambda_{\max} \left(\mathbb{E} _{\scriptscriptstyle \Ssc}\left[
\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0\geq c)\{Y_i - g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{{\sf roc}, c})\}^2 \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\alpha}_{{\sf roc}, c}^\prime)^2 \mathbf{X}_i\mathbf{X}_i^{\scriptscriptstyle \sf T}
\right] \right)$ is bounded.
Hence, by Lemma \ref{lem:continuity}, we have $\sqrt{ {\rm Var}(W)} \precsim (nm)^{-1/2}$, and consequently by Chebyshev's inequality, we have $\sqrt{n}W = o_p(1)$.
When the density ratio model is correctly specified, then
$\bar\boldsymbol{\alpha}_{{\sf roc}, t(c)}= \bar\boldsymbol{\alpha}_{{\sf roc}, c}$, $\forall c$,
\begin{align*}
W = & \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0\geq c)\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\alpha}_{{\sf roc}, c})
\{g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{{\sf roc}, c}) - g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{{\sf roc}, t(c)})\}
\\
& +
\widehat\Psc_{\scriptscriptstyle \Tsc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0\geq c)
\{g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{{\sf roc}, t(c)}) - g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{{\sf roc}, c})\},
\end{align*}
and $\mathbb{E}(W) = 0$.
Further, we have
\begin{align*}
{\rm Var}(W)
=& \frac{1}{n} (\bar\boldsymbol{\gamma}_{{\sf roc}, t(c)} - \bar\boldsymbol{\gamma}_{{\sf roc}, c} )^{\scriptscriptstyle \sf T}\mathbb{E} _{\scriptscriptstyle \Ssc}\left\{
\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0\geq c) \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\alpha}_{{\sf roc}, c})^2
\dot{g}(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{{\sf roc}, c}^\prime)^2
\mathbf{X}_i\mathbf{X}_i^{\scriptscriptstyle \sf T}
\right\} (\bar\boldsymbol{\gamma}_{{\sf roc}, t(c)} - \bar\boldsymbol{\gamma}_{{\sf roc}, c})\\
&+ \frac{1}{N} (\bar\boldsymbol{\gamma}_{{\sf roc}, t(c)} - \bar\boldsymbol{\gamma}_{{\sf roc}, c} )^{\scriptscriptstyle \sf T}
\mathbb{E} _{\scriptscriptstyle \Tsc} \left\{\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0\geq c)
\dot{g}(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{{\sf roc}, c}^\prime)^2 \mathbf{X}_i\mathbf{X}_i^{\scriptscriptstyle \sf T} \right\}
(\bar\boldsymbol{\gamma}_{{\sf roc}, t(c)} - \bar\boldsymbol{\gamma}_{{\sf roc}, c}),
\end{align*}
where $\bar\boldsymbol{\gamma}_{{\sf roc}, c}^\prime$ is on the line connecting $\bar\boldsymbol{\gamma}_{{\sf roc}, t(c)}$ and $ \bar\boldsymbol{\gamma}_{{\sf roc}, c}$.
Similarly, by Assumptions \ref{asu:1} and \ref{asu:3} and Lemma \ref{lem:continuity}, we also have $\sqrt{n}W = o_p(1)$.
}
Then we conclude that
\begin{align*}
& \sqrt{n} \big \{ \widehat \TP (c,\widehat \boldsymbol{\beta}, \widehat \boldsymbol{\theta}_{{\sf roc},t(c)}) - \TP (c, \boldsymbol{\beta}_0) \big\} \\
= & \sqrt{n} \big \{ \widehat \TP (c, \boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c}) - \TP (c, \boldsymbol{\beta}_0) \big \} + \sqrt{n} U_{31} + \sqrt{n} U_{32} + o_p(1) \\
= & \sqrt{n} (\widehat\Psc_{\scriptscriptstyle \Ssc} - \mathbb{E}_{\scriptscriptstyle \Ssc}) \big \{ \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}_{{\sf roc},c}) \{Y_i-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},c})\} \big \} \\
& + \sqrt{n} (\widehat\Psc_{\scriptscriptstyle \Tsc} - \mathbb{E}_{\scriptscriptstyle \Tsc}) \big \{ \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\gamma}_{{\sf roc},c}) \big \} + \sqrt{n} \big( \dot J_1(c,\boldsymbol{\beta}_0) + J_2(c,\boldsymbol{\beta}_0) \big)^{\scriptscriptstyle \sf T} (\widehat \boldsymbol{\beta} - \boldsymbol{\beta}_0) \\
= & \sqrt{n} \widehat\Psc_{\scriptscriptstyle \Ssc} \mathcal{W}_{{\scriptscriptstyle \Ssc} 1}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) + \sqrt{n} \widehat\Psc_{\scriptscriptstyle \Tsc} \mathcal{W}_{{\scriptscriptstyle \Tsc} 1}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) + \sqrt{n} \dot{\TP}(c, \boldsymbol{\beta}_0) ^{\scriptscriptstyle \sf T} (\widehat \boldsymbol{\beta} - \boldsymbol{\beta}_0) + o_p(1)
\end{align*}
converge to zero-mean Gaussian distribution, where
\begin{equation*}
\begin{aligned}
& \mathcal{W}_{{\scriptscriptstyle \Ssc} 1}(c,\boldsymbol{\beta}, \boldsymbol{\theta},\mathbf{D}_i) = \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\alpha}) \{Y_i-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma})\} - \mathbb{E}_{\scriptscriptstyle \Ssc}\mathbb{I}(\mathbf{A}^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) \exp(\mathbf{X}^{\scriptscriptstyle \sf T} \boldsymbol{\alpha}) \{Y-g(\mathbf{X}^{\scriptscriptstyle \sf T} \boldsymbol{\gamma})\} \,, \\
& \mathcal{W}_{{\scriptscriptstyle \Tsc} 1}(c,\boldsymbol{\beta}, \boldsymbol{\theta}, \mathbf{D}_i) = \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma}) - \mathbb{E}_{\scriptscriptstyle \Tsc}\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma}) \,, \\
& \dot{\TP}(c, \boldsymbol{\beta}) = \partial \TP (c, \boldsymbol{\beta}) / \partial \boldsymbol{\beta} = \dot J_1(c,\boldsymbol{\beta}) + \dot J_2(c,\boldsymbol{\beta}) = \partial \mathbb{P}_{\scriptscriptstyle \Tsc}(Y=1,\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta} \geq c) / \partial \boldsymbol{\beta}\,.
\end{aligned}
\end{equation*}
Hence, $\widehat \TP (c,\widehat \boldsymbol{\beta}, \widehat \boldsymbol{\theta}_{{\sf roc},c}) = \TP (c, \boldsymbol{\beta}_0) + O_p(1/\sqrt{n})$ and we obtain the conclusion by Slutsky's theorem that
\begin{equation*}
\begin{aligned}
\widehat{\mathcal{W}}_{\mathrm{TPR}}(c) := & \sqrt{n} \{ \widehat \TPR(c) - \TPR(c) \}\\
= & \sqrt{n} \left\{ \frac{\widehat \TP (c,\widehat \boldsymbol{\beta}, \widehat \boldsymbol{\theta}_{{\sf roc},c})}{\widehat \TP (- \infty,\widehat \boldsymbol{\beta}, \widehat \boldsymbol{\theta}_{{\sf roc}, -\infty})} - \frac{\TP (c, \boldsymbol{\beta}_0)}{\TP (- \infty, \boldsymbol{\beta}_0)} \right \} \\
= & \sqrt{n} \left\{ \frac{\widehat \TP (c,\widehat \boldsymbol{\beta}, \widehat \boldsymbol{\theta}_{{\sf roc},c}) - \TP (c, \boldsymbol{\beta}_0)}{\widehat \TP (- \infty,\widehat \boldsymbol{\beta}, \widehat \boldsymbol{\theta}_{{\sf roc}, - \infty})} \right\} + \sqrt{n} \left \{ \frac{\TP (c, \boldsymbol{\beta}_0)}{\widehat \TP (- \infty,\widehat \boldsymbol{\beta}, \widehat \boldsymbol{\theta}_{{\sf roc}, - \infty})} - \frac{\TP (c, \boldsymbol{\beta}_0)}{\TP (- \infty, \boldsymbol{\beta}_0)} \right \}\\
= & \frac{ \sqrt{n}\{ \widehat \TP (c,\widehat \boldsymbol{\beta}, \widehat \boldsymbol{\theta}_{{\sf roc},c}) - \TP (c,\boldsymbol{\beta}_0) \}}{\TP (- \infty, \boldsymbol{\beta}_0)} \\
& - \frac{ \sqrt{n}\TP (c,\boldsymbol{\beta}_0) \{\widehat \TP (- \infty,\widehat \boldsymbol{\beta}, \widehat \boldsymbol{\theta}_{{\sf roc},c}) - \TP (- \infty,\boldsymbol{\beta}_0) \}}{\TP (- \infty, \boldsymbol{\beta}_0)^2} + o_p(1) \\
=& \sqrt{n} \mu^{-1} \left \{ \widehat \TP (c,\widehat \boldsymbol{\beta}, \widehat \boldsymbol{\theta}_{{\sf roc},c}) - \TP (c,\boldsymbol{\beta}_0) \right\} \\
& - \sqrt{n} \mu^{-1} \TPR(c,\boldsymbol{\beta}_0) \big\{ \widehat \TP (- \infty,\widehat \boldsymbol{\beta}, \widehat \boldsymbol{\theta}_{{\sf roc},c}) - \TP (- \infty,\boldsymbol{\beta}_0) \big\} + o_p(1) \\
= & \sqrt{n} \mu^{-1} \widehat\Psc_{\scriptscriptstyle \Ssc} \left \{ \mathcal{W}_{{\scriptscriptstyle \Ssc},{\rm TPR}}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) + \dot{\TP}(c, \boldsymbol{\beta}_0)^{\scriptscriptstyle \sf T} \mathcal{W}_{{\scriptscriptstyle \Ssc},\beta}(\mathbf{D}_i) \right\} \\
& + \sqrt{n} \mu^{-1} \widehat\Psc_{\scriptscriptstyle \Tsc}\left \{ \mathcal{W}_{{\scriptscriptstyle \Tsc},{\rm TPR}}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) + \dot{\TP}(c, \boldsymbol{\beta}_0)^{\scriptscriptstyle \sf T} \mathcal{W}_{{\scriptscriptstyle \Tsc},\beta}(\mathbf{D}_i) \right\} + o_p(1)
\end{aligned}
\label{eq: TPR}
\end{equation*}
converges to mean zero Gaussian distribution where $\mu = \mathbb{P}_{\scriptscriptstyle \Tsc}(Y=1)$. Here we have
\begin{equation*}
\begin{aligned}
& \mathcal{W}_{{\scriptscriptstyle \Ssc},{\rm TPR}}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) = \mathcal{W}_{{\scriptscriptstyle \Ssc} 1}(c,\boldsymbol{\beta}, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) - \TPR(c,\boldsymbol{\beta}_0) \mathcal{W}_{{\scriptscriptstyle \Ssc} 1}(-\infty,\boldsymbol{\beta}, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) \\
& \mathcal{W}_{{\scriptscriptstyle \Tsc},{\rm TPR}}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc}, c},\mathbf{D}_i) = \mathcal{W}_{{\scriptscriptstyle \Tsc} 1}(c,\boldsymbol{\beta}, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) - \TPR(c,\boldsymbol{\beta}_0) \mathcal{W}_{{\scriptscriptstyle \Tsc} 1}(-\infty,\boldsymbol{\beta}, \bar \boldsymbol{\theta}_{{\sf roc}, c},D_i)
\end{aligned}
\end{equation*}
and $\mathcal{W}_{{\scriptscriptstyle \Ssc},\beta}(\mathbf{D}_i)$ and $\mathcal{W}_{{\scriptscriptstyle \Tsc},\beta}(\mathbf{D}_i)$ are defined in Theorem \ref{thm:1}. Analogous arguments can be used to show the uniform consistency of $\widehat \mathbf{FPR}(c)$ for $\mathbf{FPR}(c)$.
Recall the definitions of $\mathbf{FP}$ and $\mathbf{FPR}$ are
\[
\mathbf{FP}(c,\boldsymbol{\beta}_0) = \mathbb{P}_{\scriptscriptstyle \Tsc}(Y=0,\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0 \geq c) \text{ and } \mathbf{FPR}(c,\boldsymbol{\beta}_0) = \mathbf{FP}(c,\boldsymbol{\beta}_0)/\mathbf{FP}(-\infty,\boldsymbol{\beta}_0) = \mathbb{P}_{\scriptscriptstyle \Tsc}(\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0 \geq c \mid Y=0)\,.
\]
We have
\begin{align*}
& \sqrt{n} \big \{ \widehat \mathbf{FP} (c,\widehat \boldsymbol{\beta}, \widehat \boldsymbol{\theta}_{{\sf roc},c}) - \mathbf{FP} (c, \boldsymbol{\beta}_0) \big\} \\
= & \sqrt{n} \widehat\Psc_{\scriptscriptstyle \Ssc} \mathcal{W}_{{\scriptscriptstyle \Ssc} 2}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) + \sqrt{n} \widehat\Psc_{\scriptscriptstyle \Tsc} \mathcal{W}_{{\scriptscriptstyle \Tsc} 2}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) + \sqrt{n} \dot{\mathbf{FP}}(c, \boldsymbol{\beta}_0) ^{\scriptscriptstyle \sf T} (\widehat \boldsymbol{\beta} - \boldsymbol{\beta}_0) + o_p(1)
\end{align*}
converge to mean zero Gaussian distribution, where
\begin{equation*}
\begin{aligned}
& \mathcal{W}_{{\scriptscriptstyle \Ssc} 2}(c,\boldsymbol{\beta}, \boldsymbol{\theta},\mathbf{D}_i) = - \mathcal{W}_{{\scriptscriptstyle \Ssc} 1}(c,\boldsymbol{\beta}, \boldsymbol{\theta},\mathbf{D}_i)
\,, \\
& \mathcal{W}_{{\scriptscriptstyle \Tsc} 2}(c,\boldsymbol{\beta}, \boldsymbol{\theta}, \mathbf{D}_i) = \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) \{1-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma})\} - \mathbb{E}_{\scriptscriptstyle \Tsc}\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) \{1-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma})\} \,, \\
& \dot{\mathbf{FP}}(c, \boldsymbol{\beta}) = \partial \mathbf{FP} (c, \boldsymbol{\beta}) / \partial \boldsymbol{\beta} = \partial \mathbb{P}_{\scriptscriptstyle \Tsc}(Y=0,\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta} \geq c) / \partial \boldsymbol{\beta}\,,
\end{aligned}
\end{equation*}
and
\begin{equation*}
\begin{aligned}
\widehat{\mathcal{W}}_{\mathrm{FPR}}(c) := & \sqrt{n} \left\{ \widehat \mathbf{FPR}(c) - \mathbf{FPR}(c) \right \}\\
=& \sqrt{n} (1-\mu)^{-1} \left \{ \widehat \mathbf{FP} (c,\widehat \boldsymbol{\beta}, \widehat \boldsymbol{\theta}_{{\sf roc},c}) - \mathbf{FP} (c,\boldsymbol{\beta}_0) \right\} \\
& - \sqrt{n} (1-\mu)^{-1} \mathbf{FPR}(c,\boldsymbol{\beta}_0) \big\{ \widehat \mathbf{FP} (- \infty,\widehat \boldsymbol{\beta}, \widehat \boldsymbol{\theta}_{{\sf roc},c}) - \mathbf{FP} (- \infty,\boldsymbol{\beta}_0) \big\} + o_p(1) \\
= & \sqrt{n} (1-\mu)^{-1} \widehat\Psc_{\scriptscriptstyle \Ssc} \left \{ \mathcal{W}_{{\scriptscriptstyle \Ssc},{\rm FPR}}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) + \dot{\mathbf{FP}}(c, \boldsymbol{\beta}_0)^{\scriptscriptstyle \sf T} \mathcal{F}_{\beta,1}(\mathbf{D}_i) \right\} \\
& + \sqrt{n} (1-\mu)^{-1} \widehat\Psc_{\scriptscriptstyle \Tsc}\left \{ \mathcal{W}_{{\scriptscriptstyle \Tsc},{\rm FPR}}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) + \dot{\mathbf{FP}}(c, \boldsymbol{\beta}_0)^{\scriptscriptstyle \sf T} \mathcal{F}_{\beta,2}(\mathbf{D}_i) \right\} + o_p(1) \\
\end{aligned}
\end{equation*}
where
\begin{equation*}
\begin{aligned}
& \mathcal{W}_{{\scriptscriptstyle \Ssc}, {\rm FPR }}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) = \mathcal{W}_{{\scriptscriptstyle \Ssc} 2}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) - \mathbf{FPR}(c,\boldsymbol{\beta}_0) \mathcal{W}_{{\scriptscriptstyle \Ssc} 2}(-\infty,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) \,, \\
& \mathcal{W}_{{\scriptscriptstyle \Tsc}, {\rm FPR }}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) = \mathcal{W}_{{\scriptscriptstyle \Tsc} 2}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) - \mathbf{FPR}(c,\boldsymbol{\beta}_0) \mathcal{W}_{{\scriptscriptstyle \Tsc} 2}(-\infty,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) \,.
\end{aligned}
\end{equation*}
For $\widehat \ROC(u) = \widehat \TPR \{ \widehat \mathbf{FPR}^{-1}(u)\}$, we note that the uniform consistency of $\widehat \TPR (\cdot)$ and $\widehat \mathbf{FPR} (\cdot)$ directly implies the uniform consistency of $\widehat \ROC(\cdot)$ for $\ROC(\cdot)$.
The weak convergence of $\widehat{\mathcal{W}}_{ {\rm roc }}(u)=\sqrt{n}\left\{\widehat{\ROC}(u)-\ROC(u)\right\}$ also directly follows from
the weak convergences of $\widehat{\mathcal{W}}_{\mathrm{TPR}}(\cdot)$ and $\widehat{\mathcal{W}}_{\mathrm{FPR}}(\cdot)$. Specifically, for $u \in [0,1]$, let $\widehat c_{u} = \widehat{\mathbf{FPR}}^{-1}(u)$ and $c_{u} = {\mathbf{FPR}}^{-1}(u)$. We have
$$
\begin{aligned}
\widehat{\mathcal{W}}_{\mathrm{ROC}}(u) &=\widehat{\mathcal{W}}_{\mathrm{TPR}}(\widehat{c}_{u}) + \sqrt{n} \left[\ROC \left[\mathbf{FPR} \{\widehat{\mathbf{FPR}}^{-1}(u)\}\right]-\ROC (u)\right] \\
&=\widehat{\mathcal{W}}_{\mathrm{TPR}}(c_{u}) + \sqrt{n} \ROC(u)\left\{\mathbf{FPR}(c_{u})-\widehat{\mathbf{FPR}}(c_{u}) \right\}+o_{p}(1) \\
&=\widehat{\mathcal{W}}_{\mathrm{TPR}}(c_{u})- \ROC(u)\widehat{\mathcal{W}}_{\mathrm{FPR}}(c_{u}) + o_{p}(1) \\
&= \sqrt{n} \mu^{-1} \widehat\Psc_{\scriptscriptstyle \Ssc} \left \{ \mathcal{W}_{{\scriptscriptstyle \Ssc},{\rm TPR}}(c_{u},\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c_{u}},\mathbf{D}_i) + \dot{\TP}(c_u, \boldsymbol{\beta}_0)^{\scriptscriptstyle \sf T}\mathcal{W}_{{\scriptscriptstyle \Ssc}, \beta}(\mathbf{D}_i) \right\} \\
& + \sqrt{n} \mu^{-1} \widehat\Psc_{\scriptscriptstyle \Tsc}\left \{ \mathcal{W}_{{\scriptscriptstyle \Tsc},{\rm TPR}}(c_{u},\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c_{u}},\mathbf{D}_i) + \dot{\TP}(c_u, \boldsymbol{\beta}_0)^{\scriptscriptstyle \sf T} \mathcal{W}_{{\scriptscriptstyle \Tsc}, \beta}(\mathbf{D}_i) \right\} \\
& - \ROC(u) \sqrt{n} (1-\mu)^{-1} \widehat\Psc_{\scriptscriptstyle \Ssc} \left \{ \mathcal{W}_{{\scriptscriptstyle \Ssc},{\rm FPR}}(c_{u},\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c_{u}},D_i) + \dot{\mathbf{FP}}(c_u, \boldsymbol{\beta}_0)^{\scriptscriptstyle \sf T} \mathcal{W}_{{\scriptscriptstyle \Ssc},\beta}(\mathbf{D}_i) \right\} \\
& - \ROC(u) \sqrt{n} (1-\mu)^{-1} \widehat\Psc_{\scriptscriptstyle \Tsc}\left \{ \mathcal{W}_{{\scriptscriptstyle \Tsc},{\rm FPR}}(c_{u},\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc}, c_{u}},D_i) + \dot{\mathbf{FP}}(c_u, \boldsymbol{\beta}_0)^{\scriptscriptstyle \sf T} \mathcal{W}_{{\scriptscriptstyle \Tsc}, \beta}(\mathbf{D}_i) \right\} +o_{p}(1) \\
& = \sqrt{n} \widehat\Psc_{\scriptscriptstyle \Ssc} \mathcal{W}_{{\scriptscriptstyle \Ssc}, {\sf roc}}(u,\mathbf{D}_i) + \sqrt{n} \widehat\Psc_{\scriptscriptstyle \Tsc} \mathcal{W}_{{\scriptscriptstyle \Tsc}, {\sf roc}}(u,\mathbf{D}_i) + o_p(1)\,,
\end{aligned}
$$
where
\begin{align*}
\mathcal{W}_{{\scriptscriptstyle \Ssc},{\sf roc}}(u,\mathbf{D}_i) = & \mu^{-1} \mathcal{W}_{{\scriptscriptstyle \Ssc},{\rm TPR}}(c_{u},\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc}, c_{u}},\mathbf{D}_i) - (1-\mu)^{-1} \ROC(u) \mathcal{W}_{{\scriptscriptstyle \Ssc},{\rm FPR}}(c_{u},\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c_{u}},\mathbf{D}_i) \\
& + \mu^{-1} \dot{\TP}(c_u, \boldsymbol{\beta}_0) - (1-\mu)^{-1} \ROC(u) \dot{\mathbf{FP}}(c_u, \boldsymbol{\beta}_0) \} ^{\scriptscriptstyle \sf T}\mathcal{W}_{{\scriptscriptstyle \Ssc}, \beta}(\mathbf{D}_i) \\
\mathcal{W}_{{\scriptscriptstyle \Tsc},{\sf roc}}(u,\mathbf{D}_i)= & \mu^{-1} \mathcal{W}_{{\scriptscriptstyle \Tsc},{\rm TPR}}(c_{u},\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c_{u}},\mathbf{D}_i) - (1-\mu)^{-1} \ROC(u) \mathcal{W}_{{\scriptscriptstyle \Tsc},{\rm FPR}}(c_{u},\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc}, c_{u}},\mathbf{D}_i) \\
& +\{ \mu^{-1} \dot{\TP}(c_u, \boldsymbol{\beta}_0) - (1-\mu)^{-1} \ROC(u) \dot{\mathbf{FP}}(c_u, \boldsymbol{\beta}_0) \} ^{\scriptscriptstyle \sf T}\mathcal{W}_{{\scriptscriptstyle \Tsc},\beta}(\mathbf{D}_i) \,.
\end{align*}
It follows that $\widehat{\mathcal{W}}_{\mathrm{ROC}}(u)$ converges weakly to a zero-mean Gaussian process.
{
We now derive the non-asymptotic upper bound for
$\sup_{c\in\Omega_c}|\widehat\TP(c) - \TP(c)|$,
where $\widehat\TP(c) = \widehat\TP(c, \widehat \boldsymbol{\theta}_{{\sf roc}, \widehat c_j}, \widehat \boldsymbol{\beta})$.
We decompose
$$
\widehat \TP(c) - \TP(c) =
\widehat\TP(c, \widehat \boldsymbol{\theta}_{{\sf roc}, \widehat c_j}, \widehat \boldsymbol{\beta}) - \widehat\TP(c, \bar\boldsymbol{\theta}_{{\sf roc}, t(c)}, \boldsymbol{\beta}_0)
+ \widehat\TP(c, \bar\boldsymbol{\theta}_{{\sf roc}, t(c)}, \boldsymbol{\beta}_0) - \TP(c),
$$
and by Assumption \ref{asu:model} and the double robust property, we know
$$
\sup_{c\in \Omega_c} | \mathbb{E} \{ \widehat\TP(c, \bar\boldsymbol{\theta}_{{\sf roc}, t(c)}, \boldsymbol{\beta}_0) \} - \TP(c)|=0.
$$
In the proof of Theorem \ref{thm:2}, we have obtained that
$$
\sup_{c\in \Omega_c}|\widehat\TP(c, \widehat \boldsymbol{\theta}_{{\sf roc}, \widehat c_j}, \widehat \boldsymbol{\beta}) - \widehat\TP(c, \bar\boldsymbol{\theta}_{{\sf roc}, t(c)}, \boldsymbol{\beta}_0)| = O_p(n^{-1/2}).
$$
We then bound
$$
\sup_{c\in \Omega_c}|\widehat\TP(c, \bar\boldsymbol{\theta}_{{\sf roc}, t(c)}, \boldsymbol{\beta}_0) - \mathbb{E}\{\widehat\TP(c, \bar\boldsymbol{\theta}_{{\sf roc}, t(c)}, \boldsymbol{\beta}_0)\}|.
$$
Define function class
$$
\mathcal{F} = \left(\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c) \left[\rho_n S_i\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\alpha}_{{\sf roc}, t(c)})\{Y_i - g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{{\sf roc}, t(c)})\}
+ \rho_N(1 -S_i)g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{{\sf roc}, t(c)})\right]: \forall c\in \Omega_c \right).
$$
By Assumption \ref{asu:3},
$\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\alpha}_{{\sf roc}, c})$ is uniformly bounded, and hence $\mathcal{F}$ is uniformly bounded.
For the indicator function, the cardinality is less than or equal to $n+N+1$. For function $t(c)$, the cardinality is less than or equal to $m$.
Hence, $\mathcal{F}$ has polynomial discrimination.
Since $n\asymp N$ and $m \ll n$, by Lemma 4.14 in \cite{wainwright2019high}, the Rademacher complexity of $\mathcal{F}$ satisfying
$$
\mathcal{R}_n(\mathcal{F}) \precsim \sqrt{\frac{\log~n}{n}} = o(1).
$$
By Theorem 4.10 in \cite{wainwright2019high}, we have
$$
\sup_{c\in \Omega_c}|\widehat\TP(c, \bar\boldsymbol{\theta}_{{\sf roc}, t(c)}, \boldsymbol{\beta}_0) - \mathbb{E}\{\widehat\TP(c, \bar\boldsymbol{\theta}_{{\sf roc}, t(c)}, \boldsymbol{\beta}_0)\}| \precsim \sqrt{\frac{\log~n}{n}} + \delta,
$$
with probability at least $1 - \exp(- K_1n\delta^2)$,
and
$$
\sup_{c\in \Omega_c}|\widehat\TP(c) - \TP(c)| \precsim \sqrt{\frac{\log~n}{n}} + \delta,
$$
with probability at least $1 - \exp(- K_1n\delta^2)$, where $K_1$ is a positive constant.
Similarly, we have
$$
\sup_{c\in \Omega_c}|\widehat\mathbf{FP}(c) - \mathbf{FP}(c)| \precsim \sqrt{\frac{\log~n}{n}} + \delta
$$
with probability at least $1 - \exp(- K_2n\delta^2)$, where $K_2$ is a positive constant, and consequently,
$$
\sup_{u\in (0,1)}|\widehat\ROC(u) - \ROC(u)| \precsim \sqrt{\frac{\log~n}{n}} + \delta
$$
with probability at least $1 - \exp(- Kn\delta^2)$, where $K$ is a positive constant.
We set $\delta = \sqrt{n^{-1}\log~n}$, and have
$$
\sup_{u\in (0,1)}|\widehat\ROC(u) - \ROC(u)| \precsim \sqrt{\frac{\log~n}{n}}
$$
with probability at least $1 - \exp(- K\log~n)$.
}
\end{proof}
\subsection{Technical lemmas}
\subsubsection{Lemma \ref{lem:a2} and its proof}\label{sec:proof:lem:a1}
\begin{lemma}
\label{lem:a2}
Let $\breve \boldsymbol{\alpha}$ and $\breve \boldsymbol{\gamma}$ be any estimators satisfying $\|\breve \boldsymbol{\alpha} - \boldsymbol{\alpha}^{\star}\|_2 = O_p( r_\alpha ) = o_p(1)$ and $\|\breve \boldsymbol{\gamma} - \boldsymbol{\gamma}^{\star}\|_2 = O_p(r_\gamma) = o_p(1)$, where $\lim_{n\rightarrow \infty} \breve \boldsymbol{\alpha} = \boldsymbol{\alpha}^{\star}$ and $\lim_{n\rightarrow \infty} \breve \boldsymbol{\gamma} = \boldsymbol{\gamma}^{\star}$.
Under Assumptions \ref{asu:1}, \ref{asu:model}, and that $\boldsymbol{\alpha}^{\star}=\boldsymbol{\alpha}_0$ or $\boldsymbol{\gamma}^{\star}=\boldsymbol{\gamma}_0$ whenever the corresponding nuisance model is correct, $\|\breve \boldsymbol{\beta} - \boldsymbol{\beta}_0 \|_{\infty} = O_p(r_\alpha + r_\gamma+n^{-1/2})$ where $ \breve \boldsymbol{\beta}$ is the solution of $\mathcal{E}(\boldsymbol{\beta}, \breve \boldsymbol{\alpha}, \breve \boldsymbol{\gamma}) = \mathbf{0}$.
\end{lemma}
\begin{proof}
First, we show the consistency of $\breve \boldsymbol{\beta}$. Let $\breve{h}(\mathbf{x}) = \exp(\mathbf{x} ^{\scriptscriptstyle \sf T} \breve\boldsymbol{\alpha})$, $\breve{r}_{\beta}(\mathbf{x}) = g(\mathbf{x} ^{\scriptscriptstyle \sf T} \breve\boldsymbol{\gamma}_{\beta})$, ${h}^{\star}(\mathbf{x}) = \exp(\mathbf{x} ^{\scriptscriptstyle \sf T} \boldsymbol{\alpha}^\star)$ and ${r}^{\star}(\mathbf{x}) = g(\mathbf{x} ^{\scriptscriptstyle \sf T} \boldsymbol{\gamma}^{\star})$. We then expand $\mathcal{E}(\boldsymbol{\beta}, \breve \boldsymbol{\alpha}, \breve \boldsymbol{\gamma})$ as
\begin{equation}
\begin{aligned}
&\widehat\Psc_{\scriptscriptstyle \Ssc} \breve{h}(\mathbf{X}_i) \mathbf{A}_i \{ Y_i - \breve{r}(\mathbf{X}_i)\} + \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{A}_i \{ \breve{r}(\mathbf{X}_i) - g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}) \}\\
= &\widehat\Psc_{\scriptscriptstyle \Ssc} h^\star(\mathbf{X}_i) \mathbf{A}_i \{ Y_i - {r}^{\star}(\mathbf{X}_i)\} + \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{A}_i \{ {r}^{\star}(\mathbf{X}_i) - g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}) \} \\
& - \widehat\Psc_{\scriptscriptstyle \Ssc} \{\breve{h}(\mathbf{X}_i) - {h}^{\star}(\mathbf{X}_i)\} \mathbf{A}_i \{ \breve r(\mathbf{X}_i) - {r}^{\star}(\mathbf{X}_i)\}\\
& - \widehat\Psc_{\scriptscriptstyle \Ssc} {h}^{\star}(\mathbf{X}_i) \mathbf{A}_i \{ \breve r(\mathbf{X}_i) - {r}^{\star}(\mathbf{X}_i)\} + \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{A}_i \{ \breve{r}(\mathbf{X}_i) - {r}^{\star}(\mathbf{X}_i) \} \\
& + \widehat\Psc_{\scriptscriptstyle \Ssc} \{\breve{h}(\mathbf{X}_i) - {h}^{\star}(\mathbf{X}_i)\} \mathbf{A}_i \{Y_i - {r}^{\star}(\mathbf{X}_i)\} \\
=:& \mathcal{E}(\boldsymbol{\beta}, \boldsymbol{\alpha}^\star, \boldsymbol{\gamma}^\star) + e_a + e_b + e_c.
\end{aligned}
\label{expansion:consistency beta1}
\end{equation}
We can control the infinity bounds of $e_a$, $e_b$, $e_c$ as follows.
\begin{equation*}
\begin{aligned} \|e_a\|_{\infty} & \leq \widehat\Psc_{\scriptscriptstyle \Ssc} h^{\star}(\mathbf{X}_i) \|\mathbf{A}_i\|_{\infty} \big |\{ \breve h(\mathbf{X}_i)/h^{\star}(\mathbf{X}_i) - 1\} \{ \breve r(\mathbf{X}_i) - r^\star(\mathbf{X}_i)\} \big | \\
&\leq C\big[\widehat\Psc_{\scriptscriptstyle \Ssc}\{\mathbf{X}_i^{\scriptscriptstyle \sf T}(\breve \boldsymbol{\alpha} - \boldsymbol{\alpha}^\star)\}^2\big]^{1/2} \big[\widehat\Psc_{\scriptscriptstyle \Ssc}\{\mathbf{X}_i^{\scriptscriptstyle \sf T}(\breve \boldsymbol{\gamma} - \boldsymbol{\gamma}^\star)\}^2\big]^{1/2}\\
& \leq C' \| \breve \boldsymbol{\alpha} - \boldsymbol{\alpha}^\star \|_2 \| \breve \boldsymbol{\gamma} - \boldsymbol{\gamma}^\star \|_2 = O_p ( C_\alpha C_\gamma ) = o_p(1) \,,
\end{aligned}
\end{equation*}
where the first inequality comes from the definition of $e_a$, the second inequality comes from Assumption \ref{asu:1}, the boundness of $\mathbf{A}_i$, $\mathbf{X}_i$ and $h^{\star}(\mathbf{X}_i)$, and the first order Taylor's expansion of $\breve h(\mathbf{X}_i)/h^{\star}(\mathbf{X}_i) - 1$ and $\breve r(\mathbf{X}_i) - r^\star(\mathbf{X}_i)$, and the third inequality comes from Cauchy-Schwarz inequality and the boundness of $X_i$. Similarly, we have
\begin{equation*}
\begin{aligned}
\|e_b\|_{\infty} & = \| \widehat\Psc_{\scriptscriptstyle \Ssc} {h}^{\star}(\mathbf{X}_i) \mathbf{A}_i \{ \breve r(\mathbf{X}_i) - {r}^{\star}(\mathbf{X}_i)\} - \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{A}_i \{ \breve{r}(\mathbf{X}_i) - {r}^{\star}(\mathbf{X}_i) \} \|_{\infty}\\
& \leq C \| \widehat\Psc_{\scriptscriptstyle \Ssc} h^\star(\mathbf{X}_i) \mathbf{A}_i \dot g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma}^\star) \mathbf{X}_i^{\scriptscriptstyle \sf T} ( \breve \boldsymbol{\gamma} - \boldsymbol{\gamma}^\star) - \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{A}_i \dot g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma}^\star) \mathbf{X}_i^{\scriptscriptstyle \sf T} ( \breve \boldsymbol{\gamma} - \boldsymbol{\gamma}^\star) \|_{\infty} \\
&\leq C \widehat\Psc_{\scriptscriptstyle \Ssc} h^\star(\mathbf{X}_i) \| \mathbf{A}_i \|_{\infty} | \dot g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma}^\star) \mathbf{X}_i^{\scriptscriptstyle \sf T} ( \breve \boldsymbol{\gamma} - \boldsymbol{\gamma}^\star) |+ C \widehat\Psc_{\scriptscriptstyle \Tsc} \| \mathbf{A}_i \|_{\infty} |\dot g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma}^\star) \mathbf{X}_i^{\scriptscriptstyle \sf T} ( \breve \boldsymbol{\gamma} - \boldsymbol{\gamma}^\star) | \\
&\leq C' \big[\widehat\Psc_{\scriptscriptstyle \Ssc}\{\mathbf{X}_i^{\scriptscriptstyle \sf T}( \breve \boldsymbol{\gamma} - \boldsymbol{\gamma}^\star)\}^2\big]^{1/2} + C' \big[\widehat\Psc_{\scriptscriptstyle \Tsc}\{\mathbf{X}_i^{\scriptscriptstyle \sf T}( \breve \boldsymbol{\gamma} - \boldsymbol{\gamma}^\star)\}^2\big]^{1/2} \\
& = O_p( \|\breve \boldsymbol{\gamma} - \boldsymbol{\gamma}^\star\|_2 ) = O_p(C_\gamma) = o_p(1)
\end{aligned}
\end{equation*}
and
\begin{equation*}
\begin{aligned}
\|e_c\|_{\infty} & = \big\|\widehat\Psc_{\scriptscriptstyle \Ssc} h^\star(\mathbf{X}_i) \mathbf{A}_i \{Y_i -r^\star(\mathbf{X}_i)\} \big\{ \breve h(\mathbf{X}_i)/h^\star(\mathbf{X}_i) - 1\big\} \big\|_{\infty} \\
& \leq C \| \widehat\Psc_{\scriptscriptstyle \Ssc} h^\star(\mathbf{X}_i) \mathbf{A}_{i} \{Y_i - r^\star(\mathbf{X}_i)\} \mathbf{X}_i^{\scriptscriptstyle \sf T} (\breve \boldsymbol{\alpha} - \boldsymbol{\alpha}^\star)\|_{\infty}\\
& \leq C' \max_{1 \leq i \leq n} \| \mathbf{A}_i \|_{\infty} \big[\widehat\Psc_{\scriptscriptstyle \Ssc}\{\mathbf{X}_i^{\scriptscriptstyle \sf T}(\breve \boldsymbol{\alpha} - \boldsymbol{\alpha}^\star)\}^2\big]^{1/2} \\
& = O_p( \| \breve \boldsymbol{\alpha} - \boldsymbol{\alpha}^\star \|_2 ) = O_p(C_\alpha ) = o_p(1) \,.
\end{aligned}
\end{equation*}
Thus, $\breve \boldsymbol{\beta}$ solves $\mathcal{E}(\boldsymbol{\beta}, \boldsymbol{\alpha}^\star, \boldsymbol{\gamma}^\star) + o_p(1) = \mathbf{0}$. By Lemma \ref{lem:double} and Assumption \ref{asu:model}, we have that $\boldsymbol{\beta}_0$ is the solution of $\mathbb{E} \mathcal{E}(\boldsymbol{\beta}, \boldsymbol{\alpha}^\star, \boldsymbol{\gamma}^\star) = \mathbf{0}$. By the uniform law of large numbers (ULLN), as $n \to \infty$, $\sup_{\boldsymbol{\beta} \in \mathcal{B}} \| \mathcal{E}(\boldsymbol{\beta}, \boldsymbol{\alpha}^\star, \boldsymbol{\gamma}^\star) - \mathbb{E} \mathcal{E}(\boldsymbol{\beta}, \boldsymbol{\alpha}^\star, \boldsymbol{\gamma}^\star) \| = o_p(1)$ and thus $\breve{\boldsymbol{\beta}}$ converges to $\boldsymbol{\beta}_0$ in probability.
We now show the rate of $\|\breve{\boldsymbol{\beta}}- \boldsymbol{\beta}_0\|$. Noting that $\breve \boldsymbol{\beta}$ is consistent for $\boldsymbol{\beta}_0$, by Theorem 5.21 of \cite{van2000asymptotic}, we expand \eqref{expansion:consistency beta1} as:
\begin{align*}
\breve\boldsymbol{\beta}-\boldsymbol{\beta}_0 = &\widehat\Psc_{\scriptscriptstyle \Ssc} \boldsymbol{\Sigma}_{\boldsymbol{\beta}^+}^{-1} \mathbf{A}_i h^\star(\mathbf{X}_i) \{ Y_i - r^\star(\mathbf{X}_i)\} + \widehat\Psc_{\scriptscriptstyle \Tsc} \boldsymbol{\Sigma}_{\boldsymbol{\beta}^+}^{-1} \mathbf{A}_i \{ r^\star(\mathbf{X}_i) - g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0) \} \\
& + \widehat\Psc_{\scriptscriptstyle \Ssc} \boldsymbol{\Sigma}_{\boldsymbol{\beta}^+}^{-1} \mathbf{A}_i \{ \breve h(\mathbf{X}_i) - h^\star(\mathbf{X}_i) \} \{ Y_i - r^\star(\mathbf{X}_i)\}\\
& - \widehat\Psc_{\scriptscriptstyle \Ssc} \boldsymbol{\Sigma}_{\boldsymbol{\beta}^+}^{-1} \mathbf{A}_i h^\star(\mathbf{X}_i) \{ \breve{r}(\mathbf{X}_i) - r^\star(\mathbf{X}_i)\} + \widehat\Psc_{\scriptscriptstyle \Tsc} \boldsymbol{\Sigma}_{\boldsymbol{\beta}^+}^{-1} \mathbf{A}_i \{ \breve{r}(\mathbf{X}_i) - r^\star(\mathbf{X}_i) \} + o_p(1/\sqrt{n} )\\
:= & \mathbf{V}_1 + e_{11} + e_{12} + o_p(1/\sqrt{n} ),
\end{align*}
where $\boldsymbol{\beta}^+$ lies between $\breve\boldsymbol{\beta}$ and $\boldsymbol{\beta}_0$. Here we replace $\widehat {\boldsymbol{\Sigma}}_{\boldsymbol{\beta}^+} = \widehat\Psc_{\scriptscriptstyle \Tsc} \dot{g}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}^+) \mathbf{A}_i \mathbf{A}_i^{\scriptscriptstyle \sf T}$ by $\boldsymbol{\Sigma}_{\breve \boldsymbol{\beta}} = \mathbb{E} _{\scriptscriptstyle \Tsc} [\dot{g}(\mathbf{A}^{\scriptscriptstyle \sf T} \boldsymbol{\beta}^+) \mathbf{A} \mathbf{A}^{\scriptscriptstyle \sf T}]$, which will introduce an error of order $o_p(1/\sqrt{n})$. Then we will control $e_{11}$ and $e_{12}$. At first, we have
$$\|\boldsymbol{\Sigma}_{\boldsymbol{\beta}_0} -\boldsymbol{\Sigma}_{\boldsymbol{\beta}^+} \| = \big \| \mathbb{E} \mathbf{A} \mathbf{A}^{\scriptscriptstyle \sf T} \{\dot g(\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0)-\dot g(\mathbf{A}^{\scriptscriptstyle \sf T} \boldsymbol{\beta}^+)\} \big \| = O_p( \|\boldsymbol{\beta}_0 - \boldsymbol{\beta}^+ \|_2 )= O_p( \|\boldsymbol{\beta}_0 - \breve \boldsymbol{\beta}\|_2 ) = o_p(1)$$
by the consistency of $\breve \boldsymbol{\beta}$. By Assumption \ref{asu:1}, the eigenvalues of $\boldsymbol{\Sigma}_{\boldsymbol{\beta}^+}$ are also bounded with probability going to one. Then we have
\begin{align*}
\|e_{11}\|_{\infty} & \leq C \| \widehat\Psc_{\scriptscriptstyle \Ssc} \boldsymbol{\Sigma}_{\boldsymbol{\beta}^+}^{-1} \mathbf{A}_i h^\star(\mathbf{X}_i)\{ Y_i - \breve{r}(\mathbf{X}_i)\} \mathbf{X}_i^{\scriptscriptstyle \sf T} (\breve \boldsymbol{\alpha} - \boldsymbol{\alpha}^\star) \|_{\infty} \\
& \leq C' \max_{1 \leq i \leq n} \| \mathbf{A}_i \|_{\infty} \big[\widehat\Psc_{\scriptscriptstyle \Ssc}\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\breve \boldsymbol{\alpha} - \boldsymbol{\alpha}^\star) \}^2\big]^{1/2} = O_p(C_\alpha) \,.
\end{align*}
Similarly, $\|e_{12} \|_{\infty} = O_p( C_\gamma )$. In addition, $\mathbf{V}_1 = \boldsymbol{\Sigma}_{\boldsymbol{\beta}^+}^{-1} \mathbf{V}(\boldsymbol{\beta}_0)$ and $\mathbb{E}\mathbf{V}(\boldsymbol{\beta}_0) = 0$. Then by Central Limit Theorem, we have $\| \mathbf{V}_1 \|_{\infty} = O_p(1/\sqrt{n})$.
As a result,
$$\| \widehat\boldsymbol{\beta}-\boldsymbol{\beta}_0 \|_{\infty} = O_p(1/\sqrt{n}) + O_p(C_\alpha) + O_p(C_\gamma) + o_p(1/\sqrt{n}) = O_p(C_\alpha + C_\gamma+1/\sqrt{n}) \,.$$
Since the dimension of $\boldsymbol{\beta}$ is fixed, we also have $\| \breve \boldsymbol{\beta}-\boldsymbol{\beta}_0 \|_2 = O_p(C_\alpha + C_\gamma+1/\sqrt{n})$.
\end{proof}
{
\subsubsection{Lemma \ref{lem:continuity} and its proof}\label{sec:proof:lem:continuity}
\begin{lemma}
\label{lem:continuity}
Under Assumptions \ref{asu:1} and \ref{asu:3}, for any $c_1<c_2$, there exist universal positive constants $K_2$ and $K_3$ such that
$$
\|\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_1} - \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}\|_2^2
\leq K_2|c_1 - c_2|, \quad \text{and} \quad \|\bar\boldsymbol{\delta}_{{\sf roc}, \gamma, c_1} - \bar\boldsymbol{\delta}_{{\sf roc}, \gamma, c_2}\|_2^2
\leq K_3|c_1 - c_2|.
$$
\end{lemma}
\begin{proof}
By the definition of $\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c}$, for any $c_1 < c_2$, we have
$$
l(c_1, \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_1}) - l(c_1, \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}) \leq l(c_1, \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}) - l(c_2, \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}),
$$
where
\begin{align*}
l(c_1, \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}) = \rho_{\scriptscriptstyle \Ssc} \mathbb{E}_{\scriptscriptstyle \Ssc} \{\mathbb{I}(\mathbf{A}_i ^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c_1)
\mathcal{F}_{{\scriptscriptstyle \Ssc} i}(\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}; \bar \boldsymbol{\alpha}, \bar \boldsymbol{\gamma})\} - \rho_{\scriptscriptstyle \Tsc} \mathbb{E}_{\scriptscriptstyle \Tsc} \{ \mathbb{I}(\mathbf{A}_i ^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c_1)
\mathcal{F}_{{\scriptscriptstyle \Tsc} i}(\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}; \bar \boldsymbol{\alpha}, \bar \boldsymbol{\gamma})\}.
\end{align*}
By Taylor expansion, we have
\begin{align*}
l(c_1, \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_1}) - l(c_1, \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}) = &
\frac{\partial l(c_1, \boldsymbol{\delta})}{\partial \boldsymbol{\delta}}\big |_{\boldsymbol{\delta} = \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}}(\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_1} - \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2})\\ +& \frac{1}{2}(\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_1} - \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2})^{\scriptscriptstyle \sf T}
\frac{\partial^2 l(c_1, \boldsymbol{\delta})}{\partial \boldsymbol{\delta}\partial \boldsymbol{\delta}^{\scriptscriptstyle \sf T} }\big |_{\boldsymbol{\delta} = \bar\boldsymbol{\delta}_{{\sf roc}, \alpha}^\prime}
(\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_1} - \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}),
\end{align*}
where $\bar\boldsymbol{\delta}_{{\sf roc}, \alpha}^\prime$
is on the line connecting $\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_1}$ and $\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}$.
Since
\[
\frac{\partial l(c_1, \boldsymbol{\delta})}{\partial \boldsymbol{\delta}}\big |_{\boldsymbol{\delta} = \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_1}} = \boldsymbol{0},
\]
we further have
\begin{align*}
&l(c_1, \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_1})) - l(c_1, \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2})\\
= &
\left\{ \frac{\partial l(c_1, \boldsymbol{\delta})}{\partial \boldsymbol{\delta}}\big |_{\boldsymbol{\delta} = \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}} - \frac{\partial l(c_1, \boldsymbol{\delta})}{\partial \boldsymbol{\delta}}\big |_{\boldsymbol{\delta} = \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_1}}
\right\}
(\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_1} - \bar\boldsymbol{\delta}_{{\sf roc}, \alpha,c_2})\\ +& \frac{1}{2}(\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_1} - \bar\boldsymbol{\delta}_{{\sf roc}, \alpha,c_2})^{\scriptscriptstyle \sf T}
\frac{\partial^2 l(c_1, \boldsymbol{\delta})}{\partial \boldsymbol{\delta}\partial \boldsymbol{\delta}^{\scriptscriptstyle \sf T} }\big |_{\boldsymbol{\delta} = \bar\boldsymbol{\delta}_{{\sf roc}, \alpha}\prime}
(\bar\boldsymbol{\delta}_{{\sf roc}, \alpha,c_1} - \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2})\\
=& (\bar\boldsymbol{\delta}_{{\sf roc}, \alpha,c_1} - \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2})^{\scriptscriptstyle \sf T}
\left\{
\frac{\partial^2 l(c_1, \boldsymbol{\delta})}{\partial \boldsymbol{\delta}\partial \boldsymbol{\delta}^{\scriptscriptstyle \sf T} }\big |_{\boldsymbol{\delta} = \bar\boldsymbol{\delta}_{{\sf roc}, \alpha}^{\prime\prime}}
+ \frac{1}{2}
\frac{\partial^2 l(c_1, \boldsymbol{\delta})}{\partial \boldsymbol{\delta}\partial \boldsymbol{\delta}^{\scriptscriptstyle \sf T} }\big |_{\boldsymbol{\delta} = \bar\boldsymbol{\delta}_{{\sf roc}, \alpha}^\prime}
\right\}
(\bar\boldsymbol{\delta}_{{\sf roc}, \alpha,c_1} - \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}),
\end{align*}
where $\bar\boldsymbol{\delta}_{{\sf roc}, \alpha}^{\prime\prime}$
is on the line connecting $\bar\boldsymbol{\delta}_{{\sf roc}, \alpha,c_1}$ and $\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}$.
By Assumption \ref{asu:3}, $\|\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_1}\|_1$ and $\|\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}\|_1$ are bounded, then we have $\|\bar\boldsymbol{\delta}_{{\sf roc}, \alpha}^\prime\|_1$ and
$\|\bar\boldsymbol{\delta}_{{\sf roc}, \alpha}^{\prime\prime}\|_1$ are also bounded.
Since the minimal eigenvalue of $\mathbb{E}_{_{\scriptscriptstyle \Ssc}}(\mathbf{X}_i\mathbf{X}_i^{\scriptscriptstyle \sf T})$ is bounded away from zero and $\|\bar\boldsymbol{\alpha}\|_1$ is bounded by Assumptions \ref{asu:1} and \ref{asu:3}, there exists a positive constant $K_0$ such that
\[
\lambda_{\min} \left\{
\frac{\partial^2 l(c_1, \boldsymbol{\delta})}{\partial \boldsymbol{\delta}\partial \boldsymbol{\delta}^{\scriptscriptstyle \sf T} }\big |_{\boldsymbol{\delta} = \bar\boldsymbol{\delta}_{{\sf roc}, \alpha}{\prime\prime}}
+ \frac{1}{2}
\frac{\partial^2 l(c_1, \boldsymbol{\delta})}{\partial \boldsymbol{\delta}\partial \boldsymbol{\delta}^{\scriptscriptstyle \sf T} }\big |_{\boldsymbol{\delta} = \bar\boldsymbol{\delta}_{{\sf roc}, \alpha}\prime}
\right\} \geq K_0>0.
\]
Hence, we have
\begin{align*}
\|\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_1} - \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}\|_2^2
\leq \frac{1}{K_0}\{ l(c_1, \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}) - l(c_2, \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2})\}.
\end{align*}
Since the probability density of $\mathbf{A}_i$ is continuously differentiable by Assumption \ref{asu:1}, we have
\begin{align*}
&|l(c_1, \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}) - l(c_2, \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2})|\\
=& |\rho_{\scriptscriptstyle \Ssc} \mathbb{E}_{\scriptscriptstyle \Ssc} \{\mathbb{I}(c_1\leq \mathbf{A}_i ^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \leq c_2)
\mathcal{F}_{{\scriptscriptstyle \Ssc} i}(\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}; \bar \boldsymbol{\alpha}, \bar \boldsymbol{\gamma})\} - \rho_{\scriptscriptstyle \Tsc} \mathbb{E}_{\scriptscriptstyle \Tsc} \{ \mathbb{I}(c_1\leq \mathbf{A}_i ^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \leq c_2)
\mathcal{F}_{{\scriptscriptstyle \Tsc} i}(\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}; \bar \boldsymbol{\alpha}, \bar \boldsymbol{\gamma})\}| \\
\leq & \rho_{\scriptscriptstyle \Ssc} \mathbb{E}_{\scriptscriptstyle \Ssc} \{\mathbb{I}(c_1\leq \mathbf{A}_i ^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \leq c_2)
|\mathcal{F}_{{\scriptscriptstyle \Ssc} i}(\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}; \bar \boldsymbol{\alpha}, \bar \boldsymbol{\gamma})|\} + \rho_{\scriptscriptstyle \Tsc} \mathbb{E}_{\scriptscriptstyle \Tsc} \{ \mathbb{I}(c_1\leq \mathbf{A}_i ^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \leq c_2)
|\mathcal{F}_{{\scriptscriptstyle \Tsc} i}(\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}; \bar \boldsymbol{\alpha}, \bar \boldsymbol{\gamma})|\}\\
\leq & K_1|c_1-c_2|,
\end{align*}
where $K_1$ is a positive constant.
The last inequality holds because
$\|\mathbf{X}_i\|_\infty$ is uniformly bounded by Assumption \ref{asu:1} and
$\|\bar\boldsymbol{\alpha} + \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c}\|_1$ is also uniformly bounded for any $c$ by Assumption \ref{asu:3}.
Therefore,
\begin{align*}
\|\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_1} - \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_2}\|_2^2
\leq K_2|c_1 - c_2|,
\end{align*}
where $K_2 = K_1/K_0$.
By similar derivations, we can show that there exists a positive constant $K_3$ such that
\begin{align*}
\|\bar\boldsymbol{\delta}_{{\sf roc}, \gamma, c_1} - \bar\boldsymbol{\delta}_{{\sf roc}, \gamma, c_2}\|_2^2
\leq K_3|c_1 - c_2|.
\end{align*}
\end{proof}
}
\subsubsection{Proof of Lemma \ref{lemma:1}}
\begin{proof}
Since the analysis can be conducted for each $j$ for $j = 1, \ldots, q$, for notational simplicity, we leave out the subscript $j$ in the nuisance parameters and estimators in the following analysis, e.g., using $\bar \boldsymbol{\alpha}_{\beta}$ to represent $\bar \boldsymbol{\alpha}_{\beta_j}$ and $\widehat \boldsymbol{\alpha}_{\beta}$ to represent $\widehat \boldsymbol{\alpha}_{\beta_j}$. Recall $\widehat \boldsymbol{\alpha}_{\beta} - \bar \boldsymbol{\alpha}_{\beta} = \widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha} + \widehat\boldsymbol{\delta}_{\beta,\alpha} - \bar\boldsymbol{\delta}_{\beta,\alpha}$.
Although $\bar \boldsymbol{\delta}_{\beta, \alpha}$ is soft sparse, it can be approximated well by a sparse vector.
Let $\eta>0$ be a threshold and define the threshold subset
\[
S_\eta = [j\in \{1, \dots, p+q\}:\| (\bar\boldsymbol{\delta}_{\beta, \alpha})_j\|_1 > \eta].
\]
Let
\[
\widehat w_i(\boldsymbol{\beta}) = \mathbf{e}_j^{\scriptscriptstyle \sf T} \{\widehat\Psc_{\scriptscriptstyle \Tsc} \dot{g}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}) \mathbf{A}_i\mathbf{A}_i^{\scriptscriptstyle \sf T} \}^{-1} \mathbf{A}_i,
\]
and
\[
l_{\beta, \alpha}(\boldsymbol{\delta}; \widetilde\boldsymbol{\alpha}, \widetilde \boldsymbol{\gamma}, \widetilde\boldsymbol{\beta}) =
\widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}\widehat w_i(\widetilde \boldsymbol{\beta})\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} )\mathcal{F}_i(\boldsymbol{\delta};\widetilde\boldsymbol{\alpha}),\]
which is convex in $\boldsymbol{\delta}$.
By the definition of $\widehat\boldsymbol{\delta}_{\beta, \alpha}$,
for any $u\in(0,1]$ we have
\[
l_{\beta, \alpha}(\widehat\boldsymbol{\delta}_{\beta, \alpha}; \widetilde\boldsymbol{\alpha}, \widetilde\boldsymbol{\gamma}, \widetilde \boldsymbol{\beta}) + \lambda_{\beta,\alpha} \|\widehat \boldsymbol{\delta}_{\beta, \alpha}\|_1
\leq
l_{\beta, \alpha}\{(1 - u)\widehat\boldsymbol{\delta}_{\beta, \alpha} + u\bar\boldsymbol{\delta}_{\beta, \alpha}; \widetilde\boldsymbol{\alpha}, \widetilde\boldsymbol{\gamma}, \widetilde\boldsymbol{\beta}\} + \lambda_{\beta,\alpha} \|(1 - u)\widehat\boldsymbol{\delta}_{\beta, \alpha} + u\bar\boldsymbol{\delta}_{\beta, \alpha}\|_1,
\]
then
\[
l_{\beta, \alpha}(\widehat\boldsymbol{\delta}_{\beta, \alpha}; \widetilde\boldsymbol{\alpha}, \widetilde\boldsymbol{\gamma}, \widetilde \boldsymbol{\beta}) -
l_{\beta, \alpha}\{(1 - u)\widehat\boldsymbol{\delta}_{\beta, \alpha} + u\bar\boldsymbol{\delta}_{\beta, \alpha}; \widetilde\boldsymbol{\alpha}, \widetilde\boldsymbol{\gamma}, \widetilde\boldsymbol{\beta}\}
+ u \lambda_{\beta,\alpha} \|\widehat \boldsymbol{\delta}_{\beta, \alpha}\|_1 \leq u\lambda_{\beta,\alpha} \| \bar\boldsymbol{\delta}_{\beta, \alpha}\|_1.
\]
Dividing both sides of the preceding inequality by $u$ and letting $u\to 0+$ yields
\[
\widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}\widehat w_i(\widetilde \boldsymbol{\beta})\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} ) \left[ \rho_n S_i \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \widehat \boldsymbol{\delta}_{\beta, \alpha})\} - \rho_N (1 - S_i)\right] \mathbf{X}_i^{\scriptscriptstyle \sf T} (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})+ \lambda_{\beta,\alpha} \|\widehat \boldsymbol{\delta}_{\beta, \alpha}\|_1 \leq \lambda_{\beta,\alpha} \| \bar\boldsymbol{\delta}_{\beta, \alpha}\|_1.
\]
After a simple rearrangement, we have
\begin{align*}
& D_{\beta, \alpha}(\widehat \boldsymbol{\delta}_{\beta, \alpha},\bar \boldsymbol{\delta}_{\beta, \alpha}; \widetilde\boldsymbol{\alpha}, \widetilde\boldsymbol{\gamma}, \widetilde \boldsymbol{\beta})+\lambda_{\beta,\alpha} \|\widehat \boldsymbol{\delta}_{\beta, \alpha}\|_1 \\
\leq
& - \widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}\widehat w_i(\widetilde \boldsymbol{\beta})\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} ) \left[ \rho_n S_i \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} - \rho_N (1 - S_i)\right] (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T}\mathbf{X}_i + \lambda_{\beta,\alpha} \| \bar\boldsymbol{\delta}_{\beta, \alpha}\|_1,
\end{align*}
where
\begin{align*}
D_{\beta, \alpha}(\widehat \boldsymbol{\delta}_{\beta, \alpha},\bar \boldsymbol{\delta}_{\beta, \alpha}; \widetilde\boldsymbol{\alpha}, \widetilde\boldsymbol{\gamma}, \widetilde \boldsymbol{\beta}) & =
\langle \widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}, \nabla l_{\beta, \alpha}(\widehat\boldsymbol{\delta}_{\beta, \alpha}; \widetilde\boldsymbol{\alpha}, \widetilde\boldsymbol{\gamma}, \widetilde \boldsymbol{\beta}) - \nabla l_{\beta, \alpha}(\bar \boldsymbol{\delta}_{\beta, \alpha}; \widetilde\boldsymbol{\alpha}, \widetilde\boldsymbol{\gamma}, \widetilde \boldsymbol{\beta}) \rangle\\
& =
\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\widetilde \boldsymbol{\beta})\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} ) \left[\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \widehat \boldsymbol{\delta}_{\beta, \alpha})\} - \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\}\right](\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T}\mathbf{X}_i.
\end{align*}
We first deal with $\widetilde \boldsymbol{\alpha}$, $\widetilde \boldsymbol{\gamma}$ and $\widetilde \boldsymbol{\beta}$ in $D_{\beta, \alpha}(\widehat \boldsymbol{\delta}_{\beta, \alpha},\bar \boldsymbol{\delta}_{\beta, \alpha}; \widetilde\boldsymbol{\alpha}, \widetilde\boldsymbol{\gamma}, \widetilde \boldsymbol{\beta})$.
Since
$\|\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha}\|_1 = O_p(s\lambda_\alpha)$ and $\|\widetilde \boldsymbol{\gamma} - \bar \boldsymbol{\gamma}\|_1 = O_p(s\lambda_\gamma)$, by Lemma \ref{lem:a2},
we have $\|\widetilde \boldsymbol{\beta} - \boldsymbol{\beta}_0\|_1 = O_p(s\lambda_\alpha + s\lambda_\gamma)$.
By Assumption \ref{asu:1}, there exist positive constants $c_{\alpha,1}$, $c_{\gamma,1}$, and $c_{\beta,1}$ such that
\begin{equation}
\label{eq:lowerbound}
\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T}(\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha})\} \geq_p e^{-c_{\alpha,1}}, \ \
\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T}(\widetilde \boldsymbol{\gamma} - \bar \boldsymbol{\gamma})\} \geq_p e^{-c_{\gamma,1}} \ \ \text{and} \ \
\exp\{\mathbf{A}_i^{\scriptscriptstyle \sf T}(\widetilde \boldsymbol{\beta} - \boldsymbol{\beta}_0)\} \geq_p e^{-c_{\beta,1}}.
\end{equation}
In practice, samples are divided into two sets with positive and negative only $\widehat w_i$'s and solve for $\widehat \boldsymbol{\delta}_{\beta, \alpha}^s$, respectively.
For simplicity of the proofs,
we assume that $\widehat w_i$'s are all positive, i.e.,
$\widehat w_i(\widetilde\boldsymbol{\beta}) >0$ and $\widehat w_i(\boldsymbol{\beta}_0) >0$, $\forall i$.
Since
$\dot{g}(u) \leq \dot{g}(u^\prime)e^{|u - u^\prime|}$,
for any $(u, u^\prime)$, we have
\begin{equation}
\label{eq:w_bound}
e^{c_{\beta, 1}}\widehat w_i(\boldsymbol{\beta}_0) \geq_p \widehat w_i(\widetilde\boldsymbol{\beta})
\geq_p e^{-c_{\beta, 1}}\widehat w_i(\boldsymbol{\beta}_0), \ \ \forall i.
\end{equation}
By (\ref{eq:lowerbound}), (\ref{eq:w_bound})
and the positivity of $\dot{g}(\cdot)$, we obtain
\begin{align*}
&D_{\beta, \alpha}(\widehat \boldsymbol{\delta}_{\beta, \alpha},\bar \boldsymbol{\delta}_{\beta, \alpha}; \widetilde\boldsymbol{\alpha}, \widetilde\boldsymbol{\gamma}, \widetilde \boldsymbol{\beta}) \\
=& \widehat\Psc_{\scriptscriptstyle \Ssc} \widehat w_i (\widetilde\boldsymbol{\beta})\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} ) \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha}) \} \left[\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \widehat \boldsymbol{\delta}_{\beta, \alpha})\} - \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\}\right] \mathbf{X}_i^{\scriptscriptstyle \sf T} (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}) \\
\geq_p & e^{-c_{\alpha, 1}} \widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i (\widetilde\boldsymbol{\beta}) \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar \boldsymbol{\gamma} )
\exp\{- |(\widetilde \boldsymbol{\gamma} - \bar\boldsymbol{\gamma})^{\scriptscriptstyle \sf T}\mathbf{X}_i|\}
\left[\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \widehat \boldsymbol{\delta}_{\beta, \alpha})\} - \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\}\right] \mathbf{X}_i^{\scriptscriptstyle \sf T} (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})\\
\geq_p &
e^{-c_{\gamma,1}-c_{\alpha, 1}}
\widehat\Psc_{\scriptscriptstyle \Ssc} \widehat w_i (\widetilde\boldsymbol{\beta})
\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar \boldsymbol{\gamma} )
\left[\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \widehat \boldsymbol{\delta}_{\beta, \alpha})\} - \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\}\right] \mathbf{X}_i^{\scriptscriptstyle \sf T} (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})
\\
\geq_p & e^{-C_{\beta, \alpha, 1}}
D_{\beta, \alpha}(\widehat \boldsymbol{\delta}_{\beta, \alpha},\bar \boldsymbol{\delta}_{\beta, \alpha}; \bar \boldsymbol{\alpha}, \bar \boldsymbol{\gamma}, \boldsymbol{\beta}_0),
\end{align*}
where $C_{\beta, \alpha, 1} =c_{\beta, 1} + c_{\gamma, 1} + c_{\alpha, 1}$, and
\begin{align*}
D_{\beta, \alpha}(\widehat \boldsymbol{\delta}_{\beta, \alpha},\bar \boldsymbol{\delta}_{\beta, \alpha}; \bar \boldsymbol{\alpha}, \bar \boldsymbol{\gamma}, \boldsymbol{\beta}_0)
=
\widehat\Psc_{\scriptscriptstyle \Ssc} \widehat w_i(\bar\boldsymbol{\beta})
\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar \boldsymbol{\gamma} )
\left[\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \widehat \boldsymbol{\delta}_{\beta, \alpha})\} - \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\}\right] \mathbf{X}_i^{\scriptscriptstyle \sf T} (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}).
\end{align*}
Second, we deal with $\widetilde \boldsymbol{\alpha}$, $\widetilde \boldsymbol{\gamma}$ and $\widetilde \boldsymbol{\beta}$ in
$\widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}\widehat w_i(\widetilde \boldsymbol{\beta})\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} ) \left[ \rho_N (1 - S_i) - \rho_n S_i \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \right](\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i$.
We decompose
\begin{align*}
& \widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}\widehat w_i(\widetilde \boldsymbol{\beta})\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} ) \left[ \rho_N (1 - S_i) - \rho_n S_i \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\}\right] (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\\
=& T_0^{\scriptscriptstyle \sf T} (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}) + T_1 + T_2 + T_3 + T_4,
\end{align*}
where
\[
T_0 = \widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}} \widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \left[ \rho_N (1 - S_i) - \rho_n S_i \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \right] \mathbf{X}_i
\]
\[
T_1 =
\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0) \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \left[\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} - \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde\boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\}\right](\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i,
\]
\[
T_2 =
\widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}\widehat w_i(\boldsymbol{\beta}_0)\{ \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} ) - \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\} \left[\rho_N (1 - S_i) - \rho_n S_i \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \right] (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i,
\]
\[
T_3 =
\widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}\mathbf{c}^{\scriptscriptstyle \sf T} (\widehat\boldsymbol{\Sigma}_{\widetilde\beta}^{-1} -\widehat \boldsymbol{\Sigma}_{\beta_0}^{-1})\mathbf{A}_i \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \left[\rho_N (1 - S_i) - \rho_n S_i \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \right] (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i,
\]
and
\[
T_4 =
\widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}\mathbf{c}^{\scriptscriptstyle \sf T} (\widehat\boldsymbol{\Sigma}_{\widetilde\beta}^{-1} -\widehat \boldsymbol{\Sigma}_{\beta_0}^{-1})\mathbf{A}_i\{ \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} ) - \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\}\left[ \rho_N (1 - S_i) - \rho_n S_i \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \right] (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i.
\]
Since for any $u\in (0,1]$,
\begin{align*}
\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \widetilde \boldsymbol{\delta}_{\beta, \alpha})\} - \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\}& = \exp[\mathbf{X}_i^{\scriptscriptstyle \sf T} \{u\bar \boldsymbol{\alpha} + (1 - u)\widetilde \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha}\}](\bar \boldsymbol{\alpha} - \widetilde \boldsymbol{\alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i \\
& = \exp\{ \mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \exp\{(1-u)\mathbf{X}_i^{\scriptscriptstyle \sf T}(\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha}) \}
(\bar\boldsymbol{\alpha} - \widetilde \boldsymbol{\alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i,
\end{align*}
by Assumption \ref{asu:1} and Cauchy-Schwartz inequality,
we have
\begin{align*}
T_1
& \leq \exp\{\kappa \|\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha}\|_1\}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \{(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2} \\
& \times
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\}
\{(\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2}\\
& \leq_p e^{c_{\alpha, 1}} (C_{\beta, \alpha, 2}s\lambda_{\alpha}^2)^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \{(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2},
\end{align*}
where $C_{\beta, \alpha, 2}$ is a positive constant.
We show that the last inequality holds in several steps.
By union bound inequality and Bernstein inequality, we have
\begin{equation}
\label{eq:SigmaError}
\|\widehat \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0} - \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}\|_{\max} \leq_p C_{\Sigma}N^{\epsilon_0 - 1/2},
\end{equation}
where $C_{\Sigma}$ is a positive constant, and $\epsilon_0$ is positive and can be arbitrarily close to zero.
By Assumption \ref{asu:1}, there exists a positive constant $c_{\alpha, 2}^{\prime}$ such that
\begin{align*}
& \widehat\Psc_{\scriptscriptstyle \Ssc} \{\widehat w_i( \boldsymbol{\beta}_0) - \bar w_i( \boldsymbol{\beta}_0)\} \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\}\{(\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \\
=& \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\beta_0}^{-1} (\boldsymbol{\Sigma}_{\beta_0} - \widehat \boldsymbol{\Sigma}_{\beta_0})
\widehat \boldsymbol{\Sigma}_{\beta_0}^{-1} \mathbf{A}_i\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \{(\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2\\
\leq & \|\widehat\Psc_{\scriptscriptstyle \Ssc} \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\beta_0}^{-1} (\boldsymbol{\Sigma}_{\beta_0} - \widehat \boldsymbol{\Sigma}_{\beta_0})
\widehat \boldsymbol{\Sigma}_{\beta_0}^{-1} \mathbf{A}_i\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\}\mathbf{X}_i \mathbf{X}_i^{\scriptscriptstyle \sf T}\|_{\max} \|\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha}\|_1^2
\\
\leq_p & c_{\alpha, 2}^{\prime} N^{\epsilon_0 - 1/2} \|\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha}\|_1^2.
\end{align*}
Since $(\widehat\Psc_{\scriptscriptstyle \Ssc} - \mathbb{E}_{\scriptscriptstyle \Ssc})\left[\bar w_i( \boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma}) \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\}
\mathbf{X}_i\mathbf{X}_i^{\scriptscriptstyle \sf T} \right]$ is sub-exponential with mean zero,
by union bound inequality and Bernstein inequality,
there exists a positive constant $c_{\alpha, 2}^{\prime\prime}$ such that
\[
\|(\widehat\Psc_{\scriptscriptstyle \Ssc} - \mathbb{E}_{\scriptscriptstyle \Ssc})\left[\bar w_i( \boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma}) \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\}
\mathbf{X}_i\mathbf{X}_i^{\scriptscriptstyle \sf T} \right]\|_{\max} \leq_p c_{\alpha, 2}
^{\prime\prime}\lambda_{\boldsymbol{\alpha}}.
\]
Further, by Lemma 1 in \cite{tan2010bounded}, we have
\[
(\mathbb{E}_{\scriptscriptstyle \Ssc} - \widehat\Psc_{\scriptscriptstyle \Ssc})\left[\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha} )
\{(\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2\right]
\leq_p \lambda_{\alpha} \|\widetilde \boldsymbol{\alpha} - \bar\boldsymbol{\alpha}\|_1^2.
\]
By Lemma 7 in \cite{tan2010bounded}, there exists a positive constant $c_{\alpha, 3}$ such that
\[
\widehat\Psc_{\scriptscriptstyle \Ssc}\left[\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha} )
\{(\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2\right]
\leq_p c_{\alpha,3} e^{c_{\alpha, 1}}s\lambda_\alpha^2.
\]
By Assumption \ref{asu:3}, $\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\delta}_{\beta, \alpha})$ is uniformly bounded. Let $M_0$ be the uniform upper bound of
$\bar w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar \boldsymbol{\delta}_{\beta, \alpha})$.
Since $N^{\epsilon_0 - 1/2} \precsim \lambda_{\alpha}$, and $s\lambda_{\alpha} = o(1)$, we have
\begin{align}
\label{eq:alpha_l2}
& \widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\}
\{(\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \nonumber\\
\leq_p &c_{\alpha, 2}^{\prime} N^{\epsilon_0 - 1/2} \|\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha}\|_1^2 + c_{\alpha,2}^{\prime\prime}\lambda_{\alpha}\|\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha}\|_1^2 + \mathbb{E}_{\scriptscriptstyle \Ssc}\left[\bar w_i( \boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\}
\{(\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2\right]\nonumber\\
\leq_p & c_{\alpha,2}\lambda_{\alpha}\|\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha}\|_1^2 + M_0( \lambda_\alpha \|\widetilde \boldsymbol{\alpha} - \bar\boldsymbol{\alpha}\|_1^2 + c_{\alpha,3} e^{c_{\alpha, 1}}s\lambda_\alpha^2) \nonumber\\
\leq_p& C_{\beta, \alpha, 2}s \lambda_{\alpha}^2.
\end{align}
For $T_2$, We decompose it as
$T_{2,1}^{\scriptscriptstyle \sf T} (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}) + T_{2,2}$,
where \[
T_{2,1} = \widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}\widehat w_i(\boldsymbol{\beta}_0) \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )
\left[\rho_N (1 - S_i) - \rho_n S_i \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \right] \mathbf{X}_i \left\{ \frac{\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} )} {\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )} -1\right\},
\]
and
\[
T_{2,2} = \widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \left[ \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha}) - \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \right] (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\left\{ \frac{\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} )} {\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )} -1\right\}.
\]
By Assumptions \ref{asu:1}, \ref{asu:3} and Cauchy-Schwartz inequality,
we know
\[
\|T_{2,1}\|_\infty \leq c_g^{-1} L \kappa \|\widetilde \boldsymbol{\gamma} - \bar\boldsymbol{\gamma}\|_1 \|T_0\|_\infty
=o_p(\|T_0\|_\infty),
\]
and
\begin{align*}
T_{2,2}
& \leq \exp\{\kappa \|\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha}\|_1\}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \{(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2} \\
& \times
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\}
\{(\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2
\right]^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}
\left\{ \frac{\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} )} {\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )} -1\right\}^2\right]^{1/2}\\
& =o_p \left((s\lambda_{\alpha}^2)^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \{(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2}\right).
\end{align*}
We then decompose
$T_3$ as $T_{3,1}^{\scriptscriptstyle \sf T}(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}) + T_{3,2}$,
where \[
T_{3,1} = \widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )
\left[\rho_N (1 - S_i) - \rho_n S_i \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \right] \mathbf{X}_i \left\{\frac{\widehat w_i(\widetilde\boldsymbol{\beta})}{\widehat w_i(\boldsymbol{\beta}_0)} -1\right\},
\]
and
\[
T_{3,2} := \widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)
\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \left[ \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha}) - \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \right] (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\left\{\frac{\widehat w_i(\widetilde\boldsymbol{\beta})}{\widehat w_i(\boldsymbol{\beta}_0)} -1\right\}.
\]
Similarly, since $\|\widetilde \boldsymbol{\beta} - \boldsymbol{\beta}_0\|_1 = o_p(1)$, by Assumptions \ref{asu:1}, \ref{asu:3} ,
we have $\|T_{3,1}\|_\infty =o_p( \|T_0\|_\infty)$ and
\[T_{3,2} =o_p \left((s\lambda_{\alpha}^2)^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \{(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2}\right).\]
$T_4$ can be decomposed as $T_{4,1}^{\scriptscriptstyle \sf T}(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}) + T_{4,2}$, where
\[
T_{4,1} := \widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )
\left[\rho_N (1 - S_i) - \rho_n S_i \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \right] \mathbf{X}_i \left\{\frac{\widehat w_i(\widetilde\boldsymbol{\beta})}{\widehat w_i(\boldsymbol{\beta}_0)} -1\right\}
\left\{ \frac{\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} )} {\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )} -1\right\},
\]
and
\begin{align*}
T_{4,2} := \widehat\Psc_{\scriptscriptstyle \Ssc} \left(\widehat w_i(\boldsymbol{\beta}_0)
\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \left[ \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha}) - \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \right] \right.\\\left.
\times(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\left\{\frac{\widehat w_i(\widetilde\boldsymbol{\beta})}{\widehat w_i(\boldsymbol{\beta}_0)} -1\right\}
\left\{ \frac{\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} )} {\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )} -1\right\}\right).
\end{align*}
Similarly, we have $\|T_{4,1}\|_\infty =o_p( \|T_0\|_\infty)$, and
\[
T_{4,2} =o_p \left((s\lambda_{\alpha}^2)^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \{(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2}\right).
\]
Hence, there exist positive constants $C_{\beta, \alpha, 3}$ and $C_{\beta, \alpha, 4}$ such that
\begin{align*}
& e^{-C_{\beta, \alpha, 1}}D_{\beta, \alpha}(\widehat \boldsymbol{\delta}_{\beta, \alpha},\bar \boldsymbol{\delta}_{\beta, \alpha}; \bar \boldsymbol{\alpha}, \bar\boldsymbol{\gamma}, \boldsymbol{\beta}_0) + \lambda_{\beta, \alpha}\|\widehat \boldsymbol{\delta}_{\beta, \alpha}\|_1
\leq_p
C_{\beta, \alpha, 3} \|T_0 \|_\infty
\|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_1
\\
&+ \lambda_{\beta,\alpha} \| \bar\boldsymbol{\delta}_{\beta, \alpha}\|_1 +C_{\beta, \alpha, 4} (s\lambda_{\alpha}^2)^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \{(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2}.
\end{align*}
To deal with $\|T_0\|_\infty$, we first decompose $T_0$
as $T_{0,1} + T_{0,2}$, where
\[
T_{0,1}
= \widehat\Psc_{\scriptscriptstyle \Ssc} \widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \mathbf{X}_i
- \widehat\Psc_{\scriptscriptstyle \Tsc} \bar w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\mathbf{X}_i,
\]
and
\[
T_{0,2} =
\widehat\Psc_{\scriptscriptstyle \Tsc}
\mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\beta_0}^{-1}(\widehat \boldsymbol{\Sigma}_{\beta_0}
- \boldsymbol{\Sigma}_{\beta_0}) \widehat \boldsymbol{\Sigma}_{\beta_0}^{-1} \mathbf{A}_i
\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\mathbf{X}_i.
\]
Since
$T_{0,1}$
is sub-exponential with mean zero, by union bound inequality and Bernstein inequality,
there exist a small enough positive constant $ C_{\beta, \alpha, 5}^{\prime}$, such that
$\|T_{0,1}\|_\infty \leq_p C_{\beta, \alpha, 5}^{\prime}\lambda_{\beta, \alpha}$.
By Assumption \ref{asu:1}, (\ref{eq:SigmaError}) and boundedness of $\dot{g}(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma})$, we have $T_{0,2} \precsim_p N^{\epsilon_0 - 1/2}$.
Since $N^{\epsilon_0 - 1/2} =o(\lambda_{\beta, \alpha})$, there further exists a positive constant $C_{\beta, \alpha, 5} <C_{\beta, \alpha, 3}^{-1}$ such that
$\|T_0\|_\infty \leq_p C_{\beta, \alpha, 5}\lambda_{\beta, \alpha}$.
Consequently, we have
\begin{align*}
& e^{-C_{\beta, \alpha, 1}}D_{\beta, \alpha}(\widehat \boldsymbol{\delta}_{\beta, \alpha},\bar \boldsymbol{\delta}_{\beta, \alpha}; \bar \boldsymbol{\alpha}, \bar\boldsymbol{\gamma}, \boldsymbol{\beta}_0) + \lambda_{\beta, \alpha}\|\widehat \boldsymbol{\delta}_{\beta, \alpha}\|_1 \\
\leq&
C_{\beta, \alpha, 6} \lambda_{\beta, \alpha}\|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_1+ \lambda_{\beta,\alpha} \| \bar\boldsymbol{\delta}_{\beta, \alpha}\|_1 \\
+& C_{\beta, \alpha, 4} (s\lambda_{\alpha}^2)^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \{(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2},
\end{align*}
where $C_{\beta, \alpha, 6} = C_{\beta, \alpha, 3} C_{\beta, \alpha, 5} < 1$.
By triangle inequalities
\[
\lambda_{\beta, \alpha}\|(\bar \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta} \|_1 \leq \lambda_{\beta, \alpha}\|(\widehat \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta}\|_1 + \lambda_{\beta, \alpha}\|(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta}\|_1,
\]
and
\[
\lambda_{\beta, \alpha}\|(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c}\|_1 \leq \lambda_{\beta, \alpha}\|(\widehat \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c} \|_1 + \lambda_{\beta, \alpha} \|(\bar \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c} \|_1,
\]
we obtain
\begin{align*}
& e^{-C_{\beta, \alpha, 1}}D_{\beta, \alpha}(\widehat \boldsymbol{\delta}_{\beta, \alpha},\bar \boldsymbol{\delta}_{\beta, \alpha}; \bar \boldsymbol{\alpha}, \bar \boldsymbol{\gamma}, \boldsymbol{\beta}_0) + (1 - C_{\beta, \alpha, 6})\lambda_{\beta, \alpha}\|(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c}\|_1\\
\leq_p&
2\lambda_{\beta,\alpha} \| (\bar\boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c}\|_1
+ (1 + C_{\beta, \alpha, 6})\lambda_{\beta, \alpha}\|(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta}\|_1 \\
+& C_{\beta, \alpha, 4} (s\lambda_{\alpha}^2)^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \{(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2},
\end{align*}
or equivalently,
\begin{align}
\label{eq:ineq}
& e^{-C_{\beta, \alpha, 1}}D_{\beta, \alpha}(\widehat \boldsymbol{\delta}_{\beta, \alpha},\bar \boldsymbol{\delta}_{\beta, \alpha}; \bar \boldsymbol{\alpha}, \bar \boldsymbol{\gamma}, \boldsymbol{\beta}_0) + (1 - C_{\beta, \alpha, 6})\lambda_{\beta, \alpha}\|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_1\\ \nonumber
\leq_p&
2\lambda_{\beta,\alpha} \| (\bar\boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c}\|_1
+ 2\lambda_{\beta, \alpha}\|(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta}\|_1 \\ \nonumber
+& C_{\beta, \alpha, 4}(s\lambda_{\alpha}^2)^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \{(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2}. \end{align}
Note that since $\bar\boldsymbol{\delta}_{\beta, \alpha}$ is allowed to be soft sparse, $\| (\bar\boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c}\|_1 $ can be nonzero.
We now establish the relationship between $D_{\beta, \alpha}(\widehat \boldsymbol{\delta}_{\beta, \alpha},\bar \boldsymbol{\delta}_{\beta, \alpha}; \bar\boldsymbol{\alpha}, \bar\boldsymbol{\gamma}, \boldsymbol{\beta}_0)$ and the quadratic term
\[
(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T}\widehat \boldsymbol{\Sigma}_{\beta, \alpha} (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}),
\]
where
$\widehat \boldsymbol{\Sigma}_{\beta, \alpha} = \widehat\Psc_{\scriptscriptstyle \Ssc}
\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar \boldsymbol{\gamma} ) \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \mathbf{X}_i \mathbf{X}_i^{\scriptscriptstyle \sf T}$.
Under Assumption \ref{asu:1}, since $\exp(u) \geq \exp(u^\prime)\exp(-|u-u^\prime|)$ for any $(u, u^\prime)$, by Lemma 10 in \cite{tan2018model}, we have
\begin{align}
\label{eq:lemma10}
D_{\beta, \alpha}(\widehat \boldsymbol{\delta}_{\beta, \alpha},\bar \boldsymbol{\delta}_{\beta, \alpha}; \bar\boldsymbol{\alpha}, \bar \boldsymbol{\gamma}, \boldsymbol{\beta}_0)
\geq \frac{1 - \exp\{-\kappa \|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_1\}}{\kappa \|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_1}
(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T}\widehat \boldsymbol{\Sigma}_{\beta, \alpha} (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}).
\end{align}
Let $D^\prime_{\beta, \alpha} =e^{-C_{\beta, \alpha, 1}}D_{\beta, \alpha}(\widehat \boldsymbol{\delta}_{\beta, \alpha},\bar \boldsymbol{\delta}_{\beta, \alpha}; \bar \boldsymbol{\alpha}, \bar\boldsymbol{\gamma}, \boldsymbol{\beta}_0) + (1 - C_{\beta, \alpha, 6})\lambda_{\beta, \alpha}\|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_1$.
Then by (\ref{eq:ineq}), we have
\begin{align*}
D^\prime_{\beta, \alpha} \leq_p &
2\lambda_{\beta,\alpha} \| (\bar\boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c}\|_1
+ 2\lambda_{\beta, \alpha}\|(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta}\|_1\\
+& C_{\beta, \alpha, 4} (s\lambda_{\alpha}^2)^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \{(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2}.
\end{align*}
We now have two possible cases by comparing $\zeta\{2\lambda_{\beta,\alpha} \| (\bar\boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c}\|_1
+ 2\lambda_{\beta, \alpha}\|(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta}\|_1 \}$ and $(1 - \zeta)C_{\beta, \alpha, 4} (s\lambda_{\alpha}^2)^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \{(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2}$, $\zeta \in (0,1]$.
Either
\begin{equation}
\label{eq:1}
\zeta D^\prime_{\beta, \alpha} \leq_p C_{\beta, \alpha, 4} (s\lambda_{\alpha}^2)^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \{(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2},
\end{equation}
or
\begin{equation}
\label{eq:2}
(1 - \zeta)D^\prime_{\beta, \alpha} \leq_p
2\lambda_{\beta,\alpha} \| (\bar\boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c}\|_1
+ 2\lambda_{\beta, \alpha}\|(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta}\|_1.
\end{equation}
When equation (\ref{eq:2}) holds, since $e^{-C_{\beta, \alpha, 1}}D_{\beta, \alpha}(\widehat \boldsymbol{\delta}_{\beta, \alpha},\bar \boldsymbol{\delta}_{\beta, \alpha}; \bar \boldsymbol{\alpha}, \bar\boldsymbol{\gamma}, \boldsymbol{\beta}_0) \geq 0$, we have
\[
(1 - C_{\beta, \alpha, 6})\lambda_{\beta, \alpha}\|(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c}\|_1
\leq
2(1 - \zeta)^{-1}\lambda_{\beta,\alpha} \| (\bar\boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c}\|_1 +
(1 -\zeta)^{-1}(1 + C_{\beta, \alpha, 6})\lambda_{\beta, \alpha}\|(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta}\|_1.
\]
By equation (43) in \cite{negahban2012unified}, we have
\[
D_{\beta, \alpha}(\widehat \boldsymbol{\delta}_{\beta, \alpha},\bar \boldsymbol{\delta}_{\beta, \alpha}; \bar\boldsymbol{\alpha}, \bar \boldsymbol{\gamma}, \boldsymbol{\beta}_0) \geq C_1\|\widehat \boldsymbol{\delta}_{\beta, \alpha}
- \bar \boldsymbol{\delta}_{\beta, \alpha}\|^2_2 - C_2 \frac{\log ~p}{n}\|(\bar \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c}\|_1^2
\]
for some positive constant $C_1$ and $C_2$.
Then by Theorem 1 in \cite{negahban2012unified}, we have
\[
\|\widehat \boldsymbol{\delta}_{\beta, \alpha}
- \bar \boldsymbol{\delta}_{\beta, \alpha}\|^2_2
\leq_p C_4 \lambda_{\beta, \alpha}^2 |S_\eta| + C_3 \lambda_{\beta, \alpha} \left\{ \frac{\log ~p}{n} \|(\bar \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c}\|_1^2 + \|(\bar \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c}\|_1\right\}
\]
for some positive constant $C_3$ and $C_4$.
We then upper bound $|S_\eta|$ and $\|(\bar \boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c}\|_1$.
By the proof of Corollary 3 in \cite{negahban2012unified},
under Assumptions \ref{asu:1} and \ref{asu:4},
we have
\[
|S_\eta| \leq R_r \eta^{-r},
\]
and
\[
\|(\bar\boldsymbol{\delta}_{\beta, \alpha})_{S_\eta^c}\|_1\leq R_r \eta^{1-r}.
\]
Setting $\eta\asymp\sqrt{n^{-1}\log ~p}$, we have
\begin{equation}
\label{eq:scenario1}
\|\widehat \boldsymbol{\delta}_{\beta, \alpha}
- \bar \boldsymbol{\delta}_{\beta, \alpha}\|^2_2
\leq_p C_{\beta, \alpha, 7} R_r\left(\frac{\log (p+q)}{n} \right)^{1- r/2},
\end{equation}
where $C_{\beta, \alpha, 7}$ is a positive constants.
Recall that
\[
\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_i(\boldsymbol{\beta}_0)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{\beta, \alpha})\} \{(\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 = (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha})^{\scriptscriptstyle \sf T}\widehat \boldsymbol{\Sigma}_{\beta, \alpha} (\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}).
\]
When equation (\ref{eq:1}) holds, since
\[
D_{\beta, \alpha} \leq e^{c_{\beta, \alpha,1}}D^\prime_{\beta, \alpha},
\]
by equation (\ref{eq:lemma10}), we have
\[
\zeta D^\prime_{\beta, \alpha} \leq C_{\beta, \alpha, 4} (s\lambda_{\alpha}^2)^{1/2}
\left[\frac{\kappa \|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_1}{1 - \exp\{-\kappa \|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_1\}
} e^{c_{\beta, \alpha,1}}D^\prime_{\beta, \alpha}\right]^{1/2},
\]
and hence
\begin{equation}
\label{eq:3}
D^\prime_{\beta, \alpha} \leq
\zeta^{-2}e^{C_{\beta, \alpha, 1}}C_{\beta, \alpha, 4}^2 s\lambda_{\alpha}^2
\frac{\kappa \|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_1}{1 - \exp\{-\kappa \|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_1\}
}.
\end{equation}
Since
\[
(1 - C_{\beta, \alpha, 6})\lambda_{\beta, \alpha}\|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_1 \leq D^\prime_{\beta, \alpha},
\]
by (\ref{eq:3}) we have
\[
1 - \exp\{-\kappa \|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_1\} \leq \frac{\zeta^{-2}e^{C_{\beta, \alpha, 1}}C_{\beta, \alpha, 4}^2 }{1 - C_{\beta, \alpha, 6}}\frac{s\lambda_\alpha^2}{\lambda_{\beta, \alpha}}.
\]
Since we assume $s\sqrt{n^{-1}\log~p} = o(1)$, let $n$ be large enough such that
\[
\frac{\zeta^{-2}e^{C_{\beta, \alpha, 1}}C_{\beta, \alpha, 4}^2}{1 - C_{\beta, \alpha, 6}}\frac{s_\alpha\lambda_\alpha^2}{\lambda_{\beta, \alpha}} \leq C_{\beta, \alpha, 8} <1.
\]
Hence,
\[
\frac{1 - \exp\{-\kappa \|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_1\}}{
\kappa \|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_1
} = \int_0^1 \exp\{ -\kappa \|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_1 u\}du \geq \exp\{ -\kappa \|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_1\} \geq 1 - C_{\beta, \alpha, 8}.
\]
Equation (\ref{eq:3}) leads to
\begin{equation}
\label{eq:scenario2}
D^\prime_{\beta, \alpha} \leq
\zeta^{-2}e^{ C_{\beta, \alpha, 1}} C_{\beta, \alpha, 4}^2
(1 - C_{\beta, \alpha, 8})^{-1}s\lambda_{\alpha}^2.
\end{equation}
When the density ratio model is correctly specified, we have $\bar \boldsymbol{\delta}_{\beta, \alpha} = \boldsymbol{0}$, and consequently
\[
\|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_2^2
=O_p\left(
s \lambda_\alpha^2 \right), \ \ \text{and} \ \
\|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_ 1
=O_p\left(
\frac{s \lambda_\alpha^2}{\lambda_{\beta, \alpha}} \right).
\]
When the density ratio model is misspecified,
by combining (\ref{eq:scenario1}) and (\ref{eq:scenario2}), we obtain
\[
\|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_2^2
=O_p\left\{
s \lambda_\alpha^2 + R_r\left(\frac{\log ~p}{n} \right)^{1- r/2}\right\},
\]
and
\[
\|\widehat \boldsymbol{\delta}_{\beta, \alpha} - \bar \boldsymbol{\delta}_{\beta, \alpha}\|_ 1
=O_p\left\{
\frac{s \lambda_\alpha^2}{\lambda_{\beta, \alpha}} + R_r\left(\frac{\log~p}{n} \right)^{1/2- r/2}\right\}.
\]
In conclusion,
when the density ratio model is correctly specified, we have
$\bar\boldsymbol{\alpha} = \bar\boldsymbol{\alpha}_{\beta}$,
\[
\|\widehat \boldsymbol{\alpha}_{\beta} - \bar\boldsymbol{\alpha}_{\beta}\|_2^2
=O_p\left(s\frac{\log~p}{n}\right), \ \ \text{and} \ \
\|\widehat \boldsymbol{\alpha}_{\beta} - \bar\boldsymbol{\alpha}_{\beta}\|_1
=O_p\left(s\sqrt{\frac{\log~p}{n}}\right).
\]
When the density ratio model is misspecified,
we have
\[
\|\widehat \boldsymbol{\alpha}_{\beta} - \bar\boldsymbol{\alpha}_{\beta}\|_2^2 =O_p\left\{ s \frac{\log ~p}{n}+ R_r\left(\frac{\log ~p}{n} \right)^{1- r/2}\right\},
\]
and
\[
\|\widehat \boldsymbol{\alpha}_{\beta} - \bar\boldsymbol{\alpha}_{\beta}\|_1 =O_p\left\{ s \sqrt{\frac{\log~p }{n}} + R_r\left(\frac{\log ~p}{n} \right)^{1/2- r/2}\right\}.
\]
Following similar derivations, under Assumptions \ref{asu:1}, \ref{asu:3} and \ref{asu:4},
when the imputation model is correctly specified, we have
$\bar\boldsymbol{\gamma} = \bar\boldsymbol{\gamma}_{\beta}$,
\[
\|\widehat \boldsymbol{\gamma}_{\beta} - \bar\boldsymbol{\gamma}_{\beta}\|_2^2
=O_p\left(s\frac{\log~p}{n}\right), \ \ \text{and} \ \
\|\widehat \boldsymbol{\gamma}_{\beta} - \bar\boldsymbol{\gamma}_{\beta}\|_1
=O_p\left(s\sqrt{\frac{\log~p}{n}}\right).
\]
When the imputation ratio model is misspecified,
we have
\[
\|\widehat \boldsymbol{\gamma}_{\beta} - \bar\boldsymbol{\gamma}_{\beta}\|_2^2 =O_p\left\{ s \frac{\log ~p}{n}+ R_r\left(\frac{\log ~p}{n} \right)^{1- r/2}\right\},
\]
and
\[
\|\widehat \boldsymbol{\gamma}_{\beta} - \bar\boldsymbol{\gamma}_{\beta}\|_1 =O_p\left\{ s \sqrt{\frac{\log~p }{n}} + R_r\left(\frac{\log ~p}{n} \right)^{1/2- r/2}\right\}.
\]
\end{proof}
\subsubsection{Proof of Lemma \ref{lemma:2}}
\begin{proof}
By the definition of $\widehat \boldsymbol{\delta}_{{\sf roc},\alpha, \widehat c_j}$ and the convexity of the loss function, we have
\begin{align}
\label{eq:ineq_auc}
& D_{{\sf roc}, \alpha, c_j}(\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j},\bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j}; \widetilde\boldsymbol{\alpha}, \widetilde\boldsymbol{\gamma}, \widehat \boldsymbol{\beta}) +\lambda_{{\sf roc},\alpha, \widehat c_j} \|\widehat \boldsymbol{\delta}_{{\sf roc},\alpha, \widehat c_j}\|_1 \nonumber\\
\leq
& - \widehat\Psc_{{\scriptscriptstyle \Ssc} {\scriptscriptstyle \cup} {\scriptscriptstyle \Tsc}}\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq \widehat c_j) \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} ) \left[ \rho_n S_i \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc},\alpha, \widehat c_j})\} - \rho_N (1 - S_i)\right] (\widehat \boldsymbol{\delta}_{{\sf roc},\alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc},\alpha, c_j})^{\scriptscriptstyle \sf T}\mathbf{X}_i\nonumber \\
+ & \lambda_{{\sf roc},\alpha, \widehat c_j} \| \bar\boldsymbol{\delta}_{{\sf roc},\alpha, c_j}\|_1,
\end{align}
where
\begin{align*}
&D_{{\sf roc}, \alpha, c_j}(\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j},\bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j}; \widetilde\boldsymbol{\alpha}, \widetilde\boldsymbol{\gamma}, \widehat \boldsymbol{\beta})\\ = &
\widehat\Psc_{\scriptscriptstyle \Ssc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq \widehat c_j) \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} ) \left[\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j})\} - \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\}\right](\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})^{\scriptscriptstyle \sf T}\mathbf{X}_i \\
=& T_{5,1} + T_{5,2},
\end{align*}
\[
T_{5,1} = \widehat\Psc_{\scriptscriptstyle \Ssc}\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c_j) \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} ) \left[\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j})\} - \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\}\right](\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})^{\scriptscriptstyle \sf T}\mathbf{X}_i,
\]
and
\begin{align*}
T_{5,2} = \widehat\Psc_{\scriptscriptstyle \Ssc} \left( \{\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat\boldsymbol{\beta} \geq \widehat c_j) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c_j)\} \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} ) \left[\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j})\} - \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\}\right] \right.\\\left.
\times
(\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})^{\scriptscriptstyle \sf T}\mathbf{X}_i \right).
\end{align*}
By Theorem \ref{thm:1}, we know $\|\widehat \boldsymbol{\beta} - \boldsymbol{\beta}_0\|_2 = O_p(n^{-1/2})$. Together with $|\widehat c_j - c_j| = O_p(n^{-1/2})$, under Assumption \ref{asu:1}, we have
\begin{align*}
\widehat\Psc_{\scriptscriptstyle \Ssc} |\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq \widehat c_j) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0\geq c_j)|
&\leq \widehat\Psc_{\scriptscriptstyle \Ssc} |\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq \widehat c_j) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0\geq \widehat c_j)| + \widehat\Psc_{\scriptscriptstyle \Ssc} |\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq \widehat c_j) - \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0\geq c_j)| \\
&= O_p(n^{-1/2}).
\end{align*}
Since $\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} ) \left[\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j})\} - \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\}\right](\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})^{\scriptscriptstyle \sf T}\mathbf{X}_i$ is non-negative and $n^{1/2} = o(n_j)$, there exists a positive constant $C_{{\sf roc}, \alpha, 1}$ such that
\begin{align*}
D_{{\sf roc}, \alpha, \widehat c_j}(\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j},\bar \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j}; \widetilde\boldsymbol{\alpha}, \widetilde\boldsymbol{\gamma}, \widehat \boldsymbol{\beta})
\geq C_{{\sf roc},\alpha, 1}T_{5,1}.
\end{align*}
To deal with the right-hand-side of (\ref{eq:ineq_auc}), We first consider the upper bound of
\begin{align*}
\|\widehat\Psc_{\mathcal{S}\cup\mathcal{T}} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c_j)\}
\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \left[ \rho_N (1 - S_i) - \rho_n S_i \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar\boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\}\right] \mathbf{X}_i \|_\infty.
\end{align*}
Based on the definition of $\bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j}$, we know
\[
E_{\scriptscriptstyle \Ssc}[\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c_j)\}
\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar\boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\} \mathbf{X}_i]
= E_{\scriptscriptstyle \Tsc}[\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c_j)\}
\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \mathbf{X}_i].
\]
Since $\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \left[ \rho_N (1 - S_i) - \rho_n S_i \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar\boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\}\right] \mathbf{X}_i$ given $\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c_j)\} =1$ is sub-exponential and mean-zero, by union bound inequality and Bernstein inequality we have
\[
\|\widehat\Psc_{\mathcal{S}\cup\mathcal{T}} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c_j)
\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \left[ \rho_N (1 - S_i) - \rho_n S_i \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar\boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\}\right] \mathbf{X}_i \|_\infty \leq_p C_{{\sf roc}, \alpha, 2} \frac{n_j}{n}\sqrt{\frac{\log ~np}{n_j}},
\]
for $j \in \{0,\dots, m\}$.
Similarly, we have
\begin{align*}
\|\widehat\Psc_{\mathcal{S}\cup\mathcal{T}} \{\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq \widehat c_j)
- \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c_j)\}
\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \left[ \rho_N (1 - S_i) - \rho_n S_i \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar\boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\}\right] \mathbf{X}_i \|_\infty \\
\leq_p C^\prime_{{\sf roc}, \alpha, 2} \frac{n^{1/2}}{n} \sqrt{\frac{\log ~np}{n^{1/2}}} = o\left\{\frac{n_j}{n}\sqrt{\frac{\log ~np}{n_j}}\right\}.
\end{align*}
Second, by Assumption \ref{asu:1}, Cauchy-Schwartz inequality and (\ref{eq:alpha_l2}), we have
\begin{align*}
& \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c_j) \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \left[\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\} - \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde\boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\}\right](\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})^{\scriptscriptstyle \sf T} \mathbf{X}_i \\
\leq& \exp\{\kappa \|\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha}\|_1\}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c_j)^2 \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\} \{(\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2} \\
& \times
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\}
\{(\widetilde \boldsymbol{\alpha} - \bar \boldsymbol{\alpha})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2}\\
\leq_p& e^{c_{\alpha, 1}} (C_{\beta,\alpha, 2}s\lambda_{\alpha}^2)^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc} I(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c_j) \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\} \{(\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2}.
\end{align*}
Similarly, we have
\begin{align*}
& \widehat\Psc_{\scriptscriptstyle \Ssc} \{\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq \widehat c_j)
- \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c_j)\} \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} ) \left[\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\} - \exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\widetilde\boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\} \right] \\
&\times (\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})^{\scriptscriptstyle \sf T} \mathbf{X}_i \\
\leq_p& n^{-1/4} e^{c_{\alpha, 1}} (C_{\beta,\alpha, 2}s\lambda_{\alpha}^2)^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc} \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\} \{(\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2} \\
\ll_p&
e^{c_{\alpha, 1}} (C_{\beta,\alpha, 2}s\lambda_{\alpha}^2)^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc} I(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c_j) \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\} \{(\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2}.
\end{align*}
Last inequality holds because there exists a positive constant $C_{{\sf roc}, \alpha,3}$ such that
\begin{align*}
&\widehat\Psc_{\scriptscriptstyle \Ssc} \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\} \{(\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \\
\leq_p &C_{{\sf roc}, \alpha,3} \frac{n}{n_j}
\widehat\Psc_{\scriptscriptstyle \Ssc} I(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c_j) \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\} \{(\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2,
\end{align*}
and $n^{-1/4}\sqrt{n/n_j} = o(1)$.
Under Assumptions \ref{asu:1} and \ref{asu:3},
similar to the proofs regarding $\widehat \boldsymbol{\delta}_{\beta,\alpha}$, there exist positive constants $C_{{\sf roc}, \alpha, 4}$, $C_{{\sf roc}, \alpha, 5}<1$ and $C_{{\sf roc}, \alpha, 6}$ such that \begin{align*}
& e^{-C_{{\sf roc}, \alpha, 4}}
D_{{\sf roc}, \alpha, c_j}(\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j},\bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j}; \bar\boldsymbol{\alpha}, \bar\boldsymbol{\gamma}, \boldsymbol{\beta}_0)
+ \lambda_{{\sf roc}, \alpha, \widehat c_j}\|\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j}\|_1 \\
\leq_p&
C_{{\sf roc}, \alpha, 5} \lambda_{{\sf roc}, \alpha, \widehat c_j}\|\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j}\|_1+ \lambda_{{\sf roc},\alpha, \widehat c_j} \| \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_j}\|_1 \\
+& C_{{\sf roc}, \alpha, 6} (s\lambda_{\alpha}^2)^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\mathbb{I}( \mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0\geq c_j)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\} \{(\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2}.
\end{align*}
We then deal with the soft sparsity of $\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_j}$.
Let $\eta_j>0$ be a threshold and define the threshold subset
\[
S_{\eta_j} = [l\in \{1, \dots, p+q\}:\| (\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})_l\|_1 > \eta_j],
\]
and set $\eta_j \asymp \sqrt{n_j^{-1}\log ~np}$.
By triangle inequalities, we have
\begin{align*}
D^\prime_{{\sf roc}, \alpha, c_j} &\leq_p
2\lambda_{{\sf roc},\alpha,\widehat c_j} \| (\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})_{S_{\eta_j}^c}\|_1
+ 2\lambda_{{\sf roc}, \alpha, \widehat c_j}\|(\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})_{S_{\eta_j}}\|_1\\
+& C_{{\sf roc}, \alpha, 6} (s\lambda_{\alpha}^2)^{1/2}
\left[\widehat\Psc_{\scriptscriptstyle \Ssc}\mathbb{I}( \mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0\geq c_j)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\bar\boldsymbol{\gamma} )\exp\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\bar \boldsymbol{\alpha} + \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})\} \{(\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j})^{\scriptscriptstyle \sf T} \mathbf{X}_i\}^2 \right]^{1/2},
\end{align*}
where
\[
D^\prime_{{\sf roc}, \alpha, c_j} := e^{-C_{{\sf roc}, \alpha, 4}}
D_{{\sf roc}, \alpha, c_j}(\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j},\bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j}; \bar\boldsymbol{\alpha}, \bar\boldsymbol{\gamma}, \boldsymbol{\beta}_0)
+ (1 - C_{{\sf roc}, \alpha, 5})\lambda_{{\sf roc}, \alpha, \widehat c_j}\|\widehat \boldsymbol{\delta}_{{\sf roc}, \alpha, \widehat c_j} - \bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c_j}\|_1.
\]
Under Assumptions \ref{asu:1}, \ref{asu:3}, \ref{asu:4} and \ref{asu:5},
similar to the proofs regarding $\widehat \boldsymbol{\delta}_{\beta, \alpha}$ and $\widehat \boldsymbol{\delta}_{\beta, \gamma}$,
we prove the results in Lemma \ref{lemma:2}.
\end{proof}
\newpage
\section{Additional numerical implementation details and results}
\label{app:B}
\begin{algorithm}[H]
\caption{The Importance Weighting method (IW)}
\textbf{Input:} $\{\mathbf{D}_i : i = 1, 2, \dots, n+N \}$ and preliminary estimators $\widetilde\boldsymbol{\alpha}$.
\textbf{Step I:} Obtain $\widehat \boldsymbol{\beta}_{\rm iw}$ by solving
\[
\widehat\Psc_{\mathcal{S}} \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\alpha}) \mathbf{A}_i \{ Y_i - g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta})\} = \mathbf{0} \, .
\]
\textbf{Step II:} Obtain $\widehat \TPR_{\rm iw}(c)$ and $\widehat \mathbf{FPR}_{\rm iw}(c)$ by
\[
\widehat \TPR_{\rm iw}(c) = \frac{\widehat\Psc_{\mathcal{S}}\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}_{\rm iw} \geq c) \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widetilde \boldsymbol{\alpha} ) Y_i }{\widehat\Psc_{\mathcal{S}} \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widetilde \boldsymbol{\alpha} ) Y_i}\,; \quad \widehat \mathbf{FPR}_{\rm iw}(c) = \frac{\widehat\Psc_{\mathcal{S}}\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}_{\rm iw} \geq c)\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widetilde \boldsymbol{\alpha} ) \{1-Y_i\} }{\widehat\Psc_{\mathcal{S}} \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widetilde \boldsymbol{\alpha} ) \{1-Y_i\} }
\]
\textbf{Output:} $\widehat \boldsymbol{\beta}_{\rm iw}$, $\widehat \ROC_{\rm iw}(u) = \widehat \TPR_{\rm iw}\{\widehat \mathbf{FPR}_{\rm iw}^{-1}(u)\}$ and $\widehat \AUC_{\rm iw} = \int_0^1 \widehat \ROC_{\rm iw}(u) {\rm d} u$.
\label{alg iw}
\end{algorithm}
\begin{algorithm}[H]
\caption{The IMputation only method (IM)}
\textbf{Input:} $\{\mathbf{D}_i : i = 1, 2, \dots, n+N \}$ and preliminary estimators $\widetilde\boldsymbol{\gamma}$.
\textbf{Step I:} Obtain $\widehat \boldsymbol{\beta}_{\rm im}$ by solving
\[
\widehat\Psc_{\mathcal{T}} \mathbf{A}_i \{ g(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma})- g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}) \} = \mathbf{0} \, .
\] \textbf{Step II:} Obtain $\widehat \TPR_{\rm im}(c)$ and $\widehat \mathbf{FPR}_{\rm im}(c)$ by
\[
\widehat \TPR_{\rm im}(c) = \frac{ \widehat\Psc_{\mathcal{T}} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}_{\rm im} \geq c) g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widetilde \boldsymbol{\gamma}) }{ \widehat\Psc_{\mathcal{T}} g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widetilde \boldsymbol{\gamma}) }\,; \quad \widehat \mathbf{FPR}_{\rm im}(c) = \frac{ \widehat\Psc_{\mathcal{T}} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}_{\rm im} \geq c) \{1- g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widetilde \boldsymbol{\gamma}) \} }{ \widehat\Psc_{\mathcal{T}}\{1- g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widetilde \boldsymbol{\gamma}) \} }
\]
\textbf{Output:} $\widehat \boldsymbol{\beta}_{\rm im}$ , $\widehat \ROC_{\rm im}(u) = \widehat \TPR_{\rm im}\{\widehat \mathbf{FPR}_{\rm im}^{-1}(u)\}$ and $\widehat \AUC_{\rm im} = \int_0^1 \widehat \ROC_{\rm im}(u) {\rm d} u$.
\label{alg im}
\end{algorithm}
\begin{table}[H]
\centering
\caption{\label{tab:real beta} Estimators of the target model coefficients. $\beta_0,\beta_1,\beta_2,\beta_3$ represent respectively the intercept, coefficient of T2D risk score, coefficient of gender (1 for Female), coefficient of ethnicity (1 for European). Source: naive source estimator; DRAMATIC: our proposed method; IW: importance weighting method; IM: imputation based method; Target: gold standard benchmark obtained using validation labels.}
\begin{tabular}{lcccccc}
\hline
\hline
& Source & DRAMATIC & IW & IM & Target \\ \hline
$\beta_0$ (Intercept) & -1.400 & -0.750 & -1.097 & -0.967 & -0.430\\
$\beta_1$ (T2D score) & 1.793 & 1.789 & 1.510 & 1.108 & 1.309 \\
$\beta_2$ (Gender) & -0.547 & -0.808 & -0.715 & -0.505 & -0.833 \\
$\beta_3$ (Ethnicity) & -0.869 & -0.925 & -0.780 & -0.652 & -1.019 \\
\hline\hline\\[-1.8ex]
\end{tabular}
\end{table}
\section{Introduction}
\def_{\scriptscriptstyle \Tsc}{_{\scriptscriptstyle \mathcal{T}}}
\def_{\scriptscriptstyle \Ssc}{_{\scriptscriptstyle \mathcal{S}}}
\def\mathbb{P}{\mathbb{P}}
\subsection{Background}
While cohort studies remain to be critical sources for studying disease progression and treatment response, they have limitations, including the generalizability of the study findings to the real world, the limited ability to test broader hypotheses, and the cost associated with performing these studies. In recent years, due to the increasing adoption of electronic health records (EHR) and the linkage of EHR with specimen bio-repositories, integrated large datasets now exist as a new data source for deriving real-world predictive models. For example, the Massachusetts General and Brigham (MGB) Healthcare Biobank \citep{castro2022mass} and the million veteran's project \citep{gaziano2016million} contain a wealth of both clinical and genomic data for patients enrolled in the biobanks. These integrated datasets open opportunities for developing accurate EHR-based personalized risk prediction models, which can be easily integrated into clinical workflow to assist clinical decision-making. These models can also be contrasted and integrated with models derived from traditional cohort studies to improve generalizability.
Efficiently deriving and evaluating prediction models using EHR data remains challenging due to practical and methodological obstacles.
First, gold standard labels on disease outcomes are typically not readily available in EHR and require accessibility of clinical notes and labor-intensive manual annotations. For example, at MGB, only $\sim 74\%$ of patients with at least one diagnostic code of type II diabetes (T2D) have T2D \citep{liao2019high}. On the other hand, verifying true T2D status via manual chart review is laborious and thus only available for a small number of patients at MGB. Existing methods for deriving EHR risk models typically fit prediction models to inaccurate proxies of the true outcome, such as the presence or absence of a billing code or a classification derived from a machine learning algorithm \citep{kurreeman2011genetic}. The misclassification error in the proxies can lead to biases in the derived prediction model and the estimated accuracy of the trained model.
Another challenge associated with ensuring the generalizability of EHR derived model arises from the heterogeneity between different EHR cohorts. It is well-known that patient populations may differ substantially across different healthcare systems, leading to significant heterogeneity in the performance of prediction models \citep{rasmy2018study}. Risk prediction models derived to optimize one EHR cohort may not perform well for another cohort. In addition, data for EHR studies are often refreshed over time as new patients enter the study, and the data of existing patients get updated. Thus, EHR data from two different time points of the same institution can have a significant shift in their distribution.
Both the scarcity of gold standard labels on disease outcomes and the covariate shift poses great challenges to generalizing knowledge about EHR-derived risk models from a labeled source population to a target population where gold standard labels are typically not available. To transfer such knowledge, much recent literature has been focused on transfer learning techniques for model estimation \citep{liu2020doubly,geva2021high}, such as the transportability of prediction algorithms trained from EHR data \citep{weng2020deep}. On the other hand, there is a paucity of literature on transfer learning of model performance. In this paper, we propose a novel {\bf D}oubly {\bf R}obust {\bf A}ugmented {\bf M}odel {\bf A}ccuracy {\bf T}ransfer {\bf I}nferen{\bf C}e (DRAMATIC) procedure to enable point and interval estimation for commonly used classification performance measures in an unlabeled target population using labeled source data. Specifically, the proposed DRAMATIC procedure develops and evaluates a prediction model for a binary response $Y$ given low dimensional predictors $\mathbf{A}$ in the target population by leveraging source data on $Y$ and high dimensional features $\mathbf{X} = (\mathbf{A}^{\scriptscriptstyle \sf T}, \mathbf{W}^{\scriptscriptstyle \sf T})^{\scriptscriptstyle \sf T}$ and only requires to be shared between the source and target data. We next detail the problem of interest and highlight our contributions in view of existing literature.
\subsection{Problem statement}
Let $Y$ denote the binary outcome of interest and $\mathbf{A}_{q\times 1} = (A_1,\dots,A_q)^{\scriptscriptstyle \sf T}$ denote a vector of predictors used to predict $Y$ and the first element of $\mathbf{A}$ being 1. Our goal is to develop and evaluate a risk model predicting $Y$ with $\mathbf{A}$ for a target population $\mathcal{T}$ using observed data on $n$ samples of $(Y, \mathbf{X}^{\scriptscriptstyle \sf T})^{\scriptscriptstyle \sf T}$ in a source population $\mathcal{S}$ and $N$ samples of $\mathbf{X}$ in $\mathcal{T}$, where $\mathbf{X}=(\mathbf{A}^{\scriptscriptstyle \sf T},\mathbf{W}^{\scriptscriptstyle \sf T})^{\scriptscriptstyle \sf T}$ and $\mathbf{W}_{p\times 1} = (W_1,\dots,W_p)^{\scriptscriptstyle \sf T}$ denotes confounding (adjustment) covariates that may relate to $Y$ and/or the distributional shift between $\mathcal{T}$ and $\mathcal{S}$. We use $S=1$ to indicate that the sample is labeled from $\mathcal{S}$ and $S = 0$ for the unlabelled target data. We assume that $q$, the dimension of risk factors, is fixed but allow $p$, the dimension of the covariates $\mathbf{W}$ needed to enable transfer learning, to be high, i.e., $p \gg \min(n,N)$. In an EHR-driven genetic risk prediction study, examples of $\mathbf{A}$ include relevant genetic markers and demographic variables, while $\mathbf{W}$ may include a large number of EHR proxies of $Y$, including relevant diagnostic codes, medication prescriptions, laboratory tests, as well as mentions of clinical terms extracted from free text clinical notes via natural language processing. Hence $Y \mid \mathbf{A}$ may differ between the target and source populations, but $Y \mid \mathbf{A},\mathbf{W}$ is assumed to be the same.
To develop the risk model for $Y \mid \mathbf{A}$ in $\mathcal{T}$, we assume a {\em working} logistic regression model
$$
\mathbb{P}_{\scriptscriptstyle \Tsc}(Y = 1 \mid \mathbf{A}) = \mathbb{P}(Y = 1 \mid \mathbf{A},S=0) = g(\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta}) \quad \mbox{with}\quad g(x)=1/(1+e^{-x}),
$$
which is not required to hold throughout. Here we define $\mathbb{P}_{\scriptscriptstyle \Tsc}(\cdot) = \mathbb{P} ( \cdot \mid S=0)$, $\mathbb{P}_{\scriptscriptstyle \Ssc}(\cdot) = \mathbb{P} ( \cdot \mid S=1)$, $\mathbb{E}_{\scriptscriptstyle \Tsc}[\cdot] = \mathbb{E}[ \cdot \mid S=0]$ and $\mathbb{E}_{\scriptscriptstyle \Ssc}[\cdot] = \mathbb{E}[ \cdot \mid S=1]$.
Let the observed data be $\mathscr{D} = \{\mathbf{D}_i =(S_i Y_i, \mathbf{X}_i^{\scriptscriptstyle \sf T} , S_i)^{\scriptscriptstyle \sf T}: i = 1, 2, \dots, n+N \}$ where we let $S_i=\mathbb{I}(1 \le i \le n)$ without loss of generality and $\mathbb{I}(\cdot)$ is the indicator function. Assume that the samples are generated following
\[
Y,\mathbf{X}\mid S=s\sim p_{\mathbf{X}|S=s}(\mathbf{x})\cdot p_{Y|\mathbf{X}}(y),
\]
where $p_{\mathbf{X}|S=s}(\cdot)$ and $p_{Y|\mathbf{X}}(\cdot)$ represent the probability mass/density function of $\mathbf{X}$ given $S$ and $Y$ given $\mathbf{X}$ respectively. Here, our underlying assumption is that distribution of covariates $\mathbf{X}$ is different across $\mathcal{S}$ and $\mathcal{T}$ while the conditional distribution of the response $Y$ given $\mathbf{X}$ is the same across the two populations. We define the density ratio function as
\begin{equation}
h_0(\mathbf{x}) = p_{\mathbf{X}|S=0}(\mathbf{x})/p_{\mathbf{X}|S=1}(\mathbf{x})
\label{def:density}
\end{equation} and the conditional mean of $Y$ on the two populations as
\begin{equation}
r_0(\mathbf{x})=\mathbb{E}[Y\mid\mathbf{X}=\mathbf{x}]\,.
\label{def:r}
\end{equation}
Let $\boldsymbol{\beta}_0$ be the outcome parameter of our interests that solves the estimating equation:
\[
\mathbb{E}_{\scriptscriptstyle \Tsc} [\mathbf{A}\{Y-g(\mathbf{A}^{\scriptscriptstyle \sf T} \boldsymbol{\beta})\}] = \mathbf{0} \,.
\]
To evaluate the accuracy of the risk model $g(\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0)$ on $\mathcal{T}$, we aim at inferring the commonly used receiver operating characteristic (ROC) curve of $Y$ against $\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0$\footnote{Since $g(\cdot)$ is monotonically increasing, the ROC property of $Y\sim \mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0$ is the same as that of $Y\sim g(\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0)$.}. Specifically, for any given cutoff value $c$, we may classify a patient as high risk if $\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0\ge c$, and low risk otherwise. Then the true positive rate (TPR) and false positive rate (FPR) associated with the classification rule $\mathbb{I}(\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0 \ge c)$ on $\mathcal{T}$ are defined as
\[
\TPR(c) = \mathbb{P}_{\scriptscriptstyle \Tsc}(\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0 \geq c \mid Y=1) \text{ and } \mathbf{FPR}(c) = \mathbb{P}_{\scriptscriptstyle \Tsc}(\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0\geq c \mid Y=0).
\]
Then the trade-off between TPR and FPR can be characterized by the ROC curve: $\ROC(u)=\TPR\{\mathbf{FPR}^{-1}(u)\}$ \citep{gerds2008performance}. The overall classification performance of the risk model can be summarized based on the area under the ROC curve (AUC): $\AUC=\int_0^1\ROC(u) {\rm d} u$. Our goal is to estimate the model parameter $\boldsymbol{\beta}_0$ and the accuracy measures $\ROC(\cdot)$ and $\AUC$ on $\mathcal{T}$ using data $\mathscr{D}$, in which $Y$ is only observed for subjects in $\mathcal{S}$ and the covariates $\mathbf{X}$ characterizing the shift between $\mathcal{S}$ and $\mathcal{T}$ are of high or ultra-high dimensionality.
\subsection{Related literature}
Transfer learning aims at using the information from the source population to aid the task in the target population \citep{torrey2010transfer}, with diverse real-world applications \citep[e.g.]{madden2004transfer, liu2006value,zhu2011heterogeneous,wiens2014study}. Existing machine learning methods have mainly focused on algorithms with little discussion about the theoretical guarantee,
while theoretically justified transfer learning methods have largely focused on estimation of low dimensional regression models. For example, \cite{huang2007correcting} proposed an importance weighting procedure based on kernel mean matching to correct for covariate shift in regression, \cite{reddi2015doubly} proposed a robust transfer learning strategy for removing bias while retaining small variance in parametric regression models, and \cite{LiuandZhang} introduced an imputation model for the response to augment importance weighting and achieved model double robustness for M-estimation.
In the era of big data, more and more application fields like EHR studies involve a potentially high dimensional set of features in knowledge transferring to characterize the distributional shift between the source and target populations. This poses additional methodological challenges to transfer learning due to the inherent biases induced by the regularization needed to overcome the high dimensionality. Transfer learning for high dimensional regression models has been an active area of research in recent years \citep[e.g.]{li2020transfer,he2021transfer, bastani2021predicting,tian2021transfer,li2022transfer}. However, limited literature exists for transfer learning of prediction performance measures in high-dimensional settings.
Existing transfer learning methods for model prediction performance can only handle the distributional shift of low-dimensional features. For example, \cite{rotnitzky2006doubly} developed estimators of the AUC of a single diagnostic test that adjust for selection verification. Their estimator is doubly robust in the sense of being consistent and asymptotically normal if either the outcome model or the selection-into-verification model is correct. \cite{liu2010model} proposed a likelihood approach to estimate the nonignorable parameter, which is not identifiable in the previous works, and proved the asymptotic normality for their AUC estimators. Semiparametric estimation of the covariate specific ROC curves with a partial missing gold standard was developed by \cite{liu2011semiparametric}. Similarly, \cite{long2011robust} aims at estimating the AUC of a disease status given a biomarker value under a setting where the biomarker values are missing for some subjects while disease status is always confirmed and a set of auxiliary variables are fully observed.
The aforementioned methods focused on evaluating the prediction accuracy of a single predictor with low-dimensional adjustment variables for covariate shift. To the best of our knowledge, no existing work considers transfer inference for prediction accuracy of a fitted regression model in the presence of high-dimensional variables for covariates shift, which are substantially challenging due to the excessive (first-order) bias induced by high-dimensional nuisance estimators. Consequently, special debiasing techniques are needed to overcome biases induced by the high-dimensional nuisance models.
The recent works of doubly robust estimation with high-dimensional nuisance models are technically relevant to the calibrated nuisance model estimation step in our DRAMATIC method \citep{tan2020model,ning2020robust,dukes2020inference,liu2020double}. For example, \cite{tan2020model} considered the problem of estimating average treatment effects (ATE) with high dimensional confounding covariates under certain sparse conditions. They proposed using the calibrated estimation equations to reduce the bias incurred by the regularized error of high-dimensional regression. Their estimators are doubly robust, which remains consistent if the propensity score model or the outcome regression model is correctly specified. The calibration idea has also been adopted for ATE estimation in \cite{ning2020robust} via covariate balancing and for estimating the parametric components in a semiparametric model \citep{dukes2020inference}. \cite{liu2020double} estimated the parametric component of a logistic partially linear model by deriving certain moment equations to calibrate the first-order bias of the nuisance models. Further extensions have been recently studied, for example, to alleviate the ultra-sparsity assumptions of the nuisance models at the expense of requiring both models to be correct \citep{bradic2019sparsity}; and to solve general semiparametric estimating equations
\citep{smucler2019unifying,ghosh2020doubly}. However, none of the existing methods can be easily applied to estimate a (working) logistic model and its associated ROC curve. Estimation for the ROC curve is particularly challenging since the estimator needs to converge uniformly for all points on the curve, which necessitates controlling an infinite number of moment equations simultaneously.
\subsection{Our contribution}
Our proposed DRAMATIC method is an efficient transfer learning procedure for estimating the model coefficients and accuracy parameters that overcomes the high dimensionality of the nuisance models while attaining robustness to model misspecifications. It involves a calibration procedure to estimate the targeted logistic regression coefficients and the TPR, FPR, ROC, and AUC parameters and attains the desirable $n^{\frac{1}{2}}$-consistency if either the high dimensional imputation model or the high dimensional density ratio model is correctly specified.
Our work is the first one to consider a doubly robust estimation of the ROC curve of a working classification model under the high-dimensional covariate shift regime. In this process, the excessive regularization bias incurred by the high-dimensional nuisance estimators is removed by constructing a proper set of calibrated moment equations accommodating all cutoff points on the ROC curve. In specific, we propose a ``nearest quantile" strategy that ensures computation efficiency and the uniform $n^{\frac{1}{2}}$-convergence of the ROC estimators by avoiding the demand of fitting the calibrated nuisance models on the extreme left side of the ROC curve, which would result in excessive estimation errors due to the poor effective sample size. Thus, our method also exhibits novelty in view of calibrating high-dimensional nuisance models for semiparametric estimation because non-existing methods in this field can be used to calibrate the doubly robust estimators of a curve with infinite points and achieve simultaneous convergence.
\subsection{Outline of the paper}
The remainder of the paper is organized as follows. In Section \ref{sec:method}, we formally introduce the DRAMATIC estimator for $\boldsymbol{\beta}_0$ and the accuracy parameters. In Section \ref{sec:thm}, we present the
large sample properties of our proposed estimators. In Section \ref{sec:simu}, we study the finite sample performance of our method and
compare it with existing methods under various simulation settings. In Section \ref{sec:data analysis}, we apply DRAMATIC to train and evaluate a type II diabetes (T2D) genetic risk prediction model for a patient cohort of 2021 using labeled source data collected in 2014.
\def\boldsymbol{\delta}{\boldsymbol{\delta}}
\def\mathcal{F}{\mathcal{F}}
\def\mathcal{G}{\mathcal{G}}
\def\mathcal{L}{\mathcal{L}}
\section{DRAMATIC Procedure}
\label{sec:method}
\subsection{Outline and technical challenges}\label{sec:method:dr:cons}
Let $\widehat\Psc_{\scriptscriptstyle \Ssc}=n^{-1} \sum_{i=1}^n$, $\widehat\Psc_{\scriptscriptstyle \Tsc}=N^{-1} \sum_{i=n+1}^{N+n}$ and $\widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}={(N+n)}^{-1} \sum_{i=1}^{N+n}$ represent the empirical mean operators for the source samples, target samples and their union, respectively. Two common approaches to tackle the problem are (i) importance weighting and (ii) imputation. The former uses an estimated density ratio function $\widehat h(\cdot)$ to weight the source data which adjusts for covariate shift by solving the estimation equation
\begin{equation}
\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat{h}(\mathbf{X}_i) \mathbf{A}_i \{ Y_i - g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}) \} = \mathbf{0}\,,
\label{eq1}
\end{equation}
and the latter imputes the outcomes of the target data using an imputation function $\widehat r (\cdot)$ fitted by the source data, then solves the estimation equation
\begin{equation}
\widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{A}_i \{ \widehat r (\mathbf{X}_i) - g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}) \} = \mathbf{0} \,.
\label{eq2}
\end{equation}
Although the two approaches are simple and natural, they are prone to excessive bias due to the potential model misspecification and the estimation errors of the nuisance parameters. Specifically, the importance weighting method leads to an inconsistent solution when the density ratio model $h_0$ is misspecified, and the imputation method is invalid when the imputation model $r_0$ is misspecified.
We overcome these challenges by augmenting \eqref{eq1} and \eqref{eq2} to yield a consistent estimator of $\boldsymbol{\beta}_0$ when either the density ratio model or the imputation model is correctly specified. To this end, we first specify the two high dimensional nuisance models in parametric forms: $h(\mathbf{x}) = \exp(\mathbf{x}^{\scriptscriptstyle \sf T}\boldsymbol{\alpha})$ and $r(\mathbf{x})=g(\mathbf{x}^{\scriptscriptstyle \sf T}\boldsymbol{\gamma})$, which are not required to hold simultaneously. Let $\widehat\boldsymbol{\alpha}$ and $\widehat\boldsymbol{\gamma}$ be some estimators for the nuisance parameters $\boldsymbol{\alpha}$ and $\boldsymbol{\gamma}$, obtained using $\ell_1$-regularized regression. We then introduce the doubly robust estimator for $\boldsymbol{\beta}$, denoted by $\widehat\boldsymbol{\beta}$, as the solution to
\begin{equation}
\mathcal{E}(\boldsymbol{\beta}, \widehat\boldsymbol{\alpha}, \widehat\boldsymbol{\gamma} ) := \widehat\Psc_{\scriptscriptstyle \Ssc} \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\alpha}) \mathbf{A}_i \{ Y_i - g(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\gamma})\} + \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{A}_i \{ g(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\gamma}) - g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}) \} = \mathbf{0} \,.
\label{aeq}
\end{equation}
The double robustness of the estimation equation \eqref{aeq} is validated in the following lemma.
\begin{proposition}
For $\boldsymbol{\alpha}_0 \in \mathbb{R}^p$ and $\boldsymbol{\gamma}_0 \in \mathbb{R}^p$, let $\bar \boldsymbol{\beta}$ be the solution of $\mathbb{E} \mathcal{E}(\boldsymbol{\beta}, \boldsymbol{\alpha}_0, \boldsymbol{\gamma}_0) = \mathbf{0}$. If $\boldsymbol{\alpha}_0$ satisfies $h_0(\mathbf{x}) = \exp(\mathbf{x}^{\scriptscriptstyle \sf T}\boldsymbol{\alpha}_0)$ for $h_0(\mathbf{x})$ defined in \eqref{def:density} or $\boldsymbol{\gamma}_0$ satisfies $r_0(\mathbf{x})=g(\mathbf{x}^{\scriptscriptstyle \sf T}\boldsymbol{\gamma}_0)$ for $r_0(\mathbf{x})$ defined in \eqref{def:r}, then $\bar \boldsymbol{\beta} = \boldsymbol{\beta}_0$.
\label{lem:double}
\end{proposition}
\begin{proof}
When $h_0(\mathbf{x}) = \exp(\mathbf{x}^{\scriptscriptstyle \sf T}\boldsymbol{\alpha}_0)$,
\begin{equation*}
\begin{aligned}
\mathbb{E} \mathcal{E}(\boldsymbol{\beta}, \boldsymbol{\alpha}_0, \boldsymbol{\gamma}_0) = &\mathbb{E}_{\scriptscriptstyle \Ssc} h_0 (\mathbf{X}) \mathbf{A} \{ Y-g(\mathbf{A}^{\scriptscriptstyle \sf T} \boldsymbol{\beta})\} + \mathbb{E}_{\scriptscriptstyle \Ssc} h_0 (\mathbf{X}) \mathbf{A} \{ g(\mathbf{A}^{\scriptscriptstyle \sf T} \boldsymbol{\beta}) - g(\mathbf{x}^{\scriptscriptstyle \sf T}\boldsymbol{\gamma}_0)\} \\
& + \mathbb{E}_{\scriptscriptstyle \Tsc} \mathbf{A} \{ g(\mathbf{x}^{\scriptscriptstyle \sf T}\boldsymbol{\gamma}_0) - g(\mathbf{A}^{\scriptscriptstyle \sf T} \boldsymbol{\beta}) \}=\mathbb{E}_{\scriptscriptstyle \Tsc} \mathbf{A} \{ Y -g(\mathbf{A}^{\scriptscriptstyle \sf T} \boldsymbol{\beta})\} + \mathbf{0}.
\end{aligned}
\end{equation*}
When $r_0(\mathbf{x}) = g(\mathbf{x}^{\scriptscriptstyle \sf T}\boldsymbol{\gamma}_0)$,
$\mathbb{E} \mathcal{E}(\boldsymbol{\beta}, \boldsymbol{\alpha}_0, \boldsymbol{\gamma}_0) = \mathbf{0} + \mathbb{E}_{\scriptscriptstyle \Tsc} \mathbf{A} \{ r_0(\mathbf{X}) - g(\mathbf{A}^{\scriptscriptstyle \sf T} \boldsymbol{\beta}) \} = \mathbb{E}_{\scriptscriptstyle \Tsc} \mathbf{A} \{ Y- g(\mathbf{A}^{\scriptscriptstyle \sf T} \boldsymbol{\beta}) \}$. Both cases lead to solution $\boldsymbol{\beta}_0$.
\end{proof}
However, unlike the low dimensional setting, due to both the potential model misspecification and the excessive estimation errors of the $\ell_1$-regularized nuisance estimators, simply plugging in arbitrary estimators $\widehat\boldsymbol{\alpha}$ and $\widehat\boldsymbol{\beta}$ in \eqref{aeq} could result in excessive first order bias and thus estimators converging slower than the desirable parametric rate even when one nuisance model is correct. Inspired by \cite{tan2020model}, we propose a bias calibration method detailed in Section \ref{sec:method:cal:out} to handle this problem.
For the accuracy parameters $\TPR(c)$ and $\mathbf{FPR}(c)$, we propose doubly robust estimators $\widehat \TPR(c)$ and $\widehat \mathbf{FPR}(c)$ through augmentation similar to the construction in \eqref{aeq}. Specifically, for any given cutoff value $c$, we construct the doubly robust estimators for the proportion of the true positives $P(\mathbf{A}_i^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0 \ge c, Y_i=1)$ and the proportion of false positives $P(\mathbf{A}_i^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0 \ge c, Y_i = 0)$ as:
\begin{align*}
& \widehat \TP(c; \widehat \boldsymbol{\beta}, \widehat\boldsymbol{\alpha}_c, \widehat\boldsymbol{\gamma}_c) = \widehat\Psc_{\scriptscriptstyle \Ssc}\mathbb{I}( \mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq c) \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\alpha}_c) \{Y_i-g(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\gamma}_c)\} + \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c) g(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\gamma}_c) \,, \\
& \widehat \mathbf{FP}(c; \widehat \boldsymbol{\beta}, \widehat\boldsymbol{\alpha}_c, \widehat\boldsymbol{\gamma}_c) = \widehat\Psc_{\scriptscriptstyle \Ssc} \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq c) \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\alpha}) \{g(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\gamma}_c) - Y_i\} + \widehat\Psc_{\scriptscriptstyle \Tsc}\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq c) \{1-g(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\gamma}_c)\} \,,
\end{align*}
where the nuisance parameters $\widehat\boldsymbol{\alpha}_c$ and $\widehat\boldsymbol{\gamma}_c$ are dependent on $c$ as will be shown later. Then we take the TPR and FPR estimators as
\begin{align}
& \widehat \TPR(c) := \widehat \TPR(c; \widehat \boldsymbol{\beta}, \widehat\boldsymbol{\alpha}_c, \widehat\boldsymbol{\gamma}_c, \widehat\boldsymbol{\alpha}_{-\infty}, \widehat\boldsymbol{\gamma}_{-\infty}) = \frac{\widehat \TP(c; \widehat \boldsymbol{\beta}, \widehat\boldsymbol{\alpha}_c, \widehat\boldsymbol{\gamma}_c)}{\widehat \TP(-\infty; \widehat \boldsymbol{\beta}, \widehat\boldsymbol{\alpha}_{-\infty}, \widehat\boldsymbol{\gamma}_{-\infty})} \,, \label{equ:roc:dr0} \\
& \widehat \mathbf{FPR}(c) := \widehat \mathbf{FPR}(c; \widehat \boldsymbol{\beta}, \widehat\boldsymbol{\alpha}_c, \widehat\boldsymbol{\gamma}_c, \widehat\boldsymbol{\alpha}_{-\infty}, \widehat\boldsymbol{\gamma}_{-\infty}) = \frac{\widehat \mathbf{FP}(c; \widehat \boldsymbol{\beta}, \widehat\boldsymbol{\alpha}_c, \widehat\boldsymbol{\gamma}_c)}{\widehat \mathbf{FP}(-\infty; \widehat \boldsymbol{\beta}, \widehat\boldsymbol{\alpha}_{-\infty}, \widehat\boldsymbol{\gamma}_{-\infty})} \,.
\label{equ:roc:dr}
\end{align}
Here, for different cut-off $c$ (and $-\infty$), we plug-in different estimators for the nuisance parameters $\boldsymbol{\alpha}_0$ and $\boldsymbol{\gamma}_0$ in the doubly robust estimators of TPR and FPR since the calibration is specific to $c$.
The estimators for ROC and AUC can be constructed directly using the TPR and FPR estimators in \eqref{equ:roc:dr0} and \eqref{equ:roc:dr}, and maintain a similar doubly robust property.
\begin{remark}
Constructing calibrated estimators $\widehat\boldsymbol{\alpha}_c$ and $\widehat\boldsymbol{\gamma}_c$ is key to achieving doubly robust properties for the accuracy parameters in the high dimensional setting since
the nuisance estimators' excessive bias can prevent them from achieving the parametric rate. We detail in Section \ref{sec:cal:roc} the moment conditions needed to constraint the nuisance estimators to overcome the bias. Eliminating bias for the ROC estimators is substantially more challenging than that of $\boldsymbol{\beta}_0$ since the curve involves an infinite number of points, leading to both statistical and computational challenges. Specifically, the effective sample size for a large cutoff value $c$ can be much smaller than $n$, i.e., $\widehat\Psc_{\scriptscriptstyle \Ssc}\mathbb{I}( \mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq c)$ can be small, which deteriorates the convergence rate of $\widehat \TPR(c)$ and $\widehat \mathbf{FPR}(c)$.
\end{remark}
\begin{remark}
Carrying out the calibration step for all observed unique values of $\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat\boldsymbol{\beta}$ is computationally prohibitive when $n$ is not small. To solve these problems, we propose a novel ``nearest" quantile method that ensures a desirable convergence rate and requires fitting an acceptable number of calibrating equations.
\end{remark}
We outline the DRAMATIC method in Algorithm \ref{alg} and detail the key steps in sections \ref{sec:method:cal:out} and \ref{sec:cal:roc}. We further describe in section \ref{sec:method:boot} a multiplier bootstrap method for inference, whose validity is ensured by the asymptotic normality of the proposed estimators as established in section \ref{sec:thm}.
\begin{algorithm
\caption{Doubly Robust Augmented Model Accuracy Transfer InferenCe (DRAMATIC)} \label{alg}
\textbf{Input:} $\mathscr{D} = \{\mathbf{D}_i =(S_i Y_i, \mathbf{X}_i^{\scriptscriptstyle \sf T} , S_i)^{\scriptscriptstyle \sf T}: i \in [n+N] \}$;
\\
\\
Constructing DR estimator $\widehat \boldsymbol{\beta}$ of $\boldsymbol{\beta}_0$:
\quad \textbf{1.1:} Obtain the preliminary nuisance estimators $\widetilde \boldsymbol{\alpha}$ and $\widetilde \boldsymbol{\gamma}$ by \eqref{eq:initial alpha}.
\quad \textbf{1.2:} Obtain the preliminary estimator $\widetilde \boldsymbol{\beta}$ of $\boldsymbol{\beta}_0$ by solving $\mathcal{E}(\boldsymbol{\beta}, \widetilde\boldsymbol{\alpha}, \widetilde\boldsymbol{\gamma} ) = \mathbf{0}$ defined by \eqref{aeq}.
\quad \textbf{1.3:} Obtain the calibrated nuisance estimators $\widehat \boldsymbol{\alpha}_{\beta}$ and $\widehat \boldsymbol{\gamma}_{\beta}$ by \eqref{equ:cal:beta:alpha} based on $\widetilde \boldsymbol{\alpha}$, $\widetilde \boldsymbol{\gamma}$ and $\widetilde \boldsymbol{\beta}$.
\quad \textbf{1.4:} Obtain the DR estimator $\widehat \boldsymbol{\beta}$ by solving $\mathcal{E}(\boldsymbol{\beta}, \widehat\boldsymbol{\alpha}_\beta, \widehat\boldsymbol{\gamma}_\beta ) = \mathbf{0}$.
\\
\\
Constructing DR estimators $\TPR(c)$, $\mathbf{FPR}(c)$, $\ROC(u)$ and $\AUC$:
\quad \textbf{2.1:} Obtain the calibrated nuisance estimators $\widehat\boldsymbol{\alpha}_{{\sf roc},\widehat c_j}$ and $\widehat\boldsymbol{\gamma}_{{\sf roc},\widehat c_j}$ by \eqref{equ:cal:roc}.
\quad \textbf{2.2:} Obtain the DR estimators $\widehat \TPR(c)$ and $\widehat \mathbf{FPR}(c)$ of $\TPR(c)$ and $\mathbf{FPR}(c)$ by \eqref{equ:roc:dr:cal}.
\\
\\
\textbf{Output:}
$\widehat \boldsymbol{\beta}$, $\widehat \ROC(u) = \widehat \TPR\{\widehat \mathbf{FPR}^{-1}(u)\}$ for any $u\in[0,1]$, and $\widehat \AUC = \int_0^1 \widehat \ROC(u) {\rm d} u$.
\end{algorithm}
\subsection{Calibrated estimation of the outcome parameters}\label{sec:method:cal:out}
Inspired by \cite{tan2020model} and \cite{dukes2020inference}, we construct calibrated nuisance estimators according to certain moment conditions, to correct for their regularization bias. First, we obtain preliminary estimators for $\boldsymbol{\gamma}$ in the imputation model $\mathbb{E}[Y\mid \mathbf{X} = \mathbf{x}] = g(\mathbf{x}^{\scriptscriptstyle \sf T}\boldsymbol{\gamma})$ and $\boldsymbol{\alpha}$ in density ratio $h(\mathbf{x})=\exp(\mathbf{x}^{\scriptscriptstyle \sf T}\boldsymbol{\alpha})$ as
\begin{equation}
\begin{aligned}
\widetilde\boldsymbol{\alpha}=&\argmin{\boldsymbol{\alpha}\in\mathbb{R}^{q+p}}\widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}\{\rho_nS_i \exp (\mathbf{X}_i ^{\scriptscriptstyle \sf T} \boldsymbol{\alpha}) - \rho_N(1-S_i)\mathbf{X}_i ^{\scriptscriptstyle \sf T} \boldsymbol{\alpha} \} + \lambda_{\alpha} \|\boldsymbol{\alpha}\|_1;\\
\widetilde\boldsymbol{\gamma}=&\argmin{\boldsymbol{\gamma}\in\mathbb{R}^{q+p}}\widehat\Psc_{\scriptscriptstyle \Ssc} \{ -Y_i \mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma} + G(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma})\} + \lambda_{\gamma} \|\boldsymbol{\gamma}\|_1,
\end{aligned}
\label{eq:initial alpha}
\end{equation}
where $G(a)=\int_{0}^ag(u) {\rm d} u = {\rm ln}(1+e^a)- {\rm ln} 2$, $\rho_n=(N+n)/n$, $\rho_N=(N+n)/N$ and $\lambda_{\gamma}$, $\lambda_{\alpha}$ are two tuning parameters. Then we solve
\begin{equation*}
\mathcal{E}(\boldsymbol{\beta}, \widetilde\boldsymbol{\alpha}, \widetilde\boldsymbol{\gamma} ) = \mathbf{0}
\end{equation*}
to obtain a preliminary estimator $\widetilde\boldsymbol{\beta}$. According to our discussion in Section \ref{sec:method:dr:cons}, $\widetilde\boldsymbol{\beta}$ is consistent for the target $\boldsymbol{\beta}_0$ when either the density ratio or the imputation model is correctly specified, while it may not achieve the desirable parametric rate $n^{-1/2}$ that is crucial for asymptotic inference.
Denote the information matrix as $\bar\boldsymbol{\Sigma}_{\boldsymbol{\beta}} = \mathbb{E}_{\scriptscriptstyle \Ssc}[\dot{g}(\mathbf{A}^{\scriptscriptstyle \sf T} \boldsymbol{\beta}) \mathbf{A} \mathbf{A}^{\scriptscriptstyle \sf T}]$ and let $\widehat{\boldsymbol{\Sigma}}_{\boldsymbol{\beta}} = \widehat\Psc_{\scriptscriptstyle \Tsc} \dot{g}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}) \mathbf{A}_i \mathbf{A}_i^{\scriptscriptstyle \sf T}$ where $\dot g(\cdot)$ is the derivative of $g(\cdot)$. Given $\mathbf{e}_j \in \mathbb{R}^q$ the $j$th unit vector in $\mathbb{R}^q$ with its $j$th element being one and other elements being zero, we now focus on estimating $\mathbf{e}_j ^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0$. Through asymptotic expansion, we find that the first order bias of $\sqrt{n} (\mathbf{e}_j^{\scriptscriptstyle \sf T} \widehat{\boldsymbol{\beta}}- \mathbf{e}_j^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0)$ can be potentially controlled by the two terms\footnote{We subscript the moment conditions and estimators with the symbol $\beta$ to distinguish them from those used for ROC estimation in Section \ref{sec:cal:roc}.}
\begin{equation}
\begin{split}
\mathcal{M}_{\beta_j,\alpha}(\boldsymbol{\alpha})&=: \big\|\widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}\widehat w_{ji}(\widetilde\boldsymbol{\beta})\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} )\left\{\rho_n S_i \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T}\boldsymbol{\alpha})\mathbf{X}_i-(1-S_i)\rho_N \mathbf{X}_i \right\}\big\|_{\infty} \,;\\
\mathcal{M}_{\beta_j,\gamma}(\boldsymbol{\gamma})&=:\big\|\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_{ji}(\widetilde\boldsymbol{\beta})\exp(\mathbf{X}_i ^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\alpha})\{ Y_i - g(\mathbf{X}_i^{\scriptscriptstyle \sf T}\boldsymbol{\gamma})\} \mathbf{X}_i \big\|_{\infty} \,,
\end{split}
\label{equ:mom:cons}
\end{equation}
where $\widehat w_{ji}(\boldsymbol{\beta})=\mathbf{e}_j^{\scriptscriptstyle \sf T} \widehat\boldsymbol{\Sigma}_{\boldsymbol{\beta}}^{-1} \mathbf{A}_i$ and for simplicity, we denote by $\widehat w_{ji}=\widehat w_{ji}(\widetilde\boldsymbol{\beta})$. The detailed derivation is presented in Appendix~\ref{sec:app:proof}. It implies that when at least one nuisance model is correctly specified, one can reduce the impact of fitting bias of the nuisance estimators plugged in \eqref{aeq} through controlling \eqref{equ:mom:cons}. To this end, we conduct the following regularized regressions to calibrate the nuisance estimators:
\begin{equation}
\begin{split}
\widehat\boldsymbol{\delta}_{\beta_j,\alpha}&=\argmin{\boldsymbol{\delta}\in\mathbb{R}^{q+p}}\widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}\widehat w_{ji}\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} )\mathcal{F}_i(\boldsymbol{\delta};\widetilde\boldsymbol{\alpha}) + \lambda_{\beta_j,\alpha} \|\boldsymbol{\delta}\|_1 \, ;\\
\widehat\boldsymbol{\delta}_{\beta_j,\gamma}&=\argmin{\boldsymbol{\delta}\in\mathbb{R}^{q+p}}\widehat\Psc_{\scriptscriptstyle \Ssc}\widehat w_{ji}\exp(\mathbf{X}_i ^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\alpha})\mathcal{G}_i(\boldsymbol{\delta};\widetilde\boldsymbol{\gamma})+ \lambda_{\beta_j,\gamma} \|\boldsymbol{\delta}\|_1 \,,
\end{split}
\label{equ:cal:beta:alpha}
\end{equation}
and set $\widehat\boldsymbol{\alpha}_{\beta_j} = \widetilde\boldsymbol{\alpha}+ \widehat\boldsymbol{\delta}_{\beta_j,\alpha}$, $\widehat\boldsymbol{\gamma}_{\beta_j} = \widetilde\boldsymbol{\gamma} + \widehat\boldsymbol{\delta}_{\beta_j,\gamma}$, where
\begin{align*}
\mathcal{F}_i(\boldsymbol{\delta};\boldsymbol{\alpha})& = \rho_nS_i \exp \{\mathbf{X}_i ^{\scriptscriptstyle \sf T} (\boldsymbol{\alpha}+\boldsymbol{\delta})\} - \rho_N(1-S_i)\mathbf{X}_i ^{\scriptscriptstyle \sf T} (\boldsymbol{\alpha}+\boldsymbol{\delta}) \, ;\\
\mathcal{G}_i(\boldsymbol{\delta};\boldsymbol{\gamma})&=-Y_i \mathbf{X}_i^{\scriptscriptstyle \sf T} (\boldsymbol{\gamma}+\boldsymbol{\delta}) + G\{\mathbf{X}_i^{\scriptscriptstyle \sf T} (\boldsymbol{\gamma}+\boldsymbol{\delta})\}\,.
\end{align*}
Tuning strategy of the parameters $\lambda_{\beta_j,\alpha}$ and $\lambda_{\beta_j,\gamma}$ is introduced in Section \ref{sec:method:tune} and their rates are given in Section \ref{sec:thm}. Karush--Kuhn--Tucker (KKT) conditions of the LASSO problems \eqref{equ:cal:beta:alpha} guarantee the $\ell_{\infty}$-terms defined in (\ref{equ:mom:cons}) to be controlled by the parameters $\lambda_{\beta_j,\alpha}$ and $\lambda_{\beta_j,\gamma}$, i.e. $\mathcal{M}_{\beta_j,\alpha}(\widehat\boldsymbol{\alpha}_{\beta_j}) \leq \lambda_{\beta_j,\alpha}$ and $\mathcal{M}_{\beta_j,\gamma}(\widehat\boldsymbol{\gamma}_{\beta_j}) \leq \lambda_{\beta_j,\gamma}$. Then we plug $\widehat\boldsymbol{\gamma}_{\beta_j}$ and $\widehat\boldsymbol{\alpha}_{\beta_j}$ in \eqref{aeq} to solve $\widehat\boldsymbol{\beta}$ from
\begin{equation*}
\mathcal{E}(\boldsymbol{\beta}, \widehat\boldsymbol{\alpha}_{\beta_j}, \widehat\boldsymbol{\gamma}_{\beta_j}) = \mathbf{0} \,,
\end{equation*}
and obtain $\mathbf{e}_j^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\beta}$ as the estimator for $\mathbf{e}_j^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0$. To obtain the estimator for the whole $\boldsymbol{\beta}_0$, one can follow this procedure for $j = 1, \ldots, q$ to estimate each entry of $\boldsymbol{\beta}_0$ separately and concatenate them together. With a little abuse of notation, we still denote the resulted vector estimator as $\widehat\boldsymbol{\beta}$.
\begin{remark}
Different from \cite{tan2020model}, the weights $\widehat w_{ji}$'s may not be positive definite since we aim at transferring a regression model but not simple average parameters. For example, if $\mathbf{A}$ has entrywise symmetric distribution (except the intercept), it is possible that approximately half of $\widehat w_{ji}$'s are positive and the other half are negative. Hence, the loss function in (\ref{equ:cal:beta:alpha}) can be irregular and ill-posed. To handle this issue, we follow a similar idea as \cite{liu2020doubly} to divide the samples into two sets with positive and negative only $\widehat w_{ji}$'s respectively and solve
\begin{equation}
\begin{split}
\widehat\boldsymbol{\delta}_{\beta_j,\alpha}^{s}&=\argmin{\boldsymbol{\delta}\in\mathbb{R}^{q+p}}\widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}I_{ji}^s |\widehat w_{ji}| \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} )\mathcal{F}_i(\boldsymbol{\delta};\widetilde\boldsymbol{\alpha}) + \lambda_{\beta,\alpha}^s \|\boldsymbol{\delta}\|_1 \, ;\\
\widehat\boldsymbol{\delta}_{\beta_j,\gamma}^s &= \argmin{\boldsymbol{\delta}\in\mathbb{R}^{q+p}} \widehat\Psc_{\scriptscriptstyle \Ssc} I_{ji}^s |\widehat w_{ji}| \exp(\mathbf{X}_i ^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\alpha})\mathcal{G}_i(\boldsymbol{\delta};\widetilde\boldsymbol{\gamma})+ \lambda_{\beta,\gamma}^s \|\boldsymbol{\delta}\|_1 \\
\end{split}
\label{equ:cal:sign}
\end{equation}
for $s = +, -$ separately instead of \eqref{equ:cal:beta:alpha}, where $I_{ji}^+ = \mathbb{I}(\widehat w_{ji}>0)$, $I_{ji}^- = \mathbb{I}(\widehat w_{ji} \leq 0)$, and $\lambda_{\beta_j,\alpha}^s,\lambda_{\beta_j,\gamma}^s$ are tuning parameters. Correspondingly, we take $\widehat\boldsymbol{\alpha}^s_{\beta_j} = \widetilde \boldsymbol{\alpha}+\widehat\boldsymbol{\delta}_{\beta_j,\alpha}^s$ and $\widehat\boldsymbol{\gamma}^s_{\beta_j}= \widetilde\boldsymbol{\gamma}+\widehat\boldsymbol{\delta}_{\beta_j,\gamma}^s$, and solve $\mathbf{e}_j^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\beta}$ from
\[
\widehat\Psc_{\scriptscriptstyle \Ssc} \widehat h_{\beta_j}(\mathbf{X}_i) \mathbf{A}_i \{ Y_i -\widehat r_{\beta_j}(\mathbf{X}_i)\} + \widehat\Psc_{\scriptscriptstyle \Tsc} \mathbf{A}_i \{ \widehat r_{\beta_j}(\mathbf{X}_i)- g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}) \} = \mathbf{0},
\]
where $\widehat h_{\beta_j}(\mathbf{X}_i)=\mathbb{I}(\widehat w_{ji} > 0)\exp (\mathbf{X}_i ^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\alpha}^+_{\beta_j} ) + \mathbb{I}(\widehat w_{ji} \leq0)\exp ( \mathbf{X}_i ^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\alpha}^-_{\beta_j} )$ and $\widehat r_{\beta_j}(\mathbf{X}_i)=\mathbb{I}(\widehat w_{ji} > 0)g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\gamma} ^+_{\beta_j}) + \mathbb{I}(\widehat w_{ji} \leq 0)g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\gamma}^-_{\beta_j})$, for $i = 1, \ldots, N+n$.
\end{remark}
\begin{remark}
Different from \cite{tan2020model} and \cite{dukes2020inference}, we use an estimating procedure similar to Trans-Lasso proposed by \citep{li2020transfer} to construct the nuisance model estimators for both the model coefficients and ROC parameters. As is shown in our asymptotic and finite studies, this strategy can ensure more precise and stable nuisance estimators when the calibrated estimating equations have reduced sample size or effective sample size due to the sample splitting that appears in equations (\ref{equ:cal:sign}) and (\ref{equ:cal:roc}).
\label{rem:1}
\end{remark}
\subsection{Calibrated estimation of the ROC parameters}\label{sec:cal:roc}
Based on the estimator $\widehat \boldsymbol{\beta}$, we use \eqref{equ:roc:dr0} and \eqref{equ:roc:dr} to construct the doubly robust estimators for $\TPR(c)$ and $\mathbf{FPR}(c)$ simultaneously for all $c\in\mathbb{R}$.
Similar to Section \ref{sec:method:cal:out}, { our goal is to derive calibrated nuisance estimators to mitigate the first order biases of $\sqrt{n} \{ \widehat \TPR(c) - \TPR(c) \}$ and $\sqrt{n} \{ \widehat \mathbf{FPR}(c) - \mathbf{FPR}(c) \}$. By the asymptotic analysis in Appendix~\ref{proof thm2}, with the initial estimators $\widetilde\boldsymbol{\alpha}$ and $\widetilde\boldsymbol{\gamma}$, we find these biases can be controlled by constructing nuisance estimators to control the error terms}
\begin{equation}
\begin{split}
&\mathcal{M}_{{\sf roc},\alpha}(\boldsymbol{\alpha};c)=: \big\|\widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq c)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T} \widetilde\boldsymbol{\gamma}) \{\rho_n S_i \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T}\boldsymbol{\alpha})\mathbf{X}_i- \rho_N (1-S_i) \mathbf{X}_i \}\big\|_{\infty} \,;\\
&\mathcal{M}_{{\sf roc},\gamma}(\boldsymbol{\gamma};c)=:\big\|\widehat\Psc_{\scriptscriptstyle \Ssc}\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq c)\exp(\mathbf{X}_i ^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\alpha})\mathbf{X}_i\{ Y_i - g(\mathbf{X}_i^{\scriptscriptstyle \sf T}\boldsymbol{\gamma})\}\big\|_{\infty} \,.
\end{split}
\label{equ:mom:cons:roc}
\end{equation}
Controlling the terms in \eqref{equ:mom:cons:roc} is more challenging than \eqref{equ:mom:cons} since we need to control the empirical processes $\mathcal{M}_{{\sf roc},\alpha}(\boldsymbol{\alpha};c)$ and $\mathcal{M}_{{\sf roc},\gamma}(\boldsymbol{\gamma};c)$ uniformly for $c\in\mathbb{R}$.
The effective sample size for estimating TPR and FPR is also much smaller than $n$ for larger $c$, leading to additional statistical challenges. We propose a ``nearest quantile" strategy that only calibrates for a sequence of
cutoff values (quantiles) decided from the data.
Then each point on the ROC curve will be close enough to its nearest calibrated quantile. Hence the approximation bias between them could be negligible, and one can plug in the calibrated nuisance estimators of its nearest quantile to derive the doubly robust estimator for each point.
Specifically, we split the source data into $m$ segments with equal sizes $n_{\min} = \lceil n/m \rceil$ according to the order of $\mathbf{A}_i^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\beta}$, where $ \lceil x \rceil$ is the smallest integer larger or equal to $x$. Without loss of generality, we assume that $\mathbf{A}_1^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq \mathbf{A}_2^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq \ldots \mathbf{A}_n^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}$. We then choose the
quantiles as $\widehat c_j = \mathbf{A}_{n_j}^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\beta}$ for $j \in [m]$ where $n_1 = n, n_2 = (m-1) n_{\min}$, $\ldots$, $n_m=n_{\min}$ as illustrated in Figure \ref{fig:roc_ill}.
To control the moment terms \eqref{equ:mom:cons:roc} corresponding to each $\widehat c_j$, we propose to solve the following calibration equations for $c \in \mathcal{C}_{\sf cal} := \{\widehat{c}_j:j=1,\ldots,m\}$
\begin{equation}
\begin{split}
\widehat\boldsymbol{\delta}_{{\sf roc},\alpha, c}&=\argmin{\boldsymbol{\delta}\in\mathbb{R}^{q+p}}\widehat\Psc_{{\scriptscriptstyle \Ssc}{\scriptscriptstyle \cup}{\scriptscriptstyle \Tsc}}\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq c) \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} )\mathcal{F}_i(\boldsymbol{\delta};\widetilde\boldsymbol{\alpha}) + \lambda_{{\sf roc},\alpha,c} \|\boldsymbol{\delta}\|_1,\\
\widehat\boldsymbol{\delta}_{{\sf roc},\gamma,c}&=\argmin{\boldsymbol{\delta}\in\mathbb{R}^{q+p}}\widehat\Psc_{\scriptscriptstyle \Ssc}\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq c) \exp(\mathbf{X}_i ^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\alpha})\mathcal{G}_i(\boldsymbol{\delta};\widetilde\boldsymbol{\gamma})+ \lambda_{{\sf roc},\gamma,c} \|\boldsymbol{\delta}\|_1,
\end{split}
\label{equ:cal:roc}
\end{equation}
and set $\widehat\boldsymbol{\alpha}_{{\sf roc},c}=\widetilde\boldsymbol{\alpha}+\widehat\boldsymbol{\delta}_{{\sf roc},\alpha,t(c)}$, $\widehat\boldsymbol{\gamma}_{{\sf roc},c}=\widetilde\boldsymbol{\gamma}+\widehat\boldsymbol{\delta}_{{\sf roc},\gamma,t(c)}$ for any $c\in\mathbb{R}$ where $t(c)=\argmin{c'\in\mathcal{C}_{\sf cal}}|c'-c|$. Here $\lambda_{{\sf roc},\alpha,\widehat c_j}$ and $\lambda_{{\sf roc},\gamma,\widehat c_j}$ are two series of $\ell_1$-penalty parameters to be tuned using the procedure described in Section \ref{sec:method:tune}. We then construct the doubly robust estimators for $\TPR(\cdot)$ and $\mathbf{FPR}(\cdot)$ based on the calibrated nuisance estimators and plug the estimators $\widehat\boldsymbol{\alpha}_{{\sf roc},t(c)}$ and $\widehat\boldsymbol{\gamma}_{{\sf roc},t(c)}$
into \eqref{equ:roc:dr} to obtain
\begin{equation}
\begin{aligned}
& \widehat \TPR(c) = \widehat \TPR(c; \widehat \boldsymbol{\beta}, \widehat\boldsymbol{\alpha}_{{\sf roc},t(c)}, \widehat\boldsymbol{\gamma}_{{\sf roc},t(c)},\widehat\boldsymbol{\alpha}_{{\sf roc},\widehat c_1},\widehat\boldsymbol{\gamma}_{{\sf roc},\widehat c_1}),\\
& \widehat \mathbf{FPR}(c) = \widehat \mathbf{FPR}(c; \widehat \boldsymbol{\beta}, \widehat\boldsymbol{\alpha}_{{\sf roc},t(c)}, \widehat\boldsymbol{\gamma}_{{\sf roc},t(c)},\widehat\boldsymbol{\alpha}_{{\sf roc},\widehat c_1},\widehat\boldsymbol{\gamma}_{{\sf roc},\widehat c_1}) \,.
\end{aligned}
\label{equ:roc:dr:cal}
\end{equation}
\begin{remark}
The choice of $n_{\min}$ is critical. Large $n_{\min}$ can lead to non-negligible approximation error from $t(c)-c$, and hence imprecise calibration. Small $n_{\min}$ can result in overly small effective sample sizes for some calibrated estimating equations, says $\widehat c_j=\widehat c_m$ in equation \eqref{equ:cal:roc}, leading to high error rate in $\widehat\boldsymbol{\alpha}_{{\sf roc},\widehat c_j}$ and $\widehat\boldsymbol{\gamma}_{{\sf roc},\widehat c_j}$. We study in Section \ref{sec:thm} this trade-off and derive an optimal $n_{\min}$ of order $n^{\frac{1}{2}}(\log~np)^{1/3}$, which can effectively remove the excessive first order bias in the nuisance estimators simultaneously for all $c$.
\end{remark}
Finally, one can evaluate $\widehat\ROC(u)=\widehat\TPR\{{\widehat\mathbf{FPR}}^{-1}(u)\}$, and $\widehat\AUC=\int_0^1\widehat\ROC(u) {\rm d}u$, to estimate the ROC function and AUC. An illustration of this ``nearest quantile" strategy is presented in Figure \ref{fig:roc_ill}.
\begin{figure}[htbp]
\centering
\caption{\label{fig:roc_ill} The nearest quantile strategy. The quantiles $\widehat c_j$'s are set to divide the source sample equally such that $n_{j} - n_{j+1} = n_{\min}$ and $n_m = n_{\min}$. The nuisance models are calibrated at these quantiles for the doubly robust estimation of $\TPR$ and $\mathbf{FPR}$. For each $\widehat c_j$, the effective sample size of its calibrating equations is labeled as $n_j$. For any point $c\in\mathbb{R}$, the nuisance models calibrated in the nearest quantile $t(c)$ will be used to construct the doubly robust estimators of $\TPR(c)$ and $\mathbf{FPR}(c)$.}
\includegraphics[width=0.5\linewidth]{ROC_plot2.pdf}
\end{figure}
\subsection{Tuning parameters}\label{sec:method:tune}
In this section, we describe our empirical procedures to choose the tuning parameters introduced in previous sections. To select the $\ell_1$-penalty parameters $\lambda_{\alpha}$, $\lambda_{\gamma}$ for the preliminary estimators, we use cross-validation with the range stick to the rates given in Section \ref{sec:thm}.
For $\lambda_{\beta_j,\alpha}$ (or $\lambda_{\beta_j,\alpha}^{+}$ and $\lambda_{\beta_j,\alpha}^-$), $\lambda_{\beta_j,\gamma}$ (or $\lambda_{\beta_j,\gamma}^+$ and $\lambda_{\beta_j,\gamma}^-$), $\lambda_{{\sf roc},\alpha,\widehat{c}_j}$ and $\lambda_{{\sf roc},\gamma,\widehat{c}_j}$, we adopt a Gaussian multiplier bootstrap procedure \citep{chernozhukov2013gaussian} to obtain their data-driven estimations. In specific, we generate independent standard normal variables $\mathbf{\epsilon}=(\epsilon_{1},\dots,\epsilon_{n},\ldots,\epsilon_{N+n})^{\scriptscriptstyle \sf T}$, and take
\begin{equation*}
\begin{aligned}
\mathcal{M}_{\beta_j,\alpha}^{\mathbf{\epsilon}}&=\left\|\widehat\Psc_{\scriptscriptstyle \Ssc}\epsilon_i\widehat w_{ji}\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\alpha})\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} )\mathbf{X}_i-\widehat\Psc_{\scriptscriptstyle \Tsc} \epsilon_i\widehat w_i\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} )\mathbf{X}_i\right\|_{\infty};\\
\mathcal{M}_{\beta_j,\gamma}^{\mathbf{\epsilon}}&=\left\|\widehat\Psc_{\scriptscriptstyle \Ssc}\epsilon_i\widehat w_{ji}\exp(\mathbf{X}_i ^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\alpha})\mathbf{X}_i\{ Y_i - g(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma})\}\right\|_{\infty};\\
\mathcal{M}_{{\sf roc},\alpha}^{\mathbf{\epsilon}}&=\max_{\widehat c_j \in\mathcal{C}_{\sf cal}}(n/ n_j)^{\frac{1}{2}}\left\|\widehat\Psc_{\scriptscriptstyle \Ssc}\epsilon_i\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq \widehat c_j )\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\alpha})\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} )\mathbf{X}_i-\widehat\Psc_{\scriptscriptstyle \Tsc} \epsilon_i\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq \widehat c_j)\dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma} )\mathbf{X}_i\right\|_{\infty};\\
\mathcal{M}_{{\sf roc},\gamma}^{\mathbf{\epsilon}}&=\max_{ \widehat c_j \in\mathcal{C}_{\sf cal}}(n/n_j)^{\frac{1}{2}}\left\|\widehat\Psc_{\scriptscriptstyle \Ssc}\epsilon_i\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta} \geq \widehat c_j)\exp(\mathbf{X}_i ^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\alpha})\mathbf{X}_i\{ Y_i - g(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widetilde\boldsymbol{\gamma})\}\right\|_{\infty}\,.
\end{aligned}
\end{equation*}
Since $\mathbf{\epsilon}$ has mean $\mathbf{0}$ and concentrates the samples, $\mathcal{M}_{\beta_j,\alpha}^{\mathbf{\epsilon}}$, $\mathcal{M}_{\beta_j,\gamma}^{\mathbf{\epsilon}}$, $\mathcal{M}_{{\sf roc},\alpha}^{\mathbf{\epsilon}}$ and $\mathcal{M}_{{\sf roc},\gamma}^{\mathbf{\epsilon}}$ achieve the desirable rate (see Section \ref{sec:thm}) for the control of the moment terms in (\ref{equ:mom:cons}) and (\ref{equ:mom:cons:roc}) with high probability. We sample $\mathbf{\epsilon}$ to estimate the $q\in(0,1)$ quantiles of $\mathcal{M}_{\beta_j,\alpha}^{\mathbf{\epsilon}}$, $\mathcal{M}_{\beta_j,\gamma}^{\mathbf{\epsilon}}$, $\mathcal{M}_{{\sf roc},\alpha}^{\mathbf{\epsilon}}$ and $\mathcal{M}_{{\sf roc},\gamma}^{\mathbf{\epsilon}}
$, denoted as $Q_{q}(\mathcal{M}_{\beta_j,\alpha}^{\mathbf{\epsilon}})$, $Q_{q}(\mathcal{M}_{\beta_j,\gamma}^{\mathbf{\epsilon}})$, $Q_{q}(\mathcal{M}_{{\sf roc},\alpha}^{\mathbf{\epsilon}})$ and $Q_{q}(\mathcal{M}_{{\sf roc},\gamma}^{\mathbf{\epsilon}})$, respectively. We then set the tuning parameters as $\lambda_{\beta_j,\alpha}=Q_q(\mathcal{M}_{\beta_j,\alpha}^{\mathbf{\epsilon}})$, $\lambda_{\beta_j,\gamma}=Q_q(\mathcal{M}_{\beta_j,\gamma}^{\mathbf{\epsilon}})$, $\lambda_{{\sf roc},\alpha,\widehat{c}_j}=(n/ n_j)^{-1/2}Q_q(\mathcal{M}_{{\sf roc},\alpha}^{\mathbf{\epsilon}})$ and $\lambda_{{\sf roc},\gamma,\widehat{c}_j}=(n/ n_j)^{-1/2}Q_q(\mathcal{M}_{{\sf roc},\gamma}^{\mathbf{\epsilon}})$. Similar tuning strategies has been commonly used in recent high-dimensional statistics literature like \cite{battey2018distributed} who set $q=0.95$ since a relatively large $q$ can help with controlling the maximum moment terms properly. Inspired by this, we take $q=0.8$ in our numerical implementation while we find that settings like $q=0.9$ and $q=0.95$ warrant similar performances.
\subsection{Multiplier bootstrap for inference}\label{sec:method:boot}
As justified by our theoretical results in Section \ref{sec:thm}, the proposed estimators are root-$n$ consistent and asymptotically normal. Since the asymptotic variance of the accuracy parameters involves unknown derivatives of non-parametric functionals, we propose using multiplier bootstrap instead of explicit plug-in estimation to construct standard error (SE) estimators and confidence intervals (CIs).
To be specific, we generate independent Gaussian
random variables $\mathbf{b}=(b_1,\ldots,b_n,\ldots,b_{N+n})^{\scriptscriptstyle \sf T}$ with mean one and unit variance. Then we obtain the bootstrap estimator of $\boldsymbol{\beta}$ as
\[
\widehat \boldsymbol{\beta}^{\mathbf{b}} = \widehat \boldsymbol{\beta} + \widehat{\boldsymbol{\Sigma}}_{\widehat \boldsymbol{\beta}}^{-1} \big[\widehat\Psc_{\scriptscriptstyle \Ssc} (b_i-1)\exp(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\alpha}_{\beta})\mathbf{A}_i \{ Y_i - g(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\gamma}_{\beta})\} + \widehat\Psc_{\scriptscriptstyle \Tsc} (b_i-1) \mathbf{A}_i \{ g(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\gamma}_{\beta})- g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}) \} \big]
\]
and the bootstrap samples of $\TPR(c)$ and $\mathbf{FPR}(c)$ as
\begin{align*}
&\widehat \TPR^{\mathbf{b}}(c) = \frac{\widehat\Psc_{\scriptscriptstyle \Ssc} b_i\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}^{\mathbf{b}}\geq c) \widehat h_{{\sf roc}}(\mathbf{X}_i;c) \{Y_i-\widehat r_{{\sf roc}}(\mathbf{X}_i;c)\} + \widehat\Psc_{\mathcal{T}}b_i \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}^{\mathbf{b}}\geq c) \widehat r_{{\sf roc}}(\mathbf{X}_i;c)}{\widehat\Psc_{\scriptscriptstyle \Ssc} b_i\widehat h_{{\sf roc}}(\mathbf{X}_i;\widehat c_1) \{Y_i-\widehat r_{{\sf roc}}(\mathbf{X}_i; \widehat c_1) \} + \widehat\Psc_{\scriptscriptstyle \Tsc} b_i\widehat h_{{\sf roc}}(\mathbf{X}_i;\widehat c)};\\
&\widehat \mathbf{FPR}^{\mathbf{b}}(c) = \frac{\widehat\Psc_{\scriptscriptstyle \Ssc} b_i\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}^{\mathbf{b}}\geq c) \widehat h_{{\sf roc}}(\mathbf{X}_i;c) \{\widehat r_{{\sf roc}}(\mathbf{X}_i;c) -Y_i\} + \widehat\Psc_{\scriptscriptstyle \Tsc} b_i\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}^{\mathbf{b}}\geq c) \{1-\widehat r_{{\sf roc}}(\mathbf{X}_i;c) \} }{\widehat\Psc_{\scriptscriptstyle \Ssc} b_i\widehat h_{{\sf roc}}(\mathbf{X}_i;\widehat c_1) \{\widehat r_{{\sf roc}}(\mathbf{X}_i;\widehat c_1) -Y_i\} + \widehat\Psc_{\scriptscriptstyle \Tsc} b_i\{1-\widehat r_{{\sf roc}}(\mathbf{X}_i;\widehat c_1) \} },
\end{align*}
where $\widehat h_{{\sf roc}}(\mathbf{X}_i;c) = \exp( \mathbf{X}_i^{\scriptscriptstyle \sf T} \widehat\boldsymbol{\alpha}_{{\sf roc},t(c)})$ and $\widehat r_{{\sf roc}}(\mathbf{X}_i;c) = g(\mathbf{X}_i^{\scriptscriptstyle \sf T}\widehat\boldsymbol{\gamma}_{{\sf roc},t(c)})$,
and the bootstrap sample for ROC and AUC as $\widehat\ROC^{\mathbf{b}}(u)=\widehat\TPR^{\mathbf{b}}\{[{\widehat\mathbf{FPR}}^{\mathbf{b}}]^{-1}(u)\}$, and $\widehat\AUC^{\mathbf{b}}=\int_0^1\widehat\ROC^{\mathbf{b}}(u)du$. Finally, SEs of the estimators are estimated using the standard deviation of their bootstrap samples. One may subsequently construct CIs using normal approximation centering at the point estimates or taking quantiles of the bootstrap samples.
\begin{remark}
Typically we use non-negative random variables such as exponential random variables for multiplier bootstrap to maintain the range of TPR and FPR, while theoretically speaking, it does not matter to use Gaussian random variables, which are used in our subsequent numeric analysis for convenience.
\end{remark}
\def\mathcal{Q}{\mathcal{Q}}
\def\mathcal{H}{\mathcal{H}}
\def\boldsymbol{u}{\boldsymbol{u}}
\section{Theoretical justification}\label{sec:thm}
Let $\|\boldsymbol{u}\|_0$ represent the number of non-zero entries of the input vector $\boldsymbol{u}$ and $\|\boldsymbol{u}\|_q$ be the $l_q$ norm of $\boldsymbol{u}$ for any $q>0$. We use $a_n = o(b_n)$ if $\lim_{n \rightarrow \infty} a_n/b_n = 0$, $a_n = O(b_n)$ if $\lim_{n \rightarrow \infty} a_n/b_n \leq C$ for some constant $C$, and $a_n \asymp b_n$ if $a_n = O(b_n)$ and $b_n = O(a_n)$. Throughout, we denote by $\rho= \lim_{N,n \rightarrow \infty}(n/N)^{\frac{1}{2}}$ and assume $\rho=O(1)$. We denote the population parameters by
\begin{equation*}
\begin{aligned}
\bar\boldsymbol{\alpha} & = \argmin{\boldsymbol{\alpha}\in\mathbb{R}^{q+p}} \big \{ \rho_{\scriptscriptstyle \Ssc} \mathbb{E}_{\scriptscriptstyle \Ssc} [ \exp (\mathbf{X}_i ^{\scriptscriptstyle \sf T} \boldsymbol{\alpha})] - \rho_{\scriptscriptstyle \Tsc} \mathbb{E}_{\scriptscriptstyle \Tsc}[ \mathbf{X}_i ^{\scriptscriptstyle \sf T} \boldsymbol{\alpha} ] \big \} , \quad \bar\boldsymbol{\gamma} = \argmin{\boldsymbol{\gamma}\in\mathbb{R}^{q+p}}
\mathbb{E}_{\scriptscriptstyle \Ssc} [ -Y_i \mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma} + G(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma})],
\end{aligned}
\end{equation*}
where $\rho_{\scriptscriptstyle \Ssc} = 1+\rho^{-2}$ and $\rho_{\scriptscriptstyle \Tsc} = 1+\rho^{2}$, and for $j = 1,\ldots,q$,
\begin{equation*}
\begin{aligned}
\bar\boldsymbol{\alpha}_{\beta_j} & = \bar\boldsymbol{\alpha} + \argmin{\boldsymbol{\delta}\in\mathbb{R}^{q+p}} \mathbf{e}_j ^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \big\{ \rho_{\scriptscriptstyle \Ssc} \mathbb{E}_{\scriptscriptstyle \Ssc} [\mathbf{A}_i \mathcal{F}_{{\scriptscriptstyle \Ssc} i}(\boldsymbol{\delta}; \bar \boldsymbol{\alpha}, \bar \boldsymbol{\gamma})
] - \rho_{\scriptscriptstyle \Tsc} \mathbb{E}_{\scriptscriptstyle \Tsc} [ \mathbf{A}_i
\mathcal{F}_{{\scriptscriptstyle \Tsc} i}(\boldsymbol{\delta};\bar \boldsymbol{\alpha},\bar \boldsymbol{\gamma}) ] \big \} \,,\\
\bar\boldsymbol{\gamma}_{\beta_j} & = \bar\boldsymbol{\gamma}+ \argmin{\boldsymbol{\delta}\in\mathbb{R}^{q+p}} \mathbf{e}_j ^{\scriptscriptstyle \sf T} \boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbb{E}_{\scriptscriptstyle \Ssc} [ \mathbf{A}_i \exp(\mathbf{X}_i ^{\scriptscriptstyle \sf T} \bar \boldsymbol{\alpha}) \mathcal{G}_i(\boldsymbol{\delta};\bar \boldsymbol{\gamma})] \,,
\end{aligned}
\end{equation*}
where $\mathcal{F}_{{\scriptscriptstyle \Ssc} i}(\boldsymbol{\delta};\boldsymbol{\alpha},\boldsymbol{\gamma}) = \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma} ) \exp \{\mathbf{X}_i ^{\scriptscriptstyle \sf T} ( \boldsymbol{\alpha}+\boldsymbol{\delta})\}$ and $\mathcal{F}_{{\scriptscriptstyle \Tsc} i}(\boldsymbol{\delta};\boldsymbol{\alpha},\boldsymbol{\gamma}) = \dot{g}( \mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma} ) \mathbf{X}_i ^{\scriptscriptstyle \sf T} (\boldsymbol{\alpha}+\boldsymbol{\delta})$.
We define $c_j$ as $1 - (m-j+1)n_{\min}/n$ population quantile of $\mathbf{A}_i^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0$, for $j \in [m]$.
For $c \in \mathbb{R}$, we define
\begin{equation*}
\begin{split}
\bar\boldsymbol{\alpha}_{{\sf roc},c} & = \bar \boldsymbol{\alpha} + \argmin{\boldsymbol{\delta}\in\mathbb{R}^{q+p}} \big\{ \rho_{\scriptscriptstyle \Ssc} \mathbb{E}_{\scriptscriptstyle \Ssc} [\mathbb{I}(\mathbf{A}_i ^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c)
\mathcal{F}_{{\scriptscriptstyle \Ssc} i}(\boldsymbol{\delta}; \bar \boldsymbol{\alpha}, \bar \boldsymbol{\gamma}) ] - \rho_{\scriptscriptstyle \Tsc} \mathbb{E}_{\scriptscriptstyle \Tsc} [ \mathbb{I}(\mathbf{A}_i ^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c)
\mathcal{F}_{{\scriptscriptstyle \Tsc} i}(\boldsymbol{\delta}; \bar \boldsymbol{\alpha}, \bar \boldsymbol{\gamma}) ] \big \} \,, \\
\bar\boldsymbol{\gamma}_{{\sf roc},c} &= \bar \boldsymbol{\gamma} + \argmin{\boldsymbol{\delta}\in\mathbb{R}^{q+p}}\mathbb{E}_{\scriptscriptstyle \Ssc} [ \mathbb{I}(\mathbf{A}_i ^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0 \geq c ) \exp(\mathbf{X}_i ^{\scriptscriptstyle \sf T}\bar \boldsymbol{\alpha}) \mathcal{G}_i(\boldsymbol{\delta};\bar \boldsymbol{\gamma}) ]\,.
\end{split}
\end{equation*}
For simplicity, we focus on the case without the need to split the samples through (\ref{equ:cal:sign}). Our analysis can be naturally generalized to this setting. We first introduce assumptions for our theoretical analysis and remark on their appropriateness.
\begin{assume}
There exists a constant $\kappa>0$ such that $\max_{i\in\{1,\ldots,n+N\}}\|\mathbf{X}_i\|_{\infty}<\kappa$.
$\mathbf{A}_i$ belongs to a compact set, and its probability density is continuously differentiable.
The eigenvalues of $\mathbb{E}_{\scriptscriptstyle \Tsc}[\mathbf{X}_i\mathbf{X}_i^{\scriptscriptstyle \sf T}]$ and $\mathbb{E}_{\scriptscriptstyle \Ssc}[\mathbf{X}_i\mathbf{X}_i^{\scriptscriptstyle \sf T}]$ stay away from $0$ and $\infty$.
$\boldsymbol{\beta}$ belongs to a compact parameter space $\mathcal{B}$.
\label{asu:1}
\end{assume}
\begin{assume}
At least one of the following two conditions hold: (a) there exists $\boldsymbol{\alpha}_0$ such that $h_0(\mathbf{x}) = p_{\mathbf{X}|S=0}(\mathbf{x})/p_{\mathbf{X}|S=1}(\mathbf{x})=\exp(\mathbf{x}^{\scriptscriptstyle \sf T}\boldsymbol{\alpha}_0)$; or (b) there exists $\boldsymbol{\gamma}_0$ such that $r_0(\mathbf{x})=\mathbb{E}_{\scriptscriptstyle \Ssc}[Y|\mathbf{X}=\mathbf{x}]=\mathbb{E}_{\scriptscriptstyle \Tsc}[Y|\mathbf{X}=\mathbf{x}]=g(\mathbf{x}^{\scriptscriptstyle \sf T}\boldsymbol{\gamma}_0)$.
\label{asu:model}
\end{assume}
\begin{assume}
There exists a positive constant $c_g$ such that
$\min_{i\in\{1,\dots, n+N\}} \dot{g}(\mathbf{X}_i^{\scriptscriptstyle \sf T}\bar \boldsymbol{\gamma}) \geq c_g$.
{There exists a positive constant $C_{\delta}$ such that
$\|\bar\boldsymbol{\delta}_{\beta_j, \alpha}\|_1$, $\|\bar\boldsymbol{\delta}_{\beta_j, \gamma}\|_1$,
$\|\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c}\|_1$, $\|\bar\boldsymbol{\delta}_{{\sf roc}, \gamma, c}\|_1$, and $\|\bar\boldsymbol{\alpha}\|_1$ are bounded by $C_{\delta}/\kappa$ for $j\in\{1, \dots, q\}$ and any $c$.}
\label{asu:3}
\end{assume}
\begin{assume}
Assume that $\|\bar\boldsymbol{\alpha}\|_0,\|\bar\boldsymbol{\gamma}\|_0\leq s$ for some $s=o(n^{\frac{1}{2}}/\log p)$, and $\bar \boldsymbol{\delta}_{\beta_j, \alpha}$ and $\bar \boldsymbol{\delta}_{\beta_j, \gamma}$ are $l_r$ sparse, i.e., $$\|\bar\boldsymbol{\delta}_{\beta_j, \alpha}\|_r\leq R_r \quad \mbox{and}\quad \|\bar\boldsymbol{\delta}_{\beta_j, \gamma}\|_r\leq R_r$$
for some $r\in [0,1]$ and $R_r>0$ satisfying $R_r( \log p/n)^{\frac{1-r}{2}} = o\{(\log p)^{-1/2}\}$.
\label{asu:4}
\end{assume}
\begin{assume}
Assume that $s = o(n^{1/4}/\log p)$, $n_{\min} \asymp n^{\frac{1}{2}}(\log~np)^{1/3}$,
and $\bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c}, \bar \boldsymbol{\delta}_{{\sf roc}, \gamma, c}$ are $l_r$ sparse for all $c$, i.e.,
$\|\bar \boldsymbol{\delta}_{{\sf roc}, \alpha, c}\|_r \leq R_{{\sf roc}, r}$ and $\|\bar \boldsymbol{\delta}_{{\sf roc}, \gamma, c}\|_r \leq R_{{\sf roc}, r}$,
where $r\in [0,1]$ and
$$R_{{\sf roc}, r} = o \left\{ \frac{n^{1/4 - r/4} }{(\log ~np)^{2/3- r/3}}\right\}.$$
\label{asu:5}
\end{assume}
\begin{remark}
Assumption \ref{asu:1} could be relaxed to sub-gaussian $\mathbf{X}_i$ by imposing on some moment regularity conditions.
The assumption for $\mathbf{A}_i$ and $\mathbf{A}_i^{\scriptscriptstyle \sf T}\boldsymbol{\beta}_0$ are commonly founded in handling the empirical process of the $M$-estimators and their model accuracy parameters \citep{gronsbell2018semi,gronsbell2020efficient}.
The assumption for eigenvalues of the covariance matrix is standard in high-dimensional settings \citep{buhlmann2011statistics,negahban2012unified}.
Assumption \ref{asu:model} is standard in the literature of doubly robust inference \citep[e.g.]{bang2005doubly}. Technical conditions in Assumption \ref{asu:3} have been commonly used for analyzing $l_1$-regularized GLM \citep[e.g.]{tan2020model}.
It is reasonable in that when $r=1$ in Assumptions \ref{asu:4} and \ref{asu:5}, terms like $|\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\delta}_{\beta_j, \alpha}|$ will be smaller than $\|\mathbf{X}_i\|_{\infty}\|\bar\boldsymbol{\delta}_{\beta_j, \alpha}\|_1$, which is bounded under Assumptions \ref{asu:1}, \ref{asu:4}, and \ref{asu:5}.
\end{remark}
\begin{remark}
Assumptions \ref{asu:4} and \ref{asu:5} is on the approximate sparsity of
$\bar\boldsymbol{\delta}_{\beta_j, \alpha}$, $\bar\boldsymbol{\delta}_{\beta_j, \gamma, c}$, $\bar\boldsymbol{\delta}_{{\sf roc}, \alpha,c}$ and $\bar\boldsymbol{\delta}_{{\sf roc}, \gamma,c}$. The minimum effective sample size of the calibrating equations, $n_{\min}$ is picked to minimize the excessive bias rate through trading-off the estimation errors of the nuisance models and the approximation bias introduced by the samples in $\{i:\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq \widehat c_m)\neq\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}\geq c)\}$ for any $c>\widehat c_m$; see Section \ref{sec:cal:roc} and Figure \ref{fig:roc_ill}. Such optimal choice ensures the sharpness of our sparsity assumptions.
\end{remark}
\begin{remark}
Different from \cite{tan2020model} and \cite{dukes2020inference}, we only impose $l_0$ (exact) sparsity assumptions on the preliminary estimators $\bar\boldsymbol{\alpha}$ and $\bar\boldsymbol{\gamma}$ while imposing approximate sparsity assumptions on the increments of the calibrated nuisance parameters $\bar \boldsymbol{\delta}_{\beta_j, \alpha}$, $\bar \boldsymbol{\delta}_{\beta_j, \gamma}$, $\bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c}$ and $\bar\boldsymbol{\delta}_{{\sf roc}, \gamma, c}$. These assumptions could be more reasonable than directly assuming the exact sparsity of the calibrated nuisance parameters since the calibrated estimating equations involve some additional weights on the GLM. Note that when the density ratio model is correctly specified, we have $\bar\boldsymbol{\delta}_{\beta_j, \alpha} = \bar\boldsymbol{\delta}_{{\sf roc}, \alpha, c} = \mathbf{0}$, and when the imputation model is correctly specified, $\bar\boldsymbol{\delta}_{\beta_j, \gamma} = \bar\boldsymbol{\delta}_{{\sf roc}, \gamma, c} = \mathbf{0}$.
\end{remark}
Based on the above introduced assumptions, we present the main conclusions of our asymptotic analysis below. Lemmas \ref{lemma:1} and \ref{lemma:2} are two important mediate results about the convergence rates of the calibrated nuisance estimators. Theorems \ref{thm:1} and \ref{thm:2} give the $n^{\frac{1}{2}}$-convergence and asymptotic expansion of our model coefficient estimator $\widehat\boldsymbol{\beta}$ and model accuracy estimator $\widehat\ROC(u)$.
\begin{lemma}
Under Assumptions \ref{asu:1}, \ref{asu:3} and \ref{asu:4},
with the tuning parameters $\lambda_{\alpha}$, $\lambda_{\gamma}$, $\lambda_{\beta_j,\alpha}$, and $\lambda_{\beta_j,\gamma}$ of order $(\log p/n)^{\frac{1}{2}}$, for $j = 1, \ldots,q$, we have
\small{
\begin{align*}
&\small \|\widehat \boldsymbol{\alpha}_{\beta_j} - \bar\boldsymbol{\alpha}_{\beta_j}\|_2^2+\|\widehat \boldsymbol{\gamma}_{\beta_j} - \bar\boldsymbol{\gamma}_{\beta_j}\|_2^2 = O_p\left\{ \frac{s\log ~p}{n}+ R_r\left(\frac{\log ~p}{n} \right)^{1- \frac{r}{2}}\right\},\\
& \|\widehat \boldsymbol{\alpha}_{\beta_j} - \bar\boldsymbol{\alpha}_{\beta_j}\|_1 +\|\widehat \boldsymbol{\gamma}_{\beta_j} - \bar\boldsymbol{\gamma}_{\beta_j}\|_1 =O_p\left\{ s\left(\frac{ \log~p }{n}\right)^{\frac{1}{2}} + R_r\left(\frac{\log ~p}{n} \right)^{\frac{1-r}{2}}\right\}.
\end{align*}
}
\label{lemma:1}
\end{lemma}
\begin{theorem}
Under Assumptions \ref{asu:1}--\ref{asu:4}, with the choices on tuning parameters given in Lemma \ref{lemma:1}, $\widehat\boldsymbol{\beta}$ converges to $\boldsymbol{\beta}_0$ in probability, and
\[
\sqrt{n}\mathbf{e}_j^{\scriptscriptstyle \sf T}(\widehat\boldsymbol{\beta}-\boldsymbol{\beta}_0) = \frac{1}{\sqrt{n}}\sum_{i=1}^n \mathbf{e}_j ^{\scriptscriptstyle \sf T} \mathcal{W}_{{\scriptscriptstyle \Ssc},\beta_j}(\mathbf{D}_i)+\frac{\rho}{\sqrt{N}}\sum_{i=n+1}^{n+N} \mathbf{e}_j ^{\scriptscriptstyle \sf T} \mathcal{W}_{{\scriptscriptstyle \Tsc},\beta_j}(\mathbf{D}_i)+o_p(1),
\]
converges weakly to $N(0,\sigma_j^2)$ for $j = 1, \ldots, q$, where
\begin{align*}
\mathcal{W}_{{\scriptscriptstyle \Ssc}, \beta_j}(\mathbf{D}_i)&= \bar\boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbf{A}_i \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\alpha}_{\beta_j}) \{ Y_i - g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{\beta_j})\}, \
\mathcal{W}_{{\scriptscriptstyle \Tsc}, \beta_j}(\mathbf{D}_i) = \bar\boldsymbol{\Sigma}_{\boldsymbol{\beta}_0}^{-1} \mathbf{A}_i \{g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \bar\boldsymbol{\gamma}_{\beta_j}) - g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0) \} \,,
\end{align*}
and $\sigma_j^2 = {\rm Var}\big(\mathbf{e}_j ^{\scriptscriptstyle \sf T} \mathcal{W}_{{\scriptscriptstyle \Ssc},\beta_j}(\mathbf{D}_i)\big) + \rho^2 {\rm Var}\big(\mathbf{e}_j ^{\scriptscriptstyle \sf T} \mathcal{W}_{{\scriptscriptstyle \Tsc},\beta_j}(\mathbf{D}_i)\big)$.
\label{thm:1}
\end{theorem}
\begin{lemma}
Under Assumptions \ref{asu:1}--\ref{asu:5},
with the tuning parameters $\lambda_{\alpha}$, $\lambda_{\gamma}$, $\lambda_{\beta,\alpha}$, and $\lambda_{\beta,\gamma}$ of order $(\log p/n)^{\frac{1}{2}}$ as well as $\lambda_{{\sf roc},\alpha,\widehat c_j}$ and $\lambda_{{\sf roc},\gamma,\widehat c_j}$ of order $n_jn^{-1}(\log~np/n_j)^{\frac{1}{2}}$ for $j=1,\ldots,m$,
\small{
\begin{align*}
&\|\widehat \boldsymbol{\alpha}_{{\sf roc}, \widehat c_j} - \bar\boldsymbol{\alpha}_{{\sf roc}, c_j}\|_2^2+\|\widehat \boldsymbol{\gamma}_{{\sf roc}, \widehat c_j} - \bar\boldsymbol{\gamma}_{{\sf roc}, c_j}\|_2^2 = O_p\left\{ \frac{s\log ~p}{n}+ R_{{\sf roc}, r}\left(\frac{\log ~np}{n} \right)^{1-\frac{r}{2}}\right\},\\
&\|\widehat \boldsymbol{\alpha}_{{\sf roc}, \widehat c_j} - \bar\boldsymbol{\alpha}_{{\sf roc}, c_j}\|_1+\|\widehat \boldsymbol{\gamma}_{{\sf roc}, \widehat c_j} - \bar\boldsymbol{\gamma}_{{\sf roc}, c_j}\|_1 =O_p\left\{ s\left(\frac{\log ~p}{n}\right)^{\frac{1}{2}} + R_{{\sf roc}, r}\left(\frac{\log ~np}{n} \right)^{\frac{1-r}{2}}\right\}.
\end{align*}
}
\label{lemma:2}
\end{lemma}
\begin{theorem}
Under Assumptions \ref{asu:1}--\ref{asu:5}, and with the tuning parameters given in Lemma \ref{lemma:2}, { we have
$$
\sup_{u\in (0,1)}|\widehat\ROC(u) - \ROC(u)| = O_p\left\{\left(\frac{\log~n}{n}\right)^{\frac{1}{2}}\right\},
$$
} and that
$$
\sqrt{n}\{\widehat\ROC(u)-\ROC(u)\}=\frac{1}{\sqrt{n}}\sum_{i=1}^n\mathcal{W}_{{\scriptscriptstyle \Ssc},{\sf roc}}(u,\mathbf{D}_i)+\frac{\rho}{\sqrt{N}}\sum_{i=n+1}^{n+N}\mathcal{W}_{{\scriptscriptstyle \Tsc},{\sf roc}}(u,\mathbf{D}_i)+o_p(1)
$$
uniformly for all $u\in(0,1)$, where $c_{u} = {\mathbf{FPR}}^{-1}(u)$, $\mu = \mathbb{P}_{\scriptscriptstyle \Tsc}(Y=1)$, and for $\mathcal{K} = \mathcal{S}, \mathcal{T}$,
\begin{equation}
\begin{split}
\mathcal{W}_{{\scriptscriptstyle \Ksc},{\sf roc}}(u,\mathbf{D}_i) = & \mu^{-1} \mathcal{W}_{{\scriptscriptstyle \Ksc},{\rm TPR}}(c_{u},\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c_{u}},\mathbf{D}_i) - (1-\mu)^{-1} \ROC(u) \mathcal{W}_{{\scriptscriptstyle \Ksc},{\rm FPR}}(c_{u},\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c_{u}},\mathbf{D}_i) \\
& + \{ \mu^{-1} \dot{\TP}(c_u, \boldsymbol{\beta}_0) - (1-\mu)^{-1} \ROC(u) \dot{\mathbf{FP}}(c_u, \boldsymbol{\beta}_0) \} ^{\scriptscriptstyle \sf T}\mathcal{W}_{{\scriptscriptstyle \Ksc}, \beta}(\mathbf{D}_i) \,,
\end{split}
\label{equ:term1:roc}
\end{equation}
where $\bar\boldsymbol{\theta}_{{\sf roc},c_{u}}= (\bar\boldsymbol{\alpha}_{{\sf roc},c_{u}}^{\scriptscriptstyle \sf T} ,\bar\boldsymbol{\gamma}_{{\sf roc},c_{u}}^{\scriptscriptstyle \sf T})^{\scriptscriptstyle \sf T}$,
\begin{equation*}
\begin{aligned}
& \mathcal{W}_{{\scriptscriptstyle \Ksc},{\rm TPR}}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) = \mathcal{W}_{{\scriptscriptstyle \Ksc} 1}(c,\boldsymbol{\beta}, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) - \TPR(c,\boldsymbol{\beta}_0) \mathcal{W}_{{\scriptscriptstyle \Ksc} 1}(-\infty,\boldsymbol{\beta}, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) \,, \\
& \mathcal{W}_{{\scriptscriptstyle \Ksc}, {\rm FPR }}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) = \mathcal{W}_{{\scriptscriptstyle \Ksc} 2}(c,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) - \mathbf{FPR}(c,\boldsymbol{\beta}_0) \mathcal{W}_{{\scriptscriptstyle \Ksc} 2}(-\infty,\boldsymbol{\beta}_0, \bar \boldsymbol{\theta}_{{\sf roc},c},\mathbf{D}_i) \,, \\
& \mathcal{W}_{{\scriptscriptstyle \Ssc} 1}(c,\boldsymbol{\beta}, \boldsymbol{\theta},D_i) = \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) \exp(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\alpha}) \{Y_i-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma})\} - \mathbb{E}_{\scriptscriptstyle \Ssc}\mathbb{I}(\mathbf{A}^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) \exp(\mathbf{X}^{\scriptscriptstyle \sf T} \boldsymbol{\alpha}) \{Y-g(\mathbf{X}^{\scriptscriptstyle \sf T} \boldsymbol{\gamma})\} \,, \\
& \mathcal{W}_{{\scriptscriptstyle \Tsc} 1}(c,\boldsymbol{\beta}, \boldsymbol{\theta},D_i) = \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma}) - \mathbb{E}_{\scriptscriptstyle \Tsc}\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma}) \,, \\
& \mathcal{W}_{{\scriptscriptstyle \Ssc} 2}(c,\boldsymbol{\beta}, \boldsymbol{\theta},\mathbf{D}_i) = - \mathcal{W}_{{\scriptscriptstyle \Ssc} 1}(c,\boldsymbol{\beta}, \boldsymbol{\theta},\mathbf{D}_i)
\,, \\
& \mathcal{W}_{{\scriptscriptstyle \Tsc} 2}(c,\boldsymbol{\beta}, \boldsymbol{\theta}, \mathbf{D}_i) = \mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) \{1-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma})\} - \mathbb{E}_{\scriptscriptstyle \Tsc}\mathbb{I}(\mathbf{A}_i^{\scriptscriptstyle \sf T} \boldsymbol{\beta} \geq c) \{1-g(\mathbf{X}_i^{\scriptscriptstyle \sf T} \boldsymbol{\gamma})\} \,, \\
& \dot{\TP}(c, \boldsymbol{\beta}) = \partial \TP (c, \boldsymbol{\beta}) / \partial \boldsymbol{\beta} = \partial \mathbb{P}_{\scriptscriptstyle \Tsc}(Y=1,\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta} \geq c) / \partial \boldsymbol{\beta}\,, \\
& \dot{\mathbf{FP}}(c, \boldsymbol{\beta}) = \partial \mathbf{FP} (c, \boldsymbol{\beta}) / \partial \boldsymbol{\beta} = \partial \mathbb{P}_{\scriptscriptstyle \Tsc}(Y=0,\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta} \geq c) / \partial \boldsymbol{\beta}\,.
\end{aligned}
\end{equation*}
Consequently, $\{\sqrt{n}\{\widehat\ROC(u)-\ROC_0(u)\}:u\in(0,1)\}$ weakly converges to a Gaussian process with mean $0$, and $\sqrt{n}(\widehat{\AUC} - \AUC)$ weakly converges to a normal distribution with mean $0$.
\label{thm:2}
\end{theorem}
Finally, we make several comments on the relationship between the asymptotic efficiency of the estimators $\widehat\boldsymbol{\beta},\widehat\ROC(u)$ and model specification or the unlabeled sample size $N$ in Remark \ref{rem:eff}.
\begin{remark}
Similar to the results of \cite{hahn1998role}, one can show that when both nuisance models are correct, our estimators of the model coefficients and the accuracy parameters attain semiparametric efficiency. While different from doubly robust estimators constructed with $n^{\frac{1}{2}}$-consistent nuisance estimators under low dimensionality studied in previous literature \citep[e.g.]{bang2005doubly,shu2018improved}, the nuisance estimators in our case do not contribute non-negligible variability to $\widehat\boldsymbol{\beta},\widehat\ROC(u)$ even when they are misspecified. This is due to the calibrating procedures introduced to remove the excessive bias caused by them.
When $N\gg n$, i.e., the size of target samples is much larger than the labeling size, influence of the target samples, i.e., ${\rho}N^{-1/2}\sum_{i=n+1}^{n+N} \mathbf{e}_j ^{\scriptscriptstyle \sf T} \mathcal{W}_{{\scriptscriptstyle \Tsc},\beta_j}(\mathbf{D}_i)$ and ${\rho}N^{-1/2}\sum_{i=n+1}^{n+N}\mathcal{W}_{{\scriptscriptstyle \Tsc},{\sf roc}}(u,\mathbf{D}_i)$ will be asymptotically negligible as $\rho\rightarrow 0$. So the asymptotic variance of our estimators can be reduced compared with the case with $N\asymp n$. Similar phenomenons could be found in the recent literature of semi-supervised inference \citep[e.g.]{kawakita2013semi,chakrabortty2018efficient}. Inspired by \cite{gronsbell2018semi}, when $g(\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta})$ is correctly specified on the target population, i.e. $\mathbb{E}_{\scriptscriptstyle \Tsc}[Y\mid\mathbf{A}]=g(\mathbf{A}^{\scriptscriptstyle \sf T} \boldsymbol{\beta}_0)$, the coefficients of the score function $\mathcal{W}_{{\scriptscriptstyle \Ksc}, \beta_j}(\mathbf{D}_i)$ in (\ref{equ:term1:roc}) become zero. That means the variability of $\widehat\boldsymbol{\beta}$ will not contribute to the variability of $\widehat\ROC(u)$.
\label{rem:eff}
\end{remark}
\section{Simulation studies}\label{sec:simu}
We conducted simulation studies to validate the proposed point and interval estimation procedures. We set $q=4$, $p=100$, $N=3000$, and $n=600$ throughout. Let $\mathbf{U} = (U_1,U_2, \ldots , U_{p+q})^{\scriptscriptstyle \sf T}$ be independent variables, where $U_1 = 1$ and each $U_j,j\geq2$ is $N(0, 1)$ truncated to the interval $(2.5, 2.5)$ then standardized to zero mean and unit variance. We generate $Y$ and $S$ (indication of source samples) given $\mathbf{X}$ from Bernoulli distributions independently with the following three data-generating configurations:
\begin{enumerate}
\item[(i)] Both nuisance models are correctly specified:
\[\mathbb{P}(Y=1 \mid \mathbf{X}) = {\rm expit}(\mathbf{X}^{\scriptscriptstyle \sf T}\boldsymbol{\gamma})\; \text{ and } \; \mathbb{P}(S=1 \mid \mathbf{X}) = {\rm expit}(\mathbf{X}^{\scriptscriptstyle \sf T}\boldsymbol{\alpha})
\,,\]
where $X_4 = \mathbb{I}(U_1 > 0) U_1 + U_4$, $X_5 = \mathbb{I}(U_2 > 0) U_2 + 2U_5$, $X_j = U_j, j \neq 4,5$, $\boldsymbol{\gamma} = (0,0.6,-0.5,0.5,0,1,0.75,\mathbf{0}_{p-3}^{\scriptscriptstyle \sf T})^{\scriptscriptstyle \sf T}$ and $\boldsymbol{\alpha} = (0, 0.4, -0.5,0,0,0,0.75,\mathbf{0}_{p-3})^{\scriptscriptstyle \sf T}$.
\item[(ii)] The density ratio model is correct while the imputation model is misspecified:
\[\mathbb{P}(Y=1 \mid \mathbf{X}) = {\rm expit}(\boldsymbol{\Phi}^{\scriptscriptstyle \sf T}\boldsymbol{\gamma})\; \text{ and } \; \mathbb{P}(S=1 \mid \mathbf{X}) = {\rm expit}(\mathbf{X}^{\scriptscriptstyle \sf T}\boldsymbol{\alpha})
\,,\]
where $\mathbf{X} = \mathbf{U}$, $\boldsymbol{\Phi} = (1,|X_2|,X_2,X_3)^{\scriptscriptstyle \sf T}$, $\boldsymbol{\gamma} = (0.8,0.5,0.3,0.5)^{\scriptscriptstyle \sf T}$ and $$\boldsymbol{\alpha} = (0,0.6,-0.6,0.5,0,5,0.5,\mathbf{0}_{p-2}^{\scriptscriptstyle \sf T})^{\scriptscriptstyle \sf T}.$$
\item[(iii)] The imputation model is correct while the density ratio model is misspecified:
\[\mathbb{P}(Y=1 \mid \mathbf{X}) = {\rm expit}(\mathbf{X}^{\scriptscriptstyle \sf T}\boldsymbol{\gamma})\; \text{ and } \; \mathbb{P}(S=1 \mid \mathbf{X}) = {\rm expit}(\boldsymbol{\Phi}^{\scriptscriptstyle \sf T}\boldsymbol{\alpha})
\,,\]
where $\mathbf{X} = \mathbf{U}$, $\boldsymbol{\gamma} = (0,0.6,-0.5,0.5,0,1,0.75,\mathbf{0}_{p-3}^{\scriptscriptstyle \sf T})^{\scriptscriptstyle \sf T}$, $\boldsymbol{\Phi} = (1,X_2 X_5, X_3 X_6)^{\scriptscriptstyle \sf T}$ and $\boldsymbol{\alpha} = (0.4,0.4,0.25)^{\scriptscriptstyle \sf T}$.
\end{enumerate}
We compare our DRAMATIC method with two existing methods: (i) the importance weighting method (IW) detailed in Algorithm \ref{alg iw} and (ii) the imputation only method (IM) detailed in Algorithm \ref{alg im} presented in Appendix \ref{app:B}. When using our approach, we set $n_{\min}=120$, select the penalty parameters $\lambda_{\alpha}$ and $\lambda_{\gamma}$ using cross-validation from the range $[0.01(\log p/n)^{\frac{1}{2}},0.5(\log p/n)^{\frac{1}{2}}]$, and select $\lambda_{\beta,\alpha}$, $\lambda_{\beta,\gamma}$, $\lambda_{{\sf roc},\alpha,\widehat{c}_j}$ and $\lambda_{{\sf roc},\gamma,\widehat{c}_j}$ using the tuning procedure introduced in Section \ref{sec:method:tune}.
We report the average bias (Bias), root of average mean square error (rMSE), and average coverage probability (CP) of $95\%$ CI for the estimators based on $1000$ repetitions for each configuration. Results of estimating $\boldsymbol{\beta}_0$ are presented in Table~\ref{tab:beta} and those of AUC and ROC evaluated at two important points, $0.1$ and $0.2$, are presented in Table~\ref{tab:auc}. Table~\ref{tab:beta} shows that DRAMATIC performs consistently well in estimating $\boldsymbol{\beta}$. The interval estimators of DRAMATIC have proper coverage probabilities located in the $\pm0.05$ range of the nominal level. Under Configuration (ii), when the imputation model is misspecified, IM has large biases for some coefficients. For example, it has absolute bias $0.24$ for $\beta_3$, while DRAMATIC has biases smaller than $0.03$. Under Configuration (iii), when the density ratio model is misspecified, IW yields an unsatisfactory bias on $\beta_3$, which is more than $80\%$ of the MSE. On the contrary, the biases of DRAMATIC do not exceed $0.03$. Throughout the three configurations, the biases of DRAMATIC are smaller than $23\%$ of the rMSE for all coefficients $\beta_j$'s, while the performances of IM and IW are not robust, especially when the models are misspecified. Although they have low bias and rMSE under some cases, the bias of IM can be $95\%$ of the rMSE for $\beta_3$ in Configuration (ii) and of IW can be $90\%$ of the rMSE for $\beta_3$ in Configuration (iii). Even under Configuration (i), when both nuisance models are correct, DRAMATIC is still the most efficient method achieving the smallest rMSE for all $\beta_j$'s due to its debiasing correction. For example, DRAMATIC attains $169\%$ relative efficiency compared to IM on $\beta_1$ and $160\%$ relative efficiency compared to IW on $\beta_2$.
From Table~\ref{tab:auc}, we see that DRAMATIC achieves the smallest bias and rMSE for almost all model accuracy parameters in the three configurations. In addition, under all the settings, our interval estimators have proper coverage probabilities located in $\pm0.02$.
Under Configuration (i), DRAMATIC yields absolute biases no larger than $0.004$, while IM and IW can produce absolute bias larger than $0.02$, exceeding five times of DRAMATIC. Under Configurations (ii) and (iii), the performances of IM and IW are getting even worse, with absolute biases arriving at about $0.05$. Although the model of IM is correct under Configuration (ii) and the model of IW is correct under Configuration (iii), they are still not able to produce estimators with low bias due the excessive errors of the high-dimensional nuisance models. For example, under Configuration (ii), the bias of IM accounts for $68\%$ of the rMSE on AUC and $81\%$ of the rMSE on ROC($0.1$). Under Configuration (iii), the bias of IW accounts for $53\%$ of its rMSE on AUC. On the contrary, under all the configurations, DRAMATIC outputs a low bias of no more than $2\%$ of the magnitude of the true parameters and only occupy a small proportion of its rMSE. Thus, DRAMATIC is robust and efficient in model estimation and evaluation as long as at least one nuisance model is correctly specified and substantially outperforms existing choices.
\begin{table}[t]
\centering
\caption{\label{tab:beta} Average bias (Bias), root of average mean square error (rMSE), and average coverage probability (CP) of $95\%$ CI for the estimations of $\boldsymbol{\beta}_0$ with the sample size $n = 600$ and $N=3000$ under Configurations (i), (ii) and (iii) with $1000$ repetition for each configuration.}
\begin{tabular}{c l|cc|cc|ccc}
\hline \hline
\multirow{2}{*}{ Configuration } & & \multicolumn{2}{c|}{IM} & \multicolumn{2}{c|}{IW} & \multicolumn{3}{c}{DRAMATIC} \\ \cline{3-9}
& & Bias & rMSE & Bias & rMSE & Bias & rMSE & CP \\
\hline
\multirow{4}{*}{ (i)} & $\beta_1$ & 0.086 & 0.121 & 0.044 & 0.108 & \textbf{0.021} & \textbf{0.093} & 0.917 \\
& $\beta_2$ & -0.080 & 0.108 & \textbf{0.002} & 0.129 & -0.008 & \textbf{0.102} & 0.919 \\
& $\beta_3$ & 0.077 & 0.105 & \textbf{-0.001} & 0.124 & 0.016 & \textbf{0.099} & 0.927 \\
& $\beta_4$ & -0.074 & 0.102 & \textbf{0.002} & 0.120 & -0.008 & \textbf{0.093} & 0.922 \\ \hline
\multirow{4}{*}{(ii)} & $\beta_1$ & -0.013 & \textbf{0.100} & \textbf{0.003} & 0.117 & -0.008 & 0.120 & 0.944 \\
& $\beta_2$ & 0.087 & \textbf{0.126} & 0.036 & 0.163 & \textbf{0.028} & 0.162 & 0.902 \\
& $\beta_3$ & -0.244 & 0.256 & \textbf{-0.022} & 0.150 & \textbf{-0.022} & \textbf{0.148} & 0.927 \\
& $\beta_4$ & -0.089 & \textbf{0.128} & 0.012 & 0.153 & \textbf{0.004} & 0.154 & 0.934 \\ \hline
\multirow{4}{*}{ (iii)} & $\beta_1$ & \textbf{-0.002} & \textbf{0.080} & -0.003 & 0.086 & \textbf{-0.002} & 0.084 & 0.925\\
& $\beta_2$ & -0.089 & 0.119 & \textbf{-0.008} & 0.090 & -0.011 & \textbf{0.086} & 0.936 \\
& $\beta_3$ & 0.107 & 0.134 & 0.191 & 0.213 & \textbf{0.031} & \textbf{0.092} & 0.916 \\
& $\beta_4$ & -0.079 & 0.112 & \textbf{-0.003} & 0.091 & -0.004 & \textbf{0.086} & 0.931 \\ \hline
\end{tabular}
\end{table}
\begin{table}[t]
\centering
\caption{\label{tab:auc} Average bias (Bias), root of average mean square error (rMSE), and average coverage probability (CP) of $95\%$ CI for the estimations of AUC and ROC with the sample size $n = 600$ and $N=3000$ under Configurations (i), (ii) and (iii) with $1000$ repetition for each configuration.}
\begin{tabular}{c l|cc|cc|ccc}
\hline \hline
\multirow{2}{*}{Configuration } & & \multicolumn{2}{c|}{IM} & \multicolumn{2}{c|}{IW} & \multicolumn{3}{c}{DRAMATIC} \\ \cline{3-9}
& & Bias & rMSE & Bias & rMSE & Bias & rMSE & CP \\
\hline
\multirow{3}{*}{ (i)} & AUC & -0.018 & 0.025 & \textbf{-0.003} & 0.033 & 0.004 & \textbf{0.024} & 0.932 \\
& ROC(0.1) & 0.018 & 0.057 & 0.021 & 0.058 & \textbf{-0.004} & \textbf{0.036} & 0.971 \\
& ROC(0.2) & -0.021 & \textbf{0.035} & -0.018 & 0.037 & \textbf{0.001} & 0.043 & 0.954 \\ \hline
\multirow{3}{*}{ (ii)} & AUC & -0.033 & 0.040 & -0.009 & 0.038 & \textbf{0.004} & \textbf{0.032} & 0.931 \\
& ROC(0.1) & -0.054 & 0.060 & -0.050 & 0.059 & \textbf{0.008} & \textbf{0.056} & 0.947 \\
& ROC(0.2) & -0.052 & 0.061 & -0.049 & 0.061 & \textbf{0.014} & \textbf{0.060} & 0.953 \\ \hline
\multirow{3}{*}{(iii)} & AUC & -0.029 & 0.034 & -0.024 & 0.033 & \textbf{0.000} & \textbf{0.019} & 0.925 \\
& ROC(0.1) & -0.040 & 0.047 & -0.040 & 0.048 & \textbf{-0.007} & \textbf{0.039} & 0.938 \\
& ROC(0.2) & -0.042 & 0.051 & -0.042 & 0.053 & \textbf{-0.001} & \textbf{0.039} & 0.946 \\ \hline
\end{tabular}
\end{table}
\section{Application to electronic medical records studies}\label{sec:data analysis}
Type II diabetes (T2D), mainly caused by insulin resistance or insufficiency, is a common disease with an increasing prevalence over recent years. Existing genetic studies and genome-wide association studies (GWAS) have detected and identified hundreds of genetic variants associated with T2D, which enables the use of polygenic risk score (GRS) for T2D risk prediction \citep{he2021comparisons} and thus brings up the possibility of precision medicine.
We base on the Mass General Brigham (MGB) biobank data \citep{castro2022mass} to construct and evaluate the risk prediction model of T2D against GRS and basic demographic information. MGB biobank links the electronic health record (EHR) features of the enrolled subjects with their whole genome sequencing data, and thus provides the possibility of identifying the T2D status from the EHR and modeling it against the gene variants. However, in our study, there arises two main challenges including (i) distributional shift of T2D status and its relevant EHR features between the source and the target population, and (ii) high dimensionality of these EHR features.
In specific, the source samples are $n=449$ subjects collected and labeled with T2D status $Y\in\{0,1\}$ through manual chart review in 2014. To create labels for multiple phenotypes efficiently and simultaneously, patients with several other disease codes like coronary artery disease were set with larger probabilities of being sampled, which makes the sampling not completely random. While the target samples are taken as $N=2000$ subjects randomly sampled from the biobank data set collected in 2021. Both the sampling scheme and the different time window of data collection can cause shift of EHR features between the source and target samples, because around 2015, the EHR system at MGB was switched and the International Classification of Diseases (ICD) system was changed from version 9 to version 10. We aim at estimating and evaluating the logistic model of $Y\sim \mathbf{A}=(1,A_1,A_2,A_3)$ on the target population, where $A_1$ is the GRS of T2D, $A_2$ is indication of Female, $A_3$ is indication of European. To create the GRS of T2D, we utilize the GWAS results of \cite{mahajan2018fine} to firstly choose single nucleotide polymorphisms (SNPs) significantly associated with T2D (with $p$-value below $10^{-10}$), prune the highly corrected SNPs (due to linkage disequilibrium), and then combine the selected 65 SNPs weighted by their beta coefficients. The adjustment features $\mathbf{X}$ are taken as the union of $\mathbf{A}$ and the health-utilization-adjusted (through partial linear regression) log-counts ($u\xrightarrow{}\log(u+1)$) of $p=143$ T2D-relevant diagnostic and procedure codes selected among the whole EHR features using the clinical knowledge extraction approach proposed by \cite{hong2021clinical}, independently from the source and target samples. To validate the transfer learning and evaluation approaches, we randomly sample and label $267$ subjects from the target population to obtain a validation data set.
Inspired by \cite{LiuandZhang}, we fit $Y \sim \mathbf{A}$ using the validation data to obtain $\widehat\boldsymbol{\beta}_{\rm val}$ as a ``gold standard" benchmark and use the following metrics to evaluate the performance of a transfer learning estimator $\widehat \boldsymbol{\beta}$:
\begin{itemize}
\item[] \textbf{rMSPE.} Root of mean square prediction error against $\widehat\boldsymbol{\beta}_{\rm val}$ evaluated on the target data:
$$\left[{\widehat\Psc_{\mathcal{T}}\{g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat\boldsymbol{\beta}_{\rm val} )- g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat\boldsymbol{\beta})\}^2}\right]^{\frac{1}{2}} \;.$$
\item[] \textbf{CC} with $\widehat\boldsymbol{\beta}_{\rm val}$. Classifier’s correlation with that of $\widehat\boldsymbol{\beta}_{\rm val}$:
$$
{\rm Corr}_{\mathcal{T}}\big\{ \mathbb{I}( g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat\boldsymbol{\beta}_{\rm val}) \geq \widehat\Psc_{\mathcal{T}} g(\mathbf{A}_{i'}^{\scriptscriptstyle \sf T} \widehat\boldsymbol{\beta}_{\rm val} )), \mathbb{I}( g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}) \geq \widehat\Psc_{\mathcal{T}} g(\mathbf{A}_{i'}^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta})) \big\} \,,
$$
where ${\rm Corr}_{\mathcal{T}}(\cdot,\cdot)$ represents the empirical Pearson correlation on the target population.
\item[] \textbf{FCR} v.s. $\widehat\boldsymbol{\beta}_{\rm val}$. False classification rate of $\widehat \boldsymbol{\beta}$'s classifier against that of $\widehat\boldsymbol{\beta}_{\rm val}$:
$$\widehat\Psc_{\mathcal{T}}\big( \mathbb{I}( g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat\boldsymbol{\beta}_{\rm val}) \geq \widehat\Psc_{\mathcal{T}} g(\mathbf{A}_{i'}^{\scriptscriptstyle \sf T} \widehat\boldsymbol{\beta}_{\rm val}) ) \neq \mathbb{I}( g(\mathbf{A}_i^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta}) \geq \widehat\Psc_{\mathcal{T}} g(\mathbf{A}_{i'}^{\scriptscriptstyle \sf T} \widehat \boldsymbol{\beta})) \big) \,.$$
\end{itemize}
To evaluate the performance of estimating the ROC curve, we estimate the ROC and AUC of $\mathbf{A}^{\scriptscriptstyle \sf T}\boldsymbol{\beta}$ on the validation data by cross-validation and use it as the ``gold standard". The total variation (TV) of any estimated ROC curve to this gold standard curve is calculated as the evaluation metric. Methods under comparison include our proposed DRAMATIC approach, the IW and IM approaches introduced in Algorithms \ref{alg iw} and \ref{alg im}, and the naive source estimation obtained by fitting and evaluating the regression model simply using the labeled source samples.
Estimation of $\boldsymbol{\beta}$ obtained using these included methods are presented in Table \ref{tab:real beta} of Appendix \ref{app:B}, and their performance evaluated against the validation estimator $\widehat\boldsymbol{\beta}_{\rm val}$ (referred as ``Target") presented in Table \ref{tab:eval:beta}. Among all methods under comparison, the DRAMATIC estimator achieves the closest estimated prevalence, the smallest rMSPE, the largest classifier’s correlation, and the smallest false classification rate evaluated against the validation estimator. For example, in terms of rMSPE, our estimator has a $68\%$ smaller error than Source, $54\%$ smaller than IW, and $34\%$ smaller than IM.
We further plotted the estimated ROC curves in Figure \ref{fig:roc}. One can see that the ROC curve estimated using DRAMATIC (the red dashed line) is the closest one to the validation benchmark (the black solid line) among all methods. This is strictly verified by the total variation (TV) distance from the ``Target" ROC estimator presented in Table \ref{tab:real:roc}. In specific, the TV error of our approach is $41\%$ smaller than that of the Source estimator, $58\%$ smaller than IW, and $32\%$ smaller than IM. As shown in Table \ref{tab:real:roc}, our approach also produces the closest estimation in AUC, ROC$(0.1)$ and ROC$(0.2)$ to the Target than other transfer learning or source estimators. In addition, our 95\% CI estimators of AUC and ROC attain moderate and reasonable length and correctly cover the validation estimators. All these promising results illustrate that DRAMATIC is more robust and efficient than existing methods in transfer learning and model evaluation under the distributional shift of high dimensional adjustment features. This conclusion is also consistent with our asymptotic analysis and simulation studies.
\begin{table}[htpb]
\centering
\caption{Estimation performance on $\boldsymbol{\beta}$ of the source or transfer learning estimators evaluated with the validation labels and the validation estimator. Prevalence: mean of the estimated risk probabilities $g(\mathbf{A}^{\scriptscriptstyle \sf T} \widehat\boldsymbol{\beta})$ on the target samples. rMSPE: root of mean square prediction error; CC with $ \widehat\boldsymbol{\beta}_{\rm val}$: classifier's correlation with that of $\widehat\boldsymbol{\beta}_{\rm val}$; FCR v.s. $ \widehat\boldsymbol{\beta}_{\rm val}$: false classification rate against $ \widehat\boldsymbol{\beta}_{\rm val}$. Source: naive source data estimator; DRAMATIC: our proposed doubly robust method; IW: importance weighting method; IM: imputation only method; Target: gold standard benchmark obtained using the validation samples.}
\label{tab:eval:beta}
\begin{tabular}{ccccccc}
\hline
\hline\\[-1.8ex]
& Source & DRAMATIC & IW & IM & Target \\
\hline\\[-1.8ex]
Prevalence & 0.115 & 0.173 & 0.141 & 0.173 & 0.195 \\
rMSPE & 0.091 & {\bf 0.029} & 0.063 & 0.044 & 0 \\
CC with $ \widehat\boldsymbol{\beta}_{\rm val}$ & 0.365 & {\bf 0.733} & 0.507 & 0.686 & 1 \\
FCR v.s. $ \widehat\boldsymbol{\beta}_{\rm val}$ & 0.419 & {\bf 0.154} & 0.311 & 0.165 & 0 \\
\hline\hline\\[-1.8ex]
\end{tabular}
\end{table}
\begin{figure}[ht]
\centering
\caption{\label{fig:roc} Estimated ROC curves. Source: naive source data estimator; DRAMATIC: our proposed doubly robust method; IW: importance weighting method; IM: imputation based method; Target: benchmark obtained using the validation samples.}
\includegraphics[width=0.7\linewidth]{roc.pdf}
\end{figure}
\begin{table}[htpb]
\centering
\caption{\label{tab:real:roc} Total variation distance from the validation estimator (TV from Target) of the ROC curve, and estimators of AUC, ROC$(0.1)$ and ROC$(0.2)$, attached with their 95\% CIs constructed using our method. Source: naive source data estimator; DRAMATIC: our proposed doubly robust method; IW: importance weighting method; IM: imputation based method; Target: benchmark obtained using the validation samples.}
\label{tab:roc}
\begin{tabular}{ccccccc}
\hline
\hline
& Source & DRAMATIC (95\% CI) & IW & IM & Target \\ \hline
TV from Target & 0.032 & {\bf 0.019} & 0.045 & 0.028 & 0 \\
AUC & 0.677 & 0.671 (0.622, 0.720) & 0.673 & 0.626 & 0.656 \\
ROC$(0.1)$ & 0.173 & 0.217 (0.152, 0.281) & 0.152 & 0.201 & 0.218\\
ROC$(0.2)$ & 0.346 & 0.390 (0.310, 0.470) & 0.331 & 0.354 & 0.388\\
\hline\hline\\[-1.8ex]
\end{tabular}
\end{table}
\section{Discussion}
In this paper, we propose DRAMATIC, a transfer
learning approach to estimate the coefficients and evaluate the ROC of the logistic model on a target population with no observations of $Y$ under the distributional shift of high dimensional covariates. Our method is doubly robust in the sense that either a correctly specified density ratio model or a correct imputation model can lead to $n^{\frac{1}{2}}$-consistent estimators under certain sparsity conditions. The main challenge of constructing a doubly robust estimator of the ROC curve under the high dimensional nuisance model regime is to remove the excessive regularization bias for all points on the ROC curve simultaneously. We realize this through a novel ``nearest quantile" strategy that is statistically effective and computationally efficient, as introduced in Section \ref{sec:cal:roc}. In addition, unlike existing doubly robust estimators with high dimensional nuisance models, our method does not require the calibrated nuisance parameters to be exactly sparse. Instead, we make assumptions on their $l_r$ norms for arbitrary $r\in [0,1]$, which is more general and reasonable than the $l_0$ sparsity.
We note that other evaluation prediction and classification metrics, such as the overall misclassification
rate and the Brier score \citep[e.g.]{gronsbell2020efficient} can be further considered in our setting. Our high dimensional calibrated method for curve estimator involving step functions could be potentially used in other settings such as the diagnose curve evaluation of the survival models. Besides, we target a classification model with the dimension of $\mathbf{A}$, $q$ being low and fixed. While there are also real-world interests in transferring evaluation of a high dimensional sparse model where $q$ grows with $n$ and $N$. This is an open and challenging problem due to the excessive regularization error of $\boldsymbol{\beta}_0$ in the high dimensional case. Finally, we mainly focus on the robustness of our estimator but pay less attention to its asymptotic efficiency. It is of independent interest to explore the efficiency of our estimator when $N \gg n$, which is usually referred to as semi-supervised learning. In this scenario, existing results have been established when covariates in the unlabeled and labeled data sets have the same distribution \citep[e.g.]{kawakita2013semi,gronsbell2018semi,chakrabortty2018efficient,gronsbell2020efficient,azriel2021semi}. When there exists covariate shift, one may justify the efficiency gain of having $N \gg n$ in a similar way.
\bibliographystyle{apalike}
|
Subsets and Splits